nfs + zfs hangs on RELENG_9

[Andriy Gapon]

[restoring cc to fs@]

on 04/10/2012 14:32 Nikolay Denev said the following:
> I have procstat only for the nfsd threads from the moment of the IO hang.
> And this is the only one with "arc" :
> 
>      1422 138630 nfsd             nfsd: service    mi_switch+0x186
>     sleepq_wait+0x42 _sleep+0x390 arc_lowmem+0x77 kmem_malloc+0xc1
>     uma_large_malloc+0x4a malloc+0xd9 arc_get_data_buf+0xb5 arc_read_nolock+0x1ec
>     arc_read+0x93 dbuf_read+0x452 dmu_buf_hold_array_by_dnode+0x16b
>     dmu_buf_hold_array+0x67 dmu_read_uio+0x3f zfs_freebsd_read+0x3e8
>     nfsvno_read+0x2e5 nfsrvd_read+0x3ff nfsrvd_dorpc+0x3c0

Oh, very important stack trace.

Earlier Nikolay Denev said the following:
>   PID    TID COMM             TDNAME           KSTACK                       
>     7 100192 zfskern          arc_reclaim_thre mi_switch+0x186 sleepq_wait+0x42 _sx_xlock_hard+0x428
> _sx_xlock+0x51 arc_buf_remove_ref+0x8a dbuf_rele_and_unlock+0x132 dbuf_evict+0x11
> dbuf_do_evict+0x53 arc_do_user_evicts+0xb4 arc_reclaim_thread+0x263 fork_exit+0x11f
> fork_trampoline+0xe 

To me this looks like a deadlock caused by a FreeBSD add-on to ZFS: arc_lowmem
handler.
I think that this is what happens:
The nfsd thread does read, arc_read_nolock finds a buffer in a ghost cache and
calls arc_get_data_buf while holding a hash_lock (one of buffer hash locks).
arc_get_data_buf needs to allocate some memory and, as luck would have it, there
is a memory shortage.  Low memory handlers are invoked (directly) and one of them
is arc_lowmem.  arc_lowmem simply kicks arc_reclaim_thread to do its job and then
loops sleep-waiting until memory shortage is less severe.  arc_reclaim_thread
tries to evict some buffers and, as luck would have it again, it attempts to evict
either the same buffer or, most likely, a different buffer that hashes to the same
lock.
So arc_reclaim_thread is blocked on the arc buffer lock.  While the nfsd thread
holds the lock, but waits in arc_lowmem for arc_reclaim_thread to make progress.

Eventually the held lock stalls other threads that attempt to grab it, the stall
propagates to txg_sync_thread threads and all ZFS I/O stops.

-- 
Andriy Gapon