No new issues found with those patches, testing GPU reset using libdrm
deadlock detection test on Ellsmire.
The patches are Tested-By: Andrey Grodzovsky <andrey.grodzovsky at amd.com>
P.S
Noticed existing issues (before Monk's patches)
Multiple [drm:amdgpu_cs_ioctl [amdgpu]] *ERROR* Failed to initialize
parser !
And occasional unlock imbalance forn amdgpu_cs_ioctl
DEBUG_LOCKS_WARN_ON(depth <= 0)
[Â Â 93.069011 <Â Â Â 0.000017>] WARNING: CPU: 3 PID: 2215 at
kernel/locking/lockdep.c:3682 lock_release+0x2e8/0x360
On CZ full reset hangs the system.
Gonna take a look at those issues.
Thanks,
Andrey
From
On 02/28/2018 08:31 AM, Liu, Monk wrote:
> Already sent
>
> -----Original Message-----
> From: Grodzovsky, Andrey
> Sent: 2018å¹´2æ??28æ?¥ 21:31
> To: Koenig, Christian <Christian.Koenig at amd.com>; Liu, Monk <Monk.Liu at amd.com>; amd-gfx at lists.freedesktop.org
> Subject: Re: [PATCH 1/4] drm/amdgpu: stop all rings before doing gpu recover
>
> Will do once Monk sends V2 for [PATCH 4/4] drm/amdgpu: try again kiq access if not in IRQ
>
> Andrey
>
>
> On 02/28/2018 07:20 AM, Christian König wrote:
>> Andrey please give this set a good testing as well.
>>
>> Am 28.02.2018 um 08:21 schrieb Monk Liu:
>>> found recover_vram_from_shadow sometimes get executed in paralle with
>>> SDMA scheduler, should stop all schedulers before doing gpu
>>> reset/recover
>>>
>>> Change-Id: Ibaef3e3c015f3cf88f84b2eaf95cda95ae1a64e3
>>> Signed-off-by: Monk Liu <Monk.Liu at amd.com>
>> For now this patch is Reviewed-by: Christian König
>> <christian.koenig at amd.com>.
>>
>> Regards,
>> Christian.
>>
>>> ---
>>> Â drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 40
>>> +++++++++++-------------------
>>> Â 1 file changed, 15 insertions(+), 25 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>>> index 75d1733..e9d81a8 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>>> @@ -2649,22 +2649,23 @@ int amdgpu_device_gpu_recover(struct
>>> amdgpu_device *adev,
>>> Â Â Â Â Â Â /* block TTM */
>>> Â Â Â Â Â resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
>>> +
>>> Â Â Â Â Â /* store modesetting */
>>> Â Â Â Â Â if (amdgpu_device_has_dc_support(adev))
>>> Â Â Â Â Â Â Â Â Â state = drm_atomic_helper_suspend(adev->ddev);
>>> Â -Â Â Â /* block scheduler */
>>> +Â Â Â /* block all schedulers and reset given job's ring */
>>> Â Â Â Â Â for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
>>> Â Â Â Â Â Â Â Â Â struct amdgpu_ring *ring = adev->rings[i];
>>> Â Â Â Â Â Â Â Â Â Â if (!ring || !ring->sched.thread)
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â continue;
>>> Â -Â Â Â Â Â Â Â /* only focus on the ring hit timeout if &job not NULL */
>>> +Â Â Â Â Â Â Â kthread_park(ring->sched.thread);
>>> +
>>> Â Â Â Â Â Â Â Â Â if (job && job->ring->idx != i)
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â continue;
>>> Â -Â Â Â Â Â Â Â kthread_park(ring->sched.thread);
>>> Â Â Â Â Â Â Â Â Â drm_sched_hw_job_reset(&ring->sched, &job->base);
>>> Â Â Â Â Â Â Â Â Â Â /* after all hw jobs are reset, hw fence is meaningless,
>>> so force_completion */ @@ -2707,33 +2708,22 @@ int
>>> amdgpu_device_gpu_recover(struct amdgpu_device *adev,
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â }
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â dma_fence_put(fence);
>>> Â Â Â Â Â Â Â Â Â }
>>> +Â Â Â }
>>> Â -Â Â Â Â Â Â Â for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
>>> -Â Â Â Â Â Â Â Â Â Â Â struct amdgpu_ring *ring = adev->rings[i];
>>> -
>>> -Â Â Â Â Â Â Â Â Â Â Â if (!ring || !ring->sched.thread)
>>> -Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â continue;
>>> +Â Â Â for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
>>> +Â Â Â Â Â Â Â struct amdgpu_ring *ring = adev->rings[i];
>>> Â -Â Â Â Â Â Â Â Â Â Â Â /* only focus on the ring hit timeout if &job not NULL
>>> */
>>> -Â Â Â Â Â Â Â Â Â Â Â if (job && job->ring->idx != i)
>>> -Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â continue;
>>> +Â Â Â Â Â Â Â if (!ring || !ring->sched.thread)
>>> +Â Â Â Â Â Â Â Â Â Â Â continue;
>>> Â +Â Â Â Â Â Â Â /* only need recovery sched of the given job's ring
>>> +Â Â Â Â Â Â Â Â * or all rings (in the case @job is NULL)
>>> +Â Â Â Â Â Â Â Â * after above amdgpu_reset accomplished
>>> +Â Â Â Â Â Â Â Â */
>>> +Â Â Â Â Â Â Â if ((!job || job->ring->idx == i) && !r)
>>> Â Â Â Â Â Â Â Â Â Â Â Â Â drm_sched_job_recovery(&ring->sched);
>>> -Â Â Â Â Â Â Â Â Â Â Â kthread_unpark(ring->sched.thread);
>>> -Â Â Â Â Â Â Â }
>>> -Â Â Â } else {
>>> -Â Â Â Â Â Â Â for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
>>> -Â Â Â Â Â Â Â Â Â Â Â struct amdgpu_ring *ring = adev->rings[i];
>>> Â -Â Â Â Â Â Â Â Â Â Â Â if (!ring || !ring->sched.thread)
>>> -Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â continue;
>>> -
>>> -Â Â Â Â Â Â Â Â Â Â Â /* only focus on the ring hit timeout if &job not NULL
>>> */
>>> -Â Â Â Â Â Â Â Â Â Â Â if (job && job->ring->idx != i)
>>> -Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â continue;
>>> -
>>> -Â Â Â Â Â Â Â Â Â Â Â kthread_unpark(adev->rings[i]->sched.thread);
>>> -Â Â Â Â Â Â Â }
>>> +Â Â Â Â Â Â Â kthread_unpark(ring->sched.thread);
>>> Â Â Â Â Â }
>>> Â Â Â Â Â Â if (amdgpu_device_has_dc_support(adev)) {
