[Intel-gfx] [PATCH v13 01/10] drm/shmem-helper: Switch to reservation lock

[Christian König]

Am 25.03.23 um 15:58 schrieb Dmitry Osipenko:
> On 3/15/23 16:46, Dmitry Osipenko wrote:
>> On 3/14/23 05:26, Dmitry Osipenko wrote:
>>> @@ -633,7 +605,10 @@ int drm_gem_shmem_mmap(struct drm_gem_shmem_object *shmem, struct vm_area_struct
>>>   		return ret;
>>>   	}
>>>   
>>> +	dma_resv_lock(shmem->base.resv, NULL);
>>>   	ret = drm_gem_shmem_get_pages(shmem);
>>> +	dma_resv_unlock(shmem->base.resv);
>> Intel CI reported locking problem [1] here. It actually was also
>> reported for v12, but I missed that report because of the other noisy
>> reports.
>>
>> [1]
>> https://intel-gfx-ci.01.org/tree/drm-tip/Patchwork_114671v2/shard-snb5/igt@prime_vgem@sync@rcs0.html
>>
>> The test does the following:
>>
>> 1. creates vgem
>> 2. export it do dmabuf
>> 3. mmaps dmabuf
>>
>> There is an obvious deadlock there. The DRM code assumes that mmap() is
>> unlocked, while for dma-buf it's locked. I'll see how to fix it for v14.
>>
> Christian, there is a deadlock problem in drm_gem_shmem_mmap() once we
> move drm-shmem to use resv lock. The current dma-buf locking policy
> claims that importer holds the lock for mmap(), but DRM code assumes
> that obj->mmap() handles the locking itself and then the same
> obj->mmap() code path is used by both dma-buf DRM and a usual DRM object
> paths. Hence importer -> dma_buf_mmap_internal()[takes the lock] ->
> exporter -> drm_gem_shmem_mmap()[takes the lock] deadlocks.
>
> I was looking at how to fix it and to me the best option is to change
> the dma-buf locking policy, making exporter responsible for handling the
> resv lock. Changing DRM code mmap lockings might be possible too [moving
> locking to drm_gem_mmap_obj()], but will be very messy and doesn't feel
> intuitive.
>
> Want to get yours thoughts on this before sending out the dma-buf mmap()
> policy-change patch. Does the new mmap() locking policy sound good to
> you? Thanks!


IIRC we tried that before and ran into problems.

dma_buf_mmap() needs to swap the backing file of the VMA and for this 
calls fput() on the old file.

This fput() in turn could (in theory) grab the resv lock as well and 
there isn't anything we could do about that.

Just information from the back of my memory, probably best if you double 
check that.

Regards,
Christian.