Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Message-ID: <a3016b01-1456-d571-a44d-6ed6cd6a66a0@gmail.com>
Date:   Sun, 6 Feb 2022 18:11:24 +0100
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:91.0) Gecko/20100101
 Thunderbird/91.5.0
Subject: Re: [BUG] gpu: drm: radeon: two possible deadlocks involving locking
 and waiting
Content-Language: en-US
To:     Jia-Ju Bai <baijiaju1990@gmail.com>,
        =?UTF-8?Q?Christian_K=c3=b6nig?= <christian.koenig@amd.com>,
        alexander.deucher@amd.com, Xinhui.Pan@amd.com, airlied@linux.ie,
        daniel@ffwll.ch
Cc:     dri-devel@lists.freedesktop.org, amd-gfx@lists.freedesktop.org,
        linux-kernel <linux-kernel@vger.kernel.org>
References: <d5e4460f-7e26-81d2-2efe-6f47760b78d2@gmail.com>
 <7cdc2d3f-df52-f7a9-15bf-fe4bc01d3c4f@amd.com>
 <8fa82beb-468c-afb0-3eed-4240200395a3@gmail.com>
From:   =?UTF-8?Q?Christian_K=c3=b6nig?= <ckoenig.leichtzumerken@gmail.com>
In-Reply-To: <8fa82beb-468c-afb0-3eed-4240200395a3@gmail.com>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 8bit
Precedence: bulk


Am 05.02.22 um 02:55 schrieb Jia-Ju Bai:
> Hi Christian,
>
> Thanks for the reply :)
>
> On 2022/2/1 15:56, Christian König wrote:
>> Hi Jia-Ju,
>>
>> interesting that you have found those issues with an automated tool.
>>
>> And yes that is a well design flaw within the radeon driver which can 
>> happen on hardware faults, e.g. when radeon_ring_backup() needs to be 
>> called.
>
> In fact, my tool finds dozens of similar possible deadlocks caused by 
> wait_event_timeout() in radeon_fence_wait_seq_timeout().

Those are false positives.

The call to radeon_fence_process() from radeon_fence_count_emitted() for 
example is just to speed things up, it's not mandatory for correct 
operation and so it doesn't matter if it isn't called because of the 
thread is blocked on the pm.mutex.

But I also don't see how your tool should be able to figure that out 
automated.

Regards,
Christian.

> There are three other examples in Linux 5.16:
>
> #BUG 1
> radeon_dpm_change_power_state_locked()
>   mutex_lock(&rdev->ring_lock); --> Line 1133 (Lock A)
>   radeon_fence_wait_empty()
>     radeon_fence_wait_seq_timeout()
>       wait_event_timeout(rdev->fence_queue, ...) --> Line 504 (Wait X)
>
> radeon_fence_driver_fini()
>   mutex_lock(&rdev->ring_lock); --> Line 917 (Lock A)
>   wake_up_all(&rdev->fence_queue); --> Line 927 (Wake X)
>
> #BUG 2
> radeon_set_pm_profile()
>   mutex_lock(&rdev->pm.mutex); --> Line 382 (Lock A)
>   radeon_pm_set_clocks()
>     radeon_fence_wait_empty()
>       radeon_fence_wait_seq_timeout()
>         wait_event_timeout(rdev->fence_queue, ...) --> Line 504 (Wait X)
>
> radeon_dynpm_idle_work_handler()
>   mutex_lock(&rdev->pm.mutex); --> Line 1861 (Lock A)
>   radeon_fence_count_emitted()
>     radeon_fence_process()
>       wake_up_all(&rdev->fence_queue); --> Line 323 (Wake X)
>
> #BUG 3
> radeon_pm_fini_old()
>   mutex_lock(&rdev->pm.mutex); --> Line 1642 (Lock A)
>   radeon_pm_set_clocks()
>     radeon_fence_wait_empty()
>       radeon_fence_wait_seq_timeout()
>         wait_event_timeout(rdev->fence_queue, ...) --> Line 504 (Wait X)
>
> radeon_dynpm_idle_work_handler()
>   mutex_lock(&rdev->pm.mutex); --> Line 1861 (Lock A)
>   radeon_fence_count_emitted()
>     radeon_fence_process()
>       wake_up_all(&rdev->fence_queue); --> Line 323 (Wake X)
>
> Thus, to fix these possible deadlocks, we could moditify the code 
> related to radeon_fence_wait_seq_timeout().
> But I am not quite familar with the radeon driver, so I am not sure 
> how to moditify the code properly.
>
>>
>> But that happens so rarely and the driver is not developed further 
>> that we decided to not address this any more.
>
> Ah, okay.
>
>>
>> Regards,
>> Christian.
>>
>> Am 01.02.22 um 08:40 schrieb Jia-Ju Bai:
>>> Hello,
>>>
>>> My static analysis tool reports a possible deadlock in the radeon 
>>> driver in Linux 5.16:
>>>
>>> #BUG 1
>>> radeon_dpm_change_power_state_locked()
>>>   mutex_lock(&rdev->ring_lock); --> Line 1133 (Lock A)
>>>   radeon_fence_wait_empty()
>>>     radeon_fence_wait_seq_timeout()
>>>       wait_event_timeout(rdev->fence_queue, ...) --> Line 504 (Wait X)
>>>
>>> radeon_ring_backup()
>>>   mutex_lock(&rdev->ring_lock); --> Line 289(Lock A)
>>>   radeon_fence_count_emitted()
>>>     radeon_fence_process()
>>>       wake_up_all(&rdev->fence_queue); --> Line 323 (Wake X)
>>>
>>> When radeon_dpm_change_power_state_locked() is executed, "Wait X" is 
>>> performed by holding "Lock A". If radeon_ring_backup() is executed 
>>> at this time, "Wake X" cannot be performed to wake up "Wait X" in 
>>> radeon_dpm_change_power_state_locked(), because "Lock A" has been 
>>> already hold by radeon_dpm_change_power_state_locked(), causing a 
>>> possible deadlock.
>>> I find that "Wait X" is performed with a timeout 
>>> MAX_SCHEDULE_TIMEOUT, to relieve the possible deadlock; but I think 
>>> this timeout can cause inefficient execution.
>>>
>>> #BUG 2
>>> radeon_ring_lock()
>>>   mutex_lock(&rdev->ring_lock); --> Line 147 (Lock A)
>>>   radeon_ring_alloc()
>>>     radeon_fence_wait_next()
>>>       radeon_fence_wait_seq_timeout()
>>>         wait_event_timeout(rdev->fence_queue, ...) --> Line 504 
>>> (Wait X)
>>>
>>> radeon_ring_backup()
>>>   mutex_lock(&rdev->ring_lock); --> Line 289(Lock A)
>>>   radeon_fence_count_emitted()
>>>     radeon_fence_process()
>>>       wake_up_all(&rdev->fence_queue); --> Line 323 (Wake X)
>>>
>>> When radeon_ring_lock() is executed, "Wait X" is performed by 
>>> holding "Lock A". If radeon_ring_backup() is executed at this time, 
>>> "Wake X" cannot be performed to wake up "Wait X" in 
>>> radeon_ring_lock(), because "Lock A" has been already hold by 
>>> radeon_ring_lock(), causing a possible deadlock.
>>> I find that "Wait X" is performed with a timeout 
>>> MAX_SCHEDULE_TIMEOUT, to relieve the possible deadlock; but I think 
>>> this timeout can cause inefficient execution.
>>>
>>> I am not quite sure whether these possible problems are real and how 
>>> to fix them if they are real.
>>> Any feedback would be appreciated, thanks :)
>>>
>>>
>>> Best wishes,
>>> Jia-Ju Bai
>>>
>>
>