[Intel-gfx] [PATCH] drm/i915/selftests: Move test flush to outside vm->mutex

[Tvrtko Ursulin]

On 12/06/2020 15:55, Chris Wilson wrote:
> Quoting Tvrtko Ursulin (2020-06-12 15:44:51)
>> From: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>>
>> As per our locking rules it is not allowed to wait on requests while
>> holding locks. In this case we were trying to idle the GPU while holding
>> the vm->mutex.
> 
> Synchronous eviction would like to have a word.
> 
>> Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>> ---
>>   drivers/gpu/drm/i915/selftests/i915_gem_evict.c | 5 +++--
>>   1 file changed, 3 insertions(+), 2 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/i915/selftests/i915_gem_evict.c b/drivers/gpu/drm/i915/selftests/i915_gem_evict.c
>> index 028baae9631f..67f4497c8224 100644
>> --- a/drivers/gpu/drm/i915/selftests/i915_gem_evict.c
>> +++ b/drivers/gpu/drm/i915/selftests/i915_gem_evict.c
>> @@ -498,8 +498,6 @@ static int igt_evict_contexts(void *arg)
>>   
>>          mutex_lock(&ggtt->vm.mutex);
>>   out_locked:
>> -       if (igt_flush_test(i915))
>> -               err = -EIO;
>>          while (reserved) {
>>                  struct reserved *next = reserved->next;
>>   
>> @@ -513,6 +511,9 @@ static int igt_evict_contexts(void *arg)
>>          mutex_unlock(&ggtt->vm.mutex);
>>          intel_runtime_pm_put(&i915->runtime_pm, wakeref);
>>   
>> +       if (igt_flush_test(i915))
>> +               err = -EIO;
> 
> The patch is ok, since the manual drm_mm_node reservations are not used
> by the GTT, but the reason is a bit specious.

We have a comment in i915_request_wait which says:

	/*
	 * We must never wait on the GPU while holding a lock as we
	 * may need to perform a GPU reset. So while we don't need to
	 * serialise wait/reset with an explicit lock, we do want
	 * lockdep to detect potential dependency cycles.
	 */

And then there was a lockdep splat here 
https://intel-gfx-ci.01.org/tree/drm-tip/Trybot_6595/fi-skl-6700k2/igt@i915_selftest@live@evict.html, 
which although uses some extra lockdep annotation patches, seemed to 
connect the two:

<4> [258.014638] Chain exists of:
   &gt->reset.mutex --> fs_reclaim --> &vm->mutex
<4> [258.014640]  Possible unsafe locking scenario:
<4> [258.014641]        CPU0                    CPU1
<4> [258.014641]        ----                    ----
<4> [258.014642]   lock(&vm->mutex);
<4> [258.014642]                                lock(fs_reclaim);
<4> [258.014643]                                lock(&vm->mutex);
<4> [258.014644]   lock(&gt->reset.mutex);
<4> [258.014645]
  *** DEADLOCK ***
<4> [258.014646] 2 locks held by i915_selftest/5153:

Why despite the comment in request wait it does not otherwise see this I 
don't know.

Regards,

Tvrtko