Bug#584881: NeilBrown added a cond_resched
> Maybe someone can verify whether the patch contributed by Neil Brown
> [ http://www.ktrap.org/mailarchive/git-commits-head/2009/10/31/11499 ]
> fixes this? Should be available since 2.6.32.
Thanks for the suggestion, but it's not clear to me from the commit
message and the content that that's addressing a deadlock rather than
a performance problem.
When my system is in the deadlocked state, there seems to be no
problem finding CPUs to run runnable processes on - just a shortage of
runnable processes that can make progress, because everything is stuck
in disk wait. So additional reschedules wouldn't seem likely to help.
chiark has a total of 8 cores and only two md devices, only one of
which is a mirror (the other is stripe) so I don't think a single
management thread could cause total cpu starvation anyway.