sync hangs - 2.6.35.10

I've just setup a 48 core server with 128GB of memory in a typical HPC setup. The only IO-activity happens over NFS and the applications are cpu-hogs.

The system is fully working and everthing looks apparently fine, but anything that issue a sync is hung for eternity.

root fs is ext4 and it appears that sync hitting that drive get hung due to some other things going on. There is only logging activity on that drive.

[  508.778695] Btrfs loaded
[ 7208.780233] INFO: task grub-probe:14787 blocked for more than 120
seconds.
[ 7208.780316] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs"
disables this message.
[ 7208.780397] grub-probe    D 0000000000000000     0 14787  14768
0x00000000
[ 7208.780402]  ffff882005f8fbb8 0000000000000086 ffff882000000000
0000000000015880
[ 7208.780406]  ffff882005f8ffd8 0000000000015880 ffff882005f8ffd8
ffff88200cd70000
[ 7208.780410]  0000000000015880 0000000000015880 ffff882005f8ffd8
0000000000015880
[ 7208.780413] Call Trace:
[ 7208.780424]  [<ffffffff8155d3cd>] schedule_timeout+0x22d/0x310
[ 7208.780430]  [<ffffffff8102ccae>] ? physflat_send_IPI_mask+0xe/0x10
[ 7208.780433]  [<ffffffff8155c666>] wait_for_common+0xd6/0x180
[ 7208.780439]  [<ffffffff810533b0>] ? default_wake_function+0x0/0x20
[ 7208.780441]  [<ffffffff8155c7ed>] wait_for_completion+0x1d/0x20
[ 7208.780446]  [<ffffffff81160ff3>] writeback_inodes_sb+0xb3/0xe0
[ 7208.780449]  [<ffffffff81165c4e>] __sync_filesystem+0x4e/0xa0
[ 7208.780452]  [<ffffffff81165d7a>] sync_filesystem+0x3a/0x70
[ 7208.780456]  [<ffffffff8116f9fe>] fsync_bdev+0x2e/0x60
[ 7208.780460]  [<ffffffff8128e5ce>] blkdev_ioctl+0x4ee/0x820
[ 7208.780463]  [<ffffffff8116dfcc>] block_ioctl+0x3c/0x40
[ 7208.780468]  [<ffffffff8114edad>] vfs_ioctl+0x3d/0xd0
[ 7208.780471]  [<ffffffff8114f3b8>] do_vfs_ioctl+0x88/0x540
[ 7208.780475]  [<ffffffff811586fa>] ? alloc_fd+0x10a/0x150
[ 7208.780478]  [<ffffffff8114f8f1>] sys_ioctl+0x81/0xa0
[ 7208.780483]  [<ffffffff8100a032>] system_call_fastpath+0x16/0x1b

Full dmesg here: http://shrek.krogh.cc/~jesper/bonnie-dmesg.txt

It seems like the problems about broken sync writeback discussed about a year ago .. last discussions in late january this year.

http://thread.gmane.org/gmane.linux.kernel/949268/focus=1090266

Any patches that may be relevant?

Definitely. There have been several patches fixing livelock issues
in this area under various conditions. Considering your use case, you might be hit by problems being fixed by commits:
6585027a5e8cb490e3a761b2f3f3c3acf722aff2 aa373cf550994623efb5d49a4d8775bafd10bbc1 b9543dac5bbc4aef0a598965b6b34f6259ab9a9b
(went into 2.6.38-rc1)

possibly also older
7624ee72aa09334af072853457a5d46d9901c3f8
(in 2.6.36-rc1)

Quoted 7 lines · Jan Kara

Definitely. There have been several patches fixing livelock issues
in this area under various conditions. Considering your use case, you might be hit by problems being fixed by commits:
6585027a5e8cb490e3a761b2f3f3c3acf722aff2 aa373cf550994623efb5d49a4d8775bafd10bbc1 b9543dac5bbc4aef0a598965b6b34f6259ab9a9b
(went into 2.6.38-rc1)

I couldn't get above to apply on a 2.6.35.10 kernel, but applied to a 2.6.37 the problem seems solved. Have they been queued up for stable?

Quoted 3 lines

possibly also older
7624ee72aa09334af072853457a5d46d9901c3f8
(in 2.6.36-rc1)

This one alone applied to 2.6.35 didn't solve it.

jk@clyde:~$ time sync

real 19m2.650s
user 0m0.000s
sys 0m0.030s

Without notable disk-activity while waiting. (<5MB/s measured using dstat) This sample was "decent" I've seen sync times over 2 horus on this system with 128GB of memory.. again without any notable disk-activity while waiting.