[Intel-gfx] [PATCH 2/3] drm/i915: unpin backing storage in dmabuf_unmap

[Daniel Vetter]

On Wed, Aug 7, 2013 at 11:29 AM, Chris Wilson <chris at chris-wilson.co.uk> wrote:
> On Wed, Aug 07, 2013 at 11:15:07AM +0200, Daniel Vetter wrote:
>> This fixes a WARN in i915_gem_free_object when the
>> obj->pages_pin_count isn't 0.
>>
>> Reported-by: Maarten Lankhorst <maarten.lankhorst at canonical.com>
>> Cc: Maarten Lankhorst <maarten.lankhorst at canonical.com>
>> Signed-off-by: Daniel Vetter <daniel.vetter at ffwll.ch>
>
> Papers over the WARN with iffy locking. Not all callers hold
> struct_mutex, right? Worse some do, some don't...

Oops, that needs mutex locking here. I've checked the code and none of
the callers here  should ever hold our own dev->struct_mutex (due to
the self-import checks we bypass dma-buf for our own objects) so no
immediate deadlock. But it's easy to create circles and piss off
lockded ofc.

> What's the long term plan here?

Per-bo locking with ww mutexes. Locking is done by the callers of
map/unmap, sonce only those can properly do the ww locking dance on
all relevent buffers of a batch upfront. It's going to be fun ;-)

Looking closer I've also spotted that our map_buf callback has a call
to i915_mutex_lock_interruptible, which means the map can fail with
-EINTR. Currently it seems unspec'ed whether map is allowed to do
that, but current callers certainly can't cope with this. I'll throw a
2nd patch on top.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch