Am 2021-06-22 um 9:32 a.m. schrieb Philip Yang:

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@amd.com>
---
 drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 14 ++++++++++++++
 drivers/gpu/drm/amd/amdkfd/kfd_svm.c     | 24 ++++++++++++++++++++++++
 drivers/gpu/drm/amd/amdkfd/kfd_svm.h     |  2 ++
 3 files changed, 40 insertions(+)

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 5468ea4264c6..f3323328f01f 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 };
@@ -2315,6 +2333,7 @@ int
 svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
 			uint64_t addr)
 {
+	struct kfd_process_device *pdd;
 	struct mm_struct *mm = NULL;
 	struct svm_range_list *svms;
 	struct svm_range *prange;
@@ -2440,6 +2459,11 @@ svm_range_restore_pages(struct amdgpu_device *adev, unsigned int pasid,
 out_unlock_svms:
 	mutex_unlock(&svms->lock);
 	mmap_read_unlock(mm);
+
+	pdd = svm_range_get_pdd_by_adev(prange, adev);

svm_range_get_pdd_by_adev needs to do a linear search. You don't need
this here because you already know the gpuidx. I think you can just call
kfd_process_device_from_gpuidx(p, gpu_idx) here.

1. fault on different page of same range

2. fault is skipped to recover later

3. fault cannot recover because GPU no access

4. fault recover failed, may try again later

5. fault is recovered

Philip

With that fixed, the series is

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


P.S.: Thanks for catching and fixing those memory leaks in patch 2.

+	if (pdd)
+		WRITE_ONCE(pdd->faults, pdd->faults + 1);
+
 	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.