Re: [PATCH 5/9] perf/x86: implement cross-HT corruption bug workaround

[Peter Zijlstra]

On Thu, Jun 05, 2014 at 05:01:25PM +0300, Maria Dimakopoulou wrote:
> On Thu, Jun 5, 2014 at 4:48 PM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > On Wed, Jun 04, 2014 at 11:34:14PM +0200, Stephane Eranian wrote:
> >
> >> +static struct event_constraint *
> >> +intel_get_excl_constraints(struct cpu_hw_events *cpuc, struct perf_event *event,
> >> +                        struct event_constraint *c)
> >> +{
> >
> >> +     if (!(c->flags & PERF_X86_EVENT_DYNAMIC)) {
> >> +
> >> +             /*
> >> +              * in case we fail, we assume no counter
> >> +              * is supported to be on the safe side
> >> +              */
> >> +             cx = kmalloc(sizeof(*cx), GFP_KERNEL);
> >> +             if (!cx)
> >> +                     return &emptyconstraint;
> >> +
> >
> > Ok, so forgive me if I'm wrong, but the way we get here is through:
> >
> > x86_schedule_event()
> >   ->start_scheduling()
> >     spin_lock()
> >   ->get_event_constraints()
> >     intel_get_excl_constraints()
> >       kmalloc(.gfp=GFP_KERNEL)
> >
> > How can that ever work?

> Are you saying it is illegal to call kmalloc() from
> this context?

Nobody will come and arrest you for it, so no. Broken though. GFP_KERNEL
will attempt to sleep to wait for reclaim, and you're holding a
spinlock.

> kmalloc is needed because we need to allocate
> a new constraint struct since the static constraint
> cannot be modified.
> 
> Worst case we can statically allocate a second
> constraint struct in the event struct.

Nah, since you will need at most one constraint per counter, you could
preallocate num_counter constraints for each cpu.

Attachment: pgpc_NGyHX0Mn.pgp
Description: PGP signature