Skip to contentSkip to conversation

drm/amdgpu: Remove rounding from vram allocation path

PATCH In review Daniel Phillips 2 participants 3 messages Jul 29, 2022

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/2 drm/amdgpu: Remove rounding from vram allocation path

Rounding up allocations in the allocation path caused test regressions, so now just round in the availability path.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 11 +++++------
 1 file changed, 5 insertions(+), 6 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +5 -6
index 043a808c88a3..014a594899fb 100644
@@ -40,10 +40,10 @@
 #define AMDGPU_USERPTR_RESTORE_DELAY_MS 1
 
 /*
- * Align VRAM allocations to 2MB to avoid fragmentation caused by 4K allocations in the tail 2MB
+ * Align VRAM availability to 2MB to avoid fragmentation caused by 4K allocations in the tail 2MB
  * BO chunk
  */
-#define VRAM_ALLOCATION_ALIGN (1 << 21)
+#define VRAM_AVAILABLITY_ALIGN (1 << 21)
 
 /* Impose limit on how much memory KFD can use */
 static struct {
@@ -149,7 +149,7 @@ static int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
 		 * to avoid fragmentation caused by 4K allocations in the tail
 		 * 2M BO chunk.
 		 */
-		vram_needed = ALIGN(size, VRAM_ALLOCATION_ALIGN);
+		vram_needed = size;
 	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
 		system_mem_needed = size;
 	} else if (!(alloc_flag &
@@ -198,7 +198,7 @@ static void unreserve_mem_limit(struct amdgpu_device *adev,
 		kfd_mem_limit.system_mem_used -= size;
 		kfd_mem_limit.ttm_mem_used -= size;
 	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
-		adev->kfd.vram_used -= ALIGN(size, VRAM_ALLOCATION_ALIGN);
+		adev->kfd.vram_used -= size;
 	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
 		kfd_mem_limit.system_mem_used -= size;
 	} else if (!(alloc_flag &
@@ -1642,7 +1642,6 @@ size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev)
 	uint64_t reserved_for_pt =
 		ESTIMATE_PT_SIZE(amdgpu_amdkfd_total_mem_size);
 	size_t available;
-
 	spin_lock(&kfd_mem_limit.mem_limit_lock);
 	available = adev->gmc.real_vram_size
 		- adev->kfd.vram_used
@@ -1650,7 +1649,7 @@ size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev)
 		- reserved_for_pt;
 	spin_unlock(&kfd_mem_limit.mem_limit_lock);
 
-	return ALIGN_DOWN(available, VRAM_ALLOCATION_ALIGN);
+	return ALIGN_DOWN(available, VRAM_AVAILABLITY_ALIGN);
 }
 
 int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(
Signature · 1 line
2.34.1
#1 lore raw

PATCH 2/2 drm/amdgpu: Pessimistic availability based on rounded up allocations

Seperately accumulate a statistic of rounded up allocations to use to report availability, with a view to increasing the likelihood a buffer object can be successfully allocated at exactly the size reported by the availability API.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h       | 1 +
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 4 +++-
 2 files changed, 4 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +1 -0
index 73bf8b5f2aa9..781274be5f27 100644
@@ -96,6 +96,7 @@ struct amdgpu_amdkfd_fence {
 struct amdgpu_kfd_dev {
 	struct kfd_dev *dev;
 	uint64_t vram_used;
+	uint64_t vram_used_aligned;
 	bool init_complete;
 	struct work_struct reset_work;
 };
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +3 -1
index 014a594899fb..e5ea897f56bc 100644
@@ -181,6 +181,7 @@ static int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
 	 * memory, TTM memory and GPU memory as computed above
 	 */
 	adev->kfd.vram_used += vram_needed;
+	adev->kfd.vram_used_aligned += ALIGN(vram_needed, VRAM_AVAILABLITY_ALIGN);
 	kfd_mem_limit.system_mem_used += system_mem_needed;
 	kfd_mem_limit.ttm_mem_used += ttm_mem_needed;
 
@@ -199,6 +200,7 @@ static void unreserve_mem_limit(struct amdgpu_device *adev,
 		kfd_mem_limit.ttm_mem_used -= size;
 	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
 		adev->kfd.vram_used -= size;
+		adev->kfd.vram_used_aligned -= ALIGN(size, VRAM_AVAILABLITY_ALIGN);
 	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
 		kfd_mem_limit.system_mem_used -= size;
 	} else if (!(alloc_flag &
@@ -1644,7 +1646,7 @@ size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev)
 	size_t available;
 	spin_lock(&kfd_mem_limit.mem_limit_lock);
 	available = adev->gmc.real_vram_size
-		- adev->kfd.vram_used
+		- adev->kfd.vram_used_aligned
 		- atomic64_read(&adev->vram_pin_size)
 		- reserved_for_pt;
 	spin_unlock(&kfd_mem_limit.mem_limit_lock);
Signature · 1 line
2.34.1
maintainer Reviewed-by on v1 2/2
#2 lore raw

Re: [PATCH 2/2] drm/amdgpu: Pessimistic availability based on rounded up allocations

Your patches are missing Signed-off-by lines. If you use "git commit -s", git should add those automatically for your convenience. Other than that, the patches look good to me. With Signed-off-by added, the series is

Am 2022-07-28 um 23:16 schrieb Daniel Phillips:

Quoted 50 lines

Seperately accumulate a statistic of rounded up allocations to use to report availability, with a view to increasing the likelihood a buffer object can be successfully allocated at exactly the size reported by the availability API.

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h       | 1 +
  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 4 +++-
  2 files changed, 4 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h +1 -0
index 73bf8b5f2aa9..781274be5f27 100644
@@ -96,6 +96,7 @@ struct amdgpu_amdkfd_fence {
  struct amdgpu_kfd_dev {
  	struct kfd_dev *dev;
  	uint64_t vram_used;
+	uint64_t vram_used_aligned;
  	bool init_complete;
  	struct work_struct reset_work;
  };
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +3 -1
index 014a594899fb..e5ea897f56bc 100644
@@ -181,6 +181,7 @@ static int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev,
  	 * memory, TTM memory and GPU memory as computed above
  	 */
  	adev->kfd.vram_used += vram_needed;
+	adev->kfd.vram_used_aligned += ALIGN(vram_needed, VRAM_AVAILABLITY_ALIGN);
  	kfd_mem_limit.system_mem_used += system_mem_needed;
  	kfd_mem_limit.ttm_mem_used += ttm_mem_needed;
 
@@ -199,6 +200,7 @@ static void unreserve_mem_limit(struct amdgpu_device *adev,
  		kfd_mem_limit.ttm_mem_used -= size;
  	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {
  		adev->kfd.vram_used -= size;
+		adev->kfd.vram_used_aligned -= ALIGN(size, VRAM_AVAILABLITY_ALIGN);
  	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_USERPTR) {
  		kfd_mem_limit.system_mem_used -= size;
  	} else if (!(alloc_flag &
@@ -1644,7 +1646,7 @@ size_t amdgpu_amdkfd_get_available_memory(struct amdgpu_device *adev)
  	size_t available;
  	spin_lock(&kfd_mem_limit.mem_limit_lock);
  	available = adev->gmc.real_vram_size
-		- adev->kfd.vram_used
+		- adev->kfd.vram_used_aligned
  		- atomic64_read(&adev->vram_pin_size)
  		- reserved_for_pt;
  	spin_unlock(&kfd_mem_limit.mem_limit_lock);