| Message ID | 20180412181158.8884-1-alan.christopher.jenkins@gmail.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
Hi Alan, On Thu, 2018-04-12 at 19:11 +0100, Alan Jenkins wrote: > # dd if=/dev/sda of=/dev/null iflag=direct & \ > while killall -SIGUSR1 dd; do sleep 0.1; done & \ > echo mem > /sys/power/state ; \ > sleep 5; killall dd # stop after 5 seconds Can you please also add a regression test to blktests[1] for this? [1] https://github.com/osandov/blktests Thanks, Johannes
On 13/04/18 09:31, Johannes Thumshirn wrote: > Hi Alan, > > On Thu, 2018-04-12 at 19:11 +0100, Alan Jenkins wrote: >> # dd if=/dev/sda of=/dev/null iflag=direct & \ >> while killall -SIGUSR1 dd; do sleep 0.1; done & \ >> echo mem > /sys/power/state ; \ >> sleep 5; killall dd # stop after 5 seconds > Can you please also add a regression test to blktests[1] for this? > > [1] https://github.com/osandov/blktests > > Thanks, > Johannes Good question. It would be nice to promote this test. Template looks like I need the commit (sha1) first. I had some ideas about automating it, so I wrote a standalone (see end). I can automate the wakeup by using pm_test, but this is still a system suspend test. Unfortunately I don't think there's any alternative. To give the most dire example # This test is non-destructive, but it exercises suspend in all drivers. # If your system has a problem with suspend, it might not wake up again. So I'm not sure if it would be acceptable for the default set? How useful is this going to be? Is there an expanded/full set of tests that gets run somewhere? If you can't guarantee it's going to be run somewhere, I'd worry the cost/benefit feels a little narrow :-(. There were one or two further "interesting" details, and it might theoretically bitrot if it's not run periodically. If you look at the diff and title for the fix, I don't think it's at high risk of being reversed unintentionally. And I think you can trust users will notice if the fix gets merged away accidentally, before it hits -stable releases :-). The issue kills the entire GUI session on resume from suspend, say once every three days, on gnome-shell (due to Xwayland). One unfortunate user switched to Xorg only to find that was also affected. I honestly assume the issue applies generally to laptop systems. The only mitigating factor is if you have RAM to spare, so you don't hit the major pagefaults during resume. #!/bin/bash # This test is non-destructive, but it exercises suspend in all drivers. # If your system has a problem with suspend, it might not wake up again. # TEST_DEV must be SCSI (inc. libata). # # Additionally, this test will abort if $TEST_DEV is too tiny # and we finish reading it within 3 seconds. Sorry. TEST_DEV=sda # RATIONALE # # The original root cause issue was the behaviour around blk_queue_freeze(). # It put tasks into an interruptible wait, which is wrong for block devices. # # XXX Insert reference to fix commit XXX # # The freeze feature is not directly exposed to userspace, so I can not test # it directly :(. (It's used to "guarantee no request is in use, so we can # change any data structure of the queue afterward". I.e. freeze, modify the # queue structure, unfreeze). # # However, this lead to a regression with a decent reproducer. In v4.15 the # same interruptible wait was also used for SCSI suspend/resume. SCSI resume # can take a second or so... hence we like to do it asynchronously. This # means we can observe the wait at resume time, and we can test if it is # interruptible. # # Note `echo quiesce > /sys/class/scsi_device/*/device/state` can *not* # trigger the specific wait in the block layer. That code path only # sets the SCSI device state; it does not set any block device state. # (It does not call into blk_queue_freeze() or blk_set_preempt_only(); # it literally just sets sdev->sdev_state to SDEV_QUIESCE). set -o nounset abort() { echo "$*" echo "=== Test ERROR ===" exit 2 } SYSFS_PM_TEST_DELAY=/sys/module/suspend/parameters/pm_test_delay SAVED_PM_TEST_DELAY= # Child process IDs DD= SUBSHELL= cleanup() { # In many cases the subshell will already have exited... # and semantics for `wait` are crappy in shell. # Failure will be harmless in most cases. # Just try to provide enough context for the user to guess. echo "Cleaning up" if [ -n "$SUBSHELL" ]; then echo "Killing sub-shell PID $SUBSHELL..." kill $SUBSHELL wait $SUBSHELL fi if [ -n "$DD" ]; then echo "Killing 'dd' PID $DD..." kill $DD wait $DD fi echo "Resetting pm_test" echo none > /sys/power/pm_test echo "Resetting pm_test_delay" if [ -n "$SAVED_PM_TEST_DELAY" ]; then echo "$SAVED_PM_TEST_DELAY" > "$SYSFS_PM_TEST_DELAY" fi } trap cleanup EXIT # "If a user has disabled async probing a likely reason # is due to a storage enclosure that does not inject # staggered spin-ups. For safety, make resume # synchronous as well in that case." if ! SCAN="$(cat /sys/module/scsi_mod/parameters/scan)"; then abort "error reading '/sys/module/scsi_mod/parameters/scan' ?" fi if [ "$SCAN" != "async" ]; then abort "This test does not work if you have set 'scsi_mod.scan=sync'" fi # Ignore USR1, in the hope that this applies to child processes. # This allows us to safely `kill -USR1 $DD`, when we don't know # whether the child process has fully started yet. # # I think this is the only place I relied on the specific # shell (bash) behaviour. trap "" USR1 # Check dd can work if ! dd iflag=direct if="/dev/$TEST_DEV" of=/dev/null count=1 status=none; then abort "'dd'" fi # Start dd, as a background process which submits IOs and yells when one fails. # We want to hit the block layer, so use direct IO to avoid being served from # page cache. dd iflag=direct if="/dev/$TEST_DEV" of=/dev/null status=none & DD=$! if ! echo devices > /sys/power/pm_test; then abort "Setting pm_test failed, does your kernel lack CONFIG_PM_TEST?" fi if ! SAVED_PM_TEST_DELAY="$(cat "$SYSFS_PM_TEST_DELAY")"; then abort "error reading pm_test_delay" fi if ! echo 0 > "$SYSFS_PM_TEST_DELAY"; then abort "error setting pm_test_delay" fi # Just keep sending signals to 'dd' as long as it's alive. # dd accepts USR1 signal to print status. It doesn't seem to be a problem # that we told dd not to actually *print* anything ('status=none'). # # In theory this script is probably subject to various pid re-use races. # But I started in sh... so far blktests does not depend on python... # also direct IO is best to reproduce this, which is not built in to python. # ( while kill -USR1 $DD 2>/dev/null; do true; done ) & SUBSHELL=$! # Wait a second without suspending, it might pick up typos # or other unexpected errors. sleep 1 if ! kill -0 $DD; then DD= wait $DD || echo "'dd' exited with error" abort "'dd' exited early?" fi if ! kill -0 $SUBSHELL; then SUBSHELL= abort "subshell exited early?" fi # Log that we're suspending. User might not have guessed, # or maybe suspend (or pm_test suspend) is broken on this system. echo "Now simulating suspend/resume" if ! echo mem > /sys/power/state; then abort "system suspend failed or not supported?" fi # Now wait for TEST_DEV to resume asynchronously if ! dd iflag=direct if="/dev/$TEST_DEV" of=/dev/null count=1 status=none; then abort "'dd'" fi # Wait another second. This might be useful in the case dd got blocked on a # page fault during the suspend; it will have a second to get sorted out, # while potentially receiving an IO error and exiting. sleep 1 if ! kill -0 $DD 2>/dev/null; then if wait $DD; then DD= abort "'dd' exited early, without error. Device too tiny?" fi echo "'dd' exited with error" echo "=== Test FAIL ===" DD= exit 1 fi echo "=== Test PASS ==="
On 4/14/18 1:46 PM, Alan Jenkins wrote: > On 13/04/18 09:31, Johannes Thumshirn wrote: >> Hi Alan, >> >> On Thu, 2018-04-12 at 19:11 +0100, Alan Jenkins wrote: >>> # dd if=/dev/sda of=/dev/null iflag=direct & \ >>> while killall -SIGUSR1 dd; do sleep 0.1; done & \ >>> echo mem > /sys/power/state ; \ >>> sleep 5; killall dd # stop after 5 seconds >> Can you please also add a regression test to blktests[1] for this? >> >> [1] https://github.com/osandov/blktests >> >> Thanks, >> Johannes > > Good question. It would be nice to promote this test. > > Template looks like I need the commit (sha1) first. > > I had some ideas about automating it, so I wrote a standalone (see > end). I can automate the wakeup by using pm_test, but this is still a > system suspend test. Unfortunately I don't think there's any > alternative. To give the most dire example > > # This test is non-destructive, but it exercises suspend in all drivers. > # If your system has a problem with suspend, it might not wake up again. > > > So I'm not sure if it would be acceptable for the default set? > > How useful is this going to be? Is there an expanded/full set of tests > that gets run somewhere? > > If you can't guarantee it's going to be run somewhere, I'd worry the > cost/benefit feels a little narrow :-(. There were one or two further > "interesting" details, and it might theoretically bitrot if it's not run > periodically. I run it, just last week we found two new bugs with it. I'm requiring anyone that submits block patches to run the test suite, and also working towards having it be part of the 0-day runs so it gets run on posted patches automatically. So yes, it's useful and it won't bitrot. Please do turn it into a blktests test.
On 4/12/18 12:11 PM, Alan Jenkins wrote: > When blk_queue_enter() waits for a queue to unfreeze, or unset the > PREEMPT_ONLY flag, do not allow it to be interrupted by a signal. > > The PREEMPT_ONLY flag was introduced later in commit 3a0a529971ec > ("block, scsi: Make SCSI quiesce and resume work reliably"). Note the SCSI > device is resumed asynchronously, i.e. after un-freezing userspace tasks. > > So that commit exposed the bug as a regression in v4.15. A mysterious > SIGBUS (or -EIO) sometimes happened during the time the device was being > resumed. Most frequently, there was no kernel log message, and we saw Xorg > or Xwayland killed by SIGBUS.[1] > > [1] E.g. https://bugzilla.redhat.com/show_bug.cgi?id=1553979 > > Without this fix, I get an IO error in this test: > > # dd if=/dev/sda of=/dev/null iflag=direct & \ > while killall -SIGUSR1 dd; do sleep 0.1; done & \ > echo mem > /sys/power/state ; \ > sleep 5; killall dd # stop after 5 seconds > > The interruptible wait was added to blk_queue_enter in > commit 3ef28e83ab15 ("block: generic request_queue reference counting"). > Before then, the interruptible wait was only in blk-mq, but I don't think > it could ever have been correct. Applied, thanks. Still want that test in blktests, though!
On 14/04/18 20:52, Jens Axboe wrote: > On 4/14/18 1:46 PM, Alan Jenkins wrote: >> On 13/04/18 09:31, Johannes Thumshirn wrote: >>> Hi Alan, >>> >>> On Thu, 2018-04-12 at 19:11 +0100, Alan Jenkins wrote: >>>> # dd if=/dev/sda of=/dev/null iflag=direct & \ >>>> while killall -SIGUSR1 dd; do sleep 0.1; done & \ >>>> echo mem > /sys/power/state ; \ >>>> sleep 5; killall dd # stop after 5 seconds >>> Can you please also add a regression test to blktests[1] for this? >>> >>> [1] https://github.com/osandov/blktests >>> >>> Thanks, >>> Johannes >> Good question. It would be nice to promote this test. >> >> Template looks like I need the commit (sha1) first. >> >> I had some ideas about automating it, so I wrote a standalone (see >> end). I can automate the wakeup by using pm_test, but this is still a >> system suspend test. Unfortunately I don't think there's any >> alternative. To give the most dire example >> >> # This test is non-destructive, but it exercises suspend in all drivers. >> # If your system has a problem with suspend, it might not wake up again. >> >> >> So I'm not sure if it would be acceptable for the default set? >> >> How useful is this going to be? Is there an expanded/full set of tests >> that gets run somewhere? >> >> If you can't guarantee it's going to be run somewhere, I'd worry the >> cost/benefit feels a little narrow :-(. There were one or two further >> "interesting" details, and it might theoretically bitrot if it's not run >> periodically. > I run it, just last week we found two new bugs with it. I'm requiring > anyone that submits block patches to run the test suite, and also > working towards having it be part of the 0-day runs so it gets run > on posted patches automatically. > > So yes, it's useful and it won't bitrot. Please do turn it into a blktests > test. Thanks, it's really great to have a test suite. I was specifically checking in on how we can include a system suspend test. I've been thinking the suspend test could be opt-in test (e.g. ALLOW_PM_TEST=1), and then we have some infrastructure (you or 0-day runs) that does the opt-in. Without knowing anything about the infrastructure, I didn't want to assume that would work. I'm aware of one particular suspend issue; inside virt-manager VMs I see Linux crashing with a null pointer in qxl_drm_freeze. A regression soon after I learned how to use VMs for suspend tests :-( , and it's been long enough that I suspect few people use it. Partly what you saw me fishing for in the comments, is the idea of some kernel code allowing more direct testing of the queue freeze / preempt_only flag. That might be better engineering, but I don't know where I could put it. Alan
diff --git a/block/blk-core.c b/block/blk-core.c index abcb8684ba67..1a762f3980f2 100644 --- a/block/blk-core.c +++ b/block/blk-core.c @@ -915,7 +915,6 @@ int blk_queue_enter(struct request_queue *q, blk_mq_req_flags_t flags) while (true) { bool success = false; - int ret; rcu_read_lock(); if (percpu_ref_tryget_live(&q->q_usage_counter)) { @@ -947,14 +946,12 @@ int blk_queue_enter(struct request_queue *q, blk_mq_req_flags_t flags) */ smp_rmb(); - ret = wait_event_interruptible(q->mq_freeze_wq, - (atomic_read(&q->mq_freeze_depth) == 0 && - (preempt || !blk_queue_preempt_only(q))) || - blk_queue_dying(q)); + wait_event(q->mq_freeze_wq, + (atomic_read(&q->mq_freeze_depth) == 0 && + (preempt || !blk_queue_preempt_only(q))) || + blk_queue_dying(q)); if (blk_queue_dying(q)) return -ENODEV; - if (ret) - return ret; } }
When blk_queue_enter() waits for a queue to unfreeze, or unset the PREEMPT_ONLY flag, do not allow it to be interrupted by a signal. The PREEMPT_ONLY flag was introduced later in commit 3a0a529971ec ("block, scsi: Make SCSI quiesce and resume work reliably"). Note the SCSI device is resumed asynchronously, i.e. after un-freezing userspace tasks. So that commit exposed the bug as a regression in v4.15. A mysterious SIGBUS (or -EIO) sometimes happened during the time the device was being resumed. Most frequently, there was no kernel log message, and we saw Xorg or Xwayland killed by SIGBUS.[1] [1] E.g. https://bugzilla.redhat.com/show_bug.cgi?id=1553979 Without this fix, I get an IO error in this test: # dd if=/dev/sda of=/dev/null iflag=direct & \ while killall -SIGUSR1 dd; do sleep 0.1; done & \ echo mem > /sys/power/state ; \ sleep 5; killall dd # stop after 5 seconds The interruptible wait was added to blk_queue_enter in commit 3ef28e83ab15 ("block: generic request_queue reference counting"). Before then, the interruptible wait was only in blk-mq, but I don't think it could ever have been correct. Reviewed-by: Bart Van Assche <bart.vanassche@wdc.com> Cc: stable@vger.kernel.org Signed-off-by: Alan Jenkins <alan.christopher.jenkins@gmail.com> --- v2: fix indentation block/blk-core.c | 11 ++++------- 1 file changed, 4 insertions(+), 7 deletions(-)