Re: [PATCH v4 06/21] drm/i915: Prepare to dynamic dma-buf locking specification

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 9/2/22 19:26, Ruhl, Michael J wrote:
>> 02.09.2022 13:31, Dmitry Osipenko пишет:
>>> 01.09.2022 17:02, Ruhl, Michael J пишет:
>>> ...
>>>>> --- a/drivers/gpu/drm/i915/gem/i915_gem_object.c
>>>>> +++ b/drivers/gpu/drm/i915/gem/i915_gem_object.c
>>>>> @@ -331,7 +331,19 @@ static void __i915_gem_free_objects(struct
>>>>> drm_i915_private *i915,
>>>>> 			continue;
>>>>> 		}
>>>>>
>>>>> +		/*
>>>>> +		 * dma_buf_unmap_attachment() requires reservation to be
>>>>> +		 * locked. The imported GEM shouldn't share reservation lock,
>>>>> +		 * so it's safe to take the lock.
>>>>> +		 */
>>>>> +		if (obj->base.import_attach)
>>>>> +			i915_gem_object_lock(obj, NULL);
>>>>
>>>> There is a lot of stuff going here.  Taking the lock may be premature...
>>>>
>>>>> 		__i915_gem_object_pages_fini(obj);
>>>>
>>>> The i915_gem_dmabuf.c:i915_gem_object_put_pages_dmabuf is where
>>>> unmap_attachment is actually called, would it make more sense to make
>>>> do the locking there?
>>>
>>> The __i915_gem_object_put_pages() is invoked with a held reservation
>>> lock, while freeing object is a special time when we know that GEM is
>>> unused.
>>>
>>> The __i915_gem_free_objects() was taking the lock two weeks ago until
>>> the change made Chris Wilson [1] reached linux-next.
>>>
>>> [1]
>>> https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-
>> next.git/commit/?id=2826d447fbd60e6a05e53d5f918bceb8c04e315c
>>>
>>> I don't think we can take the lock within
>>> i915_gem_object_put_pages_dmabuf(), it may/should deadlock other code
>> paths.
>>
>> On the other hand, we can check whether the GEM's refcount number is
>> zero in i915_gem_object_put_pages_dmabuf() and then take the lock if
>> it's zero.
>>
>> Also, seems it should be possible just to bail out from
>> i915_gem_object_put_pages_dmabuf() if refcount=0. The further
>> drm_prime_gem_destroy() will take care of unmapping. Perhaps this could
>> be the best option, I'll give it a test.
> 
> i915_gem_object_put_pages() is uses the SG, and the usage for
> drm_prim_gem_destroy()
> 
> from __i915_gem_free_objects() doesn't use the SG because it has been "freed"
> already, I am not sure if that would work...
> 
> Hmm.. with that in mind, maybe moving the base.import_attach check to 
> __i915_gem_object_put_pages with your attach check?

I see you meant __i915_gem_object_pages_fini() here.

> 	atomic_set(&obj->mm.pages_pin_count, 0);
> 	if (obj->base.import)
> 		i915_gem_object_lock(obj, NULL);
> 
> 	__i915_gem_object_put_pages(obj);
> 
> 	if (obj->base.import)
> 		i915_gem_object_unlock(obj, NULL);
> 	GEM_BUG_ON(i915_gem_object_has_pages(obj));
> 
> Pretty much one step up from the dmabuf interface, but we are guaranteed to
> not have any pinned pages?

Importer shouldn't hold pages outside of dma-buf API, otherwise it
should be a bug.

> The other caller of __i915_gem_object_pages_fini is the i915_ttm move_notify
> which should not conflict (export side, not import side).
> 
> Since it appears that not locking during the clean up is desirable, trying to make sure take the lock
> is taken at the last moment might be the right path?

Reducing the scope of locking usually is preferred more. Yours
suggestion works okay, I couldn't spot any problems at least for a
non-TTM code paths.

It's indeed a bit not nice that __i915_gem_object_pages_fini() is used
by TTM, but should be safe for imported objects. Will be great if anyone
from i915 maintainers could ack this variant.

-- 
Best regards,
Dmitry



[Index of Archives]     [Linux DRI Users]     [Linux Intel Graphics]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]     [XFree86]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Linux Kernel]     [Linux SCSI]     [XFree86]
  Powered by Linux