回复: [PATCH] drm/amdgpu: Increase tlb flush timeout for sriov

[Chen, Horace]

[AMD Official Use Only]

Hi Christian,

I think the problem is that a non-FLRed VF will not know that another VF got an FLR, unless host triggered a whole GPU reset.
So in the worst situation, for example the VF0 to VF10 are all hang and will be FLRed one by one, the VF11 will not know that there are FLRs happened, in VF11's prespective, it just see the fence didn't come back for about 5.5(0.5*11) seconds.

Thanks & Regards,
Horace.

________________________________
发件人: Koenig, Christian <Christian.Koenig at amd.com>
发送时间: 2021年5月19日 19:49
收件人: Liu, Cheng Zhe <ChengZhe.Liu at amd.com>; Christian König <ckoenig.leichtzumerken at gmail.com>; amd-gfx at lists.freedesktop.org <amd-gfx at lists.freedesktop.org>
抄送: Xiao, Jack <Jack.Xiao at amd.com>; Xu, Feifei <Feifei.Xu at amd.com>; Wang, Kevin(Yang) <Kevin1.Wang at amd.com>; Tuikov, Luben <Luben.Tuikov at amd.com>; Deucher, Alexander <Alexander.Deucher at amd.com>; Zhang, Hawking <Hawking.Zhang at amd.com>; Chen, Horace <Horace.Chen at amd.com>
主题: Re: [PATCH] drm/amdgpu: Increase tlb flush timeout for sriov

Yeah, but you can't do that it will probably trigger the watchdog timer.

The usec_timeout is named this way because it is a usec timeout.
Anything large than 1ms is a no-go here.

When the other instances do a FLR we don't really need to wait for the
TLB flush anyway since any FLR will kill that.

Christian.

Am 19.05.21 um 13:08 schrieb Liu, Cheng Zhe:
> [AMD Official Use Only]
>
> We support 12 VF at most. In worst case, the first 11 all IDLE fail and do FLR, it will need 11 * 500ms to switch to the 12nd VF,
> so I set 12 * 500ms  for the timeout.
>
> -----Original Message-----
> From: Christian König <ckoenig.leichtzumerken at gmail.com>
> Sent: Wednesday, May 19, 2021 6:08 PM
> To: Liu, Cheng Zhe <ChengZhe.Liu at amd.com>; amd-gfx at lists.freedesktop.org
> Cc: Xiao, Jack <Jack.Xiao at amd.com>; Xu, Feifei <Feifei.Xu at amd.com>; Wang, Kevin(Yang) <Kevin1.Wang at amd.com>; Tuikov, Luben <Luben.Tuikov at amd.com>; Deucher, Alexander <Alexander.Deucher at amd.com>; Koenig, Christian <Christian.Koenig at amd.com>; Zhang, Hawking <Hawking.Zhang at amd.com>
> Subject: Re: [PATCH] drm/amdgpu: Increase tlb flush timeout for sriov
>
> Am 19.05.21 um 11:32 schrieb Chengzhe Liu:
>> When there is 12 VF, we need to increase the timeout
> NAK, 6 seconds is way to long to wait polling on a fence.
>
> Why should an invalidation take that long? The engine are per VF just to avoid exactly that problem.
>
> Christian.
>
>> Signed-off-by: Chengzhe Liu <ChengZhe.Liu at amd.com>
>> ---
>>    drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c | 6 +++++-
>>    drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c  | 6 +++++-
>>    2 files changed, 10 insertions(+), 2 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c
>> b/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c
>> index f02dc904e4cf..a5f005c5d0ec 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c
>> @@ -404,6 +404,7 @@ static int gmc_v10_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev,
>>       uint32_t seq;
>>       uint16_t queried_pasid;
>>       bool ret;
>> +    uint32_t sriov_usec_timeout = 6000000;  /* wait for 12 * 500ms for
>> +SRIOV */
>>       struct amdgpu_ring *ring = &adev->gfx.kiq.ring;
>>       struct amdgpu_kiq *kiq = &adev->gfx.kiq;
>>
>> @@ -422,7 +423,10 @@ static int gmc_v10_0_flush_gpu_tlb_pasid(struct
>> amdgpu_device *adev,
>>
>>               amdgpu_ring_commit(ring);
>>               spin_unlock(&adev->gfx.kiq.ring_lock);
>> -            r = amdgpu_fence_wait_polling(ring, seq, adev->usec_timeout);
>> +            if (amdgpu_sriov_vf(adev))
>> +                    r = amdgpu_fence_wait_polling(ring, seq, sriov_usec_timeout);
>> +            else
>> +                    r = amdgpu_fence_wait_polling(ring, seq, adev->usec_timeout);
>>               if (r < 1) {
>>                       dev_err(adev->dev, "wait for kiq fence error: %ld.\n", r);
>>                       return -ETIME;
>> diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c
>> b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c
>> index ceb3968d8326..e4a18d8f75c2 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c
>> @@ -857,6 +857,7 @@ static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev,
>>       uint32_t seq;
>>       uint16_t queried_pasid;
>>       bool ret;
>> +    uint32_t sriov_usec_timeout = 6000000;  /* wait for 12 * 500ms for
>> +SRIOV */
>>       struct amdgpu_ring *ring = &adev->gfx.kiq.ring;
>>       struct amdgpu_kiq *kiq = &adev->gfx.kiq;
>>
>> @@ -896,7 +897,10 @@ static int gmc_v9_0_flush_gpu_tlb_pasid(struct
>> amdgpu_device *adev,
>>
>>               amdgpu_ring_commit(ring);
>>               spin_unlock(&adev->gfx.kiq.ring_lock);
>> -            r = amdgpu_fence_wait_polling(ring, seq, adev->usec_timeout);
>> +            if (amdgpu_sriov_vf(adev))
>> +                    r = amdgpu_fence_wait_polling(ring, seq, sriov_usec_timeout);
>> +            else
>> +                    r = amdgpu_fence_wait_polling(ring, seq, adev->usec_timeout);
>>               if (r < 1) {
>>                       dev_err(adev->dev, "wait for kiq fence error: %ld.\n", r);
>>                       up_read(&adev->reset_sem);

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/amd-gfx/attachments/20210519/13d3504c/attachment-0001.htm>