RELENG_6 panic under heavy load

[David Xu]

On Tuesday 12 December 2006 20:22, Dmitriy Kirhlarov wrote:
> On Thu, Dec 07, 2006 at 11:18:52AM +0800, David Xu wrote:
> > On Thursday 16 November 2006 19:15, Gleb Smirnoff wrote:
> > > On Thu, Nov 16, 2006 at 01:24:36PM +0300, Gleb Smirnoff wrote:
> > > T>   I wonder why UMA was suspected to be the problem. Dima gave
> > > T> me access to the core. Here are more details from the trace:
> > >
> > > It looks like a race between two threads in one process. Look here:
> >
> > Can you try the patch ?
> > http://people.freebsd.org/~davidxu/patch/ksegrp_preempt.patch
>
> I've tested it. This patch works also, but with a little bit different
> behaviour. With patch from jhb@ I got LA 7-8, with this patch I have
> LA 5-6, same as on unpatched system. But it seems to me, that system
> is less interactive, compared to jhb@ patch.
>
> WBR
> Dmitriy

jhb patch is incomplete, it implies that every place a thread is doing state
transition and waking another thread up should be patched, there is
other code in kern_sig.c unpatched, though I don't know other places,
but the code maybe_preempt_in_ksegrp should be synced with
maybe_preempt, it should fix all problems. the LA you have seen is lower
than jhb might be a nature of KSEGRP, but I am not sure, if you program
forces all threads to be system-scope, it might fix the problem. 

David Xu