8.1R ZFS almost locking up system
Dan Nelson
dnelson at allantgroup.com
Thu Sep 2 14:57:03 UTC 2010
In the last episode (Sep 01), Tim Bishop said:
> On Tue, Aug 31, 2010 at 10:58:29AM -0500, Dan Nelson wrote:
> > In the last episode (Aug 31), Tim Bishop said:
> > > It happened again this Saturday (clearly something in the weekly
> > > periodic run is triggering the issue). procstat -kk shows the
> > > following for processes doing something zfs related (where zfs related
> > > means the string 'zfs' in the procstat -kk output):
> > >
> > > 0 100084 kernel zfs_vn_rele_task mi_switch+0x16f sleepq_wait+0x42 _sleep+0x31c taskqueue_thread_loop+0xb7 fork_exit+0x118 fork_trampoline+0xe
> > > 5 100031 zfskern arc_reclaim_thre mi_switch+0x16f sleepq_timedwait+0x42 _cv_timedwait+0x129 arc_reclaim_thread+0x2d1 fork_exit+0x118 fork_trampoline+0xe
> > > 5 100032 zfskern l2arc_feed_threa mi_switch+0x16f sleepq_timedwait+0x42 _cv_timedwait+0x129 l2arc_feed_thread+0x1be fork_exit+0x118 fork_trampoline+0xe
> > > 5 100085 zfskern txg_thread_enter mi_switch+0x16f sleepq_wait+0x42 _cv_wait+0x111 txg_thread_wait+0x79 txg_quiesce_thread+0xb5 fork_exit+0x118 fork_trampoline+0xe
> > > 5 100086 zfskern txg_thread_enter mi_switch+0x16f sleepq_wait+0x42 _cv_wait+0x111 zio_wait+0x61 dsl_pool_sync+0xea spa_sync+0x355 txg_sync_thread+0x195 fork_exit+0x118 fork_trampoline+0xe
> > > 17 100040 syncer - mi_switch+0x16f sleepq_wait+0x42 _cv_wait+0x111 txg_wait_synced+0x7c zil_commit+0x416 zfs_sync+0xa6 sync_fsync+0x184 sync_vnode+0x16b sched_sync+0x1c9 fork_exit+0x118 fork_trampoline+0xe
> > > 2210 100156 syslogd - mi_switch+0x16f sleepq_wait+0x42 _cv_wait+0x111 txg_wait_open+0x85 zfs_freebsd_write+0x378 VOP_WRITE_APV+0xb2 vn_write+0x2d7 dofilewrite+0x85 kern_writev+0x60 writev+0x41 syscall+0x1e7 Xfast_syscall+0xe1
> > > 3500 100177 syslogd - mi_switch+0x16f sleepq_wait+0x42 _cv_wait+0x111 txg_wait_open+0x85 zfs_freebsd_write+0x378 VOP_WRITE_APV+0xb2 vn_write+0x2d7 dofilewrite+0x85 kern_writev+0x60 writev+0x41 syscall+0x1e7 Xfast_syscall+0xe1
> > > 3783 100056 syslogd - mi_switch+0x16f sleepq_wait+0x42 _cv_wait+0x111 txg_wait_open+0x85 zfs_freebsd_write+0x378 VOP_WRITE_APV+0xb2 vn_write+0x2d7 dofilewrite+0x85 kern_writev+0x60 writev+0x41 syscall+0x1e7 Xfast_syscall+0xe1
> > > 4064 100165 mysqld initial thread mi_switch+0x16f sleepq_wait+0x42 _cv_wait+0x111 txg_wait_open+0x85 dmu_tx_assign+0x16c zfs_inactive+0xd9 zfs_freebsd_inactive+0x1a vinactive+0x6a vputx+0x1cc vn_close+0xa1 vn_closefile+0x5a _fdrop+0x23 closef+0x3b kern_close+0x14d syscall+0x1e7 Xfast_syscall+0xe1
> > > 4441 100224 python2.6 initial thread mi_switch+0x16f sleepq_wait+0x42 _cv_wait+0x111 txg_wait_open+0x85 dmu_tx_assign+0x16c zfs_inactive+0xd9 zfs_freebsd_inactive+0x1a vinactive+0x6a vputx+0x1cc null_reclaim+0xbc vgonel+0x12e vrecycle+0x7d null_inactive+0x1f vinactive+0x6a vputx+0x1cc vn_close+0xa1 vn_closefile+0x5a _fdrop+0x23
> > > 4444 100227 python2.6 initial thread mi_switch+0x16f sleepq_wait+0x42 _cv_wait+0x111 txg_wait_open+0x85 dmu_tx_assign+0x16c zfs_inactive+0xd9 zfs_freebsd_inactive+0x1a vinactive+0x6a vputx+0x1cc null_reclaim+0xbc vgonel+0x12e vrecycle+0x7d null_inactive+0x1f vinactive+0x6a vputx+0x1cc vn_close+0xa1 vn_closefile+0x5a _fdrop+0x23
> > > 4445 100228 python2.6 initial thread mi_switch+0x16f sleepq_wait+0x42 _cv_wait+0x111 txg_wait_open+0x85 dmu_tx_assign+0x16c zfs_inactive+0xd9 zfs_freebsd_inactive+0x1a vinactive+0x6a vputx+0x1cc null_reclaim+0xbc vgonel+0x12e vrecycle+0x7d null_inactive+0x1f vinactive+0x6a vputx+0x1cc vn_close+0xa1 vn_closefile+0x5a _fdrop+0x23
> > > 4446 100229 python2.6 initial thread mi_switch+0x16f sleepq_wait+0x42 _cv_wait+0x111 txg_wait_open+0x85 dmu_tx_assign+0x16c zfs_inactive+0xd9 zfs_freebsd_inactive+0x1a vinactive+0x6a vputx+0x1cc null_reclaim+0xbc vgonel+0x12e vrecycle+0x7d null_inactive+0x1f vinactive+0x6a vputx+0x1cc vn_close+0xa1 vn_closefile+0x5a _fdrop+0x23
> > > 4447 100089 python2.6 initial thread mi_switch+0x16f sleepq_wait+0x42 _cv_wait+0x111 txg_wait_open+0x85 dmu_tx_assign+0x16c zfs_inactive+0xd9 zfs_freebsd_inactive+0x1a vinactive+0x6a vputx+0x1cc null_reclaim+0xbc vgonel+0x12e vrecycle+0x7d null_inactive+0x1f vinactive+0x6a vputx+0x1cc vn_close+0xa1 vn_closefile+0x5a _fdrop+0x23
> > > 5352 100270 mutt - mi_switch+0x16f sleepq_wait+0x42 _cv_wait+0x111 txg_wait_synced+0x7c zil_commit+0x416 zfs_freebsd_fsync+0xd7 null_bypass+0xd3 fsync+0x161 syscall+0x1e7 Xfast_syscall+0xe1
> > > 52686 100200 tarsnap - mi_switch+0x16f sleepq_wait+0x42 _cv_wait+0x111 txg_wait_open+0x85 dmu_tx_assign+0x16c zfs_inactive+0xd9 zfs_freebsd_inactive+0x1a vinactive+0x6a vputx+0x1cc vn_close+0xa1 vn_closefile+0x5a _fdrop+0x23 closef+0x3b kern_close+0x14d syscall+0x1e7 Xfast_syscall+0xe1
> > > 59049 100207 webalizer initial thread mi_switch+0x16f sleepq_wait+0x42 _cv_wait+0x111 txg_wait_open+0x85 zfs_freebsd_write+0x378 VOP_WRITE_APV+0xb2 null_bypass+0xd3 VOP_WRITE_APV+0x141 vn_write+0x2d7 dofilewrite+0x85 kern_pwritev+0x63 pwrite+0x59 syscall+0x1e7 Xfast_syscall+0xe1
> > > 77573 100479 perl - mi_switch+0x16f sleepq_wait+0x42 _cv_wait+0x111 txg_wait_open+0x85 zfs_freebsd_write+0x378 VOP_WRITE_APV+0xb2 null_bypass+0xd3 VOP_WRITE_APV+0x141 vn_write+0x2d7 dofilewrite+0x85 kern_writev+0x60 write+0x55 syscall+0x1e7 Xfast_syscall+0xe1
> > > 78595 100275 zfs - mi_switch+0x16f sleepq_wait+0x42 _cv_wait+0x111 txg_wait_synced+0x7c dsl_sync_task_group_wait+0x11c dmu_objset_snapshot+0x1b8 zfs_ioc_snapshot+0x7c zfsdev_ioctl+0x8d devfs_ioctl_f+0x77 kern_ioctl+0xf6 ioctl+0xfd syscall+0x1e7 Xfast_syscall+0xe1
> > > 81989 100596 zfs - mi_switch+0x16f sleepq_wait+0x42 _cv_wait+0x111 txg_wait_synced+0x7c dsl_sync_task_group_wait+0x11c dmu_objset_snapshot+0x1b8 zfs_ioc_snapshot+0x7c zfsdev_ioctl+0x8d devfs_ioctl_f+0x77 kern_ioctl+0xf6 ioctl+0xfd syscall+0x1e7 Xfast_syscall+0xe1
> > >
> > > I'm not sure if this shows anything useful?
> >
> > All your userland processes are basically waiting for the kernel to
> > finish writing a ZFS transaction group to disk. mutt has called fsync,
> > which may have been the trigger. Usually writing a transaction group is
> > fast, though, because ZFS will batch up all the new data into one
> > contiguous block and write it at full speed to disk. That's why I asked
> > about full filesystems before, since if your FS has been near 99%, you
> > may not have any large runs of freespace left.
>
> Right. But I wouldn't have thought that'd be effectively terminal? It's
> not just a bit slow - the machine freezes up, sometimes for many hours
> until rebooted.
>
> > I noticed in your original post:
> >
> > capacity operations bandwidth
> > pool used avail read write read write
> > ---------- ----- ----- ----- ----- ----- -----
> > pool0 117G 16.7G 248 114 865K 269K
> > mirror 117G 16.7G 248 114 865K 269K
> > ad4s3 - - 43 56 2.47M 269K
> > ad6s3 - - 39 56 2.41M 269K
> > ---------- ----- ----- ----- ----- ----- -----
>
> I did a scrub the other day and I noticed this same pattern (reads
> happening more on the disks and the pool).
A scrub works down the ZFS tree, and due to COW the upper levels of metadata
are likely to be scattered across the disk in small chunks, so the first 10%
or so of the scrub will be tiny random reads. Once it gets down to actually
checking file data it should max out all your disks throughput.
> > # gstat
> > ...
> > L(q) ops/s r/s kBps ms/r w/s kBps ms/w %busy Name
> > 1 48 48 3042 9.8 0 0 0.0 47.6| ad4
> > 0 38 38 2406 10.5 0 0 0.0 39.5| ad6
> >
> > You have a pair of mirrored disks, each doing around 40% I/O load, which
> > is 80% load if a single-threaded task is driving all the I/O. I see the
> > syncer process is also trying to write to the ZIL. Are you running
> > something that does a lot of fsync calls (a database server for
> > example)? Is this system an NFS server maybe? Try setting the sysctl
> > vfs.zfs.zil_disable=1 and see if your performance improves.
>
> I am running both MySQL and PostgreSQL in jails, but both are extremely
> lightly loaded. No NFS.
>
> I've looked at disabling the ZIL, but it doesn't seem to be a
> recommended thing to do?
This is just for debugging, to see whether ZIL is your problem. ZIL logs
synchronous actions (fsync() and writes to O_SYNC files) as they happen, and
is always replayed on reboot (as opposed to the rest of the txg which gets
rolled back).
http://www.solarisinternals.com/wiki/index.php/ZFS_Evil_Tuning_Guide#Disabling_the_ZIL_.28Don.27t.29
> I've also just upgraded to 8-STABLE to see if the few ZFS updates in
> there make any difference.
--
Dan Nelson
dnelson at allantgroup.com
More information about the freebsd-stable
mailing list