On Tue, Jun 15, 2021 at 04:46:56PM +1000, Dave Chinner wrote:
> Hi folks,
>
> This is the first fix for the problems Brian has reported from
> generic/019. This has fixed the hang, but the other log recovery
> problem he reported is still present (seen once with these patches
> in place).
>
> I've tested these out to a couple of hundred cycles of
> continual looping generic/019 before the systems fall over with a
> perag reference count underrun at unmount after a shutdown. I'm
> pretty sure the hang is fixed, as it would manifest within 10-20
> cycles without this patch.
>
> The first patch is the iclogbuf state tracing I used to capture the
> iclogbuf wrapping state. The second patch is the fix.
I found another bug while testing for-next. If I run generic/100 more
than about ~30 times with a 1k block size:
FSTYP -- xfs (debug)
PLATFORM -- Linux/x86_64 flax-mtr00 5.13.0-rc4-djwx #rc4 SMP
PREEMPT Mon Jun 7 11:17:23 PDT 2021
MKFS_OPTIONS -- -f -b size=1024, /dev/sdf
MOUNT_OPTIONS -- -o usrquota,grpquota,prjquota, /dev/sdf /opt
I see this in dmesg:
run fstests generic/100 at 2021-06-15 10:41:45
XFS (sda): ctx ticket reservation ran out. Need to up reservation
XFS (sda): ticket reservation summary:
XFS (sda): unit res = 47168 bytes
XFS (sda): current res = -404 bytes
XFS (sda): original count = 1
XFS (sda): remaining count = 1
XFS (sda): xfs_do_force_shutdown(0x2) called from line 2440 of file fs/xfs/xfs_log.c. Return address = xlog_write+0x608/0x640 [xfs]
XFS (sda): Log I/O Error Detected. Shutting down filesystem
XFS (sda): Please unmount the filesystem and rectify the problem(s)
XFS (sda): Unmounting Filesystem
Looking up that line in gdb produces:
0xffffffffa038a0a8 is in xlog_write (fs/xfs/xfs_log.c:2439).
2434 int log_offset;
2435
2436 if (ticket->t_curr_res < 0) {
2437 xfs_alert_tag(log->l_mp, XFS_PTAG_LOGRES,
2438 "ctx ticket reservation ran out. Need to up reservation");
2439 xlog_print_tic_res(log->l_mp, ticket);
2440 xfs_force_shutdown(log->l_mp, SHUTDOWN_LOG_IO_ERROR);
2441 }
I haven't applied these two patches yet, but looking back through
fstests reports I never saw this before the recent for-next push.
I'm uncertain if it's the CIL work or the xattr refactoring that did
this, though AFAICT generic/100 itself does not generate any xattrs and
I don't have any LSMs enabled that would cause them to be created.
--D
>
> Cheers,
>
> Dave.
>
>
