Re: [Bug #20472] [2.6.34 -> 2.6.35] INFO: task rpcbind:14163 blocked for more than 120 seconds.

["Rafael J. Wysocki" <rjw@sisk.pl>]

On Wednesday, October 20, 2010, Paweł Sikora wrote:
> On Tuesday 19 October 2010 22:30:22 Rafael J. Wysocki wrote:
> > On Tuesday, October 19, 2010, Paweł Sikora wrote:
> > > On Monday 18 October 2010 23:36:51 Rafael J. Wysocki wrote:
> > > > On Monday, October 18, 2010, Paweł Sikora wrote:
> > > > > On Sunday 17 October 2010 22:21:48 Rafael J. Wysocki wrote:
> > > > > > This message has been generated automatically as a part of a summary report
> > > > > > of recent regressions.
> > > > > > 
> > > > > > The following bug entry is on the current list of known regressions
> > > > > > from 2.6.35.  Please verify if it still should be listed and let the tracking team
> > > > > > know (either way).
> > > > > 
> > > > > hi,
> > > > > 
> > > > > hot news from the front, recent tests on the pld-linux.org vendor kernel
> > > > > show that the random task blocking in 2.6.35.7 during medium/heavy load (~4..25)
> > > > > is related to the grsecurity patch. we'll forward problem report to the
> > > > > grsec maintainer asap...
> > > > 
> > > > So, to make things clear, this is not a mainline kernel issue, is it?
> > > 
> > > i thought that was a grsecurity issue but now i have a testcase that
> > > casues task blocking on my machine with the vanilla 2.6.35.7 kernel:
> > > 
> > > steps to repoduce:
> > > 
> > > on console 1:
> > > - run 'make -j32' in kernel tree and wait several second for make forks
> > >   and medium system load (~15).
> > > 
> > > on console 2:
> > > - run 'sync' from root account to bump load more.
> > > 
> > > on console 3:
> > > - observe 'tail -f /var/log/kernel' for blocking issues.
> > > 
> > > [  360.517917] INFO: task sync:6712 blocked for more than 120 seconds.
> > > [  360.517920] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
> > > [  360.517922] sync          D 00000000ffff4c7c     0  6712   2770 0x00000000
> > > [  360.517926]  ffff88021ac71d18 0000000000000086 ffff880200000000 0000000000004000
> > > [  360.517930]  0000000000013700 0000000000013700 ffff88021ac71fd8 ffff88021ac71fd8
> > > [  360.517933]  ffff880205593020 0000000000013700 ffff88021ac71fd8 0000000000004000
> > > [  360.517936] Call Trace:
> > > [  360.517944]  [<ffffffff813b8dbd>] schedule_timeout+0x20d/0x2f0
> > > [  360.517949]  [<ffffffff8103baba>] ? enqueue_entity+0xea/0x170
> > > [  360.517951]  [<ffffffff8103bc09>] ? enqueue_task_fair+0x49/0x50
> > > [  360.517954]  [<ffffffff813b8945>] wait_for_common+0xc5/0x150
> > > [  360.517957]  [<ffffffff81040430>] ? default_wake_function+0x0/0x10
> > > [  360.517959]  [<ffffffff813b8a78>] wait_for_completion+0x18/0x20
> > > [  360.517963]  [<ffffffff8113b343>] sync_inodes_sb+0x83/0x170
> > > [  360.517967]  [<ffffffff8113f890>] ? sync_one_sb+0x0/0x20
> > > [  360.517969]  [<ffffffff8113f880>] __sync_filesystem+0x80/0x90
> > > [  360.517972]  [<ffffffff8113f8ab>] sync_one_sb+0x1b/0x20
> > > [  360.517975]  [<ffffffff8111d2c7>] iterate_supers+0x77/0xc0
> > > [  360.517978]  [<ffffffff8113f7bb>] sync_filesystems+0x1b/0x20
> > > [  360.517980]  [<ffffffff8113f92c>] sys_sync+0x1c/0x40
> > > [  360.517984]  [<ffffffff81002d6b>] system_call_fastpath+0x16/0x1b
> > > 
> > > the 'make' is running on ext4 filesystem with underlying mdadm (raid10)
> > > created from 2 caviar raid edition disks. the cpu is intel quad core
> > > Q9300 with 8GB of ddr2 ram.
> > 
> > So, how is this related to the original report?
> 
> it's easier to reproduce on single test machine (no need to playing with parallel make and nfs).
> in fact this isn't a 2.6.35.5->.7 regression but 2.6.34->2.6.35. the git-bisect found
> first commit that introduces the 'sync' task blocking problem here:
> 
> commit 7c8a3554c683f512dbcee26faedb42e4c05f12fa
> Author: Jens Axboe <jens.axboe@oracle.com>
> Date:   Tue May 18 14:29:29 2010 +0200
> 
>     writeback: ensure that WB_SYNC_NONE writeback with sb pinned is sync
>     
>     Even if the writeout itself isn't a data integrity operation, we need
>     to ensure that the caller doesn't drop the sb umount sem before we
>     have actually done the writeback.
>     
>     This is a fixup for commit e913fc82.
>     
>     Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
> 
> 
> i'll try to revert this patch from the 2.6.35.7 and test make+nfs in few days...

OK, thanks.  Jens added to the CC list.

Rafael