[PATCH v2 4/4] drm/amdkfd: implement counters for vm fault and migration

[Felix Kuehling]

On 2021-06-23 11:02 a.m., Philip Yang wrote:
> Add helper function to get process device data structure from adev to
> update counters.
>
> Update vm faults, page_in, page_out counters will no be executed in
> parallel, use WRITE_ONCE to avoid any form of compiler optimizations.
>
> Signed-off-by: Philip Yang <Philip.Yang at amd.com>

Reviewed-by: Felix Kuehling <Felix.Kuehling at amd.com>


> ---
>   drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 14 ++++++++
>   drivers/gpu/drm/amd/amdkfd/kfd_svm.c     | 45 +++++++++++++++++++++++-
>   drivers/gpu/drm/amd/amdkfd/kfd_svm.h     |  2 ++
>   3 files changed, 60 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
> index fd8f544f0de2..45b5349283af 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
> @@ -413,6 +413,7 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
>   			uint64_t end)
>   {
>   	uint64_t npages = (end - start) >> PAGE_SHIFT;
> +	struct kfd_process_device *pdd;
>   	struct dma_fence *mfence = NULL;
>   	struct migrate_vma migrate;
>   	dma_addr_t *scratch;
> @@ -473,6 +474,12 @@ svm_migrate_vma_to_vram(struct amdgpu_device *adev, struct svm_range *prange,
>   out_free:
>   	kvfree(buf);
>   out:
> +	if (!r) {
> +		pdd = svm_range_get_pdd_by_adev(prange, adev);
> +		if (pdd)
> +			WRITE_ONCE(pdd->page_in, pdd->page_in + migrate.cpages);
> +	}
> +
>   	return r;
>   }
>   
> @@ -629,6 +636,7 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
>   		       struct vm_area_struct *vma, uint64_t start, uint64_t end)
>   {
>   	uint64_t npages = (end - start) >> PAGE_SHIFT;
> +	struct kfd_process_device *pdd;
>   	struct dma_fence *mfence = NULL;
>   	struct migrate_vma migrate;
>   	dma_addr_t *scratch;
> @@ -678,6 +686,12 @@ svm_migrate_vma_to_ram(struct amdgpu_device *adev, struct svm_range *prange,
>   out_free:
>   	kvfree(buf);
>   out:
> +	if (!r) {
> +		pdd = svm_range_get_pdd_by_adev(prange, adev);
> +		if (pdd)
> +			WRITE_ONCE(pdd->page_out,
> +				   pdd->page_out + migrate.cpages);
> +	}
>   	return r;
>   }
>   
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
> index 7f5ebc00b349..00b46846a5e0 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
> @@ -564,6 +564,24 @@ svm_range_get_adev_by_id(struct svm_range *prange, uint32_t gpu_id)
>   	return (struct amdgpu_device *)pdd->dev->kgd;
>   }
>   
> +struct kfd_process_device *
> +svm_range_get_pdd_by_adev(struct svm_range *prange, struct amdgpu_device *adev)
> +{
> +	struct kfd_process *p;
> +	int32_t gpu_idx, gpuid;
> +	int r;
> +
> +	p = container_of(prange->svms, struct kfd_process, svms);
> +
> +	r = kfd_process_gpuid_from_kgd(p, adev, &gpuid, &gpu_idx);
> +	if (r) {
> +		pr_debug("failed to get device id by adev %p\n", adev);
> +		return NULL;
> +	}
> +
> +	return kfd_process_device_from_gpuidx(p, gpu_idx);
> +}
> +
>   static int svm_range_bo_validate(void *param, struct amdgpu_bo *bo)
>   {
>   	struct ttm_operation_ctx ctx = { false, false };
> @@ -2311,6 +2329,27 @@ static bool svm_range_skip_recover(struct svm_range *prange)
>   	return false;
>   }
>   
> +static void
> +svm_range_count_fault(struct amdgpu_device *adev, struct kfd_process *p,
> +		      struct svm_range *prange, int32_t gpuidx)
> +{
> +	struct kfd_process_device *pdd;
> +
> +	if (gpuidx == MAX_GPU_INSTANCE)
> +		/* fault is on different page of same range
> +		 * or fault is skipped to recover later
> +		 */
> +		pdd = svm_range_get_pdd_by_adev(prange, adev);
> +	else
> +		/* fault recovered
> +		 * or fault cannot recover because GPU no access on the range
> +		 */
> +		pdd = kfd_process_device_from_gpuidx(p, gpuidx);
> +
> +	if (pdd)
> +		WRITE_ONCE(pdd->faults, pdd->faults + 1);
> +}
> +
>   int
>   svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
>   			uint64_t addr)
> @@ -2320,7 +2359,8 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
>   	struct svm_range *prange;
>   	struct kfd_process *p;
>   	uint64_t timestamp;
> -	int32_t best_loc, gpuidx;
> +	int32_t best_loc;
> +	int32_t gpuidx = MAX_GPU_INSTANCE;
>   	bool write_locked = false;
>   	int r = 0;
>   
> @@ -2440,6 +2480,9 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
>   out_unlock_svms:
>   	mutex_unlock(&svms->lock);
>   	mmap_read_unlock(mm);
> +
> +	svm_range_count_fault(adev, p, prange, gpuidx);
> +
>   	mmput(mm);
>   out:
>   	kfd_unref_process(p);
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.h b/drivers/gpu/drm/amd/amdkfd/kfd_svm.h
> index 0c0fc399395e..a9af03994d1a 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.h
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.h
> @@ -174,6 +174,8 @@ void svm_range_dma_unmap(struct device *dev, dma_addr_t *dma_addr,
>   			 unsigned long offset, unsigned long npages);
>   void svm_range_free_dma_mappings(struct svm_range *prange);
>   void svm_range_prefault(struct svm_range *prange, struct mm_struct *mm);
> +struct kfd_process_device *
> +svm_range_get_pdd_by_adev(struct svm_range *prange, struct amdgpu_device *adev);
>   
>   /* SVM API and HMM page migration work together, device memory type
>    * is initialized to not 0 when page migration register device memory.