On Wed, 2009-07-01 at 20:15 +0200, Philippe Gerum wrote:
> On Wed, 2009-07-01 at 19:56 +0200, Jan Kiszka wrote:
> > Jan Kiszka wrote:
> > > Jan Kiszka wrote:
> > >> Gilles Chanteperdrix wrote:
> > >>> Jan Kiszka wrote:
> > >>>> Jan Kiszka wrote:
> > >>>>> Gilles Chanteperdrix wrote:
> > >>>>>> Jan Kiszka wrote:
> > >>>>>>> Jan Kiszka wrote:
> > >>>>>>>> It's still unclear what goes on precisely, we are still digging, 
> > >>>>>>>> but the
> > >>>>>>>> test system that can produce this is highly contended.
> > >>>>>>> Short update: Further instrumentation revealed that cr3 differs from
> > >>>>>>> active_mm->pgd while we are looping over that fault, ie. the kernel
> > >>>>>>> tries to fixup the wrong mm. And that means we have some open race
> > >>>>>>> window between updating cr3 and active_mm somewhere (isn't 
> > >>>>>>> switch_mm run
> > >>>>>>> in a preemptible manner now?).
> > >>>>>> Maybe the rsp is wrong and leads you to the wrong active_mm ?
> > >>>>>>
> > >>>>>>> As a first shot I disabled CONFIG_IPIPE_DELAYED_ATOMICSW, and we 
> > >>>>>>> are now
> > >>>>>>> checking if it makes a difference. Digging deeper into the code in 
> > >>>>>>> the
> > >>>>>>> meanwhile...
> > >>>>>> As you have found out in the mean time, we do not use unlocked 
> > >>>>>> context
> > >>>>>> switches on x86.
> > >>>>>>
> > >>>>> Yes.
> > >>>>>
> > >>>>> The last question I asked myself (but couldn't answer yet due to other
> > >>>>> activity) was: Where are the local_irq_disable/enable_hw around
> > >>>>> switch_mm for its Linux callers?
> > >>>> Ha, that's the point: only activate_mm is protected, but we have more
> > >>>> spots in 2.6.29 and maybe other kernels, too!
> > >>> Ok, I do not see where switch_mm is called with IRQs off. What I found,
> > >> We have two direct callers of switch_mm in sched.c and one in fs/aio.c.
> > >> Both need protection (I pushed IRQ disabling into switch_mm), but that
> > >> is not enough according to current tests. It seems to reduce to
> > >> probability of corruption, though.
> > >>
> > >>> however, is that leave_mm sets the cr3 and just clears
> > >>> active_mm->cpu_vm_mask. So, at this point, we have a discrepancy between
> > >>> cr3 and active_mm. I do not know what could happen if Xenomai could
> > >>> interrupt leave_mm between the cpu_clear and the write_cr3. From what I
> > >>> understand, switch_mm called by Xenomai upon return to root would re-set
> > >>> the bit, and re-set cr3, which would be set to the kernel cr3 right
> > >>> after that, but this would result in the active_mm.cpu_vm_mask bit being
> > >>> set instead of cleared as expected. So, maybe an irqs off section is
> > >>> missing in leave_mm.
> > >> leave_mm is already protected by its caller smp_invalidate_interrupt -
> > >> but now I'm parsing context_switch /wrt to lazy tlb.
> > >>
> > > 
> > > Hmm... lazy tlb: This means a new task is switched in and has active_mm
> > > != mm. But do_page_fault reads task->mm... Just thoughts, no clear
> > > picture yet.
> > > 
> > 
> > Looking closer at the call sites of switch_mm, I think our the problem
> > is mostly related to use_mm from fs/aio.c (customer is using aio
> > heavily). But other callers need protection, too. We are going to test
> > this patch tomorrow:
> > 
> > diff --git a/fs/aio.c b/fs/aio.c
> > index 76da125..d90fca3 100644
> > --- a/fs/aio.c
> > +++ b/fs/aio.c
> > @@ -618,13 +618,16 @@ static void use_mm(struct mm_struct *mm)
> >  {
> >     struct mm_struct *active_mm;
> >     struct task_struct *tsk = current;
> > +   unsigned long flags;
> >  
> >     task_lock(tsk);
> >     active_mm = tsk->active_mm;
> >     atomic_inc(&mm->mm_count);
> > +   local_irq_save_hw_cond(flags);
> >     tsk->mm = mm;
> >     tsk->active_mm = mm;
> >     switch_mm(active_mm, mm, tsk);
> > +   local_irq_restore_hw_cond(flags);
> >     task_unlock(tsk);
> >  
> >     mmdrop(active_mm);
> > diff --git a/kernel/sched.c b/kernel/sched.c
> > index aa8f86c..8c545a4 100644
> > --- a/kernel/sched.c
> > +++ b/kernel/sched.c
> > @@ -2668,8 +2668,12 @@ context_switch(struct rq *rq, struct task_struct 
> > *prev,
> >             next->active_mm = oldmm;
> >             atomic_inc(&oldmm->mm_count);
> >             enter_lazy_tlb(oldmm, next);
> > -   } else
> > +   } else {
> > +           unsigned long flags;
> > +           local_irq_save_hw_cond(flags);
> >             switch_mm(oldmm, mm, next);
> > +           local_irq_restore_hw_cond(flags);
> > +   }
> >  
> >     if (unlikely(!prev->mm)) {
> >             prev->active_mm = NULL;
> > @@ -6406,8 +6410,12 @@ void idle_task_exit(void)
> >  
> >     BUG_ON(cpu_online(smp_processor_id()));
> >  
> > -   if (mm != &init_mm)
> > +   if (mm != &init_mm) {
> > +           unsigned long flags;
> > +           local_irq_save_hw_cond(flags);
> >             switch_mm(mm, &init_mm, current);
> > +           local_irq_restore_hw_cond(flags);
> > +   }
> >     mmdrop(mm);
> >  }
> 
> Please fix the callee instead of ironing the call sites. This would
> avoid further issues as upstream emits additional switch_mm calls over
> time, and make ironing activate_mm useless.
> 

Btw, this how the powerpc port works in locked switch mode, and this
particular bug does not apply in unlocked switch mode anyway; this is
why we don't have that problem on this arch. The ARM port always works
in unlocked switch mode IIRC for latency reasons, so this should be a
non-issue here as well. Gilles, would you confirm this?

> >  
> > 
> > Jan
> > 
-- 
Philippe.



_______________________________________________
Xenomai-core mailing list
Xenomai-core@gna.org
https://mail.gna.org/listinfo/xenomai-core

Reply via email to