On Fri, Oct 07, 2011 at 08:01:55PM +0200, Peter Zijlstra wrote:

On Fri, 2011-10-07 at 10:48 -0700, Simon Kirby wrote:

quoted

Yes, they stopped locking up with d670ec13 reverted.

quoted

quoted

[ 1717.560005]  [<ffffffff8104b8d4>] task_sched_runtime+0x24/0x90
[ 1717.560005]  [<ffffffff8107c924>] thread_group_cputime+0x74/0xb0
[ 1717.560005]  [<ffffffff8107d126>] thread_group_cputimer+0xa6/0xf0
[ 1717.560005]  [<ffffffff8107d198>] cpu_timer_sample_group+0x28/0x90
[ 1717.560005]  [<ffffffff8107d3c3>] set_process_cpu_timer+0x33/0x110
[ 1717.560005]  [<ffffffff8107d4da>] update_rlimit_cpu+0x3a/0x60
[ 1717.560005]  [<ffffffff8106fe9e>] do_prlimit+0xfe/0x1f0
[ 1717.560005]  [<ffffffff8106ffd6>] sys_setrlimit+0x46/0x60
[ 1717.560005]  [<ffffffff816be292>] system_call_fastpath+0x16/0x1b

OK so that cputimer stuff is horrid and the worst part is that I cannot
seem to trigger this. You guys must have some weird userspace stuff that
I simply don't have.

I haven't tried your patch yet, but it might help to mention that on
this particular cluster, we are using CONFIG_TASK_IO_ACCOUNTING under
CONFIG_TASKSTATS, and we have process accounting enabled (w/"accton").
Perhaps that enables some other path that makes it difficult to hit
otherwise.

You can't have clouds without weather reporting, of course. :)

Other than that, it's just a typical shared web environment.

Simon-