Skip to contentSkip to conversation

drm/amdgpu: fix lru size grouping v2

PATCH In review Christian König 2 participants 7 messages Aug 18, 2016 2 messages not in the archive

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

PATCH 1/5 drm/amdgpu: fix lru size grouping v2

From: Christian König <[email protected]>

Adding a BO can make it the insertion point for larger sizes as well.

v2: add a comment about the guard structure.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu.h     | 2 ++
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 8 ++++++++
 2 files changed, 10 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +2 -0
index a55eb6f..b230dd9 100644
@@ -428,6 +428,8 @@ struct amdgpu_mman {
 
 	/* custom LRU management */
 	struct amdgpu_mman_lru			log2_size[AMDGPU_TTM_LRU_SIZE];
+	/* guard for log2_size array, don't add anything in between */
+	struct amdgpu_mman_lru			guard;
 };
 
 int amdgpu_copy_buffer(struct amdgpu_ring *ring,
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +8 -0
index f5dc87e..1a12ce0 100644
@@ -951,6 +951,8 @@ static struct list_head *amdgpu_ttm_lru_tail(struct ttm_buffer_object *tbo)
 	struct list_head *res = lru->lru[tbo->mem.mem_type];
 
 	lru->lru[tbo->mem.mem_type] = &tbo->lru;
+	while ((++lru)->lru[tbo->mem.mem_type] == res)
+		lru->lru[tbo->mem.mem_type] = &tbo->lru;
 
 	return res;
 }
@@ -961,6 +963,8 @@ static struct list_head *amdgpu_ttm_swap_lru_tail(struct ttm_buffer_object *tbo)
 	struct list_head *res = lru->swap_lru;
 
 	lru->swap_lru = &tbo->swap;
+	while ((++lru)->swap_lru == res)
+		lru->swap_lru = &tbo->swap;
 
 	return res;
 }
@@ -1008,6 +1012,10 @@ int amdgpu_ttm_init(struct amdgpu_device *adev)
 		lru->swap_lru = &adev->mman.bdev.glob->swap_lru;
 	}
 
+	for (j = 0; j < TTM_NUM_MEM_TYPES; ++j)
+		adev->mman.guard.lru[j] = NULL;
+	adev->mman.guard.swap_lru = NULL;
+
 	adev->mman.initialized = true;
 	r = ttm_bo_init_mm(&adev->mman.bdev, TTM_PL_VRAM,
 				adev->mc.real_vram_size >> PAGE_SHIFT);
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): 1471514358-2131-1-git-send-email-deathsimple-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

#2 lore raw

PATCH 2/5 drm/amdgpu: cleanup amdgpu_ttm_placement_init

From: Christian König <[email protected]>

Make it more clear what this function does. No intendet functional change.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_object.c | 112 +++++++++++++++--------------
 1 file changed, 60 insertions(+), 52 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_object.c +60 -52
index 4d084ee..e7a6a2e 100644
@@ -111,87 +111,95 @@ bool amdgpu_ttm_bo_is_amdgpu_bo(struct ttm_buffer_object *bo)
 
 static void amdgpu_ttm_placement_init(struct amdgpu_device *adev,
 				      struct ttm_placement *placement,
-				      struct ttm_place *placements,
+				      struct ttm_place *places,
 				      u32 domain, u64 flags)
 {
 	u32 c = 0, i;
 
-	placement->placement = placements;
-	placement->busy_placement = placements;
-
 	if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
+		unsigned visible_pfn = adev->mc.visible_vram_size >> PAGE_SHIFT;
+
 		if (flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS &&
-			adev->mc.visible_vram_size < adev->mc.real_vram_size) {
-			placements[c].fpfn =
-				adev->mc.visible_vram_size >> PAGE_SHIFT;
-			placements[c++].flags = TTM_PL_FLAG_WC |
+		    adev->mc.visible_vram_size < adev->mc.real_vram_size) {
+			places[c].fpfn = visible_pfn;
+			if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
+				places[c].lpfn = visible_pfn;
+			else
+				places[c].lpfn = 0;
+			places[c].flags = TTM_PL_FLAG_WC |
 				TTM_PL_FLAG_UNCACHED | TTM_PL_FLAG_VRAM |
 				TTM_PL_FLAG_TOPDOWN;
+			c++;
 		}
-		placements[c].fpfn = 0;
-		placements[c++].flags = TTM_PL_FLAG_WC | TTM_PL_FLAG_UNCACHED |
+
+		places[c].fpfn = 0;
+		places[c].lpfn = 0;
+		places[c].flags = TTM_PL_FLAG_WC | TTM_PL_FLAG_UNCACHED |
 			TTM_PL_FLAG_VRAM;
-		if (!(flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED))
-			placements[c - 1].flags |= TTM_PL_FLAG_TOPDOWN;
+		if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
+			places[c].lpfn = visible_pfn;
+		else
+			places[c].flags |= TTM_PL_FLAG_TOPDOWN;
+		c++;
 	}
 
 	if (domain & AMDGPU_GEM_DOMAIN_GTT) {
-		if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC) {
-			placements[c].fpfn = 0;
-			placements[c++].flags = TTM_PL_FLAG_WC |
-				TTM_PL_FLAG_TT | TTM_PL_FLAG_UNCACHED;
-		} else {
-			placements[c].fpfn = 0;
-			placements[c++].flags = TTM_PL_FLAG_CACHED |
-				TTM_PL_FLAG_TT;
-		}
+		places[c].fpfn = 0;
+		places[c].lpfn = 0;
+		places[c].flags = TTM_PL_FLAG_TT;
+		if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
+			places[c].flags |= TTM_PL_FLAG_WC |
+				TTM_PL_FLAG_UNCACHED;
+		else
+			places[c].flags |= TTM_PL_FLAG_CACHED;
+		c++;
 	}
 
 	if (domain & AMDGPU_GEM_DOMAIN_CPU) {
-		if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC) {
-			placements[c].fpfn = 0;
-			placements[c++].flags = TTM_PL_FLAG_WC |
-				TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_UNCACHED;
-		} else {
-			placements[c].fpfn = 0;
-			placements[c++].flags = TTM_PL_FLAG_CACHED |
-				TTM_PL_FLAG_SYSTEM;
-		}
+		places[c].fpfn = 0;
+		places[c].lpfn = 0;
+		places[c].flags = TTM_PL_FLAG_SYSTEM;
+		if (flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC)
+			places[c].flags |= TTM_PL_FLAG_WC |
+				TTM_PL_FLAG_UNCACHED;
+		else
+			places[c].flags |= TTM_PL_FLAG_CACHED;
+		c++;
 	}
 
 	if (domain & AMDGPU_GEM_DOMAIN_GDS) {
-		placements[c].fpfn = 0;
-		placements[c++].flags = TTM_PL_FLAG_UNCACHED |
-			AMDGPU_PL_FLAG_GDS;
+		places[c].fpfn = 0;
+		places[c].lpfn = 0;
+		places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_GDS;
+		c++;
 	}
+
 	if (domain & AMDGPU_GEM_DOMAIN_GWS) {
-		placements[c].fpfn = 0;
-		placements[c++].flags = TTM_PL_FLAG_UNCACHED |
-			AMDGPU_PL_FLAG_GWS;
+		places[c].fpfn = 0;
+		places[c].lpfn = 0;
+		places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_GWS;
+		c++;
 	}
+
 	if (domain & AMDGPU_GEM_DOMAIN_OA) {
-		placements[c].fpfn = 0;
-		placements[c++].flags = TTM_PL_FLAG_UNCACHED |
-			AMDGPU_PL_FLAG_OA;
+		places[c].fpfn = 0;
+		places[c].lpfn = 0;
+		places[c].flags = TTM_PL_FLAG_UNCACHED | AMDGPU_PL_FLAG_OA;
+		c++;
 	}
 
 	if (!c) {
-		placements[c].fpfn = 0;
-		placements[c++].flags = TTM_PL_MASK_CACHING |
-			TTM_PL_FLAG_SYSTEM;
+		places[c].fpfn = 0;
+		places[c].lpfn = 0;
+		places[c].flags = TTM_PL_MASK_CACHING | TTM_PL_FLAG_SYSTEM;
+		c++;
 	}
+
 	placement->num_placement = c;
-	placement->num_busy_placement = c;
+	placement->placement = places;
 
-	for (i = 0; i < c; i++) {
-		if ((flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED) &&
-			(placements[i].flags & TTM_PL_FLAG_VRAM) &&
-			!placements[i].fpfn)
-			placements[i].lpfn =
-				adev->mc.visible_vram_size >> PAGE_SHIFT;
-		else
-			placements[i].lpfn = 0;
-	}
+	placement->num_busy_placement = c;
+	placement->busy_placement = places;
 }
 
 void amdgpu_ttm_placement_from_domain(struct amdgpu_bo *rbo, u32 domain)
Signature · 6 lines
2.5.0

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

PATCH 3/5 drm/amdgpu: stop adding dummy entry in amdgpu_ttm_placement_init

In reply to a message not in the archive

From: Christian König <[email protected]>

AMDGPU_GEM_CREATE_NO_CPU_ACCESS and AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED are obviously mutual exclusive. So stop adding a dummy entry without effect when both are specified.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_object.c | 8 +++-----
 1 file changed, 3 insertions(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_object.c +3 -5
index e7a6a2e..b8567d2 100644
@@ -114,18 +114,16 @@ static void amdgpu_ttm_placement_init(struct amdgpu_device *adev,
 				      struct ttm_place *places,
 				      u32 domain, u64 flags)
 {
-	u32 c = 0, i;
+	u32 c = 0;
 
 	if (domain & AMDGPU_GEM_DOMAIN_VRAM) {
 		unsigned visible_pfn = adev->mc.visible_vram_size >> PAGE_SHIFT;
 
 		if (flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS &&
+		    !(flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED) &&
 		    adev->mc.visible_vram_size < adev->mc.real_vram_size) {
 			places[c].fpfn = visible_pfn;
-			if (flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
-				places[c].lpfn = visible_pfn;
-			else
-				places[c].lpfn = 0;
+			places[c].lpfn = 0;
 			places[c].flags = TTM_PL_FLAG_WC |
 				TTM_PL_FLAG_UNCACHED | TTM_PL_FLAG_VRAM |
 				TTM_PL_FLAG_TOPDOWN;
Signature · 6 lines
2.5.0

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

PATCH 4/5 drm/amdgpu: cleanup amdgpu_vm_bo_update params

In reply to a message not in the archive

From: Christian König <[email protected]>

Make it more obvious what we are doing here.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu.h    |  2 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c |  2 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 21 +++++++++++----------
 3 files changed, 13 insertions(+), 12 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +1 -1
index b230dd9..ffad8d9 100644
@@ -977,7 +977,7 @@ int amdgpu_vm_clear_invalids(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 			     struct amdgpu_sync *sync);
 int amdgpu_vm_bo_update(struct amdgpu_device *adev,
 			struct amdgpu_bo_va *bo_va,
-			struct ttm_mem_reg *mem);
+			bool clear);
 void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
 			     struct amdgpu_bo *bo);
 struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +1 -1
index 3434098..d5d61a7 100644
@@ -643,7 +643,7 @@ static int amdgpu_bo_vm_update_pte(struct amdgpu_cs_parser *p,
 			if (bo_va == NULL)
 				continue;
 
-			r = amdgpu_vm_bo_update(adev, bo_va, &bo->tbo.mem);
+			r = amdgpu_vm_bo_update(adev, bo_va, false);
 			if (r)
 				return r;
 
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +11 -10
index 45b2f20..0b92d36 100644
@@ -1080,28 +1080,32 @@ static int amdgpu_vm_bo_split_mapping(struct amdgpu_device *adev,
  *
  * @adev: amdgpu_device pointer
  * @bo_va: requested BO and VM object
- * @mem: ttm mem
+ * @clear: if true clear the entries
  *
  * Fill in the page table entries for @bo_va.
  * Returns 0 for success, -EINVAL for failure.
- *
- * Object have to be reserved and mutex must be locked!
  */
 int amdgpu_vm_bo_update(struct amdgpu_device *adev,
 			struct amdgpu_bo_va *bo_va,
-			struct ttm_mem_reg *mem)
+			bool clear)
 {
 	struct amdgpu_vm *vm = bo_va->vm;
 	struct amdgpu_bo_va_mapping *mapping;
 	dma_addr_t *pages_addr = NULL;
 	uint32_t gtt_flags, flags;
+	struct ttm_mem_reg *mem;
 	struct fence *exclusive;
 	uint64_t addr;
 	int r;
 
-	if (mem) {
+	if (clear) {
+		mem = NULL;
+		addr = 0;
+		exclusive = NULL;
+	} else {
 		struct ttm_dma_tt *ttm;
 
+		mem = &bo_va->bo->tbo.mem;
 		addr = (u64)mem->start << PAGE_SHIFT;
 		switch (mem->mem_type) {
 		case TTM_PL_TT:
@@ -1119,9 +1123,6 @@ int amdgpu_vm_bo_update(struct amdgpu_device *adev,
 		}
 
 		exclusive = reservation_object_get_excl(bo_va->bo->tbo.resv);
-	} else {
-		addr = 0;
-		exclusive = NULL;
 	}
 
 	flags = amdgpu_ttm_tt_pte_flags(adev, bo_va->bo->tbo.ttm, mem);
@@ -1152,7 +1153,7 @@ int amdgpu_vm_bo_update(struct amdgpu_device *adev,
 	spin_lock(&vm->status_lock);
 	list_splice_init(&bo_va->invalids, &bo_va->valids);
 	list_del_init(&bo_va->vm_status);
-	if (!mem)
+	if (clear)
 		list_add(&bo_va->vm_status, &vm->cleared);
 	spin_unlock(&vm->status_lock);
 
@@ -1215,7 +1216,7 @@ int amdgpu_vm_clear_invalids(struct amdgpu_device *adev,
 			struct amdgpu_bo_va, vm_status);
 		spin_unlock(&vm->status_lock);
 
-		r = amdgpu_vm_bo_update(adev, bo_va, NULL);
+		r = amdgpu_vm_bo_update(adev, bo_va, true);
 		if (r)
 			return r;
 
Signature · 6 lines
2.5.0

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

PATCH 5/5 drm/amdgpu: add AMDGPU_GEM_CREATE_LINEAR flag

In reply to a message not in the archive

From: Christian König <[email protected]>

Add a flag noting that a BO must be created using linear VRAM and set this flag on all in kernel users where appropriate.

Hopefully I haven't missed anything.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu.h        |  1 +
 drivers/gpu/drm/amd/amdgpu/amdgpu_cgs.c    |  6 ++++--
 drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c     | 31 ++++++++++++++++++++++++++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_gart.c   |  3 ++-
 drivers/gpu/drm/amd/amdgpu/amdgpu_object.c | 13 +++++++++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c    | 10 ++++++++--
 drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c    |  6 +++++-
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c     |  6 ++++--
 include/uapi/drm/amdgpu_drm.h              |  2 ++
 9 files changed, 70 insertions(+), 8 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +1 -0
index ffad8d9..2f5a70a 100644
@@ -2533,6 +2533,7 @@ static inline void amdgpu_acpi_fini(struct amdgpu_device *adev) { }
 struct amdgpu_bo_va_mapping *
 amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
 		       uint64_t addr, struct amdgpu_bo **bo);
+int amdgpu_cs_make_bos_linear(struct amdgpu_cs_parser *parser);
 
 #if defined(CONFIG_DRM_AMD_DAL)
 int amdgpu_dm_display_resume(struct amdgpu_device *adev );
drivers/gpu/drm/amd/amdgpu/amdgpu_cgs.c +4 -2
index bc0440f..9a7092a 100644
@@ -146,7 +146,8 @@ static int amdgpu_cgs_alloc_gpu_mem(struct cgs_device *cgs_device,
 	switch(type) {
 	case CGS_GPU_MEM_TYPE__VISIBLE_CONTIG_FB:
 	case CGS_GPU_MEM_TYPE__VISIBLE_FB:
-		flags = AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
+		flags = AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
+			AMDGPU_GEM_CREATE_VRAM_LINEAR;
 		domain = AMDGPU_GEM_DOMAIN_VRAM;
 		if (max_offset > adev->mc.real_vram_size)
 			return -EINVAL;
@@ -157,7 +158,8 @@ static int amdgpu_cgs_alloc_gpu_mem(struct cgs_device *cgs_device,
 		break;
 	case CGS_GPU_MEM_TYPE__INVISIBLE_CONTIG_FB:
 	case CGS_GPU_MEM_TYPE__INVISIBLE_FB:
-		flags = AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
+		flags = AMDGPU_GEM_CREATE_NO_CPU_ACCESS |
+			AMDGPU_GEM_CREATE_VRAM_LINEAR;
 		domain = AMDGPU_GEM_DOMAIN_VRAM;
 		if (adev->mc.visible_vram_size < adev->mc.real_vram_size) {
 			place.fpfn =
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +31 -0
index d5d61a7..20d08ae 100644
@@ -1041,3 +1041,34 @@ amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
 
 	return NULL;
 }
+
+/**
+ * amdgpu_cs_make_bos_linear - mark all BOs in this CS as linear
+ *
+ * @parser: command submission parser context
+ *
+ * Helper for UVD/VCE VM emulation, mark all BOs in this CS as linear.
+ */
+int amdgpu_cs_make_bos_linear(struct amdgpu_cs_parser *parser)
+{
+	unsigned i;
+	int r;
+
+	if (!parser->bo_list)
+		return 0;
+
+	for (i = 0; i < parser->bo_list->num_entries; i++) {
+		struct amdgpu_bo *bo = parser->bo_list->array[i].robj;
+
+		if (bo->flags |= AMDGPU_GEM_CREATE_VRAM_LINEAR)
+			continue;
+
+		bo->flags |= AMDGPU_GEM_CREATE_VRAM_LINEAR;
+		amdgpu_ttm_placement_from_domain(bo, bo->allowed_domains);
+		r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
+		if (unlikely(r))
+			return r;
+	}
+
+	return 0;
+}
drivers/gpu/drm/amd/amdgpu/amdgpu_gart.c +2 -1
index 0feea34..b5982a9 100644
@@ -126,7 +126,8 @@ int amdgpu_gart_table_vram_alloc(struct amdgpu_device *adev)
 	if (adev->gart.robj == NULL) {
 		r = amdgpu_bo_create(adev, adev->gart.table_size,
 				     PAGE_SIZE, true, AMDGPU_GEM_DOMAIN_VRAM,
-				     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
+				     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
+				     AMDGPU_GEM_CREATE_VRAM_LINEAR,
 				     NULL, NULL, &adev->gart.robj);
 		if (r) {
 			return r;
drivers/gpu/drm/amd/amdgpu/amdgpu_object.c +13 -0
index b8567d2..3945695 100644
@@ -542,6 +542,8 @@ int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain,
 
 		return 0;
 	}
+
+	bo->flags |= AMDGPU_GEM_CREATE_VRAM_LINEAR;
 	amdgpu_ttm_placement_from_domain(bo, domain);
 	for (i = 0; i < bo->placement.num_placement; i++) {
 		/* force to pin into visible video ram */
@@ -777,6 +779,15 @@ int amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
 	if (bo->mem.mem_type != TTM_PL_VRAM)
 		return 0;
 
+	/* TODO: figure out how to map scattered VRAM to the CPU */
+	if (!(abo->flags & AMDGPU_GEM_CREATE_VRAM_LINEAR)) {
+		abo->flags |= AMDGPU_GEM_CREATE_VRAM_LINEAR;
+		amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM);
+		r = ttm_bo_validate(bo, &abo->placement, false, false);
+		if (unlikely(r))
+			return r;
+	}
+
 	size = bo->mem.num_pages << PAGE_SHIFT;
 	offset = bo->mem.start << PAGE_SHIFT;
 	if ((offset + size) <= adev->mc.visible_vram_size)
@@ -845,6 +856,8 @@ u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
 	WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_SYSTEM);
 	WARN_ON_ONCE(!ww_mutex_is_locked(&bo->tbo.resv->lock) &&
 		     !bo->pin_count);
+	WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_VRAM &&
+		     !(bo->flags & AMDGPU_GEM_CREATE_VRAM_LINEAR));
 
 	return bo->tbo.offset;
 }
drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c +8 -2
index 533d702..a425d10 100644
@@ -890,6 +890,10 @@ int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser *parser, uint32_t ib_idx)
 		return -EINVAL;
 	}
 
+	r = amdgpu_cs_make_bos_linear(parser);
+	if (r)
+		return r;
+
 	ctx.parser = parser;
 	ctx.buf_sizes = buf_sizes;
 	ctx.ib_idx = ib_idx;
@@ -1004,7 +1008,8 @@ int amdgpu_uvd_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
 
 	r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true,
 			     AMDGPU_GEM_DOMAIN_VRAM,
-			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
+			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
+			     AMDGPU_GEM_CREATE_VRAM_LINEAR,
 			     NULL, NULL, &bo);
 	if (r)
 		return r;
@@ -1053,7 +1058,8 @@ int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
 
 	r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true,
 			     AMDGPU_GEM_DOMAIN_VRAM,
-			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
+			     AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
+			     AMDGPU_GEM_CREATE_VRAM_LINEAR,
 			     NULL, NULL, &bo);
 	if (r)
 		return r;
drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c +5 -1
index 05865ce..9aa4458 100644
@@ -634,7 +634,11 @@ int amdgpu_vce_ring_parse_cs(struct amdgpu_cs_parser *p, uint32_t ib_idx)
 	uint32_t allocated = 0;
 	uint32_t tmp, handle = 0;
 	uint32_t *size = &tmp;
-	int i, r = 0, idx = 0;
+	int i, r, idx = 0;
+
+	r = amdgpu_cs_make_bos_linear(p);
+	if (r)
+		return r;
 
 	while (idx < ib->length_dw) {
 		uint32_t len = amdgpu_get_ib_value(p, ib_idx, idx);
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +4 -2
index 0b92d36..cafbadb 100644
@@ -1362,7 +1362,8 @@ int amdgpu_vm_bo_map(struct amdgpu_device *adev,
 				     AMDGPU_GPU_PAGE_SIZE, true,
 				     AMDGPU_GEM_DOMAIN_VRAM,
 				     AMDGPU_GEM_CREATE_NO_CPU_ACCESS |
-				     AMDGPU_GEM_CREATE_SHADOW,
+				     AMDGPU_GEM_CREATE_SHADOW |
+				     AMDGPU_GEM_CREATE_VRAM_LINEAR,
 				     NULL, resv, &pt);
 		if (r)
 			goto error_free;
@@ -1562,7 +1563,8 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm)
 	r = amdgpu_bo_create(adev, pd_size, align, true,
 			     AMDGPU_GEM_DOMAIN_VRAM,
 			     AMDGPU_GEM_CREATE_NO_CPU_ACCESS |
-			     AMDGPU_GEM_CREATE_SHADOW,
+			     AMDGPU_GEM_CREATE_SHADOW |
+			     AMDGPU_GEM_CREATE_VRAM_LINEAR,
 			     NULL, NULL, &vm->page_directory);
 	if (r)
 		goto error_free_sched_entity;
include/uapi/drm/amdgpu_drm.h +2 -0
index da2d3e1..1420c94 100644
@@ -79,6 +79,8 @@
 #define AMDGPU_GEM_CREATE_VRAM_CLEARED		(1 << 3)
 /* Flag that create shadow bo(GTT) while allocating vram bo */
 #define AMDGPU_GEM_CREATE_SHADOW		(1 << 4)
+/* Flag that allocating the BO should use linear VRAM */
+#define AMDGPU_GEM_CREATE_VRAM_LINEAR		(1 << 5)
 
 struct drm_amdgpu_gem_create_in  {
 	/** the requested memory size */
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): 1471514358-2131-5-git-send-email-deathsimple-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

#7 lore raw

Re: [PATCH 5/5] drm/amdgpu: add AMDGPU_GEM_CREATE_LINEAR flag

Quoted 8 lines · Christian König

From: Christian König <[email protected]>

Add a flag noting that a BO must be created using linear VRAM and set this flag on all in kernel users where appropriate.

Hopefully I haven't missed anything.

How about instead of LINEAR we use CONTIGUOUS? linear is too easily confused with tiling. What about making contiguous the default and having explicit requests for non-contiguous or scatter gather?

Alex

Quoted 231 lines
---
 drivers/gpu/drm/amd/amdgpu/amdgpu.h        |  1 +
 drivers/gpu/drm/amd/amdgpu/amdgpu_cgs.c    |  6 ++++--
 drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c     | 31 ++++++++++++++++++++++++++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_gart.c   |  3 ++-
 drivers/gpu/drm/amd/amdgpu/amdgpu_object.c | 13 +++++++++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c    | 10 ++++++++--
 drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c    |  6 +++++-
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c     |  6 ++++--
 include/uapi/drm/amdgpu_drm.h              |  2 ++
 9 files changed, 70 insertions(+), 8 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +1 -0
index ffad8d9..2f5a70a 100644
@@ -2533,6 +2533,7 @@ static inline void amdgpu_acpi_fini(struct amdgpu_device *adev) { }
 struct amdgpu_bo_va_mapping *
 amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
                       uint64_t addr, struct amdgpu_bo **bo);
+int amdgpu_cs_make_bos_linear(struct amdgpu_cs_parser *parser);
 
 #if defined(CONFIG_DRM_AMD_DAL)
 int amdgpu_dm_display_resume(struct amdgpu_device *adev );
drivers/gpu/drm/amd/amdgpu/amdgpu_cgs.c +4 -2
index bc0440f..9a7092a 100644
@@ -146,7 +146,8 @@ static int amdgpu_cgs_alloc_gpu_mem(struct cgs_device *cgs_device,
        switch(type) {
        case CGS_GPU_MEM_TYPE__VISIBLE_CONTIG_FB:
        case CGS_GPU_MEM_TYPE__VISIBLE_FB:
-               flags = AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
+               flags = AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
+                       AMDGPU_GEM_CREATE_VRAM_LINEAR;
                domain = AMDGPU_GEM_DOMAIN_VRAM;
                if (max_offset > adev->mc.real_vram_size)
                        return -EINVAL;
@@ -157,7 +158,8 @@ static int amdgpu_cgs_alloc_gpu_mem(struct cgs_device *cgs_device,
                break;
        case CGS_GPU_MEM_TYPE__INVISIBLE_CONTIG_FB:
        case CGS_GPU_MEM_TYPE__INVISIBLE_FB:
-               flags = AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
+               flags = AMDGPU_GEM_CREATE_NO_CPU_ACCESS |
+                       AMDGPU_GEM_CREATE_VRAM_LINEAR;
                domain = AMDGPU_GEM_DOMAIN_VRAM;
                if (adev->mc.visible_vram_size < adev->mc.real_vram_size) {
                        place.fpfn =
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +31 -0
index d5d61a7..20d08ae 100644
@@ -1041,3 +1041,34 @@ amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
 
        return NULL;
 }
+
+/**
+ * amdgpu_cs_make_bos_linear - mark all BOs in this CS as linear
+ *
+ * @parser: command submission parser context
+ *
+ * Helper for UVD/VCE VM emulation, mark all BOs in this CS as linear.
+ */
+int amdgpu_cs_make_bos_linear(struct amdgpu_cs_parser *parser)
+{
+       unsigned i;
+       int r;
+
+       if (!parser->bo_list)
+               return 0;
+
+       for (i = 0; i < parser->bo_list->num_entries; i++) {
+               struct amdgpu_bo *bo = parser->bo_list->array[i].robj;
+
+               if (bo->flags |= AMDGPU_GEM_CREATE_VRAM_LINEAR)
+                       continue;
+
+               bo->flags |= AMDGPU_GEM_CREATE_VRAM_LINEAR;
+               amdgpu_ttm_placement_from_domain(bo, bo->allowed_domains);
+               r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
+               if (unlikely(r))
+                       return r;
+       }
+
+       return 0;
+}
drivers/gpu/drm/amd/amdgpu/amdgpu_gart.c +2 -1
index 0feea34..b5982a9 100644
@@ -126,7 +126,8 @@ int amdgpu_gart_table_vram_alloc(struct amdgpu_device *adev)
        if (adev->gart.robj == NULL) {
                r = amdgpu_bo_create(adev, adev->gart.table_size,
                                     PAGE_SIZE, true, AMDGPU_GEM_DOMAIN_VRAM,
-                                    AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
+                                    AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
+                                    AMDGPU_GEM_CREATE_VRAM_LINEAR,
                                     NULL, NULL, &adev->gart.robj);
                if (r) {
                        return r;
drivers/gpu/drm/amd/amdgpu/amdgpu_object.c +13 -0
index b8567d2..3945695 100644
@@ -542,6 +542,8 @@ int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain,
 
                return 0;
        }
+
+       bo->flags |= AMDGPU_GEM_CREATE_VRAM_LINEAR;
        amdgpu_ttm_placement_from_domain(bo, domain);
        for (i = 0; i < bo->placement.num_placement; i++) {
                /* force to pin into visible video ram */
@@ -777,6 +779,15 @@ int amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
        if (bo->mem.mem_type != TTM_PL_VRAM)
                return 0;
 
+       /* TODO: figure out how to map scattered VRAM to the CPU */
+       if (!(abo->flags & AMDGPU_GEM_CREATE_VRAM_LINEAR)) {
+               abo->flags |= AMDGPU_GEM_CREATE_VRAM_LINEAR;
+               amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM);
+               r = ttm_bo_validate(bo, &abo->placement, false, false);
+               if (unlikely(r))
+                       return r;
+       }
+
        size = bo->mem.num_pages << PAGE_SHIFT;
        offset = bo->mem.start << PAGE_SHIFT;
        if ((offset + size) <= adev->mc.visible_vram_size)
@@ -845,6 +856,8 @@ u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)
        WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_SYSTEM);
        WARN_ON_ONCE(!ww_mutex_is_locked(&bo->tbo.resv->lock) &&
                     !bo->pin_count);
+       WARN_ON_ONCE(bo->tbo.mem.mem_type == TTM_PL_VRAM &&
+                    !(bo->flags & AMDGPU_GEM_CREATE_VRAM_LINEAR));
 
        return bo->tbo.offset;
 }
drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c +8 -2
index 533d702..a425d10 100644
@@ -890,6 +890,10 @@ int amdgpu_uvd_ring_parse_cs(struct amdgpu_cs_parser *parser, uint32_t ib_idx)
                return -EINVAL;
        }
 
+       r = amdgpu_cs_make_bos_linear(parser);
+       if (r)
+               return r;
+
        ctx.parser = parser;
        ctx.buf_sizes = buf_sizes;
        ctx.ib_idx = ib_idx;
@@ -1004,7 +1008,8 @@ int amdgpu_uvd_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
 
        r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true,
                             AMDGPU_GEM_DOMAIN_VRAM,
-                            AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
+                            AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
+                            AMDGPU_GEM_CREATE_VRAM_LINEAR,
                             NULL, NULL, &bo);
        if (r)
                return r;
@@ -1053,7 +1058,8 @@ int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
 
        r = amdgpu_bo_create(adev, 1024, PAGE_SIZE, true,
                             AMDGPU_GEM_DOMAIN_VRAM,
-                            AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
+                            AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
+                            AMDGPU_GEM_CREATE_VRAM_LINEAR,
                             NULL, NULL, &bo);
        if (r)
                return r;
drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c +5 -1
index 05865ce..9aa4458 100644
@@ -634,7 +634,11 @@ int amdgpu_vce_ring_parse_cs(struct amdgpu_cs_parser *p, uint32_t ib_idx)
        uint32_t allocated = 0;
        uint32_t tmp, handle = 0;
        uint32_t *size = &tmp;
-       int i, r = 0, idx = 0;
+       int i, r, idx = 0;
+
+       r = amdgpu_cs_make_bos_linear(p);
+       if (r)
+               return r;
 
        while (idx < ib->length_dw) {
                uint32_t len = amdgpu_get_ib_value(p, ib_idx, idx);
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +4 -2
index 0b92d36..cafbadb 100644
@@ -1362,7 +1362,8 @@ int amdgpu_vm_bo_map(struct amdgpu_device *adev,
                                     AMDGPU_GPU_PAGE_SIZE, true,
                                     AMDGPU_GEM_DOMAIN_VRAM,
                                     AMDGPU_GEM_CREATE_NO_CPU_ACCESS |
-                                    AMDGPU_GEM_CREATE_SHADOW,
+                                    AMDGPU_GEM_CREATE_SHADOW |
+                                    AMDGPU_GEM_CREATE_VRAM_LINEAR,
                                     NULL, resv, &pt);
                if (r)
                        goto error_free;
@@ -1562,7 +1563,8 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm)
        r = amdgpu_bo_create(adev, pd_size, align, true,
                             AMDGPU_GEM_DOMAIN_VRAM,
                             AMDGPU_GEM_CREATE_NO_CPU_ACCESS |
-                            AMDGPU_GEM_CREATE_SHADOW,
+                            AMDGPU_GEM_CREATE_SHADOW |
+                            AMDGPU_GEM_CREATE_VRAM_LINEAR,
                             NULL, NULL, &vm->page_directory);
        if (r)
                goto error_free_sched_entity;
include/uapi/drm/amdgpu_drm.h +2 -1
index da2d3e1..1420c94 100644
@@ -79,6 +79,8 @@
 #define AMDGPU_GEM_CREATE_VRAM_CLEARED         (1 << 3)
 /* Flag that create shadow bo(GTT) while allocating vram bo */
 #define AMDGPU_GEM_CREATE_SHADOW               (1 << 4)
+/* Flag that allocating the BO should use linear VRAM */
+#define AMDGPU_GEM_CREATE_VRAM_LINEAR          (1 << 5)
 
 struct drm_amdgpu_gem_create_in  {
        /** the requested memory size */
--

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

Reviewed-by
#8 lore raw
In reply to a message not in the archive
Quoted 8 lines · Christian König

From: Christian König <[email protected]>

Adding a BO can make it the insertion point for larger sizes as well.

v2: add a comment about the guard structure.

Patches 1-4 are:

Quoted 54 lines
---
 drivers/gpu/drm/amd/amdgpu/amdgpu.h     | 2 ++
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 8 ++++++++
 2 files changed, 10 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +2 -0
index a55eb6f..b230dd9 100644
@@ -428,6 +428,8 @@ struct amdgpu_mman {
 
        /* custom LRU management */
        struct amdgpu_mman_lru                  log2_size[AMDGPU_TTM_LRU_SIZE];
+       /* guard for log2_size array, don't add anything in between */
+       struct amdgpu_mman_lru                  guard;
 };
 
 int amdgpu_copy_buffer(struct amdgpu_ring *ring,
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +8 -1
index f5dc87e..1a12ce0 100644
@@ -951,6 +951,8 @@ static struct list_head *amdgpu_ttm_lru_tail(struct ttm_buffer_object *tbo)
        struct list_head *res = lru->lru[tbo->mem.mem_type];
 
        lru->lru[tbo->mem.mem_type] = &tbo->lru;
+       while ((++lru)->lru[tbo->mem.mem_type] == res)
+               lru->lru[tbo->mem.mem_type] = &tbo->lru;
 
        return res;
 }
@@ -961,6 +963,8 @@ static struct list_head *amdgpu_ttm_swap_lru_tail(struct ttm_buffer_object *tbo)
        struct list_head *res = lru->swap_lru;
 
        lru->swap_lru = &tbo->swap;
+       while ((++lru)->swap_lru == res)
+               lru->swap_lru = &tbo->swap;
 
        return res;
 }
@@ -1008,6 +1012,10 @@ int amdgpu_ttm_init(struct amdgpu_device *adev)
                lru->swap_lru = &adev->mman.bdev.glob->swap_lru;
        }
 
+       for (j = 0; j < TTM_NUM_MEM_TYPES; ++j)
+               adev->mman.guard.lru[j] = NULL;
+       adev->mman.guard.swap_lru = NULL;
+
        adev->mman.initialized = true;
        r = ttm_bo_init_mm(&adev->mman.bdev, TTM_PL_VRAM,
                                adev->mc.real_vram_size >> PAGE_SHIFT);
--

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