Locked up processes after upgrade to ZFS v15

[Kostik Belousov]

On Sun, Oct 10, 2010 at 12:47:50PM +0300, Andriy Gapon wrote:
> on 08/10/2010 19:24 Andriy Gapon said the following:
> > on 06/10/2010 21:51 Kai Gallasch said the following:
> >>
> >> Am 06.10.2010 um 19:32 schrieb Martin Simmons:
> >>
> >>>>>>>> On Wed, 6 Oct 2010 14:28:31 +0200, Kai Gallasch said:
> >>>>
> >>>> How can I debug this and get further information?
> >>>
> >>> procstat -k -k $pid will generate a backtrace (or replace $pid by -a for all
> >>> processes).
> >>
> >> procstat for process 12111 (state: zfs)
> >> sonnenkraft:~ # procstat -k -k 12111
> >>   PID    TID COMM             TDNAME           KSTACK                       
> >> 12111 102385 httpd            -                mi_switch+0x21b sleepq_switch+0x123 sleepq_wait+0x4d __lockmgr_args+0x7ae vop_stdlock+0x39 VOP_LOCK1_APV+0x9b _vn_lock+0x57 vget+0x7b cache_lookup+0x4e0 vfs_cache_lookup+0xc0 VOP_LOOKUP_APV+0xb7 lookup+0x3d3 namei+0x457 vn_open_cred+0x1e3 kern_openat+0x181 syscall+0x102 Xfast_syscall+0xe2
> >>
> >> procstat for process 24731 (state: zfsmrb)
> >> # procstat -k -k 24731
> >>   PID    TID COMM             TDNAME           KSTACK                       
> >> 24731 102273 httpd            -                mi_switch+0x21b sleepq_switch+0x123 sleepq_wait+0x4d _sleep+0x369 zfs_freebsd_read+0x2a6 VOP_READ_APV+0xaf vnode_pager_generic_getpages+0x3ea VOP_GETPAGES_APV+0xb5 vnode_pager_getpages+0x8c vm_fault+0x685 trap_pfault+0x128 trap+0x52c calltrap+0x8
> 
> Hm, I think that we actually shouldn't see a stack like that.
> vm_fault sets VPO_BUSY on a page before calling vnode_pager_generic_getpages, so
> the thread gets stuck forever in zfs mappedread.
> It seems like the page that was seen as invalid in vm_fault becomes valid while
> call flow reaches mappedread.
The vnode is share-locked, and vm object lock is dropped and reacquired
several times until control reaches zfs_mappedread. This indeed allows
a window during which page might be read by other thread.

There are two possible routes to solve the issue:
1. Provide zfs-specific VOP_GETPAGES().
2. Use my vm6 patch. Sigh.

> 
> >> In my original post I wrote that only apache httpd processes would lock up..
> >> This is wrong. Several other non-httpd processes also got stuck in state zfs or zfsmrb.
> > 
> > Interesting.
> > It's possible that TID 102385 might be waiting on a vnode lock held by TID 102273.
> > But TID 102273 seems to be waiting on a vnode's page lock.
> > It would be very interesting to learn what process has that page busy, for how
> > long and why.
> > Perhaps there is a code path that busies a page, but never un-busies it...
> > 
> 
> 
> -- 
> Andriy Gapon
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 196 bytes
Desc: not available
Url : http://lists.freebsd.org/pipermail/freebsd-fs/attachments/20101010/c9a0edda/attachment.pgp