[Intel-gfx] [BUG] completely bonkers use of set_need_resched + VM_FAULT_NOPAGE

[Peter Zijlstra]

On Thu, Sep 12, 2013 at 05:58:49PM +0200, Daniel Vetter wrote:
> On Thu, Sep 12, 2013 at 5:43 PM, Peter Zijlstra <peterz at infradead.org> wrote:
> >> The one in ttm is just bonghits to shut up lockdep: ttm can recurse
> >> into it's own pagefault handler and then deadlock, the trylock just
> >> keeps lockdep quiet. We've had that bug arise in drm/i915 due to some
> >> fun userspace did and now have testcases for them. The right solution
> >> to fix this is to use copy_to|from_user_atomic in ttm everywhere it
> >> holds locks and have slowpaths which drops locks, copies stuff into a
> >> temp allocation and then continues. At least that's how we've fixed
> >> all those inversions in i915-gem. I'm not volunteering to fix this ;-)
> >
> > Yikes.. so how common is it? If I simply rip the set_need_resched() out
> > it will 'spin' on the fault a little longer until a 'natural' preemption
> > point -- if such a thing is every going to happen.
> 
> It's a case of "our userspace doesn't do this", so as long as you're
> not evil and frob the drm device nodes of ttm drivers directly the
> deadlock will never happen. No idea how much contention actually
> happens on e.g. shared buffer objects - in i915 we have just one lock
> and so suffer quite a bit more from contention. So no idea how much
> removing the yield would hurt.

If 'sane' userspace is never supposed to do this, then only insane
userspace is going to hurt from this and that's a GOOD (tm) thing,
right? ;-)

And it won't actually deadlock if you don't use FIFO, for the regular
scheduler class it'll just spin a little longer before getting preempted
so no real worries there.