Skip to contentSkip to conversation

drm/amdgpu: fix amdgpu_fill_buffer

PATCH In review Christian König 2 participants 5 messages Nov 17, 2016 2 messages not in the archive

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/4 drm/amdgpu: fix amdgpu_fill_buffer

From: Christian König <[email protected]>

The mem start is relative to the domain in the address space, so this worked only when VRAM was mapped at offset 0.

It also didn't handled multiple drm_mm_nodes for split VRAM.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 70 ++++++++++++++++++++++-----------
 1 file changed, 48 insertions(+), 22 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +48 -22
index 5951666..c4dccb3 100644
@@ -1378,29 +1378,40 @@ error_free:
 	return r;
 }
 
-int amdgpu_fill_buffer(struct amdgpu_bo *bo,
-		uint32_t src_data,
-		struct reservation_object *resv,
-		struct fence **fence)
+int amdgpu_fill_buffer(struct amdgpu_bo *bo, uint32_t src_data,
+		       struct reservation_object *resv,
+		       struct fence **fence)
 {
 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
-	struct amdgpu_job *job;
+	uint32_t max_bytes = adev->mman.buffer_funcs->fill_max_bytes;
 	struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
 
-	uint32_t max_bytes, byte_count;
-	uint64_t dst_offset;
+	struct drm_mm_node *mm_node;
+	unsigned long num_pages;
 	unsigned int num_loops, num_dw;
-	unsigned int i;
+
+	struct amdgpu_job *job;
 	int r;
 
-	byte_count = bo->tbo.num_pages << PAGE_SHIFT;
-	max_bytes = adev->mman.buffer_funcs->fill_max_bytes;
-	num_loops = DIV_ROUND_UP(byte_count, max_bytes);
+	if (!ring->ready) {
+		DRM_ERROR("Trying to clear memory with ring turned off.\n");
+		return -EINVAL;
+	}
+
+	num_pages = bo->tbo.num_pages;
+	mm_node = bo->tbo.mem.mm_node;
+	num_loops = 0;
+	while (num_pages) {
+		uint32_t byte_count = mm_node->size << PAGE_SHIFT;
+
+		num_loops += DIV_ROUND_UP(byte_count, max_bytes);
+		num_pages -= mm_node->size;
+		++mm_node;
+	}
 	num_dw = num_loops * adev->mman.buffer_funcs->fill_num_dw;
 
 	/* for IB padding */
-	while (num_dw & 0x7)
-		num_dw++;
+	num_dw += 64;
 
 	r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, &job);
 	if (r)
@@ -1408,28 +1419,43 @@ int amdgpu_fill_buffer(struct amdgpu_bo *bo,
 
 	if (resv) {
 		r = amdgpu_sync_resv(adev, &job->sync, resv,
-				AMDGPU_FENCE_OWNER_UNDEFINED);
+				     AMDGPU_FENCE_OWNER_UNDEFINED);
 		if (r) {
 			DRM_ERROR("sync failed (%d).\n", r);
 			goto error_free;
 		}
 	}
 
-	dst_offset = bo->tbo.mem.start << PAGE_SHIFT;
-	for (i = 0; i < num_loops; i++) {
-		uint32_t cur_size_in_bytes = min(byte_count, max_bytes);
+	num_pages = bo->tbo.num_pages;
+	mm_node = bo->tbo.mem.mm_node;
 
-		amdgpu_emit_fill_buffer(adev, &job->ibs[0], src_data,
-				dst_offset, cur_size_in_bytes);
+	while (num_pages) {
+		uint32_t byte_count = mm_node->size << PAGE_SHIFT;
+		uint64_t dst_addr;
 
-		dst_offset += cur_size_in_bytes;
-		byte_count -= cur_size_in_bytes;
+		r = amdgpu_mm_node_addr(&bo->tbo, mm_node,
+					&bo->tbo.mem, &dst_addr);
+		if (r)
+			return r;
+
+		while (byte_count) {
+			uint32_t cur_size_in_bytes = min(byte_count, max_bytes);
+
+			amdgpu_emit_fill_buffer(adev, &job->ibs[0], src_data,
+						dst_addr, cur_size_in_bytes);
+
+			dst_addr += cur_size_in_bytes;
+			byte_count -= cur_size_in_bytes;
+		}
+
+		num_pages -= mm_node->size;
+		++mm_node;
 	}
 
 	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
 	WARN_ON(job->ibs[0].length_dw > num_dw);
 	r = amdgpu_job_submit(job, ring, &adev->mman.entity,
-			AMDGPU_FENCE_OWNER_UNDEFINED, fence);
+			      AMDGPU_FENCE_OWNER_UNDEFINED, fence);
 	if (r)
 		goto error_free;
 
Signature · 6 lines
2.5.0

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 1479394810-1887-1-git-send-email-deathsimple-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

#2 lore raw

PATCH 2/4 drm/amdgpu: fix error handling in amdgpu_bo_create_restricted

From: Christian König <[email protected]>

Manually checking why a function could fail is not a good idea if you can just check the functions return code.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_object.c | 11 ++++-------
 1 file changed, 4 insertions(+), 7 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_object.c +4 -7
index 052c1b0..2ac8d43 100644
@@ -382,12 +382,6 @@ int amdgpu_bo_create_restricted(struct amdgpu_device *adev,
 	    bo->tbo.mem.placement & TTM_PL_FLAG_VRAM) {
 		struct fence *fence;
 
-		if (adev->mman.buffer_funcs_ring == NULL ||
-		   !adev->mman.buffer_funcs_ring->ready) {
-			r = -EBUSY;
-			goto fail_free;
-		}
-
 		r = amdgpu_bo_reserve(bo, false);
 		if (unlikely(r != 0))
 			goto fail_free;
@@ -397,7 +391,10 @@ int amdgpu_bo_create_restricted(struct amdgpu_device *adev,
 		if (unlikely(r != 0))
 			goto fail_unreserve;
 
-		amdgpu_fill_buffer(bo, 0, bo->tbo.resv, &fence);
+		r = amdgpu_fill_buffer(bo, 0, bo->tbo.resv, &fence);
+		if (unlikely(r))
+			goto fail_unreserve;
+
 		amdgpu_bo_fence(bo, fence, false);
 		amdgpu_bo_unreserve(bo);
 		fence_put(bo->tbo.moving);
Signature · 6 lines
2.5.0

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
#3 lore raw

PATCH 3/4 drm/amdgpu: improve AMDGPU_GEM_CREATE_VRAM_CLEARED handling

In reply to a message not in the archive

From: Christian König <[email protected]>

Drop this whole reserve/unreserve dance, just lock the reservation object manually when creating the BO.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_object.c | 29 ++++++++++++++---------------
 1 file changed, 14 insertions(+), 15 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_object.c +14 -15
index 2ac8d43..3cbe598 100644
@@ -371,36 +371,36 @@ int amdgpu_bo_create_restricted(struct amdgpu_device *adev,
 
 	amdgpu_fill_placement_to_bo(bo, placement);
 	/* Kernel allocation are uninterruptible */
+
+	if (!resv) {
+		bool locked;
+
+		reservation_object_init(&bo->tbo.ttm_resv);
+		locked = ww_mutex_trylock(&bo->tbo.ttm_resv.lock);
+		WARN_ON(!locked);
+	}
 	r = ttm_bo_init(&adev->mman.bdev, &bo->tbo, size, type,
 			&bo->placement, page_align, !kernel, NULL,
-			acc_size, sg, resv, &amdgpu_ttm_bo_destroy);
-	if (unlikely(r != 0)) {
+			acc_size, sg, resv ? resv : &bo->tbo.ttm_resv,
+			&amdgpu_ttm_bo_destroy);
+	if (unlikely(r != 0))
 		return r;
-	}
 
 	if (flags & AMDGPU_GEM_CREATE_VRAM_CLEARED &&
 	    bo->tbo.mem.placement & TTM_PL_FLAG_VRAM) {
 		struct fence *fence;
 
-		r = amdgpu_bo_reserve(bo, false);
-		if (unlikely(r != 0))
-			goto fail_free;
-
-		amdgpu_ttm_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
-		r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
-		if (unlikely(r != 0))
-			goto fail_unreserve;
-
 		r = amdgpu_fill_buffer(bo, 0, bo->tbo.resv, &fence);
 		if (unlikely(r))
 			goto fail_unreserve;
 
 		amdgpu_bo_fence(bo, fence, false);
-		amdgpu_bo_unreserve(bo);
 		fence_put(bo->tbo.moving);
 		bo->tbo.moving = fence_get(fence);
 		fence_put(fence);
 	}
+	if (!resv)
+		ww_mutex_unlock(&bo->tbo.resv->lock);
 	*bo_ptr = bo;
 
 	trace_amdgpu_bo_create(bo);
@@ -408,8 +408,7 @@ int amdgpu_bo_create_restricted(struct amdgpu_device *adev,
 	return 0;
 
 fail_unreserve:
-	amdgpu_bo_unreserve(bo);
-fail_free:
+	ww_mutex_unlock(&bo->tbo.resv->lock);
 	amdgpu_bo_unref(&bo);
 	return r;
 }
Signature · 6 lines
2.5.0

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
#4 lore raw

PATCH 4/4 drm/amdgpu: use AMDGPU_GEM_CREATE_VRAM_CLEARED for VM PD/PTs

In reply to a message not in the archive

From: Christian König <[email protected]>

Doesn't make much sense to have the same functionality twice.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 99 ++--------------------------------
 1 file changed, 4 insertions(+), 95 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +4 -95
index d6c2839..d05546e 100644
@@ -530,70 +530,6 @@ static void amdgpu_vm_do_copy_ptes(struct amdgpu_pte_update_params *params,
 }
 
 /**
- * amdgpu_vm_clear_bo - initially clear the page dir/table
- *
- * @adev: amdgpu_device pointer
- * @bo: bo to clear
- *
- * need to reserve bo first before calling it.
- */
-static int amdgpu_vm_clear_bo(struct amdgpu_device *adev,
-			      struct amdgpu_vm *vm,
-			      struct amdgpu_bo *bo)
-{
-	struct amdgpu_ring *ring;
-	struct fence *fence = NULL;
-	struct amdgpu_job *job;
-	struct amdgpu_pte_update_params params;
-	unsigned entries;
-	uint64_t addr;
-	int r;
-
-	ring = container_of(vm->entity.sched, struct amdgpu_ring, sched);
-
-	r = reservation_object_reserve_shared(bo->tbo.resv);
-	if (r)
-		return r;
-
-	r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
-	if (r)
-		goto error;
-
-	r = amdgpu_ttm_bind(&bo->tbo, &bo->tbo.mem);
-	if (r)
-		goto error;
-
-	addr = amdgpu_bo_gpu_offset(bo);
-	entries = amdgpu_bo_size(bo) / 8;
-
-	r = amdgpu_job_alloc_with_ib(adev, 64, &job);
-	if (r)
-		goto error;
-
-	memset(&params, 0, sizeof(params));
-	params.adev = adev;
-	params.ib = &job->ibs[0];
-	amdgpu_vm_do_set_ptes(&params, addr, 0, entries, 0, 0);
-	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
-
-	WARN_ON(job->ibs[0].length_dw > 64);
-	r = amdgpu_job_submit(job, ring, &vm->entity,
-			      AMDGPU_FENCE_OWNER_VM, &fence);
-	if (r)
-		goto error_free;
-
-	amdgpu_bo_fence(bo, fence, true);
-	fence_put(fence);
-	return 0;
-
-error_free:
-	amdgpu_job_free(job);
-
-error:
-	return r;
-}
-
-/**
  * amdgpu_vm_map_gart - Resolve gart mapping of addr
  *
  * @pages_addr: optional DMA address to use for lookup
@@ -1435,7 +1371,8 @@ int amdgpu_vm_bo_map(struct amdgpu_device *adev,
 				     AMDGPU_GEM_DOMAIN_VRAM,
 				     AMDGPU_GEM_CREATE_NO_CPU_ACCESS |
 				     AMDGPU_GEM_CREATE_SHADOW |
-				     AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS,
+				     AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS |
+				     AMDGPU_GEM_CREATE_VRAM_CLEARED,
 				     NULL, resv, &pt);
 		if (r)
 			goto error_free;
@@ -1445,22 +1382,6 @@ int amdgpu_vm_bo_map(struct amdgpu_device *adev,
 		 */
 		pt->parent = amdgpu_bo_ref(vm->page_directory);
 
-		r = amdgpu_vm_clear_bo(adev, vm, pt);
-		if (r) {
-			amdgpu_bo_unref(&pt->shadow);
-			amdgpu_bo_unref(&pt);
-			goto error_free;
-		}
-
-		if (pt->shadow) {
-			r = amdgpu_vm_clear_bo(adev, vm, pt->shadow);
-			if (r) {
-				amdgpu_bo_unref(&pt->shadow);
-				amdgpu_bo_unref(&pt);
-				goto error_free;
-			}
-		}
-
 		vm->page_tables[pt_idx].bo = pt;
 		vm->page_tables[pt_idx].addr = 0;
 	}
@@ -1642,7 +1563,8 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm)
 			     AMDGPU_GEM_DOMAIN_VRAM,
 			     AMDGPU_GEM_CREATE_NO_CPU_ACCESS |
 			     AMDGPU_GEM_CREATE_SHADOW |
-			     AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS,
+			     AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS |
+			     AMDGPU_GEM_CREATE_VRAM_CLEARED,
 			     NULL, NULL, &vm->page_directory);
 	if (r)
 		goto error_free_sched_entity;
@@ -1651,24 +1573,11 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm)
 	if (r)
 		goto error_free_page_directory;
 
-	r = amdgpu_vm_clear_bo(adev, vm, vm->page_directory);
-	if (r)
-		goto error_unreserve;
-
-	if (vm->page_directory->shadow) {
-		r = amdgpu_vm_clear_bo(adev, vm, vm->page_directory->shadow);
-		if (r)
-			goto error_unreserve;
-	}
-
 	vm->last_eviction_counter = atomic64_read(&adev->num_evictions);
 	amdgpu_bo_unreserve(vm->page_directory);
 
 	return 0;
 
-error_unreserve:
-	amdgpu_bo_unreserve(vm->page_directory);
-
 error_free_page_directory:
 	amdgpu_bo_unref(&vm->page_directory->shadow);
 	amdgpu_bo_unref(&vm->page_directory);
Signature · 6 lines
2.5.0

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 1479394810-1887-4-git-send-email-deathsimple-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

Reviewed-by Alex Deucher
#6 lore raw

RE: [PATCH 4/4] drm/amdgpu: use AMDGPU_GEM_CREATE_VRAM_CLEARED for VM PD/PTs

Quoted 8 lines

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Christian König
Sent: Thursday, November 17, 2016 10:00 AM
To: [email protected]
Subject: [PATCH 4/4] drm/amdgpu: use AMDGPU_GEM_CREATE_VRAM_CLEARED for VM PD/PTs

Quoted 5 lines

From: Christian König <[email protected]>

Doesn't make much sense to have the same functionality twice.

Series is:

Quoted 163 lines
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 99 ++---------------------------
-----
 1 file changed, 4 insertions(+), 95 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -64
index d6c2839..d05546e 100644
@@ -530,70 +530,6 @@ static void amdgpu_vm_do_copy_ptes(struct
 amdgpu_pte_update_params *params,
 }
 
 /**
- * amdgpu_vm_clear_bo - initially clear the page dir/table
- *
- * @adev: amdgpu_device pointer
- * @bo: bo to clear
- *
- * need to reserve bo first before calling it.
- */
-static int amdgpu_vm_clear_bo(struct amdgpu_device *adev,
-			      struct amdgpu_vm *vm,
-			      struct amdgpu_bo *bo)
-{
-	struct amdgpu_ring *ring;
-	struct fence *fence = NULL;
-	struct amdgpu_job *job;
-	struct amdgpu_pte_update_params params;
-	unsigned entries;
-	uint64_t addr;
-	int r;
-
-	ring = container_of(vm->entity.sched, struct amdgpu_ring, sched);
-
-	r = reservation_object_reserve_shared(bo->tbo.resv);
-	if (r)
-		return r;
-
-	r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
-	if (r)
-		goto error;
-
-	r = amdgpu_ttm_bind(&bo->tbo, &bo->tbo.mem);
-	if (r)
-		goto error;
-
-	addr = amdgpu_bo_gpu_offset(bo);
-	entries = amdgpu_bo_size(bo) / 8;
-
-	r = amdgpu_job_alloc_with_ib(adev, 64, &job);
-	if (r)
-		goto error;
-
-	memset(&params, 0, sizeof(params));
-	params.adev = adev;
-	params.ib = &job->ibs[0];
-	amdgpu_vm_do_set_ptes(&params, addr, 0, entries, 0, 0);
-	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
-
-	WARN_ON(job->ibs[0].length_dw > 64);
-	r = amdgpu_job_submit(job, ring, &vm->entity,
-			      AMDGPU_FENCE_OWNER_VM, &fence);
-	if (r)
-		goto error_free;
-
-	amdgpu_bo_fence(bo, fence, true);
-	fence_put(fence);
-	return 0;
-
-error_free:
-	amdgpu_job_free(job);
-
-error:
-	return r;
-}
-
-/**
  * amdgpu_vm_map_gart - Resolve gart mapping of addr
  *

* @pages_addr: optional DMA address to use for lookup

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -0
@@ -1435,7 +1371,8 @@ int amdgpu_vm_bo_map(struct amdgpu_device
 *adev,
 				     AMDGPU_GEM_DOMAIN_VRAM,
 

AMDGPU_GEM_CREATE_NO_CPU_ACCESS |

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -1
 				     AMDGPU_GEM_CREATE_SHADOW |
-

AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS,
+
AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS |
+
AMDGPU_GEM_CREATE_VRAM_CLEARED,

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -16
 				     NULL, resv, &pt);
 		if (r)
 			goto error_free;
@@ -1445,22 +1382,6 @@ int amdgpu_vm_bo_map(struct amdgpu_device
 *adev,
 		 */
 		pt->parent = amdgpu_bo_ref(vm->page_directory);
 
-		r = amdgpu_vm_clear_bo(adev, vm, pt);
-		if (r) {
-			amdgpu_bo_unref(&pt->shadow);
-			amdgpu_bo_unref(&pt);
-			goto error_free;
-		}
-
-		if (pt->shadow) {
-			r = amdgpu_vm_clear_bo(adev, vm, pt->shadow);
-			if (r) {
-				amdgpu_bo_unref(&pt->shadow);
-				amdgpu_bo_unref(&pt);
-				goto error_free;
-			}
-		}
-
 		vm->page_tables[pt_idx].bo = pt;
 		vm->page_tables[pt_idx].addr = 0;
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +2 -1
 	}
@@ -1642,7 +1563,8 @@ int amdgpu_vm_init(struct amdgpu_device *adev,
 struct amdgpu_vm *vm)
 			     AMDGPU_GEM_DOMAIN_VRAM,
 			     AMDGPU_GEM_CREATE_NO_CPU_ACCESS |
 			     AMDGPU_GEM_CREATE_SHADOW |
-			     AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS,
+			     AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS |
+			     AMDGPU_GEM_CREATE_VRAM_CLEARED,
 			     NULL, NULL, &vm->page_directory);
 	if (r)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -13
 		goto error_free_sched_entity;
@@ -1651,24 +1573,11 @@ int amdgpu_vm_init(struct amdgpu_device
 *adev, struct amdgpu_vm *vm)
 	if (r)
 		goto error_free_page_directory;
 
-	r = amdgpu_vm_clear_bo(adev, vm, vm->page_directory);
-	if (r)
-		goto error_unreserve;
-
-	if (vm->page_directory->shadow) {
-		r = amdgpu_vm_clear_bo(adev, vm, vm->page_directory-
 >shadow);
-		if (r)
-			goto error_unreserve;
-	}
-
 	vm->last_eviction_counter = atomic64_read(&adev-
 >num_evictions);
 	amdgpu_bo_unreserve(vm->page_directory);
 
 	return 0;
 
-error_unreserve:
-	amdgpu_bo_unreserve(vm->page_directory);
-

error_free_page_directory:

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -1
 	amdgpu_bo_unref(&vm->page_directory->shadow);
 	amdgpu_bo_unref(&vm->page_directory);
--

2.5.0

Quoted 4 lines

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx