Re: regression introduced by - timers: fix itimer/many thread hang

[Petr Tesarik <ptesarik@suse.cz>]

Dne Friday 07 of November 2008 11:29:04 Peter Zijlstra napsal(a):
> (fwiw your email doesn't come across properly, evo refuses to display
> them, there's some mangling of headers which makes it think there's an
> attachment)
>
> On Thu, 2008-11-06 at 15:52 -0800, Frank Mayhar wrote:
> > On Thu, 2008-11-06 at 16:08 +0100, Peter Zijlstra wrote:
> > > On Thu, 2008-11-06 at 09:03 -0600, Christoph Lameter wrote:
> > > > On Thu, 6 Nov 2008, Peter Zijlstra wrote:
> > > > > Also, you just introduced per-cpu allocations for each
> > > > > thread-group, while Christoph is reworking the per-cpu allocator,
> > > > > with one unfortunate side-effect - its going to have a limited size
> > > > > pool. Therefore this will limit the number of thread-groups we can
> > > > > have.
> > > >
> > > > Patches exist that implement a dynamically growable percpu pool
> > > > (using virtual mappings though). If the cost of the additional
> > > > complexity / overhead is justifiable then we can make the percpu pool
> > > > dynamically extendable.
> > >
> > > Right, but I don't think the patch under consideration will fly anyway,
> > > doing a for_each_possible_cpu() loop on every tick on all cpus isn't
> > > really healthy, even for moderate sized machines.
> >
> > I personally think that you're overstating this.  First, the current
> > implementation walks all threads for each tick, which is simply not
> > scalable and results in soft lockups with large numbers of threads.
> > This patch fixes a real bug.  Second, this only happens "on every tick"
> > for processes that have more than one thread _and_ that use posix
> > interval timers.  Roland and I went to some effort to keep loops like
> > the on you're referring to out of the common paths.
> >
> > In any event, while this particular implementation may not be optimal,
> > at least it's _right_.  Whatever happened to "make it right, then make
> > it fast?"
>
> Well, I'm not thinking you did it right ;-)
>
> While I agree that the linear loop is sub-optimal, but it only really
> becomes a problem when you have hundreds or thousands of threads in your
> application, which I'll argue to be insane anyway.

This is just not true. I've seen a very real example of a lockup with a very 
sane number of threads (one per CPU), but on a very large machine (1024 CPUs 
IIRC). The application set per-process CPU profiling with an interval of 1 
tick, which translates to 1024 timers firing off with each tick...

Well, yes, that was broken, too, but that's the way one quite popular FORTRAN 
compiler works...

Petr