On Tue, Apr 19, 2005 at 11:07:01AM +1000, Benjamin Herrenschmidt was heard to
remark:
> On Mon, 2005-04-18 at 14:38 -0500, Linas Vepstas wrote:
> >
> > Hi,
> >
> > The patch below appears to fix a problem where a number of dead processes
> > linger on the system. On a highly loaded system, dozens of processes
> > were found stuck in do_exit(), calling thier very last schedule(), and
> > then being lost forever.
> >
> > Processes that are PF_DEAD are cleaned up *after* the context switch,
> > in a routine called finish_task_switch(task_t *prev). The "prev" gets
> > the value returned by _switch() in entry.S, but this value comes from
> >
> > __switch_to (struct task_struct *prev,
> > struct task_struct *new)
> > {
> > old_thread = ¤t->thread; ///XXX shouldn't this be prev, not
> > current?
> > last = _switch(old_thread, new_thread);
> > return last;
> > }
> >
> > The way I see it, "prev" and "current" are almost always going to be
> > pointing at the same thing; however, if a "need resched" happens,
> > or there's a pre-emept or some-such, then prev and current won't be
> > the same; in which case, finish_task_switch() will end up cleaning
> > up the old current, instead of prev. This will result in dead processes
> > hanging around, which will never be scheduled again, and will never
> > get a chance to have put_task_struct() called on them.
>
> Ok, thinking moer about this ... that will need maybe some help from
> Ingo so I fully understand where schedule's are allowed ... We are
> basically in the middle of the scheduler here, so I wonder how much of
> the scheduler itself can be preempted or so ...
>
> Basically, under which circumstances can prev and current be different ?
I remember finding a path through void __sched schedule(void) that
took a branch through the goto need_resched; that would result in this.
I takes a bit of mental gymnastics to see how this might happen.
FWIW, I can send you a debug session showing all cpu's idle and 44 dead
processes sitting in do_exit(). All but two of these were Java threads,
so this seems to be some sort of thread-scheduling subtlty. (the two
that weren't java threads were a find|grep pair that must have gotten
tangled in.)
Given that the patch seems to fix the problem, I didn't dig much deeper.
--linas
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [EMAIL PROTECTED]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
