OpenCloudOS-Kernel/net/rxrpc/input.c

1427 lines
37 KiB
C
Raw Normal View History

// SPDX-License-Identifier: GPL-2.0-or-later
/* RxRPC packet reception
*
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
* Copyright (C) 2007, 2016 Red Hat, Inc. All Rights Reserved.
* Written by David Howells (dhowells@redhat.com)
*/
#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
#include <linux/module.h>
#include <linux/net.h>
#include <linux/skbuff.h>
#include <linux/errqueue.h>
#include <linux/udp.h>
#include <linux/in.h>
#include <linux/in6.h>
#include <linux/icmp.h>
include cleanup: Update gfp.h and slab.h includes to prepare for breaking implicit slab.h inclusion from percpu.h percpu.h is included by sched.h and module.h and thus ends up being included when building most .c files. percpu.h includes slab.h which in turn includes gfp.h making everything defined by the two files universally available and complicating inclusion dependencies. percpu.h -> slab.h dependency is about to be removed. Prepare for this change by updating users of gfp and slab facilities include those headers directly instead of assuming availability. As this conversion needs to touch large number of source files, the following script is used as the basis of conversion. http://userweb.kernel.org/~tj/misc/slabh-sweep.py The script does the followings. * Scan files for gfp and slab usages and update includes such that only the necessary includes are there. ie. if only gfp is used, gfp.h, if slab is used, slab.h. * When the script inserts a new include, it looks at the include blocks and try to put the new include such that its order conforms to its surrounding. It's put in the include block which contains core kernel includes, in the same order that the rest are ordered - alphabetical, Christmas tree, rev-Xmas-tree or at the end if there doesn't seem to be any matching order. * If the script can't find a place to put a new include (mostly because the file doesn't have fitting include block), it prints out an error message indicating which .h file needs to be added to the file. The conversion was done in the following steps. 1. The initial automatic conversion of all .c files updated slightly over 4000 files, deleting around 700 includes and adding ~480 gfp.h and ~3000 slab.h inclusions. The script emitted errors for ~400 files. 2. Each error was manually checked. Some didn't need the inclusion, some needed manual addition while adding it to implementation .h or embedding .c file was more appropriate for others. This step added inclusions to around 150 files. 3. The script was run again and the output was compared to the edits from #2 to make sure no file was left behind. 4. Several build tests were done and a couple of problems were fixed. e.g. lib/decompress_*.c used malloc/free() wrappers around slab APIs requiring slab.h to be added manually. 5. The script was run on all .h files but without automatically editing them as sprinkling gfp.h and slab.h inclusions around .h files could easily lead to inclusion dependency hell. Most gfp.h inclusion directives were ignored as stuff from gfp.h was usually wildly available and often used in preprocessor macros. Each slab.h inclusion directive was examined and added manually as necessary. 6. percpu.h was updated not to include slab.h. 7. Build test were done on the following configurations and failures were fixed. CONFIG_GCOV_KERNEL was turned off for all tests (as my distributed build env didn't work with gcov compiles) and a few more options had to be turned off depending on archs to make things build (like ipr on powerpc/64 which failed due to missing writeq). * x86 and x86_64 UP and SMP allmodconfig and a custom test config. * powerpc and powerpc64 SMP allmodconfig * sparc and sparc64 SMP allmodconfig * ia64 SMP allmodconfig * s390 SMP allmodconfig * alpha SMP allmodconfig * um on x86_64 SMP allmodconfig 8. percpu.h modifications were reverted so that it could be applied as a separate patch and serve as bisection point. Given the fact that I had only a couple of failures from tests on step 6, I'm fairly confident about the coverage of this conversion patch. If there is a breakage, it's likely to be something in one of the arch headers which should be easily discoverable easily on most builds of the specific arch. Signed-off-by: Tejun Heo <tj@kernel.org> Guess-its-ok-by: Christoph Lameter <cl@linux-foundation.org> Cc: Ingo Molnar <mingo@redhat.com> Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
2010-03-24 16:04:11 +08:00
#include <linux/gfp.h>
#include <net/sock.h>
#include <net/af_rxrpc.h>
#include <net/ip.h>
#include <net/udp.h>
#include <net/net_namespace.h>
#include "ar-internal.h"
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
static void rxrpc_proto_abort(const char *why,
struct rxrpc_call *call, rxrpc_seq_t seq)
{
if (rxrpc_abort_call(why, call, seq, RX_PROTOCOL_ERROR, -EBADMSG)) {
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
set_bit(RXRPC_CALL_EV_ABORT, &call->events);
rxrpc_queue_call(call);
}
}
/*
* Do TCP-style congestion management [RFC 5681].
*/
static void rxrpc_congestion_management(struct rxrpc_call *call,
struct sk_buff *skb,
struct rxrpc_ack_summary *summary,
rxrpc_serial_t acked_serial)
{
enum rxrpc_congest_change change = rxrpc_cong_no_change;
unsigned int cumulative_acks = call->cong_cumul_acks;
unsigned int cwnd = call->cong_cwnd;
bool resend = false;
summary->flight_size =
(call->tx_top - call->tx_hard_ack) - summary->nr_acks;
if (test_and_clear_bit(RXRPC_CALL_RETRANS_TIMEOUT, &call->flags)) {
summary->retrans_timeo = true;
call->cong_ssthresh = max_t(unsigned int,
summary->flight_size / 2, 2);
cwnd = 1;
if (cwnd >= call->cong_ssthresh &&
call->cong_mode == RXRPC_CALL_SLOW_START) {
call->cong_mode = RXRPC_CALL_CONGEST_AVOIDANCE;
call->cong_tstamp = skb->tstamp;
cumulative_acks = 0;
}
}
cumulative_acks += summary->nr_new_acks;
cumulative_acks += summary->nr_rot_new_acks;
if (cumulative_acks > 255)
cumulative_acks = 255;
summary->mode = call->cong_mode;
summary->cwnd = call->cong_cwnd;
summary->ssthresh = call->cong_ssthresh;
summary->cumulative_acks = cumulative_acks;
summary->dup_acks = call->cong_dup_acks;
switch (call->cong_mode) {
case RXRPC_CALL_SLOW_START:
if (summary->nr_nacks > 0)
goto packet_loss_detected;
if (summary->cumulative_acks > 0)
cwnd += 1;
if (cwnd >= call->cong_ssthresh) {
call->cong_mode = RXRPC_CALL_CONGEST_AVOIDANCE;
call->cong_tstamp = skb->tstamp;
}
goto out;
case RXRPC_CALL_CONGEST_AVOIDANCE:
if (summary->nr_nacks > 0)
goto packet_loss_detected;
/* We analyse the number of packets that get ACK'd per RTT
* period and increase the window if we managed to fill it.
*/
if (call->peer->rtt_usage == 0)
goto out;
if (ktime_before(skb->tstamp,
ktime_add_ns(call->cong_tstamp,
call->peer->rtt)))
goto out_no_clear_ca;
change = rxrpc_cong_rtt_window_end;
call->cong_tstamp = skb->tstamp;
if (cumulative_acks >= cwnd)
cwnd++;
goto out;
case RXRPC_CALL_PACKET_LOSS:
if (summary->nr_nacks == 0)
goto resume_normality;
if (summary->new_low_nack) {
change = rxrpc_cong_new_low_nack;
call->cong_dup_acks = 1;
if (call->cong_extra > 1)
call->cong_extra = 1;
goto send_extra_data;
}
call->cong_dup_acks++;
if (call->cong_dup_acks < 3)
goto send_extra_data;
change = rxrpc_cong_begin_retransmission;
call->cong_mode = RXRPC_CALL_FAST_RETRANSMIT;
call->cong_ssthresh = max_t(unsigned int,
summary->flight_size / 2, 2);
cwnd = call->cong_ssthresh + 3;
call->cong_extra = 0;
call->cong_dup_acks = 0;
resend = true;
goto out;
case RXRPC_CALL_FAST_RETRANSMIT:
if (!summary->new_low_nack) {
if (summary->nr_new_acks == 0)
cwnd += 1;
call->cong_dup_acks++;
if (call->cong_dup_acks == 2) {
change = rxrpc_cong_retransmit_again;
call->cong_dup_acks = 0;
resend = true;
}
} else {
change = rxrpc_cong_progress;
cwnd = call->cong_ssthresh;
if (summary->nr_nacks == 0)
goto resume_normality;
}
goto out;
default:
BUG();
goto out;
}
resume_normality:
change = rxrpc_cong_cleared_nacks;
call->cong_dup_acks = 0;
call->cong_extra = 0;
call->cong_tstamp = skb->tstamp;
if (cwnd < call->cong_ssthresh)
call->cong_mode = RXRPC_CALL_SLOW_START;
else
call->cong_mode = RXRPC_CALL_CONGEST_AVOIDANCE;
out:
cumulative_acks = 0;
out_no_clear_ca:
if (cwnd >= RXRPC_RXTX_BUFF_SIZE - 1)
cwnd = RXRPC_RXTX_BUFF_SIZE - 1;
call->cong_cwnd = cwnd;
call->cong_cumul_acks = cumulative_acks;
trace_rxrpc_congest(call, summary, acked_serial, change);
if (resend && !test_and_set_bit(RXRPC_CALL_EV_RESEND, &call->events))
rxrpc_queue_call(call);
return;
packet_loss_detected:
change = rxrpc_cong_saw_nack;
call->cong_mode = RXRPC_CALL_PACKET_LOSS;
call->cong_dup_acks = 0;
goto send_extra_data;
send_extra_data:
/* Send some previously unsent DATA if we have some to advance the ACK
* state.
*/
if (call->rxtx_annotations[call->tx_top & RXRPC_RXTX_BUFF_MASK] &
RXRPC_TX_ANNO_LAST ||
summary->nr_acks != call->tx_top - call->tx_hard_ack) {
call->cong_extra++;
wake_up(&call->waitq);
}
goto out_no_clear_ca;
}
/*
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
* Apply a hard ACK by advancing the Tx window.
*/
static bool rxrpc_rotate_tx_window(struct rxrpc_call *call, rxrpc_seq_t to,
struct rxrpc_ack_summary *summary)
{
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
struct sk_buff *skb, *list = NULL;
bool rot_last = false;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
int ix;
rxrpc: Pass the last Tx packet marker in the annotation buffer When the last packet of data to be transmitted on a call is queued, tx_top is set and then the RXRPC_CALL_TX_LAST flag is set. Unfortunately, this leaves a race in the ACK processing side of things because the flag affects the interpretation of tx_top and also allows us to start receiving reply data before we've finished transmitting. To fix this, make the following changes: (1) rxrpc_queue_packet() now sets a marker in the annotation buffer instead of setting the RXRPC_CALL_TX_LAST flag. (2) rxrpc_rotate_tx_window() detects the marker and sets the flag in the same context as the routines that use it. (3) rxrpc_end_tx_phase() is simplified to just shift the call state. The Tx window must have been rotated before calling to discard the last packet. (4) rxrpc_receiving_reply() is added to handle the arrival of the first DATA packet of a reply to a client call (which is an implicit ACK of the Tx phase). (5) The last part of rxrpc_input_ack() is reordered to perform Tx rotation, then soft-ACK application and then to end the phase if we've rotated the last packet. In the event of a terminal ACK, the soft-ACK application will be skipped as nAcks should be 0. (6) rxrpc_input_ackall() now has to rotate as well as ending the phase. In addition: (7) Alter the transmit tracepoint to log the rotation of the last packet. (8) Remove the no-longer relevant queue_reqack tracepoint note. The ACK-REQUESTED packet header flag is now set as needed when we actually transmit the packet and may vary by retransmission. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-23 19:39:22 +08:00
u8 annotation;
if (call->acks_lowest_nak == call->tx_hard_ack) {
call->acks_lowest_nak = to;
} else if (before_eq(call->acks_lowest_nak, to)) {
summary->new_low_nack = true;
call->acks_lowest_nak = to;
}
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
spin_lock(&call->lock);
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
while (before(call->tx_hard_ack, to)) {
call->tx_hard_ack++;
ix = call->tx_hard_ack & RXRPC_RXTX_BUFF_MASK;
skb = call->rxtx_buffer[ix];
rxrpc: Pass the last Tx packet marker in the annotation buffer When the last packet of data to be transmitted on a call is queued, tx_top is set and then the RXRPC_CALL_TX_LAST flag is set. Unfortunately, this leaves a race in the ACK processing side of things because the flag affects the interpretation of tx_top and also allows us to start receiving reply data before we've finished transmitting. To fix this, make the following changes: (1) rxrpc_queue_packet() now sets a marker in the annotation buffer instead of setting the RXRPC_CALL_TX_LAST flag. (2) rxrpc_rotate_tx_window() detects the marker and sets the flag in the same context as the routines that use it. (3) rxrpc_end_tx_phase() is simplified to just shift the call state. The Tx window must have been rotated before calling to discard the last packet. (4) rxrpc_receiving_reply() is added to handle the arrival of the first DATA packet of a reply to a client call (which is an implicit ACK of the Tx phase). (5) The last part of rxrpc_input_ack() is reordered to perform Tx rotation, then soft-ACK application and then to end the phase if we've rotated the last packet. In the event of a terminal ACK, the soft-ACK application will be skipped as nAcks should be 0. (6) rxrpc_input_ackall() now has to rotate as well as ending the phase. In addition: (7) Alter the transmit tracepoint to log the rotation of the last packet. (8) Remove the no-longer relevant queue_reqack tracepoint note. The ACK-REQUESTED packet header flag is now set as needed when we actually transmit the packet and may vary by retransmission. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-23 19:39:22 +08:00
annotation = call->rxtx_annotations[ix];
rxrpc_see_skb(skb, rxrpc_skb_rotated);
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
call->rxtx_buffer[ix] = NULL;
call->rxtx_annotations[ix] = 0;
skb->next = list;
list = skb;
rxrpc: Pass the last Tx packet marker in the annotation buffer When the last packet of data to be transmitted on a call is queued, tx_top is set and then the RXRPC_CALL_TX_LAST flag is set. Unfortunately, this leaves a race in the ACK processing side of things because the flag affects the interpretation of tx_top and also allows us to start receiving reply data before we've finished transmitting. To fix this, make the following changes: (1) rxrpc_queue_packet() now sets a marker in the annotation buffer instead of setting the RXRPC_CALL_TX_LAST flag. (2) rxrpc_rotate_tx_window() detects the marker and sets the flag in the same context as the routines that use it. (3) rxrpc_end_tx_phase() is simplified to just shift the call state. The Tx window must have been rotated before calling to discard the last packet. (4) rxrpc_receiving_reply() is added to handle the arrival of the first DATA packet of a reply to a client call (which is an implicit ACK of the Tx phase). (5) The last part of rxrpc_input_ack() is reordered to perform Tx rotation, then soft-ACK application and then to end the phase if we've rotated the last packet. In the event of a terminal ACK, the soft-ACK application will be skipped as nAcks should be 0. (6) rxrpc_input_ackall() now has to rotate as well as ending the phase. In addition: (7) Alter the transmit tracepoint to log the rotation of the last packet. (8) Remove the no-longer relevant queue_reqack tracepoint note. The ACK-REQUESTED packet header flag is now set as needed when we actually transmit the packet and may vary by retransmission. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-23 19:39:22 +08:00
if (annotation & RXRPC_TX_ANNO_LAST) {
rxrpc: Pass the last Tx packet marker in the annotation buffer When the last packet of data to be transmitted on a call is queued, tx_top is set and then the RXRPC_CALL_TX_LAST flag is set. Unfortunately, this leaves a race in the ACK processing side of things because the flag affects the interpretation of tx_top and also allows us to start receiving reply data before we've finished transmitting. To fix this, make the following changes: (1) rxrpc_queue_packet() now sets a marker in the annotation buffer instead of setting the RXRPC_CALL_TX_LAST flag. (2) rxrpc_rotate_tx_window() detects the marker and sets the flag in the same context as the routines that use it. (3) rxrpc_end_tx_phase() is simplified to just shift the call state. The Tx window must have been rotated before calling to discard the last packet. (4) rxrpc_receiving_reply() is added to handle the arrival of the first DATA packet of a reply to a client call (which is an implicit ACK of the Tx phase). (5) The last part of rxrpc_input_ack() is reordered to perform Tx rotation, then soft-ACK application and then to end the phase if we've rotated the last packet. In the event of a terminal ACK, the soft-ACK application will be skipped as nAcks should be 0. (6) rxrpc_input_ackall() now has to rotate as well as ending the phase. In addition: (7) Alter the transmit tracepoint to log the rotation of the last packet. (8) Remove the no-longer relevant queue_reqack tracepoint note. The ACK-REQUESTED packet header flag is now set as needed when we actually transmit the packet and may vary by retransmission. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-23 19:39:22 +08:00
set_bit(RXRPC_CALL_TX_LAST, &call->flags);
rot_last = true;
}
if ((annotation & RXRPC_TX_ANNO_MASK) != RXRPC_TX_ANNO_ACK)
summary->nr_rot_new_acks++;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
}
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
spin_unlock(&call->lock);
trace_rxrpc_transmit(call, (rot_last ?
rxrpc: Pass the last Tx packet marker in the annotation buffer When the last packet of data to be transmitted on a call is queued, tx_top is set and then the RXRPC_CALL_TX_LAST flag is set. Unfortunately, this leaves a race in the ACK processing side of things because the flag affects the interpretation of tx_top and also allows us to start receiving reply data before we've finished transmitting. To fix this, make the following changes: (1) rxrpc_queue_packet() now sets a marker in the annotation buffer instead of setting the RXRPC_CALL_TX_LAST flag. (2) rxrpc_rotate_tx_window() detects the marker and sets the flag in the same context as the routines that use it. (3) rxrpc_end_tx_phase() is simplified to just shift the call state. The Tx window must have been rotated before calling to discard the last packet. (4) rxrpc_receiving_reply() is added to handle the arrival of the first DATA packet of a reply to a client call (which is an implicit ACK of the Tx phase). (5) The last part of rxrpc_input_ack() is reordered to perform Tx rotation, then soft-ACK application and then to end the phase if we've rotated the last packet. In the event of a terminal ACK, the soft-ACK application will be skipped as nAcks should be 0. (6) rxrpc_input_ackall() now has to rotate as well as ending the phase. In addition: (7) Alter the transmit tracepoint to log the rotation of the last packet. (8) Remove the no-longer relevant queue_reqack tracepoint note. The ACK-REQUESTED packet header flag is now set as needed when we actually transmit the packet and may vary by retransmission. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-23 19:39:22 +08:00
rxrpc_transmit_rotate_last :
rxrpc_transmit_rotate));
wake_up(&call->waitq);
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
while (list) {
skb = list;
list = skb->next;
skb_mark_not_on_list(skb);
rxrpc_free_skb(skb, rxrpc_skb_freed);
}
return rot_last;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
}
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
/*
* End the transmission phase of a call.
*
* This occurs when we get an ACKALL packet, the first DATA packet of a reply,
* or a final ACK packet.
*/
rxrpc: Pass the last Tx packet marker in the annotation buffer When the last packet of data to be transmitted on a call is queued, tx_top is set and then the RXRPC_CALL_TX_LAST flag is set. Unfortunately, this leaves a race in the ACK processing side of things because the flag affects the interpretation of tx_top and also allows us to start receiving reply data before we've finished transmitting. To fix this, make the following changes: (1) rxrpc_queue_packet() now sets a marker in the annotation buffer instead of setting the RXRPC_CALL_TX_LAST flag. (2) rxrpc_rotate_tx_window() detects the marker and sets the flag in the same context as the routines that use it. (3) rxrpc_end_tx_phase() is simplified to just shift the call state. The Tx window must have been rotated before calling to discard the last packet. (4) rxrpc_receiving_reply() is added to handle the arrival of the first DATA packet of a reply to a client call (which is an implicit ACK of the Tx phase). (5) The last part of rxrpc_input_ack() is reordered to perform Tx rotation, then soft-ACK application and then to end the phase if we've rotated the last packet. In the event of a terminal ACK, the soft-ACK application will be skipped as nAcks should be 0. (6) rxrpc_input_ackall() now has to rotate as well as ending the phase. In addition: (7) Alter the transmit tracepoint to log the rotation of the last packet. (8) Remove the no-longer relevant queue_reqack tracepoint note. The ACK-REQUESTED packet header flag is now set as needed when we actually transmit the packet and may vary by retransmission. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-23 19:39:22 +08:00
static bool rxrpc_end_tx_phase(struct rxrpc_call *call, bool reply_begun,
const char *abort_why)
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
{
unsigned int state;
rxrpc: Pass the last Tx packet marker in the annotation buffer When the last packet of data to be transmitted on a call is queued, tx_top is set and then the RXRPC_CALL_TX_LAST flag is set. Unfortunately, this leaves a race in the ACK processing side of things because the flag affects the interpretation of tx_top and also allows us to start receiving reply data before we've finished transmitting. To fix this, make the following changes: (1) rxrpc_queue_packet() now sets a marker in the annotation buffer instead of setting the RXRPC_CALL_TX_LAST flag. (2) rxrpc_rotate_tx_window() detects the marker and sets the flag in the same context as the routines that use it. (3) rxrpc_end_tx_phase() is simplified to just shift the call state. The Tx window must have been rotated before calling to discard the last packet. (4) rxrpc_receiving_reply() is added to handle the arrival of the first DATA packet of a reply to a client call (which is an implicit ACK of the Tx phase). (5) The last part of rxrpc_input_ack() is reordered to perform Tx rotation, then soft-ACK application and then to end the phase if we've rotated the last packet. In the event of a terminal ACK, the soft-ACK application will be skipped as nAcks should be 0. (6) rxrpc_input_ackall() now has to rotate as well as ending the phase. In addition: (7) Alter the transmit tracepoint to log the rotation of the last packet. (8) Remove the no-longer relevant queue_reqack tracepoint note. The ACK-REQUESTED packet header flag is now set as needed when we actually transmit the packet and may vary by retransmission. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-23 19:39:22 +08:00
ASSERT(test_bit(RXRPC_CALL_TX_LAST, &call->flags));
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
write_lock(&call->state_lock);
[AF_RXRPC]: Add an interface to the AF_RXRPC module for the AFS filesystem to use Add an interface to the AF_RXRPC module so that the AFS filesystem module can more easily make use of the services available. AFS still opens a socket but then uses the action functions in lieu of sendmsg() and registers an intercept functions to grab messages before they're queued on the socket Rx queue. This permits AFS (or whatever) to: (1) Avoid the overhead of using the recvmsg() call. (2) Use different keys directly on individual client calls on one socket rather than having to open a whole slew of sockets, one for each key it might want to use. (3) Avoid calling request_key() at the point of issue of a call or opening of a socket. This is done instead by AFS at the point of open(), unlink() or other VFS operation and the key handed through. (4) Request the use of something other than GFP_KERNEL to allocate memory. Furthermore: (*) The socket buffer markings used by RxRPC are made available for AFS so that it can interpret the cooked RxRPC messages itself. (*) rxgen (un)marshalling abort codes are made available. The following documentation for the kernel interface is added to Documentation/networking/rxrpc.txt: ========================= AF_RXRPC KERNEL INTERFACE ========================= The AF_RXRPC module also provides an interface for use by in-kernel utilities such as the AFS filesystem. This permits such a utility to: (1) Use different keys directly on individual client calls on one socket rather than having to open a whole slew of sockets, one for each key it might want to use. (2) Avoid having RxRPC call request_key() at the point of issue of a call or opening of a socket. Instead the utility is responsible for requesting a key at the appropriate point. AFS, for instance, would do this during VFS operations such as open() or unlink(). The key is then handed through when the call is initiated. (3) Request the use of something other than GFP_KERNEL to allocate memory. (4) Avoid the overhead of using the recvmsg() call. RxRPC messages can be intercepted before they get put into the socket Rx queue and the socket buffers manipulated directly. To use the RxRPC facility, a kernel utility must still open an AF_RXRPC socket, bind an addess as appropriate and listen if it's to be a server socket, but then it passes this to the kernel interface functions. The kernel interface functions are as follows: (*) Begin a new client call. struct rxrpc_call * rxrpc_kernel_begin_call(struct socket *sock, struct sockaddr_rxrpc *srx, struct key *key, unsigned long user_call_ID, gfp_t gfp); This allocates the infrastructure to make a new RxRPC call and assigns call and connection numbers. The call will be made on the UDP port that the socket is bound to. The call will go to the destination address of a connected client socket unless an alternative is supplied (srx is non-NULL). If a key is supplied then this will be used to secure the call instead of the key bound to the socket with the RXRPC_SECURITY_KEY sockopt. Calls secured in this way will still share connections if at all possible. The user_call_ID is equivalent to that supplied to sendmsg() in the control data buffer. It is entirely feasible to use this to point to a kernel data structure. If this function is successful, an opaque reference to the RxRPC call is returned. The caller now holds a reference on this and it must be properly ended. (*) End a client call. void rxrpc_kernel_end_call(struct rxrpc_call *call); This is used to end a previously begun call. The user_call_ID is expunged from AF_RXRPC's knowledge and will not be seen again in association with the specified call. (*) Send data through a call. int rxrpc_kernel_send_data(struct rxrpc_call *call, struct msghdr *msg, size_t len); This is used to supply either the request part of a client call or the reply part of a server call. msg.msg_iovlen and msg.msg_iov specify the data buffers to be used. msg_iov may not be NULL and must point exclusively to in-kernel virtual addresses. msg.msg_flags may be given MSG_MORE if there will be subsequent data sends for this call. The msg must not specify a destination address, control data or any flags other than MSG_MORE. len is the total amount of data to transmit. (*) Abort a call. void rxrpc_kernel_abort_call(struct rxrpc_call *call, u32 abort_code); This is used to abort a call if it's still in an abortable state. The abort code specified will be placed in the ABORT message sent. (*) Intercept received RxRPC messages. typedef void (*rxrpc_interceptor_t)(struct sock *sk, unsigned long user_call_ID, struct sk_buff *skb); void rxrpc_kernel_intercept_rx_messages(struct socket *sock, rxrpc_interceptor_t interceptor); This installs an interceptor function on the specified AF_RXRPC socket. All messages that would otherwise wind up in the socket's Rx queue are then diverted to this function. Note that care must be taken to process the messages in the right order to maintain DATA message sequentiality. The interceptor function itself is provided with the address of the socket and handling the incoming message, the ID assigned by the kernel utility to the call and the socket buffer containing the message. The skb->mark field indicates the type of message: MARK MEANING =============================== ======================================= RXRPC_SKB_MARK_DATA Data message RXRPC_SKB_MARK_FINAL_ACK Final ACK received for an incoming call RXRPC_SKB_MARK_BUSY Client call rejected as server busy RXRPC_SKB_MARK_REMOTE_ABORT Call aborted by peer RXRPC_SKB_MARK_NET_ERROR Network error detected RXRPC_SKB_MARK_LOCAL_ERROR Local error encountered RXRPC_SKB_MARK_NEW_CALL New incoming call awaiting acceptance The remote abort message can be probed with rxrpc_kernel_get_abort_code(). The two error messages can be probed with rxrpc_kernel_get_error_number(). A new call can be accepted with rxrpc_kernel_accept_call(). Data messages can have their contents extracted with the usual bunch of socket buffer manipulation functions. A data message can be determined to be the last one in a sequence with rxrpc_kernel_is_data_last(). When a data message has been used up, rxrpc_kernel_data_delivered() should be called on it.. Non-data messages should be handled to rxrpc_kernel_free_skb() to dispose of. It is possible to get extra refs on all types of message for later freeing, but this may pin the state of a call until the message is finally freed. (*) Accept an incoming call. struct rxrpc_call * rxrpc_kernel_accept_call(struct socket *sock, unsigned long user_call_ID); This is used to accept an incoming call and to assign it a call ID. This function is similar to rxrpc_kernel_begin_call() and calls accepted must be ended in the same way. If this function is successful, an opaque reference to the RxRPC call is returned. The caller now holds a reference on this and it must be properly ended. (*) Reject an incoming call. int rxrpc_kernel_reject_call(struct socket *sock); This is used to reject the first incoming call on the socket's queue with a BUSY message. -ENODATA is returned if there were no incoming calls. Other errors may be returned if the call had been aborted (-ECONNABORTED) or had timed out (-ETIME). (*) Record the delivery of a data message and free it. void rxrpc_kernel_data_delivered(struct sk_buff *skb); This is used to record a data message as having been delivered and to update the ACK state for the call. The socket buffer will be freed. (*) Free a message. void rxrpc_kernel_free_skb(struct sk_buff *skb); This is used to free a non-DATA socket buffer intercepted from an AF_RXRPC socket. (*) Determine if a data message is the last one on a call. bool rxrpc_kernel_is_data_last(struct sk_buff *skb); This is used to determine if a socket buffer holds the last data message to be received for a call (true will be returned if it does, false if not). The data message will be part of the reply on a client call and the request on an incoming call. In the latter case there will be more messages, but in the former case there will not. (*) Get the abort code from an abort message. u32 rxrpc_kernel_get_abort_code(struct sk_buff *skb); This is used to extract the abort code from a remote abort message. (*) Get the error number from a local or network error message. int rxrpc_kernel_get_error_number(struct sk_buff *skb); This is used to extract the error number from a message indicating either a local error occurred or a network error occurred. Signed-off-by: David Howells <dhowells@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-04-27 06:50:17 +08:00
state = call->state;
switch (state) {
rxrpc: Pass the last Tx packet marker in the annotation buffer When the last packet of data to be transmitted on a call is queued, tx_top is set and then the RXRPC_CALL_TX_LAST flag is set. Unfortunately, this leaves a race in the ACK processing side of things because the flag affects the interpretation of tx_top and also allows us to start receiving reply data before we've finished transmitting. To fix this, make the following changes: (1) rxrpc_queue_packet() now sets a marker in the annotation buffer instead of setting the RXRPC_CALL_TX_LAST flag. (2) rxrpc_rotate_tx_window() detects the marker and sets the flag in the same context as the routines that use it. (3) rxrpc_end_tx_phase() is simplified to just shift the call state. The Tx window must have been rotated before calling to discard the last packet. (4) rxrpc_receiving_reply() is added to handle the arrival of the first DATA packet of a reply to a client call (which is an implicit ACK of the Tx phase). (5) The last part of rxrpc_input_ack() is reordered to perform Tx rotation, then soft-ACK application and then to end the phase if we've rotated the last packet. In the event of a terminal ACK, the soft-ACK application will be skipped as nAcks should be 0. (6) rxrpc_input_ackall() now has to rotate as well as ending the phase. In addition: (7) Alter the transmit tracepoint to log the rotation of the last packet. (8) Remove the no-longer relevant queue_reqack tracepoint note. The ACK-REQUESTED packet header flag is now set as needed when we actually transmit the packet and may vary by retransmission. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-23 19:39:22 +08:00
case RXRPC_CALL_CLIENT_SEND_REQUEST:
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
case RXRPC_CALL_CLIENT_AWAIT_REPLY:
rxrpc: Pass the last Tx packet marker in the annotation buffer When the last packet of data to be transmitted on a call is queued, tx_top is set and then the RXRPC_CALL_TX_LAST flag is set. Unfortunately, this leaves a race in the ACK processing side of things because the flag affects the interpretation of tx_top and also allows us to start receiving reply data before we've finished transmitting. To fix this, make the following changes: (1) rxrpc_queue_packet() now sets a marker in the annotation buffer instead of setting the RXRPC_CALL_TX_LAST flag. (2) rxrpc_rotate_tx_window() detects the marker and sets the flag in the same context as the routines that use it. (3) rxrpc_end_tx_phase() is simplified to just shift the call state. The Tx window must have been rotated before calling to discard the last packet. (4) rxrpc_receiving_reply() is added to handle the arrival of the first DATA packet of a reply to a client call (which is an implicit ACK of the Tx phase). (5) The last part of rxrpc_input_ack() is reordered to perform Tx rotation, then soft-ACK application and then to end the phase if we've rotated the last packet. In the event of a terminal ACK, the soft-ACK application will be skipped as nAcks should be 0. (6) rxrpc_input_ackall() now has to rotate as well as ending the phase. In addition: (7) Alter the transmit tracepoint to log the rotation of the last packet. (8) Remove the no-longer relevant queue_reqack tracepoint note. The ACK-REQUESTED packet header flag is now set as needed when we actually transmit the packet and may vary by retransmission. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-23 19:39:22 +08:00
if (reply_begun)
call->state = state = RXRPC_CALL_CLIENT_RECV_REPLY;
rxrpc: Pass the last Tx packet marker in the annotation buffer When the last packet of data to be transmitted on a call is queued, tx_top is set and then the RXRPC_CALL_TX_LAST flag is set. Unfortunately, this leaves a race in the ACK processing side of things because the flag affects the interpretation of tx_top and also allows us to start receiving reply data before we've finished transmitting. To fix this, make the following changes: (1) rxrpc_queue_packet() now sets a marker in the annotation buffer instead of setting the RXRPC_CALL_TX_LAST flag. (2) rxrpc_rotate_tx_window() detects the marker and sets the flag in the same context as the routines that use it. (3) rxrpc_end_tx_phase() is simplified to just shift the call state. The Tx window must have been rotated before calling to discard the last packet. (4) rxrpc_receiving_reply() is added to handle the arrival of the first DATA packet of a reply to a client call (which is an implicit ACK of the Tx phase). (5) The last part of rxrpc_input_ack() is reordered to perform Tx rotation, then soft-ACK application and then to end the phase if we've rotated the last packet. In the event of a terminal ACK, the soft-ACK application will be skipped as nAcks should be 0. (6) rxrpc_input_ackall() now has to rotate as well as ending the phase. In addition: (7) Alter the transmit tracepoint to log the rotation of the last packet. (8) Remove the no-longer relevant queue_reqack tracepoint note. The ACK-REQUESTED packet header flag is now set as needed when we actually transmit the packet and may vary by retransmission. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-23 19:39:22 +08:00
else
call->state = state = RXRPC_CALL_CLIENT_AWAIT_REPLY;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
break;
rxrpc: Pass the last Tx packet marker in the annotation buffer When the last packet of data to be transmitted on a call is queued, tx_top is set and then the RXRPC_CALL_TX_LAST flag is set. Unfortunately, this leaves a race in the ACK processing side of things because the flag affects the interpretation of tx_top and also allows us to start receiving reply data before we've finished transmitting. To fix this, make the following changes: (1) rxrpc_queue_packet() now sets a marker in the annotation buffer instead of setting the RXRPC_CALL_TX_LAST flag. (2) rxrpc_rotate_tx_window() detects the marker and sets the flag in the same context as the routines that use it. (3) rxrpc_end_tx_phase() is simplified to just shift the call state. The Tx window must have been rotated before calling to discard the last packet. (4) rxrpc_receiving_reply() is added to handle the arrival of the first DATA packet of a reply to a client call (which is an implicit ACK of the Tx phase). (5) The last part of rxrpc_input_ack() is reordered to perform Tx rotation, then soft-ACK application and then to end the phase if we've rotated the last packet. In the event of a terminal ACK, the soft-ACK application will be skipped as nAcks should be 0. (6) rxrpc_input_ackall() now has to rotate as well as ending the phase. In addition: (7) Alter the transmit tracepoint to log the rotation of the last packet. (8) Remove the no-longer relevant queue_reqack tracepoint note. The ACK-REQUESTED packet header flag is now set as needed when we actually transmit the packet and may vary by retransmission. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-23 19:39:22 +08:00
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
case RXRPC_CALL_SERVER_AWAIT_ACK:
__rxrpc_call_completed(call);
rxrpc_notify_socket(call);
state = call->state;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
break;
rxrpc: Pass the last Tx packet marker in the annotation buffer When the last packet of data to be transmitted on a call is queued, tx_top is set and then the RXRPC_CALL_TX_LAST flag is set. Unfortunately, this leaves a race in the ACK processing side of things because the flag affects the interpretation of tx_top and also allows us to start receiving reply data before we've finished transmitting. To fix this, make the following changes: (1) rxrpc_queue_packet() now sets a marker in the annotation buffer instead of setting the RXRPC_CALL_TX_LAST flag. (2) rxrpc_rotate_tx_window() detects the marker and sets the flag in the same context as the routines that use it. (3) rxrpc_end_tx_phase() is simplified to just shift the call state. The Tx window must have been rotated before calling to discard the last packet. (4) rxrpc_receiving_reply() is added to handle the arrival of the first DATA packet of a reply to a client call (which is an implicit ACK of the Tx phase). (5) The last part of rxrpc_input_ack() is reordered to perform Tx rotation, then soft-ACK application and then to end the phase if we've rotated the last packet. In the event of a terminal ACK, the soft-ACK application will be skipped as nAcks should be 0. (6) rxrpc_input_ackall() now has to rotate as well as ending the phase. In addition: (7) Alter the transmit tracepoint to log the rotation of the last packet. (8) Remove the no-longer relevant queue_reqack tracepoint note. The ACK-REQUESTED packet header flag is now set as needed when we actually transmit the packet and may vary by retransmission. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-23 19:39:22 +08:00
default:
goto bad_state;
}
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
write_unlock(&call->state_lock);
if (state == RXRPC_CALL_CLIENT_AWAIT_REPLY)
rxrpc: Pass the last Tx packet marker in the annotation buffer When the last packet of data to be transmitted on a call is queued, tx_top is set and then the RXRPC_CALL_TX_LAST flag is set. Unfortunately, this leaves a race in the ACK processing side of things because the flag affects the interpretation of tx_top and also allows us to start receiving reply data before we've finished transmitting. To fix this, make the following changes: (1) rxrpc_queue_packet() now sets a marker in the annotation buffer instead of setting the RXRPC_CALL_TX_LAST flag. (2) rxrpc_rotate_tx_window() detects the marker and sets the flag in the same context as the routines that use it. (3) rxrpc_end_tx_phase() is simplified to just shift the call state. The Tx window must have been rotated before calling to discard the last packet. (4) rxrpc_receiving_reply() is added to handle the arrival of the first DATA packet of a reply to a client call (which is an implicit ACK of the Tx phase). (5) The last part of rxrpc_input_ack() is reordered to perform Tx rotation, then soft-ACK application and then to end the phase if we've rotated the last packet. In the event of a terminal ACK, the soft-ACK application will be skipped as nAcks should be 0. (6) rxrpc_input_ackall() now has to rotate as well as ending the phase. In addition: (7) Alter the transmit tracepoint to log the rotation of the last packet. (8) Remove the no-longer relevant queue_reqack tracepoint note. The ACK-REQUESTED packet header flag is now set as needed when we actually transmit the packet and may vary by retransmission. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-23 19:39:22 +08:00
trace_rxrpc_transmit(call, rxrpc_transmit_await_reply);
else
rxrpc: Pass the last Tx packet marker in the annotation buffer When the last packet of data to be transmitted on a call is queued, tx_top is set and then the RXRPC_CALL_TX_LAST flag is set. Unfortunately, this leaves a race in the ACK processing side of things because the flag affects the interpretation of tx_top and also allows us to start receiving reply data before we've finished transmitting. To fix this, make the following changes: (1) rxrpc_queue_packet() now sets a marker in the annotation buffer instead of setting the RXRPC_CALL_TX_LAST flag. (2) rxrpc_rotate_tx_window() detects the marker and sets the flag in the same context as the routines that use it. (3) rxrpc_end_tx_phase() is simplified to just shift the call state. The Tx window must have been rotated before calling to discard the last packet. (4) rxrpc_receiving_reply() is added to handle the arrival of the first DATA packet of a reply to a client call (which is an implicit ACK of the Tx phase). (5) The last part of rxrpc_input_ack() is reordered to perform Tx rotation, then soft-ACK application and then to end the phase if we've rotated the last packet. In the event of a terminal ACK, the soft-ACK application will be skipped as nAcks should be 0. (6) rxrpc_input_ackall() now has to rotate as well as ending the phase. In addition: (7) Alter the transmit tracepoint to log the rotation of the last packet. (8) Remove the no-longer relevant queue_reqack tracepoint note. The ACK-REQUESTED packet header flag is now set as needed when we actually transmit the packet and may vary by retransmission. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-23 19:39:22 +08:00
trace_rxrpc_transmit(call, rxrpc_transmit_end);
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
_leave(" = ok");
return true;
rxrpc: Pass the last Tx packet marker in the annotation buffer When the last packet of data to be transmitted on a call is queued, tx_top is set and then the RXRPC_CALL_TX_LAST flag is set. Unfortunately, this leaves a race in the ACK processing side of things because the flag affects the interpretation of tx_top and also allows us to start receiving reply data before we've finished transmitting. To fix this, make the following changes: (1) rxrpc_queue_packet() now sets a marker in the annotation buffer instead of setting the RXRPC_CALL_TX_LAST flag. (2) rxrpc_rotate_tx_window() detects the marker and sets the flag in the same context as the routines that use it. (3) rxrpc_end_tx_phase() is simplified to just shift the call state. The Tx window must have been rotated before calling to discard the last packet. (4) rxrpc_receiving_reply() is added to handle the arrival of the first DATA packet of a reply to a client call (which is an implicit ACK of the Tx phase). (5) The last part of rxrpc_input_ack() is reordered to perform Tx rotation, then soft-ACK application and then to end the phase if we've rotated the last packet. In the event of a terminal ACK, the soft-ACK application will be skipped as nAcks should be 0. (6) rxrpc_input_ackall() now has to rotate as well as ending the phase. In addition: (7) Alter the transmit tracepoint to log the rotation of the last packet. (8) Remove the no-longer relevant queue_reqack tracepoint note. The ACK-REQUESTED packet header flag is now set as needed when we actually transmit the packet and may vary by retransmission. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-23 19:39:22 +08:00
bad_state:
write_unlock(&call->state_lock);
kdebug("end_tx %s", rxrpc_call_states[call->state]);
rxrpc_proto_abort(abort_why, call, call->tx_top);
return false;
}
/*
* Begin the reply reception phase of a call.
*/
static bool rxrpc_receiving_reply(struct rxrpc_call *call)
{
struct rxrpc_ack_summary summary = { 0 };
rxrpc: Fix call timeouts Fix the rxrpc call expiration timeouts and make them settable from userspace. By analogy with other rx implementations, there should be three timeouts: (1) "Normal timeout" This is set for all calls and is triggered if we haven't received any packets from the peer in a while. It is measured from the last time we received any packet on that call. This is not reset by any connection packets (such as CHALLENGE/RESPONSE packets). If a service operation takes a long time, the server should generate PING ACKs at a duration that's substantially less than the normal timeout so is to keep both sides alive. This is set at 1/6 of normal timeout. (2) "Idle timeout" This is set only for a service call and is triggered if we stop receiving the DATA packets that comprise the request data. It is measured from the last time we received a DATA packet. (3) "Hard timeout" This can be set for a call and specified the maximum lifetime of that call. It should not be specified by default. Some operations (such as volume transfer) take a long time. Allow userspace to set/change the timeouts on a call with sendmsg, using a control message: RXRPC_SET_CALL_TIMEOUTS The data to the message is a number of 32-bit words, not all of which need be given: u32 hard_timeout; /* sec from first packet */ u32 idle_timeout; /* msec from packet Rx */ u32 normal_timeout; /* msec from data Rx */ This can be set in combination with any other sendmsg() that affects a call. Signed-off-by: David Howells <dhowells@redhat.com>
2017-11-24 18:18:41 +08:00
unsigned long now, timo;
rxrpc: Pass the last Tx packet marker in the annotation buffer When the last packet of data to be transmitted on a call is queued, tx_top is set and then the RXRPC_CALL_TX_LAST flag is set. Unfortunately, this leaves a race in the ACK processing side of things because the flag affects the interpretation of tx_top and also allows us to start receiving reply data before we've finished transmitting. To fix this, make the following changes: (1) rxrpc_queue_packet() now sets a marker in the annotation buffer instead of setting the RXRPC_CALL_TX_LAST flag. (2) rxrpc_rotate_tx_window() detects the marker and sets the flag in the same context as the routines that use it. (3) rxrpc_end_tx_phase() is simplified to just shift the call state. The Tx window must have been rotated before calling to discard the last packet. (4) rxrpc_receiving_reply() is added to handle the arrival of the first DATA packet of a reply to a client call (which is an implicit ACK of the Tx phase). (5) The last part of rxrpc_input_ack() is reordered to perform Tx rotation, then soft-ACK application and then to end the phase if we've rotated the last packet. In the event of a terminal ACK, the soft-ACK application will be skipped as nAcks should be 0. (6) rxrpc_input_ackall() now has to rotate as well as ending the phase. In addition: (7) Alter the transmit tracepoint to log the rotation of the last packet. (8) Remove the no-longer relevant queue_reqack tracepoint note. The ACK-REQUESTED packet header flag is now set as needed when we actually transmit the packet and may vary by retransmission. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-23 19:39:22 +08:00
rxrpc_seq_t top = READ_ONCE(call->tx_top);
if (call->ackr_reason) {
spin_lock_bh(&call->lock);
call->ackr_reason = 0;
spin_unlock_bh(&call->lock);
rxrpc: Fix call timeouts Fix the rxrpc call expiration timeouts and make them settable from userspace. By analogy with other rx implementations, there should be three timeouts: (1) "Normal timeout" This is set for all calls and is triggered if we haven't received any packets from the peer in a while. It is measured from the last time we received any packet on that call. This is not reset by any connection packets (such as CHALLENGE/RESPONSE packets). If a service operation takes a long time, the server should generate PING ACKs at a duration that's substantially less than the normal timeout so is to keep both sides alive. This is set at 1/6 of normal timeout. (2) "Idle timeout" This is set only for a service call and is triggered if we stop receiving the DATA packets that comprise the request data. It is measured from the last time we received a DATA packet. (3) "Hard timeout" This can be set for a call and specified the maximum lifetime of that call. It should not be specified by default. Some operations (such as volume transfer) take a long time. Allow userspace to set/change the timeouts on a call with sendmsg, using a control message: RXRPC_SET_CALL_TIMEOUTS The data to the message is a number of 32-bit words, not all of which need be given: u32 hard_timeout; /* sec from first packet */ u32 idle_timeout; /* msec from packet Rx */ u32 normal_timeout; /* msec from data Rx */ This can be set in combination with any other sendmsg() that affects a call. Signed-off-by: David Howells <dhowells@redhat.com>
2017-11-24 18:18:41 +08:00
now = jiffies;
timo = now + MAX_JIFFY_OFFSET;
WRITE_ONCE(call->resend_at, timo);
WRITE_ONCE(call->ack_at, timo);
trace_rxrpc_timer(call, rxrpc_timer_init_for_reply, now);
}
rxrpc: Pass the last Tx packet marker in the annotation buffer When the last packet of data to be transmitted on a call is queued, tx_top is set and then the RXRPC_CALL_TX_LAST flag is set. Unfortunately, this leaves a race in the ACK processing side of things because the flag affects the interpretation of tx_top and also allows us to start receiving reply data before we've finished transmitting. To fix this, make the following changes: (1) rxrpc_queue_packet() now sets a marker in the annotation buffer instead of setting the RXRPC_CALL_TX_LAST flag. (2) rxrpc_rotate_tx_window() detects the marker and sets the flag in the same context as the routines that use it. (3) rxrpc_end_tx_phase() is simplified to just shift the call state. The Tx window must have been rotated before calling to discard the last packet. (4) rxrpc_receiving_reply() is added to handle the arrival of the first DATA packet of a reply to a client call (which is an implicit ACK of the Tx phase). (5) The last part of rxrpc_input_ack() is reordered to perform Tx rotation, then soft-ACK application and then to end the phase if we've rotated the last packet. In the event of a terminal ACK, the soft-ACK application will be skipped as nAcks should be 0. (6) rxrpc_input_ackall() now has to rotate as well as ending the phase. In addition: (7) Alter the transmit tracepoint to log the rotation of the last packet. (8) Remove the no-longer relevant queue_reqack tracepoint note. The ACK-REQUESTED packet header flag is now set as needed when we actually transmit the packet and may vary by retransmission. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-23 19:39:22 +08:00
if (!test_bit(RXRPC_CALL_TX_LAST, &call->flags)) {
if (!rxrpc_rotate_tx_window(call, top, &summary)) {
rxrpc_proto_abort("TXL", call, top);
return false;
}
rxrpc: Pass the last Tx packet marker in the annotation buffer When the last packet of data to be transmitted on a call is queued, tx_top is set and then the RXRPC_CALL_TX_LAST flag is set. Unfortunately, this leaves a race in the ACK processing side of things because the flag affects the interpretation of tx_top and also allows us to start receiving reply data before we've finished transmitting. To fix this, make the following changes: (1) rxrpc_queue_packet() now sets a marker in the annotation buffer instead of setting the RXRPC_CALL_TX_LAST flag. (2) rxrpc_rotate_tx_window() detects the marker and sets the flag in the same context as the routines that use it. (3) rxrpc_end_tx_phase() is simplified to just shift the call state. The Tx window must have been rotated before calling to discard the last packet. (4) rxrpc_receiving_reply() is added to handle the arrival of the first DATA packet of a reply to a client call (which is an implicit ACK of the Tx phase). (5) The last part of rxrpc_input_ack() is reordered to perform Tx rotation, then soft-ACK application and then to end the phase if we've rotated the last packet. In the event of a terminal ACK, the soft-ACK application will be skipped as nAcks should be 0. (6) rxrpc_input_ackall() now has to rotate as well as ending the phase. In addition: (7) Alter the transmit tracepoint to log the rotation of the last packet. (8) Remove the no-longer relevant queue_reqack tracepoint note. The ACK-REQUESTED packet header flag is now set as needed when we actually transmit the packet and may vary by retransmission. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-23 19:39:22 +08:00
}
if (!rxrpc_end_tx_phase(call, true, "ETD"))
return false;
call->tx_phase = false;
return true;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
}
/*
* Scan a data packet to validate its structure and to work out how many
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
* subpackets it contains.
*
* A jumbo packet is a collection of consecutive packets glued together with
* little headers between that indicate how to change the initial header for
* each subpacket.
*
* RXRPC_JUMBO_PACKET must be set on all but the last subpacket - and all but
* the last are RXRPC_JUMBO_DATALEN in size. The last subpacket may be of any
* size.
*/
static bool rxrpc_validate_data(struct sk_buff *skb)
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
{
struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
unsigned int offset = sizeof(struct rxrpc_wire_header);
unsigned int len = skb->len;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
u8 flags = sp->hdr.flags;
for (;;) {
if (flags & RXRPC_REQUEST_ACK)
__set_bit(sp->nr_subpackets, sp->rx_req_ack);
sp->nr_subpackets++;
if (!(flags & RXRPC_JUMBO_PACKET))
break;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
if (len - offset < RXRPC_JUMBO_SUBPKTLEN)
goto protocol_error;
if (flags & RXRPC_LAST_PACKET)
goto protocol_error;
offset += RXRPC_JUMBO_DATALEN;
if (skb_copy_bits(skb, offset, &flags, 1) < 0)
goto protocol_error;
offset += sizeof(struct rxrpc_jumbo_header);
}
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
if (flags & RXRPC_LAST_PACKET)
sp->rx_flags |= RXRPC_SKB_INCL_LAST;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
return true;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
protocol_error:
return false;
}
/*
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
* Handle reception of a duplicate packet.
*
* We have to take care to avoid an attack here whereby we're given a series of
* jumbograms, each with a sequence number one before the preceding one and
* filled up to maximum UDP size. If they never send us the first packet in
* the sequence, they can cause us to have to hold on to around 2MiB of kernel
* space until the call times out.
*
* We limit the space usage by only accepting three duplicate jumbo packets per
* call. After that, we tell the other side we're no longer accepting jumbos
* (that information is encoded in the ACK packet).
*/
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
static void rxrpc_input_dup_data(struct rxrpc_call *call, rxrpc_seq_t seq,
bool is_jumbo, bool *_jumbo_bad)
{
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
/* Discard normal packets that are duplicates. */
if (is_jumbo)
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
return;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
/* Skip jumbo subpackets that are duplicates. When we've had three or
* more partially duplicate jumbo packets, we refuse to take any more
* jumbos for this call.
*/
if (!*_jumbo_bad) {
call->nr_jumbo_bad++;
*_jumbo_bad = true;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
}
}
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
/*
* Process a DATA packet, adding the packet to the Rx ring. The caller's
* packet ref must be passed on or discarded.
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
*/
static void rxrpc_input_data(struct rxrpc_call *call, struct sk_buff *skb)
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
{
struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
enum rxrpc_call_state state;
unsigned int j, nr_subpackets;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
rxrpc_serial_t serial = sp->hdr.serial, ack_serial = 0;
rxrpc_seq_t seq0 = sp->hdr.seq, hard_ack;
bool immediate_ack = false, jumbo_bad = false;
u8 ack = 0;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
_enter("{%u,%u},{%u,%u}",
call->rx_hard_ack, call->rx_top, skb->len, seq0);
_proto("Rx DATA %%%u { #%u f=%02x n=%u }",
sp->hdr.serial, seq0, sp->hdr.flags, sp->nr_subpackets);
state = READ_ONCE(call->state);
if (state >= RXRPC_CALL_COMPLETE) {
rxrpc_free_skb(skb, rxrpc_skb_freed);
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
return;
}
rxrpc: Fix call timeouts Fix the rxrpc call expiration timeouts and make them settable from userspace. By analogy with other rx implementations, there should be three timeouts: (1) "Normal timeout" This is set for all calls and is triggered if we haven't received any packets from the peer in a while. It is measured from the last time we received any packet on that call. This is not reset by any connection packets (such as CHALLENGE/RESPONSE packets). If a service operation takes a long time, the server should generate PING ACKs at a duration that's substantially less than the normal timeout so is to keep both sides alive. This is set at 1/6 of normal timeout. (2) "Idle timeout" This is set only for a service call and is triggered if we stop receiving the DATA packets that comprise the request data. It is measured from the last time we received a DATA packet. (3) "Hard timeout" This can be set for a call and specified the maximum lifetime of that call. It should not be specified by default. Some operations (such as volume transfer) take a long time. Allow userspace to set/change the timeouts on a call with sendmsg, using a control message: RXRPC_SET_CALL_TIMEOUTS The data to the message is a number of 32-bit words, not all of which need be given: u32 hard_timeout; /* sec from first packet */ u32 idle_timeout; /* msec from packet Rx */ u32 normal_timeout; /* msec from data Rx */ This can be set in combination with any other sendmsg() that affects a call. Signed-off-by: David Howells <dhowells@redhat.com>
2017-11-24 18:18:41 +08:00
if (call->state == RXRPC_CALL_SERVER_RECV_REQUEST) {
unsigned long timo = READ_ONCE(call->next_req_timo);
unsigned long now, expect_req_by;
if (timo) {
now = jiffies;
expect_req_by = now + timo;
WRITE_ONCE(call->expect_req_by, expect_req_by);
rxrpc_reduce_call_timer(call, expect_req_by, now,
rxrpc_timer_set_for_idle);
}
}
rxrpc: Fix the packet reception routine The rxrpc_input_packet() function and its call tree was built around the assumption that data_ready() handler called from UDP to inform a kernel service that there is data to be had was non-reentrant. This means that certain locking could be dispensed with. This, however, turns out not to be the case with a multi-queue network card that can deliver packets to multiple cpus simultaneously. Each of those cpus can be in the rxrpc_input_packet() function at the same time. Fix by adding or changing some structure members: (1) Add peer->rtt_input_lock to serialise access to the RTT buffer. (2) Make conn->service_id into a 32-bit variable so that it can be cmpxchg'd on all arches. (3) Add call->input_lock to serialise access to the Rx/Tx state. Note that although the Rx and Tx states are (almost) entirely separate, there's no point completing the separation and having separate locks since it's a bi-phasal RPC protocol rather than a bi-direction streaming protocol. Data transmission and data reception do not take place simultaneously on any particular call. and making the following functional changes: (1) In rxrpc_input_data(), hold call->input_lock around the core to prevent simultaneous producing of packets into the Rx ring and updating of tracking state for a particular call. (2) In rxrpc_input_ping_response(), only read call->ping_serial once, and check it before checking RXRPC_CALL_PINGING as that's a cheaper test. The bit test and bit clear can then be combined. No further locking is needed here. (3) In rxrpc_input_ack(), take call->input_lock after we've parsed much of the ACK packet. The superseded ACK check is then done both before and after the lock is taken. The handing of ackinfo data is split, parsing before the lock is taken and processing with it held. This is keyed on rxMTU being non-zero. Congestion management is also done within the locked section. (4) In rxrpc_input_ackall(), take call->input_lock around the Tx window rotation. The ACKALL packet carries no information and is only really useful after all packets have been transmitted since it's imprecise. (5) In rxrpc_input_implicit_end_call(), we use rx->incoming_lock to prevent calls being simultaneously implicitly ended on two cpus and also to prevent any races with incoming call setup. (6) In rxrpc_input_packet(), use cmpxchg() to effect the service upgrade on a connection. It is only permitted to happen once for a connection. (7) In rxrpc_new_incoming_call(), we have to recheck the routing inside rx->incoming_lock to see if someone else set up the call, connection or peer whilst we were getting there. We can't trust the values from the earlier routing check unless we pin refs on them - which we want to avoid. Further, we need to allow for an incoming call to have its state changed on another CPU between us making it live and us adjusting it because the conn is now in the RXRPC_CONN_SERVICE state. (8) In rxrpc_peer_add_rtt(), take peer->rtt_input_lock around the access to the RTT buffer. Don't need to lock around setting peer->rtt. For reference, the inventory of state-accessing or state-altering functions used by the packet input procedure is: > rxrpc_input_packet() * PACKET CHECKING * ROUTING > rxrpc_post_packet_to_local() > rxrpc_find_connection_rcu() - uses RCU > rxrpc_lookup_peer_rcu() - uses RCU > rxrpc_find_service_conn_rcu() - uses RCU > idr_find() - uses RCU * CONNECTION-LEVEL PROCESSING - Service upgrade - Can only happen once per conn ! Changed to use cmpxchg > rxrpc_post_packet_to_conn() - Setting conn->hi_serial - Probably safe not using locks - Maybe use cmpxchg * CALL-LEVEL PROCESSING > Old-call checking > rxrpc_input_implicit_end_call() > rxrpc_call_completed() > rxrpc_queue_call() ! Need to take rx->incoming_lock > __rxrpc_disconnect_call() > rxrpc_notify_socket() > rxrpc_new_incoming_call() - Uses rx->incoming_lock for the entire process - Might be able to drop this earlier in favour of the call lock > rxrpc_incoming_call() ! Conflicts with rxrpc_input_implicit_end_call() > rxrpc_send_ping() - Don't need locks to check rtt state > rxrpc_propose_ACK * PACKET DISTRIBUTION > rxrpc_input_call_packet() > rxrpc_input_data() * QUEUE DATA PACKET ON CALL > rxrpc_reduce_call_timer() - Uses timer_reduce() ! Needs call->input_lock() > rxrpc_receiving_reply() ! Needs locking around ack state > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_proto_abort() > rxrpc_input_dup_data() - Fills the Rx buffer - rxrpc_propose_ACK() - rxrpc_notify_socket() > rxrpc_input_ack() * APPLY ACK PACKET TO CALL AND DISCARD PACKET > rxrpc_input_ping_response() - Probably doesn't need any extra locking ! Need READ_ONCE() on call->ping_serial > rxrpc_input_check_for_lost_ack() - Takes call->lock to consult Tx buffer > rxrpc_peer_add_rtt() ! Needs to take a lock (peer->rtt_input_lock) ! Could perhaps manage with cmpxchg() and xadd() instead > rxrpc_input_requested_ack - Consults Tx buffer ! Probably needs a lock > rxrpc_peer_add_rtt() > rxrpc_propose_ack() > rxrpc_input_ackinfo() - Changes call->tx_winsize ! Use cmpxchg to handle change ! Should perhaps track serial number - Uses peer->lock to record MTU specification changes > rxrpc_proto_abort() ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_input_soft_acks() - Consults the Tx buffer > rxrpc_congestion_management() - Modifies the Tx annotations ! Needs call->input_lock() > rxrpc_queue_call() > rxrpc_input_abort() * APPLY ABORT PACKET TO CALL AND DISCARD PACKET > rxrpc_set_call_completion() > rxrpc_notify_socket() > rxrpc_input_ackall() * APPLY ACKALL PACKET TO CALL AND DISCARD PACKET ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_reject_packet() There are some functions used by the above that queue the packet, after which the procedure is terminated: - rxrpc_post_packet_to_local() - local->event_queue is an sk_buff_head - local->processor is a work_struct - rxrpc_post_packet_to_conn() - conn->rx_queue is an sk_buff_head - conn->processor is a work_struct - rxrpc_reject_packet() - local->reject_queue is an sk_buff_head - local->processor is a work_struct And some that offload processing to process context: - rxrpc_notify_socket() - Uses RCU lock - Uses call->notify_lock to call call->notify_rx - Uses call->recvmsg_lock to queue recvmsg side - rxrpc_queue_call() - call->processor is a work_struct - rxrpc_propose_ACK() - Uses call->lock to wrap __rxrpc_propose_ACK() And a bunch that complete a call, all of which use call->state_lock to protect the call state: - rxrpc_call_completed() - rxrpc_set_call_completion() - rxrpc_abort_call() - rxrpc_proto_abort() - Also uses rxrpc_queue_call() Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both") Signed-off-by: David Howells <dhowells@redhat.com>
2018-10-08 22:46:25 +08:00
spin_lock(&call->input_lock);
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
/* Received data implicitly ACKs all of the request packets we sent
* when we're acting as a client.
*/
if ((state == RXRPC_CALL_CLIENT_SEND_REQUEST ||
state == RXRPC_CALL_CLIENT_AWAIT_REPLY) &&
rxrpc: Pass the last Tx packet marker in the annotation buffer When the last packet of data to be transmitted on a call is queued, tx_top is set and then the RXRPC_CALL_TX_LAST flag is set. Unfortunately, this leaves a race in the ACK processing side of things because the flag affects the interpretation of tx_top and also allows us to start receiving reply data before we've finished transmitting. To fix this, make the following changes: (1) rxrpc_queue_packet() now sets a marker in the annotation buffer instead of setting the RXRPC_CALL_TX_LAST flag. (2) rxrpc_rotate_tx_window() detects the marker and sets the flag in the same context as the routines that use it. (3) rxrpc_end_tx_phase() is simplified to just shift the call state. The Tx window must have been rotated before calling to discard the last packet. (4) rxrpc_receiving_reply() is added to handle the arrival of the first DATA packet of a reply to a client call (which is an implicit ACK of the Tx phase). (5) The last part of rxrpc_input_ack() is reordered to perform Tx rotation, then soft-ACK application and then to end the phase if we've rotated the last packet. In the event of a terminal ACK, the soft-ACK application will be skipped as nAcks should be 0. (6) rxrpc_input_ackall() now has to rotate as well as ending the phase. In addition: (7) Alter the transmit tracepoint to log the rotation of the last packet. (8) Remove the no-longer relevant queue_reqack tracepoint note. The ACK-REQUESTED packet header flag is now set as needed when we actually transmit the packet and may vary by retransmission. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-23 19:39:22 +08:00
!rxrpc_receiving_reply(call))
rxrpc: Fix the packet reception routine The rxrpc_input_packet() function and its call tree was built around the assumption that data_ready() handler called from UDP to inform a kernel service that there is data to be had was non-reentrant. This means that certain locking could be dispensed with. This, however, turns out not to be the case with a multi-queue network card that can deliver packets to multiple cpus simultaneously. Each of those cpus can be in the rxrpc_input_packet() function at the same time. Fix by adding or changing some structure members: (1) Add peer->rtt_input_lock to serialise access to the RTT buffer. (2) Make conn->service_id into a 32-bit variable so that it can be cmpxchg'd on all arches. (3) Add call->input_lock to serialise access to the Rx/Tx state. Note that although the Rx and Tx states are (almost) entirely separate, there's no point completing the separation and having separate locks since it's a bi-phasal RPC protocol rather than a bi-direction streaming protocol. Data transmission and data reception do not take place simultaneously on any particular call. and making the following functional changes: (1) In rxrpc_input_data(), hold call->input_lock around the core to prevent simultaneous producing of packets into the Rx ring and updating of tracking state for a particular call. (2) In rxrpc_input_ping_response(), only read call->ping_serial once, and check it before checking RXRPC_CALL_PINGING as that's a cheaper test. The bit test and bit clear can then be combined. No further locking is needed here. (3) In rxrpc_input_ack(), take call->input_lock after we've parsed much of the ACK packet. The superseded ACK check is then done both before and after the lock is taken. The handing of ackinfo data is split, parsing before the lock is taken and processing with it held. This is keyed on rxMTU being non-zero. Congestion management is also done within the locked section. (4) In rxrpc_input_ackall(), take call->input_lock around the Tx window rotation. The ACKALL packet carries no information and is only really useful after all packets have been transmitted since it's imprecise. (5) In rxrpc_input_implicit_end_call(), we use rx->incoming_lock to prevent calls being simultaneously implicitly ended on two cpus and also to prevent any races with incoming call setup. (6) In rxrpc_input_packet(), use cmpxchg() to effect the service upgrade on a connection. It is only permitted to happen once for a connection. (7) In rxrpc_new_incoming_call(), we have to recheck the routing inside rx->incoming_lock to see if someone else set up the call, connection or peer whilst we were getting there. We can't trust the values from the earlier routing check unless we pin refs on them - which we want to avoid. Further, we need to allow for an incoming call to have its state changed on another CPU between us making it live and us adjusting it because the conn is now in the RXRPC_CONN_SERVICE state. (8) In rxrpc_peer_add_rtt(), take peer->rtt_input_lock around the access to the RTT buffer. Don't need to lock around setting peer->rtt. For reference, the inventory of state-accessing or state-altering functions used by the packet input procedure is: > rxrpc_input_packet() * PACKET CHECKING * ROUTING > rxrpc_post_packet_to_local() > rxrpc_find_connection_rcu() - uses RCU > rxrpc_lookup_peer_rcu() - uses RCU > rxrpc_find_service_conn_rcu() - uses RCU > idr_find() - uses RCU * CONNECTION-LEVEL PROCESSING - Service upgrade - Can only happen once per conn ! Changed to use cmpxchg > rxrpc_post_packet_to_conn() - Setting conn->hi_serial - Probably safe not using locks - Maybe use cmpxchg * CALL-LEVEL PROCESSING > Old-call checking > rxrpc_input_implicit_end_call() > rxrpc_call_completed() > rxrpc_queue_call() ! Need to take rx->incoming_lock > __rxrpc_disconnect_call() > rxrpc_notify_socket() > rxrpc_new_incoming_call() - Uses rx->incoming_lock for the entire process - Might be able to drop this earlier in favour of the call lock > rxrpc_incoming_call() ! Conflicts with rxrpc_input_implicit_end_call() > rxrpc_send_ping() - Don't need locks to check rtt state > rxrpc_propose_ACK * PACKET DISTRIBUTION > rxrpc_input_call_packet() > rxrpc_input_data() * QUEUE DATA PACKET ON CALL > rxrpc_reduce_call_timer() - Uses timer_reduce() ! Needs call->input_lock() > rxrpc_receiving_reply() ! Needs locking around ack state > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_proto_abort() > rxrpc_input_dup_data() - Fills the Rx buffer - rxrpc_propose_ACK() - rxrpc_notify_socket() > rxrpc_input_ack() * APPLY ACK PACKET TO CALL AND DISCARD PACKET > rxrpc_input_ping_response() - Probably doesn't need any extra locking ! Need READ_ONCE() on call->ping_serial > rxrpc_input_check_for_lost_ack() - Takes call->lock to consult Tx buffer > rxrpc_peer_add_rtt() ! Needs to take a lock (peer->rtt_input_lock) ! Could perhaps manage with cmpxchg() and xadd() instead > rxrpc_input_requested_ack - Consults Tx buffer ! Probably needs a lock > rxrpc_peer_add_rtt() > rxrpc_propose_ack() > rxrpc_input_ackinfo() - Changes call->tx_winsize ! Use cmpxchg to handle change ! Should perhaps track serial number - Uses peer->lock to record MTU specification changes > rxrpc_proto_abort() ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_input_soft_acks() - Consults the Tx buffer > rxrpc_congestion_management() - Modifies the Tx annotations ! Needs call->input_lock() > rxrpc_queue_call() > rxrpc_input_abort() * APPLY ABORT PACKET TO CALL AND DISCARD PACKET > rxrpc_set_call_completion() > rxrpc_notify_socket() > rxrpc_input_ackall() * APPLY ACKALL PACKET TO CALL AND DISCARD PACKET ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_reject_packet() There are some functions used by the above that queue the packet, after which the procedure is terminated: - rxrpc_post_packet_to_local() - local->event_queue is an sk_buff_head - local->processor is a work_struct - rxrpc_post_packet_to_conn() - conn->rx_queue is an sk_buff_head - conn->processor is a work_struct - rxrpc_reject_packet() - local->reject_queue is an sk_buff_head - local->processor is a work_struct And some that offload processing to process context: - rxrpc_notify_socket() - Uses RCU lock - Uses call->notify_lock to call call->notify_rx - Uses call->recvmsg_lock to queue recvmsg side - rxrpc_queue_call() - call->processor is a work_struct - rxrpc_propose_ACK() - Uses call->lock to wrap __rxrpc_propose_ACK() And a bunch that complete a call, all of which use call->state_lock to protect the call state: - rxrpc_call_completed() - rxrpc_set_call_completion() - rxrpc_abort_call() - rxrpc_proto_abort() - Also uses rxrpc_queue_call() Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both") Signed-off-by: David Howells <dhowells@redhat.com>
2018-10-08 22:46:25 +08:00
goto unlock;
call->ackr_prev_seq = seq0;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
hard_ack = READ_ONCE(call->rx_hard_ack);
nr_subpackets = sp->nr_subpackets;
if (nr_subpackets > 1) {
if (call->nr_jumbo_bad > 3) {
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
ack = RXRPC_ACK_NOSPACE;
ack_serial = serial;
goto ack;
}
}
for (j = 0; j < nr_subpackets; j++) {
rxrpc_serial_t serial = sp->hdr.serial + j;
rxrpc_seq_t seq = seq0 + j;
unsigned int ix = seq & RXRPC_RXTX_BUFF_MASK;
bool terminal = (j == nr_subpackets - 1);
bool last = terminal && (sp->rx_flags & RXRPC_SKB_INCL_LAST);
u8 flags, annotation = j;
_proto("Rx DATA+%u %%%u { #%x t=%u l=%u }",
j, serial, seq, terminal, last);
if (last) {
if (test_bit(RXRPC_CALL_RX_LAST, &call->flags) &&
seq != call->rx_top) {
rxrpc_proto_abort("LSN", call, seq);
goto unlock;
}
} else {
if (test_bit(RXRPC_CALL_RX_LAST, &call->flags) &&
after_eq(seq, call->rx_top)) {
rxrpc_proto_abort("LSA", call, seq);
goto unlock;
}
rxrpc: Fix the packet reception routine The rxrpc_input_packet() function and its call tree was built around the assumption that data_ready() handler called from UDP to inform a kernel service that there is data to be had was non-reentrant. This means that certain locking could be dispensed with. This, however, turns out not to be the case with a multi-queue network card that can deliver packets to multiple cpus simultaneously. Each of those cpus can be in the rxrpc_input_packet() function at the same time. Fix by adding or changing some structure members: (1) Add peer->rtt_input_lock to serialise access to the RTT buffer. (2) Make conn->service_id into a 32-bit variable so that it can be cmpxchg'd on all arches. (3) Add call->input_lock to serialise access to the Rx/Tx state. Note that although the Rx and Tx states are (almost) entirely separate, there's no point completing the separation and having separate locks since it's a bi-phasal RPC protocol rather than a bi-direction streaming protocol. Data transmission and data reception do not take place simultaneously on any particular call. and making the following functional changes: (1) In rxrpc_input_data(), hold call->input_lock around the core to prevent simultaneous producing of packets into the Rx ring and updating of tracking state for a particular call. (2) In rxrpc_input_ping_response(), only read call->ping_serial once, and check it before checking RXRPC_CALL_PINGING as that's a cheaper test. The bit test and bit clear can then be combined. No further locking is needed here. (3) In rxrpc_input_ack(), take call->input_lock after we've parsed much of the ACK packet. The superseded ACK check is then done both before and after the lock is taken. The handing of ackinfo data is split, parsing before the lock is taken and processing with it held. This is keyed on rxMTU being non-zero. Congestion management is also done within the locked section. (4) In rxrpc_input_ackall(), take call->input_lock around the Tx window rotation. The ACKALL packet carries no information and is only really useful after all packets have been transmitted since it's imprecise. (5) In rxrpc_input_implicit_end_call(), we use rx->incoming_lock to prevent calls being simultaneously implicitly ended on two cpus and also to prevent any races with incoming call setup. (6) In rxrpc_input_packet(), use cmpxchg() to effect the service upgrade on a connection. It is only permitted to happen once for a connection. (7) In rxrpc_new_incoming_call(), we have to recheck the routing inside rx->incoming_lock to see if someone else set up the call, connection or peer whilst we were getting there. We can't trust the values from the earlier routing check unless we pin refs on them - which we want to avoid. Further, we need to allow for an incoming call to have its state changed on another CPU between us making it live and us adjusting it because the conn is now in the RXRPC_CONN_SERVICE state. (8) In rxrpc_peer_add_rtt(), take peer->rtt_input_lock around the access to the RTT buffer. Don't need to lock around setting peer->rtt. For reference, the inventory of state-accessing or state-altering functions used by the packet input procedure is: > rxrpc_input_packet() * PACKET CHECKING * ROUTING > rxrpc_post_packet_to_local() > rxrpc_find_connection_rcu() - uses RCU > rxrpc_lookup_peer_rcu() - uses RCU > rxrpc_find_service_conn_rcu() - uses RCU > idr_find() - uses RCU * CONNECTION-LEVEL PROCESSING - Service upgrade - Can only happen once per conn ! Changed to use cmpxchg > rxrpc_post_packet_to_conn() - Setting conn->hi_serial - Probably safe not using locks - Maybe use cmpxchg * CALL-LEVEL PROCESSING > Old-call checking > rxrpc_input_implicit_end_call() > rxrpc_call_completed() > rxrpc_queue_call() ! Need to take rx->incoming_lock > __rxrpc_disconnect_call() > rxrpc_notify_socket() > rxrpc_new_incoming_call() - Uses rx->incoming_lock for the entire process - Might be able to drop this earlier in favour of the call lock > rxrpc_incoming_call() ! Conflicts with rxrpc_input_implicit_end_call() > rxrpc_send_ping() - Don't need locks to check rtt state > rxrpc_propose_ACK * PACKET DISTRIBUTION > rxrpc_input_call_packet() > rxrpc_input_data() * QUEUE DATA PACKET ON CALL > rxrpc_reduce_call_timer() - Uses timer_reduce() ! Needs call->input_lock() > rxrpc_receiving_reply() ! Needs locking around ack state > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_proto_abort() > rxrpc_input_dup_data() - Fills the Rx buffer - rxrpc_propose_ACK() - rxrpc_notify_socket() > rxrpc_input_ack() * APPLY ACK PACKET TO CALL AND DISCARD PACKET > rxrpc_input_ping_response() - Probably doesn't need any extra locking ! Need READ_ONCE() on call->ping_serial > rxrpc_input_check_for_lost_ack() - Takes call->lock to consult Tx buffer > rxrpc_peer_add_rtt() ! Needs to take a lock (peer->rtt_input_lock) ! Could perhaps manage with cmpxchg() and xadd() instead > rxrpc_input_requested_ack - Consults Tx buffer ! Probably needs a lock > rxrpc_peer_add_rtt() > rxrpc_propose_ack() > rxrpc_input_ackinfo() - Changes call->tx_winsize ! Use cmpxchg to handle change ! Should perhaps track serial number - Uses peer->lock to record MTU specification changes > rxrpc_proto_abort() ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_input_soft_acks() - Consults the Tx buffer > rxrpc_congestion_management() - Modifies the Tx annotations ! Needs call->input_lock() > rxrpc_queue_call() > rxrpc_input_abort() * APPLY ABORT PACKET TO CALL AND DISCARD PACKET > rxrpc_set_call_completion() > rxrpc_notify_socket() > rxrpc_input_ackall() * APPLY ACKALL PACKET TO CALL AND DISCARD PACKET ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_reject_packet() There are some functions used by the above that queue the packet, after which the procedure is terminated: - rxrpc_post_packet_to_local() - local->event_queue is an sk_buff_head - local->processor is a work_struct - rxrpc_post_packet_to_conn() - conn->rx_queue is an sk_buff_head - conn->processor is a work_struct - rxrpc_reject_packet() - local->reject_queue is an sk_buff_head - local->processor is a work_struct And some that offload processing to process context: - rxrpc_notify_socket() - Uses RCU lock - Uses call->notify_lock to call call->notify_rx - Uses call->recvmsg_lock to queue recvmsg side - rxrpc_queue_call() - call->processor is a work_struct - rxrpc_propose_ACK() - Uses call->lock to wrap __rxrpc_propose_ACK() And a bunch that complete a call, all of which use call->state_lock to protect the call state: - rxrpc_call_completed() - rxrpc_set_call_completion() - rxrpc_abort_call() - rxrpc_proto_abort() - Also uses rxrpc_queue_call() Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both") Signed-off-by: David Howells <dhowells@redhat.com>
2018-10-08 22:46:25 +08:00
}
flags = 0;
if (last)
flags |= RXRPC_LAST_PACKET;
if (!terminal)
flags |= RXRPC_JUMBO_PACKET;
if (test_bit(j, sp->rx_req_ack))
flags |= RXRPC_REQUEST_ACK;
trace_rxrpc_rx_data(call->debug_id, seq, serial, flags, annotation);
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
if (before_eq(seq, hard_ack)) {
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
ack = RXRPC_ACK_DUPLICATE;
ack_serial = serial;
continue;
}
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
if (call->rxtx_buffer[ix]) {
rxrpc_input_dup_data(call, seq, nr_subpackets > 1,
&jumbo_bad);
if (ack != RXRPC_ACK_DUPLICATE) {
ack = RXRPC_ACK_DUPLICATE;
ack_serial = serial;
}
immediate_ack = true;
continue;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
}
if (after(seq, hard_ack + call->rx_winsize)) {
ack = RXRPC_ACK_EXCEEDS_WINDOW;
ack_serial = serial;
if (flags & RXRPC_JUMBO_PACKET) {
if (!jumbo_bad) {
call->nr_jumbo_bad++;
jumbo_bad = true;
}
}
goto ack;
}
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
if (flags & RXRPC_REQUEST_ACK && !ack) {
ack = RXRPC_ACK_REQUESTED;
ack_serial = serial;
}
/* Queue the packet. We use a couple of memory barriers here as need
* to make sure that rx_top is perceived to be set after the buffer
* pointer and that the buffer pointer is set after the annotation and
* the skb data.
*
* Barriers against rxrpc_recvmsg_data() and rxrpc_rotate_rx_window()
* and also rxrpc_fill_out_ack().
*/
if (!terminal)
rxrpc_get_skb(skb, rxrpc_skb_got);
call->rxtx_annotations[ix] = annotation;
smp_wmb();
call->rxtx_buffer[ix] = skb;
if (after(seq, call->rx_top)) {
smp_store_release(&call->rx_top, seq);
} else if (before(seq, call->rx_top)) {
/* Send an immediate ACK if we fill in a hole */
if (!ack) {
ack = RXRPC_ACK_DELAY;
ack_serial = serial;
}
immediate_ack = true;
}
if (terminal) {
/* From this point on, we're not allowed to touch the
* packet any longer as its ref now belongs to the Rx
* ring.
*/
skb = NULL;
sp = NULL;
}
if (last) {
set_bit(RXRPC_CALL_RX_LAST, &call->flags);
if (!ack) {
ack = RXRPC_ACK_DELAY;
ack_serial = serial;
}
trace_rxrpc_receive(call, rxrpc_receive_queue_last, serial, seq);
} else {
trace_rxrpc_receive(call, rxrpc_receive_queue, serial, seq);
}
if (after_eq(seq, call->rx_expect_next)) {
if (after(seq, call->rx_expect_next)) {
_net("OOS %u > %u", seq, call->rx_expect_next);
ack = RXRPC_ACK_OUT_OF_SEQUENCE;
ack_serial = serial;
}
call->rx_expect_next = seq + 1;
}
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
}
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
ack:
if (ack)
rxrpc_propose_ACK(call, ack, ack_serial,
immediate_ack, true,
rxrpc_propose_ack_input_data);
else
rxrpc_propose_ACK(call, RXRPC_ACK_DELAY, serial,
false, true,
rxrpc_propose_ack_input_data);
trace_rxrpc_notify_socket(call->debug_id, serial);
rxrpc_notify_socket(call);
rxrpc: Fix the packet reception routine The rxrpc_input_packet() function and its call tree was built around the assumption that data_ready() handler called from UDP to inform a kernel service that there is data to be had was non-reentrant. This means that certain locking could be dispensed with. This, however, turns out not to be the case with a multi-queue network card that can deliver packets to multiple cpus simultaneously. Each of those cpus can be in the rxrpc_input_packet() function at the same time. Fix by adding or changing some structure members: (1) Add peer->rtt_input_lock to serialise access to the RTT buffer. (2) Make conn->service_id into a 32-bit variable so that it can be cmpxchg'd on all arches. (3) Add call->input_lock to serialise access to the Rx/Tx state. Note that although the Rx and Tx states are (almost) entirely separate, there's no point completing the separation and having separate locks since it's a bi-phasal RPC protocol rather than a bi-direction streaming protocol. Data transmission and data reception do not take place simultaneously on any particular call. and making the following functional changes: (1) In rxrpc_input_data(), hold call->input_lock around the core to prevent simultaneous producing of packets into the Rx ring and updating of tracking state for a particular call. (2) In rxrpc_input_ping_response(), only read call->ping_serial once, and check it before checking RXRPC_CALL_PINGING as that's a cheaper test. The bit test and bit clear can then be combined. No further locking is needed here. (3) In rxrpc_input_ack(), take call->input_lock after we've parsed much of the ACK packet. The superseded ACK check is then done both before and after the lock is taken. The handing of ackinfo data is split, parsing before the lock is taken and processing with it held. This is keyed on rxMTU being non-zero. Congestion management is also done within the locked section. (4) In rxrpc_input_ackall(), take call->input_lock around the Tx window rotation. The ACKALL packet carries no information and is only really useful after all packets have been transmitted since it's imprecise. (5) In rxrpc_input_implicit_end_call(), we use rx->incoming_lock to prevent calls being simultaneously implicitly ended on two cpus and also to prevent any races with incoming call setup. (6) In rxrpc_input_packet(), use cmpxchg() to effect the service upgrade on a connection. It is only permitted to happen once for a connection. (7) In rxrpc_new_incoming_call(), we have to recheck the routing inside rx->incoming_lock to see if someone else set up the call, connection or peer whilst we were getting there. We can't trust the values from the earlier routing check unless we pin refs on them - which we want to avoid. Further, we need to allow for an incoming call to have its state changed on another CPU between us making it live and us adjusting it because the conn is now in the RXRPC_CONN_SERVICE state. (8) In rxrpc_peer_add_rtt(), take peer->rtt_input_lock around the access to the RTT buffer. Don't need to lock around setting peer->rtt. For reference, the inventory of state-accessing or state-altering functions used by the packet input procedure is: > rxrpc_input_packet() * PACKET CHECKING * ROUTING > rxrpc_post_packet_to_local() > rxrpc_find_connection_rcu() - uses RCU > rxrpc_lookup_peer_rcu() - uses RCU > rxrpc_find_service_conn_rcu() - uses RCU > idr_find() - uses RCU * CONNECTION-LEVEL PROCESSING - Service upgrade - Can only happen once per conn ! Changed to use cmpxchg > rxrpc_post_packet_to_conn() - Setting conn->hi_serial - Probably safe not using locks - Maybe use cmpxchg * CALL-LEVEL PROCESSING > Old-call checking > rxrpc_input_implicit_end_call() > rxrpc_call_completed() > rxrpc_queue_call() ! Need to take rx->incoming_lock > __rxrpc_disconnect_call() > rxrpc_notify_socket() > rxrpc_new_incoming_call() - Uses rx->incoming_lock for the entire process - Might be able to drop this earlier in favour of the call lock > rxrpc_incoming_call() ! Conflicts with rxrpc_input_implicit_end_call() > rxrpc_send_ping() - Don't need locks to check rtt state > rxrpc_propose_ACK * PACKET DISTRIBUTION > rxrpc_input_call_packet() > rxrpc_input_data() * QUEUE DATA PACKET ON CALL > rxrpc_reduce_call_timer() - Uses timer_reduce() ! Needs call->input_lock() > rxrpc_receiving_reply() ! Needs locking around ack state > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_proto_abort() > rxrpc_input_dup_data() - Fills the Rx buffer - rxrpc_propose_ACK() - rxrpc_notify_socket() > rxrpc_input_ack() * APPLY ACK PACKET TO CALL AND DISCARD PACKET > rxrpc_input_ping_response() - Probably doesn't need any extra locking ! Need READ_ONCE() on call->ping_serial > rxrpc_input_check_for_lost_ack() - Takes call->lock to consult Tx buffer > rxrpc_peer_add_rtt() ! Needs to take a lock (peer->rtt_input_lock) ! Could perhaps manage with cmpxchg() and xadd() instead > rxrpc_input_requested_ack - Consults Tx buffer ! Probably needs a lock > rxrpc_peer_add_rtt() > rxrpc_propose_ack() > rxrpc_input_ackinfo() - Changes call->tx_winsize ! Use cmpxchg to handle change ! Should perhaps track serial number - Uses peer->lock to record MTU specification changes > rxrpc_proto_abort() ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_input_soft_acks() - Consults the Tx buffer > rxrpc_congestion_management() - Modifies the Tx annotations ! Needs call->input_lock() > rxrpc_queue_call() > rxrpc_input_abort() * APPLY ABORT PACKET TO CALL AND DISCARD PACKET > rxrpc_set_call_completion() > rxrpc_notify_socket() > rxrpc_input_ackall() * APPLY ACKALL PACKET TO CALL AND DISCARD PACKET ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_reject_packet() There are some functions used by the above that queue the packet, after which the procedure is terminated: - rxrpc_post_packet_to_local() - local->event_queue is an sk_buff_head - local->processor is a work_struct - rxrpc_post_packet_to_conn() - conn->rx_queue is an sk_buff_head - conn->processor is a work_struct - rxrpc_reject_packet() - local->reject_queue is an sk_buff_head - local->processor is a work_struct And some that offload processing to process context: - rxrpc_notify_socket() - Uses RCU lock - Uses call->notify_lock to call call->notify_rx - Uses call->recvmsg_lock to queue recvmsg side - rxrpc_queue_call() - call->processor is a work_struct - rxrpc_propose_ACK() - Uses call->lock to wrap __rxrpc_propose_ACK() And a bunch that complete a call, all of which use call->state_lock to protect the call state: - rxrpc_call_completed() - rxrpc_set_call_completion() - rxrpc_abort_call() - rxrpc_proto_abort() - Also uses rxrpc_queue_call() Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both") Signed-off-by: David Howells <dhowells@redhat.com>
2018-10-08 22:46:25 +08:00
unlock:
spin_unlock(&call->input_lock);
rxrpc_free_skb(skb, rxrpc_skb_freed);
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
_leave(" [queued]");
}
/*
* Process a requested ACK.
*/
static void rxrpc_input_requested_ack(struct rxrpc_call *call,
ktime_t resp_time,
rxrpc_serial_t orig_serial,
rxrpc_serial_t ack_serial)
{
struct rxrpc_skb_priv *sp;
struct sk_buff *skb;
ktime_t sent_at;
int ix;
for (ix = 0; ix < RXRPC_RXTX_BUFF_SIZE; ix++) {
skb = call->rxtx_buffer[ix];
if (!skb)
continue;
sent_at = skb->tstamp;
smp_rmb(); /* Read timestamp before serial. */
sp = rxrpc_skb(skb);
if (sp->hdr.serial != orig_serial)
continue;
goto found;
}
return;
found:
rxrpc_peer_add_rtt(call, rxrpc_rtt_rx_requested_ack,
orig_serial, ack_serial, sent_at, resp_time);
}
/*
* Process the response to a ping that we sent to find out if we lost an ACK.
*
* If we got back a ping response that indicates a lower tx_top than what we
* had at the time of the ping transmission, we adjudge all the DATA packets
* sent between the response tx_top and the ping-time tx_top to have been lost.
*/
static void rxrpc_input_check_for_lost_ack(struct rxrpc_call *call)
{
rxrpc_seq_t top, bottom, seq;
bool resend = false;
spin_lock_bh(&call->lock);
bottom = call->tx_hard_ack + 1;
top = call->acks_lost_top;
if (before(bottom, top)) {
for (seq = bottom; before_eq(seq, top); seq++) {
int ix = seq & RXRPC_RXTX_BUFF_MASK;
u8 annotation = call->rxtx_annotations[ix];
u8 anno_type = annotation & RXRPC_TX_ANNO_MASK;
if (anno_type != RXRPC_TX_ANNO_UNACK)
continue;
annotation &= ~RXRPC_TX_ANNO_MASK;
annotation |= RXRPC_TX_ANNO_RETRANS;
call->rxtx_annotations[ix] = annotation;
resend = true;
}
}
spin_unlock_bh(&call->lock);
if (resend && !test_and_set_bit(RXRPC_CALL_EV_RESEND, &call->events))
rxrpc_queue_call(call);
}
/*
* Process a ping response.
*/
static void rxrpc_input_ping_response(struct rxrpc_call *call,
ktime_t resp_time,
rxrpc_serial_t orig_serial,
rxrpc_serial_t ack_serial)
{
rxrpc_serial_t ping_serial;
ktime_t ping_time;
ping_time = call->ping_time;
smp_rmb();
rxrpc: Fix the packet reception routine The rxrpc_input_packet() function and its call tree was built around the assumption that data_ready() handler called from UDP to inform a kernel service that there is data to be had was non-reentrant. This means that certain locking could be dispensed with. This, however, turns out not to be the case with a multi-queue network card that can deliver packets to multiple cpus simultaneously. Each of those cpus can be in the rxrpc_input_packet() function at the same time. Fix by adding or changing some structure members: (1) Add peer->rtt_input_lock to serialise access to the RTT buffer. (2) Make conn->service_id into a 32-bit variable so that it can be cmpxchg'd on all arches. (3) Add call->input_lock to serialise access to the Rx/Tx state. Note that although the Rx and Tx states are (almost) entirely separate, there's no point completing the separation and having separate locks since it's a bi-phasal RPC protocol rather than a bi-direction streaming protocol. Data transmission and data reception do not take place simultaneously on any particular call. and making the following functional changes: (1) In rxrpc_input_data(), hold call->input_lock around the core to prevent simultaneous producing of packets into the Rx ring and updating of tracking state for a particular call. (2) In rxrpc_input_ping_response(), only read call->ping_serial once, and check it before checking RXRPC_CALL_PINGING as that's a cheaper test. The bit test and bit clear can then be combined. No further locking is needed here. (3) In rxrpc_input_ack(), take call->input_lock after we've parsed much of the ACK packet. The superseded ACK check is then done both before and after the lock is taken. The handing of ackinfo data is split, parsing before the lock is taken and processing with it held. This is keyed on rxMTU being non-zero. Congestion management is also done within the locked section. (4) In rxrpc_input_ackall(), take call->input_lock around the Tx window rotation. The ACKALL packet carries no information and is only really useful after all packets have been transmitted since it's imprecise. (5) In rxrpc_input_implicit_end_call(), we use rx->incoming_lock to prevent calls being simultaneously implicitly ended on two cpus and also to prevent any races with incoming call setup. (6) In rxrpc_input_packet(), use cmpxchg() to effect the service upgrade on a connection. It is only permitted to happen once for a connection. (7) In rxrpc_new_incoming_call(), we have to recheck the routing inside rx->incoming_lock to see if someone else set up the call, connection or peer whilst we were getting there. We can't trust the values from the earlier routing check unless we pin refs on them - which we want to avoid. Further, we need to allow for an incoming call to have its state changed on another CPU between us making it live and us adjusting it because the conn is now in the RXRPC_CONN_SERVICE state. (8) In rxrpc_peer_add_rtt(), take peer->rtt_input_lock around the access to the RTT buffer. Don't need to lock around setting peer->rtt. For reference, the inventory of state-accessing or state-altering functions used by the packet input procedure is: > rxrpc_input_packet() * PACKET CHECKING * ROUTING > rxrpc_post_packet_to_local() > rxrpc_find_connection_rcu() - uses RCU > rxrpc_lookup_peer_rcu() - uses RCU > rxrpc_find_service_conn_rcu() - uses RCU > idr_find() - uses RCU * CONNECTION-LEVEL PROCESSING - Service upgrade - Can only happen once per conn ! Changed to use cmpxchg > rxrpc_post_packet_to_conn() - Setting conn->hi_serial - Probably safe not using locks - Maybe use cmpxchg * CALL-LEVEL PROCESSING > Old-call checking > rxrpc_input_implicit_end_call() > rxrpc_call_completed() > rxrpc_queue_call() ! Need to take rx->incoming_lock > __rxrpc_disconnect_call() > rxrpc_notify_socket() > rxrpc_new_incoming_call() - Uses rx->incoming_lock for the entire process - Might be able to drop this earlier in favour of the call lock > rxrpc_incoming_call() ! Conflicts with rxrpc_input_implicit_end_call() > rxrpc_send_ping() - Don't need locks to check rtt state > rxrpc_propose_ACK * PACKET DISTRIBUTION > rxrpc_input_call_packet() > rxrpc_input_data() * QUEUE DATA PACKET ON CALL > rxrpc_reduce_call_timer() - Uses timer_reduce() ! Needs call->input_lock() > rxrpc_receiving_reply() ! Needs locking around ack state > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_proto_abort() > rxrpc_input_dup_data() - Fills the Rx buffer - rxrpc_propose_ACK() - rxrpc_notify_socket() > rxrpc_input_ack() * APPLY ACK PACKET TO CALL AND DISCARD PACKET > rxrpc_input_ping_response() - Probably doesn't need any extra locking ! Need READ_ONCE() on call->ping_serial > rxrpc_input_check_for_lost_ack() - Takes call->lock to consult Tx buffer > rxrpc_peer_add_rtt() ! Needs to take a lock (peer->rtt_input_lock) ! Could perhaps manage with cmpxchg() and xadd() instead > rxrpc_input_requested_ack - Consults Tx buffer ! Probably needs a lock > rxrpc_peer_add_rtt() > rxrpc_propose_ack() > rxrpc_input_ackinfo() - Changes call->tx_winsize ! Use cmpxchg to handle change ! Should perhaps track serial number - Uses peer->lock to record MTU specification changes > rxrpc_proto_abort() ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_input_soft_acks() - Consults the Tx buffer > rxrpc_congestion_management() - Modifies the Tx annotations ! Needs call->input_lock() > rxrpc_queue_call() > rxrpc_input_abort() * APPLY ABORT PACKET TO CALL AND DISCARD PACKET > rxrpc_set_call_completion() > rxrpc_notify_socket() > rxrpc_input_ackall() * APPLY ACKALL PACKET TO CALL AND DISCARD PACKET ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_reject_packet() There are some functions used by the above that queue the packet, after which the procedure is terminated: - rxrpc_post_packet_to_local() - local->event_queue is an sk_buff_head - local->processor is a work_struct - rxrpc_post_packet_to_conn() - conn->rx_queue is an sk_buff_head - conn->processor is a work_struct - rxrpc_reject_packet() - local->reject_queue is an sk_buff_head - local->processor is a work_struct And some that offload processing to process context: - rxrpc_notify_socket() - Uses RCU lock - Uses call->notify_lock to call call->notify_rx - Uses call->recvmsg_lock to queue recvmsg side - rxrpc_queue_call() - call->processor is a work_struct - rxrpc_propose_ACK() - Uses call->lock to wrap __rxrpc_propose_ACK() And a bunch that complete a call, all of which use call->state_lock to protect the call state: - rxrpc_call_completed() - rxrpc_set_call_completion() - rxrpc_abort_call() - rxrpc_proto_abort() - Also uses rxrpc_queue_call() Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both") Signed-off-by: David Howells <dhowells@redhat.com>
2018-10-08 22:46:25 +08:00
ping_serial = READ_ONCE(call->ping_serial);
if (orig_serial == call->acks_lost_ping)
rxrpc_input_check_for_lost_ack(call);
rxrpc: Fix the packet reception routine The rxrpc_input_packet() function and its call tree was built around the assumption that data_ready() handler called from UDP to inform a kernel service that there is data to be had was non-reentrant. This means that certain locking could be dispensed with. This, however, turns out not to be the case with a multi-queue network card that can deliver packets to multiple cpus simultaneously. Each of those cpus can be in the rxrpc_input_packet() function at the same time. Fix by adding or changing some structure members: (1) Add peer->rtt_input_lock to serialise access to the RTT buffer. (2) Make conn->service_id into a 32-bit variable so that it can be cmpxchg'd on all arches. (3) Add call->input_lock to serialise access to the Rx/Tx state. Note that although the Rx and Tx states are (almost) entirely separate, there's no point completing the separation and having separate locks since it's a bi-phasal RPC protocol rather than a bi-direction streaming protocol. Data transmission and data reception do not take place simultaneously on any particular call. and making the following functional changes: (1) In rxrpc_input_data(), hold call->input_lock around the core to prevent simultaneous producing of packets into the Rx ring and updating of tracking state for a particular call. (2) In rxrpc_input_ping_response(), only read call->ping_serial once, and check it before checking RXRPC_CALL_PINGING as that's a cheaper test. The bit test and bit clear can then be combined. No further locking is needed here. (3) In rxrpc_input_ack(), take call->input_lock after we've parsed much of the ACK packet. The superseded ACK check is then done both before and after the lock is taken. The handing of ackinfo data is split, parsing before the lock is taken and processing with it held. This is keyed on rxMTU being non-zero. Congestion management is also done within the locked section. (4) In rxrpc_input_ackall(), take call->input_lock around the Tx window rotation. The ACKALL packet carries no information and is only really useful after all packets have been transmitted since it's imprecise. (5) In rxrpc_input_implicit_end_call(), we use rx->incoming_lock to prevent calls being simultaneously implicitly ended on two cpus and also to prevent any races with incoming call setup. (6) In rxrpc_input_packet(), use cmpxchg() to effect the service upgrade on a connection. It is only permitted to happen once for a connection. (7) In rxrpc_new_incoming_call(), we have to recheck the routing inside rx->incoming_lock to see if someone else set up the call, connection or peer whilst we were getting there. We can't trust the values from the earlier routing check unless we pin refs on them - which we want to avoid. Further, we need to allow for an incoming call to have its state changed on another CPU between us making it live and us adjusting it because the conn is now in the RXRPC_CONN_SERVICE state. (8) In rxrpc_peer_add_rtt(), take peer->rtt_input_lock around the access to the RTT buffer. Don't need to lock around setting peer->rtt. For reference, the inventory of state-accessing or state-altering functions used by the packet input procedure is: > rxrpc_input_packet() * PACKET CHECKING * ROUTING > rxrpc_post_packet_to_local() > rxrpc_find_connection_rcu() - uses RCU > rxrpc_lookup_peer_rcu() - uses RCU > rxrpc_find_service_conn_rcu() - uses RCU > idr_find() - uses RCU * CONNECTION-LEVEL PROCESSING - Service upgrade - Can only happen once per conn ! Changed to use cmpxchg > rxrpc_post_packet_to_conn() - Setting conn->hi_serial - Probably safe not using locks - Maybe use cmpxchg * CALL-LEVEL PROCESSING > Old-call checking > rxrpc_input_implicit_end_call() > rxrpc_call_completed() > rxrpc_queue_call() ! Need to take rx->incoming_lock > __rxrpc_disconnect_call() > rxrpc_notify_socket() > rxrpc_new_incoming_call() - Uses rx->incoming_lock for the entire process - Might be able to drop this earlier in favour of the call lock > rxrpc_incoming_call() ! Conflicts with rxrpc_input_implicit_end_call() > rxrpc_send_ping() - Don't need locks to check rtt state > rxrpc_propose_ACK * PACKET DISTRIBUTION > rxrpc_input_call_packet() > rxrpc_input_data() * QUEUE DATA PACKET ON CALL > rxrpc_reduce_call_timer() - Uses timer_reduce() ! Needs call->input_lock() > rxrpc_receiving_reply() ! Needs locking around ack state > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_proto_abort() > rxrpc_input_dup_data() - Fills the Rx buffer - rxrpc_propose_ACK() - rxrpc_notify_socket() > rxrpc_input_ack() * APPLY ACK PACKET TO CALL AND DISCARD PACKET > rxrpc_input_ping_response() - Probably doesn't need any extra locking ! Need READ_ONCE() on call->ping_serial > rxrpc_input_check_for_lost_ack() - Takes call->lock to consult Tx buffer > rxrpc_peer_add_rtt() ! Needs to take a lock (peer->rtt_input_lock) ! Could perhaps manage with cmpxchg() and xadd() instead > rxrpc_input_requested_ack - Consults Tx buffer ! Probably needs a lock > rxrpc_peer_add_rtt() > rxrpc_propose_ack() > rxrpc_input_ackinfo() - Changes call->tx_winsize ! Use cmpxchg to handle change ! Should perhaps track serial number - Uses peer->lock to record MTU specification changes > rxrpc_proto_abort() ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_input_soft_acks() - Consults the Tx buffer > rxrpc_congestion_management() - Modifies the Tx annotations ! Needs call->input_lock() > rxrpc_queue_call() > rxrpc_input_abort() * APPLY ABORT PACKET TO CALL AND DISCARD PACKET > rxrpc_set_call_completion() > rxrpc_notify_socket() > rxrpc_input_ackall() * APPLY ACKALL PACKET TO CALL AND DISCARD PACKET ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_reject_packet() There are some functions used by the above that queue the packet, after which the procedure is terminated: - rxrpc_post_packet_to_local() - local->event_queue is an sk_buff_head - local->processor is a work_struct - rxrpc_post_packet_to_conn() - conn->rx_queue is an sk_buff_head - conn->processor is a work_struct - rxrpc_reject_packet() - local->reject_queue is an sk_buff_head - local->processor is a work_struct And some that offload processing to process context: - rxrpc_notify_socket() - Uses RCU lock - Uses call->notify_lock to call call->notify_rx - Uses call->recvmsg_lock to queue recvmsg side - rxrpc_queue_call() - call->processor is a work_struct - rxrpc_propose_ACK() - Uses call->lock to wrap __rxrpc_propose_ACK() And a bunch that complete a call, all of which use call->state_lock to protect the call state: - rxrpc_call_completed() - rxrpc_set_call_completion() - rxrpc_abort_call() - rxrpc_proto_abort() - Also uses rxrpc_queue_call() Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both") Signed-off-by: David Howells <dhowells@redhat.com>
2018-10-08 22:46:25 +08:00
if (before(orig_serial, ping_serial) ||
!test_and_clear_bit(RXRPC_CALL_PINGING, &call->flags))
return;
if (after(orig_serial, ping_serial))
return;
rxrpc_peer_add_rtt(call, rxrpc_rtt_rx_ping_response,
orig_serial, ack_serial, ping_time, resp_time);
}
/*
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
* Process the extra information that may be appended to an ACK packet
*/
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
static void rxrpc_input_ackinfo(struct rxrpc_call *call, struct sk_buff *skb,
struct rxrpc_ackinfo *ackinfo)
{
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
struct rxrpc_peer *peer;
unsigned int mtu;
bool wake = false;
u32 rwind = ntohl(ackinfo->rwind);
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
_proto("Rx ACK %%%u Info { rx=%u max=%u rwin=%u jm=%u }",
sp->hdr.serial,
ntohl(ackinfo->rxMTU), ntohl(ackinfo->maxMTU),
rwind, ntohl(ackinfo->jumbo_max));
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
if (call->tx_winsize != rwind) {
if (rwind > RXRPC_RXTX_BUFF_SIZE - 1)
rwind = RXRPC_RXTX_BUFF_SIZE - 1;
if (rwind > call->tx_winsize)
wake = true;
trace_rxrpc_rx_rwind_change(call, sp->hdr.serial,
ntohl(ackinfo->rwind), wake);
call->tx_winsize = rwind;
}
if (call->cong_ssthresh > rwind)
call->cong_ssthresh = rwind;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
mtu = min(ntohl(ackinfo->rxMTU), ntohl(ackinfo->maxMTU));
peer = call->peer;
if (mtu < peer->maxdata) {
spin_lock_bh(&peer->lock);
peer->maxdata = mtu;
peer->mtu = mtu + peer->hdrsize;
spin_unlock_bh(&peer->lock);
_net("Net MTU %u (maxdata %u)", peer->mtu, peer->maxdata);
}
if (wake)
wake_up(&call->waitq);
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
}
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
/*
* Process individual soft ACKs.
*
* Each ACK in the array corresponds to one packet and can be either an ACK or
* a NAK. If we get find an explicitly NAK'd packet we resend immediately;
* packets that lie beyond the end of the ACK list are scheduled for resend by
* the timer on the basis that the peer might just not have processed them at
* the time the ACK was sent.
*/
static void rxrpc_input_soft_acks(struct rxrpc_call *call, u8 *acks,
rxrpc_seq_t seq, int nr_acks,
struct rxrpc_ack_summary *summary)
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
{
int ix;
u8 annotation, anno_type;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
for (; nr_acks > 0; nr_acks--, seq++) {
ix = seq & RXRPC_RXTX_BUFF_MASK;
annotation = call->rxtx_annotations[ix];
anno_type = annotation & RXRPC_TX_ANNO_MASK;
annotation &= ~RXRPC_TX_ANNO_MASK;
switch (*acks++) {
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
case RXRPC_ACK_TYPE_ACK:
summary->nr_acks++;
if (anno_type == RXRPC_TX_ANNO_ACK)
continue;
summary->nr_new_acks++;
call->rxtx_annotations[ix] =
RXRPC_TX_ANNO_ACK | annotation;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
break;
case RXRPC_ACK_TYPE_NACK:
if (!summary->nr_nacks &&
call->acks_lowest_nak != seq) {
call->acks_lowest_nak = seq;
summary->new_low_nack = true;
}
summary->nr_nacks++;
if (anno_type == RXRPC_TX_ANNO_NAK)
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
continue;
summary->nr_new_nacks++;
if (anno_type == RXRPC_TX_ANNO_RETRANS)
continue;
call->rxtx_annotations[ix] =
RXRPC_TX_ANNO_NAK | annotation;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
break;
default:
return rxrpc_proto_abort("SFT", call, 0);
}
}
}
/*
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
* Process an ACK packet.
*
* ack.firstPacket is the sequence number of the first soft-ACK'd/NAK'd packet
* in the ACK array. Anything before that is hard-ACK'd and may be discarded.
*
* A hard-ACK means that a packet has been processed and may be discarded; a
* soft-ACK means that the packet may be discarded and retransmission
* requested. A phase is complete when all packets are hard-ACK'd.
*/
static void rxrpc_input_ack(struct rxrpc_call *call, struct sk_buff *skb)
{
struct rxrpc_ack_summary summary = { 0 };
struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
union {
struct rxrpc_ackpacket ack;
struct rxrpc_ackinfo info;
u8 acks[RXRPC_MAXACKS];
} buf;
rxrpc_serial_t acked_serial;
rxrpc_seq_t first_soft_ack, hard_ack, prev_pkt;
int nr_acks, offset, ioffset;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
_enter("");
offset = sizeof(struct rxrpc_wire_header);
if (skb_copy_bits(skb, offset, &buf.ack, sizeof(buf.ack)) < 0) {
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
_debug("extraction failure");
return rxrpc_proto_abort("XAK", call, 0);
}
offset += sizeof(buf.ack);
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
acked_serial = ntohl(buf.ack.serial);
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
first_soft_ack = ntohl(buf.ack.firstPacket);
prev_pkt = ntohl(buf.ack.previousPacket);
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
hard_ack = first_soft_ack - 1;
nr_acks = buf.ack.nAcks;
summary.ack_reason = (buf.ack.reason < RXRPC_ACK__INVALID ?
buf.ack.reason : RXRPC_ACK__INVALID);
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
trace_rxrpc_rx_ack(call, sp->hdr.serial, acked_serial,
first_soft_ack, prev_pkt,
summary.ack_reason, nr_acks);
if (buf.ack.reason == RXRPC_ACK_PING_RESPONSE)
rxrpc_input_ping_response(call, skb->tstamp, acked_serial,
sp->hdr.serial);
if (buf.ack.reason == RXRPC_ACK_REQUESTED)
rxrpc_input_requested_ack(call, skb->tstamp, acked_serial,
sp->hdr.serial);
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
if (buf.ack.reason == RXRPC_ACK_PING) {
_proto("Rx ACK %%%u PING Request", sp->hdr.serial);
rxrpc_propose_ACK(call, RXRPC_ACK_PING_RESPONSE,
sp->hdr.serial, true, true,
rxrpc_propose_ack_respond_to_ping);
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
} else if (sp->hdr.flags & RXRPC_REQUEST_ACK) {
rxrpc_propose_ACK(call, RXRPC_ACK_REQUESTED,
sp->hdr.serial, true, true,
rxrpc_propose_ack_respond_to_ack);
}
/* Discard any out-of-order or duplicate ACKs (outside lock). */
if (before(first_soft_ack, call->ackr_first_seq) ||
before(prev_pkt, call->ackr_prev_seq))
return;
rxrpc: Fix the packet reception routine The rxrpc_input_packet() function and its call tree was built around the assumption that data_ready() handler called from UDP to inform a kernel service that there is data to be had was non-reentrant. This means that certain locking could be dispensed with. This, however, turns out not to be the case with a multi-queue network card that can deliver packets to multiple cpus simultaneously. Each of those cpus can be in the rxrpc_input_packet() function at the same time. Fix by adding or changing some structure members: (1) Add peer->rtt_input_lock to serialise access to the RTT buffer. (2) Make conn->service_id into a 32-bit variable so that it can be cmpxchg'd on all arches. (3) Add call->input_lock to serialise access to the Rx/Tx state. Note that although the Rx and Tx states are (almost) entirely separate, there's no point completing the separation and having separate locks since it's a bi-phasal RPC protocol rather than a bi-direction streaming protocol. Data transmission and data reception do not take place simultaneously on any particular call. and making the following functional changes: (1) In rxrpc_input_data(), hold call->input_lock around the core to prevent simultaneous producing of packets into the Rx ring and updating of tracking state for a particular call. (2) In rxrpc_input_ping_response(), only read call->ping_serial once, and check it before checking RXRPC_CALL_PINGING as that's a cheaper test. The bit test and bit clear can then be combined. No further locking is needed here. (3) In rxrpc_input_ack(), take call->input_lock after we've parsed much of the ACK packet. The superseded ACK check is then done both before and after the lock is taken. The handing of ackinfo data is split, parsing before the lock is taken and processing with it held. This is keyed on rxMTU being non-zero. Congestion management is also done within the locked section. (4) In rxrpc_input_ackall(), take call->input_lock around the Tx window rotation. The ACKALL packet carries no information and is only really useful after all packets have been transmitted since it's imprecise. (5) In rxrpc_input_implicit_end_call(), we use rx->incoming_lock to prevent calls being simultaneously implicitly ended on two cpus and also to prevent any races with incoming call setup. (6) In rxrpc_input_packet(), use cmpxchg() to effect the service upgrade on a connection. It is only permitted to happen once for a connection. (7) In rxrpc_new_incoming_call(), we have to recheck the routing inside rx->incoming_lock to see if someone else set up the call, connection or peer whilst we were getting there. We can't trust the values from the earlier routing check unless we pin refs on them - which we want to avoid. Further, we need to allow for an incoming call to have its state changed on another CPU between us making it live and us adjusting it because the conn is now in the RXRPC_CONN_SERVICE state. (8) In rxrpc_peer_add_rtt(), take peer->rtt_input_lock around the access to the RTT buffer. Don't need to lock around setting peer->rtt. For reference, the inventory of state-accessing or state-altering functions used by the packet input procedure is: > rxrpc_input_packet() * PACKET CHECKING * ROUTING > rxrpc_post_packet_to_local() > rxrpc_find_connection_rcu() - uses RCU > rxrpc_lookup_peer_rcu() - uses RCU > rxrpc_find_service_conn_rcu() - uses RCU > idr_find() - uses RCU * CONNECTION-LEVEL PROCESSING - Service upgrade - Can only happen once per conn ! Changed to use cmpxchg > rxrpc_post_packet_to_conn() - Setting conn->hi_serial - Probably safe not using locks - Maybe use cmpxchg * CALL-LEVEL PROCESSING > Old-call checking > rxrpc_input_implicit_end_call() > rxrpc_call_completed() > rxrpc_queue_call() ! Need to take rx->incoming_lock > __rxrpc_disconnect_call() > rxrpc_notify_socket() > rxrpc_new_incoming_call() - Uses rx->incoming_lock for the entire process - Might be able to drop this earlier in favour of the call lock > rxrpc_incoming_call() ! Conflicts with rxrpc_input_implicit_end_call() > rxrpc_send_ping() - Don't need locks to check rtt state > rxrpc_propose_ACK * PACKET DISTRIBUTION > rxrpc_input_call_packet() > rxrpc_input_data() * QUEUE DATA PACKET ON CALL > rxrpc_reduce_call_timer() - Uses timer_reduce() ! Needs call->input_lock() > rxrpc_receiving_reply() ! Needs locking around ack state > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_proto_abort() > rxrpc_input_dup_data() - Fills the Rx buffer - rxrpc_propose_ACK() - rxrpc_notify_socket() > rxrpc_input_ack() * APPLY ACK PACKET TO CALL AND DISCARD PACKET > rxrpc_input_ping_response() - Probably doesn't need any extra locking ! Need READ_ONCE() on call->ping_serial > rxrpc_input_check_for_lost_ack() - Takes call->lock to consult Tx buffer > rxrpc_peer_add_rtt() ! Needs to take a lock (peer->rtt_input_lock) ! Could perhaps manage with cmpxchg() and xadd() instead > rxrpc_input_requested_ack - Consults Tx buffer ! Probably needs a lock > rxrpc_peer_add_rtt() > rxrpc_propose_ack() > rxrpc_input_ackinfo() - Changes call->tx_winsize ! Use cmpxchg to handle change ! Should perhaps track serial number - Uses peer->lock to record MTU specification changes > rxrpc_proto_abort() ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_input_soft_acks() - Consults the Tx buffer > rxrpc_congestion_management() - Modifies the Tx annotations ! Needs call->input_lock() > rxrpc_queue_call() > rxrpc_input_abort() * APPLY ABORT PACKET TO CALL AND DISCARD PACKET > rxrpc_set_call_completion() > rxrpc_notify_socket() > rxrpc_input_ackall() * APPLY ACKALL PACKET TO CALL AND DISCARD PACKET ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_reject_packet() There are some functions used by the above that queue the packet, after which the procedure is terminated: - rxrpc_post_packet_to_local() - local->event_queue is an sk_buff_head - local->processor is a work_struct - rxrpc_post_packet_to_conn() - conn->rx_queue is an sk_buff_head - conn->processor is a work_struct - rxrpc_reject_packet() - local->reject_queue is an sk_buff_head - local->processor is a work_struct And some that offload processing to process context: - rxrpc_notify_socket() - Uses RCU lock - Uses call->notify_lock to call call->notify_rx - Uses call->recvmsg_lock to queue recvmsg side - rxrpc_queue_call() - call->processor is a work_struct - rxrpc_propose_ACK() - Uses call->lock to wrap __rxrpc_propose_ACK() And a bunch that complete a call, all of which use call->state_lock to protect the call state: - rxrpc_call_completed() - rxrpc_set_call_completion() - rxrpc_abort_call() - rxrpc_proto_abort() - Also uses rxrpc_queue_call() Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both") Signed-off-by: David Howells <dhowells@redhat.com>
2018-10-08 22:46:25 +08:00
buf.info.rxMTU = 0;
ioffset = offset + nr_acks + 3;
rxrpc: Fix the packet reception routine The rxrpc_input_packet() function and its call tree was built around the assumption that data_ready() handler called from UDP to inform a kernel service that there is data to be had was non-reentrant. This means that certain locking could be dispensed with. This, however, turns out not to be the case with a multi-queue network card that can deliver packets to multiple cpus simultaneously. Each of those cpus can be in the rxrpc_input_packet() function at the same time. Fix by adding or changing some structure members: (1) Add peer->rtt_input_lock to serialise access to the RTT buffer. (2) Make conn->service_id into a 32-bit variable so that it can be cmpxchg'd on all arches. (3) Add call->input_lock to serialise access to the Rx/Tx state. Note that although the Rx and Tx states are (almost) entirely separate, there's no point completing the separation and having separate locks since it's a bi-phasal RPC protocol rather than a bi-direction streaming protocol. Data transmission and data reception do not take place simultaneously on any particular call. and making the following functional changes: (1) In rxrpc_input_data(), hold call->input_lock around the core to prevent simultaneous producing of packets into the Rx ring and updating of tracking state for a particular call. (2) In rxrpc_input_ping_response(), only read call->ping_serial once, and check it before checking RXRPC_CALL_PINGING as that's a cheaper test. The bit test and bit clear can then be combined. No further locking is needed here. (3) In rxrpc_input_ack(), take call->input_lock after we've parsed much of the ACK packet. The superseded ACK check is then done both before and after the lock is taken. The handing of ackinfo data is split, parsing before the lock is taken and processing with it held. This is keyed on rxMTU being non-zero. Congestion management is also done within the locked section. (4) In rxrpc_input_ackall(), take call->input_lock around the Tx window rotation. The ACKALL packet carries no information and is only really useful after all packets have been transmitted since it's imprecise. (5) In rxrpc_input_implicit_end_call(), we use rx->incoming_lock to prevent calls being simultaneously implicitly ended on two cpus and also to prevent any races with incoming call setup. (6) In rxrpc_input_packet(), use cmpxchg() to effect the service upgrade on a connection. It is only permitted to happen once for a connection. (7) In rxrpc_new_incoming_call(), we have to recheck the routing inside rx->incoming_lock to see if someone else set up the call, connection or peer whilst we were getting there. We can't trust the values from the earlier routing check unless we pin refs on them - which we want to avoid. Further, we need to allow for an incoming call to have its state changed on another CPU between us making it live and us adjusting it because the conn is now in the RXRPC_CONN_SERVICE state. (8) In rxrpc_peer_add_rtt(), take peer->rtt_input_lock around the access to the RTT buffer. Don't need to lock around setting peer->rtt. For reference, the inventory of state-accessing or state-altering functions used by the packet input procedure is: > rxrpc_input_packet() * PACKET CHECKING * ROUTING > rxrpc_post_packet_to_local() > rxrpc_find_connection_rcu() - uses RCU > rxrpc_lookup_peer_rcu() - uses RCU > rxrpc_find_service_conn_rcu() - uses RCU > idr_find() - uses RCU * CONNECTION-LEVEL PROCESSING - Service upgrade - Can only happen once per conn ! Changed to use cmpxchg > rxrpc_post_packet_to_conn() - Setting conn->hi_serial - Probably safe not using locks - Maybe use cmpxchg * CALL-LEVEL PROCESSING > Old-call checking > rxrpc_input_implicit_end_call() > rxrpc_call_completed() > rxrpc_queue_call() ! Need to take rx->incoming_lock > __rxrpc_disconnect_call() > rxrpc_notify_socket() > rxrpc_new_incoming_call() - Uses rx->incoming_lock for the entire process - Might be able to drop this earlier in favour of the call lock > rxrpc_incoming_call() ! Conflicts with rxrpc_input_implicit_end_call() > rxrpc_send_ping() - Don't need locks to check rtt state > rxrpc_propose_ACK * PACKET DISTRIBUTION > rxrpc_input_call_packet() > rxrpc_input_data() * QUEUE DATA PACKET ON CALL > rxrpc_reduce_call_timer() - Uses timer_reduce() ! Needs call->input_lock() > rxrpc_receiving_reply() ! Needs locking around ack state > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_proto_abort() > rxrpc_input_dup_data() - Fills the Rx buffer - rxrpc_propose_ACK() - rxrpc_notify_socket() > rxrpc_input_ack() * APPLY ACK PACKET TO CALL AND DISCARD PACKET > rxrpc_input_ping_response() - Probably doesn't need any extra locking ! Need READ_ONCE() on call->ping_serial > rxrpc_input_check_for_lost_ack() - Takes call->lock to consult Tx buffer > rxrpc_peer_add_rtt() ! Needs to take a lock (peer->rtt_input_lock) ! Could perhaps manage with cmpxchg() and xadd() instead > rxrpc_input_requested_ack - Consults Tx buffer ! Probably needs a lock > rxrpc_peer_add_rtt() > rxrpc_propose_ack() > rxrpc_input_ackinfo() - Changes call->tx_winsize ! Use cmpxchg to handle change ! Should perhaps track serial number - Uses peer->lock to record MTU specification changes > rxrpc_proto_abort() ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_input_soft_acks() - Consults the Tx buffer > rxrpc_congestion_management() - Modifies the Tx annotations ! Needs call->input_lock() > rxrpc_queue_call() > rxrpc_input_abort() * APPLY ABORT PACKET TO CALL AND DISCARD PACKET > rxrpc_set_call_completion() > rxrpc_notify_socket() > rxrpc_input_ackall() * APPLY ACKALL PACKET TO CALL AND DISCARD PACKET ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_reject_packet() There are some functions used by the above that queue the packet, after which the procedure is terminated: - rxrpc_post_packet_to_local() - local->event_queue is an sk_buff_head - local->processor is a work_struct - rxrpc_post_packet_to_conn() - conn->rx_queue is an sk_buff_head - conn->processor is a work_struct - rxrpc_reject_packet() - local->reject_queue is an sk_buff_head - local->processor is a work_struct And some that offload processing to process context: - rxrpc_notify_socket() - Uses RCU lock - Uses call->notify_lock to call call->notify_rx - Uses call->recvmsg_lock to queue recvmsg side - rxrpc_queue_call() - call->processor is a work_struct - rxrpc_propose_ACK() - Uses call->lock to wrap __rxrpc_propose_ACK() And a bunch that complete a call, all of which use call->state_lock to protect the call state: - rxrpc_call_completed() - rxrpc_set_call_completion() - rxrpc_abort_call() - rxrpc_proto_abort() - Also uses rxrpc_queue_call() Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both") Signed-off-by: David Howells <dhowells@redhat.com>
2018-10-08 22:46:25 +08:00
if (skb->len >= ioffset + sizeof(buf.info) &&
skb_copy_bits(skb, ioffset, &buf.info, sizeof(buf.info)) < 0)
return rxrpc_proto_abort("XAI", call, 0);
spin_lock(&call->input_lock);
/* Discard any out-of-order or duplicate ACKs (inside lock). */
if (before(first_soft_ack, call->ackr_first_seq) ||
before(prev_pkt, call->ackr_prev_seq))
rxrpc: Fix the packet reception routine The rxrpc_input_packet() function and its call tree was built around the assumption that data_ready() handler called from UDP to inform a kernel service that there is data to be had was non-reentrant. This means that certain locking could be dispensed with. This, however, turns out not to be the case with a multi-queue network card that can deliver packets to multiple cpus simultaneously. Each of those cpus can be in the rxrpc_input_packet() function at the same time. Fix by adding or changing some structure members: (1) Add peer->rtt_input_lock to serialise access to the RTT buffer. (2) Make conn->service_id into a 32-bit variable so that it can be cmpxchg'd on all arches. (3) Add call->input_lock to serialise access to the Rx/Tx state. Note that although the Rx and Tx states are (almost) entirely separate, there's no point completing the separation and having separate locks since it's a bi-phasal RPC protocol rather than a bi-direction streaming protocol. Data transmission and data reception do not take place simultaneously on any particular call. and making the following functional changes: (1) In rxrpc_input_data(), hold call->input_lock around the core to prevent simultaneous producing of packets into the Rx ring and updating of tracking state for a particular call. (2) In rxrpc_input_ping_response(), only read call->ping_serial once, and check it before checking RXRPC_CALL_PINGING as that's a cheaper test. The bit test and bit clear can then be combined. No further locking is needed here. (3) In rxrpc_input_ack(), take call->input_lock after we've parsed much of the ACK packet. The superseded ACK check is then done both before and after the lock is taken. The handing of ackinfo data is split, parsing before the lock is taken and processing with it held. This is keyed on rxMTU being non-zero. Congestion management is also done within the locked section. (4) In rxrpc_input_ackall(), take call->input_lock around the Tx window rotation. The ACKALL packet carries no information and is only really useful after all packets have been transmitted since it's imprecise. (5) In rxrpc_input_implicit_end_call(), we use rx->incoming_lock to prevent calls being simultaneously implicitly ended on two cpus and also to prevent any races with incoming call setup. (6) In rxrpc_input_packet(), use cmpxchg() to effect the service upgrade on a connection. It is only permitted to happen once for a connection. (7) In rxrpc_new_incoming_call(), we have to recheck the routing inside rx->incoming_lock to see if someone else set up the call, connection or peer whilst we were getting there. We can't trust the values from the earlier routing check unless we pin refs on them - which we want to avoid. Further, we need to allow for an incoming call to have its state changed on another CPU between us making it live and us adjusting it because the conn is now in the RXRPC_CONN_SERVICE state. (8) In rxrpc_peer_add_rtt(), take peer->rtt_input_lock around the access to the RTT buffer. Don't need to lock around setting peer->rtt. For reference, the inventory of state-accessing or state-altering functions used by the packet input procedure is: > rxrpc_input_packet() * PACKET CHECKING * ROUTING > rxrpc_post_packet_to_local() > rxrpc_find_connection_rcu() - uses RCU > rxrpc_lookup_peer_rcu() - uses RCU > rxrpc_find_service_conn_rcu() - uses RCU > idr_find() - uses RCU * CONNECTION-LEVEL PROCESSING - Service upgrade - Can only happen once per conn ! Changed to use cmpxchg > rxrpc_post_packet_to_conn() - Setting conn->hi_serial - Probably safe not using locks - Maybe use cmpxchg * CALL-LEVEL PROCESSING > Old-call checking > rxrpc_input_implicit_end_call() > rxrpc_call_completed() > rxrpc_queue_call() ! Need to take rx->incoming_lock > __rxrpc_disconnect_call() > rxrpc_notify_socket() > rxrpc_new_incoming_call() - Uses rx->incoming_lock for the entire process - Might be able to drop this earlier in favour of the call lock > rxrpc_incoming_call() ! Conflicts with rxrpc_input_implicit_end_call() > rxrpc_send_ping() - Don't need locks to check rtt state > rxrpc_propose_ACK * PACKET DISTRIBUTION > rxrpc_input_call_packet() > rxrpc_input_data() * QUEUE DATA PACKET ON CALL > rxrpc_reduce_call_timer() - Uses timer_reduce() ! Needs call->input_lock() > rxrpc_receiving_reply() ! Needs locking around ack state > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_proto_abort() > rxrpc_input_dup_data() - Fills the Rx buffer - rxrpc_propose_ACK() - rxrpc_notify_socket() > rxrpc_input_ack() * APPLY ACK PACKET TO CALL AND DISCARD PACKET > rxrpc_input_ping_response() - Probably doesn't need any extra locking ! Need READ_ONCE() on call->ping_serial > rxrpc_input_check_for_lost_ack() - Takes call->lock to consult Tx buffer > rxrpc_peer_add_rtt() ! Needs to take a lock (peer->rtt_input_lock) ! Could perhaps manage with cmpxchg() and xadd() instead > rxrpc_input_requested_ack - Consults Tx buffer ! Probably needs a lock > rxrpc_peer_add_rtt() > rxrpc_propose_ack() > rxrpc_input_ackinfo() - Changes call->tx_winsize ! Use cmpxchg to handle change ! Should perhaps track serial number - Uses peer->lock to record MTU specification changes > rxrpc_proto_abort() ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_input_soft_acks() - Consults the Tx buffer > rxrpc_congestion_management() - Modifies the Tx annotations ! Needs call->input_lock() > rxrpc_queue_call() > rxrpc_input_abort() * APPLY ABORT PACKET TO CALL AND DISCARD PACKET > rxrpc_set_call_completion() > rxrpc_notify_socket() > rxrpc_input_ackall() * APPLY ACKALL PACKET TO CALL AND DISCARD PACKET ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_reject_packet() There are some functions used by the above that queue the packet, after which the procedure is terminated: - rxrpc_post_packet_to_local() - local->event_queue is an sk_buff_head - local->processor is a work_struct - rxrpc_post_packet_to_conn() - conn->rx_queue is an sk_buff_head - conn->processor is a work_struct - rxrpc_reject_packet() - local->reject_queue is an sk_buff_head - local->processor is a work_struct And some that offload processing to process context: - rxrpc_notify_socket() - Uses RCU lock - Uses call->notify_lock to call call->notify_rx - Uses call->recvmsg_lock to queue recvmsg side - rxrpc_queue_call() - call->processor is a work_struct - rxrpc_propose_ACK() - Uses call->lock to wrap __rxrpc_propose_ACK() And a bunch that complete a call, all of which use call->state_lock to protect the call state: - rxrpc_call_completed() - rxrpc_set_call_completion() - rxrpc_abort_call() - rxrpc_proto_abort() - Also uses rxrpc_queue_call() Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both") Signed-off-by: David Howells <dhowells@redhat.com>
2018-10-08 22:46:25 +08:00
goto out;
call->acks_latest_ts = skb->tstamp;
call->acks_latest = sp->hdr.serial;
call->ackr_first_seq = first_soft_ack;
call->ackr_prev_seq = prev_pkt;
/* Parse rwind and mtu sizes if provided. */
rxrpc: Fix the packet reception routine The rxrpc_input_packet() function and its call tree was built around the assumption that data_ready() handler called from UDP to inform a kernel service that there is data to be had was non-reentrant. This means that certain locking could be dispensed with. This, however, turns out not to be the case with a multi-queue network card that can deliver packets to multiple cpus simultaneously. Each of those cpus can be in the rxrpc_input_packet() function at the same time. Fix by adding or changing some structure members: (1) Add peer->rtt_input_lock to serialise access to the RTT buffer. (2) Make conn->service_id into a 32-bit variable so that it can be cmpxchg'd on all arches. (3) Add call->input_lock to serialise access to the Rx/Tx state. Note that although the Rx and Tx states are (almost) entirely separate, there's no point completing the separation and having separate locks since it's a bi-phasal RPC protocol rather than a bi-direction streaming protocol. Data transmission and data reception do not take place simultaneously on any particular call. and making the following functional changes: (1) In rxrpc_input_data(), hold call->input_lock around the core to prevent simultaneous producing of packets into the Rx ring and updating of tracking state for a particular call. (2) In rxrpc_input_ping_response(), only read call->ping_serial once, and check it before checking RXRPC_CALL_PINGING as that's a cheaper test. The bit test and bit clear can then be combined. No further locking is needed here. (3) In rxrpc_input_ack(), take call->input_lock after we've parsed much of the ACK packet. The superseded ACK check is then done both before and after the lock is taken. The handing of ackinfo data is split, parsing before the lock is taken and processing with it held. This is keyed on rxMTU being non-zero. Congestion management is also done within the locked section. (4) In rxrpc_input_ackall(), take call->input_lock around the Tx window rotation. The ACKALL packet carries no information and is only really useful after all packets have been transmitted since it's imprecise. (5) In rxrpc_input_implicit_end_call(), we use rx->incoming_lock to prevent calls being simultaneously implicitly ended on two cpus and also to prevent any races with incoming call setup. (6) In rxrpc_input_packet(), use cmpxchg() to effect the service upgrade on a connection. It is only permitted to happen once for a connection. (7) In rxrpc_new_incoming_call(), we have to recheck the routing inside rx->incoming_lock to see if someone else set up the call, connection or peer whilst we were getting there. We can't trust the values from the earlier routing check unless we pin refs on them - which we want to avoid. Further, we need to allow for an incoming call to have its state changed on another CPU between us making it live and us adjusting it because the conn is now in the RXRPC_CONN_SERVICE state. (8) In rxrpc_peer_add_rtt(), take peer->rtt_input_lock around the access to the RTT buffer. Don't need to lock around setting peer->rtt. For reference, the inventory of state-accessing or state-altering functions used by the packet input procedure is: > rxrpc_input_packet() * PACKET CHECKING * ROUTING > rxrpc_post_packet_to_local() > rxrpc_find_connection_rcu() - uses RCU > rxrpc_lookup_peer_rcu() - uses RCU > rxrpc_find_service_conn_rcu() - uses RCU > idr_find() - uses RCU * CONNECTION-LEVEL PROCESSING - Service upgrade - Can only happen once per conn ! Changed to use cmpxchg > rxrpc_post_packet_to_conn() - Setting conn->hi_serial - Probably safe not using locks - Maybe use cmpxchg * CALL-LEVEL PROCESSING > Old-call checking > rxrpc_input_implicit_end_call() > rxrpc_call_completed() > rxrpc_queue_call() ! Need to take rx->incoming_lock > __rxrpc_disconnect_call() > rxrpc_notify_socket() > rxrpc_new_incoming_call() - Uses rx->incoming_lock for the entire process - Might be able to drop this earlier in favour of the call lock > rxrpc_incoming_call() ! Conflicts with rxrpc_input_implicit_end_call() > rxrpc_send_ping() - Don't need locks to check rtt state > rxrpc_propose_ACK * PACKET DISTRIBUTION > rxrpc_input_call_packet() > rxrpc_input_data() * QUEUE DATA PACKET ON CALL > rxrpc_reduce_call_timer() - Uses timer_reduce() ! Needs call->input_lock() > rxrpc_receiving_reply() ! Needs locking around ack state > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_proto_abort() > rxrpc_input_dup_data() - Fills the Rx buffer - rxrpc_propose_ACK() - rxrpc_notify_socket() > rxrpc_input_ack() * APPLY ACK PACKET TO CALL AND DISCARD PACKET > rxrpc_input_ping_response() - Probably doesn't need any extra locking ! Need READ_ONCE() on call->ping_serial > rxrpc_input_check_for_lost_ack() - Takes call->lock to consult Tx buffer > rxrpc_peer_add_rtt() ! Needs to take a lock (peer->rtt_input_lock) ! Could perhaps manage with cmpxchg() and xadd() instead > rxrpc_input_requested_ack - Consults Tx buffer ! Probably needs a lock > rxrpc_peer_add_rtt() > rxrpc_propose_ack() > rxrpc_input_ackinfo() - Changes call->tx_winsize ! Use cmpxchg to handle change ! Should perhaps track serial number - Uses peer->lock to record MTU specification changes > rxrpc_proto_abort() ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_input_soft_acks() - Consults the Tx buffer > rxrpc_congestion_management() - Modifies the Tx annotations ! Needs call->input_lock() > rxrpc_queue_call() > rxrpc_input_abort() * APPLY ABORT PACKET TO CALL AND DISCARD PACKET > rxrpc_set_call_completion() > rxrpc_notify_socket() > rxrpc_input_ackall() * APPLY ACKALL PACKET TO CALL AND DISCARD PACKET ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_reject_packet() There are some functions used by the above that queue the packet, after which the procedure is terminated: - rxrpc_post_packet_to_local() - local->event_queue is an sk_buff_head - local->processor is a work_struct - rxrpc_post_packet_to_conn() - conn->rx_queue is an sk_buff_head - conn->processor is a work_struct - rxrpc_reject_packet() - local->reject_queue is an sk_buff_head - local->processor is a work_struct And some that offload processing to process context: - rxrpc_notify_socket() - Uses RCU lock - Uses call->notify_lock to call call->notify_rx - Uses call->recvmsg_lock to queue recvmsg side - rxrpc_queue_call() - call->processor is a work_struct - rxrpc_propose_ACK() - Uses call->lock to wrap __rxrpc_propose_ACK() And a bunch that complete a call, all of which use call->state_lock to protect the call state: - rxrpc_call_completed() - rxrpc_set_call_completion() - rxrpc_abort_call() - rxrpc_proto_abort() - Also uses rxrpc_queue_call() Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both") Signed-off-by: David Howells <dhowells@redhat.com>
2018-10-08 22:46:25 +08:00
if (buf.info.rxMTU)
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
rxrpc_input_ackinfo(call, skb, &buf.info);
rxrpc: Fix the packet reception routine The rxrpc_input_packet() function and its call tree was built around the assumption that data_ready() handler called from UDP to inform a kernel service that there is data to be had was non-reentrant. This means that certain locking could be dispensed with. This, however, turns out not to be the case with a multi-queue network card that can deliver packets to multiple cpus simultaneously. Each of those cpus can be in the rxrpc_input_packet() function at the same time. Fix by adding or changing some structure members: (1) Add peer->rtt_input_lock to serialise access to the RTT buffer. (2) Make conn->service_id into a 32-bit variable so that it can be cmpxchg'd on all arches. (3) Add call->input_lock to serialise access to the Rx/Tx state. Note that although the Rx and Tx states are (almost) entirely separate, there's no point completing the separation and having separate locks since it's a bi-phasal RPC protocol rather than a bi-direction streaming protocol. Data transmission and data reception do not take place simultaneously on any particular call. and making the following functional changes: (1) In rxrpc_input_data(), hold call->input_lock around the core to prevent simultaneous producing of packets into the Rx ring and updating of tracking state for a particular call. (2) In rxrpc_input_ping_response(), only read call->ping_serial once, and check it before checking RXRPC_CALL_PINGING as that's a cheaper test. The bit test and bit clear can then be combined. No further locking is needed here. (3) In rxrpc_input_ack(), take call->input_lock after we've parsed much of the ACK packet. The superseded ACK check is then done both before and after the lock is taken. The handing of ackinfo data is split, parsing before the lock is taken and processing with it held. This is keyed on rxMTU being non-zero. Congestion management is also done within the locked section. (4) In rxrpc_input_ackall(), take call->input_lock around the Tx window rotation. The ACKALL packet carries no information and is only really useful after all packets have been transmitted since it's imprecise. (5) In rxrpc_input_implicit_end_call(), we use rx->incoming_lock to prevent calls being simultaneously implicitly ended on two cpus and also to prevent any races with incoming call setup. (6) In rxrpc_input_packet(), use cmpxchg() to effect the service upgrade on a connection. It is only permitted to happen once for a connection. (7) In rxrpc_new_incoming_call(), we have to recheck the routing inside rx->incoming_lock to see if someone else set up the call, connection or peer whilst we were getting there. We can't trust the values from the earlier routing check unless we pin refs on them - which we want to avoid. Further, we need to allow for an incoming call to have its state changed on another CPU between us making it live and us adjusting it because the conn is now in the RXRPC_CONN_SERVICE state. (8) In rxrpc_peer_add_rtt(), take peer->rtt_input_lock around the access to the RTT buffer. Don't need to lock around setting peer->rtt. For reference, the inventory of state-accessing or state-altering functions used by the packet input procedure is: > rxrpc_input_packet() * PACKET CHECKING * ROUTING > rxrpc_post_packet_to_local() > rxrpc_find_connection_rcu() - uses RCU > rxrpc_lookup_peer_rcu() - uses RCU > rxrpc_find_service_conn_rcu() - uses RCU > idr_find() - uses RCU * CONNECTION-LEVEL PROCESSING - Service upgrade - Can only happen once per conn ! Changed to use cmpxchg > rxrpc_post_packet_to_conn() - Setting conn->hi_serial - Probably safe not using locks - Maybe use cmpxchg * CALL-LEVEL PROCESSING > Old-call checking > rxrpc_input_implicit_end_call() > rxrpc_call_completed() > rxrpc_queue_call() ! Need to take rx->incoming_lock > __rxrpc_disconnect_call() > rxrpc_notify_socket() > rxrpc_new_incoming_call() - Uses rx->incoming_lock for the entire process - Might be able to drop this earlier in favour of the call lock > rxrpc_incoming_call() ! Conflicts with rxrpc_input_implicit_end_call() > rxrpc_send_ping() - Don't need locks to check rtt state > rxrpc_propose_ACK * PACKET DISTRIBUTION > rxrpc_input_call_packet() > rxrpc_input_data() * QUEUE DATA PACKET ON CALL > rxrpc_reduce_call_timer() - Uses timer_reduce() ! Needs call->input_lock() > rxrpc_receiving_reply() ! Needs locking around ack state > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_proto_abort() > rxrpc_input_dup_data() - Fills the Rx buffer - rxrpc_propose_ACK() - rxrpc_notify_socket() > rxrpc_input_ack() * APPLY ACK PACKET TO CALL AND DISCARD PACKET > rxrpc_input_ping_response() - Probably doesn't need any extra locking ! Need READ_ONCE() on call->ping_serial > rxrpc_input_check_for_lost_ack() - Takes call->lock to consult Tx buffer > rxrpc_peer_add_rtt() ! Needs to take a lock (peer->rtt_input_lock) ! Could perhaps manage with cmpxchg() and xadd() instead > rxrpc_input_requested_ack - Consults Tx buffer ! Probably needs a lock > rxrpc_peer_add_rtt() > rxrpc_propose_ack() > rxrpc_input_ackinfo() - Changes call->tx_winsize ! Use cmpxchg to handle change ! Should perhaps track serial number - Uses peer->lock to record MTU specification changes > rxrpc_proto_abort() ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_input_soft_acks() - Consults the Tx buffer > rxrpc_congestion_management() - Modifies the Tx annotations ! Needs call->input_lock() > rxrpc_queue_call() > rxrpc_input_abort() * APPLY ABORT PACKET TO CALL AND DISCARD PACKET > rxrpc_set_call_completion() > rxrpc_notify_socket() > rxrpc_input_ackall() * APPLY ACKALL PACKET TO CALL AND DISCARD PACKET ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_reject_packet() There are some functions used by the above that queue the packet, after which the procedure is terminated: - rxrpc_post_packet_to_local() - local->event_queue is an sk_buff_head - local->processor is a work_struct - rxrpc_post_packet_to_conn() - conn->rx_queue is an sk_buff_head - conn->processor is a work_struct - rxrpc_reject_packet() - local->reject_queue is an sk_buff_head - local->processor is a work_struct And some that offload processing to process context: - rxrpc_notify_socket() - Uses RCU lock - Uses call->notify_lock to call call->notify_rx - Uses call->recvmsg_lock to queue recvmsg side - rxrpc_queue_call() - call->processor is a work_struct - rxrpc_propose_ACK() - Uses call->lock to wrap __rxrpc_propose_ACK() And a bunch that complete a call, all of which use call->state_lock to protect the call state: - rxrpc_call_completed() - rxrpc_set_call_completion() - rxrpc_abort_call() - rxrpc_proto_abort() - Also uses rxrpc_queue_call() Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both") Signed-off-by: David Howells <dhowells@redhat.com>
2018-10-08 22:46:25 +08:00
if (first_soft_ack == 0) {
rxrpc_proto_abort("AK0", call, 0);
goto out;
}
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
/* Ignore ACKs unless we are or have just been transmitting. */
switch (READ_ONCE(call->state)) {
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
case RXRPC_CALL_CLIENT_SEND_REQUEST:
case RXRPC_CALL_CLIENT_AWAIT_REPLY:
case RXRPC_CALL_SERVER_SEND_REPLY:
case RXRPC_CALL_SERVER_AWAIT_ACK:
break;
default:
rxrpc: Fix the packet reception routine The rxrpc_input_packet() function and its call tree was built around the assumption that data_ready() handler called from UDP to inform a kernel service that there is data to be had was non-reentrant. This means that certain locking could be dispensed with. This, however, turns out not to be the case with a multi-queue network card that can deliver packets to multiple cpus simultaneously. Each of those cpus can be in the rxrpc_input_packet() function at the same time. Fix by adding or changing some structure members: (1) Add peer->rtt_input_lock to serialise access to the RTT buffer. (2) Make conn->service_id into a 32-bit variable so that it can be cmpxchg'd on all arches. (3) Add call->input_lock to serialise access to the Rx/Tx state. Note that although the Rx and Tx states are (almost) entirely separate, there's no point completing the separation and having separate locks since it's a bi-phasal RPC protocol rather than a bi-direction streaming protocol. Data transmission and data reception do not take place simultaneously on any particular call. and making the following functional changes: (1) In rxrpc_input_data(), hold call->input_lock around the core to prevent simultaneous producing of packets into the Rx ring and updating of tracking state for a particular call. (2) In rxrpc_input_ping_response(), only read call->ping_serial once, and check it before checking RXRPC_CALL_PINGING as that's a cheaper test. The bit test and bit clear can then be combined. No further locking is needed here. (3) In rxrpc_input_ack(), take call->input_lock after we've parsed much of the ACK packet. The superseded ACK check is then done both before and after the lock is taken. The handing of ackinfo data is split, parsing before the lock is taken and processing with it held. This is keyed on rxMTU being non-zero. Congestion management is also done within the locked section. (4) In rxrpc_input_ackall(), take call->input_lock around the Tx window rotation. The ACKALL packet carries no information and is only really useful after all packets have been transmitted since it's imprecise. (5) In rxrpc_input_implicit_end_call(), we use rx->incoming_lock to prevent calls being simultaneously implicitly ended on two cpus and also to prevent any races with incoming call setup. (6) In rxrpc_input_packet(), use cmpxchg() to effect the service upgrade on a connection. It is only permitted to happen once for a connection. (7) In rxrpc_new_incoming_call(), we have to recheck the routing inside rx->incoming_lock to see if someone else set up the call, connection or peer whilst we were getting there. We can't trust the values from the earlier routing check unless we pin refs on them - which we want to avoid. Further, we need to allow for an incoming call to have its state changed on another CPU between us making it live and us adjusting it because the conn is now in the RXRPC_CONN_SERVICE state. (8) In rxrpc_peer_add_rtt(), take peer->rtt_input_lock around the access to the RTT buffer. Don't need to lock around setting peer->rtt. For reference, the inventory of state-accessing or state-altering functions used by the packet input procedure is: > rxrpc_input_packet() * PACKET CHECKING * ROUTING > rxrpc_post_packet_to_local() > rxrpc_find_connection_rcu() - uses RCU > rxrpc_lookup_peer_rcu() - uses RCU > rxrpc_find_service_conn_rcu() - uses RCU > idr_find() - uses RCU * CONNECTION-LEVEL PROCESSING - Service upgrade - Can only happen once per conn ! Changed to use cmpxchg > rxrpc_post_packet_to_conn() - Setting conn->hi_serial - Probably safe not using locks - Maybe use cmpxchg * CALL-LEVEL PROCESSING > Old-call checking > rxrpc_input_implicit_end_call() > rxrpc_call_completed() > rxrpc_queue_call() ! Need to take rx->incoming_lock > __rxrpc_disconnect_call() > rxrpc_notify_socket() > rxrpc_new_incoming_call() - Uses rx->incoming_lock for the entire process - Might be able to drop this earlier in favour of the call lock > rxrpc_incoming_call() ! Conflicts with rxrpc_input_implicit_end_call() > rxrpc_send_ping() - Don't need locks to check rtt state > rxrpc_propose_ACK * PACKET DISTRIBUTION > rxrpc_input_call_packet() > rxrpc_input_data() * QUEUE DATA PACKET ON CALL > rxrpc_reduce_call_timer() - Uses timer_reduce() ! Needs call->input_lock() > rxrpc_receiving_reply() ! Needs locking around ack state > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_proto_abort() > rxrpc_input_dup_data() - Fills the Rx buffer - rxrpc_propose_ACK() - rxrpc_notify_socket() > rxrpc_input_ack() * APPLY ACK PACKET TO CALL AND DISCARD PACKET > rxrpc_input_ping_response() - Probably doesn't need any extra locking ! Need READ_ONCE() on call->ping_serial > rxrpc_input_check_for_lost_ack() - Takes call->lock to consult Tx buffer > rxrpc_peer_add_rtt() ! Needs to take a lock (peer->rtt_input_lock) ! Could perhaps manage with cmpxchg() and xadd() instead > rxrpc_input_requested_ack - Consults Tx buffer ! Probably needs a lock > rxrpc_peer_add_rtt() > rxrpc_propose_ack() > rxrpc_input_ackinfo() - Changes call->tx_winsize ! Use cmpxchg to handle change ! Should perhaps track serial number - Uses peer->lock to record MTU specification changes > rxrpc_proto_abort() ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_input_soft_acks() - Consults the Tx buffer > rxrpc_congestion_management() - Modifies the Tx annotations ! Needs call->input_lock() > rxrpc_queue_call() > rxrpc_input_abort() * APPLY ABORT PACKET TO CALL AND DISCARD PACKET > rxrpc_set_call_completion() > rxrpc_notify_socket() > rxrpc_input_ackall() * APPLY ACKALL PACKET TO CALL AND DISCARD PACKET ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_reject_packet() There are some functions used by the above that queue the packet, after which the procedure is terminated: - rxrpc_post_packet_to_local() - local->event_queue is an sk_buff_head - local->processor is a work_struct - rxrpc_post_packet_to_conn() - conn->rx_queue is an sk_buff_head - conn->processor is a work_struct - rxrpc_reject_packet() - local->reject_queue is an sk_buff_head - local->processor is a work_struct And some that offload processing to process context: - rxrpc_notify_socket() - Uses RCU lock - Uses call->notify_lock to call call->notify_rx - Uses call->recvmsg_lock to queue recvmsg side - rxrpc_queue_call() - call->processor is a work_struct - rxrpc_propose_ACK() - Uses call->lock to wrap __rxrpc_propose_ACK() And a bunch that complete a call, all of which use call->state_lock to protect the call state: - rxrpc_call_completed() - rxrpc_set_call_completion() - rxrpc_abort_call() - rxrpc_proto_abort() - Also uses rxrpc_queue_call() Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both") Signed-off-by: David Howells <dhowells@redhat.com>
2018-10-08 22:46:25 +08:00
goto out;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
}
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
if (before(hard_ack, call->tx_hard_ack) ||
rxrpc: Fix the packet reception routine The rxrpc_input_packet() function and its call tree was built around the assumption that data_ready() handler called from UDP to inform a kernel service that there is data to be had was non-reentrant. This means that certain locking could be dispensed with. This, however, turns out not to be the case with a multi-queue network card that can deliver packets to multiple cpus simultaneously. Each of those cpus can be in the rxrpc_input_packet() function at the same time. Fix by adding or changing some structure members: (1) Add peer->rtt_input_lock to serialise access to the RTT buffer. (2) Make conn->service_id into a 32-bit variable so that it can be cmpxchg'd on all arches. (3) Add call->input_lock to serialise access to the Rx/Tx state. Note that although the Rx and Tx states are (almost) entirely separate, there's no point completing the separation and having separate locks since it's a bi-phasal RPC protocol rather than a bi-direction streaming protocol. Data transmission and data reception do not take place simultaneously on any particular call. and making the following functional changes: (1) In rxrpc_input_data(), hold call->input_lock around the core to prevent simultaneous producing of packets into the Rx ring and updating of tracking state for a particular call. (2) In rxrpc_input_ping_response(), only read call->ping_serial once, and check it before checking RXRPC_CALL_PINGING as that's a cheaper test. The bit test and bit clear can then be combined. No further locking is needed here. (3) In rxrpc_input_ack(), take call->input_lock after we've parsed much of the ACK packet. The superseded ACK check is then done both before and after the lock is taken. The handing of ackinfo data is split, parsing before the lock is taken and processing with it held. This is keyed on rxMTU being non-zero. Congestion management is also done within the locked section. (4) In rxrpc_input_ackall(), take call->input_lock around the Tx window rotation. The ACKALL packet carries no information and is only really useful after all packets have been transmitted since it's imprecise. (5) In rxrpc_input_implicit_end_call(), we use rx->incoming_lock to prevent calls being simultaneously implicitly ended on two cpus and also to prevent any races with incoming call setup. (6) In rxrpc_input_packet(), use cmpxchg() to effect the service upgrade on a connection. It is only permitted to happen once for a connection. (7) In rxrpc_new_incoming_call(), we have to recheck the routing inside rx->incoming_lock to see if someone else set up the call, connection or peer whilst we were getting there. We can't trust the values from the earlier routing check unless we pin refs on them - which we want to avoid. Further, we need to allow for an incoming call to have its state changed on another CPU between us making it live and us adjusting it because the conn is now in the RXRPC_CONN_SERVICE state. (8) In rxrpc_peer_add_rtt(), take peer->rtt_input_lock around the access to the RTT buffer. Don't need to lock around setting peer->rtt. For reference, the inventory of state-accessing or state-altering functions used by the packet input procedure is: > rxrpc_input_packet() * PACKET CHECKING * ROUTING > rxrpc_post_packet_to_local() > rxrpc_find_connection_rcu() - uses RCU > rxrpc_lookup_peer_rcu() - uses RCU > rxrpc_find_service_conn_rcu() - uses RCU > idr_find() - uses RCU * CONNECTION-LEVEL PROCESSING - Service upgrade - Can only happen once per conn ! Changed to use cmpxchg > rxrpc_post_packet_to_conn() - Setting conn->hi_serial - Probably safe not using locks - Maybe use cmpxchg * CALL-LEVEL PROCESSING > Old-call checking > rxrpc_input_implicit_end_call() > rxrpc_call_completed() > rxrpc_queue_call() ! Need to take rx->incoming_lock > __rxrpc_disconnect_call() > rxrpc_notify_socket() > rxrpc_new_incoming_call() - Uses rx->incoming_lock for the entire process - Might be able to drop this earlier in favour of the call lock > rxrpc_incoming_call() ! Conflicts with rxrpc_input_implicit_end_call() > rxrpc_send_ping() - Don't need locks to check rtt state > rxrpc_propose_ACK * PACKET DISTRIBUTION > rxrpc_input_call_packet() > rxrpc_input_data() * QUEUE DATA PACKET ON CALL > rxrpc_reduce_call_timer() - Uses timer_reduce() ! Needs call->input_lock() > rxrpc_receiving_reply() ! Needs locking around ack state > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_proto_abort() > rxrpc_input_dup_data() - Fills the Rx buffer - rxrpc_propose_ACK() - rxrpc_notify_socket() > rxrpc_input_ack() * APPLY ACK PACKET TO CALL AND DISCARD PACKET > rxrpc_input_ping_response() - Probably doesn't need any extra locking ! Need READ_ONCE() on call->ping_serial > rxrpc_input_check_for_lost_ack() - Takes call->lock to consult Tx buffer > rxrpc_peer_add_rtt() ! Needs to take a lock (peer->rtt_input_lock) ! Could perhaps manage with cmpxchg() and xadd() instead > rxrpc_input_requested_ack - Consults Tx buffer ! Probably needs a lock > rxrpc_peer_add_rtt() > rxrpc_propose_ack() > rxrpc_input_ackinfo() - Changes call->tx_winsize ! Use cmpxchg to handle change ! Should perhaps track serial number - Uses peer->lock to record MTU specification changes > rxrpc_proto_abort() ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_input_soft_acks() - Consults the Tx buffer > rxrpc_congestion_management() - Modifies the Tx annotations ! Needs call->input_lock() > rxrpc_queue_call() > rxrpc_input_abort() * APPLY ABORT PACKET TO CALL AND DISCARD PACKET > rxrpc_set_call_completion() > rxrpc_notify_socket() > rxrpc_input_ackall() * APPLY ACKALL PACKET TO CALL AND DISCARD PACKET ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_reject_packet() There are some functions used by the above that queue the packet, after which the procedure is terminated: - rxrpc_post_packet_to_local() - local->event_queue is an sk_buff_head - local->processor is a work_struct - rxrpc_post_packet_to_conn() - conn->rx_queue is an sk_buff_head - conn->processor is a work_struct - rxrpc_reject_packet() - local->reject_queue is an sk_buff_head - local->processor is a work_struct And some that offload processing to process context: - rxrpc_notify_socket() - Uses RCU lock - Uses call->notify_lock to call call->notify_rx - Uses call->recvmsg_lock to queue recvmsg side - rxrpc_queue_call() - call->processor is a work_struct - rxrpc_propose_ACK() - Uses call->lock to wrap __rxrpc_propose_ACK() And a bunch that complete a call, all of which use call->state_lock to protect the call state: - rxrpc_call_completed() - rxrpc_set_call_completion() - rxrpc_abort_call() - rxrpc_proto_abort() - Also uses rxrpc_queue_call() Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both") Signed-off-by: David Howells <dhowells@redhat.com>
2018-10-08 22:46:25 +08:00
after(hard_ack, call->tx_top)) {
rxrpc_proto_abort("AKW", call, 0);
goto out;
}
if (nr_acks > call->tx_top - hard_ack) {
rxrpc_proto_abort("AKN", call, 0);
goto out;
}
if (after(hard_ack, call->tx_hard_ack)) {
if (rxrpc_rotate_tx_window(call, hard_ack, &summary)) {
rxrpc_end_tx_phase(call, false, "ETA");
rxrpc: Fix the packet reception routine The rxrpc_input_packet() function and its call tree was built around the assumption that data_ready() handler called from UDP to inform a kernel service that there is data to be had was non-reentrant. This means that certain locking could be dispensed with. This, however, turns out not to be the case with a multi-queue network card that can deliver packets to multiple cpus simultaneously. Each of those cpus can be in the rxrpc_input_packet() function at the same time. Fix by adding or changing some structure members: (1) Add peer->rtt_input_lock to serialise access to the RTT buffer. (2) Make conn->service_id into a 32-bit variable so that it can be cmpxchg'd on all arches. (3) Add call->input_lock to serialise access to the Rx/Tx state. Note that although the Rx and Tx states are (almost) entirely separate, there's no point completing the separation and having separate locks since it's a bi-phasal RPC protocol rather than a bi-direction streaming protocol. Data transmission and data reception do not take place simultaneously on any particular call. and making the following functional changes: (1) In rxrpc_input_data(), hold call->input_lock around the core to prevent simultaneous producing of packets into the Rx ring and updating of tracking state for a particular call. (2) In rxrpc_input_ping_response(), only read call->ping_serial once, and check it before checking RXRPC_CALL_PINGING as that's a cheaper test. The bit test and bit clear can then be combined. No further locking is needed here. (3) In rxrpc_input_ack(), take call->input_lock after we've parsed much of the ACK packet. The superseded ACK check is then done both before and after the lock is taken. The handing of ackinfo data is split, parsing before the lock is taken and processing with it held. This is keyed on rxMTU being non-zero. Congestion management is also done within the locked section. (4) In rxrpc_input_ackall(), take call->input_lock around the Tx window rotation. The ACKALL packet carries no information and is only really useful after all packets have been transmitted since it's imprecise. (5) In rxrpc_input_implicit_end_call(), we use rx->incoming_lock to prevent calls being simultaneously implicitly ended on two cpus and also to prevent any races with incoming call setup. (6) In rxrpc_input_packet(), use cmpxchg() to effect the service upgrade on a connection. It is only permitted to happen once for a connection. (7) In rxrpc_new_incoming_call(), we have to recheck the routing inside rx->incoming_lock to see if someone else set up the call, connection or peer whilst we were getting there. We can't trust the values from the earlier routing check unless we pin refs on them - which we want to avoid. Further, we need to allow for an incoming call to have its state changed on another CPU between us making it live and us adjusting it because the conn is now in the RXRPC_CONN_SERVICE state. (8) In rxrpc_peer_add_rtt(), take peer->rtt_input_lock around the access to the RTT buffer. Don't need to lock around setting peer->rtt. For reference, the inventory of state-accessing or state-altering functions used by the packet input procedure is: > rxrpc_input_packet() * PACKET CHECKING * ROUTING > rxrpc_post_packet_to_local() > rxrpc_find_connection_rcu() - uses RCU > rxrpc_lookup_peer_rcu() - uses RCU > rxrpc_find_service_conn_rcu() - uses RCU > idr_find() - uses RCU * CONNECTION-LEVEL PROCESSING - Service upgrade - Can only happen once per conn ! Changed to use cmpxchg > rxrpc_post_packet_to_conn() - Setting conn->hi_serial - Probably safe not using locks - Maybe use cmpxchg * CALL-LEVEL PROCESSING > Old-call checking > rxrpc_input_implicit_end_call() > rxrpc_call_completed() > rxrpc_queue_call() ! Need to take rx->incoming_lock > __rxrpc_disconnect_call() > rxrpc_notify_socket() > rxrpc_new_incoming_call() - Uses rx->incoming_lock for the entire process - Might be able to drop this earlier in favour of the call lock > rxrpc_incoming_call() ! Conflicts with rxrpc_input_implicit_end_call() > rxrpc_send_ping() - Don't need locks to check rtt state > rxrpc_propose_ACK * PACKET DISTRIBUTION > rxrpc_input_call_packet() > rxrpc_input_data() * QUEUE DATA PACKET ON CALL > rxrpc_reduce_call_timer() - Uses timer_reduce() ! Needs call->input_lock() > rxrpc_receiving_reply() ! Needs locking around ack state > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_proto_abort() > rxrpc_input_dup_data() - Fills the Rx buffer - rxrpc_propose_ACK() - rxrpc_notify_socket() > rxrpc_input_ack() * APPLY ACK PACKET TO CALL AND DISCARD PACKET > rxrpc_input_ping_response() - Probably doesn't need any extra locking ! Need READ_ONCE() on call->ping_serial > rxrpc_input_check_for_lost_ack() - Takes call->lock to consult Tx buffer > rxrpc_peer_add_rtt() ! Needs to take a lock (peer->rtt_input_lock) ! Could perhaps manage with cmpxchg() and xadd() instead > rxrpc_input_requested_ack - Consults Tx buffer ! Probably needs a lock > rxrpc_peer_add_rtt() > rxrpc_propose_ack() > rxrpc_input_ackinfo() - Changes call->tx_winsize ! Use cmpxchg to handle change ! Should perhaps track serial number - Uses peer->lock to record MTU specification changes > rxrpc_proto_abort() ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_input_soft_acks() - Consults the Tx buffer > rxrpc_congestion_management() - Modifies the Tx annotations ! Needs call->input_lock() > rxrpc_queue_call() > rxrpc_input_abort() * APPLY ABORT PACKET TO CALL AND DISCARD PACKET > rxrpc_set_call_completion() > rxrpc_notify_socket() > rxrpc_input_ackall() * APPLY ACKALL PACKET TO CALL AND DISCARD PACKET ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_reject_packet() There are some functions used by the above that queue the packet, after which the procedure is terminated: - rxrpc_post_packet_to_local() - local->event_queue is an sk_buff_head - local->processor is a work_struct - rxrpc_post_packet_to_conn() - conn->rx_queue is an sk_buff_head - conn->processor is a work_struct - rxrpc_reject_packet() - local->reject_queue is an sk_buff_head - local->processor is a work_struct And some that offload processing to process context: - rxrpc_notify_socket() - Uses RCU lock - Uses call->notify_lock to call call->notify_rx - Uses call->recvmsg_lock to queue recvmsg side - rxrpc_queue_call() - call->processor is a work_struct - rxrpc_propose_ACK() - Uses call->lock to wrap __rxrpc_propose_ACK() And a bunch that complete a call, all of which use call->state_lock to protect the call state: - rxrpc_call_completed() - rxrpc_set_call_completion() - rxrpc_abort_call() - rxrpc_proto_abort() - Also uses rxrpc_queue_call() Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both") Signed-off-by: David Howells <dhowells@redhat.com>
2018-10-08 22:46:25 +08:00
goto out;
}
}
rxrpc: Pass the last Tx packet marker in the annotation buffer When the last packet of data to be transmitted on a call is queued, tx_top is set and then the RXRPC_CALL_TX_LAST flag is set. Unfortunately, this leaves a race in the ACK processing side of things because the flag affects the interpretation of tx_top and also allows us to start receiving reply data before we've finished transmitting. To fix this, make the following changes: (1) rxrpc_queue_packet() now sets a marker in the annotation buffer instead of setting the RXRPC_CALL_TX_LAST flag. (2) rxrpc_rotate_tx_window() detects the marker and sets the flag in the same context as the routines that use it. (3) rxrpc_end_tx_phase() is simplified to just shift the call state. The Tx window must have been rotated before calling to discard the last packet. (4) rxrpc_receiving_reply() is added to handle the arrival of the first DATA packet of a reply to a client call (which is an implicit ACK of the Tx phase). (5) The last part of rxrpc_input_ack() is reordered to perform Tx rotation, then soft-ACK application and then to end the phase if we've rotated the last packet. In the event of a terminal ACK, the soft-ACK application will be skipped as nAcks should be 0. (6) rxrpc_input_ackall() now has to rotate as well as ending the phase. In addition: (7) Alter the transmit tracepoint to log the rotation of the last packet. (8) Remove the no-longer relevant queue_reqack tracepoint note. The ACK-REQUESTED packet header flag is now set as needed when we actually transmit the packet and may vary by retransmission. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-23 19:39:22 +08:00
if (nr_acks > 0) {
rxrpc: Fix the packet reception routine The rxrpc_input_packet() function and its call tree was built around the assumption that data_ready() handler called from UDP to inform a kernel service that there is data to be had was non-reentrant. This means that certain locking could be dispensed with. This, however, turns out not to be the case with a multi-queue network card that can deliver packets to multiple cpus simultaneously. Each of those cpus can be in the rxrpc_input_packet() function at the same time. Fix by adding or changing some structure members: (1) Add peer->rtt_input_lock to serialise access to the RTT buffer. (2) Make conn->service_id into a 32-bit variable so that it can be cmpxchg'd on all arches. (3) Add call->input_lock to serialise access to the Rx/Tx state. Note that although the Rx and Tx states are (almost) entirely separate, there's no point completing the separation and having separate locks since it's a bi-phasal RPC protocol rather than a bi-direction streaming protocol. Data transmission and data reception do not take place simultaneously on any particular call. and making the following functional changes: (1) In rxrpc_input_data(), hold call->input_lock around the core to prevent simultaneous producing of packets into the Rx ring and updating of tracking state for a particular call. (2) In rxrpc_input_ping_response(), only read call->ping_serial once, and check it before checking RXRPC_CALL_PINGING as that's a cheaper test. The bit test and bit clear can then be combined. No further locking is needed here. (3) In rxrpc_input_ack(), take call->input_lock after we've parsed much of the ACK packet. The superseded ACK check is then done both before and after the lock is taken. The handing of ackinfo data is split, parsing before the lock is taken and processing with it held. This is keyed on rxMTU being non-zero. Congestion management is also done within the locked section. (4) In rxrpc_input_ackall(), take call->input_lock around the Tx window rotation. The ACKALL packet carries no information and is only really useful after all packets have been transmitted since it's imprecise. (5) In rxrpc_input_implicit_end_call(), we use rx->incoming_lock to prevent calls being simultaneously implicitly ended on two cpus and also to prevent any races with incoming call setup. (6) In rxrpc_input_packet(), use cmpxchg() to effect the service upgrade on a connection. It is only permitted to happen once for a connection. (7) In rxrpc_new_incoming_call(), we have to recheck the routing inside rx->incoming_lock to see if someone else set up the call, connection or peer whilst we were getting there. We can't trust the values from the earlier routing check unless we pin refs on them - which we want to avoid. Further, we need to allow for an incoming call to have its state changed on another CPU between us making it live and us adjusting it because the conn is now in the RXRPC_CONN_SERVICE state. (8) In rxrpc_peer_add_rtt(), take peer->rtt_input_lock around the access to the RTT buffer. Don't need to lock around setting peer->rtt. For reference, the inventory of state-accessing or state-altering functions used by the packet input procedure is: > rxrpc_input_packet() * PACKET CHECKING * ROUTING > rxrpc_post_packet_to_local() > rxrpc_find_connection_rcu() - uses RCU > rxrpc_lookup_peer_rcu() - uses RCU > rxrpc_find_service_conn_rcu() - uses RCU > idr_find() - uses RCU * CONNECTION-LEVEL PROCESSING - Service upgrade - Can only happen once per conn ! Changed to use cmpxchg > rxrpc_post_packet_to_conn() - Setting conn->hi_serial - Probably safe not using locks - Maybe use cmpxchg * CALL-LEVEL PROCESSING > Old-call checking > rxrpc_input_implicit_end_call() > rxrpc_call_completed() > rxrpc_queue_call() ! Need to take rx->incoming_lock > __rxrpc_disconnect_call() > rxrpc_notify_socket() > rxrpc_new_incoming_call() - Uses rx->incoming_lock for the entire process - Might be able to drop this earlier in favour of the call lock > rxrpc_incoming_call() ! Conflicts with rxrpc_input_implicit_end_call() > rxrpc_send_ping() - Don't need locks to check rtt state > rxrpc_propose_ACK * PACKET DISTRIBUTION > rxrpc_input_call_packet() > rxrpc_input_data() * QUEUE DATA PACKET ON CALL > rxrpc_reduce_call_timer() - Uses timer_reduce() ! Needs call->input_lock() > rxrpc_receiving_reply() ! Needs locking around ack state > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_proto_abort() > rxrpc_input_dup_data() - Fills the Rx buffer - rxrpc_propose_ACK() - rxrpc_notify_socket() > rxrpc_input_ack() * APPLY ACK PACKET TO CALL AND DISCARD PACKET > rxrpc_input_ping_response() - Probably doesn't need any extra locking ! Need READ_ONCE() on call->ping_serial > rxrpc_input_check_for_lost_ack() - Takes call->lock to consult Tx buffer > rxrpc_peer_add_rtt() ! Needs to take a lock (peer->rtt_input_lock) ! Could perhaps manage with cmpxchg() and xadd() instead > rxrpc_input_requested_ack - Consults Tx buffer ! Probably needs a lock > rxrpc_peer_add_rtt() > rxrpc_propose_ack() > rxrpc_input_ackinfo() - Changes call->tx_winsize ! Use cmpxchg to handle change ! Should perhaps track serial number - Uses peer->lock to record MTU specification changes > rxrpc_proto_abort() ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_input_soft_acks() - Consults the Tx buffer > rxrpc_congestion_management() - Modifies the Tx annotations ! Needs call->input_lock() > rxrpc_queue_call() > rxrpc_input_abort() * APPLY ABORT PACKET TO CALL AND DISCARD PACKET > rxrpc_set_call_completion() > rxrpc_notify_socket() > rxrpc_input_ackall() * APPLY ACKALL PACKET TO CALL AND DISCARD PACKET ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_reject_packet() There are some functions used by the above that queue the packet, after which the procedure is terminated: - rxrpc_post_packet_to_local() - local->event_queue is an sk_buff_head - local->processor is a work_struct - rxrpc_post_packet_to_conn() - conn->rx_queue is an sk_buff_head - conn->processor is a work_struct - rxrpc_reject_packet() - local->reject_queue is an sk_buff_head - local->processor is a work_struct And some that offload processing to process context: - rxrpc_notify_socket() - Uses RCU lock - Uses call->notify_lock to call call->notify_rx - Uses call->recvmsg_lock to queue recvmsg side - rxrpc_queue_call() - call->processor is a work_struct - rxrpc_propose_ACK() - Uses call->lock to wrap __rxrpc_propose_ACK() And a bunch that complete a call, all of which use call->state_lock to protect the call state: - rxrpc_call_completed() - rxrpc_set_call_completion() - rxrpc_abort_call() - rxrpc_proto_abort() - Also uses rxrpc_queue_call() Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both") Signed-off-by: David Howells <dhowells@redhat.com>
2018-10-08 22:46:25 +08:00
if (skb_copy_bits(skb, offset, buf.acks, nr_acks) < 0) {
rxrpc_proto_abort("XSA", call, 0);
goto out;
}
rxrpc_input_soft_acks(call, buf.acks, first_soft_ack, nr_acks,
&summary);
rxrpc: Pass the last Tx packet marker in the annotation buffer When the last packet of data to be transmitted on a call is queued, tx_top is set and then the RXRPC_CALL_TX_LAST flag is set. Unfortunately, this leaves a race in the ACK processing side of things because the flag affects the interpretation of tx_top and also allows us to start receiving reply data before we've finished transmitting. To fix this, make the following changes: (1) rxrpc_queue_packet() now sets a marker in the annotation buffer instead of setting the RXRPC_CALL_TX_LAST flag. (2) rxrpc_rotate_tx_window() detects the marker and sets the flag in the same context as the routines that use it. (3) rxrpc_end_tx_phase() is simplified to just shift the call state. The Tx window must have been rotated before calling to discard the last packet. (4) rxrpc_receiving_reply() is added to handle the arrival of the first DATA packet of a reply to a client call (which is an implicit ACK of the Tx phase). (5) The last part of rxrpc_input_ack() is reordered to perform Tx rotation, then soft-ACK application and then to end the phase if we've rotated the last packet. In the event of a terminal ACK, the soft-ACK application will be skipped as nAcks should be 0. (6) rxrpc_input_ackall() now has to rotate as well as ending the phase. In addition: (7) Alter the transmit tracepoint to log the rotation of the last packet. (8) Remove the no-longer relevant queue_reqack tracepoint note. The ACK-REQUESTED packet header flag is now set as needed when we actually transmit the packet and may vary by retransmission. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-23 19:39:22 +08:00
}
if (call->rxtx_annotations[call->tx_top & RXRPC_RXTX_BUFF_MASK] &
RXRPC_TX_ANNO_LAST &&
summary.nr_acks == call->tx_top - hard_ack &&
rxrpc_is_client_call(call))
rxrpc_propose_ACK(call, RXRPC_ACK_PING, sp->hdr.serial,
false, true,
rxrpc_propose_ack_ping_for_lost_reply);
rxrpc: Fix the packet reception routine The rxrpc_input_packet() function and its call tree was built around the assumption that data_ready() handler called from UDP to inform a kernel service that there is data to be had was non-reentrant. This means that certain locking could be dispensed with. This, however, turns out not to be the case with a multi-queue network card that can deliver packets to multiple cpus simultaneously. Each of those cpus can be in the rxrpc_input_packet() function at the same time. Fix by adding or changing some structure members: (1) Add peer->rtt_input_lock to serialise access to the RTT buffer. (2) Make conn->service_id into a 32-bit variable so that it can be cmpxchg'd on all arches. (3) Add call->input_lock to serialise access to the Rx/Tx state. Note that although the Rx and Tx states are (almost) entirely separate, there's no point completing the separation and having separate locks since it's a bi-phasal RPC protocol rather than a bi-direction streaming protocol. Data transmission and data reception do not take place simultaneously on any particular call. and making the following functional changes: (1) In rxrpc_input_data(), hold call->input_lock around the core to prevent simultaneous producing of packets into the Rx ring and updating of tracking state for a particular call. (2) In rxrpc_input_ping_response(), only read call->ping_serial once, and check it before checking RXRPC_CALL_PINGING as that's a cheaper test. The bit test and bit clear can then be combined. No further locking is needed here. (3) In rxrpc_input_ack(), take call->input_lock after we've parsed much of the ACK packet. The superseded ACK check is then done both before and after the lock is taken. The handing of ackinfo data is split, parsing before the lock is taken and processing with it held. This is keyed on rxMTU being non-zero. Congestion management is also done within the locked section. (4) In rxrpc_input_ackall(), take call->input_lock around the Tx window rotation. The ACKALL packet carries no information and is only really useful after all packets have been transmitted since it's imprecise. (5) In rxrpc_input_implicit_end_call(), we use rx->incoming_lock to prevent calls being simultaneously implicitly ended on two cpus and also to prevent any races with incoming call setup. (6) In rxrpc_input_packet(), use cmpxchg() to effect the service upgrade on a connection. It is only permitted to happen once for a connection. (7) In rxrpc_new_incoming_call(), we have to recheck the routing inside rx->incoming_lock to see if someone else set up the call, connection or peer whilst we were getting there. We can't trust the values from the earlier routing check unless we pin refs on them - which we want to avoid. Further, we need to allow for an incoming call to have its state changed on another CPU between us making it live and us adjusting it because the conn is now in the RXRPC_CONN_SERVICE state. (8) In rxrpc_peer_add_rtt(), take peer->rtt_input_lock around the access to the RTT buffer. Don't need to lock around setting peer->rtt. For reference, the inventory of state-accessing or state-altering functions used by the packet input procedure is: > rxrpc_input_packet() * PACKET CHECKING * ROUTING > rxrpc_post_packet_to_local() > rxrpc_find_connection_rcu() - uses RCU > rxrpc_lookup_peer_rcu() - uses RCU > rxrpc_find_service_conn_rcu() - uses RCU > idr_find() - uses RCU * CONNECTION-LEVEL PROCESSING - Service upgrade - Can only happen once per conn ! Changed to use cmpxchg > rxrpc_post_packet_to_conn() - Setting conn->hi_serial - Probably safe not using locks - Maybe use cmpxchg * CALL-LEVEL PROCESSING > Old-call checking > rxrpc_input_implicit_end_call() > rxrpc_call_completed() > rxrpc_queue_call() ! Need to take rx->incoming_lock > __rxrpc_disconnect_call() > rxrpc_notify_socket() > rxrpc_new_incoming_call() - Uses rx->incoming_lock for the entire process - Might be able to drop this earlier in favour of the call lock > rxrpc_incoming_call() ! Conflicts with rxrpc_input_implicit_end_call() > rxrpc_send_ping() - Don't need locks to check rtt state > rxrpc_propose_ACK * PACKET DISTRIBUTION > rxrpc_input_call_packet() > rxrpc_input_data() * QUEUE DATA PACKET ON CALL > rxrpc_reduce_call_timer() - Uses timer_reduce() ! Needs call->input_lock() > rxrpc_receiving_reply() ! Needs locking around ack state > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_proto_abort() > rxrpc_input_dup_data() - Fills the Rx buffer - rxrpc_propose_ACK() - rxrpc_notify_socket() > rxrpc_input_ack() * APPLY ACK PACKET TO CALL AND DISCARD PACKET > rxrpc_input_ping_response() - Probably doesn't need any extra locking ! Need READ_ONCE() on call->ping_serial > rxrpc_input_check_for_lost_ack() - Takes call->lock to consult Tx buffer > rxrpc_peer_add_rtt() ! Needs to take a lock (peer->rtt_input_lock) ! Could perhaps manage with cmpxchg() and xadd() instead > rxrpc_input_requested_ack - Consults Tx buffer ! Probably needs a lock > rxrpc_peer_add_rtt() > rxrpc_propose_ack() > rxrpc_input_ackinfo() - Changes call->tx_winsize ! Use cmpxchg to handle change ! Should perhaps track serial number - Uses peer->lock to record MTU specification changes > rxrpc_proto_abort() ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_input_soft_acks() - Consults the Tx buffer > rxrpc_congestion_management() - Modifies the Tx annotations ! Needs call->input_lock() > rxrpc_queue_call() > rxrpc_input_abort() * APPLY ABORT PACKET TO CALL AND DISCARD PACKET > rxrpc_set_call_completion() > rxrpc_notify_socket() > rxrpc_input_ackall() * APPLY ACKALL PACKET TO CALL AND DISCARD PACKET ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_reject_packet() There are some functions used by the above that queue the packet, after which the procedure is terminated: - rxrpc_post_packet_to_local() - local->event_queue is an sk_buff_head - local->processor is a work_struct - rxrpc_post_packet_to_conn() - conn->rx_queue is an sk_buff_head - conn->processor is a work_struct - rxrpc_reject_packet() - local->reject_queue is an sk_buff_head - local->processor is a work_struct And some that offload processing to process context: - rxrpc_notify_socket() - Uses RCU lock - Uses call->notify_lock to call call->notify_rx - Uses call->recvmsg_lock to queue recvmsg side - rxrpc_queue_call() - call->processor is a work_struct - rxrpc_propose_ACK() - Uses call->lock to wrap __rxrpc_propose_ACK() And a bunch that complete a call, all of which use call->state_lock to protect the call state: - rxrpc_call_completed() - rxrpc_set_call_completion() - rxrpc_abort_call() - rxrpc_proto_abort() - Also uses rxrpc_queue_call() Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both") Signed-off-by: David Howells <dhowells@redhat.com>
2018-10-08 22:46:25 +08:00
rxrpc_congestion_management(call, skb, &summary, acked_serial);
out:
spin_unlock(&call->input_lock);
}
/*
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
* Process an ACKALL packet.
*/
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
static void rxrpc_input_ackall(struct rxrpc_call *call, struct sk_buff *skb)
{
struct rxrpc_ack_summary summary = { 0 };
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
_proto("Rx ACKALL %%%u", sp->hdr.serial);
rxrpc: Fix the packet reception routine The rxrpc_input_packet() function and its call tree was built around the assumption that data_ready() handler called from UDP to inform a kernel service that there is data to be had was non-reentrant. This means that certain locking could be dispensed with. This, however, turns out not to be the case with a multi-queue network card that can deliver packets to multiple cpus simultaneously. Each of those cpus can be in the rxrpc_input_packet() function at the same time. Fix by adding or changing some structure members: (1) Add peer->rtt_input_lock to serialise access to the RTT buffer. (2) Make conn->service_id into a 32-bit variable so that it can be cmpxchg'd on all arches. (3) Add call->input_lock to serialise access to the Rx/Tx state. Note that although the Rx and Tx states are (almost) entirely separate, there's no point completing the separation and having separate locks since it's a bi-phasal RPC protocol rather than a bi-direction streaming protocol. Data transmission and data reception do not take place simultaneously on any particular call. and making the following functional changes: (1) In rxrpc_input_data(), hold call->input_lock around the core to prevent simultaneous producing of packets into the Rx ring and updating of tracking state for a particular call. (2) In rxrpc_input_ping_response(), only read call->ping_serial once, and check it before checking RXRPC_CALL_PINGING as that's a cheaper test. The bit test and bit clear can then be combined. No further locking is needed here. (3) In rxrpc_input_ack(), take call->input_lock after we've parsed much of the ACK packet. The superseded ACK check is then done both before and after the lock is taken. The handing of ackinfo data is split, parsing before the lock is taken and processing with it held. This is keyed on rxMTU being non-zero. Congestion management is also done within the locked section. (4) In rxrpc_input_ackall(), take call->input_lock around the Tx window rotation. The ACKALL packet carries no information and is only really useful after all packets have been transmitted since it's imprecise. (5) In rxrpc_input_implicit_end_call(), we use rx->incoming_lock to prevent calls being simultaneously implicitly ended on two cpus and also to prevent any races with incoming call setup. (6) In rxrpc_input_packet(), use cmpxchg() to effect the service upgrade on a connection. It is only permitted to happen once for a connection. (7) In rxrpc_new_incoming_call(), we have to recheck the routing inside rx->incoming_lock to see if someone else set up the call, connection or peer whilst we were getting there. We can't trust the values from the earlier routing check unless we pin refs on them - which we want to avoid. Further, we need to allow for an incoming call to have its state changed on another CPU between us making it live and us adjusting it because the conn is now in the RXRPC_CONN_SERVICE state. (8) In rxrpc_peer_add_rtt(), take peer->rtt_input_lock around the access to the RTT buffer. Don't need to lock around setting peer->rtt. For reference, the inventory of state-accessing or state-altering functions used by the packet input procedure is: > rxrpc_input_packet() * PACKET CHECKING * ROUTING > rxrpc_post_packet_to_local() > rxrpc_find_connection_rcu() - uses RCU > rxrpc_lookup_peer_rcu() - uses RCU > rxrpc_find_service_conn_rcu() - uses RCU > idr_find() - uses RCU * CONNECTION-LEVEL PROCESSING - Service upgrade - Can only happen once per conn ! Changed to use cmpxchg > rxrpc_post_packet_to_conn() - Setting conn->hi_serial - Probably safe not using locks - Maybe use cmpxchg * CALL-LEVEL PROCESSING > Old-call checking > rxrpc_input_implicit_end_call() > rxrpc_call_completed() > rxrpc_queue_call() ! Need to take rx->incoming_lock > __rxrpc_disconnect_call() > rxrpc_notify_socket() > rxrpc_new_incoming_call() - Uses rx->incoming_lock for the entire process - Might be able to drop this earlier in favour of the call lock > rxrpc_incoming_call() ! Conflicts with rxrpc_input_implicit_end_call() > rxrpc_send_ping() - Don't need locks to check rtt state > rxrpc_propose_ACK * PACKET DISTRIBUTION > rxrpc_input_call_packet() > rxrpc_input_data() * QUEUE DATA PACKET ON CALL > rxrpc_reduce_call_timer() - Uses timer_reduce() ! Needs call->input_lock() > rxrpc_receiving_reply() ! Needs locking around ack state > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_proto_abort() > rxrpc_input_dup_data() - Fills the Rx buffer - rxrpc_propose_ACK() - rxrpc_notify_socket() > rxrpc_input_ack() * APPLY ACK PACKET TO CALL AND DISCARD PACKET > rxrpc_input_ping_response() - Probably doesn't need any extra locking ! Need READ_ONCE() on call->ping_serial > rxrpc_input_check_for_lost_ack() - Takes call->lock to consult Tx buffer > rxrpc_peer_add_rtt() ! Needs to take a lock (peer->rtt_input_lock) ! Could perhaps manage with cmpxchg() and xadd() instead > rxrpc_input_requested_ack - Consults Tx buffer ! Probably needs a lock > rxrpc_peer_add_rtt() > rxrpc_propose_ack() > rxrpc_input_ackinfo() - Changes call->tx_winsize ! Use cmpxchg to handle change ! Should perhaps track serial number - Uses peer->lock to record MTU specification changes > rxrpc_proto_abort() ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_input_soft_acks() - Consults the Tx buffer > rxrpc_congestion_management() - Modifies the Tx annotations ! Needs call->input_lock() > rxrpc_queue_call() > rxrpc_input_abort() * APPLY ABORT PACKET TO CALL AND DISCARD PACKET > rxrpc_set_call_completion() > rxrpc_notify_socket() > rxrpc_input_ackall() * APPLY ACKALL PACKET TO CALL AND DISCARD PACKET ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_reject_packet() There are some functions used by the above that queue the packet, after which the procedure is terminated: - rxrpc_post_packet_to_local() - local->event_queue is an sk_buff_head - local->processor is a work_struct - rxrpc_post_packet_to_conn() - conn->rx_queue is an sk_buff_head - conn->processor is a work_struct - rxrpc_reject_packet() - local->reject_queue is an sk_buff_head - local->processor is a work_struct And some that offload processing to process context: - rxrpc_notify_socket() - Uses RCU lock - Uses call->notify_lock to call call->notify_rx - Uses call->recvmsg_lock to queue recvmsg side - rxrpc_queue_call() - call->processor is a work_struct - rxrpc_propose_ACK() - Uses call->lock to wrap __rxrpc_propose_ACK() And a bunch that complete a call, all of which use call->state_lock to protect the call state: - rxrpc_call_completed() - rxrpc_set_call_completion() - rxrpc_abort_call() - rxrpc_proto_abort() - Also uses rxrpc_queue_call() Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both") Signed-off-by: David Howells <dhowells@redhat.com>
2018-10-08 22:46:25 +08:00
spin_lock(&call->input_lock);
if (rxrpc_rotate_tx_window(call, call->tx_top, &summary))
rxrpc: Pass the last Tx packet marker in the annotation buffer When the last packet of data to be transmitted on a call is queued, tx_top is set and then the RXRPC_CALL_TX_LAST flag is set. Unfortunately, this leaves a race in the ACK processing side of things because the flag affects the interpretation of tx_top and also allows us to start receiving reply data before we've finished transmitting. To fix this, make the following changes: (1) rxrpc_queue_packet() now sets a marker in the annotation buffer instead of setting the RXRPC_CALL_TX_LAST flag. (2) rxrpc_rotate_tx_window() detects the marker and sets the flag in the same context as the routines that use it. (3) rxrpc_end_tx_phase() is simplified to just shift the call state. The Tx window must have been rotated before calling to discard the last packet. (4) rxrpc_receiving_reply() is added to handle the arrival of the first DATA packet of a reply to a client call (which is an implicit ACK of the Tx phase). (5) The last part of rxrpc_input_ack() is reordered to perform Tx rotation, then soft-ACK application and then to end the phase if we've rotated the last packet. In the event of a terminal ACK, the soft-ACK application will be skipped as nAcks should be 0. (6) rxrpc_input_ackall() now has to rotate as well as ending the phase. In addition: (7) Alter the transmit tracepoint to log the rotation of the last packet. (8) Remove the no-longer relevant queue_reqack tracepoint note. The ACK-REQUESTED packet header flag is now set as needed when we actually transmit the packet and may vary by retransmission. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-23 19:39:22 +08:00
rxrpc_end_tx_phase(call, false, "ETL");
rxrpc: Fix the packet reception routine The rxrpc_input_packet() function and its call tree was built around the assumption that data_ready() handler called from UDP to inform a kernel service that there is data to be had was non-reentrant. This means that certain locking could be dispensed with. This, however, turns out not to be the case with a multi-queue network card that can deliver packets to multiple cpus simultaneously. Each of those cpus can be in the rxrpc_input_packet() function at the same time. Fix by adding or changing some structure members: (1) Add peer->rtt_input_lock to serialise access to the RTT buffer. (2) Make conn->service_id into a 32-bit variable so that it can be cmpxchg'd on all arches. (3) Add call->input_lock to serialise access to the Rx/Tx state. Note that although the Rx and Tx states are (almost) entirely separate, there's no point completing the separation and having separate locks since it's a bi-phasal RPC protocol rather than a bi-direction streaming protocol. Data transmission and data reception do not take place simultaneously on any particular call. and making the following functional changes: (1) In rxrpc_input_data(), hold call->input_lock around the core to prevent simultaneous producing of packets into the Rx ring and updating of tracking state for a particular call. (2) In rxrpc_input_ping_response(), only read call->ping_serial once, and check it before checking RXRPC_CALL_PINGING as that's a cheaper test. The bit test and bit clear can then be combined. No further locking is needed here. (3) In rxrpc_input_ack(), take call->input_lock after we've parsed much of the ACK packet. The superseded ACK check is then done both before and after the lock is taken. The handing of ackinfo data is split, parsing before the lock is taken and processing with it held. This is keyed on rxMTU being non-zero. Congestion management is also done within the locked section. (4) In rxrpc_input_ackall(), take call->input_lock around the Tx window rotation. The ACKALL packet carries no information and is only really useful after all packets have been transmitted since it's imprecise. (5) In rxrpc_input_implicit_end_call(), we use rx->incoming_lock to prevent calls being simultaneously implicitly ended on two cpus and also to prevent any races with incoming call setup. (6) In rxrpc_input_packet(), use cmpxchg() to effect the service upgrade on a connection. It is only permitted to happen once for a connection. (7) In rxrpc_new_incoming_call(), we have to recheck the routing inside rx->incoming_lock to see if someone else set up the call, connection or peer whilst we were getting there. We can't trust the values from the earlier routing check unless we pin refs on them - which we want to avoid. Further, we need to allow for an incoming call to have its state changed on another CPU between us making it live and us adjusting it because the conn is now in the RXRPC_CONN_SERVICE state. (8) In rxrpc_peer_add_rtt(), take peer->rtt_input_lock around the access to the RTT buffer. Don't need to lock around setting peer->rtt. For reference, the inventory of state-accessing or state-altering functions used by the packet input procedure is: > rxrpc_input_packet() * PACKET CHECKING * ROUTING > rxrpc_post_packet_to_local() > rxrpc_find_connection_rcu() - uses RCU > rxrpc_lookup_peer_rcu() - uses RCU > rxrpc_find_service_conn_rcu() - uses RCU > idr_find() - uses RCU * CONNECTION-LEVEL PROCESSING - Service upgrade - Can only happen once per conn ! Changed to use cmpxchg > rxrpc_post_packet_to_conn() - Setting conn->hi_serial - Probably safe not using locks - Maybe use cmpxchg * CALL-LEVEL PROCESSING > Old-call checking > rxrpc_input_implicit_end_call() > rxrpc_call_completed() > rxrpc_queue_call() ! Need to take rx->incoming_lock > __rxrpc_disconnect_call() > rxrpc_notify_socket() > rxrpc_new_incoming_call() - Uses rx->incoming_lock for the entire process - Might be able to drop this earlier in favour of the call lock > rxrpc_incoming_call() ! Conflicts with rxrpc_input_implicit_end_call() > rxrpc_send_ping() - Don't need locks to check rtt state > rxrpc_propose_ACK * PACKET DISTRIBUTION > rxrpc_input_call_packet() > rxrpc_input_data() * QUEUE DATA PACKET ON CALL > rxrpc_reduce_call_timer() - Uses timer_reduce() ! Needs call->input_lock() > rxrpc_receiving_reply() ! Needs locking around ack state > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_proto_abort() > rxrpc_input_dup_data() - Fills the Rx buffer - rxrpc_propose_ACK() - rxrpc_notify_socket() > rxrpc_input_ack() * APPLY ACK PACKET TO CALL AND DISCARD PACKET > rxrpc_input_ping_response() - Probably doesn't need any extra locking ! Need READ_ONCE() on call->ping_serial > rxrpc_input_check_for_lost_ack() - Takes call->lock to consult Tx buffer > rxrpc_peer_add_rtt() ! Needs to take a lock (peer->rtt_input_lock) ! Could perhaps manage with cmpxchg() and xadd() instead > rxrpc_input_requested_ack - Consults Tx buffer ! Probably needs a lock > rxrpc_peer_add_rtt() > rxrpc_propose_ack() > rxrpc_input_ackinfo() - Changes call->tx_winsize ! Use cmpxchg to handle change ! Should perhaps track serial number - Uses peer->lock to record MTU specification changes > rxrpc_proto_abort() ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_input_soft_acks() - Consults the Tx buffer > rxrpc_congestion_management() - Modifies the Tx annotations ! Needs call->input_lock() > rxrpc_queue_call() > rxrpc_input_abort() * APPLY ABORT PACKET TO CALL AND DISCARD PACKET > rxrpc_set_call_completion() > rxrpc_notify_socket() > rxrpc_input_ackall() * APPLY ACKALL PACKET TO CALL AND DISCARD PACKET ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_reject_packet() There are some functions used by the above that queue the packet, after which the procedure is terminated: - rxrpc_post_packet_to_local() - local->event_queue is an sk_buff_head - local->processor is a work_struct - rxrpc_post_packet_to_conn() - conn->rx_queue is an sk_buff_head - conn->processor is a work_struct - rxrpc_reject_packet() - local->reject_queue is an sk_buff_head - local->processor is a work_struct And some that offload processing to process context: - rxrpc_notify_socket() - Uses RCU lock - Uses call->notify_lock to call call->notify_rx - Uses call->recvmsg_lock to queue recvmsg side - rxrpc_queue_call() - call->processor is a work_struct - rxrpc_propose_ACK() - Uses call->lock to wrap __rxrpc_propose_ACK() And a bunch that complete a call, all of which use call->state_lock to protect the call state: - rxrpc_call_completed() - rxrpc_set_call_completion() - rxrpc_abort_call() - rxrpc_proto_abort() - Also uses rxrpc_queue_call() Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both") Signed-off-by: David Howells <dhowells@redhat.com>
2018-10-08 22:46:25 +08:00
spin_unlock(&call->input_lock);
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
}
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
/*
* Process an ABORT packet directed at a call.
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
*/
static void rxrpc_input_abort(struct rxrpc_call *call, struct sk_buff *skb)
{
struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
__be32 wtmp;
u32 abort_code = RX_CALL_DEAD;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
_enter("");
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
if (skb->len >= 4 &&
skb_copy_bits(skb, sizeof(struct rxrpc_wire_header),
&wtmp, sizeof(wtmp)) >= 0)
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
abort_code = ntohl(wtmp);
trace_rxrpc_rx_abort(call, sp->hdr.serial, abort_code);
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
_proto("Rx ABORT %%%u { %x }", sp->hdr.serial, abort_code);
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
if (rxrpc_set_call_completion(call, RXRPC_CALL_REMOTELY_ABORTED,
abort_code, -ECONNABORTED))
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
rxrpc_notify_socket(call);
}
/*
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
* Process an incoming call packet.
*/
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
static void rxrpc_input_call_packet(struct rxrpc_call *call,
struct sk_buff *skb)
{
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
struct rxrpc_skb_priv *sp = rxrpc_skb(skb);
rxrpc: Fix call timeouts Fix the rxrpc call expiration timeouts and make them settable from userspace. By analogy with other rx implementations, there should be three timeouts: (1) "Normal timeout" This is set for all calls and is triggered if we haven't received any packets from the peer in a while. It is measured from the last time we received any packet on that call. This is not reset by any connection packets (such as CHALLENGE/RESPONSE packets). If a service operation takes a long time, the server should generate PING ACKs at a duration that's substantially less than the normal timeout so is to keep both sides alive. This is set at 1/6 of normal timeout. (2) "Idle timeout" This is set only for a service call and is triggered if we stop receiving the DATA packets that comprise the request data. It is measured from the last time we received a DATA packet. (3) "Hard timeout" This can be set for a call and specified the maximum lifetime of that call. It should not be specified by default. Some operations (such as volume transfer) take a long time. Allow userspace to set/change the timeouts on a call with sendmsg, using a control message: RXRPC_SET_CALL_TIMEOUTS The data to the message is a number of 32-bit words, not all of which need be given: u32 hard_timeout; /* sec from first packet */ u32 idle_timeout; /* msec from packet Rx */ u32 normal_timeout; /* msec from data Rx */ This can be set in combination with any other sendmsg() that affects a call. Signed-off-by: David Howells <dhowells@redhat.com>
2017-11-24 18:18:41 +08:00
unsigned long timo;
_enter("%p,%p", call, skb);
rxrpc: Fix call timeouts Fix the rxrpc call expiration timeouts and make them settable from userspace. By analogy with other rx implementations, there should be three timeouts: (1) "Normal timeout" This is set for all calls and is triggered if we haven't received any packets from the peer in a while. It is measured from the last time we received any packet on that call. This is not reset by any connection packets (such as CHALLENGE/RESPONSE packets). If a service operation takes a long time, the server should generate PING ACKs at a duration that's substantially less than the normal timeout so is to keep both sides alive. This is set at 1/6 of normal timeout. (2) "Idle timeout" This is set only for a service call and is triggered if we stop receiving the DATA packets that comprise the request data. It is measured from the last time we received a DATA packet. (3) "Hard timeout" This can be set for a call and specified the maximum lifetime of that call. It should not be specified by default. Some operations (such as volume transfer) take a long time. Allow userspace to set/change the timeouts on a call with sendmsg, using a control message: RXRPC_SET_CALL_TIMEOUTS The data to the message is a number of 32-bit words, not all of which need be given: u32 hard_timeout; /* sec from first packet */ u32 idle_timeout; /* msec from packet Rx */ u32 normal_timeout; /* msec from data Rx */ This can be set in combination with any other sendmsg() that affects a call. Signed-off-by: David Howells <dhowells@redhat.com>
2017-11-24 18:18:41 +08:00
timo = READ_ONCE(call->next_rx_timo);
if (timo) {
unsigned long now = jiffies, expect_rx_by;
expect_rx_by = now + timo;
rxrpc: Fix call timeouts Fix the rxrpc call expiration timeouts and make them settable from userspace. By analogy with other rx implementations, there should be three timeouts: (1) "Normal timeout" This is set for all calls and is triggered if we haven't received any packets from the peer in a while. It is measured from the last time we received any packet on that call. This is not reset by any connection packets (such as CHALLENGE/RESPONSE packets). If a service operation takes a long time, the server should generate PING ACKs at a duration that's substantially less than the normal timeout so is to keep both sides alive. This is set at 1/6 of normal timeout. (2) "Idle timeout" This is set only for a service call and is triggered if we stop receiving the DATA packets that comprise the request data. It is measured from the last time we received a DATA packet. (3) "Hard timeout" This can be set for a call and specified the maximum lifetime of that call. It should not be specified by default. Some operations (such as volume transfer) take a long time. Allow userspace to set/change the timeouts on a call with sendmsg, using a control message: RXRPC_SET_CALL_TIMEOUTS The data to the message is a number of 32-bit words, not all of which need be given: u32 hard_timeout; /* sec from first packet */ u32 idle_timeout; /* msec from packet Rx */ u32 normal_timeout; /* msec from data Rx */ This can be set in combination with any other sendmsg() that affects a call. Signed-off-by: David Howells <dhowells@redhat.com>
2017-11-24 18:18:41 +08:00
WRITE_ONCE(call->expect_rx_by, expect_rx_by);
rxrpc_reduce_call_timer(call, expect_rx_by, now,
rxrpc_timer_set_for_normal);
}
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
switch (sp->hdr.type) {
case RXRPC_PACKET_TYPE_DATA:
rxrpc_input_data(call, skb);
goto no_free;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
case RXRPC_PACKET_TYPE_ACK:
rxrpc_input_ack(call, skb);
break;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
case RXRPC_PACKET_TYPE_BUSY:
_proto("Rx BUSY %%%u", sp->hdr.serial);
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
/* Just ignore BUSY packets from the server; the retry and
* lifespan timers will take care of business. BUSY packets
* from the client don't make sense.
*/
break;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
case RXRPC_PACKET_TYPE_ABORT:
rxrpc_input_abort(call, skb);
break;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
case RXRPC_PACKET_TYPE_ACKALL:
rxrpc_input_ackall(call, skb);
break;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
default:
break;
}
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
rxrpc_free_skb(skb, rxrpc_skb_freed);
no_free:
_leave("");
}
/*
rxrpc: Fix the packet reception routine The rxrpc_input_packet() function and its call tree was built around the assumption that data_ready() handler called from UDP to inform a kernel service that there is data to be had was non-reentrant. This means that certain locking could be dispensed with. This, however, turns out not to be the case with a multi-queue network card that can deliver packets to multiple cpus simultaneously. Each of those cpus can be in the rxrpc_input_packet() function at the same time. Fix by adding or changing some structure members: (1) Add peer->rtt_input_lock to serialise access to the RTT buffer. (2) Make conn->service_id into a 32-bit variable so that it can be cmpxchg'd on all arches. (3) Add call->input_lock to serialise access to the Rx/Tx state. Note that although the Rx and Tx states are (almost) entirely separate, there's no point completing the separation and having separate locks since it's a bi-phasal RPC protocol rather than a bi-direction streaming protocol. Data transmission and data reception do not take place simultaneously on any particular call. and making the following functional changes: (1) In rxrpc_input_data(), hold call->input_lock around the core to prevent simultaneous producing of packets into the Rx ring and updating of tracking state for a particular call. (2) In rxrpc_input_ping_response(), only read call->ping_serial once, and check it before checking RXRPC_CALL_PINGING as that's a cheaper test. The bit test and bit clear can then be combined. No further locking is needed here. (3) In rxrpc_input_ack(), take call->input_lock after we've parsed much of the ACK packet. The superseded ACK check is then done both before and after the lock is taken. The handing of ackinfo data is split, parsing before the lock is taken and processing with it held. This is keyed on rxMTU being non-zero. Congestion management is also done within the locked section. (4) In rxrpc_input_ackall(), take call->input_lock around the Tx window rotation. The ACKALL packet carries no information and is only really useful after all packets have been transmitted since it's imprecise. (5) In rxrpc_input_implicit_end_call(), we use rx->incoming_lock to prevent calls being simultaneously implicitly ended on two cpus and also to prevent any races with incoming call setup. (6) In rxrpc_input_packet(), use cmpxchg() to effect the service upgrade on a connection. It is only permitted to happen once for a connection. (7) In rxrpc_new_incoming_call(), we have to recheck the routing inside rx->incoming_lock to see if someone else set up the call, connection or peer whilst we were getting there. We can't trust the values from the earlier routing check unless we pin refs on them - which we want to avoid. Further, we need to allow for an incoming call to have its state changed on another CPU between us making it live and us adjusting it because the conn is now in the RXRPC_CONN_SERVICE state. (8) In rxrpc_peer_add_rtt(), take peer->rtt_input_lock around the access to the RTT buffer. Don't need to lock around setting peer->rtt. For reference, the inventory of state-accessing or state-altering functions used by the packet input procedure is: > rxrpc_input_packet() * PACKET CHECKING * ROUTING > rxrpc_post_packet_to_local() > rxrpc_find_connection_rcu() - uses RCU > rxrpc_lookup_peer_rcu() - uses RCU > rxrpc_find_service_conn_rcu() - uses RCU > idr_find() - uses RCU * CONNECTION-LEVEL PROCESSING - Service upgrade - Can only happen once per conn ! Changed to use cmpxchg > rxrpc_post_packet_to_conn() - Setting conn->hi_serial - Probably safe not using locks - Maybe use cmpxchg * CALL-LEVEL PROCESSING > Old-call checking > rxrpc_input_implicit_end_call() > rxrpc_call_completed() > rxrpc_queue_call() ! Need to take rx->incoming_lock > __rxrpc_disconnect_call() > rxrpc_notify_socket() > rxrpc_new_incoming_call() - Uses rx->incoming_lock for the entire process - Might be able to drop this earlier in favour of the call lock > rxrpc_incoming_call() ! Conflicts with rxrpc_input_implicit_end_call() > rxrpc_send_ping() - Don't need locks to check rtt state > rxrpc_propose_ACK * PACKET DISTRIBUTION > rxrpc_input_call_packet() > rxrpc_input_data() * QUEUE DATA PACKET ON CALL > rxrpc_reduce_call_timer() - Uses timer_reduce() ! Needs call->input_lock() > rxrpc_receiving_reply() ! Needs locking around ack state > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_proto_abort() > rxrpc_input_dup_data() - Fills the Rx buffer - rxrpc_propose_ACK() - rxrpc_notify_socket() > rxrpc_input_ack() * APPLY ACK PACKET TO CALL AND DISCARD PACKET > rxrpc_input_ping_response() - Probably doesn't need any extra locking ! Need READ_ONCE() on call->ping_serial > rxrpc_input_check_for_lost_ack() - Takes call->lock to consult Tx buffer > rxrpc_peer_add_rtt() ! Needs to take a lock (peer->rtt_input_lock) ! Could perhaps manage with cmpxchg() and xadd() instead > rxrpc_input_requested_ack - Consults Tx buffer ! Probably needs a lock > rxrpc_peer_add_rtt() > rxrpc_propose_ack() > rxrpc_input_ackinfo() - Changes call->tx_winsize ! Use cmpxchg to handle change ! Should perhaps track serial number - Uses peer->lock to record MTU specification changes > rxrpc_proto_abort() ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_input_soft_acks() - Consults the Tx buffer > rxrpc_congestion_management() - Modifies the Tx annotations ! Needs call->input_lock() > rxrpc_queue_call() > rxrpc_input_abort() * APPLY ABORT PACKET TO CALL AND DISCARD PACKET > rxrpc_set_call_completion() > rxrpc_notify_socket() > rxrpc_input_ackall() * APPLY ACKALL PACKET TO CALL AND DISCARD PACKET ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_reject_packet() There are some functions used by the above that queue the packet, after which the procedure is terminated: - rxrpc_post_packet_to_local() - local->event_queue is an sk_buff_head - local->processor is a work_struct - rxrpc_post_packet_to_conn() - conn->rx_queue is an sk_buff_head - conn->processor is a work_struct - rxrpc_reject_packet() - local->reject_queue is an sk_buff_head - local->processor is a work_struct And some that offload processing to process context: - rxrpc_notify_socket() - Uses RCU lock - Uses call->notify_lock to call call->notify_rx - Uses call->recvmsg_lock to queue recvmsg side - rxrpc_queue_call() - call->processor is a work_struct - rxrpc_propose_ACK() - Uses call->lock to wrap __rxrpc_propose_ACK() And a bunch that complete a call, all of which use call->state_lock to protect the call state: - rxrpc_call_completed() - rxrpc_set_call_completion() - rxrpc_abort_call() - rxrpc_proto_abort() - Also uses rxrpc_queue_call() Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both") Signed-off-by: David Howells <dhowells@redhat.com>
2018-10-08 22:46:25 +08:00
* Handle a new service call on a channel implicitly completing the preceding
* call on that channel. This does not apply to client conns.
*
* TODO: If callNumber > call_id + 1, renegotiate security.
*/
rxrpc: Fix the packet reception routine The rxrpc_input_packet() function and its call tree was built around the assumption that data_ready() handler called from UDP to inform a kernel service that there is data to be had was non-reentrant. This means that certain locking could be dispensed with. This, however, turns out not to be the case with a multi-queue network card that can deliver packets to multiple cpus simultaneously. Each of those cpus can be in the rxrpc_input_packet() function at the same time. Fix by adding or changing some structure members: (1) Add peer->rtt_input_lock to serialise access to the RTT buffer. (2) Make conn->service_id into a 32-bit variable so that it can be cmpxchg'd on all arches. (3) Add call->input_lock to serialise access to the Rx/Tx state. Note that although the Rx and Tx states are (almost) entirely separate, there's no point completing the separation and having separate locks since it's a bi-phasal RPC protocol rather than a bi-direction streaming protocol. Data transmission and data reception do not take place simultaneously on any particular call. and making the following functional changes: (1) In rxrpc_input_data(), hold call->input_lock around the core to prevent simultaneous producing of packets into the Rx ring and updating of tracking state for a particular call. (2) In rxrpc_input_ping_response(), only read call->ping_serial once, and check it before checking RXRPC_CALL_PINGING as that's a cheaper test. The bit test and bit clear can then be combined. No further locking is needed here. (3) In rxrpc_input_ack(), take call->input_lock after we've parsed much of the ACK packet. The superseded ACK check is then done both before and after the lock is taken. The handing of ackinfo data is split, parsing before the lock is taken and processing with it held. This is keyed on rxMTU being non-zero. Congestion management is also done within the locked section. (4) In rxrpc_input_ackall(), take call->input_lock around the Tx window rotation. The ACKALL packet carries no information and is only really useful after all packets have been transmitted since it's imprecise. (5) In rxrpc_input_implicit_end_call(), we use rx->incoming_lock to prevent calls being simultaneously implicitly ended on two cpus and also to prevent any races with incoming call setup. (6) In rxrpc_input_packet(), use cmpxchg() to effect the service upgrade on a connection. It is only permitted to happen once for a connection. (7) In rxrpc_new_incoming_call(), we have to recheck the routing inside rx->incoming_lock to see if someone else set up the call, connection or peer whilst we were getting there. We can't trust the values from the earlier routing check unless we pin refs on them - which we want to avoid. Further, we need to allow for an incoming call to have its state changed on another CPU between us making it live and us adjusting it because the conn is now in the RXRPC_CONN_SERVICE state. (8) In rxrpc_peer_add_rtt(), take peer->rtt_input_lock around the access to the RTT buffer. Don't need to lock around setting peer->rtt. For reference, the inventory of state-accessing or state-altering functions used by the packet input procedure is: > rxrpc_input_packet() * PACKET CHECKING * ROUTING > rxrpc_post_packet_to_local() > rxrpc_find_connection_rcu() - uses RCU > rxrpc_lookup_peer_rcu() - uses RCU > rxrpc_find_service_conn_rcu() - uses RCU > idr_find() - uses RCU * CONNECTION-LEVEL PROCESSING - Service upgrade - Can only happen once per conn ! Changed to use cmpxchg > rxrpc_post_packet_to_conn() - Setting conn->hi_serial - Probably safe not using locks - Maybe use cmpxchg * CALL-LEVEL PROCESSING > Old-call checking > rxrpc_input_implicit_end_call() > rxrpc_call_completed() > rxrpc_queue_call() ! Need to take rx->incoming_lock > __rxrpc_disconnect_call() > rxrpc_notify_socket() > rxrpc_new_incoming_call() - Uses rx->incoming_lock for the entire process - Might be able to drop this earlier in favour of the call lock > rxrpc_incoming_call() ! Conflicts with rxrpc_input_implicit_end_call() > rxrpc_send_ping() - Don't need locks to check rtt state > rxrpc_propose_ACK * PACKET DISTRIBUTION > rxrpc_input_call_packet() > rxrpc_input_data() * QUEUE DATA PACKET ON CALL > rxrpc_reduce_call_timer() - Uses timer_reduce() ! Needs call->input_lock() > rxrpc_receiving_reply() ! Needs locking around ack state > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_proto_abort() > rxrpc_input_dup_data() - Fills the Rx buffer - rxrpc_propose_ACK() - rxrpc_notify_socket() > rxrpc_input_ack() * APPLY ACK PACKET TO CALL AND DISCARD PACKET > rxrpc_input_ping_response() - Probably doesn't need any extra locking ! Need READ_ONCE() on call->ping_serial > rxrpc_input_check_for_lost_ack() - Takes call->lock to consult Tx buffer > rxrpc_peer_add_rtt() ! Needs to take a lock (peer->rtt_input_lock) ! Could perhaps manage with cmpxchg() and xadd() instead > rxrpc_input_requested_ack - Consults Tx buffer ! Probably needs a lock > rxrpc_peer_add_rtt() > rxrpc_propose_ack() > rxrpc_input_ackinfo() - Changes call->tx_winsize ! Use cmpxchg to handle change ! Should perhaps track serial number - Uses peer->lock to record MTU specification changes > rxrpc_proto_abort() ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_input_soft_acks() - Consults the Tx buffer > rxrpc_congestion_management() - Modifies the Tx annotations ! Needs call->input_lock() > rxrpc_queue_call() > rxrpc_input_abort() * APPLY ABORT PACKET TO CALL AND DISCARD PACKET > rxrpc_set_call_completion() > rxrpc_notify_socket() > rxrpc_input_ackall() * APPLY ACKALL PACKET TO CALL AND DISCARD PACKET ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_reject_packet() There are some functions used by the above that queue the packet, after which the procedure is terminated: - rxrpc_post_packet_to_local() - local->event_queue is an sk_buff_head - local->processor is a work_struct - rxrpc_post_packet_to_conn() - conn->rx_queue is an sk_buff_head - conn->processor is a work_struct - rxrpc_reject_packet() - local->reject_queue is an sk_buff_head - local->processor is a work_struct And some that offload processing to process context: - rxrpc_notify_socket() - Uses RCU lock - Uses call->notify_lock to call call->notify_rx - Uses call->recvmsg_lock to queue recvmsg side - rxrpc_queue_call() - call->processor is a work_struct - rxrpc_propose_ACK() - Uses call->lock to wrap __rxrpc_propose_ACK() And a bunch that complete a call, all of which use call->state_lock to protect the call state: - rxrpc_call_completed() - rxrpc_set_call_completion() - rxrpc_abort_call() - rxrpc_proto_abort() - Also uses rxrpc_queue_call() Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both") Signed-off-by: David Howells <dhowells@redhat.com>
2018-10-08 22:46:25 +08:00
static void rxrpc_input_implicit_end_call(struct rxrpc_sock *rx,
struct rxrpc_connection *conn,
struct rxrpc_call *call)
{
switch (READ_ONCE(call->state)) {
case RXRPC_CALL_SERVER_AWAIT_ACK:
rxrpc_call_completed(call);
rxrpc: Fix the packet reception routine The rxrpc_input_packet() function and its call tree was built around the assumption that data_ready() handler called from UDP to inform a kernel service that there is data to be had was non-reentrant. This means that certain locking could be dispensed with. This, however, turns out not to be the case with a multi-queue network card that can deliver packets to multiple cpus simultaneously. Each of those cpus can be in the rxrpc_input_packet() function at the same time. Fix by adding or changing some structure members: (1) Add peer->rtt_input_lock to serialise access to the RTT buffer. (2) Make conn->service_id into a 32-bit variable so that it can be cmpxchg'd on all arches. (3) Add call->input_lock to serialise access to the Rx/Tx state. Note that although the Rx and Tx states are (almost) entirely separate, there's no point completing the separation and having separate locks since it's a bi-phasal RPC protocol rather than a bi-direction streaming protocol. Data transmission and data reception do not take place simultaneously on any particular call. and making the following functional changes: (1) In rxrpc_input_data(), hold call->input_lock around the core to prevent simultaneous producing of packets into the Rx ring and updating of tracking state for a particular call. (2) In rxrpc_input_ping_response(), only read call->ping_serial once, and check it before checking RXRPC_CALL_PINGING as that's a cheaper test. The bit test and bit clear can then be combined. No further locking is needed here. (3) In rxrpc_input_ack(), take call->input_lock after we've parsed much of the ACK packet. The superseded ACK check is then done both before and after the lock is taken. The handing of ackinfo data is split, parsing before the lock is taken and processing with it held. This is keyed on rxMTU being non-zero. Congestion management is also done within the locked section. (4) In rxrpc_input_ackall(), take call->input_lock around the Tx window rotation. The ACKALL packet carries no information and is only really useful after all packets have been transmitted since it's imprecise. (5) In rxrpc_input_implicit_end_call(), we use rx->incoming_lock to prevent calls being simultaneously implicitly ended on two cpus and also to prevent any races with incoming call setup. (6) In rxrpc_input_packet(), use cmpxchg() to effect the service upgrade on a connection. It is only permitted to happen once for a connection. (7) In rxrpc_new_incoming_call(), we have to recheck the routing inside rx->incoming_lock to see if someone else set up the call, connection or peer whilst we were getting there. We can't trust the values from the earlier routing check unless we pin refs on them - which we want to avoid. Further, we need to allow for an incoming call to have its state changed on another CPU between us making it live and us adjusting it because the conn is now in the RXRPC_CONN_SERVICE state. (8) In rxrpc_peer_add_rtt(), take peer->rtt_input_lock around the access to the RTT buffer. Don't need to lock around setting peer->rtt. For reference, the inventory of state-accessing or state-altering functions used by the packet input procedure is: > rxrpc_input_packet() * PACKET CHECKING * ROUTING > rxrpc_post_packet_to_local() > rxrpc_find_connection_rcu() - uses RCU > rxrpc_lookup_peer_rcu() - uses RCU > rxrpc_find_service_conn_rcu() - uses RCU > idr_find() - uses RCU * CONNECTION-LEVEL PROCESSING - Service upgrade - Can only happen once per conn ! Changed to use cmpxchg > rxrpc_post_packet_to_conn() - Setting conn->hi_serial - Probably safe not using locks - Maybe use cmpxchg * CALL-LEVEL PROCESSING > Old-call checking > rxrpc_input_implicit_end_call() > rxrpc_call_completed() > rxrpc_queue_call() ! Need to take rx->incoming_lock > __rxrpc_disconnect_call() > rxrpc_notify_socket() > rxrpc_new_incoming_call() - Uses rx->incoming_lock for the entire process - Might be able to drop this earlier in favour of the call lock > rxrpc_incoming_call() ! Conflicts with rxrpc_input_implicit_end_call() > rxrpc_send_ping() - Don't need locks to check rtt state > rxrpc_propose_ACK * PACKET DISTRIBUTION > rxrpc_input_call_packet() > rxrpc_input_data() * QUEUE DATA PACKET ON CALL > rxrpc_reduce_call_timer() - Uses timer_reduce() ! Needs call->input_lock() > rxrpc_receiving_reply() ! Needs locking around ack state > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_proto_abort() > rxrpc_input_dup_data() - Fills the Rx buffer - rxrpc_propose_ACK() - rxrpc_notify_socket() > rxrpc_input_ack() * APPLY ACK PACKET TO CALL AND DISCARD PACKET > rxrpc_input_ping_response() - Probably doesn't need any extra locking ! Need READ_ONCE() on call->ping_serial > rxrpc_input_check_for_lost_ack() - Takes call->lock to consult Tx buffer > rxrpc_peer_add_rtt() ! Needs to take a lock (peer->rtt_input_lock) ! Could perhaps manage with cmpxchg() and xadd() instead > rxrpc_input_requested_ack - Consults Tx buffer ! Probably needs a lock > rxrpc_peer_add_rtt() > rxrpc_propose_ack() > rxrpc_input_ackinfo() - Changes call->tx_winsize ! Use cmpxchg to handle change ! Should perhaps track serial number - Uses peer->lock to record MTU specification changes > rxrpc_proto_abort() ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_input_soft_acks() - Consults the Tx buffer > rxrpc_congestion_management() - Modifies the Tx annotations ! Needs call->input_lock() > rxrpc_queue_call() > rxrpc_input_abort() * APPLY ABORT PACKET TO CALL AND DISCARD PACKET > rxrpc_set_call_completion() > rxrpc_notify_socket() > rxrpc_input_ackall() * APPLY ACKALL PACKET TO CALL AND DISCARD PACKET ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_reject_packet() There are some functions used by the above that queue the packet, after which the procedure is terminated: - rxrpc_post_packet_to_local() - local->event_queue is an sk_buff_head - local->processor is a work_struct - rxrpc_post_packet_to_conn() - conn->rx_queue is an sk_buff_head - conn->processor is a work_struct - rxrpc_reject_packet() - local->reject_queue is an sk_buff_head - local->processor is a work_struct And some that offload processing to process context: - rxrpc_notify_socket() - Uses RCU lock - Uses call->notify_lock to call call->notify_rx - Uses call->recvmsg_lock to queue recvmsg side - rxrpc_queue_call() - call->processor is a work_struct - rxrpc_propose_ACK() - Uses call->lock to wrap __rxrpc_propose_ACK() And a bunch that complete a call, all of which use call->state_lock to protect the call state: - rxrpc_call_completed() - rxrpc_set_call_completion() - rxrpc_abort_call() - rxrpc_proto_abort() - Also uses rxrpc_queue_call() Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both") Signed-off-by: David Howells <dhowells@redhat.com>
2018-10-08 22:46:25 +08:00
/* Fall through */
case RXRPC_CALL_COMPLETE:
break;
default:
if (rxrpc_abort_call("IMP", call, 0, RX_CALL_DEAD, -ESHUTDOWN)) {
set_bit(RXRPC_CALL_EV_ABORT, &call->events);
rxrpc_queue_call(call);
}
rxrpc: Fix the packet reception routine The rxrpc_input_packet() function and its call tree was built around the assumption that data_ready() handler called from UDP to inform a kernel service that there is data to be had was non-reentrant. This means that certain locking could be dispensed with. This, however, turns out not to be the case with a multi-queue network card that can deliver packets to multiple cpus simultaneously. Each of those cpus can be in the rxrpc_input_packet() function at the same time. Fix by adding or changing some structure members: (1) Add peer->rtt_input_lock to serialise access to the RTT buffer. (2) Make conn->service_id into a 32-bit variable so that it can be cmpxchg'd on all arches. (3) Add call->input_lock to serialise access to the Rx/Tx state. Note that although the Rx and Tx states are (almost) entirely separate, there's no point completing the separation and having separate locks since it's a bi-phasal RPC protocol rather than a bi-direction streaming protocol. Data transmission and data reception do not take place simultaneously on any particular call. and making the following functional changes: (1) In rxrpc_input_data(), hold call->input_lock around the core to prevent simultaneous producing of packets into the Rx ring and updating of tracking state for a particular call. (2) In rxrpc_input_ping_response(), only read call->ping_serial once, and check it before checking RXRPC_CALL_PINGING as that's a cheaper test. The bit test and bit clear can then be combined. No further locking is needed here. (3) In rxrpc_input_ack(), take call->input_lock after we've parsed much of the ACK packet. The superseded ACK check is then done both before and after the lock is taken. The handing of ackinfo data is split, parsing before the lock is taken and processing with it held. This is keyed on rxMTU being non-zero. Congestion management is also done within the locked section. (4) In rxrpc_input_ackall(), take call->input_lock around the Tx window rotation. The ACKALL packet carries no information and is only really useful after all packets have been transmitted since it's imprecise. (5) In rxrpc_input_implicit_end_call(), we use rx->incoming_lock to prevent calls being simultaneously implicitly ended on two cpus and also to prevent any races with incoming call setup. (6) In rxrpc_input_packet(), use cmpxchg() to effect the service upgrade on a connection. It is only permitted to happen once for a connection. (7) In rxrpc_new_incoming_call(), we have to recheck the routing inside rx->incoming_lock to see if someone else set up the call, connection or peer whilst we were getting there. We can't trust the values from the earlier routing check unless we pin refs on them - which we want to avoid. Further, we need to allow for an incoming call to have its state changed on another CPU between us making it live and us adjusting it because the conn is now in the RXRPC_CONN_SERVICE state. (8) In rxrpc_peer_add_rtt(), take peer->rtt_input_lock around the access to the RTT buffer. Don't need to lock around setting peer->rtt. For reference, the inventory of state-accessing or state-altering functions used by the packet input procedure is: > rxrpc_input_packet() * PACKET CHECKING * ROUTING > rxrpc_post_packet_to_local() > rxrpc_find_connection_rcu() - uses RCU > rxrpc_lookup_peer_rcu() - uses RCU > rxrpc_find_service_conn_rcu() - uses RCU > idr_find() - uses RCU * CONNECTION-LEVEL PROCESSING - Service upgrade - Can only happen once per conn ! Changed to use cmpxchg > rxrpc_post_packet_to_conn() - Setting conn->hi_serial - Probably safe not using locks - Maybe use cmpxchg * CALL-LEVEL PROCESSING > Old-call checking > rxrpc_input_implicit_end_call() > rxrpc_call_completed() > rxrpc_queue_call() ! Need to take rx->incoming_lock > __rxrpc_disconnect_call() > rxrpc_notify_socket() > rxrpc_new_incoming_call() - Uses rx->incoming_lock for the entire process - Might be able to drop this earlier in favour of the call lock > rxrpc_incoming_call() ! Conflicts with rxrpc_input_implicit_end_call() > rxrpc_send_ping() - Don't need locks to check rtt state > rxrpc_propose_ACK * PACKET DISTRIBUTION > rxrpc_input_call_packet() > rxrpc_input_data() * QUEUE DATA PACKET ON CALL > rxrpc_reduce_call_timer() - Uses timer_reduce() ! Needs call->input_lock() > rxrpc_receiving_reply() ! Needs locking around ack state > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_proto_abort() > rxrpc_input_dup_data() - Fills the Rx buffer - rxrpc_propose_ACK() - rxrpc_notify_socket() > rxrpc_input_ack() * APPLY ACK PACKET TO CALL AND DISCARD PACKET > rxrpc_input_ping_response() - Probably doesn't need any extra locking ! Need READ_ONCE() on call->ping_serial > rxrpc_input_check_for_lost_ack() - Takes call->lock to consult Tx buffer > rxrpc_peer_add_rtt() ! Needs to take a lock (peer->rtt_input_lock) ! Could perhaps manage with cmpxchg() and xadd() instead > rxrpc_input_requested_ack - Consults Tx buffer ! Probably needs a lock > rxrpc_peer_add_rtt() > rxrpc_propose_ack() > rxrpc_input_ackinfo() - Changes call->tx_winsize ! Use cmpxchg to handle change ! Should perhaps track serial number - Uses peer->lock to record MTU specification changes > rxrpc_proto_abort() ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_input_soft_acks() - Consults the Tx buffer > rxrpc_congestion_management() - Modifies the Tx annotations ! Needs call->input_lock() > rxrpc_queue_call() > rxrpc_input_abort() * APPLY ABORT PACKET TO CALL AND DISCARD PACKET > rxrpc_set_call_completion() > rxrpc_notify_socket() > rxrpc_input_ackall() * APPLY ACKALL PACKET TO CALL AND DISCARD PACKET ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_reject_packet() There are some functions used by the above that queue the packet, after which the procedure is terminated: - rxrpc_post_packet_to_local() - local->event_queue is an sk_buff_head - local->processor is a work_struct - rxrpc_post_packet_to_conn() - conn->rx_queue is an sk_buff_head - conn->processor is a work_struct - rxrpc_reject_packet() - local->reject_queue is an sk_buff_head - local->processor is a work_struct And some that offload processing to process context: - rxrpc_notify_socket() - Uses RCU lock - Uses call->notify_lock to call call->notify_rx - Uses call->recvmsg_lock to queue recvmsg side - rxrpc_queue_call() - call->processor is a work_struct - rxrpc_propose_ACK() - Uses call->lock to wrap __rxrpc_propose_ACK() And a bunch that complete a call, all of which use call->state_lock to protect the call state: - rxrpc_call_completed() - rxrpc_set_call_completion() - rxrpc_abort_call() - rxrpc_proto_abort() - Also uses rxrpc_queue_call() Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both") Signed-off-by: David Howells <dhowells@redhat.com>
2018-10-08 22:46:25 +08:00
trace_rxrpc_improper_term(call);
break;
}
rxrpc: Fix the packet reception routine The rxrpc_input_packet() function and its call tree was built around the assumption that data_ready() handler called from UDP to inform a kernel service that there is data to be had was non-reentrant. This means that certain locking could be dispensed with. This, however, turns out not to be the case with a multi-queue network card that can deliver packets to multiple cpus simultaneously. Each of those cpus can be in the rxrpc_input_packet() function at the same time. Fix by adding or changing some structure members: (1) Add peer->rtt_input_lock to serialise access to the RTT buffer. (2) Make conn->service_id into a 32-bit variable so that it can be cmpxchg'd on all arches. (3) Add call->input_lock to serialise access to the Rx/Tx state. Note that although the Rx and Tx states are (almost) entirely separate, there's no point completing the separation and having separate locks since it's a bi-phasal RPC protocol rather than a bi-direction streaming protocol. Data transmission and data reception do not take place simultaneously on any particular call. and making the following functional changes: (1) In rxrpc_input_data(), hold call->input_lock around the core to prevent simultaneous producing of packets into the Rx ring and updating of tracking state for a particular call. (2) In rxrpc_input_ping_response(), only read call->ping_serial once, and check it before checking RXRPC_CALL_PINGING as that's a cheaper test. The bit test and bit clear can then be combined. No further locking is needed here. (3) In rxrpc_input_ack(), take call->input_lock after we've parsed much of the ACK packet. The superseded ACK check is then done both before and after the lock is taken. The handing of ackinfo data is split, parsing before the lock is taken and processing with it held. This is keyed on rxMTU being non-zero. Congestion management is also done within the locked section. (4) In rxrpc_input_ackall(), take call->input_lock around the Tx window rotation. The ACKALL packet carries no information and is only really useful after all packets have been transmitted since it's imprecise. (5) In rxrpc_input_implicit_end_call(), we use rx->incoming_lock to prevent calls being simultaneously implicitly ended on two cpus and also to prevent any races with incoming call setup. (6) In rxrpc_input_packet(), use cmpxchg() to effect the service upgrade on a connection. It is only permitted to happen once for a connection. (7) In rxrpc_new_incoming_call(), we have to recheck the routing inside rx->incoming_lock to see if someone else set up the call, connection or peer whilst we were getting there. We can't trust the values from the earlier routing check unless we pin refs on them - which we want to avoid. Further, we need to allow for an incoming call to have its state changed on another CPU between us making it live and us adjusting it because the conn is now in the RXRPC_CONN_SERVICE state. (8) In rxrpc_peer_add_rtt(), take peer->rtt_input_lock around the access to the RTT buffer. Don't need to lock around setting peer->rtt. For reference, the inventory of state-accessing or state-altering functions used by the packet input procedure is: > rxrpc_input_packet() * PACKET CHECKING * ROUTING > rxrpc_post_packet_to_local() > rxrpc_find_connection_rcu() - uses RCU > rxrpc_lookup_peer_rcu() - uses RCU > rxrpc_find_service_conn_rcu() - uses RCU > idr_find() - uses RCU * CONNECTION-LEVEL PROCESSING - Service upgrade - Can only happen once per conn ! Changed to use cmpxchg > rxrpc_post_packet_to_conn() - Setting conn->hi_serial - Probably safe not using locks - Maybe use cmpxchg * CALL-LEVEL PROCESSING > Old-call checking > rxrpc_input_implicit_end_call() > rxrpc_call_completed() > rxrpc_queue_call() ! Need to take rx->incoming_lock > __rxrpc_disconnect_call() > rxrpc_notify_socket() > rxrpc_new_incoming_call() - Uses rx->incoming_lock for the entire process - Might be able to drop this earlier in favour of the call lock > rxrpc_incoming_call() ! Conflicts with rxrpc_input_implicit_end_call() > rxrpc_send_ping() - Don't need locks to check rtt state > rxrpc_propose_ACK * PACKET DISTRIBUTION > rxrpc_input_call_packet() > rxrpc_input_data() * QUEUE DATA PACKET ON CALL > rxrpc_reduce_call_timer() - Uses timer_reduce() ! Needs call->input_lock() > rxrpc_receiving_reply() ! Needs locking around ack state > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_proto_abort() > rxrpc_input_dup_data() - Fills the Rx buffer - rxrpc_propose_ACK() - rxrpc_notify_socket() > rxrpc_input_ack() * APPLY ACK PACKET TO CALL AND DISCARD PACKET > rxrpc_input_ping_response() - Probably doesn't need any extra locking ! Need READ_ONCE() on call->ping_serial > rxrpc_input_check_for_lost_ack() - Takes call->lock to consult Tx buffer > rxrpc_peer_add_rtt() ! Needs to take a lock (peer->rtt_input_lock) ! Could perhaps manage with cmpxchg() and xadd() instead > rxrpc_input_requested_ack - Consults Tx buffer ! Probably needs a lock > rxrpc_peer_add_rtt() > rxrpc_propose_ack() > rxrpc_input_ackinfo() - Changes call->tx_winsize ! Use cmpxchg to handle change ! Should perhaps track serial number - Uses peer->lock to record MTU specification changes > rxrpc_proto_abort() ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_input_soft_acks() - Consults the Tx buffer > rxrpc_congestion_management() - Modifies the Tx annotations ! Needs call->input_lock() > rxrpc_queue_call() > rxrpc_input_abort() * APPLY ABORT PACKET TO CALL AND DISCARD PACKET > rxrpc_set_call_completion() > rxrpc_notify_socket() > rxrpc_input_ackall() * APPLY ACKALL PACKET TO CALL AND DISCARD PACKET ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_reject_packet() There are some functions used by the above that queue the packet, after which the procedure is terminated: - rxrpc_post_packet_to_local() - local->event_queue is an sk_buff_head - local->processor is a work_struct - rxrpc_post_packet_to_conn() - conn->rx_queue is an sk_buff_head - conn->processor is a work_struct - rxrpc_reject_packet() - local->reject_queue is an sk_buff_head - local->processor is a work_struct And some that offload processing to process context: - rxrpc_notify_socket() - Uses RCU lock - Uses call->notify_lock to call call->notify_rx - Uses call->recvmsg_lock to queue recvmsg side - rxrpc_queue_call() - call->processor is a work_struct - rxrpc_propose_ACK() - Uses call->lock to wrap __rxrpc_propose_ACK() And a bunch that complete a call, all of which use call->state_lock to protect the call state: - rxrpc_call_completed() - rxrpc_set_call_completion() - rxrpc_abort_call() - rxrpc_proto_abort() - Also uses rxrpc_queue_call() Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both") Signed-off-by: David Howells <dhowells@redhat.com>
2018-10-08 22:46:25 +08:00
spin_lock(&rx->incoming_lock);
__rxrpc_disconnect_call(conn, call);
rxrpc: Fix the packet reception routine The rxrpc_input_packet() function and its call tree was built around the assumption that data_ready() handler called from UDP to inform a kernel service that there is data to be had was non-reentrant. This means that certain locking could be dispensed with. This, however, turns out not to be the case with a multi-queue network card that can deliver packets to multiple cpus simultaneously. Each of those cpus can be in the rxrpc_input_packet() function at the same time. Fix by adding or changing some structure members: (1) Add peer->rtt_input_lock to serialise access to the RTT buffer. (2) Make conn->service_id into a 32-bit variable so that it can be cmpxchg'd on all arches. (3) Add call->input_lock to serialise access to the Rx/Tx state. Note that although the Rx and Tx states are (almost) entirely separate, there's no point completing the separation and having separate locks since it's a bi-phasal RPC protocol rather than a bi-direction streaming protocol. Data transmission and data reception do not take place simultaneously on any particular call. and making the following functional changes: (1) In rxrpc_input_data(), hold call->input_lock around the core to prevent simultaneous producing of packets into the Rx ring and updating of tracking state for a particular call. (2) In rxrpc_input_ping_response(), only read call->ping_serial once, and check it before checking RXRPC_CALL_PINGING as that's a cheaper test. The bit test and bit clear can then be combined. No further locking is needed here. (3) In rxrpc_input_ack(), take call->input_lock after we've parsed much of the ACK packet. The superseded ACK check is then done both before and after the lock is taken. The handing of ackinfo data is split, parsing before the lock is taken and processing with it held. This is keyed on rxMTU being non-zero. Congestion management is also done within the locked section. (4) In rxrpc_input_ackall(), take call->input_lock around the Tx window rotation. The ACKALL packet carries no information and is only really useful after all packets have been transmitted since it's imprecise. (5) In rxrpc_input_implicit_end_call(), we use rx->incoming_lock to prevent calls being simultaneously implicitly ended on two cpus and also to prevent any races with incoming call setup. (6) In rxrpc_input_packet(), use cmpxchg() to effect the service upgrade on a connection. It is only permitted to happen once for a connection. (7) In rxrpc_new_incoming_call(), we have to recheck the routing inside rx->incoming_lock to see if someone else set up the call, connection or peer whilst we were getting there. We can't trust the values from the earlier routing check unless we pin refs on them - which we want to avoid. Further, we need to allow for an incoming call to have its state changed on another CPU between us making it live and us adjusting it because the conn is now in the RXRPC_CONN_SERVICE state. (8) In rxrpc_peer_add_rtt(), take peer->rtt_input_lock around the access to the RTT buffer. Don't need to lock around setting peer->rtt. For reference, the inventory of state-accessing or state-altering functions used by the packet input procedure is: > rxrpc_input_packet() * PACKET CHECKING * ROUTING > rxrpc_post_packet_to_local() > rxrpc_find_connection_rcu() - uses RCU > rxrpc_lookup_peer_rcu() - uses RCU > rxrpc_find_service_conn_rcu() - uses RCU > idr_find() - uses RCU * CONNECTION-LEVEL PROCESSING - Service upgrade - Can only happen once per conn ! Changed to use cmpxchg > rxrpc_post_packet_to_conn() - Setting conn->hi_serial - Probably safe not using locks - Maybe use cmpxchg * CALL-LEVEL PROCESSING > Old-call checking > rxrpc_input_implicit_end_call() > rxrpc_call_completed() > rxrpc_queue_call() ! Need to take rx->incoming_lock > __rxrpc_disconnect_call() > rxrpc_notify_socket() > rxrpc_new_incoming_call() - Uses rx->incoming_lock for the entire process - Might be able to drop this earlier in favour of the call lock > rxrpc_incoming_call() ! Conflicts with rxrpc_input_implicit_end_call() > rxrpc_send_ping() - Don't need locks to check rtt state > rxrpc_propose_ACK * PACKET DISTRIBUTION > rxrpc_input_call_packet() > rxrpc_input_data() * QUEUE DATA PACKET ON CALL > rxrpc_reduce_call_timer() - Uses timer_reduce() ! Needs call->input_lock() > rxrpc_receiving_reply() ! Needs locking around ack state > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_proto_abort() > rxrpc_input_dup_data() - Fills the Rx buffer - rxrpc_propose_ACK() - rxrpc_notify_socket() > rxrpc_input_ack() * APPLY ACK PACKET TO CALL AND DISCARD PACKET > rxrpc_input_ping_response() - Probably doesn't need any extra locking ! Need READ_ONCE() on call->ping_serial > rxrpc_input_check_for_lost_ack() - Takes call->lock to consult Tx buffer > rxrpc_peer_add_rtt() ! Needs to take a lock (peer->rtt_input_lock) ! Could perhaps manage with cmpxchg() and xadd() instead > rxrpc_input_requested_ack - Consults Tx buffer ! Probably needs a lock > rxrpc_peer_add_rtt() > rxrpc_propose_ack() > rxrpc_input_ackinfo() - Changes call->tx_winsize ! Use cmpxchg to handle change ! Should perhaps track serial number - Uses peer->lock to record MTU specification changes > rxrpc_proto_abort() ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_input_soft_acks() - Consults the Tx buffer > rxrpc_congestion_management() - Modifies the Tx annotations ! Needs call->input_lock() > rxrpc_queue_call() > rxrpc_input_abort() * APPLY ABORT PACKET TO CALL AND DISCARD PACKET > rxrpc_set_call_completion() > rxrpc_notify_socket() > rxrpc_input_ackall() * APPLY ACKALL PACKET TO CALL AND DISCARD PACKET ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_reject_packet() There are some functions used by the above that queue the packet, after which the procedure is terminated: - rxrpc_post_packet_to_local() - local->event_queue is an sk_buff_head - local->processor is a work_struct - rxrpc_post_packet_to_conn() - conn->rx_queue is an sk_buff_head - conn->processor is a work_struct - rxrpc_reject_packet() - local->reject_queue is an sk_buff_head - local->processor is a work_struct And some that offload processing to process context: - rxrpc_notify_socket() - Uses RCU lock - Uses call->notify_lock to call call->notify_rx - Uses call->recvmsg_lock to queue recvmsg side - rxrpc_queue_call() - call->processor is a work_struct - rxrpc_propose_ACK() - Uses call->lock to wrap __rxrpc_propose_ACK() And a bunch that complete a call, all of which use call->state_lock to protect the call state: - rxrpc_call_completed() - rxrpc_set_call_completion() - rxrpc_abort_call() - rxrpc_proto_abort() - Also uses rxrpc_queue_call() Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both") Signed-off-by: David Howells <dhowells@redhat.com>
2018-10-08 22:46:25 +08:00
spin_unlock(&rx->incoming_lock);
rxrpc_notify_socket(call);
}
/*
* post connection-level events to the connection
* - this includes challenges, responses, some aborts and call terminal packet
* retransmission.
*/
static void rxrpc_post_packet_to_conn(struct rxrpc_connection *conn,
struct sk_buff *skb)
{
_enter("%p,%p", conn, skb);
skb_queue_tail(&conn->rx_queue, skb);
rxrpc_queue_conn(conn);
}
/*
* post endpoint-level events to the local endpoint
* - this includes debug and version messages
*/
static void rxrpc_post_packet_to_local(struct rxrpc_local *local,
struct sk_buff *skb)
{
_enter("%p,%p", local, skb);
if (rxrpc_get_local_maybe(local)) {
skb_queue_tail(&local->event_queue, skb);
rxrpc_queue_local(local);
} else {
rxrpc_free_skb(skb, rxrpc_skb_freed);
}
}
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
/*
* put a packet up for transport-level abort
*/
static void rxrpc_reject_packet(struct rxrpc_local *local, struct sk_buff *skb)
{
CHECK_SLAB_OKAY(&local->usage);
if (rxrpc_get_local_maybe(local)) {
skb_queue_tail(&local->reject_queue, skb);
rxrpc_queue_local(local);
} else {
rxrpc_free_skb(skb, rxrpc_skb_freed);
}
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
}
/*
* Extract the wire header from a packet and translate the byte order.
*/
static noinline
int rxrpc_extract_header(struct rxrpc_skb_priv *sp, struct sk_buff *skb)
{
struct rxrpc_wire_header whdr;
/* dig out the RxRPC connection details */
if (skb_copy_bits(skb, 0, &whdr, sizeof(whdr)) < 0) {
trace_rxrpc_rx_eproto(NULL, sp->hdr.serial,
tracepoint_string("bad_hdr"));
return -EBADMSG;
}
memset(sp, 0, sizeof(*sp));
sp->hdr.epoch = ntohl(whdr.epoch);
sp->hdr.cid = ntohl(whdr.cid);
sp->hdr.callNumber = ntohl(whdr.callNumber);
sp->hdr.seq = ntohl(whdr.seq);
sp->hdr.serial = ntohl(whdr.serial);
sp->hdr.flags = whdr.flags;
sp->hdr.type = whdr.type;
sp->hdr.userStatus = whdr.userStatus;
sp->hdr.securityIndex = whdr.securityIndex;
sp->hdr._rsvd = ntohs(whdr._rsvd);
sp->hdr.serviceId = ntohs(whdr.serviceId);
return 0;
}
/*
* handle data received on the local endpoint
* - may be called in interrupt context
rxrpc: Rework local endpoint management Rework the local RxRPC endpoint management. Local endpoint objects are maintained in a flat list as before. This should be okay as there shouldn't be more than one per open AF_RXRPC socket (there can be fewer as local endpoints can be shared if their local service ID is 0 and they share the same local transport parameters). Changes: (1) Local endpoints may now only be shared if they have local service ID 0 (ie. they're not being used for listening). This prevents a scenario where process A is listening of the Cache Manager port and process B contacts a fileserver - which may then attempt to send CM requests back to B. But if A and B are sharing a local endpoint, A will get the CM requests meant for B. (2) We use a mutex to handle lookups and don't provide RCU-only lookups since we only expect to access the list when opening a socket or destroying an endpoint. The local endpoint object is pointed to by the transport socket's sk_user_data for the life of the transport socket - allowing us to refer to it directly from the sk_data_ready and sk_error_report callbacks. (3) atomic_inc_not_zero() now exists and can be used to only share a local endpoint if the last reference hasn't yet gone. (4) We can remove rxrpc_local_lock - a spinlock that had to be taken with BH processing disabled given that we assume sk_user_data won't change under us. (5) The transport socket is shut down before we clear the sk_user_data pointer so that we can be sure that the transport socket's callbacks won't be invoked once the RCU destruction is scheduled. (6) Local endpoints have a work item that handles both destruction and event processing. The means that destruction doesn't then need to wait for event processing. The event queues can then be cleared after the transport socket is shut down. (7) Local endpoints are no longer available for resurrection beyond the life of the sockets that had them open. As soon as their last ref goes, they are scheduled for destruction and may not have their usage count moved from 0. Signed-off-by: David Howells <dhowells@redhat.com>
2016-04-04 21:00:35 +08:00
*
rxrpc: fix race condition in rxrpc_input_packet() After commit 5271953cad31 ("rxrpc: Use the UDP encap_rcv hook"), rxrpc_input_packet() is directly called from lockless UDP receive path, under rcu_read_lock() protection. It must therefore use RCU rules : - udp_sk->sk_user_data can be cleared at any point in this function. rcu_dereference_sk_user_data() is what we need here. - Also, since sk_user_data might have been set in rxrpc_open_socket() we must observe a proper RCU grace period before kfree(local) in rxrpc_lookup_local() v4: @local can be NULL in xrpc_lookup_local() as reported by kbuild test robot <lkp@intel.com> and Julia Lawall <julia.lawall@lip6.fr>, thanks ! v3,v2 : addressed David Howells feedback, thanks ! syzbot reported : kasan: CONFIG_KASAN_INLINE enabled kasan: GPF could be caused by NULL-ptr deref or user memory access general protection fault: 0000 [#1] PREEMPT SMP KASAN CPU: 0 PID: 19236 Comm: syz-executor703 Not tainted 5.1.0-rc6 #79 Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011 RIP: 0010:__lock_acquire+0xbef/0x3fb0 kernel/locking/lockdep.c:3573 Code: 00 0f 85 a5 1f 00 00 48 81 c4 10 01 00 00 5b 41 5c 41 5d 41 5e 41 5f 5d c3 48 b8 00 00 00 00 00 fc ff df 4c 89 ea 48 c1 ea 03 <80> 3c 02 00 0f 85 4a 21 00 00 49 81 7d 00 20 54 9c 89 0f 84 cf f4 RSP: 0018:ffff88809d7aef58 EFLAGS: 00010002 RAX: dffffc0000000000 RBX: 0000000000000000 RCX: 0000000000000000 RDX: 0000000000000026 RSI: 0000000000000000 RDI: 0000000000000001 RBP: ffff88809d7af090 R08: 0000000000000001 R09: 0000000000000001 R10: ffffed1015d05bc7 R11: ffff888089428600 R12: 0000000000000000 R13: 0000000000000130 R14: 0000000000000001 R15: 0000000000000001 FS: 00007f059044d700(0000) GS:ffff8880ae800000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 00000000004b6040 CR3: 00000000955ca000 CR4: 00000000001406f0 Call Trace: lock_acquire+0x16f/0x3f0 kernel/locking/lockdep.c:4211 __raw_spin_lock_irqsave include/linux/spinlock_api_smp.h:110 [inline] _raw_spin_lock_irqsave+0x95/0xcd kernel/locking/spinlock.c:152 skb_queue_tail+0x26/0x150 net/core/skbuff.c:2972 rxrpc_reject_packet net/rxrpc/input.c:1126 [inline] rxrpc_input_packet+0x4a0/0x5536 net/rxrpc/input.c:1414 udp_queue_rcv_one_skb+0xaf2/0x1780 net/ipv4/udp.c:2011 udp_queue_rcv_skb+0x128/0x730 net/ipv4/udp.c:2085 udp_unicast_rcv_skb.isra.0+0xb9/0x360 net/ipv4/udp.c:2245 __udp4_lib_rcv+0x701/0x2ca0 net/ipv4/udp.c:2301 udp_rcv+0x22/0x30 net/ipv4/udp.c:2482 ip_protocol_deliver_rcu+0x60/0x8f0 net/ipv4/ip_input.c:208 ip_local_deliver_finish+0x23b/0x390 net/ipv4/ip_input.c:234 NF_HOOK include/linux/netfilter.h:289 [inline] NF_HOOK include/linux/netfilter.h:283 [inline] ip_local_deliver+0x1e9/0x520 net/ipv4/ip_input.c:255 dst_input include/net/dst.h:450 [inline] ip_rcv_finish+0x1e1/0x300 net/ipv4/ip_input.c:413 NF_HOOK include/linux/netfilter.h:289 [inline] NF_HOOK include/linux/netfilter.h:283 [inline] ip_rcv+0xe8/0x3f0 net/ipv4/ip_input.c:523 __netif_receive_skb_one_core+0x115/0x1a0 net/core/dev.c:4987 __netif_receive_skb+0x2c/0x1c0 net/core/dev.c:5099 netif_receive_skb_internal+0x117/0x660 net/core/dev.c:5202 napi_frags_finish net/core/dev.c:5769 [inline] napi_gro_frags+0xade/0xd10 net/core/dev.c:5843 tun_get_user+0x2f24/0x3fb0 drivers/net/tun.c:1981 tun_chr_write_iter+0xbd/0x156 drivers/net/tun.c:2027 call_write_iter include/linux/fs.h:1866 [inline] do_iter_readv_writev+0x5e1/0x8e0 fs/read_write.c:681 do_iter_write fs/read_write.c:957 [inline] do_iter_write+0x184/0x610 fs/read_write.c:938 vfs_writev+0x1b3/0x2f0 fs/read_write.c:1002 do_writev+0x15e/0x370 fs/read_write.c:1037 __do_sys_writev fs/read_write.c:1110 [inline] __se_sys_writev fs/read_write.c:1107 [inline] __x64_sys_writev+0x75/0xb0 fs/read_write.c:1107 do_syscall_64+0x103/0x610 arch/x86/entry/common.c:290 entry_SYSCALL_64_after_hwframe+0x49/0xbe Fixes: 5271953cad31 ("rxrpc: Use the UDP encap_rcv hook") Signed-off-by: Eric Dumazet <edumazet@google.com> Reported-by: syzbot <syzkaller@googlegroups.com> Acked-by: David Howells <dhowells@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-04-25 00:44:11 +08:00
* [!] Note that as this is called from the encap_rcv hook, the socket is not
* held locked by the caller and nothing prevents sk_user_data on the UDP from
* being cleared in the middle of processing this function.
*
* Called with the RCU read lock held from the IP layer via UDP.
*/
int rxrpc_input_packet(struct sock *udp_sk, struct sk_buff *skb)
{
rxrpc: fix race condition in rxrpc_input_packet() After commit 5271953cad31 ("rxrpc: Use the UDP encap_rcv hook"), rxrpc_input_packet() is directly called from lockless UDP receive path, under rcu_read_lock() protection. It must therefore use RCU rules : - udp_sk->sk_user_data can be cleared at any point in this function. rcu_dereference_sk_user_data() is what we need here. - Also, since sk_user_data might have been set in rxrpc_open_socket() we must observe a proper RCU grace period before kfree(local) in rxrpc_lookup_local() v4: @local can be NULL in xrpc_lookup_local() as reported by kbuild test robot <lkp@intel.com> and Julia Lawall <julia.lawall@lip6.fr>, thanks ! v3,v2 : addressed David Howells feedback, thanks ! syzbot reported : kasan: CONFIG_KASAN_INLINE enabled kasan: GPF could be caused by NULL-ptr deref or user memory access general protection fault: 0000 [#1] PREEMPT SMP KASAN CPU: 0 PID: 19236 Comm: syz-executor703 Not tainted 5.1.0-rc6 #79 Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011 RIP: 0010:__lock_acquire+0xbef/0x3fb0 kernel/locking/lockdep.c:3573 Code: 00 0f 85 a5 1f 00 00 48 81 c4 10 01 00 00 5b 41 5c 41 5d 41 5e 41 5f 5d c3 48 b8 00 00 00 00 00 fc ff df 4c 89 ea 48 c1 ea 03 <80> 3c 02 00 0f 85 4a 21 00 00 49 81 7d 00 20 54 9c 89 0f 84 cf f4 RSP: 0018:ffff88809d7aef58 EFLAGS: 00010002 RAX: dffffc0000000000 RBX: 0000000000000000 RCX: 0000000000000000 RDX: 0000000000000026 RSI: 0000000000000000 RDI: 0000000000000001 RBP: ffff88809d7af090 R08: 0000000000000001 R09: 0000000000000001 R10: ffffed1015d05bc7 R11: ffff888089428600 R12: 0000000000000000 R13: 0000000000000130 R14: 0000000000000001 R15: 0000000000000001 FS: 00007f059044d700(0000) GS:ffff8880ae800000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 00000000004b6040 CR3: 00000000955ca000 CR4: 00000000001406f0 Call Trace: lock_acquire+0x16f/0x3f0 kernel/locking/lockdep.c:4211 __raw_spin_lock_irqsave include/linux/spinlock_api_smp.h:110 [inline] _raw_spin_lock_irqsave+0x95/0xcd kernel/locking/spinlock.c:152 skb_queue_tail+0x26/0x150 net/core/skbuff.c:2972 rxrpc_reject_packet net/rxrpc/input.c:1126 [inline] rxrpc_input_packet+0x4a0/0x5536 net/rxrpc/input.c:1414 udp_queue_rcv_one_skb+0xaf2/0x1780 net/ipv4/udp.c:2011 udp_queue_rcv_skb+0x128/0x730 net/ipv4/udp.c:2085 udp_unicast_rcv_skb.isra.0+0xb9/0x360 net/ipv4/udp.c:2245 __udp4_lib_rcv+0x701/0x2ca0 net/ipv4/udp.c:2301 udp_rcv+0x22/0x30 net/ipv4/udp.c:2482 ip_protocol_deliver_rcu+0x60/0x8f0 net/ipv4/ip_input.c:208 ip_local_deliver_finish+0x23b/0x390 net/ipv4/ip_input.c:234 NF_HOOK include/linux/netfilter.h:289 [inline] NF_HOOK include/linux/netfilter.h:283 [inline] ip_local_deliver+0x1e9/0x520 net/ipv4/ip_input.c:255 dst_input include/net/dst.h:450 [inline] ip_rcv_finish+0x1e1/0x300 net/ipv4/ip_input.c:413 NF_HOOK include/linux/netfilter.h:289 [inline] NF_HOOK include/linux/netfilter.h:283 [inline] ip_rcv+0xe8/0x3f0 net/ipv4/ip_input.c:523 __netif_receive_skb_one_core+0x115/0x1a0 net/core/dev.c:4987 __netif_receive_skb+0x2c/0x1c0 net/core/dev.c:5099 netif_receive_skb_internal+0x117/0x660 net/core/dev.c:5202 napi_frags_finish net/core/dev.c:5769 [inline] napi_gro_frags+0xade/0xd10 net/core/dev.c:5843 tun_get_user+0x2f24/0x3fb0 drivers/net/tun.c:1981 tun_chr_write_iter+0xbd/0x156 drivers/net/tun.c:2027 call_write_iter include/linux/fs.h:1866 [inline] do_iter_readv_writev+0x5e1/0x8e0 fs/read_write.c:681 do_iter_write fs/read_write.c:957 [inline] do_iter_write+0x184/0x610 fs/read_write.c:938 vfs_writev+0x1b3/0x2f0 fs/read_write.c:1002 do_writev+0x15e/0x370 fs/read_write.c:1037 __do_sys_writev fs/read_write.c:1110 [inline] __se_sys_writev fs/read_write.c:1107 [inline] __x64_sys_writev+0x75/0xb0 fs/read_write.c:1107 do_syscall_64+0x103/0x610 arch/x86/entry/common.c:290 entry_SYSCALL_64_after_hwframe+0x49/0xbe Fixes: 5271953cad31 ("rxrpc: Use the UDP encap_rcv hook") Signed-off-by: Eric Dumazet <edumazet@google.com> Reported-by: syzbot <syzkaller@googlegroups.com> Acked-by: David Howells <dhowells@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-04-25 00:44:11 +08:00
struct rxrpc_local *local = rcu_dereference_sk_user_data(udp_sk);
struct rxrpc_connection *conn;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
struct rxrpc_channel *chan;
struct rxrpc_call *call = NULL;
struct rxrpc_skb_priv *sp;
rxrpc: Make service call handling more robust Make the following changes to improve the robustness of the code that sets up a new service call: (1) Cache the rxrpc_sock struct obtained in rxrpc_data_ready() to do a service ID check and pass that along to rxrpc_new_incoming_call(). This means that I can remove the check from rxrpc_new_incoming_call() without the need to worry about the socket attached to the local endpoint getting replaced - which would invalidate the check. (2) Cache the rxrpc_peer struct, thereby allowing the peer search to be done once. The peer is passed to rxrpc_new_incoming_call(), thereby saving the need to repeat the search. This also reduces the possibility of rxrpc_publish_service_conn() BUG()'ing due to the detection of a duplicate connection, despite the initial search done by rxrpc_find_connection_rcu() having turned up nothing. This BUG() shouldn't ever get hit since rxrpc_data_ready() *should* be non-reentrant and the result of the initial search should still hold true, but it has proven possible to hit. I *think* this may be due to __rxrpc_lookup_peer_rcu() cutting short the iteration over the hash table if it finds a matching peer with a zero usage count, but I don't know for sure since it's only ever been hit once that I know of. Another possibility is that a bug in rxrpc_data_ready() that checked the wrong byte in the header for the RXRPC_CLIENT_INITIATED flag might've let through a packet that caused a spurious and invalid call to be set up. That is addressed in another patch. (3) Fix __rxrpc_lookup_peer_rcu() to skip peer records that have a zero usage count rather than stopping and returning not found, just in case there's another peer record behind it in the bucket. (4) Don't search the peer records in rxrpc_alloc_incoming_call(), but rather either use the peer cached in (2) or, if one wasn't found, preemptively install a new one. Fixes: 8496af50eb38 ("rxrpc: Use RCU to access a peer's service connection tree") Signed-off-by: David Howells <dhowells@redhat.com>
2018-09-27 22:13:09 +08:00
struct rxrpc_peer *peer = NULL;
struct rxrpc_sock *rx = NULL;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
unsigned int channel;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
_enter("%p", udp_sk);
rxrpc: fix race condition in rxrpc_input_packet() After commit 5271953cad31 ("rxrpc: Use the UDP encap_rcv hook"), rxrpc_input_packet() is directly called from lockless UDP receive path, under rcu_read_lock() protection. It must therefore use RCU rules : - udp_sk->sk_user_data can be cleared at any point in this function. rcu_dereference_sk_user_data() is what we need here. - Also, since sk_user_data might have been set in rxrpc_open_socket() we must observe a proper RCU grace period before kfree(local) in rxrpc_lookup_local() v4: @local can be NULL in xrpc_lookup_local() as reported by kbuild test robot <lkp@intel.com> and Julia Lawall <julia.lawall@lip6.fr>, thanks ! v3,v2 : addressed David Howells feedback, thanks ! syzbot reported : kasan: CONFIG_KASAN_INLINE enabled kasan: GPF could be caused by NULL-ptr deref or user memory access general protection fault: 0000 [#1] PREEMPT SMP KASAN CPU: 0 PID: 19236 Comm: syz-executor703 Not tainted 5.1.0-rc6 #79 Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011 RIP: 0010:__lock_acquire+0xbef/0x3fb0 kernel/locking/lockdep.c:3573 Code: 00 0f 85 a5 1f 00 00 48 81 c4 10 01 00 00 5b 41 5c 41 5d 41 5e 41 5f 5d c3 48 b8 00 00 00 00 00 fc ff df 4c 89 ea 48 c1 ea 03 <80> 3c 02 00 0f 85 4a 21 00 00 49 81 7d 00 20 54 9c 89 0f 84 cf f4 RSP: 0018:ffff88809d7aef58 EFLAGS: 00010002 RAX: dffffc0000000000 RBX: 0000000000000000 RCX: 0000000000000000 RDX: 0000000000000026 RSI: 0000000000000000 RDI: 0000000000000001 RBP: ffff88809d7af090 R08: 0000000000000001 R09: 0000000000000001 R10: ffffed1015d05bc7 R11: ffff888089428600 R12: 0000000000000000 R13: 0000000000000130 R14: 0000000000000001 R15: 0000000000000001 FS: 00007f059044d700(0000) GS:ffff8880ae800000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 00000000004b6040 CR3: 00000000955ca000 CR4: 00000000001406f0 Call Trace: lock_acquire+0x16f/0x3f0 kernel/locking/lockdep.c:4211 __raw_spin_lock_irqsave include/linux/spinlock_api_smp.h:110 [inline] _raw_spin_lock_irqsave+0x95/0xcd kernel/locking/spinlock.c:152 skb_queue_tail+0x26/0x150 net/core/skbuff.c:2972 rxrpc_reject_packet net/rxrpc/input.c:1126 [inline] rxrpc_input_packet+0x4a0/0x5536 net/rxrpc/input.c:1414 udp_queue_rcv_one_skb+0xaf2/0x1780 net/ipv4/udp.c:2011 udp_queue_rcv_skb+0x128/0x730 net/ipv4/udp.c:2085 udp_unicast_rcv_skb.isra.0+0xb9/0x360 net/ipv4/udp.c:2245 __udp4_lib_rcv+0x701/0x2ca0 net/ipv4/udp.c:2301 udp_rcv+0x22/0x30 net/ipv4/udp.c:2482 ip_protocol_deliver_rcu+0x60/0x8f0 net/ipv4/ip_input.c:208 ip_local_deliver_finish+0x23b/0x390 net/ipv4/ip_input.c:234 NF_HOOK include/linux/netfilter.h:289 [inline] NF_HOOK include/linux/netfilter.h:283 [inline] ip_local_deliver+0x1e9/0x520 net/ipv4/ip_input.c:255 dst_input include/net/dst.h:450 [inline] ip_rcv_finish+0x1e1/0x300 net/ipv4/ip_input.c:413 NF_HOOK include/linux/netfilter.h:289 [inline] NF_HOOK include/linux/netfilter.h:283 [inline] ip_rcv+0xe8/0x3f0 net/ipv4/ip_input.c:523 __netif_receive_skb_one_core+0x115/0x1a0 net/core/dev.c:4987 __netif_receive_skb+0x2c/0x1c0 net/core/dev.c:5099 netif_receive_skb_internal+0x117/0x660 net/core/dev.c:5202 napi_frags_finish net/core/dev.c:5769 [inline] napi_gro_frags+0xade/0xd10 net/core/dev.c:5843 tun_get_user+0x2f24/0x3fb0 drivers/net/tun.c:1981 tun_chr_write_iter+0xbd/0x156 drivers/net/tun.c:2027 call_write_iter include/linux/fs.h:1866 [inline] do_iter_readv_writev+0x5e1/0x8e0 fs/read_write.c:681 do_iter_write fs/read_write.c:957 [inline] do_iter_write+0x184/0x610 fs/read_write.c:938 vfs_writev+0x1b3/0x2f0 fs/read_write.c:1002 do_writev+0x15e/0x370 fs/read_write.c:1037 __do_sys_writev fs/read_write.c:1110 [inline] __se_sys_writev fs/read_write.c:1107 [inline] __x64_sys_writev+0x75/0xb0 fs/read_write.c:1107 do_syscall_64+0x103/0x610 arch/x86/entry/common.c:290 entry_SYSCALL_64_after_hwframe+0x49/0xbe Fixes: 5271953cad31 ("rxrpc: Use the UDP encap_rcv hook") Signed-off-by: Eric Dumazet <edumazet@google.com> Reported-by: syzbot <syzkaller@googlegroups.com> Acked-by: David Howells <dhowells@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2019-04-25 00:44:11 +08:00
if (unlikely(!local)) {
kfree_skb(skb);
return 0;
}
if (skb->tstamp == 0)
skb->tstamp = ktime_get_real();
rxrpc_new_skb(skb, rxrpc_skb_received);
skb_pull(skb, sizeof(struct udphdr));
/* The UDP protocol already released all skb resources;
* we are free to add our own data there.
*/
sp = rxrpc_skb(skb);
/* dig out the RxRPC connection details */
if (rxrpc_extract_header(sp, skb) < 0)
goto bad_message;
if (IS_ENABLED(CONFIG_AF_RXRPC_INJECT_LOSS)) {
static int lose;
if ((lose++ & 7) == 7) {
trace_rxrpc_rx_lose(sp);
rxrpc_free_skb(skb, rxrpc_skb_lost);
return 0;
}
}
rxrpc: Fix the data_ready handler Fix the rxrpc_data_ready() function to pick up all packets and to not miss any. There are two problems: (1) The sk_data_ready pointer on the UDP socket is set *after* it is bound. This means that it's open for business before we're ready to dequeue packets and there's a tiny window exists in which a packet can sneak onto the receive queue, but we never know about it. Fix this by setting the pointers on the socket prior to binding it. (2) skb_recv_udp() will return an error (such as ENETUNREACH) if there was an error on the transmission side, even though we set the sk_error_report hook. Because rxrpc_data_ready() returns immediately in such a case, it never actually removes its packet from the receive queue. Fix this by abstracting out the UDP dequeuing and checksumming into a separate function that keeps hammering on skb_recv_udp() until it returns -EAGAIN, passing the packets extracted to the remainder of the function. and two potential problems: (3) It might be possible in some circumstances or in the future for packets to be being added to the UDP receive queue whilst rxrpc is running consuming them, so the data_ready() handler might get called less often than once per packet. Allow for this by fully draining the queue on each call as (2). (4) If a packet fails the checksum check, the code currently returns after discarding the packet without checking for more. Allow for this by fully draining the queue on each call as (2). Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both") Signed-off-by: David Howells <dhowells@redhat.com> Acked-by: Paolo Abeni <pabeni@redhat.com>
2018-10-05 21:05:35 +08:00
if (skb->tstamp == 0)
skb->tstamp = ktime_get_real();
trace_rxrpc_rx_packet(sp);
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
switch (sp->hdr.type) {
case RXRPC_PACKET_TYPE_VERSION:
if (rxrpc_to_client(sp))
goto discard;
rxrpc_post_packet_to_local(local, skb);
goto out;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
case RXRPC_PACKET_TYPE_BUSY:
if (rxrpc_to_server(sp))
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
goto discard;
/* Fall through */
case RXRPC_PACKET_TYPE_ACK:
case RXRPC_PACKET_TYPE_ACKALL:
if (sp->hdr.callNumber == 0)
goto bad_message;
/* Fall through */
case RXRPC_PACKET_TYPE_ABORT:
break;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
case RXRPC_PACKET_TYPE_DATA:
if (sp->hdr.callNumber == 0 ||
sp->hdr.seq == 0)
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
goto bad_message;
if (!rxrpc_validate_data(skb))
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
goto bad_message;
rxrpc: Use skb_unshare() rather than skb_cow_data() The in-place decryption routines in AF_RXRPC's rxkad security module currently call skb_cow_data() to make sure the data isn't shared and that the skb can be written over. This has a problem, however, as the softirq handler may be still holding a ref or the Rx ring may be holding multiple refs when skb_cow_data() is called in rxkad_verify_packet() - and so skb_shared() returns true and __pskb_pull_tail() dislikes that. If this occurs, something like the following report will be generated. kernel BUG at net/core/skbuff.c:1463! ... RIP: 0010:pskb_expand_head+0x253/0x2b0 ... Call Trace: __pskb_pull_tail+0x49/0x460 skb_cow_data+0x6f/0x300 rxkad_verify_packet+0x18b/0xb10 [rxrpc] rxrpc_recvmsg_data.isra.11+0x4a8/0xa10 [rxrpc] rxrpc_kernel_recv_data+0x126/0x240 [rxrpc] afs_extract_data+0x51/0x2d0 [kafs] afs_deliver_fs_fetch_data+0x188/0x400 [kafs] afs_deliver_to_call+0xac/0x430 [kafs] afs_wait_for_call_to_complete+0x22f/0x3d0 [kafs] afs_make_call+0x282/0x3f0 [kafs] afs_fs_fetch_data+0x164/0x300 [kafs] afs_fetch_data+0x54/0x130 [kafs] afs_readpages+0x20d/0x340 [kafs] read_pages+0x66/0x180 __do_page_cache_readahead+0x188/0x1a0 ondemand_readahead+0x17d/0x2e0 generic_file_read_iter+0x740/0xc10 __vfs_read+0x145/0x1a0 vfs_read+0x8c/0x140 ksys_read+0x4a/0xb0 do_syscall_64+0x43/0xf0 entry_SYSCALL_64_after_hwframe+0x44/0xa9 Fix this by using skb_unshare() instead in the input path for DATA packets that have a security index != 0. Non-DATA packets don't need in-place encryption and neither do unencrypted DATA packets. Fixes: 248f219cb8bc ("rxrpc: Rewrite the data and ack handling code") Reported-by: Julian Wollrath <jwollrath@web.de> Signed-off-by: David Howells <dhowells@redhat.com>
2019-08-27 17:13:46 +08:00
/* Unshare the packet so that it can be modified for in-place
* decryption.
*/
if (sp->hdr.securityIndex != 0) {
struct sk_buff *nskb = skb_unshare(skb, GFP_ATOMIC);
if (!nskb) {
rxrpc_eaten_skb(skb, rxrpc_skb_unshared_nomem);
goto out;
}
if (nskb != skb) {
rxrpc_eaten_skb(skb, rxrpc_skb_received);
skb = nskb;
rxrpc_new_skb(skb, rxrpc_skb_unshared);
rxrpc: Use skb_unshare() rather than skb_cow_data() The in-place decryption routines in AF_RXRPC's rxkad security module currently call skb_cow_data() to make sure the data isn't shared and that the skb can be written over. This has a problem, however, as the softirq handler may be still holding a ref or the Rx ring may be holding multiple refs when skb_cow_data() is called in rxkad_verify_packet() - and so skb_shared() returns true and __pskb_pull_tail() dislikes that. If this occurs, something like the following report will be generated. kernel BUG at net/core/skbuff.c:1463! ... RIP: 0010:pskb_expand_head+0x253/0x2b0 ... Call Trace: __pskb_pull_tail+0x49/0x460 skb_cow_data+0x6f/0x300 rxkad_verify_packet+0x18b/0xb10 [rxrpc] rxrpc_recvmsg_data.isra.11+0x4a8/0xa10 [rxrpc] rxrpc_kernel_recv_data+0x126/0x240 [rxrpc] afs_extract_data+0x51/0x2d0 [kafs] afs_deliver_fs_fetch_data+0x188/0x400 [kafs] afs_deliver_to_call+0xac/0x430 [kafs] afs_wait_for_call_to_complete+0x22f/0x3d0 [kafs] afs_make_call+0x282/0x3f0 [kafs] afs_fs_fetch_data+0x164/0x300 [kafs] afs_fetch_data+0x54/0x130 [kafs] afs_readpages+0x20d/0x340 [kafs] read_pages+0x66/0x180 __do_page_cache_readahead+0x188/0x1a0 ondemand_readahead+0x17d/0x2e0 generic_file_read_iter+0x740/0xc10 __vfs_read+0x145/0x1a0 vfs_read+0x8c/0x140 ksys_read+0x4a/0xb0 do_syscall_64+0x43/0xf0 entry_SYSCALL_64_after_hwframe+0x44/0xa9 Fix this by using skb_unshare() instead in the input path for DATA packets that have a security index != 0. Non-DATA packets don't need in-place encryption and neither do unencrypted DATA packets. Fixes: 248f219cb8bc ("rxrpc: Rewrite the data and ack handling code") Reported-by: Julian Wollrath <jwollrath@web.de> Signed-off-by: David Howells <dhowells@redhat.com>
2019-08-27 17:13:46 +08:00
sp = rxrpc_skb(skb);
}
}
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
break;
case RXRPC_PACKET_TYPE_CHALLENGE:
if (rxrpc_to_server(sp))
goto discard;
break;
case RXRPC_PACKET_TYPE_RESPONSE:
if (rxrpc_to_client(sp))
goto discard;
break;
/* Packet types 9-11 should just be ignored. */
case RXRPC_PACKET_TYPE_PARAMS:
case RXRPC_PACKET_TYPE_10:
case RXRPC_PACKET_TYPE_11:
goto discard;
default:
_proto("Rx Bad Packet Type %u", sp->hdr.type);
goto bad_message;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
}
if (sp->hdr.serviceId == 0)
goto bad_message;
if (rxrpc_to_server(sp)) {
/* Weed out packets to services we're not offering. Packets
* that would begin a call are explicitly rejected and the rest
* are just discarded.
*/
rx = rcu_dereference(local->service);
if (!rx || (sp->hdr.serviceId != rx->srx.srx_service &&
sp->hdr.serviceId != rx->second_service)) {
if (sp->hdr.type == RXRPC_PACKET_TYPE_DATA &&
sp->hdr.seq == 1)
goto unsupported_service;
goto discard;
}
}
rxrpc: Make service call handling more robust Make the following changes to improve the robustness of the code that sets up a new service call: (1) Cache the rxrpc_sock struct obtained in rxrpc_data_ready() to do a service ID check and pass that along to rxrpc_new_incoming_call(). This means that I can remove the check from rxrpc_new_incoming_call() without the need to worry about the socket attached to the local endpoint getting replaced - which would invalidate the check. (2) Cache the rxrpc_peer struct, thereby allowing the peer search to be done once. The peer is passed to rxrpc_new_incoming_call(), thereby saving the need to repeat the search. This also reduces the possibility of rxrpc_publish_service_conn() BUG()'ing due to the detection of a duplicate connection, despite the initial search done by rxrpc_find_connection_rcu() having turned up nothing. This BUG() shouldn't ever get hit since rxrpc_data_ready() *should* be non-reentrant and the result of the initial search should still hold true, but it has proven possible to hit. I *think* this may be due to __rxrpc_lookup_peer_rcu() cutting short the iteration over the hash table if it finds a matching peer with a zero usage count, but I don't know for sure since it's only ever been hit once that I know of. Another possibility is that a bug in rxrpc_data_ready() that checked the wrong byte in the header for the RXRPC_CLIENT_INITIATED flag might've let through a packet that caused a spurious and invalid call to be set up. That is addressed in another patch. (3) Fix __rxrpc_lookup_peer_rcu() to skip peer records that have a zero usage count rather than stopping and returning not found, just in case there's another peer record behind it in the bucket. (4) Don't search the peer records in rxrpc_alloc_incoming_call(), but rather either use the peer cached in (2) or, if one wasn't found, preemptively install a new one. Fixes: 8496af50eb38 ("rxrpc: Use RCU to access a peer's service connection tree") Signed-off-by: David Howells <dhowells@redhat.com>
2018-09-27 22:13:09 +08:00
conn = rxrpc_find_connection_rcu(local, skb, &peer);
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
if (conn) {
if (sp->hdr.securityIndex != conn->security_ix)
goto wrong_security;
if (sp->hdr.serviceId != conn->service_id) {
rxrpc: Fix the packet reception routine The rxrpc_input_packet() function and its call tree was built around the assumption that data_ready() handler called from UDP to inform a kernel service that there is data to be had was non-reentrant. This means that certain locking could be dispensed with. This, however, turns out not to be the case with a multi-queue network card that can deliver packets to multiple cpus simultaneously. Each of those cpus can be in the rxrpc_input_packet() function at the same time. Fix by adding or changing some structure members: (1) Add peer->rtt_input_lock to serialise access to the RTT buffer. (2) Make conn->service_id into a 32-bit variable so that it can be cmpxchg'd on all arches. (3) Add call->input_lock to serialise access to the Rx/Tx state. Note that although the Rx and Tx states are (almost) entirely separate, there's no point completing the separation and having separate locks since it's a bi-phasal RPC protocol rather than a bi-direction streaming protocol. Data transmission and data reception do not take place simultaneously on any particular call. and making the following functional changes: (1) In rxrpc_input_data(), hold call->input_lock around the core to prevent simultaneous producing of packets into the Rx ring and updating of tracking state for a particular call. (2) In rxrpc_input_ping_response(), only read call->ping_serial once, and check it before checking RXRPC_CALL_PINGING as that's a cheaper test. The bit test and bit clear can then be combined. No further locking is needed here. (3) In rxrpc_input_ack(), take call->input_lock after we've parsed much of the ACK packet. The superseded ACK check is then done both before and after the lock is taken. The handing of ackinfo data is split, parsing before the lock is taken and processing with it held. This is keyed on rxMTU being non-zero. Congestion management is also done within the locked section. (4) In rxrpc_input_ackall(), take call->input_lock around the Tx window rotation. The ACKALL packet carries no information and is only really useful after all packets have been transmitted since it's imprecise. (5) In rxrpc_input_implicit_end_call(), we use rx->incoming_lock to prevent calls being simultaneously implicitly ended on two cpus and also to prevent any races with incoming call setup. (6) In rxrpc_input_packet(), use cmpxchg() to effect the service upgrade on a connection. It is only permitted to happen once for a connection. (7) In rxrpc_new_incoming_call(), we have to recheck the routing inside rx->incoming_lock to see if someone else set up the call, connection or peer whilst we were getting there. We can't trust the values from the earlier routing check unless we pin refs on them - which we want to avoid. Further, we need to allow for an incoming call to have its state changed on another CPU between us making it live and us adjusting it because the conn is now in the RXRPC_CONN_SERVICE state. (8) In rxrpc_peer_add_rtt(), take peer->rtt_input_lock around the access to the RTT buffer. Don't need to lock around setting peer->rtt. For reference, the inventory of state-accessing or state-altering functions used by the packet input procedure is: > rxrpc_input_packet() * PACKET CHECKING * ROUTING > rxrpc_post_packet_to_local() > rxrpc_find_connection_rcu() - uses RCU > rxrpc_lookup_peer_rcu() - uses RCU > rxrpc_find_service_conn_rcu() - uses RCU > idr_find() - uses RCU * CONNECTION-LEVEL PROCESSING - Service upgrade - Can only happen once per conn ! Changed to use cmpxchg > rxrpc_post_packet_to_conn() - Setting conn->hi_serial - Probably safe not using locks - Maybe use cmpxchg * CALL-LEVEL PROCESSING > Old-call checking > rxrpc_input_implicit_end_call() > rxrpc_call_completed() > rxrpc_queue_call() ! Need to take rx->incoming_lock > __rxrpc_disconnect_call() > rxrpc_notify_socket() > rxrpc_new_incoming_call() - Uses rx->incoming_lock for the entire process - Might be able to drop this earlier in favour of the call lock > rxrpc_incoming_call() ! Conflicts with rxrpc_input_implicit_end_call() > rxrpc_send_ping() - Don't need locks to check rtt state > rxrpc_propose_ACK * PACKET DISTRIBUTION > rxrpc_input_call_packet() > rxrpc_input_data() * QUEUE DATA PACKET ON CALL > rxrpc_reduce_call_timer() - Uses timer_reduce() ! Needs call->input_lock() > rxrpc_receiving_reply() ! Needs locking around ack state > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_proto_abort() > rxrpc_input_dup_data() - Fills the Rx buffer - rxrpc_propose_ACK() - rxrpc_notify_socket() > rxrpc_input_ack() * APPLY ACK PACKET TO CALL AND DISCARD PACKET > rxrpc_input_ping_response() - Probably doesn't need any extra locking ! Need READ_ONCE() on call->ping_serial > rxrpc_input_check_for_lost_ack() - Takes call->lock to consult Tx buffer > rxrpc_peer_add_rtt() ! Needs to take a lock (peer->rtt_input_lock) ! Could perhaps manage with cmpxchg() and xadd() instead > rxrpc_input_requested_ack - Consults Tx buffer ! Probably needs a lock > rxrpc_peer_add_rtt() > rxrpc_propose_ack() > rxrpc_input_ackinfo() - Changes call->tx_winsize ! Use cmpxchg to handle change ! Should perhaps track serial number - Uses peer->lock to record MTU specification changes > rxrpc_proto_abort() ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_input_soft_acks() - Consults the Tx buffer > rxrpc_congestion_management() - Modifies the Tx annotations ! Needs call->input_lock() > rxrpc_queue_call() > rxrpc_input_abort() * APPLY ABORT PACKET TO CALL AND DISCARD PACKET > rxrpc_set_call_completion() > rxrpc_notify_socket() > rxrpc_input_ackall() * APPLY ACKALL PACKET TO CALL AND DISCARD PACKET ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_reject_packet() There are some functions used by the above that queue the packet, after which the procedure is terminated: - rxrpc_post_packet_to_local() - local->event_queue is an sk_buff_head - local->processor is a work_struct - rxrpc_post_packet_to_conn() - conn->rx_queue is an sk_buff_head - conn->processor is a work_struct - rxrpc_reject_packet() - local->reject_queue is an sk_buff_head - local->processor is a work_struct And some that offload processing to process context: - rxrpc_notify_socket() - Uses RCU lock - Uses call->notify_lock to call call->notify_rx - Uses call->recvmsg_lock to queue recvmsg side - rxrpc_queue_call() - call->processor is a work_struct - rxrpc_propose_ACK() - Uses call->lock to wrap __rxrpc_propose_ACK() And a bunch that complete a call, all of which use call->state_lock to protect the call state: - rxrpc_call_completed() - rxrpc_set_call_completion() - rxrpc_abort_call() - rxrpc_proto_abort() - Also uses rxrpc_queue_call() Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both") Signed-off-by: David Howells <dhowells@redhat.com>
2018-10-08 22:46:25 +08:00
int old_id;
if (!test_bit(RXRPC_CONN_PROBING_FOR_UPGRADE, &conn->flags))
goto reupgrade;
old_id = cmpxchg(&conn->service_id, conn->params.service_id,
sp->hdr.serviceId);
if (old_id != conn->params.service_id &&
old_id != sp->hdr.serviceId)
goto reupgrade;
}
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
if (sp->hdr.callNumber == 0) {
/* Connection-level packet */
_debug("CONN %p {%d}", conn, conn->debug_id);
rxrpc_post_packet_to_conn(conn, skb);
goto out;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
}
if ((int)sp->hdr.serial - (int)conn->hi_serial > 0)
conn->hi_serial = sp->hdr.serial;
/* Call-bound packets are routed by connection channel. */
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
channel = sp->hdr.cid & RXRPC_CHANNELMASK;
chan = &conn->channels[channel];
/* Ignore really old calls */
if (sp->hdr.callNumber < chan->last_call)
goto discard;
if (sp->hdr.callNumber == chan->last_call) {
if (chan->call ||
sp->hdr.type == RXRPC_PACKET_TYPE_ABORT)
goto discard;
/* For the previous service call, if completed
* successfully, we discard all further packets.
*/
if (rxrpc_conn_is_service(conn) &&
chan->last_type == RXRPC_PACKET_TYPE_ACK)
goto discard;
/* But otherwise we need to retransmit the final packet
* from data cached in the connection record.
*/
if (sp->hdr.type == RXRPC_PACKET_TYPE_DATA)
trace_rxrpc_rx_data(chan->call_debug_id,
sp->hdr.seq,
sp->hdr.serial,
sp->hdr.flags, 0);
rxrpc_post_packet_to_conn(conn, skb);
goto out;
}
call = rcu_dereference(chan->call);
if (sp->hdr.callNumber > chan->call_id) {
if (rxrpc_to_client(sp))
goto reject_packet;
if (call)
rxrpc: Fix the packet reception routine The rxrpc_input_packet() function and its call tree was built around the assumption that data_ready() handler called from UDP to inform a kernel service that there is data to be had was non-reentrant. This means that certain locking could be dispensed with. This, however, turns out not to be the case with a multi-queue network card that can deliver packets to multiple cpus simultaneously. Each of those cpus can be in the rxrpc_input_packet() function at the same time. Fix by adding or changing some structure members: (1) Add peer->rtt_input_lock to serialise access to the RTT buffer. (2) Make conn->service_id into a 32-bit variable so that it can be cmpxchg'd on all arches. (3) Add call->input_lock to serialise access to the Rx/Tx state. Note that although the Rx and Tx states are (almost) entirely separate, there's no point completing the separation and having separate locks since it's a bi-phasal RPC protocol rather than a bi-direction streaming protocol. Data transmission and data reception do not take place simultaneously on any particular call. and making the following functional changes: (1) In rxrpc_input_data(), hold call->input_lock around the core to prevent simultaneous producing of packets into the Rx ring and updating of tracking state for a particular call. (2) In rxrpc_input_ping_response(), only read call->ping_serial once, and check it before checking RXRPC_CALL_PINGING as that's a cheaper test. The bit test and bit clear can then be combined. No further locking is needed here. (3) In rxrpc_input_ack(), take call->input_lock after we've parsed much of the ACK packet. The superseded ACK check is then done both before and after the lock is taken. The handing of ackinfo data is split, parsing before the lock is taken and processing with it held. This is keyed on rxMTU being non-zero. Congestion management is also done within the locked section. (4) In rxrpc_input_ackall(), take call->input_lock around the Tx window rotation. The ACKALL packet carries no information and is only really useful after all packets have been transmitted since it's imprecise. (5) In rxrpc_input_implicit_end_call(), we use rx->incoming_lock to prevent calls being simultaneously implicitly ended on two cpus and also to prevent any races with incoming call setup. (6) In rxrpc_input_packet(), use cmpxchg() to effect the service upgrade on a connection. It is only permitted to happen once for a connection. (7) In rxrpc_new_incoming_call(), we have to recheck the routing inside rx->incoming_lock to see if someone else set up the call, connection or peer whilst we were getting there. We can't trust the values from the earlier routing check unless we pin refs on them - which we want to avoid. Further, we need to allow for an incoming call to have its state changed on another CPU between us making it live and us adjusting it because the conn is now in the RXRPC_CONN_SERVICE state. (8) In rxrpc_peer_add_rtt(), take peer->rtt_input_lock around the access to the RTT buffer. Don't need to lock around setting peer->rtt. For reference, the inventory of state-accessing or state-altering functions used by the packet input procedure is: > rxrpc_input_packet() * PACKET CHECKING * ROUTING > rxrpc_post_packet_to_local() > rxrpc_find_connection_rcu() - uses RCU > rxrpc_lookup_peer_rcu() - uses RCU > rxrpc_find_service_conn_rcu() - uses RCU > idr_find() - uses RCU * CONNECTION-LEVEL PROCESSING - Service upgrade - Can only happen once per conn ! Changed to use cmpxchg > rxrpc_post_packet_to_conn() - Setting conn->hi_serial - Probably safe not using locks - Maybe use cmpxchg * CALL-LEVEL PROCESSING > Old-call checking > rxrpc_input_implicit_end_call() > rxrpc_call_completed() > rxrpc_queue_call() ! Need to take rx->incoming_lock > __rxrpc_disconnect_call() > rxrpc_notify_socket() > rxrpc_new_incoming_call() - Uses rx->incoming_lock for the entire process - Might be able to drop this earlier in favour of the call lock > rxrpc_incoming_call() ! Conflicts with rxrpc_input_implicit_end_call() > rxrpc_send_ping() - Don't need locks to check rtt state > rxrpc_propose_ACK * PACKET DISTRIBUTION > rxrpc_input_call_packet() > rxrpc_input_data() * QUEUE DATA PACKET ON CALL > rxrpc_reduce_call_timer() - Uses timer_reduce() ! Needs call->input_lock() > rxrpc_receiving_reply() ! Needs locking around ack state > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_proto_abort() > rxrpc_input_dup_data() - Fills the Rx buffer - rxrpc_propose_ACK() - rxrpc_notify_socket() > rxrpc_input_ack() * APPLY ACK PACKET TO CALL AND DISCARD PACKET > rxrpc_input_ping_response() - Probably doesn't need any extra locking ! Need READ_ONCE() on call->ping_serial > rxrpc_input_check_for_lost_ack() - Takes call->lock to consult Tx buffer > rxrpc_peer_add_rtt() ! Needs to take a lock (peer->rtt_input_lock) ! Could perhaps manage with cmpxchg() and xadd() instead > rxrpc_input_requested_ack - Consults Tx buffer ! Probably needs a lock > rxrpc_peer_add_rtt() > rxrpc_propose_ack() > rxrpc_input_ackinfo() - Changes call->tx_winsize ! Use cmpxchg to handle change ! Should perhaps track serial number - Uses peer->lock to record MTU specification changes > rxrpc_proto_abort() ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_input_soft_acks() - Consults the Tx buffer > rxrpc_congestion_management() - Modifies the Tx annotations ! Needs call->input_lock() > rxrpc_queue_call() > rxrpc_input_abort() * APPLY ABORT PACKET TO CALL AND DISCARD PACKET > rxrpc_set_call_completion() > rxrpc_notify_socket() > rxrpc_input_ackall() * APPLY ACKALL PACKET TO CALL AND DISCARD PACKET ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_reject_packet() There are some functions used by the above that queue the packet, after which the procedure is terminated: - rxrpc_post_packet_to_local() - local->event_queue is an sk_buff_head - local->processor is a work_struct - rxrpc_post_packet_to_conn() - conn->rx_queue is an sk_buff_head - conn->processor is a work_struct - rxrpc_reject_packet() - local->reject_queue is an sk_buff_head - local->processor is a work_struct And some that offload processing to process context: - rxrpc_notify_socket() - Uses RCU lock - Uses call->notify_lock to call call->notify_rx - Uses call->recvmsg_lock to queue recvmsg side - rxrpc_queue_call() - call->processor is a work_struct - rxrpc_propose_ACK() - Uses call->lock to wrap __rxrpc_propose_ACK() And a bunch that complete a call, all of which use call->state_lock to protect the call state: - rxrpc_call_completed() - rxrpc_set_call_completion() - rxrpc_abort_call() - rxrpc_proto_abort() - Also uses rxrpc_queue_call() Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both") Signed-off-by: David Howells <dhowells@redhat.com>
2018-10-08 22:46:25 +08:00
rxrpc_input_implicit_end_call(rx, conn, call);
call = NULL;
}
rxrpc: Fix handling of call quietly cancelled out on server Sometimes an in-progress call will stop responding on the fileserver when the fileserver quietly cancels the call with an internally marked abort (RX_CALL_DEAD), without sending an ABORT to the client. This causes the client's call to eventually expire from lack of incoming packets directed its way, which currently leads to it being cancelled locally with ETIME. Note that it's not currently clear as to why this happens as it's really hard to reproduce. The rotation policy implement by kAFS, however, doesn't differentiate between ETIME meaning we didn't get any response from the server and ETIME meaning the call got cancelled mid-flow. The latter leads to an oops when fetching data as the rotation partially resets the afs_read descriptor, which can result in a cleared page pointer being dereferenced because that page has already been filled. Handle this by the following means: (1) Set a flag on a call when we receive a packet for it. (2) Store the highest packet serial number so far received for a call (bearing in mind this may wrap). (3) If, when the "not received anything recently" timeout expires on a call, we've received at least one packet for a call and the connection as a whole has received packets more recently than that call, then cancel the call locally with ECONNRESET rather than ETIME. This indicates that the call was definitely in progress on the server. (4) In kAFS, if the rotation algorithm sees ECONNRESET rather than ETIME, don't try the next server, but rather abort the call. This avoids the oops as we don't try to reuse the afs_read struct. Rather, as-yet ungotten pages will be reread at a later data. Also: (5) Add an rxrpc tracepoint to log detection of the call being reset. Without this, I occasionally see an oops like the following: general protection fault: 0000 [#1] SMP PTI ... RIP: 0010:_copy_to_iter+0x204/0x310 RSP: 0018:ffff8800cae0f828 EFLAGS: 00010206 RAX: 0000000000000560 RBX: 0000000000000560 RCX: 0000000000000560 RDX: ffff8800cae0f968 RSI: ffff8800d58b3312 RDI: 0005080000000000 RBP: ffff8800cae0f968 R08: 0000000000000560 R09: ffff8800ca00f400 R10: ffff8800c36f28d4 R11: 00000000000008c4 R12: ffff8800cae0f958 R13: 0000000000000560 R14: ffff8800d58b3312 R15: 0000000000000560 FS: 00007fdaef108080(0000) GS:ffff8800ca680000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 00007fb28a8fa000 CR3: 00000000d2a76002 CR4: 00000000001606e0 Call Trace: skb_copy_datagram_iter+0x14e/0x289 rxrpc_recvmsg_data.isra.0+0x6f3/0xf68 ? trace_buffer_unlock_commit_regs+0x4f/0x89 rxrpc_kernel_recv_data+0x149/0x421 afs_extract_data+0x1e0/0x798 ? afs_wait_for_call_to_complete+0xc9/0x52e afs_deliver_fs_fetch_data+0x33a/0x5ab afs_deliver_to_call+0x1ee/0x5e0 ? afs_wait_for_call_to_complete+0xc9/0x52e afs_wait_for_call_to_complete+0x12b/0x52e ? wake_up_q+0x54/0x54 afs_make_call+0x287/0x462 ? afs_fs_fetch_data+0x3e6/0x3ed ? rcu_read_lock_sched_held+0x5d/0x63 afs_fs_fetch_data+0x3e6/0x3ed afs_fetch_data+0xbb/0x14a afs_readpages+0x317/0x40d __do_page_cache_readahead+0x203/0x2ba ? ondemand_readahead+0x3a7/0x3c1 ondemand_readahead+0x3a7/0x3c1 generic_file_buffered_read+0x18b/0x62f __vfs_read+0xdb/0xfe vfs_read+0xb2/0x137 ksys_read+0x50/0x8c do_syscall_64+0x7d/0x1a0 entry_SYSCALL_64_after_hwframe+0x49/0xbe Note the weird value in RDI which is a result of trying to kmap() a NULL page pointer. Signed-off-by: David Howells <dhowells@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2018-06-03 09:17:39 +08:00
if (call) {
if (sp->hdr.serviceId != call->service_id)
call->service_id = sp->hdr.serviceId;
if ((int)sp->hdr.serial - (int)call->rx_serial > 0)
call->rx_serial = sp->hdr.serial;
if (!test_bit(RXRPC_CALL_RX_HEARD, &call->flags))
set_bit(RXRPC_CALL_RX_HEARD, &call->flags);
}
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
}
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
if (!call || atomic_read(&call->usage) == 0) {
if (rxrpc_to_client(sp) ||
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
sp->hdr.type != RXRPC_PACKET_TYPE_DATA)
goto bad_message;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
if (sp->hdr.seq != 1)
goto discard;
rxrpc: Fix the packet reception routine The rxrpc_input_packet() function and its call tree was built around the assumption that data_ready() handler called from UDP to inform a kernel service that there is data to be had was non-reentrant. This means that certain locking could be dispensed with. This, however, turns out not to be the case with a multi-queue network card that can deliver packets to multiple cpus simultaneously. Each of those cpus can be in the rxrpc_input_packet() function at the same time. Fix by adding or changing some structure members: (1) Add peer->rtt_input_lock to serialise access to the RTT buffer. (2) Make conn->service_id into a 32-bit variable so that it can be cmpxchg'd on all arches. (3) Add call->input_lock to serialise access to the Rx/Tx state. Note that although the Rx and Tx states are (almost) entirely separate, there's no point completing the separation and having separate locks since it's a bi-phasal RPC protocol rather than a bi-direction streaming protocol. Data transmission and data reception do not take place simultaneously on any particular call. and making the following functional changes: (1) In rxrpc_input_data(), hold call->input_lock around the core to prevent simultaneous producing of packets into the Rx ring and updating of tracking state for a particular call. (2) In rxrpc_input_ping_response(), only read call->ping_serial once, and check it before checking RXRPC_CALL_PINGING as that's a cheaper test. The bit test and bit clear can then be combined. No further locking is needed here. (3) In rxrpc_input_ack(), take call->input_lock after we've parsed much of the ACK packet. The superseded ACK check is then done both before and after the lock is taken. The handing of ackinfo data is split, parsing before the lock is taken and processing with it held. This is keyed on rxMTU being non-zero. Congestion management is also done within the locked section. (4) In rxrpc_input_ackall(), take call->input_lock around the Tx window rotation. The ACKALL packet carries no information and is only really useful after all packets have been transmitted since it's imprecise. (5) In rxrpc_input_implicit_end_call(), we use rx->incoming_lock to prevent calls being simultaneously implicitly ended on two cpus and also to prevent any races with incoming call setup. (6) In rxrpc_input_packet(), use cmpxchg() to effect the service upgrade on a connection. It is only permitted to happen once for a connection. (7) In rxrpc_new_incoming_call(), we have to recheck the routing inside rx->incoming_lock to see if someone else set up the call, connection or peer whilst we were getting there. We can't trust the values from the earlier routing check unless we pin refs on them - which we want to avoid. Further, we need to allow for an incoming call to have its state changed on another CPU between us making it live and us adjusting it because the conn is now in the RXRPC_CONN_SERVICE state. (8) In rxrpc_peer_add_rtt(), take peer->rtt_input_lock around the access to the RTT buffer. Don't need to lock around setting peer->rtt. For reference, the inventory of state-accessing or state-altering functions used by the packet input procedure is: > rxrpc_input_packet() * PACKET CHECKING * ROUTING > rxrpc_post_packet_to_local() > rxrpc_find_connection_rcu() - uses RCU > rxrpc_lookup_peer_rcu() - uses RCU > rxrpc_find_service_conn_rcu() - uses RCU > idr_find() - uses RCU * CONNECTION-LEVEL PROCESSING - Service upgrade - Can only happen once per conn ! Changed to use cmpxchg > rxrpc_post_packet_to_conn() - Setting conn->hi_serial - Probably safe not using locks - Maybe use cmpxchg * CALL-LEVEL PROCESSING > Old-call checking > rxrpc_input_implicit_end_call() > rxrpc_call_completed() > rxrpc_queue_call() ! Need to take rx->incoming_lock > __rxrpc_disconnect_call() > rxrpc_notify_socket() > rxrpc_new_incoming_call() - Uses rx->incoming_lock for the entire process - Might be able to drop this earlier in favour of the call lock > rxrpc_incoming_call() ! Conflicts with rxrpc_input_implicit_end_call() > rxrpc_send_ping() - Don't need locks to check rtt state > rxrpc_propose_ACK * PACKET DISTRIBUTION > rxrpc_input_call_packet() > rxrpc_input_data() * QUEUE DATA PACKET ON CALL > rxrpc_reduce_call_timer() - Uses timer_reduce() ! Needs call->input_lock() > rxrpc_receiving_reply() ! Needs locking around ack state > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_proto_abort() > rxrpc_input_dup_data() - Fills the Rx buffer - rxrpc_propose_ACK() - rxrpc_notify_socket() > rxrpc_input_ack() * APPLY ACK PACKET TO CALL AND DISCARD PACKET > rxrpc_input_ping_response() - Probably doesn't need any extra locking ! Need READ_ONCE() on call->ping_serial > rxrpc_input_check_for_lost_ack() - Takes call->lock to consult Tx buffer > rxrpc_peer_add_rtt() ! Needs to take a lock (peer->rtt_input_lock) ! Could perhaps manage with cmpxchg() and xadd() instead > rxrpc_input_requested_ack - Consults Tx buffer ! Probably needs a lock > rxrpc_peer_add_rtt() > rxrpc_propose_ack() > rxrpc_input_ackinfo() - Changes call->tx_winsize ! Use cmpxchg to handle change ! Should perhaps track serial number - Uses peer->lock to record MTU specification changes > rxrpc_proto_abort() ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_input_soft_acks() - Consults the Tx buffer > rxrpc_congestion_management() - Modifies the Tx annotations ! Needs call->input_lock() > rxrpc_queue_call() > rxrpc_input_abort() * APPLY ABORT PACKET TO CALL AND DISCARD PACKET > rxrpc_set_call_completion() > rxrpc_notify_socket() > rxrpc_input_ackall() * APPLY ACKALL PACKET TO CALL AND DISCARD PACKET ! Need to take call->input_lock > rxrpc_rotate_tx_window() > rxrpc_end_tx_phase() > rxrpc_reject_packet() There are some functions used by the above that queue the packet, after which the procedure is terminated: - rxrpc_post_packet_to_local() - local->event_queue is an sk_buff_head - local->processor is a work_struct - rxrpc_post_packet_to_conn() - conn->rx_queue is an sk_buff_head - conn->processor is a work_struct - rxrpc_reject_packet() - local->reject_queue is an sk_buff_head - local->processor is a work_struct And some that offload processing to process context: - rxrpc_notify_socket() - Uses RCU lock - Uses call->notify_lock to call call->notify_rx - Uses call->recvmsg_lock to queue recvmsg side - rxrpc_queue_call() - call->processor is a work_struct - rxrpc_propose_ACK() - Uses call->lock to wrap __rxrpc_propose_ACK() And a bunch that complete a call, all of which use call->state_lock to protect the call state: - rxrpc_call_completed() - rxrpc_set_call_completion() - rxrpc_abort_call() - rxrpc_proto_abort() - Also uses rxrpc_queue_call() Fixes: 17926a79320a ("[AF_RXRPC]: Provide secure RxRPC sockets for use by userspace and kernel both") Signed-off-by: David Howells <dhowells@redhat.com>
2018-10-08 22:46:25 +08:00
call = rxrpc_new_incoming_call(local, rx, skb);
if (!call)
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
goto reject_packet;
}
/* Process a call packet; this either discards or passes on the ref
* elsewhere.
*/
rxrpc_input_call_packet(call, skb);
goto out;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
discard:
rxrpc_free_skb(skb, rxrpc_skb_freed);
out:
trace_rxrpc_rx_done(0, 0);
return 0;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
wrong_security:
trace_rxrpc_abort(0, "SEC", sp->hdr.cid, sp->hdr.callNumber, sp->hdr.seq,
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
RXKADINCONSISTENCY, EBADMSG);
skb->priority = RXKADINCONSISTENCY;
goto post_abort;
unsupported_service:
trace_rxrpc_abort(0, "INV", sp->hdr.cid, sp->hdr.callNumber, sp->hdr.seq,
RX_INVALID_OPERATION, EOPNOTSUPP);
skb->priority = RX_INVALID_OPERATION;
goto post_abort;
reupgrade:
trace_rxrpc_abort(0, "UPG", sp->hdr.cid, sp->hdr.callNumber, sp->hdr.seq,
RX_PROTOCOL_ERROR, EBADMSG);
goto protocol_error;
bad_message:
trace_rxrpc_abort(0, "BAD", sp->hdr.cid, sp->hdr.callNumber, sp->hdr.seq,
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
RX_PROTOCOL_ERROR, EBADMSG);
protocol_error:
skb->priority = RX_PROTOCOL_ERROR;
rxrpc: Rewrite the data and ack handling code Rewrite the data and ack handling code such that: (1) Parsing of received ACK and ABORT packets and the distribution and the filing of DATA packets happens entirely within the data_ready context called from the UDP socket. This allows us to process and discard ACK and ABORT packets much more quickly (they're no longer stashed on a queue for a background thread to process). (2) We avoid calling skb_clone(), pskb_pull() and pskb_trim(). We instead keep track of the offset and length of the content of each packet in the sk_buff metadata. This means we don't do any allocation in the receive path. (3) Jumbo DATA packet parsing is now done in data_ready context. Rather than cloning the packet once for each subpacket and pulling/trimming it, we file the packet multiple times with an annotation for each indicating which subpacket is there. From that we can directly calculate the offset and length. (4) A call's receive queue can be accessed without taking locks (memory barriers do have to be used, though). (5) Incoming calls are set up from preallocated resources and immediately made live. They can than have packets queued upon them and ACKs generated. If insufficient resources exist, DATA packet #1 is given a BUSY reply and other DATA packets are discarded). (6) sk_buffs no longer take a ref on their parent call. To make this work, the following changes are made: (1) Each call's receive buffer is now a circular buffer of sk_buff pointers (rxtx_buffer) rather than a number of sk_buff_heads spread between the call and the socket. This permits each sk_buff to be in the buffer multiple times. The receive buffer is reused for the transmit buffer. (2) A circular buffer of annotations (rxtx_annotations) is kept parallel to the data buffer. Transmission phase annotations indicate whether a buffered packet has been ACK'd or not and whether it needs retransmission. Receive phase annotations indicate whether a slot holds a whole packet or a jumbo subpacket and, if the latter, which subpacket. They also note whether the packet has been decrypted in place. (3) DATA packet window tracking is much simplified. Each phase has just two numbers representing the window (rx_hard_ack/rx_top and tx_hard_ack/tx_top). The hard_ack number is the sequence number before base of the window, representing the last packet the other side says it has consumed. hard_ack starts from 0 and the first packet is sequence number 1. The top number is the sequence number of the highest-numbered packet residing in the buffer. Packets between hard_ack+1 and top are soft-ACK'd to indicate they've been received, but not yet consumed. Four macros, before(), before_eq(), after() and after_eq() are added to compare sequence numbers within the window. This allows for the top of the window to wrap when the hard-ack sequence number gets close to the limit. Two flags, RXRPC_CALL_RX_LAST and RXRPC_CALL_TX_LAST, are added also to indicate when rx_top and tx_top point at the packets with the LAST_PACKET bit set, indicating the end of the phase. (4) Calls are queued on the socket 'receive queue' rather than packets. This means that we don't need have to invent dummy packets to queue to indicate abnormal/terminal states and we don't have to keep metadata packets (such as ABORTs) around (5) The offset and length of a (sub)packet's content are now passed to the verify_packet security op. This is currently expected to decrypt the packet in place and validate it. However, there's now nowhere to store the revised offset and length of the actual data within the decrypted blob (there may be a header and padding to skip) because an sk_buff may represent multiple packets, so a locate_data security op is added to retrieve these details from the sk_buff content when needed. (6) recvmsg() now has to handle jumbo subpackets, where each subpacket is individually secured and needs to be individually decrypted. The code to do this is broken out into rxrpc_recvmsg_data() and shared with the kernel API. It now iterates over the call's receive buffer rather than walking the socket receive queue. Additional changes: (1) The timers are condensed to a single timer that is set for the soonest of three timeouts (delayed ACK generation, DATA retransmission and call lifespan). (2) Transmission of ACK and ABORT packets is effected immediately from process-context socket ops/kernel API calls that cause them instead of them being punted off to a background work item. The data_ready handler still has to defer to the background, though. (3) A shutdown op is added to the AF_RXRPC socket so that the AFS filesystem can shut down the socket and flush its own work items before closing the socket to deal with any in-progress service calls. Future additional changes that will need to be considered: (1) Make sure that a call doesn't hog the front of the queue by receiving data from the network as fast as userspace is consuming it to the exclusion of other calls. (2) Transmit delayed ACKs from within recvmsg() when we've consumed sufficiently more packets to avoid the background work item needing to run. Signed-off-by: David Howells <dhowells@redhat.com>
2016-09-08 18:10:12 +08:00
post_abort:
skb->mark = RXRPC_SKB_MARK_REJECT_ABORT;
reject_packet:
trace_rxrpc_rx_done(skb->mark, skb->priority);
rxrpc_reject_packet(local, skb);
_leave(" [badmsg]");
return 0;
}