Skip to contentSkip to conversation

drm/amdgpu: move ring helpers to amdgpu_ring.h

PATCH In review Christian König 5 participants 27 messages Jun 30, 2017 – Jul 3, 2017 12 messages not in the archive

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

PATCH 1/12 drm/amdgpu: move ring helpers to amdgpu_ring.h

From: Christian König <[email protected]>

Keep them where they belong.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu.h      | 44 --------------------------------
 drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h | 42 ++++++++++++++++++++++++++++++
 2 files changed, 42 insertions(+), 44 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +0 -44
index ab1dad2..810796a 100644
@@ -1801,50 +1801,6 @@ bool amdgpu_device_has_dc_support(struct amdgpu_device *adev);
 #define RBIOS16(i) (RBIOS8(i) | (RBIOS8((i)+1) << 8))
 #define RBIOS32(i) ((RBIOS16(i)) | (RBIOS16((i)+2) << 16))
 
-/*
- * RING helpers.
- */
-static inline void amdgpu_ring_write(struct amdgpu_ring *ring, uint32_t v)
-{
-	if (ring->count_dw <= 0)
-		DRM_ERROR("amdgpu: writing more dwords to the ring than expected!\n");
-	ring->ring[ring->wptr++ & ring->buf_mask] = v;
-	ring->wptr &= ring->ptr_mask;
-	ring->count_dw--;
-}
-
-static inline void amdgpu_ring_write_multiple(struct amdgpu_ring *ring, void *src, int count_dw)
-{
-	unsigned occupied, chunk1, chunk2;
-	void *dst;
-
-	if (unlikely(ring->count_dw < count_dw)) {
-		DRM_ERROR("amdgpu: writing more dwords to the ring than expected!\n");
-		return;
-	}
-
-	occupied = ring->wptr & ring->buf_mask;
-	dst = (void *)&ring->ring[occupied];
-	chunk1 = ring->buf_mask + 1 - occupied;
-	chunk1 = (chunk1 >= count_dw) ? count_dw: chunk1;
-	chunk2 = count_dw - chunk1;
-	chunk1 <<= 2;
-	chunk2 <<= 2;
-
-	if (chunk1)
-		memcpy(dst, src, chunk1);
-
-	if (chunk2) {
-		src += chunk1;
-		dst = (void *)ring->ring;
-		memcpy(dst, src, chunk2);
-	}
-
-	ring->wptr += count_dw;
-	ring->wptr &= ring->ptr_mask;
-	ring->count_dw -= count_dw;
-}
-
 static inline struct amdgpu_sdma_instance *
 amdgpu_get_sdma_instance(struct amdgpu_ring *ring)
 {
drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h +42 -0

46 lines not on this page · raw message

Signature · 6 lines
2.7.4

_______________________________________________
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): 1498821733-1545-1-git-send-email-deathsimple-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

#2 lore raw

PATCH 2/12 drm/amdgpu: fix amdgpu_ring_write_multiple

From: Christian König <[email protected]>

Overwriting still used ring content has a low probability to cause problems, not writing at all has 100% probability to cause problems.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h +1 -3
index 04cbc3a..322d2529 100644
@@ -227,10 +227,8 @@ static inline void amdgpu_ring_write_multiple(struct amdgpu_ring *ring,
 	unsigned occupied, chunk1, chunk2;
 	void *dst;
 
-	if (unlikely(ring->count_dw < count_dw)) {
+	if (unlikely(ring->count_dw < count_dw))
 		DRM_ERROR("amdgpu: writing more dwords to the ring than expected!\n");
-		return;
-	}
 
 	occupied = ring->wptr & ring->buf_mask;
 	dst = (void *)&ring->ring[occupied];
Signature · 6 lines
2.7.4

_______________________________________________
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): 1498821733-1545-2-git-send-email-deathsimple-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

Reviewed-by
#4 lore raw

Re: [PATCH 02/12] drm/amdgpu: fix amdgpu_ring_write_multiple

Quoted 6 lines · Christian König

From: Christian König <[email protected]>

Overwriting still used ring content has a low probability to cause problems, not writing at all has 100% probability to cause problems.

Quoted 23 lines
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h +1 -4
index 04cbc3a..322d2529 100644
@@ -227,10 +227,8 @@ static inline void amdgpu_ring_write_multiple(struct amdgpu_ring *ring,
        unsigned occupied, chunk1, chunk2;
        void *dst;
 
-       if (unlikely(ring->count_dw < count_dw)) {
+       if (unlikely(ring->count_dw < count_dw))
                DRM_ERROR("amdgpu: writing more dwords to the ring than expected!\n");
-               return;
-       }
 
        occupied = ring->wptr & ring->buf_mask;
        dst = (void *)&ring->ring[occupied];
--

2.7.4

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

#5 lore raw

PATCH 3/12 drm/amdgpu: allow flushing VMID0 before IB execution as well

In reply to a message not in the archive

From: Christian König <[email protected]>

This allows us to queue IBs which needs an up to date system domain as well.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c  | 2 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 2 ++
 2 files changed, 3 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c +1 -1
index f774b3f..1b30d2a 100644
@@ -172,7 +172,7 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
 	if (ring->funcs->insert_start)
 		ring->funcs->insert_start(ring);
 
-	if (vm) {
+	if (job) {
 		r = amdgpu_vm_flush(ring, job);
 		if (r) {
 			amdgpu_ring_undo(ring);
drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +2 -0
index 3d641e1..4510627 100644
@@ -81,6 +81,8 @@ int amdgpu_job_alloc_with_ib(struct amdgpu_device *adev, unsigned size,
 	r = amdgpu_ib_get(adev, NULL, size, &(*job)->ibs[0]);
 	if (r)
 		kfree(*job);
+	else
+		(*job)->vm_pd_addr = adev->gart.table_addr;
 
 	return r;
 }
Signature · 6 lines
2.7.4

_______________________________________________
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): 1498821733-1545-3-git-send-email-deathsimple-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

Reviewed-by
#7 lore raw

Re: [PATCH 03/12] drm/amdgpu: allow flushing VMID0 before IB execution as well

Quoted 5 lines · Christian König

From: Christian König <[email protected]>

This allows us to queue IBs which needs an up to date system domain as well.

Quoted 34 lines
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c  | 2 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 2 ++
 2 files changed, 3 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c +1 -1
index f774b3f..1b30d2a 100644
@@ -172,7 +172,7 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
        if (ring->funcs->insert_start)
                ring->funcs->insert_start(ring);
 
-       if (vm) {
+       if (job) {
                r = amdgpu_vm_flush(ring, job);
                if (r) {
                        amdgpu_ring_undo(ring);
drivers/gpu/drm/amd/amdgpu/amdgpu_job.c +2 -1
index 3d641e1..4510627 100644
@@ -81,6 +81,8 @@ int amdgpu_job_alloc_with_ib(struct amdgpu_device *adev, unsigned size,
        r = amdgpu_ib_get(adev, NULL, size, &(*job)->ibs[0]);
        if (r)
                kfree(*job);
+       else
+               (*job)->vm_pd_addr = adev->gart.table_addr;
 
        return r;
 }
--

2.7.4

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

#8 lore raw

PATCH 4/12 drm/amdgpu: add vm_needs_flush parameter to amdgpu_copy_buffer

In reply to a message not in the archive

From: Christian König <[email protected]>

This allows us to flush the system VM here.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_benchmark.c |  2 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_object.c    |  4 ++--
 drivers/gpu/drm/amd/amdgpu/amdgpu_test.c      |  4 ++--
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c       | 12 ++++++------
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h       |  9 ++++-----
 5 files changed, 15 insertions(+), 16 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_benchmark.c +1 -1
index 1beae5b..2fb299a 100644
@@ -40,7 +40,7 @@ static int amdgpu_benchmark_do_move(struct amdgpu_device *adev, unsigned size,
 	for (i = 0; i < n; i++) {
 		struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
 		r = amdgpu_copy_buffer(ring, saddr, daddr, size, NULL, &fence,
-				       false);
+				       false, false);
 		if (r)
 			goto exit_do_move;
 		r = dma_fence_wait(fence, false);
drivers/gpu/drm/amd/amdgpu/amdgpu_object.c +2 -2
index 8ee6965..c34cf2c 100644
@@ -535,7 +535,7 @@ int amdgpu_bo_backup_to_shadow(struct amdgpu_device *adev,
 
 	r = amdgpu_copy_buffer(ring, bo_addr, shadow_addr,
 			       amdgpu_bo_size(bo), resv, fence,
-			       direct);
+			       direct, false);
 	if (!r)
 		amdgpu_bo_fence(bo, *fence, true);
 
@@ -588,7 +588,7 @@ int amdgpu_bo_restore_from_shadow(struct amdgpu_device *adev,
 
 	r = amdgpu_copy_buffer(ring, shadow_addr, bo_addr,
 			       amdgpu_bo_size(bo), resv, fence,
-			       direct);
+			       direct, false);
 	if (!r)
 		amdgpu_bo_fence(bo, *fence, true);
 
drivers/gpu/drm/amd/amdgpu/amdgpu_test.c +2 -2
index 15510da..d02e611 100644
@@ -111,7 +111,7 @@ static void amdgpu_do_test_moves(struct amdgpu_device *adev)
 		amdgpu_bo_kunmap(gtt_obj[i]);
 
 		r = amdgpu_copy_buffer(ring, gtt_addr, vram_addr,
-				       size, NULL, &fence, false);
+				       size, NULL, &fence, false, false);
 
 		if (r) {
 			DRM_ERROR("Failed GTT->VRAM copy %d\n", i);
@@ -156,7 +156,7 @@ static void amdgpu_do_test_moves(struct amdgpu_device *adev)
 		amdgpu_bo_kunmap(vram_obj);
 
 		r = amdgpu_copy_buffer(ring, vram_addr, gtt_addr,
-				       size, NULL, &fence, false);
+				       size, NULL, &fence, false, false);
 
 		if (r) {
 			DRM_ERROR("Failed VRAM->GTT copy %d\n", i);
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +6 -6
index e4860ac..bbe1639 100644
@@ -318,7 +318,7 @@ static int amdgpu_move_blit(struct ttm_buffer_object *bo,
 
 		r = amdgpu_copy_buffer(ring, old_start, new_start,
 				       cur_pages * PAGE_SIZE,
-				       bo->resv, &next, false);
+				       bo->resv, &next, false, false);
 		if (r)
 			goto error;
 
@@ -1256,12 +1256,11 @@ int amdgpu_mmap(struct file *filp, struct vm_area_struct *vma)
 	return ttm_bo_mmap(filp, vma, &adev->mman.bdev);
 }
 
-int amdgpu_copy_buffer(struct amdgpu_ring *ring,
-		       uint64_t src_offset,
-		       uint64_t dst_offset,
-		       uint32_t byte_count,
+int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
+		       uint64_t dst_offset, uint32_t byte_count,
 		       struct reservation_object *resv,
-		       struct dma_fence **fence, bool direct_submit)
+		       struct dma_fence **fence, bool direct_submit,
+		       bool vm_needs_flush)
 {
 	struct amdgpu_device *adev = ring->adev;
 	struct amdgpu_job *job;
@@ -1283,6 +1282,7 @@ int amdgpu_copy_buffer(struct amdgpu_ring *ring,
 	if (r)
 		return r;
 
+	job->vm_needs_flush = vm_needs_flush;
 	if (resv) {
 		r = amdgpu_sync_resv(adev, &job->sync, resv,
 				     AMDGPU_FENCE_OWNER_UNDEFINED);
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h +4 -5
index 6bdede8..cd5bbfa 100644
@@ -61,12 +61,11 @@ int amdgpu_gtt_mgr_alloc(struct ttm_mem_type_manager *man,
 			 const struct ttm_place *place,
 			 struct ttm_mem_reg *mem);
 
-int amdgpu_copy_buffer(struct amdgpu_ring *ring,
-		       uint64_t src_offset,
-		       uint64_t dst_offset,
-		       uint32_t byte_count,
+int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
+		       uint64_t dst_offset, uint32_t byte_count,
 		       struct reservation_object *resv,
-		       struct dma_fence **fence, bool direct_submit);
+		       struct dma_fence **fence, bool direct_submit,
+		       bool vm_needs_flush);
 int amdgpu_fill_buffer(struct amdgpu_bo *bo,
 			uint32_t src_data,
 			struct reservation_object *resv,
Signature · 6 lines
2.7.4

_______________________________________________
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): 1498821733-1545-4-git-send-email-deathsimple-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

Reviewed-by
#10 lore raw

Re: [PATCH 04/12] drm/amdgpu: add vm_needs_flush parameter to amdgpu_copy_buffer

Quoted 5 lines · Christian König

From: Christian König <[email protected]>

This allows us to flush the system VM here.

Quoted 127 lines
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_benchmark.c |  2 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_object.c    |  4 ++--
 drivers/gpu/drm/amd/amdgpu/amdgpu_test.c      |  4 ++--
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c       | 12 ++++++------
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h       |  9 ++++-----
 5 files changed, 15 insertions(+), 16 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_benchmark.c +1 -1
index 1beae5b..2fb299a 100644
@@ -40,7 +40,7 @@ static int amdgpu_benchmark_do_move(struct amdgpu_device *adev, unsigned size,
        for (i = 0; i < n; i++) {
                struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
                r = amdgpu_copy_buffer(ring, saddr, daddr, size, NULL, &fence,
-                                      false);
+                                      false, false);
                if (r)
                        goto exit_do_move;
                r = dma_fence_wait(fence, false);
drivers/gpu/drm/amd/amdgpu/amdgpu_object.c +2 -2
index 8ee6965..c34cf2c 100644
@@ -535,7 +535,7 @@ int amdgpu_bo_backup_to_shadow(struct amdgpu_device *adev,
 
        r = amdgpu_copy_buffer(ring, bo_addr, shadow_addr,
                               amdgpu_bo_size(bo), resv, fence,
-                              direct);
+                              direct, false);
        if (!r)
                amdgpu_bo_fence(bo, *fence, true);
 
@@ -588,7 +588,7 @@ int amdgpu_bo_restore_from_shadow(struct amdgpu_device *adev,
 
        r = amdgpu_copy_buffer(ring, shadow_addr, bo_addr,
                               amdgpu_bo_size(bo), resv, fence,
-                              direct);
+                              direct, false);
        if (!r)
                amdgpu_bo_fence(bo, *fence, true);
 
drivers/gpu/drm/amd/amdgpu/amdgpu_test.c +2 -2
index 15510da..d02e611 100644
@@ -111,7 +111,7 @@ static void amdgpu_do_test_moves(struct amdgpu_device *adev)
                amdgpu_bo_kunmap(gtt_obj[i]);
 
                r = amdgpu_copy_buffer(ring, gtt_addr, vram_addr,
-                                      size, NULL, &fence, false);
+                                      size, NULL, &fence, false, false);
 
                if (r) {
                        DRM_ERROR("Failed GTT->VRAM copy %d\n", i);
@@ -156,7 +156,7 @@ static void amdgpu_do_test_moves(struct amdgpu_device *adev)
                amdgpu_bo_kunmap(vram_obj);
 
                r = amdgpu_copy_buffer(ring, vram_addr, gtt_addr,
-                                      size, NULL, &fence, false);
+                                      size, NULL, &fence, false, false);
 
                if (r) {
                        DRM_ERROR("Failed VRAM->GTT copy %d\n", i);
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +6 -6
index e4860ac..bbe1639 100644
@@ -318,7 +318,7 @@ static int amdgpu_move_blit(struct ttm_buffer_object *bo,
 
                r = amdgpu_copy_buffer(ring, old_start, new_start,
                                       cur_pages * PAGE_SIZE,
-                                      bo->resv, &next, false);
+                                      bo->resv, &next, false, false);
                if (r)
                        goto error;
 
@@ -1256,12 +1256,11 @@ int amdgpu_mmap(struct file *filp, struct vm_area_struct *vma)
        return ttm_bo_mmap(filp, vma, &adev->mman.bdev);
 }
 
-int amdgpu_copy_buffer(struct amdgpu_ring *ring,
-                      uint64_t src_offset,
-                      uint64_t dst_offset,
-                      uint32_t byte_count,
+int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
+                      uint64_t dst_offset, uint32_t byte_count,
                       struct reservation_object *resv,
-                      struct dma_fence **fence, bool direct_submit)
+                      struct dma_fence **fence, bool direct_submit,
+                      bool vm_needs_flush)
 {
        struct amdgpu_device *adev = ring->adev;
        struct amdgpu_job *job;
@@ -1283,6 +1282,7 @@ int amdgpu_copy_buffer(struct amdgpu_ring *ring,
        if (r)
                return r;
 
+       job->vm_needs_flush = vm_needs_flush;
        if (resv) {
                r = amdgpu_sync_resv(adev, &job->sync, resv,
                                     AMDGPU_FENCE_OWNER_UNDEFINED);
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h +4 -6
index 6bdede8..cd5bbfa 100644
@@ -61,12 +61,11 @@ int amdgpu_gtt_mgr_alloc(struct ttm_mem_type_manager *man,
                         const struct ttm_place *place,
                         struct ttm_mem_reg *mem);
 
-int amdgpu_copy_buffer(struct amdgpu_ring *ring,
-                      uint64_t src_offset,
-                      uint64_t dst_offset,
-                      uint32_t byte_count,
+int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
+                      uint64_t dst_offset, uint32_t byte_count,
                       struct reservation_object *resv,
-                      struct dma_fence **fence, bool direct_submit);
+                      struct dma_fence **fence, bool direct_submit,
+                      bool vm_needs_flush);
 int amdgpu_fill_buffer(struct amdgpu_bo *bo,
                        uint32_t src_data,
                        struct reservation_object *resv,
--

2.7.4

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

#11 lore raw

PATCH 5/12 drm/amdgpu: bind BOs to TTM only once

In reply to a message not in the archive

From: Christian König <[email protected]>

No need to do this on every round.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 70 ++++++++++++++-------------------
 1 file changed, 29 insertions(+), 41 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +29 -41
index bbe1639..5bfe7f6 100644
@@ -252,29 +252,15 @@ static void amdgpu_move_null(struct ttm_buffer_object *bo,
 	new_mem->mm_node = NULL;
 }
 
-static int amdgpu_mm_node_addr(struct ttm_buffer_object *bo,
-			       struct drm_mm_node *mm_node,
-			       struct ttm_mem_reg *mem,
-			       uint64_t *addr)
+static uint64_t amdgpu_mm_node_addr(struct ttm_buffer_object *bo,
+				    struct drm_mm_node *mm_node,
+				    struct ttm_mem_reg *mem)
 {
-	int r;
-
-	switch (mem->mem_type) {
-	case TTM_PL_TT:
-		r = amdgpu_ttm_bind(bo, mem);
-		if (r)
-			return r;
-
-	case TTM_PL_VRAM:
-		*addr = mm_node->start << PAGE_SHIFT;
-		*addr += bo->bdev->man[mem->mem_type].gpu_offset;
-		break;
-	default:
-		DRM_ERROR("Unknown placement %d\n", mem->mem_type);
-		return -EINVAL;
-	}
+	uint64_t addr;
 
-	return 0;
+	addr = mm_node->start << PAGE_SHIFT;
+	addr += bo->bdev->man[mem->mem_type].gpu_offset;
+	return addr;
 }
 
 static int amdgpu_move_blit(struct ttm_buffer_object *bo,
@@ -298,18 +284,25 @@ static int amdgpu_move_blit(struct ttm_buffer_object *bo,
 		return -EINVAL;
 	}
 
+	if (old_mem->mem_type == TTM_PL_TT) {
+		r = amdgpu_ttm_bind(bo, old_mem);
+		if (r)
+			return r;
+	}
+
 	old_mm = old_mem->mm_node;
-	r = amdgpu_mm_node_addr(bo, old_mm, old_mem, &old_start);
-	if (r)
-		return r;
 	old_size = old_mm->size;
+	old_start = amdgpu_mm_node_addr(bo, old_mm, old_mem);
 
+	if (new_mem->mem_type == TTM_PL_TT) {
+		r = amdgpu_ttm_bind(bo, new_mem);
+		if (r)
+			return r;
+	}
 
 	new_mm = new_mem->mm_node;
-	r = amdgpu_mm_node_addr(bo, new_mm, new_mem, &new_start);
-	if (r)
-		return r;
 	new_size = new_mm->size;
+	new_start = amdgpu_mm_node_addr(bo, new_mm, new_mem);
 
 	num_pages = new_mem->num_pages;
 	while (num_pages) {
@@ -331,10 +324,7 @@ static int amdgpu_move_blit(struct ttm_buffer_object *bo,
 
 		old_size -= cur_pages;
 		if (!old_size) {
-			r = amdgpu_mm_node_addr(bo, ++old_mm, old_mem,
-						&old_start);
-			if (r)
-				goto error;
+			old_start = amdgpu_mm_node_addr(bo, ++old_mm, old_mem);
 			old_size = old_mm->size;
 		} else {
 			old_start += cur_pages * PAGE_SIZE;
@@ -342,11 +332,7 @@ static int amdgpu_move_blit(struct ttm_buffer_object *bo,
 
 		new_size -= cur_pages;
 		if (!new_size) {
-			r = amdgpu_mm_node_addr(bo, ++new_mm, new_mem,
-						&new_start);
-			if (r)
-				goto error;
-
+			new_start = amdgpu_mm_node_addr(bo, ++new_mm, new_mem);
 			new_size = new_mm->size;
 		} else {
 			new_start += cur_pages * PAGE_SIZE;
@@ -1347,6 +1333,12 @@ int amdgpu_fill_buffer(struct amdgpu_bo *bo,
 		return -EINVAL;
 	}
 
+	if (bo->tbo.mem.mem_type == TTM_PL_TT) {
+		r = amdgpu_ttm_bind(&bo->tbo, &bo->tbo.mem);
+		if (r)
+			return r;
+	}
+
 	num_pages = bo->tbo.num_pages;
 	mm_node = bo->tbo.mem.mm_node;
 	num_loops = 0;
@@ -1382,11 +1374,7 @@ int amdgpu_fill_buffer(struct amdgpu_bo *bo,
 		uint32_t byte_count = mm_node->size << PAGE_SHIFT;
 		uint64_t dst_addr;
 
-		r = amdgpu_mm_node_addr(&bo->tbo, mm_node,
-					&bo->tbo.mem, &dst_addr);
-		if (r)
-			return r;
-
+		dst_addr = amdgpu_mm_node_addr(&bo->tbo, mm_node, &bo->tbo.mem);
 		while (byte_count) {
 			uint32_t cur_size_in_bytes = min(byte_count, max_bytes);
 
Signature · 6 lines
2.7.4

_______________________________________________
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): 1498821733-1545-5-git-send-email-deathsimple-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

Reviewed-by
#13 lore raw

Re: [PATCH 05/12] drm/amdgpu: bind BOs to TTM only once

Quoted 5 lines · Christian König

From: Christian König <[email protected]>

No need to do this on every round.

Quoted 131 lines
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 70 ++++++++++++++-------------------
 1 file changed, 29 insertions(+), 41 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +29 -42
index bbe1639..5bfe7f6 100644
@@ -252,29 +252,15 @@ static void amdgpu_move_null(struct ttm_buffer_object *bo,
        new_mem->mm_node = NULL;
 }
 
-static int amdgpu_mm_node_addr(struct ttm_buffer_object *bo,
-                              struct drm_mm_node *mm_node,
-                              struct ttm_mem_reg *mem,
-                              uint64_t *addr)
+static uint64_t amdgpu_mm_node_addr(struct ttm_buffer_object *bo,
+                                   struct drm_mm_node *mm_node,
+                                   struct ttm_mem_reg *mem)
 {
-       int r;
-
-       switch (mem->mem_type) {
-       case TTM_PL_TT:
-               r = amdgpu_ttm_bind(bo, mem);
-               if (r)
-                       return r;
-
-       case TTM_PL_VRAM:
-               *addr = mm_node->start << PAGE_SHIFT;
-               *addr += bo->bdev->man[mem->mem_type].gpu_offset;
-               break;
-       default:
-               DRM_ERROR("Unknown placement %d\n", mem->mem_type);
-               return -EINVAL;
-       }
+       uint64_t addr;
 
-       return 0;
+       addr = mm_node->start << PAGE_SHIFT;
+       addr += bo->bdev->man[mem->mem_type].gpu_offset;
+       return addr;
 }
 
 static int amdgpu_move_blit(struct ttm_buffer_object *bo,
@@ -298,18 +284,25 @@ static int amdgpu_move_blit(struct ttm_buffer_object *bo,
                return -EINVAL;
        }
 
+       if (old_mem->mem_type == TTM_PL_TT) {
+               r = amdgpu_ttm_bind(bo, old_mem);
+               if (r)
+                       return r;
+       }
+
        old_mm = old_mem->mm_node;
-       r = amdgpu_mm_node_addr(bo, old_mm, old_mem, &old_start);
-       if (r)
-               return r;
        old_size = old_mm->size;
+       old_start = amdgpu_mm_node_addr(bo, old_mm, old_mem);
 
+       if (new_mem->mem_type == TTM_PL_TT) {
+               r = amdgpu_ttm_bind(bo, new_mem);
+               if (r)
+                       return r;
+       }
 
        new_mm = new_mem->mm_node;
-       r = amdgpu_mm_node_addr(bo, new_mm, new_mem, &new_start);
-       if (r)
-               return r;
        new_size = new_mm->size;
+       new_start = amdgpu_mm_node_addr(bo, new_mm, new_mem);
 
        num_pages = new_mem->num_pages;
        while (num_pages) {
@@ -331,10 +324,7 @@ static int amdgpu_move_blit(struct ttm_buffer_object *bo,
 
                old_size -= cur_pages;
                if (!old_size) {
-                       r = amdgpu_mm_node_addr(bo, ++old_mm, old_mem,
-                                               &old_start);
-                       if (r)
-                               goto error;
+                       old_start = amdgpu_mm_node_addr(bo, ++old_mm, old_mem);
                        old_size = old_mm->size;
                } else {
                        old_start += cur_pages * PAGE_SIZE;
@@ -342,11 +332,7 @@ static int amdgpu_move_blit(struct ttm_buffer_object *bo,
 
                new_size -= cur_pages;
                if (!new_size) {
-                       r = amdgpu_mm_node_addr(bo, ++new_mm, new_mem,
-                                               &new_start);
-                       if (r)
-                               goto error;
-
+                       new_start = amdgpu_mm_node_addr(bo, ++new_mm, new_mem);
                        new_size = new_mm->size;
                } else {
                        new_start += cur_pages * PAGE_SIZE;
@@ -1347,6 +1333,12 @@ int amdgpu_fill_buffer(struct amdgpu_bo *bo,
                return -EINVAL;
        }
 
+       if (bo->tbo.mem.mem_type == TTM_PL_TT) {
+               r = amdgpu_ttm_bind(&bo->tbo, &bo->tbo.mem);
+               if (r)
+                       return r;
+       }
+
        num_pages = bo->tbo.num_pages;
        mm_node = bo->tbo.mem.mm_node;
        num_loops = 0;
@@ -1382,11 +1374,7 @@ int amdgpu_fill_buffer(struct amdgpu_bo *bo,
                uint32_t byte_count = mm_node->size << PAGE_SHIFT;
                uint64_t dst_addr;
 
-               r = amdgpu_mm_node_addr(&bo->tbo, mm_node,
-                                       &bo->tbo.mem, &dst_addr);
-               if (r)
-                       return r;
-
+               dst_addr = amdgpu_mm_node_addr(&bo->tbo, mm_node, &bo->tbo.mem);
                while (byte_count) {
                        uint32_t cur_size_in_bytes = min(byte_count, max_bytes);
 
--

2.7.4

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

#14 lore raw

PATCH 6/12 drm/amdgpu: bind BOs with GTT space allocated directly

In reply to a message not in the archive

From: Christian König <[email protected]>

This avoids binding them later on.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_gtt_mgr.c | 16 +++++++++-
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c     | 49 ++++++++++++++++++-----------
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h     |  1 +
 3 files changed, 46 insertions(+), 20 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gtt_mgr.c +15 -1
index f7d22c4..6fdf83a 100644
@@ -81,6 +81,20 @@ static int amdgpu_gtt_mgr_fini(struct ttm_mem_type_manager *man)
 }
 
 /**
+ * amdgpu_gtt_mgr_is_allocated - Check if mem has address space
+ *
+ * @mem: the mem object to check
+ *
+ * Check if a mem object has already address space allocated.
+ */
+bool amdgpu_gtt_mgr_is_alloced(struct ttm_mem_reg *mem)
+{
+	struct drm_mm_node *node = mem->mm_node;
+
+	return (node->start != AMDGPU_BO_INVALID_OFFSET);
+}
+
+/**
  * amdgpu_gtt_mgr_alloc - allocate new ranges
  *
  * @man: TTM memory type manager
@@ -101,7 +115,7 @@ int amdgpu_gtt_mgr_alloc(struct ttm_mem_type_manager *man,
 	unsigned long fpfn, lpfn;
 	int r;
 
-	if (node->start != AMDGPU_BO_INVALID_OFFSET)
+	if (amdgpu_gtt_mgr_is_alloced(mem))
 		return 0;
 
 	if (place)
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +30 -19

73 lines not on this page · raw message

drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h +1 -0
index cd5bbfa..2ade5c5 100644
@@ -56,6 +56,7 @@ struct amdgpu_mman {
 extern const struct ttm_mem_type_manager_func amdgpu_gtt_mgr_func;
 extern const struct ttm_mem_type_manager_func amdgpu_vram_mgr_func;
 
+bool amdgpu_gtt_mgr_is_alloced(struct ttm_mem_reg *mem);
 int amdgpu_gtt_mgr_alloc(struct ttm_mem_type_manager *man,
 			 struct ttm_buffer_object *tbo,
 			 const struct ttm_place *place,
Signature · 6 lines
2.7.4

_______________________________________________
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): 1498821733-1545-6-git-send-email-deathsimple-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

Reviewed-by
#16 lore raw

Re: [PATCH 06/12] drm/amdgpu: bind BOs with GTT space allocated directly

Quoted 16 lines · Christian König

From: Christian König <[email protected]>

This avoids binding them later on.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_gtt_mgr.c | 16 +++++++++-
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c     | 49 ++++++++++++++++++-----------
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h     |  1 +
 3 files changed, 46 insertions(+), 20 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gtt_mgr.c +0 -0
index f7d22c4..6fdf83a 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_gtt_mgr.c static int amdgpu_gtt_mgr_fini(struct ttm_mem_type_manager *man) · line 90

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gtt_mgr.c +7 -0
@@ -81,6 +81,20 @@ static int amdgpu_gtt_mgr_fini(struct ttm_mem_type_manager *man)
 }
 
 /**
+ * amdgpu_gtt_mgr_is_allocated - Check if mem has address space
+ *
+ * @mem: the mem object to check
+ *
+ * Check if a mem object has already address space allocated.
+ */
+bool amdgpu_gtt_mgr_is_alloced(struct ttm_mem_reg *mem)

mismatch between documentation and function name. I prefer the full amdgpu_gtt_mgr_is_allocated or even better amdgpu_gtt_mgr_addr_is_allocated. With that fixed up:

Quoted 115 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gtt_mgr.c +8 -1
+{
+       struct drm_mm_node *node = mem->mm_node;
+
+       return (node->start != AMDGPU_BO_INVALID_OFFSET);
+}
+
+/**
  * amdgpu_gtt_mgr_alloc - allocate new ranges
  *
  * @man: TTM memory type manager
@@ -101,7 +115,7 @@ int amdgpu_gtt_mgr_alloc(struct ttm_mem_type_manager *man,
        unsigned long fpfn, lpfn;
        int r;
 
-       if (node->start != AMDGPU_BO_INVALID_OFFSET)
+       if (amdgpu_gtt_mgr_is_alloced(mem))
                return 0;
 
        if (place)
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +30 -19
index 5bfe7f6..eb0d7d7 100644
@@ -681,6 +681,31 @@ static void amdgpu_ttm_tt_unpin_userptr(struct ttm_tt *ttm)
        sg_free_table(ttm->sg);
 }
 
+static int amdgpu_ttm_do_bind(struct ttm_tt *ttm, struct ttm_mem_reg *mem)
+{
+       struct amdgpu_ttm_tt *gtt = (void *)ttm;
+       uint64_t flags;
+       int r;
+
+       spin_lock(&gtt->adev->gtt_list_lock);
+       flags = amdgpu_ttm_tt_pte_flags(gtt->adev, ttm, mem);
+       gtt->offset = (u64)mem->start << PAGE_SHIFT;
+       r = amdgpu_gart_bind(gtt->adev, gtt->offset, ttm->num_pages,
+               ttm->pages, gtt->ttm.dma_address, flags);
+
+       if (r) {
+               DRM_ERROR("failed to bind %lu pages at 0x%08llX\n",
+                         ttm->num_pages, gtt->offset);
+               goto error_gart_bind;
+       }
+
+       list_add_tail(&gtt->list, &gtt->adev->gtt_list);
+error_gart_bind:
+       spin_unlock(&gtt->adev->gtt_list_lock);
+       return r;
+
+}
+
 static int amdgpu_ttm_backend_bind(struct ttm_tt *ttm,
                                   struct ttm_mem_reg *bo_mem)
 {
@@ -704,7 +729,10 @@ static int amdgpu_ttm_backend_bind(struct ttm_tt *ttm,
            bo_mem->mem_type == AMDGPU_PL_OA)
                return -EINVAL;
 
-       return 0;
+       if (amdgpu_gtt_mgr_is_alloced(bo_mem))
+           r = amdgpu_ttm_do_bind(ttm, bo_mem);
+
+       return r;
 }
 
 bool amdgpu_ttm_is_bound(struct ttm_tt *ttm)
@@ -717,8 +745,6 @@ bool amdgpu_ttm_is_bound(struct ttm_tt *ttm)
 int amdgpu_ttm_bind(struct ttm_buffer_object *bo, struct ttm_mem_reg *bo_mem)
 {
        struct ttm_tt *ttm = bo->ttm;
-       struct amdgpu_ttm_tt *gtt = (void *)bo->ttm;
-       uint64_t flags;
        int r;
 
        if (!ttm || amdgpu_ttm_is_bound(ttm))
@@ -731,22 +757,7 @@ int amdgpu_ttm_bind(struct ttm_buffer_object *bo, struct ttm_mem_reg *bo_mem)
                return r;
        }
 
-       spin_lock(&gtt->adev->gtt_list_lock);
-       flags = amdgpu_ttm_tt_pte_flags(gtt->adev, ttm, bo_mem);
-       gtt->offset = (u64)bo_mem->start << PAGE_SHIFT;
-       r = amdgpu_gart_bind(gtt->adev, gtt->offset, ttm->num_pages,
-               ttm->pages, gtt->ttm.dma_address, flags);
-
-       if (r) {
-               DRM_ERROR("failed to bind %lu pages at 0x%08llX\n",
-                         ttm->num_pages, gtt->offset);
-               goto error_gart_bind;
-       }
-
-       list_add_tail(&gtt->list, &gtt->adev->gtt_list);
-error_gart_bind:
-       spin_unlock(&gtt->adev->gtt_list_lock);
-       return r;
+       return amdgpu_ttm_do_bind(ttm, bo_mem);
 }
 
 int amdgpu_ttm_recover_gart(struct amdgpu_device *adev)
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h +1 -1
index cd5bbfa..2ade5c5 100644
@@ -56,6 +56,7 @@ struct amdgpu_mman {
 extern const struct ttm_mem_type_manager_func amdgpu_gtt_mgr_func;
 extern const struct ttm_mem_type_manager_func amdgpu_vram_mgr_func;
 
+bool amdgpu_gtt_mgr_is_alloced(struct ttm_mem_reg *mem);
 int amdgpu_gtt_mgr_alloc(struct ttm_mem_type_manager *man,
                         struct ttm_buffer_object *tbo,
                         const struct ttm_place *place,
--

2.7.4

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

#17 lore raw

PATCH 7/12 drm/amdgpu: reserve the first 2x2MB of GART

In reply to a message not in the archive

From: Christian König <[email protected]>

We want to use them as remap address space.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_gtt_mgr.c | 5 ++++-
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h     | 3 +++
 2 files changed, 7 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gtt_mgr.c +4 -1
index 6fdf83a..a0976dc 100644
@@ -43,12 +43,15 @@ static int amdgpu_gtt_mgr_init(struct ttm_mem_type_manager *man,
 			       unsigned long p_size)
 {
 	struct amdgpu_gtt_mgr *mgr;
+	uint64_t start, size;
 
 	mgr = kzalloc(sizeof(*mgr), GFP_KERNEL);
 	if (!mgr)
 		return -ENOMEM;
 
-	drm_mm_init(&mgr->mm, 0, p_size);
+	start = AMDGPU_GTT_MAX_TRANSFER_SIZE * AMDGPU_GTT_NUM_TRANSFER_WINDOWS;
+	size = p_size - start;
+	drm_mm_init(&mgr->mm, start, size);
 	spin_lock_init(&mgr->lock);
 	mgr->available = p_size;
 	man->priv = mgr;
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h +3 -0
index 2ade5c5..9c4da0a 100644
@@ -34,6 +34,9 @@
 #define AMDGPU_PL_FLAG_GWS		(TTM_PL_FLAG_PRIV << 1)
 #define AMDGPU_PL_FLAG_OA		(TTM_PL_FLAG_PRIV << 2)
 
+#define AMDGPU_GTT_MAX_TRANSFER_SIZE	512
+#define AMDGPU_GTT_NUM_TRANSFER_WINDOWS	2
+
 struct amdgpu_mman {
 	struct ttm_bo_global_ref        bo_global_ref;
 	struct drm_global_reference	mem_global_ref;
Signature · 6 lines
2.7.4

_______________________________________________
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): 1498821733-1545-7-git-send-email-deathsimple-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

#19 lore raw

Re: [PATCH 07/12] drm/amdgpu: reserve the first 2x2MB of GART

Quoted 36 lines · Christian König

From: Christian König <[email protected]>

We want to use them as remap address space.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_gtt_mgr.c | 5 ++++-
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h     | 3 +++
 2 files changed, 7 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gtt_mgr.c +4 -1
index 6fdf83a..a0976dc 100644
@@ -43,12 +43,15 @@ static int amdgpu_gtt_mgr_init(struct ttm_mem_type_manager *man,
                               unsigned long p_size)
 {
        struct amdgpu_gtt_mgr *mgr;
+       uint64_t start, size;
 
        mgr = kzalloc(sizeof(*mgr), GFP_KERNEL);
        if (!mgr)
                return -ENOMEM;
 
-       drm_mm_init(&mgr->mm, 0, p_size);
+       start = AMDGPU_GTT_MAX_TRANSFER_SIZE * AMDGPU_GTT_NUM_TRANSFER_WINDOWS;
+       size = p_size - start;
+       drm_mm_init(&mgr->mm, start, size);
        spin_lock_init(&mgr->lock);
        mgr->available = p_size;
        man->priv = mgr;
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h +0 -0
index 2ade5c5..9c4da0a 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h line 37

Quoted 4 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h +1 -0
@@ -34,6 +34,9 @@
 #define AMDGPU_PL_FLAG_GWS             (TTM_PL_FLAG_PRIV << 1)
 #define AMDGPU_PL_FLAG_OA              (TTM_PL_FLAG_PRIV << 2)
 
+#define AMDGPU_GTT_MAX_TRANSFER_SIZE   512

Maybe AMDGPU_GTT_MAX_TRANSFER_SIZE_PAGES? Also you may want to update the patch title to say 2x512 pages rather than 2x2MB.

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h +2 -1
+#define AMDGPU_GTT_NUM_TRANSFER_WINDOWS        2
+
 struct amdgpu_mman {
        struct ttm_bo_global_ref        bo_global_ref;
        struct drm_global_reference     mem_global_ref;
--

2.7.4

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

#20 lore raw

PATCH 8/12 drm/amdgpu: add amdgpu_gart_map function

In reply to a message not in the archive

From: Christian König <[email protected]>

This allows us to write the mapped PTEs into
an IB instead of the table directly.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu.h      |  3 ++
 drivers/gpu/drm/amd/amdgpu/amdgpu_gart.c | 64 ++++++++++++++++++++++++--------
 2 files changed, 52 insertions(+), 15 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +3 -0
index 810796a..4a2b33d 100644
@@ -572,6 +572,9 @@ int amdgpu_gart_init(struct amdgpu_device *adev);
 void amdgpu_gart_fini(struct amdgpu_device *adev);
 int amdgpu_gart_unbind(struct amdgpu_device *adev, uint64_t offset,
 			int pages);
+int amdgpu_gart_map(struct amdgpu_device *adev, uint64_t offset,
+		    int pages, dma_addr_t *dma_addr, uint64_t flags,
+		    void *dst);
 int amdgpu_gart_bind(struct amdgpu_device *adev, uint64_t offset,
 		     int pages, struct page **pagelist,
 		     dma_addr_t *dma_addr, uint64_t flags);
drivers/gpu/drm/amd/amdgpu/amdgpu_gart.c +49 -15

86 lines not on this page · raw message

Signature · 6 lines
2.7.4

_______________________________________________
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): 1498821733-1545-8-git-send-email-deathsimple-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

Reviewed-by
#22 lore raw

Re: [PATCH 08/12] drm/amdgpu: add amdgpu_gart_map function

Quoted 6 lines · Christian König

From: Christian König <[email protected]>

This allows us to write the mapped PTEs into
an IB instead of the table directly.

Quoted 114 lines
---
 drivers/gpu/drm/amd/amdgpu/amdgpu.h      |  3 ++
 drivers/gpu/drm/amd/amdgpu/amdgpu_gart.c | 64 ++++++++++++++++++++++++--------
 2 files changed, 52 insertions(+), 15 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +3 -0
index 810796a..4a2b33d 100644
@@ -572,6 +572,9 @@ int amdgpu_gart_init(struct amdgpu_device *adev);
 void amdgpu_gart_fini(struct amdgpu_device *adev);
 int amdgpu_gart_unbind(struct amdgpu_device *adev, uint64_t offset,
                        int pages);
+int amdgpu_gart_map(struct amdgpu_device *adev, uint64_t offset,
+                   int pages, dma_addr_t *dma_addr, uint64_t flags,
+                   void *dst);
 int amdgpu_gart_bind(struct amdgpu_device *adev, uint64_t offset,
                     int pages, struct page **pagelist,
                     dma_addr_t *dma_addr, uint64_t flags);
drivers/gpu/drm/amd/amdgpu/amdgpu_gart.c +49 -16
index 8877015..d99b2b2 100644
@@ -280,6 +280,43 @@ int amdgpu_gart_unbind(struct amdgpu_device *adev, uint64_t offset,
 }
 
 /**
+ * amdgpu_gart_map - map dma_addresses into GART entries
+ *
+ * @adev: amdgpu_device pointer
+ * @offset: offset into the GPU's gart aperture
+ * @pages: number of pages to bind
+ * @dma_addr: DMA addresses of pages
+ *
+ * Map the dma_addresses into GART entries (all asics).
+ * Returns 0 for success, -EINVAL for failure.
+ */
+int amdgpu_gart_map(struct amdgpu_device *adev, uint64_t offset,
+                   int pages, dma_addr_t *dma_addr, uint64_t flags,
+                   void *dst)
+{
+       uint64_t page_base;
+       unsigned t, p;
+       int i, j;
+
+       if (!adev->gart.ready) {
+               WARN(1, "trying to bind memory to uninitialized GART !\n");
+               return -EINVAL;
+       }
+
+       t = offset / AMDGPU_GPU_PAGE_SIZE;
+       p = t / (PAGE_SIZE / AMDGPU_GPU_PAGE_SIZE);
+
+       for (i = 0; i < pages; i++, p++) {
+               page_base = dma_addr[i];
+               for (j = 0; j < (PAGE_SIZE / AMDGPU_GPU_PAGE_SIZE); j++, t++) {
+                       amdgpu_gart_set_pte_pde(adev, dst, t, page_base, flags);
+                       page_base += AMDGPU_GPU_PAGE_SIZE;
+               }
+       }
+       return 0;
+}
+
+/**
  * amdgpu_gart_bind - bind pages into the gart page table
  *
  * @adev: amdgpu_device pointer
@@ -296,31 +333,28 @@ int amdgpu_gart_bind(struct amdgpu_device *adev, uint64_t offset,
                     int pages, struct page **pagelist, dma_addr_t *dma_addr,
                     uint64_t flags)
 {
-       unsigned t;
-       unsigned p;
-       uint64_t page_base;
-       int i, j;
+#ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
+       unsigned i;
+#endif
+       int r;
 
        if (!adev->gart.ready) {
                WARN(1, "trying to bind memory to uninitialized GART !\n");
                return -EINVAL;
        }
 
-       t = offset / AMDGPU_GPU_PAGE_SIZE;
-       p = t / (PAGE_SIZE / AMDGPU_GPU_PAGE_SIZE);
-
-       for (i = 0; i < pages; i++, p++) {
 #ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
+       for (i = 0; i < pages; i++, p++)
                adev->gart.pages[p] = pagelist[i];
 #endif
-               if (adev->gart.ptr) {
-                       page_base = dma_addr[i];
-                       for (j = 0; j < (PAGE_SIZE / AMDGPU_GPU_PAGE_SIZE); j++, t++) {
-                               amdgpu_gart_set_pte_pde(adev, adev->gart.ptr, t, page_base, flags);
-                               page_base += AMDGPU_GPU_PAGE_SIZE;
-                       }
-               }
+
+       if (adev->gart.ptr) {
+               r = amdgpu_gart_map(adev, offset, pages, dma_addr, flags,
+                           adev->gart.ptr);
+               if (r)
+                       return r;
        }
+
        mb();
        amdgpu_gart_flush_gpu_tlb(adev, 0);
        return 0;
--

2.7.4

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

#23 lore raw

Re: [PATCH 08/12] drm/amdgpu: add amdgpu_gart_map function

In reply to a message not in the archive

Hi Christian, two comments below

Den 30 juni 2017 1:22 em skrev "Christian König" <[email protected]>:

From: Christian König <[email protected]>

This allows us to write the mapped PTEs into
an IB instead of the table directly.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu.h      |  3 ++
 drivers/gpu/drm/amd/amdgpu/amdgpu_gart.c | 64
++++++++++++++++++++++++--------
 2 files changed, 52 insertions(+), 15 deletions(-)
(quoted diff) +0 -0

b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
index 810796a..4a2b33d 100644

drivers/gpu/drm/amd/amdgpu/amdgpu.h +3 -0
@@ -572,6 +572,9 @@ int amdgpu_gart_init(struct amdgpu_device *adev);
 void amdgpu_gart_fini(struct amdgpu_device *adev);
 int amdgpu_gart_unbind(struct amdgpu_device *adev, uint64_t offset,
                        int pages);
+int amdgpu_gart_map(struct amdgpu_device *adev, uint64_t offset,
+                   int pages, dma_addr_t *dma_addr, uint64_t flags,
+                   void *dst);
 int amdgpu_gart_bind(struct amdgpu_device *adev, uint64_t offset,
                     int pages, struct page **pagelist,
                     dma_addr_t *dma_addr, uint64_t flags);
(quoted diff) +0 -0

b/drivers/gpu/drm/amd/amdgpu/amdgpu_gart.c
index 8877015..d99b2b2 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_gart.c +0 -0
@@ -280,6 +280,43 @@ int amdgpu_gart_unbind(struct amdgpu_device *adev,

uint64_t offset,
}

/**
+ * amdgpu_gart_map - map dma_addresses into GART entries
+ *
+ * @adev: amdgpu_device pointer
+ * @offset: offset into the GPU's gart aperture
+ * @pages: number of pages to bind
+ * @dma_addr: DMA addresses of pages
+ *
+ * Map the dma_addresses into GART entries (all asics).
+ * Returns 0 for success, -EINVAL for failure.
+ */
+int amdgpu_gart_map(struct amdgpu_device *adev, uint64_t offset,
+ int pages, dma_addr_t *dma_addr, uint64_t flags,
+ void *dst)
+{

+       uint64_t page_base;
+       unsigned t, p;
+       int i, j;

+
+ if (!adev->gart.ready) {
+ WARN(1, "trying to bind memory to uninitialized GART !\n");
+ return -EINVAL;
+ }
+
+ t = offset / AMDGPU_GPU_PAGE_SIZE;
+ p = t / (PAGE_SIZE / AMDGPU_GPU_PAGE_SIZE);
+
+ for (i = 0; i < pages; i++, p++) {
+ page_base = dma_addr[i];
+ for (j = 0; j < (PAGE_SIZE / AMDGPU_GPU_PAGE_SIZE); j++,
t++) {
+ amdgpu_gart_set_pte_pde(adev, dst, t, page_base,
flags);
+ page_base += AMDGPU_GPU_PAGE_SIZE;
+ }
+ }
+ return 0;
+}

The p variable seems set but not used in the above function.

+
+/**
* amdgpu_gart_bind - bind pages into the gart page table *
* @adev: amdgpu_device pointer
@@ -296,31 +333,28 @@ int amdgpu_gart_bind(struct amdgpu_device *adev, uint64_t offset,
int pages, struct page **pagelist, dma_addr_t
*dma_addr,

                     uint64_t flags)
 {
-       unsigned t;
-       unsigned p;
-       uint64_t page_base;
-       int i, j;

+#ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
+ unsigned i;
+#endif
+ int r;

        if (!adev->gart.ready) {
                WARN(1, "trying to bind memory to uninitialized GART !\n");
                return -EINVAL;
        }

- t = offset / AMDGPU_GPU_PAGE_SIZE;
- p = t / (PAGE_SIZE / AMDGPU_GPU_PAGE_SIZE);
-
- for (i = 0; i < pages; i++, p++) {
#ifdef CONFIG_DRM_AMDGPU_GART_DEBUGFS
+ for (i = 0; i < pages; i++, p++)

                adev->gart.pages[p] = pagelist[i];

#endif

The p variable is not declared in this function so building with CONFIG_DRM_AMDGPU_GART_DEBUGFS will fail.

Regards,
Nils

- if (adev->gart.ptr) {
- page_base = dma_addr[i];
- for (j = 0; j < (PAGE_SIZE / AMDGPU_GPU_PAGE_SIZE);
j++, t++) {
- amdgpu_gart_set_pte_pde(adev,
adev->gart.ptr, t, page_base, flags);
- page_base += AMDGPU_GPU_PAGE_SIZE;
- }
- }
+
+ if (adev->gart.ptr) {
+ r = amdgpu_gart_map(adev, offset, pages, dma_addr, flags,
+ adev->gart.ptr);
+ if (r)

+                       return r;
        }
+
        mb();
        amdgpu_gart_flush_gpu_tlb(adev, 0);
        return 0;

--
2.7.4

_______________________________________________
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

#24 lore raw

PATCH 9/12 drm/amdgpu: use the GTT windows for BO moves

In reply to a message not in the archive

From: Christian König <[email protected]>

This way we don't need to map the full BO at a time any more.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 127 +++++++++++++++++++++++++++-----
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h |   3 +
 2 files changed, 111 insertions(+), 19 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +108 -19
index eb0d7d7..247ce21 100644
@@ -47,10 +47,15 @@
 
 #define DRM_FILE_PAGE_OFFSET (0x100000000ULL >> PAGE_SHIFT)
 
+static int amdgpu_map_buffer(struct ttm_buffer_object *bo,
+			     struct ttm_mem_reg *mem,
+			     unsigned num_pages, uint64_t offset,
+			     struct amdgpu_ring *ring,
+			     uint64_t *addr);
+
 static int amdgpu_ttm_debugfs_init(struct amdgpu_device *adev);
 static void amdgpu_ttm_debugfs_fini(struct amdgpu_device *adev);
 
-
 /*
  * Global memory.
  */
@@ -97,6 +102,9 @@ static int amdgpu_ttm_global_init(struct amdgpu_device *adev)
 		goto error_bo;
 	}
 
+	mutex_init(&adev->mman.gtt_window_lock);
+	adev->mman.gtt_index = 0;
+
 	ring = adev->mman.buffer_funcs_ring;
 	rq = &ring->sched.sched_rq[AMD_SCHED_PRIORITY_KERNEL];
 	r = amd_sched_entity_init(&ring->sched, &adev->mman.entity,
@@ -123,6 +131,7 @@ static void amdgpu_ttm_global_fini(struct amdgpu_device *adev)
 	if (adev->mman.mem_global_referenced) {
 		amd_sched_entity_fini(adev->mman.entity.sched,
 				      &adev->mman.entity);
+		mutex_destroy(&adev->mman.gtt_window_lock);
 		drm_global_item_unref(&adev->mman.bo_global_ref.ref);
 		drm_global_item_unref(&adev->mman.mem_global_ref);
 		adev->mman.mem_global_referenced = false;
@@ -256,10 +265,12 @@ static uint64_t amdgpu_mm_node_addr(struct ttm_buffer_object *bo,
 				    struct drm_mm_node *mm_node,
 				    struct ttm_mem_reg *mem)
 {
-	uint64_t addr;
+	uint64_t addr = 0;
 
-	addr = mm_node->start << PAGE_SHIFT;
-	addr += bo->bdev->man[mem->mem_type].gpu_offset;
+	if (mm_node->start != AMDGPU_BO_INVALID_OFFSET) {
+		addr = mm_node->start << PAGE_SHIFT;
+		addr += bo->bdev->man[mem->mem_type].gpu_offset;
+	}
 	return addr;
 }
 
@@ -284,34 +295,41 @@ static int amdgpu_move_blit(struct ttm_buffer_object *bo,
 		return -EINVAL;
 	}
 
-	if (old_mem->mem_type == TTM_PL_TT) {
-		r = amdgpu_ttm_bind(bo, old_mem);
-		if (r)
-			return r;
-	}
-
 	old_mm = old_mem->mm_node;
 	old_size = old_mm->size;
 	old_start = amdgpu_mm_node_addr(bo, old_mm, old_mem);
 
-	if (new_mem->mem_type == TTM_PL_TT) {
-		r = amdgpu_ttm_bind(bo, new_mem);
-		if (r)
-			return r;
-	}
-
 	new_mm = new_mem->mm_node;
 	new_size = new_mm->size;
 	new_start = amdgpu_mm_node_addr(bo, new_mm, new_mem);
 
 	num_pages = new_mem->num_pages;
+	mutex_lock(&adev->mman.gtt_window_lock);
 	while (num_pages) {
-		unsigned long cur_pages = min(old_size, new_size);
+		unsigned long cur_pages = min(min(old_size, new_size),
+					      (u64)AMDGPU_GTT_MAX_TRANSFER_SIZE);
+		uint64_t from = old_start, to = new_start;
 		struct dma_fence *next;
 
-		r = amdgpu_copy_buffer(ring, old_start, new_start,
+		if (old_mem->mem_type == TTM_PL_TT &&
+		    !amdgpu_gtt_mgr_is_alloced(old_mem)) {
+			r = amdgpu_map_buffer(bo, old_mem, cur_pages,
+					      old_start, ring, &from);
+			if (r)
+				goto error;
+		}
+
+		if (new_mem->mem_type == TTM_PL_TT &&
+		    !amdgpu_gtt_mgr_is_alloced(new_mem)) {
+			r = amdgpu_map_buffer(bo, new_mem, cur_pages,
+					      new_start, ring, &to);
+			if (r)
+				goto error;
+		}
+
+		r = amdgpu_copy_buffer(ring, from, to,
 				       cur_pages * PAGE_SIZE,
-				       bo->resv, &next, false, false);
+				       bo->resv, &next, false, true);
 		if (r)
 			goto error;
 
@@ -338,12 +356,15 @@ static int amdgpu_move_blit(struct ttm_buffer_object *bo,
 			new_start += cur_pages * PAGE_SIZE;
 		}
 	}
+	mutex_unlock(&adev->mman.gtt_window_lock);
 
 	r = ttm_bo_pipeline_move(bo, fence, evict, new_mem);
 	dma_fence_put(fence);
 	return r;
 
 error:
+	mutex_unlock(&adev->mman.gtt_window_lock);
+
 	if (fence)
 		dma_fence_wait(fence, false);
 	dma_fence_put(fence);
@@ -1253,6 +1274,74 @@ int amdgpu_mmap(struct file *filp, struct vm_area_struct *vma)
 	return ttm_bo_mmap(filp, vma, &adev->mman.bdev);
 }
 
+static int amdgpu_map_buffer(struct ttm_buffer_object *bo,
+			     struct ttm_mem_reg *mem,
+			     unsigned num_pages, uint64_t offset,
+			     struct amdgpu_ring *ring,
+			     uint64_t *addr)
+{
+	struct amdgpu_ttm_tt *gtt = (void *)bo->ttm;
+	struct amdgpu_device *adev = ring->adev;
+	struct ttm_tt *ttm = bo->ttm;
+	struct amdgpu_job *job;
+	unsigned num_dw, num_bytes, idx;
+	dma_addr_t *dma_address;
+	struct dma_fence *fence;
+	uint64_t src_addr, dst_addr;
+	uint64_t flags;
+	int r;
+
+	BUG_ON(adev->mman.buffer_funcs->copy_max_bytes <
+	       AMDGPU_GTT_MAX_TRANSFER_SIZE * 8);
+
+	idx = adev->mman.gtt_index++;
+	adev->mman.gtt_index %= AMDGPU_GTT_NUM_TRANSFER_WINDOWS;
+
+	*addr = adev->mc.gtt_start;
+	*addr += (u64)idx * AMDGPU_GTT_MAX_TRANSFER_SIZE * AMDGPU_GPU_PAGE_SIZE;
+
+	num_dw = adev->mman.buffer_funcs->copy_num_dw;
+	while (num_dw & 0x7)
+		num_dw++;
+
+	num_bytes = num_pages * 8;
+
+	r = amdgpu_job_alloc_with_ib(adev, num_dw * 4 + num_bytes, &job);
+	if (r)
+		return r;
+
+	src_addr = num_dw * 4;
+	src_addr += job->ibs[0].gpu_addr;
+
+	dst_addr = adev->gart.table_addr;
+	dst_addr += idx * AMDGPU_GTT_MAX_TRANSFER_SIZE * 8;
+	amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr,
+				dst_addr, num_bytes);
+
+	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
+	WARN_ON(job->ibs[0].length_dw > num_dw);
+
+	dma_address = &gtt->ttm.dma_address[offset >> PAGE_SHIFT];
+	flags = amdgpu_ttm_tt_pte_flags(adev, ttm, mem);
+	r = amdgpu_gart_map(adev, 0, num_pages, dma_address, flags,
+			    &job->ibs[0].ptr[num_dw]);
+	if (r)
+		goto error_free;
+
+	r = amdgpu_job_submit(job, ring, &adev->mman.entity,
+			      AMDGPU_FENCE_OWNER_UNDEFINED, &fence);
+	if (r)
+		goto error_free;
+
+	dma_fence_put(fence);
+
+	return r;
+
+error_free:
+	amdgpu_job_free(job);
+	return r;
+}
+
 int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
 		       uint64_t dst_offset, uint32_t byte_count,
 		       struct reservation_object *resv,
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h +3 -0
index 9c4da0a..2e1cce5 100644
@@ -52,6 +52,9 @@ struct amdgpu_mman {
 	/* buffer handling */
 	const struct amdgpu_buffer_funcs	*buffer_funcs;
 	struct amdgpu_ring			*buffer_funcs_ring;
+
+	struct mutex				gtt_window_lock;
+	unsigned				gtt_index;
 	/* Scheduler entity for buffer moves */
 	struct amd_sched_entity			entity;
 };
Signature · 6 lines
2.7.4

_______________________________________________
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): 1498821733-1545-9-git-send-email-deathsimple-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

Reviewed-by
#26 lore raw

Re: [PATCH 09/12] drm/amdgpu: use the GTT windows for BO moves

Quoted 5 lines · Christian König

From: Christian König <[email protected]>

This way we don't need to map the full BO at a time any more.

Quoted 225 lines
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 127 +++++++++++++++++++++++++++-----
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h |   3 +
 2 files changed, 111 insertions(+), 19 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +108 -19
index eb0d7d7..247ce21 100644
@@ -47,10 +47,15 @@
 
 #define DRM_FILE_PAGE_OFFSET (0x100000000ULL >> PAGE_SHIFT)
 
+static int amdgpu_map_buffer(struct ttm_buffer_object *bo,
+                            struct ttm_mem_reg *mem,
+                            unsigned num_pages, uint64_t offset,
+                            struct amdgpu_ring *ring,
+                            uint64_t *addr);
+
 static int amdgpu_ttm_debugfs_init(struct amdgpu_device *adev);
 static void amdgpu_ttm_debugfs_fini(struct amdgpu_device *adev);
 
-
 /*
  * Global memory.
  */
@@ -97,6 +102,9 @@ static int amdgpu_ttm_global_init(struct amdgpu_device *adev)
                goto error_bo;
        }
 
+       mutex_init(&adev->mman.gtt_window_lock);
+       adev->mman.gtt_index = 0;
+
        ring = adev->mman.buffer_funcs_ring;
        rq = &ring->sched.sched_rq[AMD_SCHED_PRIORITY_KERNEL];
        r = amd_sched_entity_init(&ring->sched, &adev->mman.entity,
@@ -123,6 +131,7 @@ static void amdgpu_ttm_global_fini(struct amdgpu_device *adev)
        if (adev->mman.mem_global_referenced) {
                amd_sched_entity_fini(adev->mman.entity.sched,
                                      &adev->mman.entity);
+               mutex_destroy(&adev->mman.gtt_window_lock);
                drm_global_item_unref(&adev->mman.bo_global_ref.ref);
                drm_global_item_unref(&adev->mman.mem_global_ref);
                adev->mman.mem_global_referenced = false;
@@ -256,10 +265,12 @@ static uint64_t amdgpu_mm_node_addr(struct ttm_buffer_object *bo,
                                    struct drm_mm_node *mm_node,
                                    struct ttm_mem_reg *mem)
 {
-       uint64_t addr;
+       uint64_t addr = 0;
 
-       addr = mm_node->start << PAGE_SHIFT;
-       addr += bo->bdev->man[mem->mem_type].gpu_offset;
+       if (mm_node->start != AMDGPU_BO_INVALID_OFFSET) {
+               addr = mm_node->start << PAGE_SHIFT;
+               addr += bo->bdev->man[mem->mem_type].gpu_offset;
+       }
        return addr;
 }
 
@@ -284,34 +295,41 @@ static int amdgpu_move_blit(struct ttm_buffer_object *bo,
                return -EINVAL;
        }
 
-       if (old_mem->mem_type == TTM_PL_TT) {
-               r = amdgpu_ttm_bind(bo, old_mem);
-               if (r)
-                       return r;
-       }
-
        old_mm = old_mem->mm_node;
        old_size = old_mm->size;
        old_start = amdgpu_mm_node_addr(bo, old_mm, old_mem);
 
-       if (new_mem->mem_type == TTM_PL_TT) {
-               r = amdgpu_ttm_bind(bo, new_mem);
-               if (r)
-                       return r;
-       }
-
        new_mm = new_mem->mm_node;
        new_size = new_mm->size;
        new_start = amdgpu_mm_node_addr(bo, new_mm, new_mem);
 
        num_pages = new_mem->num_pages;
+       mutex_lock(&adev->mman.gtt_window_lock);
        while (num_pages) {
-               unsigned long cur_pages = min(old_size, new_size);
+               unsigned long cur_pages = min(min(old_size, new_size),
+                                             (u64)AMDGPU_GTT_MAX_TRANSFER_SIZE);
+               uint64_t from = old_start, to = new_start;
                struct dma_fence *next;
 
-               r = amdgpu_copy_buffer(ring, old_start, new_start,
+               if (old_mem->mem_type == TTM_PL_TT &&
+                   !amdgpu_gtt_mgr_is_alloced(old_mem)) {
+                       r = amdgpu_map_buffer(bo, old_mem, cur_pages,
+                                             old_start, ring, &from);
+                       if (r)
+                               goto error;
+               }
+
+               if (new_mem->mem_type == TTM_PL_TT &&
+                   !amdgpu_gtt_mgr_is_alloced(new_mem)) {
+                       r = amdgpu_map_buffer(bo, new_mem, cur_pages,
+                                             new_start, ring, &to);
+                       if (r)
+                               goto error;
+               }
+
+               r = amdgpu_copy_buffer(ring, from, to,
                                       cur_pages * PAGE_SIZE,
-                                      bo->resv, &next, false, false);
+                                      bo->resv, &next, false, true);
                if (r)
                        goto error;
 
@@ -338,12 +356,15 @@ static int amdgpu_move_blit(struct ttm_buffer_object *bo,
                        new_start += cur_pages * PAGE_SIZE;
                }
        }
+       mutex_unlock(&adev->mman.gtt_window_lock);
 
        r = ttm_bo_pipeline_move(bo, fence, evict, new_mem);
        dma_fence_put(fence);
        return r;
 
 error:
+       mutex_unlock(&adev->mman.gtt_window_lock);
+
        if (fence)
                dma_fence_wait(fence, false);
        dma_fence_put(fence);
@@ -1253,6 +1274,74 @@ int amdgpu_mmap(struct file *filp, struct vm_area_struct *vma)
        return ttm_bo_mmap(filp, vma, &adev->mman.bdev);
 }
 
+static int amdgpu_map_buffer(struct ttm_buffer_object *bo,
+                            struct ttm_mem_reg *mem,
+                            unsigned num_pages, uint64_t offset,
+                            struct amdgpu_ring *ring,
+                            uint64_t *addr)
+{
+       struct amdgpu_ttm_tt *gtt = (void *)bo->ttm;
+       struct amdgpu_device *adev = ring->adev;
+       struct ttm_tt *ttm = bo->ttm;
+       struct amdgpu_job *job;
+       unsigned num_dw, num_bytes, idx;
+       dma_addr_t *dma_address;
+       struct dma_fence *fence;
+       uint64_t src_addr, dst_addr;
+       uint64_t flags;
+       int r;
+
+       BUG_ON(adev->mman.buffer_funcs->copy_max_bytes <
+              AMDGPU_GTT_MAX_TRANSFER_SIZE * 8);
+
+       idx = adev->mman.gtt_index++;
+       adev->mman.gtt_index %= AMDGPU_GTT_NUM_TRANSFER_WINDOWS;
+
+       *addr = adev->mc.gtt_start;
+       *addr += (u64)idx * AMDGPU_GTT_MAX_TRANSFER_SIZE * AMDGPU_GPU_PAGE_SIZE;
+
+       num_dw = adev->mman.buffer_funcs->copy_num_dw;
+       while (num_dw & 0x7)
+               num_dw++;
+
+       num_bytes = num_pages * 8;
+
+       r = amdgpu_job_alloc_with_ib(adev, num_dw * 4 + num_bytes, &job);
+       if (r)
+               return r;
+
+       src_addr = num_dw * 4;
+       src_addr += job->ibs[0].gpu_addr;
+
+       dst_addr = adev->gart.table_addr;
+       dst_addr += idx * AMDGPU_GTT_MAX_TRANSFER_SIZE * 8;
+       amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr,
+                               dst_addr, num_bytes);
+
+       amdgpu_ring_pad_ib(ring, &job->ibs[0]);
+       WARN_ON(job->ibs[0].length_dw > num_dw);
+
+       dma_address = &gtt->ttm.dma_address[offset >> PAGE_SHIFT];
+       flags = amdgpu_ttm_tt_pte_flags(adev, ttm, mem);
+       r = amdgpu_gart_map(adev, 0, num_pages, dma_address, flags,
+                           &job->ibs[0].ptr[num_dw]);
+       if (r)
+               goto error_free;
+
+       r = amdgpu_job_submit(job, ring, &adev->mman.entity,
+                             AMDGPU_FENCE_OWNER_UNDEFINED, &fence);
+       if (r)
+               goto error_free;
+
+       dma_fence_put(fence);
+
+       return r;
+
+error_free:
+       amdgpu_job_free(job);
+       return r;
+}
+
 int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t src_offset,
                       uint64_t dst_offset, uint32_t byte_count,
                       struct reservation_object *resv,
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h +3 -1
index 9c4da0a..2e1cce5 100644
@@ -52,6 +52,9 @@ struct amdgpu_mman {
        /* buffer handling */
        const struct amdgpu_buffer_funcs        *buffer_funcs;
        struct amdgpu_ring                      *buffer_funcs_ring;
+
+       struct mutex                            gtt_window_lock;
+       unsigned                                gtt_index;
        /* Scheduler entity for buffer moves */
        struct amd_sched_entity                 entity;
 };
--

2.7.4

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

#27 lore raw

PATCH 10/12 drm/amdgpu: stop mapping BOs to GTT

In reply to a message not in the archive

From: Christian König <[email protected]>

No need to map BOs to GTT on eviction and intermediate transfers any more.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 19 ++-----------------
 1 file changed, 2 insertions(+), 17 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +2 -17
index 247ce21..e1ebcba 100644
@@ -200,7 +200,6 @@ static void amdgpu_evict_flags(struct ttm_buffer_object *bo,
 		.lpfn = 0,
 		.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM
 	};
-	unsigned i;
 
 	if (!amdgpu_ttm_bo_is_amdgpu_bo(bo)) {
 		placement->placement = &placements;
@@ -218,20 +217,6 @@ static void amdgpu_evict_flags(struct ttm_buffer_object *bo,
 			amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU);
 		} else {
 			amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_GTT);
-			for (i = 0; i < abo->placement.num_placement; ++i) {
-				if (!(abo->placements[i].flags &
-				      TTM_PL_FLAG_TT))
-					continue;
-
-				if (abo->placements[i].lpfn)
-					continue;
-
-				/* set an upper limit to force directly
-				 * allocating address space for the BO.
-				 */
-				abo->placements[i].lpfn =
-					adev->mc.gtt_size >> PAGE_SHIFT;
-			}
 		}
 		break;
 	case TTM_PL_TT:
@@ -391,7 +376,7 @@ static int amdgpu_move_vram_ram(struct ttm_buffer_object *bo,
 	placement.num_busy_placement = 1;
 	placement.busy_placement = &placements;
 	placements.fpfn = 0;
-	placements.lpfn = adev->mc.gtt_size >> PAGE_SHIFT;
+	placements.lpfn = 0;
 	placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
 	r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
 			     interruptible, no_wait_gpu);
@@ -438,7 +423,7 @@ static int amdgpu_move_ram_vram(struct ttm_buffer_object *bo,
 	placement.num_busy_placement = 1;
 	placement.busy_placement = &placements;
 	placements.fpfn = 0;
-	placements.lpfn = adev->mc.gtt_size >> PAGE_SHIFT;
+	placements.lpfn = 0;
 	placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
 	r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
 			     interruptible, no_wait_gpu);
Signature · 6 lines
2.7.4

_______________________________________________
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): 1498821733-1545-10-git-send-email-deathsimple-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

Acked-by
#29 lore raw

Re: [PATCH 10/12] drm/amdgpu: stop mapping BOs to GTT

Quoted 5 lines · Christian König

From: Christian König <[email protected]>

No need to map BOs to GTT on eviction and intermediate transfers any more.

Quoted 58 lines
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 19 ++-----------------
 1 file changed, 2 insertions(+), 17 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +2 -18
index 247ce21..e1ebcba 100644
@@ -200,7 +200,6 @@ static void amdgpu_evict_flags(struct ttm_buffer_object *bo,
                .lpfn = 0,
                .flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM
        };
-       unsigned i;
 
        if (!amdgpu_ttm_bo_is_amdgpu_bo(bo)) {
                placement->placement = &placements;
@@ -218,20 +217,6 @@ static void amdgpu_evict_flags(struct ttm_buffer_object *bo,
                        amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU);
                } else {
                        amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_GTT);
-                       for (i = 0; i < abo->placement.num_placement; ++i) {
-                               if (!(abo->placements[i].flags &
-                                     TTM_PL_FLAG_TT))
-                                       continue;
-
-                               if (abo->placements[i].lpfn)
-                                       continue;
-
-                               /* set an upper limit to force directly
-                                * allocating address space for the BO.
-                                */
-                               abo->placements[i].lpfn =
-                                       adev->mc.gtt_size >> PAGE_SHIFT;
-                       }
                }
                break;
        case TTM_PL_TT:
@@ -391,7 +376,7 @@ static int amdgpu_move_vram_ram(struct ttm_buffer_object *bo,
        placement.num_busy_placement = 1;
        placement.busy_placement = &placements;
        placements.fpfn = 0;
-       placements.lpfn = adev->mc.gtt_size >> PAGE_SHIFT;
+       placements.lpfn = 0;
        placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
        r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
                             interruptible, no_wait_gpu);
@@ -438,7 +423,7 @@ static int amdgpu_move_ram_vram(struct ttm_buffer_object *bo,
        placement.num_busy_placement = 1;
        placement.busy_placement = &placements;
        placements.fpfn = 0;
-       placements.lpfn = adev->mc.gtt_size >> PAGE_SHIFT;
+       placements.lpfn = 0;
        placements.flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_TT;
        r = ttm_bo_mem_space(bo, &placement, &tmp_mem,
                             interruptible, no_wait_gpu);
--

2.7.4

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

#30 lore raw

PATCH 11/12 drm/amdgpu: remove maximum BO size limitation.

In reply to a message not in the archive

From: Christian König <[email protected]>

We can finally remove this now.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 11 -----------
 1 file changed, 11 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +0 -11
index 96c4493..2382785 100644
@@ -58,17 +58,6 @@ int amdgpu_gem_object_create(struct amdgpu_device *adev, unsigned long size,
 		alignment = PAGE_SIZE;
 	}
 
-	if (!(initial_domain & (AMDGPU_GEM_DOMAIN_GDS | AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA))) {
-		/* Maximum bo size is the unpinned gtt size since we use the gtt to
-		 * handle vram to system pool migrations.
-		 */
-		max_size = adev->mc.gtt_size - adev->gart_pin_size;
-		if (size > max_size) {
-			DRM_DEBUG("Allocation size %ldMb bigger than %ldMb limit\n",
-				  size >> 20, max_size >> 20);
-			return -ENOMEM;
-		}
-	}
 retry:
 	r = amdgpu_bo_create(adev, size, alignment, kernel, initial_domain,
 			     flags, NULL, NULL, &robj);
Signature · 6 lines
2.7.4

_______________________________________________
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): 1498821733-1545-11-git-send-email-deathsimple-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

Reviewed-by
#32 lore raw

Re: [PATCH 11/12] drm/amdgpu: remove maximum BO size limitation.

Quoted 5 lines · Christian König

From: Christian König <[email protected]>

We can finally remove this now.

Woot!

Quoted 29 lines
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 11 -----------
 1 file changed, 11 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +0 -12
index 96c4493..2382785 100644
@@ -58,17 +58,6 @@ int amdgpu_gem_object_create(struct amdgpu_device *adev, unsigned long size,
                alignment = PAGE_SIZE;
        }
 
-       if (!(initial_domain & (AMDGPU_GEM_DOMAIN_GDS | AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA))) {
-               /* Maximum bo size is the unpinned gtt size since we use the gtt to
-                * handle vram to system pool migrations.
-                */
-               max_size = adev->mc.gtt_size - adev->gart_pin_size;
-               if (size > max_size) {
-                       DRM_DEBUG("Allocation size %ldMb bigger than %ldMb limit\n",
-                                 size >> 20, max_size >> 20);
-                       return -ENOMEM;
-               }
-       }
 retry:
        r = amdgpu_bo_create(adev, size, alignment, kernel, initial_domain,
                             flags, NULL, NULL, &robj);
--

2.7.4

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

#33 lore raw

Re: [PATCH 11/12] drm/amdgpu: remove maximum BO size limitation.

In reply to a message not in the archive
Quoted 18 lines · Christian König

From: Christian König <[email protected]>

We can finally remove this now.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 11 -----------
 1 file changed, 11 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +0 -1
index 96c4493..2382785 100644
@@ -58,17 +58,6 @@ int amdgpu_gem_object_create(struct amdgpu_device *adev, unsigned long size,
 		alignment = PAGE_SIZE;
 	}
 
-	if (!(initial_domain & (AMDGPU_GEM_DOMAIN_GDS | AMDGPU_GEM_DOMAIN_GWS | AMDGPU_GEM_DOMAIN_OA))) {

drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c int amdgpu_gem_object_create(struct amdgpu_device *adev, unsigned long size, · line 71

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +0 -10
-		/* Maximum bo size is the unpinned gtt size since we use the gtt to
-		 * handle vram to system pool migrations.
-		 */
-		max_size = adev->mc.gtt_size - adev->gart_pin_size;
-		if (size > max_size) {
-			DRM_DEBUG("Allocation size %ldMb bigger than %ldMb limit\n",
-				  size >> 20, max_size >> 20);
-			return -ENOMEM;
-		}
-	}

Do we still need to limit the size of VRAM BOs?

If not, remove the now unused max_size local.

Signature · 6 lines
Earthling Michel Dänzer               |               http://www.amd.com
Libre software enthusiast             |             Mesa and X developer
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
#34 lore raw

PATCH 12/12 drm/amdgpu: add gtt_sys_limit

In reply to a message not in the archive

From: Christian König <[email protected]>

Limit the size of the GART table for the system domain.

This saves us a bunch of visible VRAM, but also limitates the maximum BO size we can swap out.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu.h         | 2 ++
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c  | 6 ++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c     | 4 ++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_gart.c    | 8 ++++++--
 drivers/gpu/drm/amd/amdgpu/amdgpu_gtt_mgr.c | 6 ++++--
 5 files changed, 22 insertions(+), 4 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +2 -0
index 4a2b33d..ef8e6b9 100644
@@ -76,6 +76,7 @@
 extern int amdgpu_modeset;
 extern int amdgpu_vram_limit;
 extern int amdgpu_gart_size;
+extern unsigned amdgpu_gart_sys_limit;
 extern int amdgpu_moverate;
 extern int amdgpu_benchmarking;
 extern int amdgpu_testing;
@@ -605,6 +606,7 @@ struct amdgpu_mc {
 	u64			mc_vram_size;
 	u64			visible_vram_size;
 	u64			gtt_size;
+	u64			gtt_sys_limit;
 	u64			gtt_start;
 	u64			gtt_end;
 	u64			vram_start;
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +6 -0
index 5b1220f..7e3f8cb 100644
@@ -1122,6 +1122,12 @@ static void amdgpu_check_arguments(struct amdgpu_device *adev)
 		}
 	}
 
+	if (amdgpu_gart_sys_limit < 32) {
+		dev_warn(adev->dev, "gart sys limit (%d) too small\n",
+				 amdgpu_gart_sys_limit);
+			amdgpu_gart_sys_limit = 32;
+	}
+
 	amdgpu_check_vm_size(adev);
 
 	amdgpu_check_block_size(adev);
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +4 -0
index 5a1d794..907ae5e 100644
@@ -75,6 +75,7 @@
 
 int amdgpu_vram_limit = 0;
 int amdgpu_gart_size = -1; /* auto */
+unsigned amdgpu_gart_sys_limit = 256;
 int amdgpu_moverate = -1; /* auto */
 int amdgpu_benchmarking = 0;
 int amdgpu_testing = 0;
@@ -124,6 +125,9 @@ module_param_named(vramlimit, amdgpu_vram_limit, int, 0600);
 MODULE_PARM_DESC(gartsize, "Size of PCIE/IGP gart to setup in megabytes (32, 64, etc., -1 = auto)");
 module_param_named(gartsize, amdgpu_gart_size, int, 0600);
 
+MODULE_PARM_DESC(gartlimit, "GART limit for the system domain in megabytes (default 256)");
+module_param_named(gartlimit, amdgpu_gart_sys_limit, int, 0600);
+
 MODULE_PARM_DESC(moverate, "Maximum buffer migration rate in MB/s. (32, 64, etc., -1=auto, 0=1=disabled)");
 module_param_named(moverate, amdgpu_moverate, int, 0600);
 
drivers/gpu/drm/amd/amdgpu/amdgpu_gart.c +6 -2
index d99b2b2..f82eeaa 100644
@@ -70,6 +70,9 @@ void amdgpu_gart_set_defaults(struct amdgpu_device *adev)
 					adev->mc.mc_vram_size);
 	else
 		adev->mc.gtt_size = (uint64_t)amdgpu_gart_size << 20;
+
+	adev->mc.gtt_sys_limit = min((uint64_t)amdgpu_gart_sys_limit << 20,
+				     adev->mc.gtt_size);
 }
 
 /**
@@ -384,8 +387,9 @@ int amdgpu_gart_init(struct amdgpu_device *adev)
 	if (r)
 		return r;
 	/* Compute table size */
-	adev->gart.num_cpu_pages = adev->mc.gtt_size / PAGE_SIZE;
-	adev->gart.num_gpu_pages = adev->mc.gtt_size / AMDGPU_GPU_PAGE_SIZE;
+	adev->gart.num_cpu_pages = adev->mc.gtt_sys_limit / PAGE_SIZE;
+	adev->gart.num_gpu_pages = adev->mc.gtt_sys_limit /
+		AMDGPU_GPU_PAGE_SIZE;
 	DRM_INFO("GART: num cpu pages %u, num gpu pages %u\n",
 		 adev->gart.num_cpu_pages, adev->gart.num_gpu_pages);
 
drivers/gpu/drm/amd/amdgpu/amdgpu_gtt_mgr.c +4 -2
index a0976dc..9b516c5 100644
@@ -42,6 +42,7 @@ struct amdgpu_gtt_mgr {
 static int amdgpu_gtt_mgr_init(struct ttm_mem_type_manager *man,
 			       unsigned long p_size)
 {
+	struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev);
 	struct amdgpu_gtt_mgr *mgr;
 	uint64_t start, size;
 
@@ -50,7 +51,7 @@ static int amdgpu_gtt_mgr_init(struct ttm_mem_type_manager *man,
 		return -ENOMEM;
 
 	start = AMDGPU_GTT_MAX_TRANSFER_SIZE * AMDGPU_GTT_NUM_TRANSFER_WINDOWS;
-	size = p_size - start;
+	size = (adev->mc.gtt_sys_limit >> PAGE_SHIFT) - start;
 	drm_mm_init(&mgr->mm, start, size);
 	spin_lock_init(&mgr->lock);
 	mgr->available = p_size;
@@ -112,6 +113,7 @@ int amdgpu_gtt_mgr_alloc(struct ttm_mem_type_manager *man,
 			 const struct ttm_place *place,
 			 struct ttm_mem_reg *mem)
 {
+	struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev);
 	struct amdgpu_gtt_mgr *mgr = man->priv;
 	struct drm_mm_node *node = mem->mm_node;
 	enum drm_mm_insert_mode mode;
@@ -129,7 +131,7 @@ int amdgpu_gtt_mgr_alloc(struct ttm_mem_type_manager *man,
 	if (place && place->lpfn)
 		lpfn = place->lpfn;
 	else
-		lpfn = man->size;
+		lpfn = adev->gart.num_cpu_pages;
 
 	mode = DRM_MM_INSERT_BEST;
 	if (place && place->flags & TTM_PL_FLAG_TOPDOWN)
Signature · 6 lines
2.7.4

_______________________________________________
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): 1498821733-1545-12-git-send-email-deathsimple-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

Acked-by
#36 lore raw

Re: [PATCH 12/12] drm/amdgpu: add gtt_sys_limit

Quoted 5 lines · Christian König

From: Christian König <[email protected]>

Limit the size of the GART table for the system domain.

This saves us a bunch of visible VRAM, but also limitates the maximum BO size we can swap out.

The last phrase can be dropped as it's no longer relevant.

Quoted 136 lines
---
 drivers/gpu/drm/amd/amdgpu/amdgpu.h         | 2 ++
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c  | 6 ++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c     | 4 ++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_gart.c    | 8 ++++++--
 drivers/gpu/drm/amd/amdgpu/amdgpu_gtt_mgr.c | 6 ++++--
 5 files changed, 22 insertions(+), 4 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +2 -0
index 4a2b33d..ef8e6b9 100644
@@ -76,6 +76,7 @@
 extern int amdgpu_modeset;
 extern int amdgpu_vram_limit;
 extern int amdgpu_gart_size;
+extern unsigned amdgpu_gart_sys_limit;
 extern int amdgpu_moverate;
 extern int amdgpu_benchmarking;
 extern int amdgpu_testing;
@@ -605,6 +606,7 @@ struct amdgpu_mc {
        u64                     mc_vram_size;
        u64                     visible_vram_size;
        u64                     gtt_size;
+       u64                     gtt_sys_limit;
        u64                     gtt_start;
        u64                     gtt_end;
        u64                     vram_start;
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +6 -0
index 5b1220f..7e3f8cb 100644
@@ -1122,6 +1122,12 @@ static void amdgpu_check_arguments(struct amdgpu_device *adev)
                }
        }
 
+       if (amdgpu_gart_sys_limit < 32) {
+               dev_warn(adev->dev, "gart sys limit (%d) too small\n",
+                                amdgpu_gart_sys_limit);
+                       amdgpu_gart_sys_limit = 32;
+       }
+
        amdgpu_check_vm_size(adev);
 
        amdgpu_check_block_size(adev);
drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +4 -0
index 5a1d794..907ae5e 100644
@@ -75,6 +75,7 @@
 
 int amdgpu_vram_limit = 0;
 int amdgpu_gart_size = -1; /* auto */
+unsigned amdgpu_gart_sys_limit = 256;
 int amdgpu_moverate = -1; /* auto */
 int amdgpu_benchmarking = 0;
 int amdgpu_testing = 0;
@@ -124,6 +125,9 @@ module_param_named(vramlimit, amdgpu_vram_limit, int, 0600);
 MODULE_PARM_DESC(gartsize, "Size of PCIE/IGP gart to setup in megabytes (32, 64, etc., -1 = auto)");
 module_param_named(gartsize, amdgpu_gart_size, int, 0600);
 
+MODULE_PARM_DESC(gartlimit, "GART limit for the system domain in megabytes (default 256)");
+module_param_named(gartlimit, amdgpu_gart_sys_limit, int, 0600);
+
 MODULE_PARM_DESC(moverate, "Maximum buffer migration rate in MB/s. (32, 64, etc., -1=auto, 0=1=disabled)");
 module_param_named(moverate, amdgpu_moverate, int, 0600);
 
drivers/gpu/drm/amd/amdgpu/amdgpu_gart.c +6 -2
index d99b2b2..f82eeaa 100644
@@ -70,6 +70,9 @@ void amdgpu_gart_set_defaults(struct amdgpu_device *adev)
                                        adev->mc.mc_vram_size);
        else
                adev->mc.gtt_size = (uint64_t)amdgpu_gart_size << 20;
+
+       adev->mc.gtt_sys_limit = min((uint64_t)amdgpu_gart_sys_limit << 20,
+                                    adev->mc.gtt_size);
 }
 
 /**
@@ -384,8 +387,9 @@ int amdgpu_gart_init(struct amdgpu_device *adev)
        if (r)
                return r;
        /* Compute table size */
-       adev->gart.num_cpu_pages = adev->mc.gtt_size / PAGE_SIZE;
-       adev->gart.num_gpu_pages = adev->mc.gtt_size / AMDGPU_GPU_PAGE_SIZE;
+       adev->gart.num_cpu_pages = adev->mc.gtt_sys_limit / PAGE_SIZE;
+       adev->gart.num_gpu_pages = adev->mc.gtt_sys_limit /
+               AMDGPU_GPU_PAGE_SIZE;
        DRM_INFO("GART: num cpu pages %u, num gpu pages %u\n",
                 adev->gart.num_cpu_pages, adev->gart.num_gpu_pages);
 
drivers/gpu/drm/amd/amdgpu/amdgpu_gtt_mgr.c +4 -3
index a0976dc..9b516c5 100644
@@ -42,6 +42,7 @@ struct amdgpu_gtt_mgr {
 static int amdgpu_gtt_mgr_init(struct ttm_mem_type_manager *man,
                               unsigned long p_size)
 {
+       struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev);
        struct amdgpu_gtt_mgr *mgr;
        uint64_t start, size;
 
@@ -50,7 +51,7 @@ static int amdgpu_gtt_mgr_init(struct ttm_mem_type_manager *man,
                return -ENOMEM;
 
        start = AMDGPU_GTT_MAX_TRANSFER_SIZE * AMDGPU_GTT_NUM_TRANSFER_WINDOWS;
-       size = p_size - start;
+       size = (adev->mc.gtt_sys_limit >> PAGE_SHIFT) - start;
        drm_mm_init(&mgr->mm, start, size);
        spin_lock_init(&mgr->lock);
        mgr->available = p_size;
@@ -112,6 +113,7 @@ int amdgpu_gtt_mgr_alloc(struct ttm_mem_type_manager *man,
                         const struct ttm_place *place,
                         struct ttm_mem_reg *mem)
 {
+       struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev);
        struct amdgpu_gtt_mgr *mgr = man->priv;
        struct drm_mm_node *node = mem->mm_node;
        enum drm_mm_insert_mode mode;
@@ -129,7 +131,7 @@ int amdgpu_gtt_mgr_alloc(struct ttm_mem_type_manager *man,
        if (place && place->lpfn)
                lpfn = place->lpfn;
        else
-               lpfn = man->size;
+               lpfn = adev->gart.num_cpu_pages;
 
        mode = DRM_MM_INSERT_BEST;
        if (place && place->flags & TTM_PL_FLAG_TOPDOWN)
--

2.7.4

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

Reviewed-by
#37 lore raw
In reply to a message not in the archive
Quoted 5 lines · Christian König

From: Christian König <[email protected]>

Keep them where they belong.

Quoted 114 lines
---
 drivers/gpu/drm/amd/amdgpu/amdgpu.h      | 44 --------------------------------
 drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h | 42 ++++++++++++++++++++++++++++++
 2 files changed, 42 insertions(+), 44 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +0 -44
index ab1dad2..810796a 100644
@@ -1801,50 +1801,6 @@ bool amdgpu_device_has_dc_support(struct amdgpu_device *adev);
 #define RBIOS16(i) (RBIOS8(i) | (RBIOS8((i)+1) << 8))
 #define RBIOS32(i) ((RBIOS16(i)) | (RBIOS16((i)+2) << 16))
 
-/*
- * RING helpers.
- */
-static inline void amdgpu_ring_write(struct amdgpu_ring *ring, uint32_t v)
-{
-       if (ring->count_dw <= 0)
-               DRM_ERROR("amdgpu: writing more dwords to the ring than expected!\n");
-       ring->ring[ring->wptr++ & ring->buf_mask] = v;
-       ring->wptr &= ring->ptr_mask;
-       ring->count_dw--;
-}
-
-static inline void amdgpu_ring_write_multiple(struct amdgpu_ring *ring, void *src, int count_dw)
-{
-       unsigned occupied, chunk1, chunk2;
-       void *dst;
-
-       if (unlikely(ring->count_dw < count_dw)) {
-               DRM_ERROR("amdgpu: writing more dwords to the ring than expected!\n");
-               return;
-       }
-
-       occupied = ring->wptr & ring->buf_mask;
-       dst = (void *)&ring->ring[occupied];
-       chunk1 = ring->buf_mask + 1 - occupied;
-       chunk1 = (chunk1 >= count_dw) ? count_dw: chunk1;
-       chunk2 = count_dw - chunk1;
-       chunk1 <<= 2;
-       chunk2 <<= 2;
-
-       if (chunk1)
-               memcpy(dst, src, chunk1);
-
-       if (chunk2) {
-               src += chunk1;
-               dst = (void *)ring->ring;
-               memcpy(dst, src, chunk2);
-       }
-
-       ring->wptr += count_dw;
-       ring->wptr &= ring->ptr_mask;
-       ring->count_dw -= count_dw;
-}
-
 static inline struct amdgpu_sdma_instance *
 amdgpu_get_sdma_instance(struct amdgpu_ring *ring)
 {
drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h +42 -1
index bc8dec9..04cbc3a 100644
@@ -212,4 +212,46 @@ static inline void amdgpu_ring_clear_ring(struct amdgpu_ring *ring)
 
 }
 
+static inline void amdgpu_ring_write(struct amdgpu_ring *ring, uint32_t v)
+{
+       if (ring->count_dw <= 0)
+               DRM_ERROR("amdgpu: writing more dwords to the ring than expected!\n");
+       ring->ring[ring->wptr++ & ring->buf_mask] = v;
+       ring->wptr &= ring->ptr_mask;
+       ring->count_dw--;
+}
+
+static inline void amdgpu_ring_write_multiple(struct amdgpu_ring *ring,
+                                             void *src, int count_dw)
+{
+       unsigned occupied, chunk1, chunk2;
+       void *dst;
+
+       if (unlikely(ring->count_dw < count_dw)) {
+               DRM_ERROR("amdgpu: writing more dwords to the ring than expected!\n");
+               return;
+       }
+
+       occupied = ring->wptr & ring->buf_mask;
+       dst = (void *)&ring->ring[occupied];
+       chunk1 = ring->buf_mask + 1 - occupied;
+       chunk1 = (chunk1 >= count_dw) ? count_dw: chunk1;
+       chunk2 = count_dw - chunk1;
+       chunk1 <<= 2;
+       chunk2 <<= 2;
+
+       if (chunk1)
+               memcpy(dst, src, chunk1);
+
+       if (chunk2) {
+               src += chunk1;
+               dst = (void *)ring->ring;
+               memcpy(dst, src, chunk2);
+       }
+
+       ring->wptr += count_dw;
+       ring->wptr &= ring->ptr_mask;
+       ring->count_dw -= count_dw;
+}
+
 #endif
--

2.7.4

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

#38 lore raw
In reply to a message not in the archive

Thank you Christian! The series is Acked-by: Felix Kuehling <[email protected]>

Minor nit-pick in patch 6: I spotted 4-space indentation in amdgpu_ttm_backend_bind.

I'm looking at patches 5 and 9 more closely, because I'll need to make similar changes to the KFD IPC copy code.

Regards,
Felix

Quoted 106 lines · Christian König

From: Christian König <[email protected]>

Keep them where they belong.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu.h      | 44 --------------------------------
 drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h | 42 ++++++++++++++++++++++++++++++
 2 files changed, 42 insertions(+), 44 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +0 -44
index ab1dad2..810796a 100644
@@ -1801,50 +1801,6 @@ bool amdgpu_device_has_dc_support(struct amdgpu_device *adev);
 #define RBIOS16(i) (RBIOS8(i) | (RBIOS8((i)+1) << 8))
 #define RBIOS32(i) ((RBIOS16(i)) | (RBIOS16((i)+2) << 16))
 
-/*
- * RING helpers.
- */
-static inline void amdgpu_ring_write(struct amdgpu_ring *ring, uint32_t v)
-{
-	if (ring->count_dw <= 0)
-		DRM_ERROR("amdgpu: writing more dwords to the ring than expected!\n");
-	ring->ring[ring->wptr++ & ring->buf_mask] = v;
-	ring->wptr &= ring->ptr_mask;
-	ring->count_dw--;
-}
-
-static inline void amdgpu_ring_write_multiple(struct amdgpu_ring *ring, void *src, int count_dw)
-{
-	unsigned occupied, chunk1, chunk2;
-	void *dst;
-
-	if (unlikely(ring->count_dw < count_dw)) {
-		DRM_ERROR("amdgpu: writing more dwords to the ring than expected!\n");
-		return;
-	}
-
-	occupied = ring->wptr & ring->buf_mask;
-	dst = (void *)&ring->ring[occupied];
-	chunk1 = ring->buf_mask + 1 - occupied;
-	chunk1 = (chunk1 >= count_dw) ? count_dw: chunk1;
-	chunk2 = count_dw - chunk1;
-	chunk1 <<= 2;
-	chunk2 <<= 2;
-
-	if (chunk1)
-		memcpy(dst, src, chunk1);
-
-	if (chunk2) {
-		src += chunk1;
-		dst = (void *)ring->ring;
-		memcpy(dst, src, chunk2);
-	}
-
-	ring->wptr += count_dw;
-	ring->wptr &= ring->ptr_mask;
-	ring->count_dw -= count_dw;
-}
-
 static inline struct amdgpu_sdma_instance *
 amdgpu_get_sdma_instance(struct amdgpu_ring *ring)
 {
drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h +33 -0
index bc8dec9..04cbc3a 100644
@@ -212,4 +212,46 @@ static inline void amdgpu_ring_clear_ring(struct amdgpu_ring *ring)
 
 }
 
+static inline void amdgpu_ring_write(struct amdgpu_ring *ring, uint32_t v)
+{
+	if (ring->count_dw <= 0)
+		DRM_ERROR("amdgpu: writing more dwords to the ring than expected!\n");
+	ring->ring[ring->wptr++ & ring->buf_mask] = v;
+	ring->wptr &= ring->ptr_mask;
+	ring->count_dw--;
+}
+
+static inline void amdgpu_ring_write_multiple(struct amdgpu_ring *ring,
+					      void *src, int count_dw)
+{
+	unsigned occupied, chunk1, chunk2;
+	void *dst;
+
+	if (unlikely(ring->count_dw < count_dw)) {
+		DRM_ERROR("amdgpu: writing more dwords to the ring than expected!\n");
+		return;
+	}
+
+	occupied = ring->wptr & ring->buf_mask;
+	dst = (void *)&ring->ring[occupied];
+	chunk1 = ring->buf_mask + 1 - occupied;
+	chunk1 = (chunk1 >= count_dw) ? count_dw: chunk1;
+	chunk2 = count_dw - chunk1;
+	chunk1 <<= 2;
+	chunk2 <<= 2;
+
+	if (chunk1)
+		memcpy(dst, src, chunk1);
+
+	if (chunk2) {
+		src += chunk1;

drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h static inline void amdgpu_ring_clear_ring(struct amdgpu_ring *ring) · line 257

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h +9 -0
+		dst = (void *)ring->ring;
+		memcpy(dst, src, chunk2);
+	}
+
+	ring->wptr += count_dw;
+	ring->wptr &= ring->ptr_mask;
+	ring->count_dw -= count_dw;
+}
+
 #endif

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