[PATCH v16 15/20] drm/shmem-helper: Add memory shrinker

[Dmitry Osipenko]

On 9/14/23 11:27, Boris Brezillon wrote:
> On Thu, 14 Sep 2023 10:50:32 +0300
> Dmitry Osipenko <dmitry.osipenko at collabora.com> wrote:
> 
>> On 9/14/23 10:36, Boris Brezillon wrote:
>>> On Thu, 14 Sep 2023 07:02:52 +0300
>>> Dmitry Osipenko <dmitry.osipenko at collabora.com> wrote:
>>>   
>>>> On 9/13/23 10:48, Boris Brezillon wrote:  
>>>>> On Wed, 13 Sep 2023 03:56:14 +0300
>>>>> Dmitry Osipenko <dmitry.osipenko at collabora.com> wrote:
>>>>>     
>>>>>> On 9/5/23 11:03, Boris Brezillon wrote:    
>>>>>>>>                * But
>>>>>>>> +		 * acquiring the obj lock in drm_gem_shmem_release_pages_locked() can
>>>>>>>> +		 * cause a locking order inversion between reservation_ww_class_mutex
>>>>>>>> +		 * and fs_reclaim.
>>>>>>>> +		 *
>>>>>>>> +		 * This deadlock is not actually possible, because no one should
>>>>>>>> +		 * be already holding the lock when drm_gem_shmem_free() is called.
>>>>>>>> +		 * Unfortunately lockdep is not aware of this detail.  So when the
>>>>>>>> +		 * refcount drops to zero, don't touch the reservation lock.
>>>>>>>> +		 */
>>>>>>>> +		if (shmem->got_pages_sgt &&
>>>>>>>> +		    refcount_dec_and_test(&shmem->pages_use_count)) {
>>>>>>>> +			drm_gem_shmem_do_release_pages_locked(shmem);
>>>>>>>> +			shmem->got_pages_sgt = false;
>>>>>>>>  		}      
>>>>>>> Leaking memory is the right thing to do if pages_use_count > 1 (it's
>>>>>>> better to leak than having someone access memory it no longer owns), but
>>>>>>> I think it's worth mentioning in the above comment.      
>>>>>>
>>>>>> It's unlikely that it will be only a leak without a following up
>>>>>> use-after-free. Neither is acceptable.    
>>>>>
>>>>> Not necessarily, if you have a page leak, it could be that the GPU has
>>>>> access to those pages, but doesn't need the GEM object anymore
>>>>> (pages are mapped by the iommu, which doesn't need shmem->sgt or
>>>>> shmem->pages after the mapping is created). Without a WARN_ON(), this
>>>>> can go unnoticed and lead to memory corruptions/information leaks.
>>>>>     
>>>>>>
>>>>>> The drm_gem_shmem_free() could be changed such that kernel won't blow up
>>>>>> on a refcnt bug, but that's not worthwhile doing because drivers
>>>>>> shouldn't have silly bugs.    
>>>>>
>>>>> We definitely don't want to fix that, but we want to complain loudly
>>>>> (WARN_ON()), and make sure the risk is limited (preventing memory from
>>>>> being re-assigned to someone else by not freeing it).    
>>>>
>>>> That's what the code did and continues to do here. Not exactly sure what
>>>> you're trying to say. I'm going to relocate the comment in v17 to
>>>> put_pages(), we can continue discussing it there if I'm missing yours point.
>>>>  
>>>
>>> I'm just saying it would be worth mentioning that we're intentionally
>>> leaking memory if shmem->pages_use_count > 1. Something like:
>>>
>>> 	/**
>>> 	 * shmem->pages_use_count should be 1 when ->sgt != NULL and
>>> 	 * zero otherwise. If some users still hold a pages reference
>>> 	 * that's a bug, and we intentionally leak the pages so they
>>> 	 * can't be re-allocated to someone else while the GPU/CPU
>>> 	 * still have access to it.
>>> 	 */
>>> 	drm_WARN_ON(drm,
>>> 		    refcount_read(&shmem->pages_use_count) == (shmem->sgt ? 1 : 0));
>>> 	if (shmem->sgt && refcount_dec_and_test(&shmem->pages_use_count))
>>> 		drm_gem_shmem_free_pages(shmem);  
>>
>> That may be acceptable, but only once there will a driver using this
>> feature.
> 
> Which feature? That's not related to a specific feature, that's just
> how drm_gem_shmem_get_pages_sgt() works, it takes a pages ref that can
> only be released in drm_gem_shmem_free(), because sgt users are not
> refcounted and the sgt stays around until the GEM object is freed or
> its pages are evicted. The only valid cases we have at the moment are:
> 
> - pages_use_count == 1 && sgt != NULL
> - pages_use_count == 0
> 
> any other situations are buggy.

sgt may belong to dma-buf for which pages_use_count=0, this can't be
done until sgt mess is sorted out

-- 
Best regards,
Dmitry