We tried very hard to remove all possible dev_hold()/dev_put() pairs in
network stack, using RCU conversions.
There is still an unavoidable device refcount change for every dst we
create/destroy, and this can slow down some workloads (routers or some
app servers, mmap af_packet)
We can switch to a percpu refcount implementation, now dynamic per_cpu
infrastructure is mature. On a 64 cpus machine, this consumes 256 bytes
per device.
On x86, dev_hold(dev) code :
before
lock incl 0x280(%ebx)
after:
movl 0x260(%ebx),%eax
incl fs:(%eax)
Stress bench :
(Sending 160.000.000 UDP frames,
IP route cache disabled, dual E5540 @2.53GHz,
32bit kernel, FIB_TRIE)
Before:
real 1m1.662s
user 0m14.373s
sys 12m55.960s
After:
real 0m51.179s
user 0m15.329s
sys 10m15.942s
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The HW by default has RX coalescing on. For iWARP connections, this
causes a 100ms delay in connection establishement due to the ingress
MPA Start message being stalled in HW. So explicitly turn RX
coalescing off when setting up iWARP connections.
This was causing very bad performance for NP64 gather operations using
Open MPI, due to the way it sets up connections on larger jobs.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Cc: <stable@kernel.org>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Changing state to CLOSING when FIN is received causes A0 cards to
hang. Fix this by checking for A0 cards in FIN handling.
Signed-off-by: Faisal Latif <faisal.latif@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
When the driver receives an AE for FIN received, it closes the
connection without changing the state of the connection in the
hardware to closing. By changing the state to closing, hardware will
do a normal close sequence.
Signed-off-by: Faisal Latif <faisal.latif@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
During a stress testing in a large cluster, multiple close event are
detected and BUG() is hit in the iWARP core. The cause is that the
active node gave up while waiting for an MPA response from the peer
and tried to close the connection by sending RST. The passive node
driver receives the RST but is waiting for MPA response from the user.
When the MPA accept is received, the driver offloads the connection
and sends a CLOSE event. The driver gets an AE indicating RESET
received and also sends a CLOSE event, hitting a BUG().
Fix this by correcting RESET handling and sending CLOSE events.
Signed-off-by: Faisal Latif <faisal.latif@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Setting TX pause param writes to the wrong register location causing
the adapter to hang. Correct the define used to write the reigster.
Addresses: https://bugs.openfabrics.org/show_bug.cgi?id=2116
Reported-by: Shiri Franchi <shirif@voltaire.com>
Signed-off-by: Chien Tung <chien.tin.tung@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
The max depth supported by T3 is 64K entries. This fixes a bug
introduced in commit 9918b28d ("RDMA/cxgb3: Increase the max CQ
depth") that causes stalls and possibly crashes in large MPI clusters.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Cc: <stable@kernel.org>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/roland/infiniband: (42 commits)
IB/qib: Add missing <linux/slab.h> include
IB/ehca: Drop unnecessary NULL test
RDMA/nes: Fix confusing if statement indentation
IB/ehca: Init irq tasklet before irq can happen
RDMA/nes: Fix misindented code
RDMA/nes: Fix showing wqm_quanta
RDMA/nes: Get rid of "set but not used" variables
RDMA/nes: Read firmware version from correct place
IB/srp: Export req_lim via sysfs
IB/srp: Make receive buffer handling more robust
IB/srp: Use print_hex_dump()
IB: Rename RAW_ETY to RAW_ETHERTYPE
RDMA/nes: Fix two sparse warnings
RDMA/cxgb3: Make needlessly global iwch_l2t_send() static
IB/iser: Make needlessly global iser_alloc_rx_descriptors() static
RDMA/cxgb4: Add timeouts when waiting for FW responses
IB/qib: Fix race between qib_error_qp() and receive packet processing
IB/qib: Limit the number of packets processed per interrupt
IB/qib: Allow writes to the diag_counters to be able to clear them
IB/qib: Set cfgctxts to number of CPUs by default
...
of_device is just an alias for platform_device, so remove it entirely. Also
replace to_of_device() with to_platform_device() and update comment blocks.
This patch was initially generated from the following semantic patch, and then
edited by hand to pick up the bits that coccinelle didn't catch.
@@
@@
-struct of_device
+struct platform_device
Signed-off-by: Grant Likely <grant.likely@secretlab.ca>
Reviewed-by: David S. Miller <davem@davemloft.net>
Fix build failure on sparc64 which is missing the include of
<linux/slab.h> via <asm/pci.h> that x86, powerpc, ia64, etc. have.
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
list_for_each_entry binds its first argument to a non-null value, and thus
any null test on the value of that argument is superfluous.
The semantic patch that makes this change is as follows:
(http://coccinelle.lip6.fr/)
// <smpl>
@@
iterator I;
expression x;
statement S,S1,S2;
@@
I(x,...) { <...
- if (x == NULL && ...) S
...> }
// </smpl>
Signed-off-by: Julia Lawall <julia@diku.dk>
Acked-by: Alexander Schmidt <alexs@linux.vnet.ibm.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Fix confusing indentation that makes a statement look as if it's part of
an if statement when in fact it isn't.
Reported-by: Julia Lawall <julia@diku.dk>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Initialize tasklet before interrupts are requested to prevent
scheduling of an uninitialized tasklet.
Signed-off-by: Alexander Schmidt <alexs@linux.vnet.ibm.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Export req_lim via sysfs for debugging.
Signed-off-by: Bart Van Assche <bart.vanassche@gmail.com>
Acked-by: David Dillow <dave@thedillows.org>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-next-2.6: (1443 commits)
phy/marvell: add 88ec048 support
igb: Program MDICNFG register prior to PHY init
e1000e: correct MAC-PHY interconnect register offset for 82579
hso: Add new product ID
can: Add driver for esd CAN-USB/2 device
l2tp: fix export of header file for userspace
can-raw: Fix skb_orphan_try handling
Revert "net: remove zap_completion_queue"
net: cleanup inclusion
phy/marvell: add 88e1121 interface mode support
u32: negative offset fix
net: Fix a typo from "dev" to "ndev"
igb: Use irq_synchronize per vector when using MSI-X
ixgbevf: fix null pointer dereference due to filter being set for VLAN 0
e1000e: Fix irq_synchronize in MSI-X case
e1000e: register pm_qos request on hardware activation
ip_fragment: fix subtracting PPPOE_SES_HLEN from mtu twice
net: Add getsockopt support for TCP thin-streams
cxgb4: update driver version
cxgb4: add new PCI IDs
...
Manually fix up conflicts in:
- drivers/net/e1000e/netdev.c: due to pm_qos registration
infrastructure changes
- drivers/net/phy/marvell.c: conflict between adding 88ec048 support
and cleaning up the IDs
- drivers/net/wireless/ipw2x00/ipw2100.c: trivial ipw2100_pm_qos_req
conflict (registration change vs marking it static)
The current strategy in ib_srp for posting receive buffers is:
* Post one buffer after channel establishment.
* Post one buffer before sending an SRP_CMD or SRP_TSK_MGMT to the target.
As a result, only the first non-SRP_RSP information unit from the
target will be processed. If that first information unit is an
SRP_T_LOGOUT, it will be processed. On the other hand, if the
initiator receives an SRP_CRED_REQ or SRP_AER_REQ before it receives a
SRP_T_LOGOUT, the SRP_T_LOGOUT won't be processed.
We can fix this inconsistency by changing the strategy for posting
receive buffers to:
* Post all receive buffers after channel establishment.
* After a receive buffer has been consumed and processed, post it again.
A side effect is that the ib_post_recv() call is moved out of the SCSI
command processing path. Since __srp_post_recv() is not called
directly any more, get rid of it and move the code directly into
srp_post_recv(). Also, move srp_post_recv() up in the file to avoid a
forward declaration.
Signed-off-by: Bart Van Assche <bart.vanassche@gmail.com>
Acked-by: David Dillow <dave@thedillows.org>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Replace an open-coded dump of the receive buffer with a call to
print_hex_dump().
Signed-off-by: Bart Van Assche <bart.vanassche@gmail.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Change abbreviated IB_QPT_RAW_ETY to IB_QPT_RAW_ETHERTYPE to make
the special QP type easier to understand.
cf http://www.mail-archive.com/linux-rdma@vger.kernel.org/msg04530.html
Signed-off-by: Aleksey Senin <alekseys@voltaire.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Simple changes to fix warnings:
CHECK drivers/infiniband/hw/nes/nes_verbs.c
nes_verbs.c:1944:45: warning: Using plain integer as NULL pointer
nes_verbs.c:1944:48: warning: Using plain integer as NULL pointer
CHECK drivers/infiniband/hw/nes/nes_cm.c
nes_cm.c:2645:43: warning: mixing different enum types
nes_cm.c:2645:43: int enum iw_cm_event_type versus
nes_cm.c:2645:43: int enum iw_cm_event_status
Signed-off-by: Or Gerlitz <ogerlitz@voltaire.com>
Acked-by: Chien Tung <chien.tin.tung@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Don't hang a host thread if the FW stops responding.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
When transitioning a QP to the error state, in progress RWQEs need to
be marked complete. This also involves releasing the reference count
to the memory regions referenced in the SGEs. The locking in the
receive packet processing wasn't sufficient to prevent qib_error_qp()
from modifying the r_sge state at the same time, thus leading to
kernel panics.
Signed-off-by: Ralph Campbell <ralph.campbell@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Don't processes too many packets without allowing other IRQ functions
a chance to run. Otherwise, there is a chance of getting a "soft
lockup" messages and poor application response times.
Signed-off-by: Ralph Campbell <ralph.campbell@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Up to now, we have set the number of available user contexts based on
the number of hardware contexts which is set according to the number
of available CPUs. This was fine since most CPUs had a power of two
number of cores and the chip supported 4, 8, or 16 user contexts. Now
that some systems have 12 cores, the default isn't optimal and should
be set to 12 even though 16 hardware contexts need to be enabled.
Signed-off-by: Ralph Campbell <ralph.campbell@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Endpoint timer manipulation needs to be done inside the lock. Otherwise
we can get into a situation where a timer is stopped before it is started,
which hits the WARN_ON() in stop_ep_timer().
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
There is only one control txq per tx channel. So use the port number
as the queue index when sending.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
There exists a race condition where the app disconnects, which
initiates an orderly close (via rdma_fini()), concurrently with an
ingress abort condition, which initiates an abortive close operation.
Since rdma_fini() must be called without IRQs disabled, the fini can
be called after the QP has been transitioned to ERROR. This is ok,
but we need to protect against qp->ep getting NULLed.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
NULL pointer dereferences in ib_cm_init_qp_attr() were seen by some
users. From a crash dump, I determined that we died in
cm_init_qp_rts_attr() (it's inlined, so it doesn't show up in the
traceback) on the line labeled below:
static int cm_init_qp_rts_attr(struct cm_id_private *cm_id_priv,
struct ib_qp_attr *qp_attr,
int *qp_attr_mask)
{
........
if (cm_id_priv->id.lap_state == IB_CM_LAP_UNINIT) {
.....
} else {
*qp_attr_mask = IB_QP_ALT_PATH | IB_QP_PATH_MIG_STATE;
qp_attr->alt_port_num = cm_id_priv->alt_av.port->port_num; <-die
The problem is that the rdma_cm can call ib_send_cm_mra() after a
connection has been established. The ib_cm incorrectly assumes that
the MRA is in response to a LAP (load alternate path) message, even
though no LAP message has been received. The ib_cm needs to check the
lap_state before sending an MRA if the cm_id state is established.
Reported-by: Arthur Kepner <akepner@sgi.com>
Reported-by: Josh England <jjengla@gmail.com>
Signed-off-by: Sean Hefty <sean.hefty@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
When ib_unregister_device() is called from netdev stop during ifdown,
it sometimes hangs. Changes made to indicate port_err to ib_dispatch_event()
during netdev stop and port_active during netdev open. The
ib_unregister_device() is only called during remove of the module.
Signed-off-by: Faisal Latif <faisal.latif@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Read and print eeprom version and save it off for later use.
Also delete a tab.
Signed-off-by: Chien Tung <chien.tin.tung@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
This patch converts pci_table entries, where .subvendor=PCI_ANY_ID and
.subdevice=PCI_ANY_ID, .class=0 and .class_mask=0, to use the
PCI_VDEVICE macro, and thus improves readability.
Signed-off-by: Peter Huewe <peterhuewe@gmx.de>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
When ioremap() fails with a NULL pointer, catch the error and pass it
to the caller of create_qp() or create_cq() instead of trying to
dereference the NULL pointer later on.
Signed-off-by: Alexander Schmidt <alexs@linux.vnet.ibm.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
We used to allow only full specification, or using all contexts within
an HCA before moving to the next HCA. We now allow an additional
method -- round-robining through HCAs -- and make that the default.
Signed-off-by: Dave Olson <dave.olson@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Turn off IB latency mode. This improves link quality for slower
process chips.
Signed-off-by: Ralph Campbell <ralph.campbell@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
When the default llseek action gets changed to no_llseek, all file
systems relying on the current behaviour need to set explicit .llseek
operations.
In case of qib_fs, we want the files to be seekable, so
generic_file_llseek fits best.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
T4 EQ entries are in multiples of 64 bytes. Currently the RDMA SQ and
RQ use fixed sized entries composed of 4 EQ entries for the SQ and 2
EQ entries for the RQ. For optimial latency with small IO, we need to
change this so the HW only needs to DMA the EQ entries actually used
by a given work request.
Implementation:
- add wq_pidx counter to track where we are in the EQ. cidx/pidx are
used for the sw sq/rq tracking and flow control.
- the variable part of work requests is the SGL. Add new functions to
build the SGL and/or immediate data directly in the EQ memory
wrapping when needed.
- adjust the min burst size for the EQ contexts to 64B.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Q_FREECNT() returns the number of spaces free. This should never be a
negative amount. Also the num_wrs is an unsigned int so it can never
be less than zero.
Signed-off-by: Dan Carpenter <error27@gmail.com>
Acked-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Randy Dunlap