Security Notification
#22 — zfs-fuse unstable under load
| State | Resolved |
|---|---|
| Version: |
—
|
| Area | User interface |
| Issue type | Patch |
| Severity | Important |
| Submitted by | Seth Heeren |
| Submitted on | Jan 25, 2010 |
| Responsible | Seth Heeren |
| Target release: |
Last modified on
May 22, 2010
by
Seth Heeren
I'm going to provide more details. For now, refer to group posting over at http://groups.google.com/[…]/8401baca9693691a
- Steps to reproduce:
- manyfold, e.g.
bonnie++ -d /tank &
sudo zpool iostat -v 1
Observe daemon crash or deadlock within 10s (on my system). Performance and resource usage may influence the timing.
fuse-threading.patch
—
differences between files,
0Kb
Added by
Seth Heeren
on
Jan 25, 2010 06:15 PM
Target release:
0.6.0 → 0.6.1
fix available. this will in likelihood fix most of the symptoms listed. Of course, some instability may be unrelated and therefore unaffected
# git clone http://git.zfs-fuse.net/official
# cd official/
# git checkout origin/critical
Please retest
# git clone http://git.zfs-fuse.net/official
# cd official/
# git checkout origin/critical
Please retest
Added by
Seth Heeren
on
Jan 26, 2010 06:54 PM
Severity:
Critical → Important
Emmanuel announced he has an independent fix in 67aca02ee
I confirm that it allows bonnie++ to run to completion while running 'zpool iostat -v 1' as well.
Ricardo Correia has confirmed that 67aca02ee is safe (and indicates probable changes needed to cater for zero-copy txg optimization as present in upstream).
Ricardo also validated the notion of making cur_fd thread-local.
I will retest the other instability symptoms with just 67aca02ee and see if they reoccur. If not, we can probably toss my patch.
I'm not convinced why my patch (attached) fixed the same symptoms at all... I want to review the code to see whether we might still want it for generic reasons (e.g. in case we want to re-enable those uiocopy bits in zfs-fuse style).
I confirm that it allows bonnie++ to run to completion while running 'zpool iostat -v 1' as well.
Ricardo Correia has confirmed that 67aca02ee is safe (and indicates probable changes needed to cater for zero-copy txg optimization as present in upstream).
Ricardo also validated the notion of making cur_fd thread-local.
I will retest the other instability symptoms with just 67aca02ee and see if they reoccur. If not, we can probably toss my patch.
I'm not convinced why my patch (attached) fixed the same symptoms at all... I want to review the code to see whether we might still want it for generic reasons (e.g. in case we want to re-enable those uiocopy bits in zfs-fuse style).
Added by
Seth Heeren
on
May 22, 2010 01:27 PM
Issue state:
unconfirmed → resolved
well tested
0.6.9 will have multithreaded ioctl dispatch anyway, combining these ideas (__thread cur_fd is back).
closing this because tests ok
0.6.9 will have multithreaded ioctl dispatch anyway, combining these ideas (__thread cur_fd is back).
closing this because tests ok
