nfsd loads 90% CPU, client hangs

[Sergey Lapin <slapin@ossfans.org>]

Hi, all!

With recent kernels I see a problem with using NFS. It was broken
somewhere after 2.6.27.

I have ARM board with several hard drives connected over USB 1.1
dongles (USB->IDE, USB->SATA). And I have lvm2 over them.
They produce 2 logical volumes with data, which are exported
over NFS to PC host. ARM box runs vanilla kernel 2.6.30-rc8,
and PC host runs Debian kernel 2.6.24. After some bigger file writes
(when large amounts of data are written to disks) I experience the
following error in logs on ARM nfsd server host. I use kernel nfsd here,
to be clear. I use NFSv3.

INFO: task nfsd:1933 blocked for more than 120 seconds.
"echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this
message.
nfsd          D c02e29e8     0  1933      2
[<c02e29e8>] (__schedule+0x2d8/0x348) from [<c02e374c>]
(__mutex_lock_slowpath+0x8c/0xfc)
[<c02e374c>] (__mutex_lock_slowpath+0x8c/0xfc) from [<c006d670>]
(generic_file_aio_write+0x58/0xe8)
[<c006d670>] (generic_file_aio_write+0x58/0xe8) from [<c00dc1ec>]
(ext3_file_write+0x20/0xa0)
[<c00dc1ec>] (ext3_file_write+0x20/0xa0) from [<c0093cc8>]
(do_sync_readv_writev+0xac/0x100)
[<c0093cc8>] (do_sync_readv_writev+0xac/0x100) from [<c00943e4>]
(do_readv_writev+0xac/0x1b0)
[<c00943e4>] (do_readv_writev+0xac/0x1b0) from [<c009454c>]
(vfs_writev+0x64/0x74)
[<c009454c>] (vfs_writev+0x64/0x74) from [<c0124c88>]
(nfsd_vfs_write+0x10c/0x350)
[<c0124c88>] (nfsd_vfs_write+0x10c/0x350) from [<c01257b4>]
(nfsd_write+0xc0/0xd8)
[<c01257b4>] (nfsd_write+0xc0/0xd8) from [<c012c354>]
(nfsd3_proc_write+0xe8/0x114)
[<c012c354>] (nfsd3_proc_write+0xe8/0x114) from [<c0120f90>]
(nfsd_dispatch+0xcc/0x1e4)
[<c0120f90>] (nfsd_dispatch+0xcc/0x1e4) from [<c02d3f34>]
(svc_process+0x42c/0x7a8)
[<c02d3f34>] (svc_process+0x42c/0x7a8) from [<c0121640>]
(nfsd+0xe4/0x148)
[<c0121640>] (nfsd+0xe4/0x148) from [<c0056720>] (kthread+0x58/0x90)
[<c0056720>] (kthread+0x58/0x90) from [<c0044e90>] (do_exit+0x0/0x620)
[<c0044e90>] (do_exit+0x0/0x620) from [<ffffffff>] (0xffffffff)

And then NFS doesn't work at all with nfsd consuming all of CPU it can.
I see no hardware problems here, because files are perfectly accessible
locally or over HTTP, and no USB or disk error messages.

If I reboot ARM box without unmounting NFS shares on PC, the same
situation occurs as soon as ARM box boots (excessively loaded CPU with
nfsd at top, and NFS doesn't work and doesn't recover). If I unmount
them, box boots fine, but fails again as soon as I repeat file
operation.
So, the question is - what causes it and if it is possible to fix this
problem or work it around?

Thanks a lot,
S.