[RFC] drm/amdgpu: flush gpu tlb using sdma ring

[Christian König]

Am 09.06.23 um 15:32 schrieb Philip Yang:
>
> On 2023-06-09 06:02, Shashank Sharma wrote:
>> This RFC patch moves the code to flush TLB using SDMA ring to
>> a different place like a separate function. The idea is that
>> KFD/KGD code should be able to flush GPU TLB using SDMA if they
>> want to.
>>
>> Cc: Christian Koenig <christian.koenig at amd.com>
>> Signed-off-by: Shashank Sharma <shashank.sharma at amd.com>
>> ---
>>   drivers/gpu/drm/amd/amdgpu/amdgpu_sdma.c | 41 ++++++++++++++++++++++++
>>   drivers/gpu/drm/amd/amdgpu/amdgpu_sdma.h |  1 +
>>   drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c   | 36 +++++----------------
>>   3 files changed, 50 insertions(+), 28 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_sdma.c 
>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_sdma.c
>> index 231ca06bc9c7..20a5dad32bfc 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_sdma.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_sdma.c
>> @@ -30,6 +30,47 @@
>>   /* SDMA CSA reside in the 3rd page of CSA */
>>   #define AMDGPU_CSA_SDMA_OFFSET (4096 * 2)
>>   +/**
>> + * amdgpu_sdma_flush_tlb - flush gpu TLB using SDMA ring
>> + *
>> + * @adev: amdgpu device pointer
>> + *
>> + * return: returns dma fence which must be put by caller
>> + *
>> + * The SDMA on Navi has a bug which can theoretically result in memory
>> + * corruption if an invalidation happens at the same time as an VA
>> + * translation. Avoid this by doing the invalidation from the SDMA
>> + * itself.
>> + */
>> +struct dma_fence *amdgpu_sdma_flush_tlb(struct amdgpu_device *adev)
>> +{
>> +    int r;
>> +    struct dma_fence *fence;
>> +    struct amdgpu_job *job;
>> +    struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
>> +
>> +    mutex_lock(&adev->mman.gtt_window_lock);
>
> gtt_window_lock is for GART copy buffer/window sharing, here we use 
> copy ring but don't use copy buffer, and ring ib sa and job submit has 
> spinlock, so I think gtt_window_lock is not needed.

We need something to protect the entity while pushing jobs to it and 
that's the gtt_window_lock, so that can't be removed here.

Christian.

>
> gtt_window_lock should be removed from gmc_v10_0_flush_gpu_tlb too.
>
> Regards,
>
> Philip
>
>> +    r = amdgpu_job_alloc_with_ib(ring->adev, &adev->mman.entity,
>> +                     AMDGPU_FENCE_OWNER_UNDEFINED,
>> +                     16 * 4, AMDGPU_IB_POOL_IMMEDIATE,
>> +                     &job);
>> +    if (r)
>> +        goto error_alloc;
>> +
>> +    job->vm_pd_addr = amdgpu_gmc_pd_addr(adev->gart.bo);
>> +    job->vm_needs_flush = true;
>> +    job->ibs->ptr[job->ibs->length_dw++] = ring->funcs->nop;
>> +    amdgpu_ring_pad_ib(ring, &job->ibs[0]);
>> +    fence = amdgpu_job_submit(job);
>> +
>> +    mutex_unlock(&adev->mman.gtt_window_lock);
>> +    return fence;
>> +
>> +error_alloc:
>> +    mutex_unlock(&adev->mman.gtt_window_lock);
>> +    return NULL;
>> +}
>> +
>>   /*
>>    * GPU SDMA IP block helpers function.
>>    */
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_sdma.h 
>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_sdma.h
>> index fc8528812598..c895948f6e82 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_sdma.h
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_sdma.h
>> @@ -130,5 +130,6 @@ void amdgpu_sdma_destroy_inst_ctx(struct 
>> amdgpu_device *adev,
>>           bool duplicate);
>>   void amdgpu_sdma_unset_buffer_funcs_helper(struct amdgpu_device 
>> *adev);
>>   int amdgpu_sdma_ras_sw_init(struct amdgpu_device *adev);
>> +struct dma_fence *amdgpu_sdma_flush_tlb(struct amdgpu_device *adev);
>>     #endif
>> diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c 
>> b/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c
>> index ab2556ca984e..0bfaee00a838 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c
>> @@ -52,6 +52,7 @@
>>   #include "athub_v2_1.h"
>>     #include "amdgpu_reset.h"
>> +#include "amdgpu_sdma.h"
>>     #if 0
>>   static const struct soc15_reg_golden golden_settings_navi10_hdp[] =
>> @@ -333,9 +334,6 @@ static void gmc_v10_0_flush_gpu_tlb(struct 
>> amdgpu_device *adev, uint32_t vmid,
>>   {
>>       struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
>>       struct dma_fence *fence;
>> -    struct amdgpu_job *job;
>> -
>> -    int r;
>>         /* flush hdp cache */
>>       adev->hdp.funcs->flush_hdp(adev, NULL);
>> @@ -378,34 +376,16 @@ static void gmc_v10_0_flush_gpu_tlb(struct 
>> amdgpu_device *adev, uint32_t vmid,
>>           return;
>>       }
>>   -    /* The SDMA on Navi has a bug which can theoretically result 
>> in memory
>> -     * corruption if an invalidation happens at the same time as an VA
>> -     * translation. Avoid this by doing the invalidation from the SDMA
>> -     * itself.
>> -     */
>> -    r = amdgpu_job_alloc_with_ib(ring->adev, &adev->mman.entity,
>> -                     AMDGPU_FENCE_OWNER_UNDEFINED,
>> -                     16 * 4, AMDGPU_IB_POOL_IMMEDIATE,
>> -                     &job);
>> -    if (r)
>> -        goto error_alloc;
>> -
>> -    job->vm_pd_addr = amdgpu_gmc_pd_addr(adev->gart.bo);
>> -    job->vm_needs_flush = true;
>> -    job->ibs->ptr[job->ibs->length_dw++] = ring->funcs->nop;
>> -    amdgpu_ring_pad_ib(ring, &job->ibs[0]);
>> -    fence = amdgpu_job_submit(job);
>> -
>>       mutex_unlock(&adev->mman.gtt_window_lock);
>>   -    dma_fence_wait(fence, false);
>> -    dma_fence_put(fence);
>> -
>> -    return;
>> +    fence = amdgpu_sdma_flush_tlb(adev);
>> +    if (fence) {
>> +        dma_fence_wait(fence, false);
>> +        dma_fence_put(fence);
>> +        return;
>> +    }
>>   -error_alloc:
>> -    mutex_unlock(&adev->mman.gtt_window_lock);
>> -    DRM_ERROR("Error flushing GPU TLB using the SDMA (%d)!\n", r);
>> +    DRM_ERROR("Error flushing GPU TLB using the SDMA !\n");
>>   }
>>     /**