[PATCH v2 1/2] drm/amdgpu: fix a potential deadlock in amdgpu_bo_create_restricted()

Tue Feb 14 02:56:32 UTC 2017


On 2017年02月14日 03:03, Nicolai Hähnle wrote:
> On 13.02.2017 19:58, Nicolai Hähnle wrote:
>> On 13.02.2017 19:38, Samuel Pitoiset wrote:
>>>
>>>
>>> On 02/13/2017 07:09 PM, Nicolai Hähnle wrote:
>>>> On 13.02.2017 19:04, Nicolai Hähnle wrote:
>>>>> On 13.02.2017 18:49, Samuel Pitoiset wrote:
>>>>>>
>>>>>>
>>>>>> On 02/13/2017 05:25 PM, Nicolai Hähnle wrote:
>>>>>>> On 09.02.2017 11:33, Samuel Pitoiset wrote:
>>>>>>>> When ttm_bo_init() fails, the reservation mutex should be 
>>>>>>>> unlocked.
>>>>>>>>
>>>>>>>> In debug build, the kernel reported "possible recursive locking
>>>>>>>> detected" in this codepath. For debugging purposes, I also added
>>>>>>>> a "WARN_ON(ww_mutex_is_locked())" when ttm_bo_init() fails and the
>>>>>>>> mutex was locked as expected.
>>>>>>>>
>>>>>>>> This should fix (random) GPU hangs. The easy way to reproduce the
>>>>>>>> issue is to change the "Super Sampling" option from 1.0 to 2.0 in
>>>>>>>> Hitman. It will create a huge buffer, evict a bunch of buffers
>>>>>>>> (around ~5k) and deadlock.
>>>>>>>>
>>>>>>>> This regression has been introduced pretty recently.
>>>>>>>>
>>>>>>>> v2: only release the mutex if resv is NULL
>>>>>>>>
>>>>>>>> Fixes: 12a852219583 ("drm/amdgpu: improve
>>>>>>>> AMDGPU_GEM_CREATE_VRAM_CLEARED handling (v2)")
>>>>>>>> Signed-off-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>
>>>>>>>> ---
>>>>>>>>  drivers/gpu/drm/amd/amdgpu/amdgpu_object.c | 5 ++++-
>>>>>>>>  1 file changed, 4 insertions(+), 1 deletion(-)
>>>>>>>>
>>>>>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
>>>>>>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
>>>>>>>> index d1ef1d064de4..556236a112c1 100644
>>>>>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
>>>>>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
>>>>>>>> @@ -403,8 +403,11 @@ int amdgpu_bo_create_restricted(struct
>>>>>>>> amdgpu_device *adev,
>>>>>>>>              &bo->placement, page_align, !kernel, NULL,
>>>>>>>>              acc_size, sg, resv ? resv : &bo->tbo.ttm_resv,
>>>>>>>>              &amdgpu_ttm_bo_destroy);
>>>>>>>> -    if (unlikely(r != 0))
>>>>>>>> +    if (unlikely(r != 0)) {
>>>>>>>> +        if (!resv)
>>>>>>>> + ww_mutex_unlock(&bo->tbo.resv->lock);
>>>>>>>>          return r;
>>>>>>>> +    }
>>>>>>>
>>>>>>> I was looking at this myself a couple of weeks back, and I'm pretty
>>>>>>> sure
>>>>>>> I had this exact same patch just to realize that it's actually
>>>>>>> incorrect.
>>>>>>>
>>>>>>> The problem is that ttm_bo_init will actually call the destroy
>>>>>>> function
>>>>>>> (in our case, amdgpu_ttm_bo_destroy), so at this point, bo has been
>>>>>>> freed.
>>>>>>>
>>>>>>> This code is a huge mess. I'm surprised though: have you verified
>>>>>>> that
>>>>>>> this patch actually fixes a hang?
>>>>>>
>>>>>> Yes, I triple-checked. I can't reproduce the hangs with Hitman.
>>>>>
>>>>> That's surprising, but a relief. Maybe it ties into some of the other
>>>>> problems I'm seeing as well.
>>>>>
>>>>> This means we need a real fix for this; I still think the current 
>>>>> patch
>>>>> is broken.
>>>>>
>>>>>
>>>>>> This fixes a deadlock, here's the report:
>>>>>> https://hastebin.com/durodivoma.xml
>>>>>>
>>>>>> The resv->lock has to be unlocked when ttm_bo_init() fails (I 
>>>>>> checked
>>>>>> with a WARN_ON(is_locked)) because it doesn't call the destroy
>>>>>> function
>>>>>> in all situations. Presumably, when drm_vma_offset_add() fails and
>>>>>> resv
>>>>>> is not NULL, the mutex is not unlocked.
>>>>>
>>>>> On which code path is the destroy function not called? If that is the
>>>>> case, we're leaking memory.
>>>>>
>>>>> With the patch as-is, the error paths are either leaking memory (if
>>>>> you're right) or accessing memory after it's freed (otherwise).
>>>>> Obviously, neither is good.
>>>>
>>>> Actually, I find it extremely suspicious that this patch resolves 
>>>> hangs.
>>>> By all rights, no other task should have a pointer to this bo left. It
>>>> points at problems elsewhere in the code, possibly the precise problem
>>>> I've been trying to track down.
>>>
>>> Well, maybe we are just lucky but as I said, I checked many times to
>>> reproduce the issue with that patch applied without any success, you 
>>> can
>>> trust me. Although I'm also starting to think that's not the right
>>> solution (and could introduce other ones).
>>>
>>>>
>>>> Could you please revert the patch, reproduce the hang, and report
>>>> /proc/$pid/stack for all the hung tasks?
>>>
>>> Sure. The thing is: Hitman's branch has been updated during the weekend
>>> and my local installation is broken. I need to re-download the whole
>>> game (will take a while).
>>>
>>> I will let you know when I'm able to grab that report.
>>
>> Hmm, so I thought about this some more, and I'm no longer so sure that
>> your bug and mine are the same. If it was related, I'd somehow expect
>> you to get an error about a mutex being destroyed while it's held (at
>> least with lock debugging enabled).
>>
>> Anyway... we need to change the contract of ttm_bo_init, I'm just not
>> yet sure how, because there are two points of failure: one quite early
>> on, and the second rather late which gets cleaned up by ttm_bo_unref.
>
> Maybe it would actually be best to split ttm_bo_init into two parts: 
> the initial bulk of structure initialization as the first half, and 
> the ttm_bo_validate in the second half.
Agreed. Have you gone ahead with your proposal?
Although Samuel's patch isn't best way, it indeed fix a OCL bug which is 
trying to allocate multiple big buffers.

Thanks,
David Zhou
>
> Cheers,
> Nicolai
>
>>
>> Cheers,
>> Nicolai
>>
>>> Thanks Nicolai.
>>>>
>>>> Thanks,
>>>> Nicolai
>>
>
> _______________________________________________
> amd-gfx mailing list
> amd-gfx at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/amd-gfx