Xilinx Zynq pl330 dma driver has 9 irqs which all have to
be used by the driver to get it work properly.
Signed-off-by: Michal Simek <michal.simek@xilinx.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
The IB spec does not guarantee that the opcode is available in error
completions. Hence do not rely on it. See also commit 948d1e889e
("IB/srp: Introduce srp_handle_qp_err()").
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Cc: <stable@vger.kernel.org> # v3.8
Signed-off-by: Roland Dreier <roland@purestorage.com>
If SCSI commands are submitted with a SCSI request timeout that is
lower than the the IB RC timeout, it can happen that the SCSI error
handler has already started device recovery before transport layer
error handling starts. So it can happen that the SCSI error handler
tries to abort a SCSI command after it has been reset by
srp_rport_reconnect().
Tell the SCSI error handler that such commands have finished and that
it is not necessary to continue its recovery strategy for commands
that have been reset by srp_rport_reconnect().
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Cc: <stable@vger.kernel.org>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Remove an SRP target from the SRP target list before invoking the last
scsi_host_put() call. This change is necessary because that last put
frees the memory that holds the srp_target_port structure.
This patch prevents the following kernel oops:
RIP: 0010:[<ffffffff810b00d0>] __lock_acquire+0x500/0x1570
Call Trace:
[<ffffffff810b11e4>] lock_acquire+0xa4/0x120
[<ffffffff81531206>] _spin_lock+0x36/0x70
[<ffffffffa01b6d8f>] srp_remove_work+0xef/0x180 [ib_srp]
[<ffffffff8109125c>] worker_thread+0x21c/0x3d0
[<ffffffff81096e86>] kthread+0x96/0xa0
[<ffffffff8100c20a>] child_rip+0xa/0x20
Signed-off-by: Vu Pham <vuhuong@mellanox.com>
[ bvanassche - Modified path description and CC'ed stable. ]
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Cc: <stable@vger.kernel.org>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Currently, it's not possible to change queue depth for a device behind
SRP host. Sometimes, we need to adjust queue_depth for performance
reason (eg storage busy, we need lower queue_depth to avoid running
into SCSI error handler), so this patch add support for SRP driver.
Signed-off-by: Jack Wang <jinpu.wang@profitbricks.com>
Tested-by: Bart Van Assche <bvanassche@acm.org>
Acked-by: David Dillow <dillowda@ornl.gov>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Certain storage configurations, e.g. a sufficiently large array of
hard disks in a RAID configuration, need a queue depth above 64 to
achieve optimal performance. Hence make the queue depth configurable.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Acked-by: David Dillow <dillowda@ornl.gov>
Tested-by: Jack Wang <xjtuwjp@gmail.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
This patch does not change any functionality.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Acked-by: David Dillow <dillowda@ornl.gov>
Cc: Roland Dreier <roland@purestorage.com>
Cc: Vu Pham <vu@mellanox.com>
Cc: Sebastian Riemer <sebastian.riemer@profitbricks.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
On an initiator system with multiple IB ports it is not yet possible
to figure out what the originating port of an SRP connection is. Hence
make the source GID available in sysfs.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Acked-by: David Dillow <dillowda@ornl.gov>
Signed-off-by: Roland Dreier <roland@purestorage.com>
After a transport layer occurred, periodically try to reconnect
to the target until the dev_loss timer expires. Protect the
callback functions that can be invoked from inside the SCSI EH
against concurrent invocation with srp_reconnect_rport() via the
rport mutex. Change the default dev_loss_tmo from 60s into 600s
to give the reconnect mechanism a chance to kick in.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Acked-by: David Dillow <dillowda@ornl.gov>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Add support for periodically reconnecting to an SRP target until
the dev_loss timer expires. After the tenth reconnection attempt,
gradually slow down subsequent reconnect attempts.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Acked-by: David Dillow <dillowda@ornl.gov>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Start the reconnect timer, fast_io_fail timer and dev_loss timers if a
transport layer error occurs.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Acked-by: David Dillow <dillowda@ornl.gov>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Enable fast_io_fail_tmo and dev_loss_tmo functionality for the IB SRP
initiator. Add kernel module parameters that allow to specify default
values for these parameters.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Acked-by: David Dillow <dillowda@ornl.gov>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Add the necessary functions in the SRP transport module to allow an
SRP initiator driver to implement transport layer error handling
similar to the functionality already provided by the FC transport
layer. This includes:
- Support for implementing fast_io_fail_tmo, the time that should
elapse after having detected a transport layer problem and
before failing I/O.
- Support for implementing dev_loss_tmo, the time that should
elapse after having detected a transport layer problem and
before removing a remote port.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Acked-by: David Dillow <dillowda@ornl.gov>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Keep the rport data structure around after srp_remove_host() has
finished until cleanup of the IB transport layer has finished
completely. This is necessary because later patches use the rport
pointer inside the queuecommand callback. Without this patch
accessing the rport from inside a queuecommand callback is racy
because srp_remove_host() must be invoked before scsi_remove_host()
and because the queuecommand callback could get invoked after
srp_remove_host() has finished. In other words, without this patch
the queuecommand callback can get invoked after the rport data
structure has been freed.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Acked-by: David Dillow <dillowda@ornl.gov>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Allow the InfiniBand RC retry count to be configured by the user as an
option in the target login string. Reducing this retry count allows to
reduce the path failover time.
Signed-off-by: Vu Pham <vu@mellanox.com>
[ bvanassche: Rewrote patch description / changed default retry count ]
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Acked-by: David Dillow <dillowda@ornl.gov>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Commit 7fac33014f54("IB/qib: checkpatch fixes") was overzealous in
removing a simple_strtoul for a parse routine, setup_txselect(). That
routine is required to handle a multi-value string.
Unwind that aspect of the fix.
Cc: <stable@vger.kernel.org>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Convert __attribute__ ((packed)) to __packed.
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
qib_user_sdma_queue_pkts() gets called with mmap_sem held for
writing. Except for get_user_pages() deep down in
qib_user_sdma_pin_pages() we don't seem to need mmap_sem at all. Even
more interestingly the function qib_user_sdma_queue_pkts() (and also
qib_user_sdma_coalesce() called somewhat later) call copy_from_user()
which can hit a page fault and we deadlock on trying to get mmap_sem
when handling that fault.
So just make qib_user_sdma_pin_pages() use get_user_pages_fast() and
leave mmap_sem locking for mm.
This deadlock has actually been observed in the wild when the node
is under memory pressure.
Cc: <stable@vger.kernel.org>
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Roland Dreier <roland@purestorage.com>
ipath_user_sdma_queue_pkts() gets called with mmap_sem held for
writing. Except for get_user_pages() deep down in
ipath_user_sdma_pin_pages() we don't seem to need mmap_sem at all.
Even more interestingly the function ipath_user_sdma_queue_pkts() (and
also ipath_user_sdma_coalesce() called somewhat later) call
copy_from_user() which can hit a page fault and we deadlock on trying
to get mmap_sem when handling that fault. So just make
ipath_user_sdma_pin_pages() use get_user_pages_fast() and leave
mmap_sem locking for mm.
This deadlock has actually been observed in the wild when the node
is under memory pressure.
Cc: <stable@vger.kernel.org>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
[ Merged in fix for call to get_user_pages_fast from Tetsuo Handa
<penguin-kernel@I-love.SAKURA.ne.jp>. - Roland ]
Signed-off-by: Roland Dreier <roland@purestorage.com>
Remove the redundant check of comparing if a 32-bit value is greater
than 0xffffffffULL.
Reported by Dan Carpenter.
Signed-off-by: Naresh Gottumukkala <bgottumukkala@emulex.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
1) ocrdma_remove_free() is called from a call_rcu callback funtion
context, which can be a bottom-half context. So the code in
ocrdma_remove_free should not sleep.
But ocrdma_cleanup_hw() can sleep, So move it ocrdma_remove()
instead of ocrdma_remove_free.
2) Fix a couple of kbuild test robot warnings.
Signed-off-by: Naresh Gottumukkala <bgottumukkala@emulex.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
We recently added a cap on "max_wqe_allocated" in 43a6b4025c
('RDMA/ocrdma: Create IRD queue fix').
My static checker complains that the cap has a problem because it
casts large values to negative. "attrs->cap.max_send_wr" is a u32.
It comes from the user, but it's capped in ocrdma_check_qp_params() so
it can't wrap here.
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
The Connect-IB adapter has an inherent page size which equals 4K.
Define an new enum that equals the page shift and use it instead of
using the value 12 throughout the code.
Signed-off-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
RTS to RTS transition should allow update of alternate path.
Signed-off-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
mlx5_cur and mlx5_new cannot have negative values so remove the
redundant condition.
Signed-off-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
In mlx5_mr_cache_init() the size variable is not used so remove it to
avoid compiler warnings when running with make W=1.
Signed-off-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Change optimal_reclaimed_pages() to increase the output size of each
reclaim pages command. This change reduces significantly the amount of
reclaim pages commands issued to FW when the driver is unloaded which
reduces the overall driver unload time.
Signed-off-by: Moshe Lazer <moshel@mellanox.com>
Signed-off-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Firmware spec requires reserved fields to be cleared when calling
set_hca_cap. Current code queries and copy to the set area, possibly
resulting in reserved bits not cleared. This patch copies only
writable fields to the set area.
Fix also typo - msx => max
Signed-off-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Connect-IB firmware requires 4K pages to be communicated with the
driver. This patch breaks larger pages to 4K units to enable support
for architectures utilizing larger page size, such as PowerPC. This
patch also fixes several places that referred to PAGE_SHIFT instead of
explicit 12 which is the inherent page shift on Connect-IB.
Signed-off-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
If DMA mapping fails, the driver cleared the object that holds the
previously DMA mapped pages. Fix this by allocating a new object for
the command that reports back to firmware that pages can't be
supplied.
Signed-off-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
On destroy QP the driver walks over the relevant CQ and removes CQEs
reported for the destroyed QP. It also frees the related SRQ entry
without checking that this is actually an SRQ-related CQE. In case of
a CQ used for both send and receive QP, we could free SRQ entries for
send CQEs. This patch resolves this issue by verifying that this is a
SRQ related CQE by checking the SRQ number in the CQE is not zero.
Signed-off-by: Moshe Lazer <moshel@mellanox.com>
Signed-off-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Make use of destroy_srq_kernel() to clear SRQ resouces.
Signed-off-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Make sure not to overflow when reading the page list from struct
ib_fast_reg_page_list.
Signed-off-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Use asynchronous commands to execute up to eight concurrent create MR
commands. This is to fill memory caches faster so we keep consuming
from there. Also, increase timeout for shrinking caches to five
minutes.
Signed-off-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Verify that the value is non negative before rounding up to power of 2.
Signed-off-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
'op' is the already RDMA_NL_GET_OP() masked 'type'. No need to mask it again.
Signed-off-by: Mathias Krause <minipli@googlemail.com>
Reviewed-by: Yann Droneaud <ydroneaud@opteya.com>
Acked-by: Sean Hefty <sean.hefty@intel.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Currently, we don't copy the immediate data from the userspace struct
to the kernel one when UD messages are being sent.
This patch makes sure that the immediate data is set correctly.
Signed-off-by: Latchesar Ionkov <lucho@ionkov.net>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Since commit 82dc3c63c6 ("net: introduce NAPI_POLL_WEIGHT")
netif_napi_add() produces an error message if a NAPI poll weight
greater than 64 is requested.
Use the standard NAPI weight.
Signed-off-by: Michal Schmidt <mschmidt@redhat.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
The driver starts the mcast_join task whenever the netdev interface is
UP without relation to the underlying IB port state.
Until the port state is ACTIVE all the join requests are irrelevant,
and the IB core returns -EINVAL. So the user will see errors such as:
"multicast join failed for ff12:401b:... , status -22".
Instead, have ipoib_mcast_join_task() return when the port is not active.
It will be called again when the port state is changed and the
low-level driver triggers the IB_EVENT_PORT_ACTIVE event or the
IB_EVENT_CLIENT_REREGISTER event.
Signed-off-by: Erez Shitrit <erezsh@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
The path_rec_completion() callback may be invoked asynchronously even
at the middle of "driver uninit" process. This can lead to scheduling
a task that tries to touch members of the priv object that are no
longer valid. For example the function cm_create_tx_qp can attempt to
create qp with no valid priv->pd object.
The following crash is one of the results:
RIP: 0010:[<ffffffffa021bb47>] [<ffffffffa021bb47>] ipoib_cm_create_tx_qp+0x57/0x90 [ib_ipoib]
Process ipoib (pid: 5916, threadinfo ffff8803786e4000, task ffff8804150e1500)
Stack:
Call Trace:
[<ffffffff81309ef0>] ? get_random_bytes+0x20/0x30
[<ffffffffa021be2a>] ipoib_cm_tx_init+0xca/0x340 [ib_ipoib]
[<ffffffffa021f765>] ipoib_cm_tx_start+0x215/0x3f0 [ib_ipoib]
[<ffffffffa021f550>] ? ipoib_cm_tx_start+0x0/0x3f0 [ib_ipoib]
[<ffffffff8108b2b0>] worker_thread+0x170/0x2a0
[<ffffffff81090bf0>] ? autoremove_wake_function+0x0/0x40
[<ffffffff8108b140>] ? worker_thread+0x0/0x2a0
[<ffffffff81090886>] kthread+0x96/0xa0
[<ffffffff8100c14a>] child_rip+0xa/0x20
[<ffffffff810907f0>] ? kthread+0x0/0xa0
[<ffffffff8100c140>] ? child_rip+0x0/0x20
Fix that by flushing all pending path queries at this point.
Signed-off-by: Alex Markuze <markuze@mellanox.com>
Signed-off-by: Erez Shitrit <erezsh@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
The driver should avoid calling ib_sa_free_multicast on the mcast->mc
object until it finishes its initialization state. Otherwise we can
crash when ipoib_mcast_dev_flush() attempts to use the uninitialized
multicast object.
Instead, only call wait_for_completion() for multicast entries that
started the join process, meaning that ib_sa_join_multicast() finished.
Signed-off-by: Erez Shitrit <erezsh@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
The driver should not flush the whole workqueue when only one work (the
pkey poll one) needs to be cancelled. Use cancel_delayed_work_sync()
instead.
Signed-off-by: Erez Shitrit <erezsh@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
When ipoib interface is going down it takes all of its children with
it, under mutex.
For each child, dev_change_flags() is called. That function calls
ipoib_stop() via the ndo, and causes flush of the workqueue.
Sometimes in the workqueue an __ipoib_dev_flush work() is waiting and
when invoked tries to get the same mutex, which leads to a deadlock,
as seen below.
The solution is to switch to rw-sem instead of mutex.
The deadlock:
[11028.165303] [<ffffffff812b0977>] ? vgacon_scroll+0x107/0x2e0
[11028.171844] [<ffffffff814eaac5>] schedule_timeout+0x215/0x2e0
[11028.178465] [<ffffffff8105a5c3>] ? perf_event_task_sched_out+0x33/0x80
[11028.185962] [<ffffffff814ea743>] wait_for_common+0x123/0x180
[11028.192491] [<ffffffff8105fa40>] ? default_wake_function+0x0/0x20
[11028.199504] [<ffffffff814ea85d>] wait_for_completion+0x1d/0x20
[11028.206224] [<ffffffff8108b4f1>] flush_cpu_workqueue+0x61/0x90
[11028.212948] [<ffffffff8108b5a0>] ? wq_barrier_func+0x0/0x20
[11028.219375] [<ffffffff8108bfc4>] flush_workqueue+0x54/0x80
[11028.225712] [<ffffffffa05a0576>] ipoib_mcast_stop_thread+0x66/0x90 [ib_ipoib]
[11028.233988] [<ffffffffa059ccea>] ipoib_ib_dev_down+0x6a/0x100 [ib_ipoib]
[11028.241678] [<ffffffffa059849a>] ipoib_stop+0x8a/0x140 [ib_ipoib]
[11028.248692] [<ffffffff8142adf1>] dev_close+0x71/0xc0
[11028.254447] [<ffffffff8142a631>] dev_change_flags+0xa1/0x1d0
[11028.261062] [<ffffffffa059851b>] ipoib_stop+0x10b/0x140 [ib_ipoib]
[11028.268172] [<ffffffff8142adf1>] dev_close+0x71/0xc0
[11028.273922] [<ffffffff8142a631>] dev_change_flags+0xa1/0x1d0
[11028.280452] [<ffffffff8148f20b>] devinet_ioctl+0x5eb/0x6a0
[11028.286786] [<ffffffff814903b8>] inet_ioctl+0x88/0xa0
[11028.292633] [<ffffffff8141591a>] sock_ioctl+0x7a/0x280
[11028.298576] [<ffffffff81189012>] vfs_ioctl+0x22/0xa0
[11028.304326] [<ffffffff81140540>] ? unmap_region+0x110/0x130
[11028.310756] [<ffffffff811891b4>] do_vfs_ioctl+0x84/0x580
[11028.316897] [<ffffffff81189731>] sys_ioctl+0x81/0xa0
and
11028.017533] [<ffffffff8105a5c3>] ? perf_event_task_sched_out+0x33/0x80
[11028.025030] [<ffffffff8100bb8e>] ? apic_timer_interrupt+0xe/0x20
[11028.031945] [<ffffffff814eb2ae>] __mutex_lock_slowpath+0x13e/0x180
[11028.039053] [<ffffffff814eb14b>] mutex_lock+0x2b/0x50
[11028.044910] [<ffffffffa059f7e7>] __ipoib_ib_dev_flush+0x37/0x210 [ib_ipoib]
[11028.052894] [<ffffffffa059fa00>] ? ipoib_ib_dev_flush_light+0x0/0x20 [ib_ipoib]
[11028.061363] [<ffffffffa059fa17>] ipoib_ib_dev_flush_light+0x17/0x20 [ib_ipoib]
[11028.069738] [<ffffffff8108b120>] worker_thread+0x170/0x2a0
[11028.076068] [<ffffffff81090990>] ? autoremove_wake_function+0x0/0x40
[11028.083374] [<ffffffff8108afb0>] ? worker_thread+0x0/0x2a0
[11028.089709] [<ffffffff81090626>] kthread+0x96/0xa0
[11028.095266] [<ffffffff8100c0ca>] child_rip+0xa/0x20
[11028.100921] [<ffffffff81090590>] ? kthread+0x0/0xa0
[11028.106573] [<ffffffff8100c0c0>] ? child_rip+0x0/0x20
[11028.112423] INFO: task ifconfig:23640 blocked for more than 120 seconds.
Signed-off-by: Erez Shitrit <erezsh@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Change CM skb memory allocation to use GFP_KERNEL when possible.
During device init there's no need to use GFP_ATOMIC when allocating
memory for the CM skbs -- use GFP_KERNEL instead.
Signed-off-by: Tal Alon <talal@mellanox.com>
Signed-off-by: Erez Shitrit <erezsh@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
If napi has never been enabled when calling ipoib_ib_dev_stop, a
kernel crash occurs, because the verbs layer completion handler
(ipoib_ib_completion) calls napi_schedule unconditionally.
If the napi structure passed in the napi_schedule call has not
been initialized, napi will crash.
The cleanest solution is to simply enable napi before calling
ipoib_ib_dev_stop in the dev_open error flow. (dev_stop then
immediately disables napi).
Signed-off-by: Jack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: Erez Shitrit <erezsh@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Physical addresses may be wider than virtual addresses (e.g. on i386
with PAE) and must not be formatted with %p.
Compile-tested only.
Signed-off-by: Ben Hutchings <ben@decadent.org.uk>
Signed-off-by: Roland Dreier <roland@purestorage.com>
As a simple optimization that should speed up the vast majority of
connect attemps on IB devices, when we are searching for the GID of an
incoming connection in the cached GID lists of devices, search the
device that received the incoming connection request first. If we
don't find it there, then move on to other devices.
This reduces the time to perform 10,000 connections considerably.
Prior to this patch, a bad run of cmtime would look like this:
connect : 12399.26 12351.10 8609.00 1239.93
With this patch, it looks more like this:
connect : 5864.86 5799.80 8876.00 586.49
Signed-off-by: Doug Ledford <dledford@redhat.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
The cma_acquire_dev function was changed by commit 3c86aa70bf
("RDMA/cm: Add RDMA CM support for IBoE devices") to use find_gid_port()
because multiport devices might have either IB or IBoE formatted gids.
The old function assumed that all ports on the same device used the
same GID format.
However, when it was changed to use find_gid_port(), we inadvertently
lost usage of the GID cache. This turned out to be a very costly
change. In our testing, each iteration through each index of the GID
table takes roughly 35us. When you have multiple devices in a system,
and the GID you are looking for is on one of the later devices, the
code loops through all of the GID indexes on all of the early devices
before it finally succeeds on the target device. This pathological
search behavior combined with 35us per GID table index retrieval
results in results such as the following from the cmtime application
that's part of the latest librdmacm git repo:
ib1:
step total ms max ms min us us / conn
create id : 29.42 0.04 1.00 2.94
bind addr : 186705.66 19.00 18556.00 18670.57
resolve addr : 41.93 9.68 619.00 4.19
resolve route: 486.93 0.48 101.00 48.69
create qp : 4021.95 6.18 330.00 402.20
connect : 68350.39 68588.17 24632.00 6835.04
disconnect : 1460.43 252.65-1862269.00 146.04
destroy : 41.16 0.04 2.00 4.12
ib0:
step total ms max ms min us us / conn
create id : 28.61 0.68 1.00 2.86
bind addr : 2178.86 2.95 201.00 217.89
resolve addr : 51.26 16.85 845.00 5.13
resolve route: 620.08 0.43 92.00 62.01
create qp : 3344.40 6.36 273.00 334.44
connect : 6435.99 6368.53 7844.00 643.60
disconnect : 5095.38 321.90 757.00 509.54
destroy : 37.13 0.02 2.00 3.71
Clearly, both the bind address and connect operations suffer
a huge penalty for being anything other than the default
GID on the first port in the system.
After applying this patch, the numbers now look like this:
ib1:
step total ms max ms min us us / conn
create id : 30.15 0.03 1.00 3.01
bind addr : 80.27 0.04 7.00 8.03
resolve addr : 43.02 13.53 589.00 4.30
resolve route: 482.90 0.45 100.00 48.29
create qp : 3986.55 5.80 330.00 398.66
connect : 7141.53 7051.29 5005.00 714.15
disconnect : 5038.85 193.63 918.00 503.88
destroy : 37.02 0.04 2.00 3.70
ib0:
step total ms max ms min us us / conn
create id : 34.27 0.05 1.00 3.43
bind addr : 26.45 0.04 1.00 2.64
resolve addr : 38.25 10.54 760.00 3.82
resolve route: 604.79 0.43 97.00 60.48
create qp : 3314.95 6.34 273.00 331.49
connect : 12399.26 12351.10 8609.00 1239.93
disconnect : 5096.76 270.72 1015.00 509.68
destroy : 37.10 0.03 2.00 3.71
It's worth noting that we still suffer a bit of a penalty on
connect to the wrong device, but the penalty is much less than
it used to be. Follow on patches deal with this penalty.
Many thanks to Neil Horman for helping to track the source of
slow function that allowed us to track down the fact that
the original patch I mentioned above backed out cache usage
and identify just how much that impacted the system.
Signed-off-by: Doug Ledford <dledford@redhat.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
The following scenario can cause silent data corruption when doing
NFS writes. It has mainly been observed when doing database writes
using O_DIRECT.
1) The RPC client uses sendpage() to do zero-copy of the page data.
2) Due to networking issues, the reply from the server is delayed,
and so the RPC client times out.
3) The client issues a second sendpage of the page data as part of
an RPC call retransmission.
4) The reply to the first transmission arrives from the server
_before_ the client hardware has emptied the TCP socket send
buffer.
5) After processing the reply, the RPC state machine rules that
the call to be done, and triggers the completion callbacks.
6) The application notices the RPC call is done, and reuses the
pages to store something else (e.g. a new write).
7) The client NIC drains the TCP socket send buffer. Since the
page data has now changed, it reads a corrupted version of the
initial RPC call, and puts it on the wire.
This patch fixes the problem in the following manner:
The ordering guarantees of TCP ensure that when the server sends a
reply, then we know that the _first_ transmission has completed. Using
zero-copy in that situation is therefore safe.
If a time out occurs, we then send the retransmission using sendmsg()
(i.e. no zero-copy), We then know that the socket contains a full copy of
the data, and so it will retransmit a faithful reproduction even if the
RPC call completes, and the application reuses the O_DIRECT buffer in
the meantime.
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
Cc: stable@vger.kernel.org
