[Intel-gfx] [RFC PATCH v3 4/4] tests/drv_module_reload: add ipvr support

[Daniel Vetter]

On Sun, Dec 21, 2014 at 02:40:24PM +0000, Cheng, Yao wrote:
> > -----Original Message-----
> > From: Daniel Vetter [mailto:daniel.vetter at ffwll.ch]
> > Sent: Thursday, December 18, 2014 19:21
> > To: Thierry Reding
> > Cc: Cheng, Yao; intel-gfx at lists.freedesktop.org; dri-
> > devel at lists.freedesktop.org; Kelley, Sean V; Chehab, John;
> > emil.l.velikov at gmail.com; Jiang, Fei
> > Subject: Re: [RFC PATCH v3 4/4] tests/drv_module_reload: add ipvr support
> > 
> > On Thu, Dec 18, 2014 at 11:04 AM, Thierry Reding <thierry.reding at gmail.com>
> > wrote:
> > >> I double checked the symptom and found it was a deadlock on
> > drm_global_mutex.
> > >> When i915_driver_load() registers the platform device while ipvr module
> > is in the system, ipvr's probe() function tries to lock drm_global_mutex which
> > was already held by i915.
> > >> I think either of the following 2 actions need to be moved to a bottom half
> > e.g. a work queue:
> > >>       platform_device_add () call in i915_ved.c (called during
> > i915_driver_load())
> > >>       drm_dev_register() call during ipvr's probe() Which one makes
> > >> more sense? pls kindly advise (I personally prefer the former one.).
> > >
> > > Yes, that's somewhat ugly, but I don't see a way around that. I'd also
> > > think that moving platform_device_add() to a workqueue would be the
> > > best option here.
> > 
> > Or we simply kill drm_global_mutex for platform drivers that don't use the -
> > >probe hook. It should work when they have a correct order betwen
> > drm_dev_alloc and _register and all the code in between. So just ditch the -
> > >load callback in teh ipvr driver and rework the load sequence as suggested
> > somewhere else and this is fixed already. No need for bottom halfs I think.
> 
> Daniel, sorry I didn't quite understand "platform drivers that don't use
> the probe hook". For initialization, the ipvr platform driver's probe()
> is called in following 2 possible paths:
> 1. ipvr installed before i915. In this case, ipvr's probe() is called
> inside i915_driver_load() and falls into the drm_global_mutex dead lock.
> 2. i915 installed before ipvr. In this case, ipvr's probe() is called
> without drm_global_mutex held by i915 and no dead lock issue.
> If we kill drm_global_mutex, will path 2 run into issue? And in your
> suggestion, how to rework the load sequence? Do you mean calling ipvr's
> load() callback directly during platform driver probe()?

Hm right it's not that simple really. What we need in more detail is:
- Move the mutex_lock(&drm_global_mutex) out of drm_dev_register into
  all the callers. If a driver has a ->load() callback it most likely is
  racy with the usual load ordering issues.

- Rework ipvr to no longer have a ->load callback. Insteaed use the
  following sequence (in the platform ->probe callback):

  drm_dev_alloc();
  ipvr_load();
  drm_dev_register();

  With that ordering we don't need the additional guarantees that
  drm_global_mutex provides and we can avoid to take that lock around
  drm_dev_registrer() call in the ipvr code.

This should resolve the deadlock I hope.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch