Re: v5.10.1 xfs deadlock

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 21.12.20 13:22, Donald Buczek wrote:
On 18.12.20 22:49, Dave Chinner wrote:
On Thu, Dec 17, 2020 at 06:44:51PM +0100, Donald Buczek wrote:
Dear xfs developer,

I was doing some testing on a Linux 5.10.1 system with two 100 TB xfs filesystems on md raid6 raids.

The stress test was essentially `cp -a`ing a Linux source repository with two threads in parallel on each filesystem.

After about on hour, the processes to one filesystem (md1) blocked, 30 minutes later the process to the other filesystem (md0) did.

     root      7322  2167  0 Dec16 pts/1    00:00:06 cp -a /jbod/M8068/scratch/linux /jbod/M8068/scratch/1/linux.018.TMP
     root      7329  2169  0 Dec16 pts/1    00:00:05 cp -a /jbod/M8068/scratch/linux /jbod/M8068/scratch/2/linux.019.TMP
     root     13856  2170  0 Dec16 pts/1    00:00:08 cp -a /jbod/M8067/scratch/linux /jbod/M8067/scratch/2/linux.028.TMP
     root     13899  2168  0 Dec16 pts/1    00:00:05 cp -a /jbod/M8067/scratch/linux /jbod/M8067/scratch/1/linux.027.TMP

Some info from the system (all stack traces, slabinfo) is available here: https://owww.molgen.mpg.de/~buczek/2020-12-16.info.txt

It stands out, that there are many (549 for md0, but only 10 for md1)  "xfs-conv" threads all with stacks like this

     [<0>] xfs_log_commit_cil+0x6cc/0x7c0
     [<0>] __xfs_trans_commit+0xab/0x320
     [<0>] xfs_iomap_write_unwritten+0xcb/0x2e0
     [<0>] xfs_end_ioend+0xc6/0x110
     [<0>] xfs_end_io+0xad/0xe0
     [<0>] process_one_work+0x1dd/0x3e0
     [<0>] worker_thread+0x2d/0x3b0
     [<0>] kthread+0x118/0x130
     [<0>] ret_from_fork+0x22/0x30

xfs_log_commit_cil+0x6cc is

   xfs_log_commit_cil()
     xlog_cil_push_background(log)
       xlog_wait(&cil->xc_push_wait, &cil->xc_push_lock);

Some other threads, including the four "cp" commands are also blocking at xfs_log_commit_cil+0x6cc

There are also single "flush" process for each md device with this stack signature:

     [<0>] xfs_map_blocks+0xbf/0x400
     [<0>] iomap_do_writepage+0x15e/0x880
     [<0>] write_cache_pages+0x175/0x3f0
     [<0>] iomap_writepages+0x1c/0x40
     [<0>] xfs_vm_writepages+0x59/0x80
     [<0>] do_writepages+0x4b/0xe0
     [<0>] __writeback_single_inode+0x42/0x300
     [<0>] writeback_sb_inodes+0x198/0x3f0
     [<0>] __writeback_inodes_wb+0x5e/0xc0
     [<0>] wb_writeback+0x246/0x2d0
     [<0>] wb_workfn+0x26e/0x490
     [<0>] process_one_work+0x1dd/0x3e0
     [<0>] worker_thread+0x2d/0x3b0
     [<0>] kthread+0x118/0x130
     [<0>] ret_from_fork+0x22/0x30

xfs_map_blocks+0xbf is the

     xfs_ilock(ip, XFS_ILOCK_SHARED);

in xfs_map_blocks().

Can you post the entire dmesg output after running
'echo w > /proc/sysrq-trigger' to dump all the block threads to
dmesg?

I have an out of tree driver for the HBA ( smartpqi 2.1.6-005
pulled from linux-scsi) , but it is unlikely that this blocking is
related to that, because the md block devices itself are
responsive (`xxd /dev/md0` )

My bet is that the OOT driver/hardware had dropped a log IO on the
floor - XFS is waiting for the CIL push to complete, and I'm betting
that is stuck waiting for iclog IO completion while writing the CIL
to the journal. The sysrq output will tell us if this is the case,
so that's the first place to look.

I think you are right here, and I'm sorry for blaming the wrong layer.

I was to fast to accept that. The display of the non-zero "inflight" counters of the underlying block devices couldn't be reproduced. They usually are "0 0" when the filesystem is deadlocked.

To make sure, I've also added my on "bio requested" and "bio completed" atomic counters to the two submit_bio calls and to xlog_bio_end_io in xfs_log.c and when the system is deadlocked, the requests and completions do match:

    sudo gdb linux/vmlinux /proc/kcore  -ex 'print xxx_bio_requested' -ex 'print xxx_bio_completed'
    [...]
    $1 = {counter = 34723}
    $2 = {counter = 34723}

So at least log writes are not lost.

Best
  Donald

I've got the system into another (though little different) zero-progress situation. This time it happened on md0 only while md1 was still working.

I think, this should be prove, that the failure is on the block layer of the member disks:

     root:deadbird:/scratch/local/# for f in /sys/devices/virtual/block/md?/md/rd*/block/inflight;do echo $f: $(cat $f);done
     /sys/devices/virtual/block/md0/md/rd0/block/inflight: 1 0
     /sys/devices/virtual/block/md0/md/rd1/block/inflight: 1 0
     /sys/devices/virtual/block/md0/md/rd10/block/inflight: 1 0
     /sys/devices/virtual/block/md0/md/rd11/block/inflight: 1 0
     /sys/devices/virtual/block/md0/md/rd12/block/inflight: 1 0
     /sys/devices/virtual/block/md0/md/rd13/block/inflight: 1 0
     /sys/devices/virtual/block/md0/md/rd14/block/inflight: 1 0
     /sys/devices/virtual/block/md0/md/rd15/block/inflight: 1 0
     /sys/devices/virtual/block/md0/md/rd2/block/inflight: 1 0
     /sys/devices/virtual/block/md0/md/rd3/block/inflight: 1 0
     /sys/devices/virtual/block/md0/md/rd4/block/inflight: 1 0
     /sys/devices/virtual/block/md0/md/rd5/block/inflight: 1 0
     /sys/devices/virtual/block/md0/md/rd6/block/inflight: 1 0
     /sys/devices/virtual/block/md0/md/rd7/block/inflight: 1 0
     /sys/devices/virtual/block/md0/md/rd8/block/inflight: 1 0
     /sys/devices/virtual/block/md0/md/rd9/block/inflight: 1 0
     /sys/devices/virtual/block/md1/md/rd0/block/inflight: 0 0
     /sys/devices/virtual/block/md1/md/rd1/block/inflight: 0 0
     /sys/devices/virtual/block/md1/md/rd10/block/inflight: 0 0
     /sys/devices/virtual/block/md1/md/rd11/block/inflight: 0 0
     /sys/devices/virtual/block/md1/md/rd12/block/inflight: 0 0
     /sys/devices/virtual/block/md1/md/rd13/block/inflight: 0 0
     /sys/devices/virtual/block/md1/md/rd14/block/inflight: 0 0
     /sys/devices/virtual/block/md1/md/rd15/block/inflight: 0 0
     /sys/devices/virtual/block/md1/md/rd2/block/inflight: 0 0
     /sys/devices/virtual/block/md1/md/rd3/block/inflight: 0 0
     /sys/devices/virtual/block/md1/md/rd4/block/inflight: 0 0
     /sys/devices/virtual/block/md1/md/rd5/block/inflight: 0 0
     /sys/devices/virtual/block/md1/md/rd6/block/inflight: 0 0
     /sys/devices/virtual/block/md1/md/rd7/block/inflight: 0 0
     /sys/devices/virtual/block/md1/md/rd8/block/inflight: 0 0
     /sys/devices/virtual/block/md1/md/rd9/block/inflight: 0 0

Best

   Donald


Cheers,

Dave.


--
Donald Buczek
buczek@xxxxxxxxxxxxx
Tel: +49 30 8413 1433



[Index of Archives]     [XFS Filesystem Development (older mail)]     [Linux Filesystem Development]     [Linux Audio Users]     [Yosemite Trails]     [Linux Kernel]     [Linux RAID]     [Linux SCSI]


  Powered by Linux