Commit 1ac354be authored by Daniel Phillips's avatar Daniel Phillips Committed by Alex Deucher
Browse files

drm/amdgpu: Pessimistic availability based on rounded up allocations 



Separately 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.

Signed-off-by: default avatarDaniel Phillips <daniel.phillips@amd.com>
Reviewed-by: default avatarFelix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: default avatarAlex Deucher <alexander.deucher@amd.com>
parent aec208ee

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h

+1 −0
Original line number Diff line number Diff line
@@ -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

+7 −3
Original line number Diff line number Diff line
@@ -182,8 +182,10 @@ int amdgpu_amdkfd_reserve_mem_limit(struct amdgpu_device *adev, 
	 */

	WARN_ONCE(vram_needed && !adev,

		  "adev reference can't be null when vram is used");

	if (adev)

	if (adev) {

		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;
@@ -203,8 +205,10 @@ void amdgpu_amdkfd_unreserve_mem_limit(struct amdgpu_device *adev, 
	} else if (alloc_flag & KFD_IOC_ALLOC_MEM_FLAGS_VRAM) {

		WARN_ONCE(!adev,

			  "adev reference can't be null when alloc mem flags vram is set");

		if (adev)

		if (adev) {

			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 &
@@ -1610,7 +1614,7 @@ 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);