Re: sched: spinlock recursion in sched_rr_get_interval

[Davidlohr Bueso <dave@stgolabs.net>]

On Sat, 2014-12-27 at 10:52 -0500, Sasha Levin wrote:
> On 12/27/2014 04:52 AM, Davidlohr Bueso wrote:
> >> Hello，
> >> > Does ACCESS_ONCE() can help this issue? I have no evidence that its lack is
> >> > responsible for the issue, but I think here need it indeed. Is that right?
> >> > 
> >> > SPIN_BUG_ON(ACCESS_ONCE(lock->owner) == current, "recursion");
> > Hmm I guess on a contended spinlock, there's a chance that lock->owner
> > can change, if the contended lock is acquired, right between the 'cond'
> > and spin_debug(), which would explain the bogus ->owner related
> > messages. Of course the same applies to ->owner_cpu. Your ACCESS_ONCE,
> > however, doesn't really change anything since we still read ->owner
> > again in spin_debug; How about something like this (untested)?

I guess we'd need a writer rwlock counterpart too.

> There's a chance that lock->owner would change, but how would you explain
> it changing to 'current'?

So yeah, the above only deals with the weird printk values, not the
actual issue that triggers the BUG_ON. Lets sort this out first and at
least get correct data.

> That is, what race condition specifically creates the
> 'lock->owner == current' situation in the debug check?

Why do you suspect a race as opposed to a legitimate recursion issue?
Although after staring at the code for a while, I cannot see foul play
in sched_rr_get_interval.

Given that all reports show bogus contending CPU and .owner_cpu, I do
wonder if this is actually a symptom of the BUG_ON where something fishy
is going on.. although I have no evidence to support that. I also ran
into this https://lkml.org/lkml/2014/11/7/762 which shows the same bogus
values yet a totally different stack.

Sasha, I ran trinity with CONFIG_DEBUG_SPINLOCK=y all night without
triggering anything. How are you hitting this?

Thanks,
Davidlohr