Josef Bacik diagnosed following problem :
I was seeing random disconnects while testing NBD over loopback.
This turned out to be because NBD sets pfmemalloc on it's socket,
however the receiving side is a user space application so does not
have pfmemalloc set on its socket. This means that
sk_filter_trim_cap will simply drop this packet, under the
assumption that the other side will simply retransmit. Well we do
retransmit, and then the packet is just dropped again for the same
reason.
It seems the better way to address this problem is to clear pfmemalloc
in the TCP transmit path. pfmemalloc strict control really makes sense
on the receive path.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Small cleanup factorizing code doing the TCP_MAXSEG clamping.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
syszkaller fuzzer was able to trigger a divide by zero, when
TCP window scaling is not enabled.
SO_RCVBUF can be used not only to increase sk_rcvbuf, also
to decrease it below current receive buffers utilization.
If mss is negative or 0, just return a zero TCP window.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Dmitry Vyukov <dvyukov@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently the retransmission stats are not incremented if the
retransmit fails locally. But we always increment the other packet
counters that track total packet/bytes sent. Awkwardly while we
don't count these failed retransmits in RETRANSSEGS, we do count
them in FAILEDRETRANS.
If the qdisc is dropping many packets this could under-estimate
TCP retransmission rate substantially from both SNMP or per-socket
TCP_INFO stats. This patch changes this by always incrementing
retransmission stats on retransmission attempts and failures.
Another motivation is to properly track retransmists in
SCM_TIMESTAMPING_OPT_STATS. Since SCM_TSTAMP_SCHED collection is
triggered in tcp_transmit_skb(), If tp->total_retrans is incremented
after the function, we'll always mis-count by the amount of the
latest retransmission.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Refactor the cookie check logic in tcp_send_syn_data() into a function.
This function will be called else where in later changes.
Signed-off-by: Wei Wang <weiwan@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch removes the support of RFC5827 early retransmit (i.e.,
fast recovery on small inflight with <3 dupacks) because it is
subsumed by the new RACK loss detection. More specifically when
RACK receives DUPACKs, it'll arm a reordering timer to start fast
recovery after a quarter of (min)RTT, hence it covers the early
retransmit except RACK does not limit itself to specific inflight
or dupack numbers.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Forward retransmit is an esoteric feature in RFC3517 (condition(3)
in the NextSeg()). Basically if a packet is not considered lost by
the current criteria (# of dupacks etc), but the congestion window
has room for more packets, then retransmit this packet.
However it actually conflicts with the rest of recovery design. For
example, when reordering is detected we want to be conservative
in retransmitting packets but forward-retransmit feature would
break that to force more retransmission. Also the implementation is
fairly complicated inside the retransmission logic inducing extra
iterations in the write queue. With RACK losses are being detected
timely and this heuristic is no longer necessary. There this patch
removes the feature.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch makes RACK install a reordering timer when it suspects
some packets might be lost, but wants to delay the decision
a little bit to accomodate reordering.
It does not create a new timer but instead repurposes the existing
RTO timer, because both are meant to retransmit packets.
Specifically it arms a timer ICSK_TIME_REO_TIMEOUT when
the RACK timing check fails. The wait time is set to
RACK.RTT + RACK.reo_wnd - (NOW - Packet.xmit_time) + fudge
This translates to expecting a packet (Packet) should take
(RACK.RTT + RACK.reo_wnd + fudge) to deliver after it was sent.
When there are multiple packets that need a timer, we use one timer
with the maximum timeout. Therefore the timer conservatively uses
the maximum window to expire N packets by one timeout, instead of
N timeouts to expire N packets sent at different times.
The fudge factor is 2 jiffies to ensure when the timer fires, all
the suspected packets would exceed the deadline and be marked lost
by tcp_rack_detect_loss(). It has to be at least 1 jiffy because the
clock may tick between calling icsk_reset_xmit_timer(timeout) and
actually hang the timer. The next jiffy is to lower-bound the timeout
to 2 jiffies when reo_wnd is < 1ms.
When the reordering timer fires (tcp_rack_reo_timeout): If we aren't
in Recovery we'll enter fast recovery and force fast retransmit.
This is very similar to the early retransmit (RFC5827) except RACK
is not constrained to only enter recovery for small outstanding
flights.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
ktime is a union because the initial implementation stored the time in
scalar nanoseconds on 64 bit machine and in a endianess optimized timespec
variant for 32bit machines. The Y2038 cleanup removed the timespec variant
and switched everything to scalar nanoseconds. The union remained, but
become completely pointless.
Get rid of the union and just keep ktime_t as simple typedef of type s64.
The conversion was done with coccinelle and some manual mopping up.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Madalin reported crashes happening in tcp_tasklet_func() on powerpc64
Before TSQ_QUEUED bit is cleared, we must ensure the changes done
by list_del(&tp->tsq_node); are committed to memory, otherwise
corruption might happen, as an other cpu could catch TSQ_QUEUED
clearance too soon.
We can notice that old kernels were immune to this bug, because
TSQ_QUEUED was cleared after a bh_lock_sock(sk)/bh_unlock_sock(sk)
section, but they could have missed a kick to write additional bytes,
when NIC interrupts for a given flow are spread to multiple cpus.
Affected TCP flows would need an incoming ACK or RTO timer to add more
packets to the pipe. So overall situation should be better now.
Fixes: b223feb9de ("tcp: tsq: add shortcut in tcp_tasklet_func()")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Madalin Bucur <madalin.bucur@nxp.com>
Tested-by: Madalin Bucur <madalin.bucur@nxp.com>
Tested-by: Xing Lei <xing.lei@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
tsq_flags being in the same cache line than sk_wmem_alloc
makes a lot of sense. Both fields are changed from tcp_wfree()
and more generally by various TSQ related functions.
Prior patch made room in struct sock and added sk_tsq_flags,
this patch deletes tsq_flags from struct tcp_sock.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Adding a likely() in tcp_mtu_probe() moves its code which used to
be inlined in front of tcp_write_xmit()
We still have a cache line miss to access icsk->icsk_mtup.enabled,
we will probably have to reorganize fields to help data locality.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Always allow the two first skbs in write queue to be sent,
regardless of sk_wmem_alloc/sk_pacing_rate values.
This helps a lot in situations where TX completions are delayed either
because of driver latencies or softirq latencies.
Test is done with no cache line misses.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Under high load, tcp_wfree() has an atomic operation trying
to schedule a tasklet over and over.
We can schedule it only if our per cpu list was empty.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Under high stress, I've seen tcp_tasklet_func() consuming
~700 usec, handling ~150 tcp sockets.
By setting TCP_TSQ_DEFERRED in tcp_wfree(), we give a chance
for other cpus/threads entering tcp_write_xmit() to grab it,
allowing tcp_tasklet_func() to skip sockets that already did
an xmit cycle.
In the future, we might give to ACK processing an increased
budget to reduce even more tcp_tasklet_func() amount of work.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Instead of atomically clear TSQ_THROTTLED and atomically set TSQ_QUEUED
bits, use one cmpxchg() to perform a single locked operation.
Since the following patch will also set TCP_TSQ_DEFERRED here,
this cmpxchg() will make this addition free.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This is a cleanup, to ease code review of following patches.
Old 'enum tsq_flags' is renamed, and a new enumeration is added
with the flags used in cmpxchg() operations as opposed to
single bit operations.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
jiffies based timestamps allow for easy inference of number of devices
behind NAT translators and also makes tracking of hosts simpler.
commit ceaa1fef65 ("tcp: adding a per-socket timestamp offset")
added the main infrastructure that is needed for per-connection ts
randomization, in particular writing/reading the on-wire tcp header
format takes the offset into account so rest of stack can use normal
tcp_time_stamp (jiffies).
So only two items are left:
- add a tsoffset for request sockets
- extend the tcp isn generator to also return another 32bit number
in addition to the ISN.
Re-use of ISN generator also means timestamps are still monotonically
increasing for same connection quadruple, i.e. PAWS will still work.
Includes fixes from Eric Dumazet.
Signed-off-by: Florian Westphal <fw@strlen.de>
Acked-by: Eric Dumazet <edumazet@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch measures the amount of time when TCP runs out of new data
to send to the network due to insufficient send buffer, while TCP
is still busy delivering (i.e. write queue is not empty). The goal
is to indicate either the send buffer autotuning or user SO_SNDBUF
setting has resulted network under-utilization.
The measurement starts conservatively by checking various conditions
to minimize false claims (i.e. under-estimation is more likely).
The measurement stops when the SOCK_NOSPACE flag is cleared. But it
does not account the time elapsed till the next application write.
Also the measurement only starts if the sender is still busy sending
data, s.t. the limit accounted is part of the total busy time.
Signed-off-by: Francis Yan <francisyyan@gmail.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch measures the total time when the TCP stops sending because
the receiver's advertised window is not large enough. Note that
once the limit is lifted we are likely in the busy status if we
have data pending.
Signed-off-by: Francis Yan <francisyyan@gmail.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch measures TCP busy time, which is defined as the period
of time when sender has data (or FIN) to send. The time starts when
data is buffered and stops when the write queue is flushed by ACKs
or error events.
Note the busy time does not include SYN time, unless data is
included in SYN (i.e. Fast Open). It does include FIN time even
if the FIN carries no payload. Excluding pure FIN is possible but
would incur one additional test in the fast path, which may not
be worth it.
Signed-off-by: Francis Yan <francisyyan@gmail.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch implements the skeleton of the TCP chronograph
instrumentation on sender side limits:
1) idle (unspec)
2) busy sending data other than 3-4 below
3) rwnd-limited
4) sndbuf-limited
The limits are enumerated 'tcp_chrono'. Since a connection in
theory can idle forever, we do not track the actual length of this
uninteresting idle period. For the rest we track how long the sender
spends in each limit. At any point during the life time of a
connection, the sender must be in one of the four states.
If there are multiple conditions worthy of tracking in a chronograph
then the highest priority enum takes precedence over
the other conditions. So that if something "more interesting"
starts happening, stop the previous chrono and start a new one.
The time unit is jiffy(u32) in order to save space in tcp_sock.
This implies application must sample the stats no longer than every
49 days of 1ms jiffy.
Signed-off-by: Francis Yan <francisyyan@gmail.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
With TCP MTU probing enabled and offload TX checksumming disabled,
tcp_mtu_probe() calculated the wrong checksum when a fragment being copied
into the probe's SKB had an odd length. This was caused by the direct use
of skb_copy_and_csum_bits() to calculate the checksum, as it pads the
fragment being copied, if needed. When this fragment was not the last, a
subsequent call used the previous checksum without considering this
padding.
The effect was a stale connection in one way, as even retransmissions
wouldn't solve the problem, because the checksum was never recalculated for
the full SKB length.
Signed-off-by: Douglas Caetano dos Santos <douglascs@taghos.com.br>
Signed-off-by: David S. Miller <davem@davemloft.net>
Since the TFO socket is accepted right off SYN-data, the socket
owner can call getsockopt(TCP_INFO) to collect ongoing SYN-ACK
retransmission or timeout stats (i.e., tcpi_total_retrans,
tcpi_retransmits). Currently those stats are only updated
upon handshake completes. This patch fixes it.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch fixes these under-accounting SNMP rtx stats
LINUX_MIB_TCPFORWARDRETRANS
LINUX_MIB_TCPFASTRETRANS
LINUX_MIB_TCPSLOWSTARTRETRANS
when retransmitting TSO packets
Fixes: 10d3be5692 ("tcp-tso: do not split TSO packets at retransmit time")
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We saw sch_fq drops caused by the per flow limit of 100 packets and TCP
when dealing with large cwnd and bursts of retransmits.
Even after increasing the limit to 1000, and even after commit
10d3be5692 ("tcp-tso: do not split TSO packets at retransmit time"),
we can still have these drops.
Under certain conditions, TCP can spend a considerable amount of
time queuing thousands of skbs in a single tcp_xmit_retransmit_queue()
invocation, incurring latency spikes and stalls of other softirq
handlers.
This patch implements TSQ for retransmits, limiting number of packets
and giving more chance for scheduling packets in both ways.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Export tcp_mss_to_mtu(), so that congestion control modules can use
this to help calculate a pacing rate.
Signed-off-by: Van Jacobson <vanj@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Nandita Dukkipati <nanditad@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
To allow congestion control modules to use the default TSO auto-sizing
algorithm as one of the ingredients in their own decision about TSO sizing:
1) Export tcp_tso_autosize() so that CC modules can use it.
2) Change tcp_tso_autosize() to allow callers to specify a minimum
number of segments per TSO skb, in case the congestion control
module has a different notion of the best floor for TSO skbs for
the connection right now. For very low-rate paths or policed
connections it can be appropriate to use smaller TSO skbs.
Signed-off-by: Van Jacobson <vanj@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Nandita Dukkipati <nanditad@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add the tso_segs_goal() function in tcp_congestion_ops to allow the
congestion control module to specify the number of segments that
should be in a TSO skb sent by tcp_write_xmit() and
tcp_xmit_retransmit_queue(). The congestion control module can either
request a particular number of segments in TSO skb that we transmit,
or return 0 if it doesn't care.
This allows the upcoming BBR congestion control module to select small
TSO skb sizes if the module detects that the bottleneck bandwidth is
very low, or that the connection is policed to a low rate.
Signed-off-by: Van Jacobson <vanj@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Nandita Dukkipati <nanditad@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch generates data delivery rate (throughput) samples on a
per-ACK basis. These rate samples can be used by congestion control
modules, and specifically will be used by TCP BBR in later patches in
this series.
Key state:
tp->delivered: Tracks the total number of data packets (original or not)
delivered so far. This is an already-existing field.
tp->delivered_mstamp: the last time tp->delivered was updated.
Algorithm:
A rate sample is calculated as (d1 - d0)/(t1 - t0) on a per-ACK basis:
d1: the current tp->delivered after processing the ACK
t1: the current time after processing the ACK
d0: the prior tp->delivered when the acked skb was transmitted
t0: the prior tp->delivered_mstamp when the acked skb was transmitted
When an skb is transmitted, we snapshot d0 and t0 in its control
block in tcp_rate_skb_sent().
When an ACK arrives, it may SACK and ACK some skbs. For each SACKed
or ACKed skb, tcp_rate_skb_delivered() updates the rate_sample struct
to reflect the latest (d0, t0).
Finally, tcp_rate_gen() generates a rate sample by storing
(d1 - d0) in rs->delivered and (t1 - t0) in rs->interval_us.
One caveat: if an skb was sent with no packets in flight, then
tp->delivered_mstamp may be either invalid (if the connection is
starting) or outdated (if the connection was idle). In that case,
we'll re-stamp tp->delivered_mstamp.
At first glance it seems t0 should always be the time when an skb was
transmitted, but actually this could over-estimate the rate due to
phase mismatch between transmit and ACK events. To track the delivery
rate, we ensure that if packets are in flight then t0 and and t1 are
times at which packets were marked delivered.
If the initial and final RTTs are different then one may be corrupted
by some sort of noise. The noise we see most often is sending gaps
caused by delayed, compressed, or stretched acks. This either affects
both RTTs equally or artificially reduces the final RTT. We approach
this by recording the info we need to compute the initial RTT
(duration of the "send phase" of the window) when we recorded the
associated inflight. Then, for a filter to avoid bandwidth
overestimates, we generalize the per-sample bandwidth computation
from:
bw = delivered / ack_phase_rtt
to the following:
bw = delivered / max(send_phase_rtt, ack_phase_rtt)
In large-scale experiments, this filtering approach incorporating
send_phase_rtt is effective at avoiding bandwidth overestimates due to
ACK compression or stretched ACKs.
Signed-off-by: Van Jacobson <vanj@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Nandita Dukkipati <nanditad@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
If a TCP socket gets a large write queue, an overflow can happen
in a test in __tcp_retransmit_skb() preventing all retransmits.
The flow then stalls and resets after timeouts.
Tested:
sysctl -w net.core.wmem_max=1000000000
netperf -H dest -- -s 1000000000
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
While chasing tcp_xmit_retransmit_queue() kasan issue, I found
that we could avoid reading sacked field of skb that we wont send,
possibly removing one cache line miss.
Very minor change in slow path, but why not ? ;)
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
tcp_select_initial_window() intends to advertise a window
scaling for the maximum possible window size. To do so,
it considers the maximum of net.ipv4.tcp_rmem[2] and
net.core.rmem_max as the only possible upper-bounds.
However, users with CAP_NET_ADMIN can use SO_RCVBUFFORCE
to set the socket's receive buffer size to values
larger than net.ipv4.tcp_rmem[2] and net.core.rmem_max.
Thus, SO_RCVBUFFORCE is effectively ignored by
tcp_select_initial_window().
To fix this, consider the maximum of net.ipv4.tcp_rmem[2],
net.core.rmem_max and socket's initial buffer space.
Fixes: b0573dea1f ("[NET]: Introduce SO_{SND,RCV}BUFFORCE socket options")
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Suggested-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Several cases of overlapping changes, except the packet scheduler
conflicts which deal with the addition of the free list parameter
to qdisc_enqueue().
Signed-off-by: David S. Miller <davem@davemloft.net>
Arjun reported a bug in TCP stack and bisected it to a recent commit.
In case where we process SACK, we can coalesce multiple skbs
into fat ones (tcp_shift_skb_data()), to lower write queue
overhead, because we do not expect to retransmit these packets.
However, SACK reneging can happen, forcing the sender to retransmit
all these packets. If skb->len is above 64KB, we then send buggy
IP packets that could hang TSO engine on cxgb4.
Neal suggested to use tcp_tso_autosize() instead of tp->gso_segs
so that we cook packets of optimal size vs TCP/pacing.
Thanks to Arjun for reporting the bug and running the tests !
Fixes: 10d3be5692 ("tcp-tso: do not split TSO packets at retransmit time")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Arjun V <arjun@chelsio.com>
Tested-by: Arjun V <arjun@chelsio.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add in_flight (bytes in flight when packet was sent) field
to tx component of tcp_skb_cb and make it available to
congestion modules' pkts_acked() function through the
ack_sample function argument.
Signed-off-by: Lawrence Brakmo <brakmo@fb.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
tcp_hdr() is slightly more expensive than using skb->data in contexts
where we know they point to the same byte.
In receive path, tcp_v4_rcv() and tcp_v6_rcv() are in this situation,
as tcp header has not been pulled yet.
In output path, the same can be said when we just pushed the tcp header
in the skb, in tcp_transmit_skb() and tcp_make_synack()
Also factorize the two checks for tcb->tcp_flags & TCPHDR_SYN in
tcp_transmit_skb() and pass tcp header pointer to tcp_ecn_send(),
so that compiler can further optimize and avoid a reload.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The nf_conntrack_core.c fix in 'net' is not relevant in 'net-next'
because we no longer have a per-netns conntrack hash.
The ip_gre.c conflict as well as the iwlwifi ones were cases of
overlapping changes.
Conflicts:
drivers/net/wireless/intel/iwlwifi/mvm/tx.c
net/ipv4/ip_gre.c
net/netfilter/nf_conntrack_core.c
Signed-off-by: David S. Miller <davem@davemloft.net>
In the very unlikely case __tcp_retransmit_skb() can not use the cloning
done in tcp_transmit_skb(), we need to refresh skb_mstamp before doing
the copy and transmit, otherwise TCP TS val will be an exact copy of
original transmit.
Fixes: 7faee5c0d5 ("tcp: remove TCP_SKB_CB(skb)->when")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Hosts sending lot of ACK packets exhibit high sock_wfree() cost
because of cache line miss to test SOCK_USE_WRITE_QUEUE
We could move this flag close to sk_wmem_alloc but it is better
to perform the atomic_sub_and_test() on a clean cache line,
as it avoid one extra bus transaction.
skb_orphan_partial() can also have a fast track for packets that either
are TCP acks, or already went through another skb_orphan_partial()
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We want to to make TCP stack preemptible, as draining prequeue
and backlog queues can take lot of time.
Many SNMP updates were assuming that BH (and preemption) was disabled.
Need to convert some __NET_INC_STATS() calls to NET_INC_STATS()
and some __TCP_INC_STATS() to TCP_INC_STATS()
Before using this_cpu_ptr(net->ipv4.tcp_sk) in tcp_v4_send_reset()
and tcp_v4_send_ack(), we add an explicit preempt disabled section.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The SKBTX_ACK_TSTAMP flag is set in skb_shinfo->tx_flags when
the timestamp of the TCP acknowledgement should be reported on
error queue. Since accessing skb_shinfo is likely to incur a
cache-line miss at the time of receiving the ack, the
txstamp_ack bit was added in tcp_skb_cb, which is set iff
the SKBTX_ACK_TSTAMP flag is set for an skb. This makes
SKBTX_ACK_TSTAMP flag redundant.
Remove the SKBTX_ACK_TSTAMP and instead use the txstamp_ack bit
everywhere.
Note that this frees one bit in shinfo->tx_flags.
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Acked-by: Martin KaFai Lau <kafai@fb.com>
Suggested-by: Willem de Bruijn <willemb@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Rename NET_INC_STATS_BH() to __NET_INC_STATS()
and NET_ADD_STATS_BH() to __NET_ADD_STATS()
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Linux TCP stack painfully segments all TSO/GSO packets before retransmits.
This was fine back in the days when TSO/GSO were emerging, with their
bugs, but we believe the dark age is over.
Keeping big packets in write queues, but also in stack traversal
has a lot of benefits.
- Less memory overhead, because write queues have less skbs
- Less cpu overhead at ACK processing.
- Better SACK processing, as lot of studies mentioned how
awful linux was at this ;)
- Less cpu overhead to send the rtx packets
(IP stack traversal, netfilter traversal, drivers...)
- Better latencies in presence of losses.
- Smaller spikes in fq like packet schedulers, as retransmits
are not constrained by TCP Small Queues.
1 % packet losses are common today, and at 100Gbit speeds, this
translates to ~80,000 losses per second.
Losses are often correlated, and we see many retransmit events
leading to 1-MSS train of packets, at the time hosts are already
under stress.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Conflicts were two cases of simple overlapping changes,
nothing serious.
In the UDP case, we need to add a hlist_add_tail_rcu()
to linux/rculist.h, because we've moved UDP socket handling
away from using nulls lists.
Signed-off-by: David S. Miller <davem@davemloft.net>
When removing sk_refcnt manipulation on synflood, I missed that
using skb_set_owner_w() was racy, if sk->sk_wmem_alloc had already
transitioned to 0.
We should hold sk_refcnt instead, but this is a big deal under attack.
(Doing so increase performance from 3.2 Mpps to 3.8 Mpps only)
In this patch, I chose to not attach a socket to syncookies skb.
Performance is now 5 Mpps instead of 3.2 Mpps.
Following patch will remove last known false sharing in
tcp_rcv_state_process()
Fixes: 3b24d854cb ("tcp/dccp: do not touch listener sk_refcnt under synflood")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Per RFC4898, they count segments sent/received
containing a positive length data segment (that includes
retransmission segments carrying data). Unlike
tcpi_segs_out/in, tcpi_data_segs_out/in excludes segments
carrying no data (e.g. pure ack).
The patch also updates the segs_in in tcp_fastopen_add_skb()
so that segs_in >= data_segs_in property is kept.
Together with retransmission data, tcpi_data_segs_out
gives a better signal on the rxmit rate.
v6: Rebase on the latest net-next
v5: Eric pointed out that checking skb->len is still needed in
tcp_fastopen_add_skb() because skb can carry a FIN without data.
Hence, instead of open coding segs_in and data_segs_in, tcp_segs_in()
helper is used. Comment is added to the fastopen case to explain why
segs_in has to be reset and tcp_segs_in() has to be called before
__skb_pull().
v4: Add comment to the changes in tcp_fastopen_add_skb()
and also add remark on this case in the commit message.
v3: Add const modifier to the skb parameter in tcp_segs_in()
v2: Rework based on recent fix by Eric:
commit a9d99ce28e ("tcp: fix tcpi_segs_in after connection establishment")
Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Cc: Chris Rapier <rapier@psc.edu>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Marcelo Ricardo Leitner <mleitner@redhat.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
There won't be any separate counters for socket memory consumed by
protocols other than TCP in the future. Remove the indirection and link
sockets directly to their owning memory cgroup.
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Vladimir Davydov <vdavydov@virtuozzo.com>
Acked-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
There won't be a tcp control soft limit, so integrating the memcg code
into the global skmem limiting scheme complicates things unnecessarily.
Replace this with simple and clear charge and uncharge calls--hidden
behind a jump label--to account skb memory.
Note that this is not purely aesthetic: as a result of shoehorning the
per-memcg code into the same memory accounting functions that handle the
global level, the old code would compare the per-memcg consumption
against the smaller of the per-memcg limit and the global limit. This
allowed the total consumption of multiple sockets to exceed the global
limit, as long as the individual sockets stayed within bounds. After
this change, the code will always compare the per-memcg consumption to
the per-memcg limit, and the global consumption to the global limit, and
thus close this loophole.
Without a soft limit, the per-memcg memory pressure state in sockets is
generally questionable. However, we did it until now, so we continue to
enter it when the hard limit is hit, and packets are dropped, to let
other sockets in the cgroup know that they shouldn't grow their transmit
windows, either. However, keep it simple in the new callback model and
leave memory pressure lazily when the next packet is accepted (as
opposed to doing it synchroneously when packets are processed). When
packets are dropped, network performance will already be in the toilet,
so that should be a reasonable trade-off.
As described above, consumption is now checked on the per-memcg level
and the global level separately. Likewise, memory pressure states are
maintained on both the per-memcg level and the global level, and a
socket is considered under pressure when either level asserts as much.
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Reviewed-by: Vladimir Davydov <vdavydov@virtuozzo.com>
Acked-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Conflicts:
drivers/net/geneve.c
Here we had an overlapping change, where in 'net' the extraneous stats
bump was being removed whilst in 'net-next' the final argument to
udp_tunnel6_xmit_skb() was being changed.
Signed-off-by: David S. Miller <davem@davemloft.net>
Yuchung tracked a regression caused by commit 57be5bdad7 ("ip: convert
tcp_sendmsg() to iov_iter primitives") for TCP Fast Open.
Some Fast Open users do not actually add any data in the SYN packet.
Fixes: 57be5bdad7 ("ip: convert tcp_sendmsg() to iov_iter primitives")
Reported-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Acked-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
If tcp_send_ack() can not allocate skb, we properly handle this
and setup a timer to try later.
Use __GFP_NOWARN to avoid polluting syslog in the case host is
under memory pressure, so that pertinent messages are not lost under
a flood of useless information.
sk_gfp_atomic() can use its gfp_mask argument (all callers currently
were using GFP_ATOMIC before this patch)
We rename sk_gfp_atomic() to sk_gfp_mask() to clearly express this
function now takes into account its second argument (gfp_mask)
Note that when tcp_transmit_skb() is called with clone_it set to false,
we do not attempt memory allocations, so can pass a 0 gfp_mask, which
most compilers can emit faster than a non zero or constant value.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
skb_set_owner_w() is called from various places that assume
skb->sk always point to a full blown socket (as it changes
sk->sk_wmem_alloc)
We'd like to attach skb to request sockets, and in the future
to timewait sockets as well. For these kind of pseudo sockets,
we need to take a traditional refcount and use sock_edemux()
as the destructor.
It is now time to un-inline skb_set_owner_w(), being too big.
Fixes: ca6fb06518 ("tcp: attach SYNACK messages to request sockets instead of listener")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Bisected-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Conflicts:
net/ipv6/xfrm6_output.c
net/openvswitch/flow_netlink.c
net/openvswitch/vport-gre.c
net/openvswitch/vport-vxlan.c
net/openvswitch/vport.c
net/openvswitch/vport.h
The openvswitch conflicts were overlapping changes. One was
the egress tunnel info fix in 'net' and the other was the
vport ->send() op simplification in 'net-next'.
The xfrm6_output.c conflicts was also a simplification
overlapping a bug fix.
Signed-off-by: David S. Miller <davem@davemloft.net>
Remove the existing lost retransmit detection because RACK subsumes
it completely. This also stops the overloading the ack_seq field of
the skb control block.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
At the time of commit fff3269907 ("tcp: reflect SYN queue_mapping into
SYNACK packets") we had little ways to cope with SYN floods.
We no longer need to reflect incoming skb queue mappings, and instead
can pick a TX queue based on cpu cooking the SYNACK, with normal XPS
affinities.
Note that all SYNACK retransmits were picking TX queue 0, this no longer
is a win given that SYNACK rtx are now distributed on all cpus.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
One 32bit hole is following skc_refcnt, use it.
skc_incoming_cpu can also be an union for request_sock rcv_wnd.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
If a listen backlog is very big (to avoid syncookies), then
the listener sk->sk_wmem_alloc is the main source of false
sharing, as we need to touch it twice per SYNACK re-transmit
and TX completion.
(One SYN packet takes listener lock once, but up to 6 SYNACK
are generated)
By attaching the skb to the request socket, we remove this
source of contention.
Tested:
listen(fd, 10485760); // single listener (no SO_REUSEPORT)
16 RX/TX queue NIC
Sustain a SYNFLOOD attack of ~320,000 SYN per second,
Sending ~1,400,000 SYNACK per second.
Perf profiles now show listener spinlock being next bottleneck.
20.29% [kernel] [k] queued_spin_lock_slowpath
10.06% [kernel] [k] __inet_lookup_established
5.12% [kernel] [k] reqsk_timer_handler
3.22% [kernel] [k] get_next_timer_interrupt
3.00% [kernel] [k] tcp_make_synack
2.77% [kernel] [k] ipt_do_table
2.70% [kernel] [k] run_timer_softirq
2.50% [kernel] [k] ip_finish_output
2.04% [kernel] [k] cascade
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Application limited streams such as thin streams, that transmit small
amounts of payload in relatively few packets per RTT, can be prevented
from growing the CWND when in congestion avoidance. This leads to
increased sojourn times for data segments in streams that often transmit
time-dependent data.
Currently, a connection is considered CWND limited only after having
successfully transmitted at least one packet with new data, while at the
same time failing to transmit some unsent data from the output queue
because the CWND is full. Applications that produce small amounts of
data may be left in a state where it is never considered to be CWND
limited, because all unsent data is successfully transmitted each time
an incoming ACK opens up for more data to be transmitted in the send
window.
Fix by always testing whether the CWND is fully used after successful
packet transmissions, such that a connection is considered CWND limited
whenever the CWND has been filled. This is the correct behavior as
specified in RFC2861 (section 3.1).
Cc: Andreas Petlund <apetlund@simula.no>
Cc: Carsten Griwodz <griff@simula.no>
Cc: Jonas Markussen <jonassm@ifi.uio.no>
Cc: Kenneth Klette Jonassen <kennetkl@ifi.uio.no>
Cc: Mads Johannessen <madsjoh@ifi.uio.no>
Signed-off-by: Bendik Rønning Opstad <bro.devel+kernel@gmail.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Tested-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Tested-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Conflicts:
net/ipv4/arp.c
The net/ipv4/arp.c conflict was one commit adding a new
local variable while another commit was deleting one.
Signed-off-by: David S. Miller <davem@davemloft.net>
This is done to make sure we do not change listener socket
while sending SYNACK packets while socket lock is not held.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
listener socket is not locked when tcp_make_synack() is called.
We better make sure no field is written.
There is one exception : Since SYNACK packets are attached to the listener
at this moment (or SYN_RECV child in case of Fast Open),
sock_wmalloc() needs to update sk->sk_wmem_alloc, but this is done using
atomic operations so this is safe.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
SYNACK packets might be sent without holding socket lock.
For DCTCP/ECN sake, we should call INET_ECN_xmit() while
socket lock is owned, and only when we init/change congestion control.
This also fixies a bug if congestion module is changed from
dctcp to another one on a listener : we now clear ECN bits
properly.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
RST packets sent on behalf of TCP connections with TS option (RFC 7323
TCP timestamps) have incorrect TS val (set to 0), but correct TS ecr.
A > B: Flags [S], seq 0, win 65535, options [mss 1000,nop,nop,TS val 100
ecr 0], length 0
B > A: Flags [S.], seq 2444755794, ack 1, win 28960, options [mss
1460,nop,nop,TS val 7264344 ecr 100], length 0
A > B: Flags [.], ack 1, win 65535, options [nop,nop,TS val 110 ecr
7264344], length 0
B > A: Flags [R.], seq 1, ack 1, win 28960, options [nop,nop,TS val 0
ecr 110], length 0
We need to call skb_mstamp_get() to get proper TS val,
derived from skb->skb_mstamp
Note that RFC 1323 was advocating to not send TS option in RST segment,
but RFC 7323 recommends the opposite :
Once TSopt has been successfully negotiated, that is both <SYN> and
<SYN,ACK> contain TSopt, the TSopt MUST be sent in every non-<RST>
segment for the duration of the connection, and SHOULD be sent in an
<RST> segment (see Section 5.2 for details)
Note this RFC recommends to send TS val = 0, but we believe it is
premature : We do not know if all TCP stacks are properly
handling the receive side :
When an <RST> segment is
received, it MUST NOT be subjected to the PAWS check by verifying an
acceptable value in SEG.TSval, and information from the Timestamps
option MUST NOT be used to update connection state information.
SEG.TSecr MAY be used to provide stricter <RST> acceptance checks.
In 5 years, if/when all TCP stack are RFC 7323 ready, we might consider
to decide to send TS val = 0, if it buys something.
Fixes: 7faee5c0d5 ("tcp: remove TCP_SKB_CB(skb)->when")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch makes TLP to use 1 sec timer by default when RTT is
not available due to SYN/ACK retransmission or SYN cookies.
Prior to this change, the lack of RTT prevents TLP so the first
data packets sent can only be recovered by fast recovery or RTO.
If the fast recovery fails to trigger the RTO is 3 second when
SYN/ACK is retransmitted. With this patch we can trigger fast
recovery in 1sec instead.
Note that we need to check Fast Open more properly. A Fast Open
connection could be (accepted then) closed before it receives
the final ACK of 3WHS so the state is FIN_WAIT_1. Without the
new check, TLP will retransmit FIN instead of SYN/ACK.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Nandita Dukkipati <nanditad@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In commit b73c3d0e4f ("net: Save TX flow hash in sock and set in skbuf
on xmit"), Tom provided a l4 hash to most outgoing TCP packets.
We'd like to provide one as well for SYNACK packets, so that all packets
of a given flow share same txhash, to later enable bonding driver to
also use skb->hash to perform slave selection.
Note that a SYNACK retransmit shuffles the tx hash, as Tom did
in commit 265f94ff54 ("net: Recompute sk_txhash on negative routing
advice") for established sockets.
This has nice effect making TCP flows resilient to some kind of black
holes, even at connection establish phase.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Tom Herbert <tom@herbertland.com>
Cc: Mahesh Bandewar <maheshb@google.com>
Acked-by: Tom Herbert <tom@herbertland.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Issuing a CC TX_START event on control frames like pure ACK
is a waste of time, as a CC should not care.
Following patch needs this change, as we want CUBIC to properly track
idle time at a low cost, with a single TX_START being generated.
Yuchung might slightly refine the condition triggering TX_START
on a followup patch.
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Cc: Jana Iyengar <jri@google.com>
Cc: Stephen Hemminger <stephen@networkplumber.org>
Cc: Sangtae Ha <sangtae.ha@gmail.com>
Cc: Lawrence Brakmo <lawrence@brakmo.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
slow start after idle might reduce cwnd, but we perform this
after first packet was cooked and sent.
With TSO/GSO, it means that we might send a full TSO packet
even if cwnd should have been reduced to IW10.
Moving the SSAI check in skb_entail() makes sense, because
we slightly reduce number of times this check is done,
especially for large send() and TCP Small queue callbacks from
softirq context.
As Neal pointed out, we also need to perform the check
if/when receive window opens.
Tested:
Following packetdrill test demonstrates the problem
// Test of slow start after idle
`sysctl -q net.ipv4.tcp_slow_start_after_idle=1`
0.000 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3
+0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0
+0 bind(3, ..., ...) = 0
+0 listen(3, 1) = 0
+0 < S 0:0(0) win 65535 <mss 1000,sackOK,nop,nop,nop,wscale 7>
+0 > S. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 6>
+.100 < . 1:1(0) ack 1 win 511
+0 accept(3, ..., ...) = 4
+0 setsockopt(4, SOL_SOCKET, SO_SNDBUF, [200000], 4) = 0
+0 write(4, ..., 26000) = 26000
+0 > . 1:5001(5000) ack 1
+0 > . 5001:10001(5000) ack 1
+0 %{ assert tcpi_snd_cwnd == 10 }%
+.100 < . 1:1(0) ack 10001 win 511
+0 %{ assert tcpi_snd_cwnd == 20, tcpi_snd_cwnd }%
+0 > . 10001:20001(10000) ack 1
+0 > P. 20001:26001(6000) ack 1
+.100 < . 1:1(0) ack 26001 win 511
+0 %{ assert tcpi_snd_cwnd == 36, tcpi_snd_cwnd }%
+4 write(4, ..., 20000) = 20000
// If slow start after idle works properly, we should send 5 MSS here (cwnd/2)
+0 > . 26001:31001(5000) ack 1
+0 %{ assert tcpi_snd_cwnd == 10, tcpi_snd_cwnd }%
+0 > . 31001:36001(5000) ack 1
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Yuchung Cheng <ycheng@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When TLP fails to send new packet because of receive window
limit, it should fall back to retransmit the last packet instead.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Nandita Dukkipati <nanditad@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
If TLP was unable to send a probe, it extended the RTO to
now + icsk_rto. But extending the RTO makes little sense
if no TLP probe went out. With this commit, instead of
extending the RTO we re-arm it relative to the transmit time
of the write queue head.
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Nandita Dukkipati <nanditad@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
While doing experiments with reordering resilience, we found
linux senders were not able to send at full speed under reordering,
because every incoming SACK was releasing one MSS.
This patch removes the limitation, as we did for CWR state
in commit a0ea700e40 ("tcp: tso: allow CA_CWR state in
tcp_tso_should_defer()")
Neal Cardwell had a concern about limited transmit so
Yuchung conducted experiments on GFE and found nothing
worth adding an extra check on fast path :
if (icsk->icsk_ca_state == TCP_CA_Disorder &&
tcp_sk(sk)->reordering == sysctl_tcp_reordering)
goto send_now;
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
V1 of this patch contains Eric Dumazet's suggestion to move the per
dst RTAX_QUICKACK check into tcp_in_quickack_mode(). Thanks Eric.
I ran some tests and after setting the "ip route change quickack 1"
knob there were still many delayed ACKs sent. This occured
because when icsk_ack.quick=0 the !icsk_ack.pingpong value is
subsequently ignored as tcp_in_quickack_mode() checks both these
values. The condition for a quick ack to trigger requires
that both icsk_ack.quick != 0 and icsk_ack.pingpong=0. Currently
only icsk_ack.pingpong is controlled by the knob. But the
icsk_ack.quick value changes dynamically depending on heuristics.
The crux of the matter is that delayed acks still cannot be entirely
disabled even with the RTAX_QUICKACK per dst knob enabled. This
patch ensures that a quick ack is always sent when the RTAX_QUICKACK
per dst knob is turned on.
The "ip route change quickack 1" knob was recently added to enable
quickacks. It was modeled around the TCP_QUICKACK setsockopt() option.
This issue is that even with "ip route change quickack 1" enabled
we still see delayed ACKs under some conditions. It would be nice
to be able to completely disable delayed ACKs.
Here is an example:
# netstat -s|grep dela
3 delayed acks sent
For all routes enable the knob
# ip route change quickack 1
Generate some traffic across a slow link and we still see the delayed
acks.
# netstat -s|grep dela
106 delayed acks sent
1 delayed acks further delayed because of locked socket
The issue is that both the "ip route change quickack 1" knob and
the TCP_QUICKACK option set the icsk_ack.pingpong variable to 0.
However at the business end in the __tcp_ack_snd_check() routine,
tcp_in_quickack_mode() checks that both icsk_ack.quick != 0
and icsk_ack.pingpong=0 in order to trigger a quickack. As
icsk_ack.quick is determined by heuristics it can be 0. When
that occurs the icsk_ack.pingpong value is ignored and a delayed
ACK is sent regardless.
This patch moves the RTAX_QUICKACK per dst check into the
tcp_in_quickack_mode() routine which ensures that a quickack is
always sent when the quickack knob is enabled for that dst.
Signed-off-by: Jon Maxwell <jmaxwell37@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We had various issues in the past when TCP stack was modifying
gso_size/gso_segs while clones were in flight.
Commit c52e2421f7 ("tcp: must unclone packets before mangling them")
fixed these bugs and added a WARN_ON_ONCE(skb_cloned(skb)); in
tcp_set_skb_tso_segs()
These bugs are now fixed, and because TCP stack now only sets
shinfo->gso_size|segs on the clone itself, the check can be removed.
As a result of this change, compiler inlines tcp_set_skb_tso_segs() in
tcp_init_tso_segs()
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In commit cd7d8498c9 ("tcp: change tcp_skb_pcount() location") we stored
gso_segs in a temporary cache hot location.
This patch does the same for gso_size.
This allows to save 2 cache line misses in tcp xmit path for
the last packet that is considered but not sent because of
various conditions (cwnd, tso defer, receiver window, TSQ...)
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
tcp_set_skb_tso_segs() & tcp_init_tso_segs() no longer
use the sock pointer.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Our goal is to touch skb_shinfo(skb) only when absolutely needed,
to avoid two cache line misses in TCP output path for last skb
that is considered but not sent because of various conditions
(cwnd, tso defer, receiver window, TSQ...)
A packet is GSO only when skb_shinfo(skb)->gso_size is not zero.
We can set skb_shinfo(skb)->gso_type to sk->sk_gso_type even for
non GSO packets.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Xen virtual network driver has higher latency than a physical NIC.
Having only 128K as limit for TSQ introduced 30% regression in guest
throughput.
This patch raises the limit to 256K. This reduces the regression to 8%.
This buys us more time to work out a proper solution in the long run.
Signed-off-by: Wei Liu <wei.liu2@citrix.com>
Cc: David Miller <davem@davemloft.net>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
By making sure sk->sk_gso_max_segs minimal value is one,
and sysctl_tcp_min_tso_segs minimal value is one as well,
tcp_tso_autosize() will return a non zero value.
We can then revert 843925f33f
("tcp: Do not apply TSO segment limit to non-TSO packets")
and save few cpu cycles in fast path.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Acked-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch tracks the total number of inbound and outbound segments on a
TCP socket. One may use this number to have an idea on connection
quality when compared against the retransmissions.
RFC4898 named these : tcpEStatsPerfSegsIn and tcpEStatsPerfSegsOut
These are a 32bit field each and can be fetched both from TCP_INFO
getsockopt() if one has a handle on a TCP socket, or from inet_diag
netlink facility (iproute2/ss patch will follow)
Note that tp->segs_out was placed near tp->snd_nxt for good data
locality and minimal performance impact, while tp->segs_in was placed
near tp->bytes_received for the same reason.
Join work with Eric Dumazet.
Note that received SYN are accounted on the listener, but sent SYNACK
are not accounted.
Signed-off-by: Marcelo Ricardo Leitner <mleitner@redhat.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In commit 8e4d980ac2 ("tcp: fix behavior for epoll edge trigger")
we fixed a possible hang of TCP sockets under memory pressure,
by allowing sk_stream_alloc_skb() to use sk_forced_mem_schedule()
if no packet is in socket write queue.
It turns out there are other cases where we want to force memory
schedule :
tcp_fragment() & tso_fragment() need to split a big TSO packet into
two smaller ones. If we block here because of TCP memory pressure,
we can effectively block TCP socket from sending new data.
If no further ACK is coming, this hang would be definitive, and socket
has no chance to effectively reduce its memory usage.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This work as a follow-up of commit f7b3bec6f5 ("net: allow setting ecn
via routing table") and adds RFC3168 section 6.1.1.1. fallback for outgoing
ECN connections. In other words, this work adds a retry with a non-ECN
setup SYN packet, as suggested from the RFC on the first timeout:
[...] A host that receives no reply to an ECN-setup SYN within the
normal SYN retransmission timeout interval MAY resend the SYN and
any subsequent SYN retransmissions with CWR and ECE cleared. [...]
Schematic client-side view when assuming the server is in tcp_ecn=2 mode,
that is, Linux default since 2009 via commit 255cac91c3 ("tcp: extend
ECN sysctl to allow server-side only ECN"):
1) Normal ECN-capable path:
SYN ECE CWR ----->
<----- SYN ACK ECE
ACK ----->
2) Path with broken middlebox, when client has fallback:
SYN ECE CWR ----X crappy middlebox drops packet
(timeout, rtx)
SYN ----->
<----- SYN ACK
ACK ----->
In case we would not have the fallback implemented, the middlebox drop
point would basically end up as:
SYN ECE CWR ----X crappy middlebox drops packet
(timeout, rtx)
SYN ECE CWR ----X crappy middlebox drops packet
(timeout, rtx)
SYN ECE CWR ----X crappy middlebox drops packet
(timeout, rtx)
In any case, it's rather a smaller percentage of sites where there would
occur such additional setup latency: it was found in end of 2014 that ~56%
of IPv4 and 65% of IPv6 servers of Alexa 1 million list would negotiate
ECN (aka tcp_ecn=2 default), 0.42% of these webservers will fail to connect
when trying to negotiate with ECN (tcp_ecn=1) due to timeouts, which the
fallback would mitigate with a slight latency trade-off. Recent related
paper on this topic:
Brian Trammell, Mirja Kühlewind, Damiano Boppart, Iain Learmonth,
Gorry Fairhurst, and Richard Scheffenegger:
"Enabling Internet-Wide Deployment of Explicit Congestion Notification."
Proc. PAM 2015, New York.
http://ecn.ethz.ch/ecn-pam15.pdf
Thus, when net.ipv4.tcp_ecn=1 is being set, the patch will perform RFC3168,
section 6.1.1.1. fallback on timeout. For users explicitly not wanting this
which can be in DC use case, we add a net.ipv4.tcp_ecn_fallback knob that
allows for disabling the fallback.
tp->ecn_flags are not being cleared in tcp_ecn_clear_syn() on output, but
rather we let tcp_ecn_rcv_synack() take that over on input path in case a
SYN ACK ECE was delayed. Thus a spurious SYN retransmission will not prevent
ECN being negotiated eventually in that case.
Reference: https://www.ietf.org/proceedings/92/slides/slides-92-iccrg-1.pdf
Reference: https://www.ietf.org/proceedings/89/slides/slides-89-tsvarea-1.pdf
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Mirja Kühlewind <mirja.kuehlewind@tik.ee.ethz.ch>
Signed-off-by: Brian Trammell <trammell@tik.ee.ethz.ch>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Dave That <dave.taht@gmail.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Introduce an optimized version of sk_under_memory_pressure()
for TCP. Our intent is to use it in fast paths.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We plan to use sk_forced_wmem_schedule() in input path as well,
so make it non static and rename it.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Diagnosing problems related to Window Probes has been hard because
we lack a counter.
TCPWinProbe counts the number of ACK packets a sender has to send
at regular intervals to make sure a reverse ACK packet opening back
a window had not been lost.
TCPKeepAlive counts the number of ACK packets sent to keep TCP
flows alive (SO_KEEPALIVE)
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Nandita Dukkipati <nanditad@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
With the advent of small rto timers in datacenter TCP,
(ip route ... rto_min x), the following can happen :
1) Qdisc is full, transmit fails.
TCP sets a timer based on icsk_rto to retry the transmit, without
exponential backoff.
With low icsk_rto, and lot of sockets, all cpus are servicing timer
interrupts like crazy.
Intent of the code was to retry with a timer between 200 (TCP_RTO_MIN)
and 500ms (TCP_RESOURCE_PROBE_INTERVAL)
2) Receivers can send zero windows if they don't drain their receive queue.
TCP sends zero window probes, based on icsk_rto current value, with
exponential backoff.
With /proc/sys/net/ipv4/tcp_retries2 being 15 (or even smaller in
some cases), sender can abort in less than one or two minutes !
If receiver stops the sender, it obviously doesn't care of very tight
rto. Probability of dropping the ACK reopening the window is not
worth the risk.
Lets change the base timer to be at least 200ms (TCP_RTO_MIN) for these
events (but not normal RTO based retransmits)
A followup patch adds a new SNMP counter, as it would have helped a lot
diagnosing this issue.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Presence of an unbound loop in tcp_send_fin() had always been hard
to explain when analyzing crash dumps involving gigantic dying processes
with millions of sockets.
Lets try a different strategy :
In case of memory pressure, try to add the FIN flag to last packet
in write queue, even if packet was already sent. TCP stack will
be able to deliver this FIN after a timeout event. Note that this
FIN being delivered by a retransmit, it also carries a Push flag
given our current implementation.
By checking sk_under_memory_pressure(), we anticipate that cooking
many FIN packets might deplete tcp memory.
In the case we could not allocate a packet, even with __GFP_WAIT
allocation, then not sending a FIN seems quite reasonable if it allows
to get rid of this socket, free memory, and not block the process from
eventually doing other useful work.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Using sk_stream_alloc_skb() in tcp_send_fin() is dangerous in
case a huge process is killed by OOM, and tcp_mem[2] is hit.
To be able to free memory we need to make progress, so this
patch allows FIN packets to not care about tcp_mem[2], if
skb allocation succeeded.
In a follow-up patch, we might abort tcp_send_fin() infinite loop
in case TIF_MEMDIE is set on this thread, as memory allocator
did its best getting extra memory already.
This patch reverts d22e153718 ("tcp: fix tcp fin memory accounting")
Fixes: d22e153718 ("tcp: fix tcp fin memory accounting")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The dwmac-socfpga.c conflict was a case of a bug fix overlapping
changes in net-next to handle an error pointer differently.
Signed-off-by: David S. Miller <davem@davemloft.net>
I noticed tcpdump was giving funky timestamps for locally
generated SYNACK messages on loopback interface.
11:42:46.938990 IP 127.0.0.1.48245 > 127.0.0.2.23850: S
945476042:945476042(0) win 43690 <mss 65495,nop,nop,sackOK,nop,wscale 7>
20:28:58.502209 IP 127.0.0.2.23850 > 127.0.0.1.48245: S
3160535375:3160535375(0) ack 945476043 win 43690 <mss
65495,nop,nop,sackOK,nop,wscale 7>
This is because we need to clear skb->tstamp before
entering lower stack, otherwise net_timestamp_check()
does not set skb->tstamp.
Fixes: 7faee5c0d5 ("tcp: remove TCP_SKB_CB(skb)->when")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fast Open has been using an experimental option with a magic number
(RFC6994). This patch makes the client by default use the RFC7413
option (34) to get and send Fast Open cookies. This patch makes
the client solicit cookies from a given server first with the
RFC7413 option. If that fails to elicit a cookie, then it tries
the RFC6994 experimental option. If that also fails, it uses the
RFC7413 option on all subsequent connect attempts. If the server
returns a Fast Open cookie then the client caches the form of the
option that successfully elicited a cookie, and uses that form on
later connects when it presents that cookie.
The idea is to gradually obsolete the use of experimental options as
the servers and clients upgrade, while keeping the interoperability
meanwhile.
Signed-off-by: Daniel Lee <Longinus00@gmail.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Fast Open has been using the experimental option with a magic number
(RFC6994) to request and grant Fast Open cookies. This patch enables
the server to support the official IANA option 34 in RFC7413 in
addition.
The change has passed all existing Fast Open tests with both
old and new options at Google.
Signed-off-by: Daniel Lee <Longinus00@gmail.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The ipv4 code uses a mixture of coding styles. In some instances check
for non-NULL pointer is done as x != NULL and sometimes as x. x is
preferred according to checkpatch and this patch makes the code
consistent by adopting the latter form.
No changes detected by objdiff.
Signed-off-by: Ian Morris <ipm@chirality.org.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
The ipv4 code uses a mixture of coding styles. In some instances check
for NULL pointer is done as x == NULL and sometimes as !x. !x is
preferred according to checkpatch and this patch makes the code
consistent by adopting the latter form.
No changes detected by objdiff.
Signed-off-by: Ian Morris <ipm@chirality.org.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
With request socks convergence, we no longer need
different lookup methods. A request socket can
use generic lookup function.
Add const qualifier to 2nd tcp_v[46]_md5_lookup() parameter.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Since request and established sockets now have same base,
there is no need to pass two pointers to tcp_v4_md5_hash_skb()
or tcp_v6_md5_hash_skb()
Also add a const qualifier to their struct tcp_md5sig_key argument.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
While timer handler effectively runs a rcu read locked section,
there is no explicit rcu_read_lock()/rcu_read_unlock() annotations
and lockdep can be confused here :
net/ipv4/tcp_ipv4.c-906- /* caller either holds rcu_read_lock() or socket lock */
net/ipv4/tcp_ipv4.c:907: md5sig = rcu_dereference_check(tp->md5sig_info,
net/ipv4/tcp_ipv4.c-908- sock_owned_by_user(sk) ||
net/ipv4/tcp_ipv4.c-909- lockdep_is_held(&sk->sk_lock.slock));
Let's explicitely acquire rcu_read_lock() in tcp_make_synack()
Before commit fa76ce7328 ("inet: get rid of central tcp/dccp listener
timer"), we were holding listener lock so lockdep was happy.
Fixes: fa76ce7328 ("inet: get rid of central tcp/dccp listener timer")
Signed-off-by: Eric DUmazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This reverts commit ca10b9e9a8.
No longer needed after commit eb8895debe
("tcp: tcp_make_synack() should use sock_wmalloc")
When under SYNFLOOD, we build lot of SYNACK and hit false sharing
because of multiple modifications done on sk_listener->sk_wmem_alloc
Since tcp_make_synack() uses sock_wmalloc(), there is no need
to call skb_set_owner_w() again, as this adds two atomic operations.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Conflicts:
drivers/net/ethernet/emulex/benet/be_main.c
net/core/sysctl_net_core.c
net/ipv4/inet_diag.c
The be_main.c conflict resolution was really tricky. The conflict
hunks generated by GIT were very unhelpful, to say the least. It
split functions in half and moved them around, when the real actual
conflict only existed solely inside of one function, that being
be_map_pci_bars().
So instead, to resolve this, I checked out be_main.c from the top
of net-next, then I applied the be_main.c changes from 'net' since
the last time I merged. And this worked beautifully.
The inet_diag.c and sysctl_net_core.c conflicts were simple
overlapping changes, and were easily to resolve.
Signed-off-by: David S. Miller <davem@davemloft.net>
tcp_send_fin() does not account for the memory it allocates properly, so
sk_forward_alloc can be negative in cases where we've sent a FIN:
ss example output (ss -amn | grep -B1 f4294):
tcp FIN-WAIT-1 0 1 192.168.0.1:45520 192.0.2.1:8080
skmem:(r0,rb87380,t0,tb87380,f4294966016,w1280,o0,bl0)
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
As per RFC4821 7.3. Selecting Probe Size, a probe timer should
be armed once probing has converged. Once this timer expired,
probing again to take advantage of any path PMTU change. The
recommended probing interval is 10 minutes per RFC1981. Probing
interval could be sysctled by sysctl_tcp_probe_interval.
Eric Dumazet suggested to implement pseudo timer based on 32bits
jiffies tcp_time_stamp instead of using classic timer for such
rare event.
Signed-off-by: Fan Du <fan.du@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Current probe_size is chosen by doubling mss_cache,
the probing process will end shortly with a sub-optimal
mss size, and the link mtu will not be taken full
advantage of, in return, this will make user to tweak
tcp_base_mss with care.
Use binary search to choose probe_size in a fine
granularity manner, an optimal mss will be found
to boost performance as its maxmium.
In addition, introduce a sysctl_tcp_probe_threshold
to control when probing will stop in respect to
the width of search range.
Test env:
Docker instance with vxlan encapuslation(82599EB)
iperf -c 10.0.0.24 -t 60
before this patch:
1.26 Gbits/sec
After this patch: increase 26%
1.59 Gbits/sec
Signed-off-by: Fan Du <fan.du@intel.com>
Acked-by: John Heffner <johnwheffner@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Another TCP issue is triggered by ECN.
Under pressure, receiver gets ECN marks, and send back ACK packets
with ECE TCP flag. Senders enter CA_CWR state.
In this state, tcp_tso_should_defer() is short cut :
if (icsk->icsk_ca_state != TCP_CA_Open)
goto send_now;
This means that about all ACK packets we receive are triggering
a partial send, and because cwnd is kept small, we can only send
a small amount of data for each incoming ACK,
which in return generate more ACK packets.
Allowing CA_Open and CA_CWR states to enable TSO defer in
tcp_tso_should_defer() brings performance back :
TSO autodefer has more chance to defer under pressure.
This patch increases TSO and LRO/GRO efficiency back to normal levels,
and does not impact overall ECN behavior.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
With sysctl_tcp_min_tso_segs being 4, it is very possible
that tcp_tso_should_defer() decides not sending last 2 MSS
of initial window of 10 packets. This also applies if
autosizing decides to send X MSS per GSO packet, and cwnd
is not a multiple of X.
This patch implements an heuristic based on age of first
skb in write queue : If it was sent very recently (less than half srtt),
we can predict that no ACK packet will come in less than half rtt,
so deferring might cause an under utilization of our window.
This is visible on initial send (IW10) on web servers,
but more generally on some RPC, as the last part of the message
might need an extra RTT to get delivered.
Tested:
Ran following packetdrill test
// A simple server-side test that sends exactly an initial window (IW10)
// worth of packets.
`sysctl -e -q net.ipv4.tcp_min_tso_segs=4`
0.000 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3
+0 setsockopt(3, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0
+0 bind(3, ..., ...) = 0
+0 listen(3, 1) = 0
+.1 < S 0:0(0) win 32792 <mss 1460,sackOK,nop,nop,nop,wscale 7>
+0 > S. 0:0(0) ack 1 <mss 1460,nop,nop,sackOK,nop,wscale 6>
+.1 < . 1:1(0) ack 1 win 257
+0 accept(3, ..., ...) = 4
+0 write(4, ..., 14600) = 14600
+0 > . 1:5841(5840) ack 1 win 457
+0 > . 5841:11681(5840) ack 1 win 457
// Following packet should be sent right now.
+0 > P. 11681:14601(2920) ack 1 win 457
+.1 < . 1:1(0) ack 14601 win 257
+0 close(4) = 0
+0 > F. 14601:14601(0) ack 1
+.1 < F. 1:1(0) ack 14602 win 257
+0 > . 14602:14602(0) ack 2
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
TSO relies on ability to defer sending a small amount of packets.
Heuristic is to wait for future ACKS in hope to send more packets at once.
Current algorithm uses a per socket tso_deferred field as a pseudo timer.
This pseudo timer relies on future ACK, but there is no guarantee
we receive them in time.
Fix would be to use a real timer, but cost of such timer is probably too
expensive for typical cases.
This patch changes the logic to test the time of last transmit,
because we should not add bursts of more than 1ms for any given flow.
We've used this patch for about two years at Google, before FQ/pacing
as it would reduce a fair amount of bursts.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Packetization Layer Path MTU Discovery works separately beside
Path MTU Discovery at IP level, different net namespace has
various requirements on which one to chose, e.g., a virutalized
container instance would require TCP PMTU to probe an usable
effective mtu for underlying tunnel, while the host would
employ classical ICMP based PMTU to function.
Hence making TCP PMTU mechanism per net namespace to decouple
two functionality. Furthermore the probe base MSS should also
be configured separately for each namespace.
Signed-off-by: Fan Du <fan.du@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When we added pacing to TCP, we decided to let sch_fq take care
of actual pacing.
All TCP had to do was to compute sk->pacing_rate using simple formula:
sk->pacing_rate = 2 * cwnd * mss / rtt
It works well for senders (bulk flows), but not very well for receivers
or even RPC :
cwnd on the receiver can be less than 10, rtt can be around 100ms, so we
can end up pacing ACK packets, slowing down the sender.
Really, only the sender should pace, according to its own logic.
Instead of adding a new bit in skb, or call yet another flow
dissection, we tweak skb->truesize to a small value (2), and
we instruct sch_fq to use new helper and not pace pure ack.
Note this also helps TCP small queue, as ack packets present
in qdisc/NIC do not prevent sending a data packet (RPC workload)
This helps to reduce tx completion overhead, ack packets can use regular
sock_wfree() instead of tcp_wfree() which is a bit more expensive.
This has no impact in the case packets are sent to loopback interface,
as we do not coalesce ack packets (were we would detect skb->truesize
lie)
In case netem (with a delay) is used, skb_orphan_partial() also sets
skb->truesize to 1.
This patch is a combination of two patches we used for about one year at
Google.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
patch is actually smaller than it seems to be - most of it is unindenting
the inner loop body in tcp_sendmsg() itself...
the bit in tcp_input.c is going to get reverted very soon - that's what
memcpy_from_msg() will become, but not in this commit; let's keep it
reasonably contained...
There's one potentially subtle change here: in case of short copy from
userland, mainline tcp_send_syn_data() discards the skb it has allocated
and falls back to normal path, where we'll send as much as possible after
rereading the same data again. This patch trims SYN+data skb instead -
that way we don't need to copy from the same place twice.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
This work adds the possibility to define a per route/destination
congestion control algorithm. Generally, this opens up the possibility
for a machine with different links to enforce specific congestion
control algorithms with optimal strategies for each of them based
on their network characteristics, even transparently for a single
application listening on all links.
For our specific use case, this additionally facilitates deployment
of DCTCP, for example, applications can easily serve internal
traffic/dsts in DCTCP and external one with CUBIC. Other scenarios
would also allow for utilizing e.g. long living, low priority
background flows for certain destinations/routes while still being
able for normal traffic to utilize the default congestion control
algorithm. We also thought about a per netns setting (where different
defaults are possible), but given its actually a link specific
property, we argue that a per route/destination setting is the most
natural and flexible.
The administrator can utilize this through ip-route(8) by appending
"congctl [lock] <name>", where <name> denotes the name of a
congestion control algorithm and the optional lock parameter allows
to enforce the given algorithm so that applications in user space
would not be allowed to overwrite that algorithm for that destination.
The dst metric lookups are being done when a dst entry is already
available in order to avoid a costly lookup and still before the
algorithms are being initialized, thus overhead is very low when the
feature is not being used. While the client side would need to drop
the current reference on the module, on server side this can actually
even be avoided as we just got a flat-copied socket clone.
Joint work with Florian Westphal.
Suggested-by: Hannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Thomas Jarosch reported IPsec TCP stalls when a PMTU event occurs.
In fact the problem was completely unrelated to IPsec. The bug is
also reproducible if you just disable TSO/GSO.
The problem is that when the MSS goes down, existing queued packet
on the TX queue that have not been transmitted yet all look like
TSO packets and get treated as such.
This then triggers a bug where tcp_mss_split_point tells us to
generate a zero-sized packet on the TX queue. Once that happens
we're screwed because the zero-sized packet can never be removed
by ACKs.
Fixes: 1485348d24 ("tcp: Apply device TSO segment limit earlier")
Reported-by: Thomas Jarosch <thomas.jarosch@intra2net.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Cheers,
Signed-off-by: David S. Miller <davem@davemloft.net>
Commit 95bd09eb27 ("tcp: TSO packets automatic sizing") tried to
control TSO size, but did this at the wrong place (sendmsg() time)
At sendmsg() time, we might have a pessimistic view of flow rate,
and we end up building very small skbs (with 2 MSS per skb).
This is bad because :
- It sends small TSO packets even in Slow Start where rate quickly
increases.
- It tends to make socket write queue very big, increasing tcp_ack()
processing time, but also increasing memory needs, not necessarily
accounted for, as fast clones overhead is currently ignored.
- Lower GRO efficiency and more ACK packets.
Servers with a lot of small lived connections suffer from this.
Lets instead fill skbs as much as possible (64KB of payload), but split
them at xmit time, when we have a precise idea of the flow rate.
skb split is actually quite efficient.
Patch looks bigger than necessary, because TCP Small Queue decision now
has to take place after the eventual split.
As Neal suggested, introduce a new tcp_tso_autosize() helper, so that
tcp_tso_should_defer() can be synchronized on same goal.
Rename tp->xmit_size_goal_segs to tp->gso_segs, as this variable
contains number of mss that we can put in GSO packet, and is not
related to the autosizing goal anymore.
Tested:
40 ms rtt link
nstat >/dev/null
netperf -H remote -l -2000000 -- -s 1000000
nstat | egrep "IpInReceives|IpOutRequests|TcpOutSegs|IpExtOutOctets"
Before patch :
Recv Send Send
Socket Socket Message Elapsed
Size Size Size Time Throughput
bytes bytes bytes secs. 10^6bits/s
87380 2000000 2000000 0.36 44.22
IpInReceives 600 0.0
IpOutRequests 599 0.0
TcpOutSegs 1397 0.0
IpExtOutOctets 2033249 0.0
After patch :
Recv Send Send
Socket Socket Message Elapsed
Size Size Size Time Throughput
bytes bytes bytes secs. 10^6bits/sec
87380 2000000 2000000 0.36 44.27
IpInReceives 221 0.0
IpOutRequests 232 0.0
TcpOutSegs 1397 0.0
IpExtOutOctets 2013953 0.0
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Note that the code _using_ ->msg_iter at that point will be very
unhappy with anything other than unshifted iovec-backed iov_iter.
We still need to convert users to proper primitives.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
While working on sk_forward_alloc problems reported by Denys
Fedoryshchenko, we found that tcp connect() (and fastopen) do not call
sk_wmem_schedule() for SYN packet (and/or SYN/DATA packet), so
sk_forward_alloc is negative while connect is in progress.
We can fix this by calling regular sk_stream_alloc_skb() both for the
SYN packet (in tcp_connect()) and the syn_data packet in
tcp_send_syn_data()
Then, tcp_send_syn_data() can avoid copying syn_data as we simply
can manipulate syn_data->cb[] to remove SYN flag (and increment seq)
Instead of open coding memcpy_fromiovecend(), simply use this helper.
This leaves in socket write queue clean fast clone skbs.
This was tested against our fastopen packetdrill tests.
Reported-by: Denys Fedoryshchenko <nuclearcat@nuclearcat.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In DC world, GSO packets initially cooked by tcp_sendmsg() are usually
big, as sk_pacing_rate is high.
When network is congested, cwnd can be smaller than the GSO packets
found in socket write queue. tcp_write_xmit() splits GSO packets
using the available cwnd, and we end up sending a single GSO packet,
consuming all available cwnd.
With GRO aggregation on the receiver, we might handle a single GRO
packet, sending back a single ACK.
1) This single ACK might be lost
TLP or RTO are forced to attempt a retransmit.
2) This ACK releases a full cwnd, sender sends another big GSO packet,
in a ping pong mode.
This behavior does not fill the pipes in the best way, because of
scheduling artifacts.
Make sure we always have at least two GSO packets in flight.
This allows us to safely increase GRO efficiency without risking
spurious retransmits.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch allows to set ECN on a per-route basis in case the sysctl
tcp_ecn is not set to 1. In other words, when ECN is set for specific
routes, it provides a tcp_ecn=1 behaviour for that route while the rest
of the stack acts according to the global settings.
One can use 'ip route change dev $dev $net features ecn' to toggle this.
Having a more fine-grained per-route setting can be beneficial for various
reasons, for example, 1) within data centers, or 2) local ISPs may deploy
ECN support for their own video/streaming services [1], etc.
There was a recent measurement study/paper [2] which scanned the Alexa's
publicly available top million websites list from a vantage point in US,
Europe and Asia:
Half of the Alexa list will now happily use ECN (tcp_ecn=2, most likely
blamed to commit 255cac91c3 ("tcp: extend ECN sysctl to allow server-side
only ECN") ;)); the break in connectivity on-path was found is about
1 in 10,000 cases. Timeouts rather than receiving back RSTs were much
more common in the negotiation phase (and mostly seen in the Alexa
middle band, ranks around 50k-150k): from 12-thousand hosts on which
there _may_ be ECN-linked connection failures, only 79 failed with RST
when _not_ failing with RST when ECN is not requested.
It's unclear though, how much equipment in the wild actually marks CE
when buffers start to fill up.
We thought about a fallback to non-ECN for retransmitted SYNs as another
global option (which could perhaps one day be made default), but as Eric
points out, there's much more work needed to detect broken middleboxes.
Two examples Eric mentioned are buggy firewalls that accept only a single
SYN per flow, and middleboxes that successfully let an ECN flow establish,
but later mark CE for all packets (so cwnd converges to 1).
[1] http://www.ietf.org/proceedings/89/slides/slides-89-tsvarea-1.pdf, p.15
[2] http://ecn.ethz.ch/
Joint work with Daniel Borkmann.
Reference: http://thread.gmane.org/gmane.linux.network/335797
Suggested-by: Hannes Frederic Sowa <hannes@stressinduktion.org>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
Some drivers are unable to perform TX completions in a bound time.
They instead call skb_orphan()
Problem is skb_fclone_busy() has to detect this case, otherwise
we block TCP retransmits and can freeze unlucky tcp sessions on
mostly idle hosts.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Fixes: 1f3279ae0c ("tcp: avoid retransmits of TCP packets hanging in host queues")
Signed-off-by: David S. Miller <davem@davemloft.net>
Pull networking fixes from David Miller:
1) Include fixes for netrom and dsa (Fabian Frederick and Florian
Fainelli)
2) Fix FIXED_PHY support in stmmac, from Giuseppe CAVALLARO.
3) Several SKB use after free fixes (vxlan, openvswitch, vxlan,
ip_tunnel, fou), from Li ROngQing.
4) fec driver PTP support fixes from Luwei Zhou and Nimrod Andy.
5) Use after free in virtio_net, from Michael S Tsirkin.
6) Fix flow mask handling for megaflows in openvswitch, from Pravin B
Shelar.
7) ISDN gigaset and capi bug fixes from Tilman Schmidt.
8) Fix route leak in ip_send_unicast_reply(), from Vasily Averin.
9) Fix two eBPF JIT bugs on x86, from Alexei Starovoitov.
10) TCP_SKB_CB() reorganization caused a few regressions, fixed by Cong
Wang and Eric Dumazet.
11) Don't overwrite end of SKB when parsing malformed sctp ASCONF
chunks, from Daniel Borkmann.
12) Don't call sock_kfree_s() with NULL pointers, this function also has
the side effect of adjusting the socket memory usage. From Cong Wang.
* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net: (90 commits)
bna: fix skb->truesize underestimation
net: dsa: add includes for ethtool and phy_fixed definitions
openvswitch: Set flow-key members.
netrom: use linux/uaccess.h
dsa: Fix conversion from host device to mii bus
tipc: fix bug in bundled buffer reception
ipv6: introduce tcp_v6_iif()
sfc: add support for skb->xmit_more
r8152: return -EBUSY for runtime suspend
ipv4: fix a potential use after free in fou.c
ipv4: fix a potential use after free in ip_tunnel_core.c
hyperv: Add handling of IP header with option field in netvsc_set_hash()
openvswitch: Create right mask with disabled megaflows
vxlan: fix a free after use
openvswitch: fix a use after free
ipv4: dst_entry leak in ip_send_unicast_reply()
ipv4: clean up cookie_v4_check()
ipv4: share tcp_v4_save_options() with cookie_v4_check()
ipv4: call __ip_options_echo() in cookie_v4_check()
atm: simplify lanai.c by using module_pci_driver
...
Pull percpu consistent-ops changes from Tejun Heo:
"Way back, before the current percpu allocator was implemented, static
and dynamic percpu memory areas were allocated and handled separately
and had their own accessors. The distinction has been gone for many
years now; however, the now duplicate two sets of accessors remained
with the pointer based ones - this_cpu_*() - evolving various other
operations over time. During the process, we also accumulated other
inconsistent operations.
This pull request contains Christoph's patches to clean up the
duplicate accessor situation. __get_cpu_var() uses are replaced with
with this_cpu_ptr() and __this_cpu_ptr() with raw_cpu_ptr().
Unfortunately, the former sometimes is tricky thanks to C being a bit
messy with the distinction between lvalues and pointers, which led to
a rather ugly solution for cpumask_var_t involving the introduction of
this_cpu_cpumask_var_ptr().
This converts most of the uses but not all. Christoph will follow up
with the remaining conversions in this merge window and hopefully
remove the obsolete accessors"
* 'for-3.18-consistent-ops' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/percpu: (38 commits)
irqchip: Properly fetch the per cpu offset
percpu: Resolve ambiguities in __get_cpu_var/cpumask_var_t -fix
ia64: sn_nodepda cannot be assigned to after this_cpu conversion. Use __this_cpu_write.
percpu: Resolve ambiguities in __get_cpu_var/cpumask_var_t
Revert "powerpc: Replace __get_cpu_var uses"
percpu: Remove __this_cpu_ptr
clocksource: Replace __this_cpu_ptr with raw_cpu_ptr
sparc: Replace __get_cpu_var uses
avr32: Replace __get_cpu_var with __this_cpu_write
blackfin: Replace __get_cpu_var uses
tile: Use this_cpu_ptr() for hardware counters
tile: Replace __get_cpu_var uses
powerpc: Replace __get_cpu_var uses
alpha: Replace __get_cpu_var
ia64: Replace __get_cpu_var uses
s390: cio driver &__get_cpu_var replacements
s390: Replace __get_cpu_var uses
mips: Replace __get_cpu_var uses
MIPS: Replace __get_cpu_var uses in FPU emulator.
arm: Replace __this_cpu_ptr with raw_cpu_ptr
...
TCP Small queues tries to keep number of packets in qdisc
as small as possible, and depends on a tasklet to feed following
packets at TX completion time.
Choice of tasklet was driven by latencies requirements.
Then, TCP stack tries to avoid reorders, by locking flows with
outstanding packets in qdisc in a given TX queue.
What can happen is that many flows get attracted by a low performing
TX queue, and cpu servicing TX completion has to feed packets for all of
them, making this cpu 100% busy in softirq mode.
This became particularly visible with latest skb->xmit_more support
Strategy adopted in this patch is to detect when tcp_wfree() is called
from ksoftirqd and let the outstanding queue for this flow being drained
before feeding additional packets, so that skb->ooo_okay can be set
to allow select_queue() to select the optimal queue :
Incoming ACKS are normally handled by different cpus, so this patch
gives more chance for these cpus to take over the burden of feeding
qdisc with future packets.
Tested:
lpaa23:~# ./super_netperf 1400 --google-pacing-rate 3028000 -H lpaa24 -l 3600 &
lpaa23:~# sar -n DEV 1 10 | grep eth1
06:16:18 AM eth1 595448.00 1190564.00 38381.09 1760253.12 0.00 0.00 1.00
06:16:19 AM eth1 594858.00 1189686.00 38340.76 1758952.72 0.00 0.00 0.00
06:16:20 AM eth1 597017.00 1194019.00 38480.79 1765370.29 0.00 0.00 1.00
06:16:21 AM eth1 595450.00 1190936.00 38380.19 1760805.05 0.00 0.00 0.00
06:16:22 AM eth1 596385.00 1193096.00 38442.56 1763976.29 0.00 0.00 1.00
06:16:23 AM eth1 598155.00 1195978.00 38552.97 1768264.60 0.00 0.00 0.00
06:16:24 AM eth1 594405.00 1188643.00 38312.57 1757414.89 0.00 0.00 1.00
06:16:25 AM eth1 593366.00 1187154.00 38252.16 1755195.83 0.00 0.00 0.00
06:16:26 AM eth1 593188.00 1186118.00 38232.88 1753682.57 0.00 0.00 1.00
06:16:27 AM eth1 596301.00 1192241.00 38440.94 1762733.09 0.00 0.00 0.00
Average: eth1 595457.30 1190843.50 38381.69 1760664.84 0.00 0.00 0.50
lpaa23:~# ./tc -s -d qd sh dev eth1 | grep backlog
backlog 7606336b 2513p requeues 167982
backlog 224072b 74p requeues 566
backlog 581376b 192p requeues 5598
backlog 181680b 60p requeues 1070
backlog 5305056b 1753p requeues 110166 // Here, this TX queue is attracting flows
backlog 157456b 52p requeues 1758
backlog 672216b 222p requeues 3025
backlog 60560b 20p requeues 24541
backlog 448144b 148p requeues 21258
lpaa23:~# echo 1 >/proc/sys/net/ipv4/tcp_tsq_enable_tcp_wfree_ksoftirqd_detect
Immediate jump to full bandwidth, and traffic is properly
shard on all tx queues.
lpaa23:~# sar -n DEV 1 10 | grep eth1
06:16:46 AM eth1 1397632.00 2795397.00 90081.87 4133031.26 0.00 0.00 1.00
06:16:47 AM eth1 1396874.00 2793614.00 90032.99 4130385.46 0.00 0.00 0.00
06:16:48 AM eth1 1395842.00 2791600.00 89966.46 4127409.67 0.00 0.00 1.00
06:16:49 AM eth1 1395528.00 2791017.00 89946.17 4126551.24 0.00 0.00 0.00
06:16:50 AM eth1 1397891.00 2795716.00 90098.74 4133497.39 0.00 0.00 1.00
06:16:51 AM eth1 1394951.00 2789984.00 89908.96 4125022.51 0.00 0.00 0.00
06:16:52 AM eth1 1394608.00 2789190.00 89886.90 4123851.36 0.00 0.00 1.00
06:16:53 AM eth1 1395314.00 2790653.00 89934.33 4125983.09 0.00 0.00 0.00
06:16:54 AM eth1 1396115.00 2792276.00 89984.25 4128411.21 0.00 0.00 1.00
06:16:55 AM eth1 1396829.00 2793523.00 90030.19 4130250.28 0.00 0.00 0.00
Average: eth1 1396158.40 2792297.00 89987.09 4128439.35 0.00 0.00 0.50
lpaa23:~# tc -s -d qd sh dev eth1 | grep backlog
backlog 7900052b 2609p requeues 173287
backlog 878120b 290p requeues 589
backlog 1068884b 354p requeues 5621
backlog 996212b 329p requeues 1088
backlog 984100b 325p requeues 115316
backlog 956848b 316p requeues 1781
backlog 1080996b 357p requeues 3047
backlog 975016b 322p requeues 24571
backlog 990156b 327p requeues 21274
(All 8 TX queues get a fair share of the traffic)
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
TCP Small Queues (tcp_tsq_handler()) can hold one reference on
sk->sk_wmem_alloc, preventing skb->ooo_okay being set.
We should relax test done to set skb->ooo_okay to take care
of this extra reference.
Minimal truesize of skb containing one byte of payload is
SKB_TRUESIZE(1)
Without this fix, we have more chance locking flows into the wrong
transmit queue.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Lets use a proper structure to clearly document and implement
skb fast clones.
Then, we might experiment more easily alternative layouts.
This patch adds a new skb_fclone_busy() helper, used by tcp and xfrm,
to stop leaking of implementation details.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Suggested by Stephen. Also drop inline keyword and let compiler decide.
gcc 4.7.3 decides to no longer inline tcp_ecn_check_ce, so split it up.
The actual evaluation is not inlined anymore while the ECN_OK test is.
Suggested-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
This work adds the DataCenter TCP (DCTCP) congestion control
algorithm [1], which has been first published at SIGCOMM 2010 [2],
resp. follow-up analysis at SIGMETRICS 2011 [3] (and also, more
recently as an informational IETF draft available at [4]).
DCTCP is an enhancement to the TCP congestion control algorithm for
data center networks. Typical data center workloads are i.e.
i) partition/aggregate (queries; bursty, delay sensitive), ii) short
messages e.g. 50KB-1MB (for coordination and control state; delay
sensitive), and iii) large flows e.g. 1MB-100MB (data update;
throughput sensitive). DCTCP has therefore been designed for such
environments to provide/achieve the following three requirements:
* High burst tolerance (incast due to partition/aggregate)
* Low latency (short flows, queries)
* High throughput (continuous data updates, large file
transfers) with commodity, shallow buffered switches
The basic idea of its design consists of two fundamentals: i) on the
switch side, packets are being marked when its internal queue
length > threshold K (K is chosen so that a large enough headroom
for marked traffic is still available in the switch queue); ii) the
sender/host side maintains a moving average of the fraction of marked
packets, so each RTT, F is being updated as follows:
F := X / Y, where X is # of marked ACKs, Y is total # of ACKs
alpha := (1 - g) * alpha + g * F, where g is a smoothing constant
The resulting alpha (iow: probability that switch queue is congested)
is then being used in order to adaptively decrease the congestion
window W:
W := (1 - (alpha / 2)) * W
The means for receiving marked packets resp. marking them on switch
side in DCTCP is the use of ECN.
RFC3168 describes a mechanism for using Explicit Congestion Notification
from the switch for early detection of congestion, rather than waiting
for segment loss to occur.
However, this method only detects the presence of congestion, not
the *extent*. In the presence of mild congestion, it reduces the TCP
congestion window too aggressively and unnecessarily affects the
throughput of long flows [4].
DCTCP, as mentioned, enhances Explicit Congestion Notification (ECN)
processing to estimate the fraction of bytes that encounter congestion,
rather than simply detecting that some congestion has occurred. DCTCP
then scales the TCP congestion window based on this estimate [4],
thus it can derive multibit feedback from the information present in
the single-bit sequence of marks in its control law. And thus act in
*proportion* to the extent of congestion, not its *presence*.
Switches therefore set the Congestion Experienced (CE) codepoint in
packets when internal queue lengths exceed threshold K. Resulting,
DCTCP delivers the same or better throughput than normal TCP, while
using 90% less buffer space.
It was found in [2] that DCTCP enables the applications to handle 10x
the current background traffic, without impacting foreground traffic.
Moreover, a 10x increase in foreground traffic did not cause any
timeouts, and thus largely eliminates TCP incast collapse problems.
The algorithm itself has already seen deployments in large production
data centers since then.
We did a long-term stress-test and analysis in a data center, short
summary of our TCP incast tests with iperf compared to cubic:
This test measured DCTCP throughput and latency and compared it with
CUBIC throughput and latency for an incast scenario. In this test, 19
senders sent at maximum rate to a single receiver. The receiver simply
ran iperf -s.
The senders ran iperf -c <receiver> -t 30. All senders started
simultaneously (using local clocks synchronized by ntp).
This test was repeated multiple times. Below shows the results from a
single test. Other tests are similar. (DCTCP results were extremely
consistent, CUBIC results show some variance induced by the TCP timeouts
that CUBIC encountered.)
For this test, we report statistics on the number of TCP timeouts,
flow throughput, and traffic latency.
1) Timeouts (total over all flows, and per flow summaries):
CUBIC DCTCP
Total 3227 25
Mean 169.842 1.316
Median 183 1
Max 207 5
Min 123 0
Stddev 28.991 1.600
Timeout data is taken by measuring the net change in netstat -s
"other TCP timeouts" reported. As a result, the timeout measurements
above are not restricted to the test traffic, and we believe that it
is likely that all of the "DCTCP timeouts" are actually timeouts for
non-test traffic. We report them nevertheless. CUBIC will also include
some non-test timeouts, but they are drawfed by bona fide test traffic
timeouts for CUBIC. Clearly DCTCP does an excellent job of preventing
TCP timeouts. DCTCP reduces timeouts by at least two orders of
magnitude and may well have eliminated them in this scenario.
2) Throughput (per flow in Mbps):
CUBIC DCTCP
Mean 521.684 521.895
Median 464 523
Max 776 527
Min 403 519
Stddev 105.891 2.601
Fairness 0.962 0.999
Throughput data was simply the average throughput for each flow
reported by iperf. By avoiding TCP timeouts, DCTCP is able to
achieve much better per-flow results. In CUBIC, many flows
experience TCP timeouts which makes flow throughput unpredictable and
unfair. DCTCP, on the other hand, provides very clean predictable
throughput without incurring TCP timeouts. Thus, the standard deviation
of CUBIC throughput is dramatically higher than the standard deviation
of DCTCP throughput.
Mean throughput is nearly identical because even though cubic flows
suffer TCP timeouts, other flows will step in and fill the unused
bandwidth. Note that this test is something of a best case scenario
for incast under CUBIC: it allows other flows to fill in for flows
experiencing a timeout. Under situations where the receiver is issuing
requests and then waiting for all flows to complete, flows cannot fill
in for timed out flows and throughput will drop dramatically.
3) Latency (in ms):
CUBIC DCTCP
Mean 4.0088 0.04219
Median 4.055 0.0395
Max 4.2 0.085
Min 3.32 0.028
Stddev 0.1666 0.01064
Latency for each protocol was computed by running "ping -i 0.2
<receiver>" from a single sender to the receiver during the incast
test. For DCTCP, "ping -Q 0x6 -i 0.2 <receiver>" was used to ensure
that traffic traversed the DCTCP queue and was not dropped when the
queue size was greater than the marking threshold. The summary
statistics above are over all ping metrics measured between the single
sender, receiver pair.
The latency results for this test show a dramatic difference between
CUBIC and DCTCP. CUBIC intentionally overflows the switch buffer
which incurs the maximum queue latency (more buffer memory will lead
to high latency.) DCTCP, on the other hand, deliberately attempts to
keep queue occupancy low. The result is a two orders of magnitude
reduction of latency with DCTCP - even with a switch with relatively
little RAM. Switches with larger amounts of RAM will incur increasing
amounts of latency for CUBIC, but not for DCTCP.
4) Convergence and stability test:
This test measured the time that DCTCP took to fairly redistribute
bandwidth when a new flow commences. It also measured DCTCP's ability
to remain stable at a fair bandwidth distribution. DCTCP is compared
with CUBIC for this test.
At the commencement of this test, a single flow is sending at maximum
rate (near 10 Gbps) to a single receiver. One second after that first
flow commences, a new flow from a distinct server begins sending to
the same receiver as the first flow. After the second flow has sent
data for 10 seconds, the second flow is terminated. The first flow
sends for an additional second. Ideally, the bandwidth would be evenly
shared as soon as the second flow starts, and recover as soon as it
stops.
The results of this test are shown below. Note that the flow bandwidth
for the two flows was measured near the same time, but not
simultaneously.
DCTCP performs nearly perfectly within the measurement limitations
of this test: bandwidth is quickly distributed fairly between the two
flows, remains stable throughout the duration of the test, and
recovers quickly. CUBIC, in contrast, is slow to divide the bandwidth
fairly, and has trouble remaining stable.
CUBIC DCTCP
Seconds Flow 1 Flow 2 Seconds Flow 1 Flow 2
0 9.93 0 0 9.92 0
0.5 9.87 0 0.5 9.86 0
1 8.73 2.25 1 6.46 4.88
1.5 7.29 2.8 1.5 4.9 4.99
2 6.96 3.1 2 4.92 4.94
2.5 6.67 3.34 2.5 4.93 5
3 6.39 3.57 3 4.92 4.99
3.5 6.24 3.75 3.5 4.94 4.74
4 6 3.94 4 5.34 4.71
4.5 5.88 4.09 4.5 4.99 4.97
5 5.27 4.98 5 4.83 5.01
5.5 4.93 5.04 5.5 4.89 4.99
6 4.9 4.99 6 4.92 5.04
6.5 4.93 5.1 6.5 4.91 4.97
7 4.28 5.8 7 4.97 4.97
7.5 4.62 4.91 7.5 4.99 4.82
8 5.05 4.45 8 5.16 4.76
8.5 5.93 4.09 8.5 4.94 4.98
9 5.73 4.2 9 4.92 5.02
9.5 5.62 4.32 9.5 4.87 5.03
10 6.12 3.2 10 4.91 5.01
10.5 6.91 3.11 10.5 4.87 5.04
11 8.48 0 11 8.49 4.94
11.5 9.87 0 11.5 9.9 0
SYN/ACK ECT test:
This test demonstrates the importance of ECT on SYN and SYN-ACK packets
by measuring the connection probability in the presence of competing
flows for a DCTCP connection attempt *without* ECT in the SYN packet.
The test was repeated five times for each number of competing flows.
Competing Flows 1 | 2 | 4 | 8 | 16
------------------------------
Mean Connection Probability 1 | 0.67 | 0.45 | 0.28 | 0
Median Connection Probability 1 | 0.65 | 0.45 | 0.25 | 0
As the number of competing flows moves beyond 1, the connection
probability drops rapidly.
Enabling DCTCP with this patch requires the following steps:
DCTCP must be running both on the sender and receiver side in your
data center, i.e.:
sysctl -w net.ipv4.tcp_congestion_control=dctcp
Also, ECN functionality must be enabled on all switches in your
data center for DCTCP to work. The default ECN marking threshold (K)
heuristic on the switch for DCTCP is e.g., 20 packets (30KB) at
1Gbps, and 65 packets (~100KB) at 10Gbps (K > 1/7 * C * RTT, [4]).
In above tests, for each switch port, traffic was segregated into two
queues. For any packet with a DSCP of 0x01 - or equivalently a TOS of
0x04 - the packet was placed into the DCTCP queue. All other packets
were placed into the default drop-tail queue. For the DCTCP queue,
RED/ECN marking was enabled, here, with a marking threshold of 75 KB.
More details however, we refer you to the paper [2] under section 3).
There are no code changes required to applications running in user
space. DCTCP has been implemented in full *isolation* of the rest of
the TCP code as its own congestion control module, so that it can run
without a need to expose code to the core of the TCP stack, and thus
nothing changes for non-DCTCP users.
Changes in the CA framework code are minimal, and DCTCP algorithm
operates on mechanisms that are already available in most Silicon.
The gain (dctcp_shift_g) is currently a fixed constant (1/16) from
the paper, but we leave the option that it can be chosen carefully
to a different value by the user.
In case DCTCP is being used and ECN support on peer site is off,
DCTCP falls back after 3WHS to operate in normal TCP Reno mode.
ss {-4,-6} -t -i diag interface:
... dctcp wscale:7,7 rto:203 rtt:2.349/0.026 mss:1448 cwnd:2054
ssthresh:1102 ce_state 0 alpha 15 ab_ecn 0 ab_tot 735584
send 10129.2Mbps pacing_rate 20254.1Mbps unacked:1822 retrans:0/15
reordering:101 rcv_space:29200
... dctcp-reno wscale:7,7 rto:201 rtt:0.711/1.327 ato:40 mss:1448
cwnd:10 ssthresh:1102 fallback_mode send 162.9Mbps pacing_rate
325.5Mbps rcv_rtt:1.5 rcv_space:29200
More information about DCTCP can be found in [1-4].
[1] http://simula.stanford.edu/~alizade/Site/DCTCP.html
[2] http://simula.stanford.edu/~alizade/Site/DCTCP_files/dctcp-final.pdf
[3] http://simula.stanford.edu/~alizade/Site/DCTCP_files/dctcp_analysis-full.pdf
[4] http://tools.ietf.org/html/draft-bensley-tcpm-dctcp-00
Joint work with Florian Westphal and Glenn Judd.
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Glenn Judd <glenn.judd@morganstanley.com>
Acked-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
DataCenter TCP (DCTCP) determines cwnd growth based on ECN information
and ACK properties, e.g. ACK that updates window is treated differently
than DUPACK.
Also DCTCP needs information whether ACK was delayed ACK. Furthermore,
DCTCP also implements a CE state machine that keeps track of CE markings
of incoming packets.
Therefore, extend the congestion control framework to provide these
event types, so that DCTCP can be properly implemented as a normal
congestion algorithm module outside of the core stack.
Joint work with Daniel Borkmann and Glenn Judd.
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Signed-off-by: Glenn Judd <glenn.judd@morganstanley.com>
Acked-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch adds a flag to TCP congestion algorithms that allows
for requesting to mark IPv4/IPv6 sockets with transport as ECN
capable, that is, ECT(0), when required by a congestion algorithm.
It is currently used and needed in DataCenter TCP (DCTCP), as it
requires both peers to assert ECT on all IP packets sent - it
uses ECN feedback (i.e. CE, Congestion Encountered information)
from switches inside the data center to derive feedback to the
end hosts.
Therefore, simply add a new flag to icsk_ca_ops. Note that DCTCP's
algorithm/behaviour slightly diverges from RFC3168, therefore this
is only (!) enabled iff the assigned congestion control ops module
has requested this. By that, we can tightly couple this logic really
only to the provided congestion control ops.
Joint work with Florian Westphal and Glenn Judd.
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Glenn Judd <glenn.judd@morganstanley.com>
Acked-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Our goal is to access no more than one cache line access per skb in
a write or receive queue when doing the various walks.
After recent TCP_SKB_CB() reorganizations, it is almost done.
Last part is tcp_skb_pcount() which currently uses
skb_shinfo(skb)->gso_segs, which is a terrible choice, because it needs
3 cache lines in current kernel (skb->head, skb->end, and
shinfo->gso_segs are all in 3 different cache lines, far from skb->cb)
This very simple patch reuses space currently taken by tcp_tw_isn
only in input path, as tcp_skb_pcount is only needed for skb stored in
write queue.
This considerably speeds up tcp_ack(), granted we avoid shinfo->tx_flags
to get SKBTX_ACK_TSTAMP, which seems possible.
This also speeds up all sack processing in general.
This speeds up tcp_sendmsg() because it no longer has to access/dirty
shinfo.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
TCP maintains lists of skb in write queue, and in receive queues
(in order and out of order queues)
Scanning these lists both in input and output path usually requires
access to skb->next, TCP_SKB_CB(skb)->seq, and TCP_SKB_CB(skb)->end_seq
These fields are currently in two different cache lines, meaning we
waste lot of memory bandwidth when these queues are big and flows
have either packet drops or packet reorders.
We can move TCP_SKB_CB(skb)->header at the end of TCP_SKB_CB, because
this header is not used in fast path. This allows TCP to search much faster
in the skb lists.
Even with regular flows, we save one cache line miss in fast path.
Thanks to Christoph Paasch for noticing we need to cleanup
skb->cb[] (IPCB/IP6CB) before entering IP stack in tx path,
and that I forgot IPCB use in tcp_v4_hnd_req() and tcp_v4_save_options().
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
While profiling TCP stack, I noticed one useless atomic operation
in tcp_sendmsg(), caused by skb_header_release().
It turns out all current skb_header_release() users have a fresh skb,
that no other user can see, so we can avoid one atomic operation.
Introduce __skb_header_release() to clearly document this.
This gave me a 1.5 % improvement on TCP_RR workload.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
icsk_rto is a 32bit field, and icsk_backoff can reach 15 by default,
or more if some sysctl (eg tcp_retries2) are changed.
Better use 64bit to perform icsk_rto << icsk_backoff operations
As Joe Perches suggested, add a helper for this.
Yuchung spotted the tcp_v4_err() case.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The TCP_SKB_CB(skb)->when field no longer exists as of recent change
7faee5c0d5 ("tcp: remove TCP_SKB_CB(skb)->when"). And in any case,
tcp_fragment() is called on already-transmitted packets from the
__tcp_retransmit_skb() call site, so copying timestamps of any kind
in this spot is quite sensible.
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Reported-by: Yuchung Cheng <ycheng@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
After commit 740b0f1841 ("tcp: switch rtt estimations to usec resolution"),
we no longer need to maintain timestamps in two different fields.
TCP_SKB_CB(skb)->when can be removed, as same information sits in skb_mstamp.stamp_jiffies
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Replace uses of get_cpu_var for address calculation through this_cpu_ptr.
Cc: netdev@vger.kernel.org
Cc: Eric Dumazet <edumazet@google.com>
Acked-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Christoph Lameter <cl@linux.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Eric Dumazet