[Intel-gfx] [PATCH] drm/i915: remove user GTT mappings early during runtime suspend

[Imre Deak]

On Tue, 2014-05-06 at 21:27 +0200, Daniel Vetter wrote:
> On Tue, May 06, 2014 at 05:46:01PM +0300, Imre Deak wrote:
> > On Tue, 2014-05-06 at 12:59 +0100, Chris Wilson wrote:
> > > On Tue, May 06, 2014 at 02:42:26PM +0300, Imre Deak wrote:
> > > > On Tue, 2014-05-06 at 12:40 +0100, Chris Wilson wrote:
> > > > > On Tue, May 06, 2014 at 02:28:50PM +0300, Imre Deak wrote:
> > > > > > Currently user space can access GEM buffers mapped to GTT through
> > > > > > existing mappings concurrently while the platform specific suspend
> > > > > > handlers are running.  Since these handlers may change the HW state in a
> > > > > > way that would break such accesses, remove the mappings before calling
> > > > > > the handlers.
> > > > > 
> > > > > Hmm, but you never locked the device, so what is preventing those
> > > > > concurrent accesses from refaulting in GTT entires anyway. Please explain
> > > > > the context under which the runtime suspend code executes, and leave
> > > > > that explanation within easy reach of intel_runtime_suspend() -
> > > > > preferrably with testing of those assumptions.
> > > > 
> > > > During faulting we take an RPM reference, so that avoids concurrent
> > > > re-faults. I could add a comment about this to the code.
> > > 
> > > You are still iterating over lists that are not safe, right? Or are
> > > there more magic rpm references that prevent ioctls whilst
> > > intel_runtime_suspend is being called?
> > 
> > Tbh I haven't checked this, since moving the unmapping earlier doesn't
> > make a difference in this regard.
> > 
> > But it's a good point, I tried to audit now those paths. Currently the
> > assumption is that we hold an RPM reference everywhere where those lists
> > are changed. On the exec buffer path this is true, but for example in
> > i915_drop_caches_set() it's not.
> > 
> > We could fix this by taking struct_mutex around
> > i915_gem_release_all_mmaps() in intel_runtime_suspend(). Doing that
> > needs some more auditing to make sure we can't deadlock somehow. At
> > first glance it seems that the driver always schedules a deferred work
> > and calls intel_runtime_suspend() from that, so I think it's fine.
> > 
> > I suggest applying this patch regardless since the two issues are
> > separate.
> 
> If I understand the situation correctly the runtime suspend function only
> ever gets called from a worker thread after the hysteris timeout expired.
> Which means we should be able to wrap _just_ the gtt pte shotdown with
> dev->struct_mutex and nothing else. Which is good since profileration of
> dev->struct_mutex is awful.
>
> On the resume side we don't need any locking since the gtt fault handler
> will first grab the runtime reference and also dev->struct_mutex.
> 
> One issue which is looming is that this might deadlock. We might need a
> trylock in the runtime suspend function and abort the runtime suspend if
> we can't get the lock. Please test that lockdep catches this before we
> commit to a design.

After looking some more at different possible solutions and the RPM core
this looks the easiest way. There could be cleaner ones, for example
rearranging the order everywhere of getting struct_mutex and RPM ref in
the same order, so that we always get the RPM ref outside of
struct_mutex. By this we could just take the mutex during runtime
suspend without the possibility of deadlocking. But this would need a
lot of change due to the RPM get in i915_gem_free_object(). 

It's also clear that we need the trylock, since an RPM get with a struct
mutex already held can can block on a pending RPM put, and so getting
the mutex in the suspend handler could deadlock. In this case we can
fail the suspend with EAGAIN, so it'll get scheduled again. 

--Imre

> Just a very quick analysis, I didn't check the details so this might be
> horribly wrong.

> -Daniel