Skip to contentSkip to conversation

drm/amdgpu: SHADOW and VRAM_CONTIGUOUS flags shouldn't be used by userspace

PATCH In review Christian König 3 participants 12 messages Aug 10, 2017 4 messages not in the archive

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

PATCH 1/8 drm/amdgpu: SHADOW and VRAM_CONTIGUOUS flags shouldn't be used by userspace

From: Christian König <[email protected]>

The shadow handling isn't implemented completely for userspace BOs and the kernel sets the VRAM_CONTIGUOUS as necessary.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +1 -3
index 81127ff..5ae9941 100644
@@ -225,9 +225,7 @@ int amdgpu_gem_create_ioctl(struct drm_device *dev, void *data,
 	if (args->in.domain_flags & ~(AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
 				      AMDGPU_GEM_CREATE_NO_CPU_ACCESS |
 				      AMDGPU_GEM_CREATE_CPU_GTT_USWC |
-				      AMDGPU_GEM_CREATE_VRAM_CLEARED|
-				      AMDGPU_GEM_CREATE_SHADOW |
-				      AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS))
+				      AMDGPU_GEM_CREATE_VRAM_CLEARED))
 		return -EINVAL;
 
 	/* reject invalid gem domains */
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): 1502376647-4519-1-git-send-email-deathsimple-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

#2 lore raw

PATCH 2/8 drm/amdgpu: cleanup static CSA handling

From: Christian König <[email protected]>

Move the CSA bo_va from the VM to the fpriv structure.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu.h      |  1 +
 drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c   |  3 ++-
 drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c  |  6 +++---
 drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c | 25 ++++++++++++-------------
 drivers/gpu/drm/amd/amdgpu/amdgpu_virt.h |  3 ++-
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h   |  2 --
 6 files changed, 20 insertions(+), 20 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +1 -0
index d2aaad7..5e8593c 100644
@@ -750,6 +750,7 @@ void amdgpu_ctx_mgr_fini(struct amdgpu_ctx_mgr *mgr);
 struct amdgpu_fpriv {
 	struct amdgpu_vm	vm;
 	struct amdgpu_bo_va	*prt_va;
+	struct amdgpu_bo_va	*csa_va;
 	struct mutex		bo_list_lock;
 	struct idr		bo_list_handles;
 	struct amdgpu_ctx_mgr	ctx_mgr;
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +2 -1
index 517fde7..52c3d27 100644
@@ -778,7 +778,8 @@ static int amdgpu_bo_vm_update_pte(struct amdgpu_cs_parser *p)
 
 	if (amdgpu_sriov_vf(adev)) {
 		struct dma_fence *f;
-		bo_va = vm->csa_bo_va;
+
+		bo_va = fpriv->csa_va;
 		BUG_ON(!bo_va);
 		r = amdgpu_vm_bo_update(adev, bo_va, false);
 		if (r)
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +3 -3
index 4a6407d..6c92e99 100644
@@ -843,7 +843,7 @@ int amdgpu_driver_open_kms(struct drm_device *dev, struct drm_file *file_priv)
 	}
 
 	if (amdgpu_sriov_vf(adev)) {
-		r = amdgpu_map_static_csa(adev, &fpriv->vm);
+		r = amdgpu_map_static_csa(adev, &fpriv->vm, &fpriv->csa_va);
 		if (r)
 			goto out_suspend;
 	}
@@ -896,8 +896,8 @@ void amdgpu_driver_postclose_kms(struct drm_device *dev,
 	if (amdgpu_sriov_vf(adev)) {
 		/* TODO: how to handle reserve failure */
 		BUG_ON(amdgpu_bo_reserve(adev->virt.csa_obj, true));
-		amdgpu_vm_bo_rmv(adev, fpriv->vm.csa_bo_va);
-		fpriv->vm.csa_bo_va = NULL;
+		amdgpu_vm_bo_rmv(adev, fpriv->csa_va);
+		fpriv->csa_va = NULL;
 		amdgpu_bo_unreserve(adev->virt.csa_obj);
 	}
 
drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c +12 -13
index 8a081e1..8920845 100644
@@ -46,14 +46,14 @@ int amdgpu_allocate_static_csa(struct amdgpu_device *adev)
  * address within META_DATA init package to support SRIOV gfx preemption.
  */
 
-int amdgpu_map_static_csa(struct amdgpu_device *adev, struct amdgpu_vm *vm)
+int amdgpu_map_static_csa(struct amdgpu_device *adev, struct amdgpu_vm *vm,
+			  struct amdgpu_bo_va **bo_va)
 {
-	int r;
-	struct amdgpu_bo_va *bo_va;
 	struct ww_acquire_ctx ticket;
 	struct list_head list;
 	struct amdgpu_bo_list_entry pd;
 	struct ttm_validate_buffer csa_tv;
+	int r;
 
 	INIT_LIST_HEAD(&list);
 	INIT_LIST_HEAD(&csa_tv.head);
@@ -69,34 +69,33 @@ int amdgpu_map_static_csa(struct amdgpu_device *adev, struct amdgpu_vm *vm)
 		return r;
 	}
 
-	bo_va = amdgpu_vm_bo_add(adev, vm, adev->virt.csa_obj);
-	if (!bo_va) {
+	*bo_va = amdgpu_vm_bo_add(adev, vm, adev->virt.csa_obj);
+	if (!*bo_va) {
 		ttm_eu_backoff_reservation(&ticket, &list);
 		DRM_ERROR("failed to create bo_va for static CSA\n");
 		return -ENOMEM;
 	}
 
-	r = amdgpu_vm_alloc_pts(adev, bo_va->vm, AMDGPU_CSA_VADDR,
-				   AMDGPU_CSA_SIZE);
+	r = amdgpu_vm_alloc_pts(adev, (*bo_va)->vm, AMDGPU_CSA_VADDR,
+				AMDGPU_CSA_SIZE);
 	if (r) {
 		DRM_ERROR("failed to allocate pts for static CSA, err=%d\n", r);
-		amdgpu_vm_bo_rmv(adev, bo_va);
+		amdgpu_vm_bo_rmv(adev, *bo_va);
 		ttm_eu_backoff_reservation(&ticket, &list);
 		return r;
 	}
 
-	r = amdgpu_vm_bo_map(adev, bo_va, AMDGPU_CSA_VADDR, 0,AMDGPU_CSA_SIZE,
-						AMDGPU_PTE_READABLE | AMDGPU_PTE_WRITEABLE |
-						AMDGPU_PTE_EXECUTABLE);
+	r = amdgpu_vm_bo_map(adev, *bo_va, AMDGPU_CSA_VADDR, 0, AMDGPU_CSA_SIZE,
+			     AMDGPU_PTE_READABLE | AMDGPU_PTE_WRITEABLE |
+			     AMDGPU_PTE_EXECUTABLE);
 
 	if (r) {
 		DRM_ERROR("failed to do bo_map on static CSA, err=%d\n", r);
-		amdgpu_vm_bo_rmv(adev, bo_va);
+		amdgpu_vm_bo_rmv(adev, *bo_va);
 		ttm_eu_backoff_reservation(&ticket, &list);
 		return r;
 	}
 
-	vm->csa_bo_va = bo_va;
 	ttm_eu_backoff_reservation(&ticket, &list);
 	return 0;
 }
drivers/gpu/drm/amd/amdgpu/amdgpu_virt.h +2 -1
index e5b1baf..afcfb8b 100644
@@ -90,7 +90,8 @@ static inline bool is_virtual_machine(void)
 
 struct amdgpu_vm;
 int amdgpu_allocate_static_csa(struct amdgpu_device *adev);
-int amdgpu_map_static_csa(struct amdgpu_device *adev, struct amdgpu_vm *vm);
+int amdgpu_map_static_csa(struct amdgpu_device *adev, struct amdgpu_vm *vm,
+			  struct amdgpu_bo_va **bo_va);
 void amdgpu_virt_init_setting(struct amdgpu_device *adev);
 uint32_t amdgpu_virt_kiq_rreg(struct amdgpu_device *adev, uint32_t reg);
 void amdgpu_virt_kiq_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v);
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +0 -2
index 6e94cd2..9c309c5 100644
@@ -141,8 +141,6 @@ struct amdgpu_vm {
 	u64                     client_id;
 	/* dedicated to vm */
 	struct amdgpu_vm_id	*reserved_vmid[AMDGPU_MAX_VMHUBS];
-	/* each VM will map on CSA */
-	struct amdgpu_bo_va *csa_bo_va;
 
 	/* Flag to indicate if VM tables are updated by CPU or GPU (SDMA) */
 	bool                    use_cpu_for_update;
Signature · 6 lines
2.7.4

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

PATCH 3/8 drm/amdgpu: remove superflous amdgpu_bo_kmap in the VM

In reply to a message not in the archive

From: Christian König <[email protected]>

We now properly kmap all BOs after validation.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 28 +++++++++-------------------
 1 file changed, 9 insertions(+), 19 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +9 -19
index 4b964f5..d982ba7 100644
@@ -1264,15 +1264,14 @@ void amdgpu_vm_get_entry(struct amdgpu_pte_update_params *p, uint64_t addr,
  *
  * Check if we can update the PD with a huge page.
  */
-static int amdgpu_vm_handle_huge_pages(struct amdgpu_pte_update_params *p,
-				       struct amdgpu_vm_pt *entry,
-				       struct amdgpu_vm_pt *parent,
-				       unsigned nptes, uint64_t dst,
-				       uint64_t flags)
+static void amdgpu_vm_handle_huge_pages(struct amdgpu_pte_update_params *p,
+					struct amdgpu_vm_pt *entry,
+					struct amdgpu_vm_pt *parent,
+					unsigned nptes, uint64_t dst,
+					uint64_t flags)
 {
 	bool use_cpu_update = (p->func == amdgpu_vm_cpu_set_ptes);
 	uint64_t pd_addr, pde;
-	int r;
 
 	/* In the case of a mixed PT the PDE must point to it*/
 	if (p->adev->asic_type < CHIP_VEGA10 ||
@@ -1289,16 +1288,13 @@ static int amdgpu_vm_handle_huge_pages(struct amdgpu_pte_update_params *p,
 
 	if (entry->addr == dst &&
 	    entry->huge_page == !!(flags & AMDGPU_PDE_PTE))
-		return 0;
+		return;
 
 	entry->addr = dst;
 	entry->huge_page = !!(flags & AMDGPU_PDE_PTE);
 
 	if (use_cpu_update) {
-		r = amdgpu_bo_kmap(parent->bo, (void *)&pd_addr);
-		if (r)
-			return r;
-
+		pd_addr = (unsigned long)amdgpu_bo_kptr(parent->bo);
 		pde = pd_addr + (entry - parent->entries) * 8;
 		amdgpu_vm_cpu_set_ptes(p, pde, dst, 1, 0, flags);
 	} else {
@@ -1311,8 +1307,6 @@ static int amdgpu_vm_handle_huge_pages(struct amdgpu_pte_update_params *p,
 		pde = pd_addr + (entry - parent->entries) * 8;
 		amdgpu_vm_do_set_ptes(p, pde, dst, 1, 0, flags);
 	}
-
-	return 0;
 }
 
 /**
@@ -1339,7 +1333,6 @@ static int amdgpu_vm_update_ptes(struct amdgpu_pte_update_params *params,
 	struct amdgpu_bo *pt;
 	unsigned nptes;
 	bool use_cpu_update = (params->func == amdgpu_vm_cpu_set_ptes);
-	int r;
 
 	/* walk over the address space and update the page tables */
 	for (addr = start; addr < end; addr += nptes,
@@ -1355,11 +1348,8 @@ static int amdgpu_vm_update_ptes(struct amdgpu_pte_update_params *params,
 		else
 			nptes = AMDGPU_VM_PTE_COUNT(adev) - (addr & mask);
 
-		r = amdgpu_vm_handle_huge_pages(params, entry, parent,
-						nptes, dst, flags);
-		if (r)
-			return r;
-
+		amdgpu_vm_handle_huge_pages(params, entry, parent,
+					    nptes, dst, flags);
 		if (entry->huge_page)
 			continue;
 
Signature · 6 lines
2.7.4

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

PATCH 4/8 drm/amdgpu: drop the extra VM huge page flag

In reply to a message not in the archive

From: Christian König <[email protected]>

Just add the flags to the addr field as well.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 15 ++++++---------
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h |  1 -
 2 files changed, 6 insertions(+), 10 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +6 -9
index d982ba7..de652bf 100644
@@ -330,7 +330,6 @@ static int amdgpu_vm_alloc_levels(struct amdgpu_device *adev,
 
 			entry->bo = pt;
 			entry->addr = 0;
-			entry->huge_page = false;
 		}
 
 		if (level < adev->vm_manager.num_level) {
@@ -1082,11 +1081,11 @@ static int amdgpu_vm_update_level(struct amdgpu_device *adev,
 
 		pt = amdgpu_bo_gpu_offset(bo);
 		pt = amdgpu_gart_get_vm_pde(adev, pt);
-		if (parent->entries[pt_idx].addr == pt ||
-		    parent->entries[pt_idx].huge_page)
+		if ((parent->entries[pt_idx].addr & AMDGPU_PDE_PTE) ||
+		    parent->entries[pt_idx].addr == (pt | AMDGPU_PTE_VALID))
 			continue;
 
-		parent->entries[pt_idx].addr = pt;
+		parent->entries[pt_idx].addr = pt | AMDGPU_PTE_VALID;
 
 		pde = pd_addr + pt_idx * 8;
 		if (((last_pde + 8 * count) != pde) ||
@@ -1286,12 +1285,10 @@ static void amdgpu_vm_handle_huge_pages(struct amdgpu_pte_update_params *p,
 		flags |= AMDGPU_PDE_PTE;
 	}
 
-	if (entry->addr == dst &&
-	    entry->huge_page == !!(flags & AMDGPU_PDE_PTE))
+	if (entry->addr == (dst | flags))
 		return;
 
-	entry->addr = dst;
-	entry->huge_page = !!(flags & AMDGPU_PDE_PTE);
+	entry->addr = (dst | flags);
 
 	if (use_cpu_update) {
 		pd_addr = (unsigned long)amdgpu_bo_kptr(parent->bo);
@@ -1350,7 +1347,7 @@ static int amdgpu_vm_update_ptes(struct amdgpu_pte_update_params *params,
 
 		amdgpu_vm_handle_huge_pages(params, entry, parent,
 					    nptes, dst, flags);
-		if (entry->huge_page)
+		if (entry->addr & AMDGPU_PDE_PTE)
 			continue;
 
 		pt = entry->bo;
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +0 -1
index 9c309c5..f12c12f 100644
@@ -103,7 +103,6 @@ struct amdgpu_bo_list_entry;
 struct amdgpu_vm_pt {
 	struct amdgpu_bo	*bo;
 	uint64_t		addr;
-	bool			huge_page;
 
 	/* array of page tables, one for each directory entry */
 	struct amdgpu_vm_pt	*entries;
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): 1502376647-4519-4-git-send-email-deathsimple-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

Reviewed-by Alex Deucher
#6 lore raw

RE: [PATCH 4/8] drm/amdgpu: drop the extra VM huge page flag

Quoted 10 lines

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Christian König
Sent: Thursday, August 10, 2017 10:51 AM
To: [email protected]
Subject: [PATCH 4/8] drm/amdgpu: drop the extra VM huge page flag

From: Christian König <[email protected]>

Just add the flags to the addr field as well.

The flag makes it easier to understand what is going on. Can you add some comments to the code when you set the flags to describe that this is for huge pages? That should make it clear why we need to set them the way we do. With that fixed:

Quoted 80 lines
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 15 ++++++---------
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h |  1 -
 2 files changed, 6 insertions(+), 10 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -1
index d982ba7..de652bf 100644
@@ -330,7 +330,6 @@ static int amdgpu_vm_alloc_levels(struct
 amdgpu_device *adev,
 
 			entry->bo = pt;
 			entry->addr = 0;
-			entry->huge_page = false;
 		}
 
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +3 -3
 		if (level < adev->vm_manager.num_level) {
@@ -1082,11 +1081,11 @@ static int amdgpu_vm_update_level(struct
 amdgpu_device *adev,
 
 		pt = amdgpu_bo_gpu_offset(bo);
 		pt = amdgpu_gart_get_vm_pde(adev, pt);
-		if (parent->entries[pt_idx].addr == pt ||
-		    parent->entries[pt_idx].huge_page)
+		if ((parent->entries[pt_idx].addr & AMDGPU_PDE_PTE) ||
+		    parent->entries[pt_idx].addr == (pt |
 AMDGPU_PTE_VALID))
 			continue;
 
-		parent->entries[pt_idx].addr = pt;
+		parent->entries[pt_idx].addr = pt | AMDGPU_PTE_VALID;
 
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +2 -4
 		pde = pd_addr + pt_idx * 8;
 		if (((last_pde + 8 * count) != pde) ||
@@ -1286,12 +1285,10 @@ static void
 amdgpu_vm_handle_huge_pages(struct amdgpu_pte_update_params *p,
 		flags |= AMDGPU_PDE_PTE;
 	}
 
-	if (entry->addr == dst &&
-	    entry->huge_page == !!(flags & AMDGPU_PDE_PTE))
+	if (entry->addr == (dst | flags))
 		return;
 
-	entry->addr = dst;
-	entry->huge_page = !!(flags & AMDGPU_PDE_PTE);
+	entry->addr = (dst | flags);
 
 	if (use_cpu_update) {
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +1 -1
 		pd_addr = (unsigned long)amdgpu_bo_kptr(parent->bo);
@@ -1350,7 +1347,7 @@ static int amdgpu_vm_update_ptes(struct
 amdgpu_pte_update_params *params,
 
 		amdgpu_vm_handle_huge_pages(params, entry, parent,
 					    nptes, dst, flags);
-		if (entry->huge_page)
+		if (entry->addr & AMDGPU_PDE_PTE)
 			continue;
 
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -0
 		pt = entry->bo;
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +0 -2
index 9c309c5..f12c12f 100644
@@ -103,7 +103,6 @@ struct amdgpu_bo_list_entry;
 struct amdgpu_vm_pt {
 	struct amdgpu_bo	*bo;
 	uint64_t		addr;
-	bool			huge_page;
 
 	/* array of page tables, one for each directory entry */
 	struct amdgpu_vm_pt	*entries;
--

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

#7 lore raw

PATCH 5/8 drm/amdgpu: separate bo_va structure

In reply to a message not in the archive

From: Christian König <[email protected]>

Split that into vm_bo_base and bo_va to allow other uses as well.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c     |  4 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c    |  4 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_object.h | 14 ++----
 drivers/gpu/drm/amd/amdgpu/amdgpu_trace.h  |  4 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c   |  2 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c     | 81 ++++++++++++++++--------------
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h     | 12 +++++
 7 files changed, 66 insertions(+), 55 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +2 -2
index 52c3d27..fecd30e 100644
@@ -1402,7 +1402,7 @@ amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
 			    addr > mapping->last)
 				continue;
 
-			*bo = lobj->bo_va->bo;
+			*bo = lobj->bo_va->base.bo;
 			return mapping;
 		}
 
@@ -1411,7 +1411,7 @@ amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
 			    addr > mapping->last)
 				continue;
 
-			*bo = lobj->bo_va->bo;
+			*bo = lobj->bo_va->base.bo;
 			return mapping;
 		}
 	}
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +2 -2
index 5ae9941..7171968 100644
@@ -621,7 +621,7 @@ int amdgpu_gem_va_ioctl(struct drm_device *dev, void *data,
 
 	switch (args->operation) {
 	case AMDGPU_VA_OP_MAP:
-		r = amdgpu_vm_alloc_pts(adev, bo_va->vm, args->va_address,
+		r = amdgpu_vm_alloc_pts(adev, bo_va->base.vm, args->va_address,
 					args->map_size);
 		if (r)
 			goto error_backoff;
@@ -641,7 +641,7 @@ int amdgpu_gem_va_ioctl(struct drm_device *dev, void *data,
 						args->map_size);
 		break;
 	case AMDGPU_VA_OP_REPLACE:
-		r = amdgpu_vm_alloc_pts(adev, bo_va->vm, args->va_address,
+		r = amdgpu_vm_alloc_pts(adev, bo_va->base.vm, args->va_address,
 					args->map_size);
 		if (r)
 			goto error_backoff;
drivers/gpu/drm/amd/amdgpu/amdgpu_object.h +4 -10
index 9b7b4fc..a288fa6 100644
@@ -33,6 +33,7 @@
 
 #define AMDGPU_BO_INVALID_OFFSET	LONG_MAX
 
+/* bo virtual addresses in a vm */
 struct amdgpu_bo_va_mapping {
 	struct list_head		list;
 	struct rb_node			rb;
@@ -43,26 +44,19 @@ struct amdgpu_bo_va_mapping {
 	uint64_t			flags;
 };
 
-/* bo virtual addresses in a specific vm */
+/* User space allocated BO in a VM */
 struct amdgpu_bo_va {
+	struct amdgpu_vm_bo_base	base;
+
 	/* protected by bo being reserved */
-	struct list_head		bo_list;
 	struct dma_fence	        *last_pt_update;
 	unsigned			ref_count;
 
-	/* protected by vm mutex and spinlock */
-	struct list_head		vm_status;
-
 	/* mappings for this bo_va */
 	struct list_head		invalids;
 	struct list_head		valids;
-
-	/* constant after initialization */
-	struct amdgpu_vm		*vm;
-	struct amdgpu_bo		*bo;
 };
 
-
 struct amdgpu_bo {
 	/* Protected by tbo.reserved */
 	u32				preferred_domains;
drivers/gpu/drm/amd/amdgpu/amdgpu_trace.h +2 -2
index d8cd3e5..1c88bd5 100644
@@ -284,7 +284,7 @@ TRACE_EVENT(amdgpu_vm_bo_map,
 			     ),
 
 	    TP_fast_assign(
-			   __entry->bo = bo_va ? bo_va->bo : NULL;
+			   __entry->bo = bo_va ? bo_va->base.bo : NULL;
 			   __entry->start = mapping->start;
 			   __entry->last = mapping->last;
 			   __entry->offset = mapping->offset;
@@ -308,7 +308,7 @@ TRACE_EVENT(amdgpu_vm_bo_unmap,
 			     ),
 
 	    TP_fast_assign(
-			   __entry->bo = bo_va->bo;
+			   __entry->bo = bo_va->base.bo;
 			   __entry->start = mapping->start;
 			   __entry->last = mapping->last;
 			   __entry->offset = mapping->offset;
drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c +1 -1
index 8920845..ab05121 100644
@@ -76,7 +76,7 @@ int amdgpu_map_static_csa(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 		return -ENOMEM;
 	}
 
-	r = amdgpu_vm_alloc_pts(adev, (*bo_va)->vm, AMDGPU_CSA_VADDR,
+	r = amdgpu_vm_alloc_pts(adev, (*bo_va)->base.vm, AMDGPU_CSA_VADDR,
 				AMDGPU_CSA_SIZE);
 	if (r) {
 		DRM_ERROR("failed to allocate pts for static CSA, err=%d\n", r);
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +43 -38

198 lines not on this page · raw message

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +12 -0
index f12c12f..95e5e81 100644
@@ -99,6 +99,18 @@ struct amdgpu_bo_list_entry;
 #define AMDGPU_VM_USE_CPU_FOR_GFX (1 << 0)
 #define AMDGPU_VM_USE_CPU_FOR_COMPUTE (1 << 1)
 
+/* base structure for tracking BO usage in a VM */
+struct amdgpu_vm_bo_base {
+	/* constant after initialization */
+	struct amdgpu_vm		*vm;
+	struct amdgpu_bo		*bo;
+
+	/* protected by bo being reserved */
+	struct list_head		bo_list;
+
+	/* protected by spinlock */
+	struct list_head		vm_status;
+};
 
 struct amdgpu_vm_pt {
 	struct amdgpu_bo	*bo;
Signature · 6 lines
2.7.4

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

PATCH 6/8 drm/amdgpu: rename VM invalidated to moved

In reply to a message not in the archive

From: Christian König <[email protected]>

That better describes what happens here with the BO.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c |  2 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 16 ++++++++--------
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h |  6 +++---
 3 files changed, 12 insertions(+), 12 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +1 -1
index fecd30e..3f46b5a 100644
@@ -816,7 +816,7 @@ static int amdgpu_bo_vm_update_pte(struct amdgpu_cs_parser *p)
 
 	}
 
-	r = amdgpu_vm_clear_invalids(adev, vm, &p->job->sync);
+	r = amdgpu_vm_clear_moved(adev, vm, &p->job->sync);
 
 	if (amdgpu_vm_debug && p->bo_list) {
 		/* Invalidate all BOs to test for userspace bugs */
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +8 -8
index 93bcff2..b183965 100644
@@ -1979,25 +1979,25 @@ int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
 }
 
 /**
- * amdgpu_vm_clear_invalids - clear invalidated BOs in the PT
+ * amdgpu_vm_clear_moved - clear moved BOs in the PT
  *
  * @adev: amdgpu_device pointer
  * @vm: requested vm
  *
- * Make sure all invalidated BOs are cleared in the PT.
+ * Make sure all moved BOs are cleared in the PT.
  * Returns 0 for success.
  *
  * PTs have to be reserved and mutex must be locked!
  */
-int amdgpu_vm_clear_invalids(struct amdgpu_device *adev,
-			     struct amdgpu_vm *vm, struct amdgpu_sync *sync)
+int amdgpu_vm_clear_moved(struct amdgpu_device *adev, struct amdgpu_vm *vm,
+			    struct amdgpu_sync *sync)
 {
 	struct amdgpu_bo_va *bo_va = NULL;
 	int r = 0;
 
 	spin_lock(&vm->status_lock);
-	while (!list_empty(&vm->invalidated)) {
-		bo_va = list_first_entry(&vm->invalidated,
+	while (!list_empty(&vm->moved)) {
+		bo_va = list_first_entry(&vm->moved,
 			struct amdgpu_bo_va, base.vm_status);
 		spin_unlock(&vm->status_lock);
 
@@ -2392,7 +2392,7 @@ void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
 		spin_lock(&bo_base->vm->status_lock);
 		if (list_empty(&bo_base->vm_status))
 			list_add(&bo_base->vm_status,
-				 &bo_base->vm->invalidated);
+				 &bo_base->vm->moved);
 		spin_unlock(&bo_base->vm->status_lock);
 	}
 }
@@ -2461,7 +2461,7 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 	for (i = 0; i < AMDGPU_MAX_VMHUBS; i++)
 		vm->reserved_vmid[i] = NULL;
 	spin_lock_init(&vm->status_lock);
-	INIT_LIST_HEAD(&vm->invalidated);
+	INIT_LIST_HEAD(&vm->moved);
 	INIT_LIST_HEAD(&vm->cleared);
 	INIT_LIST_HEAD(&vm->freed);
 
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +3 -3
index 95e5e81..a740b57 100644
@@ -129,7 +129,7 @@ struct amdgpu_vm {
 	spinlock_t		status_lock;
 
 	/* BOs moved, but not yet updated in the PT */
-	struct list_head	invalidated;
+	struct list_head	moved;
 
 	/* BOs cleared in the PT because of a move */
 	struct list_head	cleared;
@@ -247,8 +247,8 @@ int amdgpu_vm_update_directories(struct amdgpu_device *adev,
 int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
 			  struct amdgpu_vm *vm,
 			  struct dma_fence **fence);
-int amdgpu_vm_clear_invalids(struct amdgpu_device *adev, struct amdgpu_vm *vm,
-			     struct amdgpu_sync *sync);
+int amdgpu_vm_clear_moved(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,
 			bool clear);
Signature · 6 lines
2.7.4

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

PATCH 7/8 drm/amdgpu: track evicted page tables

In reply to a message not in the archive

From: Christian König <[email protected]>

Instead of validating all page tables when one was evicted, track which one needs a validation.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c     |   7 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_object.c |   2 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c     | 217 ++++++++++++++---------------
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h     |  14 +-
 4 files changed, 116 insertions(+), 124 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +2 -5
index 3f46b5a..f68ac56 100644
@@ -632,9 +632,6 @@ static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
 
 	amdgpu_cs_report_moved_bytes(p->adev, p->bytes_moved,
 				     p->bytes_moved_vis);
-	fpriv->vm.last_eviction_counter =
-		atomic64_read(&p->adev->num_evictions);
-
 	if (p->bo_list) {
 		struct amdgpu_bo *gds = p->bo_list->gds_obj;
 		struct amdgpu_bo *gws = p->bo_list->gws_obj;
@@ -826,7 +823,7 @@ static int amdgpu_bo_vm_update_pte(struct amdgpu_cs_parser *p)
 			if (!bo)
 				continue;
 
-			amdgpu_vm_bo_invalidate(adev, bo);
+			amdgpu_vm_bo_invalidate(adev, bo, false);
 		}
 	}
 
@@ -851,7 +848,7 @@ static int amdgpu_cs_ib_vm_chunk(struct amdgpu_device *adev,
 	}
 
 	if (p->job->vm) {
-		p->job->vm_pd_addr = amdgpu_bo_gpu_offset(vm->root.bo);
+		p->job->vm_pd_addr = amdgpu_bo_gpu_offset(vm->root.base.bo);
 
 		r = amdgpu_bo_vm_update_pte(p);
 		if (r)
drivers/gpu/drm/amd/amdgpu/amdgpu_object.c +1 -1
index e7e8991..9a41488 100644
@@ -929,7 +929,7 @@ void amdgpu_bo_move_notify(struct ttm_buffer_object *bo,
 		return;
 
 	abo = container_of(bo, struct amdgpu_bo, tbo);
-	amdgpu_vm_bo_invalidate(adev, abo);
+	amdgpu_vm_bo_invalidate(adev, abo, evict);
 
 	amdgpu_bo_kunmap(abo);
 
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +105 -112

411 lines not on this page · raw message

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +8 -6
index a740b57..a649950 100644
@@ -113,12 +113,12 @@ struct amdgpu_vm_bo_base {
 };
 
 struct amdgpu_vm_pt {
-	struct amdgpu_bo	*bo;
-	uint64_t		addr;
+	struct amdgpu_vm_bo_base	base;
+	uint64_t			addr;
 
 	/* array of page tables, one for each directory entry */
-	struct amdgpu_vm_pt	*entries;
-	unsigned		last_entry_used;
+	struct amdgpu_vm_pt		*entries;
+	unsigned			last_entry_used;
 };
 
 struct amdgpu_vm {
@@ -128,6 +128,9 @@ struct amdgpu_vm {
 	/* protecting invalidated */
 	spinlock_t		status_lock;
 
+	/* BOs who needs a validation */
+	struct list_head	evicted;
+
 	/* BOs moved, but not yet updated in the PT */
 	struct list_head	moved;
 
@@ -140,7 +143,6 @@ struct amdgpu_vm {
 	/* contains the page directory */
 	struct amdgpu_vm_pt     root;
 	struct dma_fence	*last_dir_update;
-	uint64_t		last_eviction_counter;
 
 	/* protecting freed */
 	spinlock_t		freed_lock;
@@ -253,7 +255,7 @@ int amdgpu_vm_bo_update(struct amdgpu_device *adev,
 			struct amdgpu_bo_va *bo_va,
 			bool clear);
 void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
-			     struct amdgpu_bo *bo);
+			     struct amdgpu_bo *bo, bool evicted);
 struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
 				       struct amdgpu_bo *bo);
 struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev,
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): 1502376647-4519-7-git-send-email-deathsimple-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

#11 lore raw

Re: [PATCH 7/8] drm/amdgpu: track evicted page tables

Quoted 13 lines · Christian König
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +4 -1
@@ -2384,11 +2360,24 @@ void amdgpu_vm_bo_rmv(struct amdgpu_device *adev,
  * Mark @bo as invalid.
  */
 void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
-			     struct amdgpu_bo *bo)
+			     struct amdgpu_bo *bo, bool evicted)
 {
 	struct amdgpu_vm_bo_base *bo_base;
 
 	list_for_each_entry(bo_base, &bo->va, bo_list) {
+		struct amdgpu_vm *vm = bo_base->vm;
+
+		if (evicted && bo->tbo.resv == vm->root.base.bo->tbo.resv) {

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c void amdgpu_vm_bo_rmv(struct amdgpu_device *adev, · line 2380

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +10 -0
+			spin_lock(&bo_base->vm->status_lock);
+			list_move(&bo_base->vm_status, &vm->evicted);
+			spin_unlock(&bo_base->vm->status_lock);
+			continue;
+		}
+
+		/* Don't add page tables to the moved state */
+		if (bo->tbo.type == ttm_bo_type_kernel)
+			continue;
+

It looks like you're using two different conditions to detect page tables. When evicted is true, you use (bo->tbo.resv == vm->root.base.bo->tbo.resv). Otherwise you use (bo->tbo.type == ttm_bo_type_kernel). Are these really the same or is there a subtle difference I'm missing? I'd find the following logic easier to read, but I'm not sure if that's really your intention:

	if (bo->tbo.resv == vm->root.base.bo->tbo.resv) {
		spin_lock(&bo_base->vm->status_lock);

		if (evicted)
			list_move(&bo_base->vm_status, &vm->evicted);
		else if (list_empty(&bo_base->vm_status))
			/* this is from the next commit */
			list_add(&bo_base->vm_status, &vm->relocated);

		spin_unlock(&bo_base->vm->status_lock);
		continue;
	}

Regards,
Felix

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c void amdgpu_vm_bo_rmv(struct amdgpu_device *adev, · line 2383

Quoted 3 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -0
 		spin_lock(&bo_base->vm->status_lock);
 		if (list_empty(&bo_base->vm_status))
 			list_add(&bo_base->vm_status,

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

#12 lore raw

PATCH 8/8 drm/amdgpu: rework page directory filling

In reply to a message not in the archive

From: Christian König <[email protected]>

Keep track off relocated PDs/PTs instead of walking and checking all PDs.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 85 ++++++++++++++++++++++------------
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h |  3 ++
 2 files changed, 59 insertions(+), 29 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +56 -29
index 8e2ce91..8056e4a 100644
@@ -204,7 +204,10 @@ int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 		}
 
 		spin_lock(&vm->status_lock);
-		list_del_init(&bo_base->vm_status);
+		if (bo->parent)
+			list_move(&bo_base->vm_status, &vm->relocated);
+		else
+			list_del_init(&bo_base->vm_status);
 	}
 	spin_unlock(&vm->status_lock);
 
@@ -303,8 +306,10 @@ static int amdgpu_vm_alloc_levels(struct amdgpu_device *adev,
 			entry->base.vm = vm;
 			entry->base.bo = pt;
 			list_add_tail(&entry->base.bo_list, &pt->va);
-			INIT_LIST_HEAD(&entry->base.vm_status);
-			entry->addr = 0;
+			spin_lock(&vm->status_lock);
+			list_add(&entry->base.vm_status, &vm->relocated);
+			spin_unlock(&vm->status_lock);
+			entry->addr = ~0ULL;
 		}
 
 		if (level < adev->vm_manager.num_level) {
@@ -989,18 +994,17 @@ static int amdgpu_vm_wait_pd(struct amdgpu_device *adev, struct amdgpu_vm *vm,
  */
 static int amdgpu_vm_update_level(struct amdgpu_device *adev,
 				  struct amdgpu_vm *vm,
-				  struct amdgpu_vm_pt *parent,
-				  unsigned level)
+				  struct amdgpu_vm_pt *parent)
 {
 	struct amdgpu_bo *shadow;
 	struct amdgpu_ring *ring = NULL;
 	uint64_t pd_addr, shadow_addr = 0;
-	uint32_t incr = amdgpu_vm_bo_size(adev, level + 1);
 	uint64_t last_pde = ~0, last_pt = ~0, last_shadow = ~0;
 	unsigned count = 0, pt_idx, ndw = 0;
 	struct amdgpu_job *job;
 	struct amdgpu_pte_update_params params;
 	struct dma_fence *fence = NULL;
+	uint32_t incr;
 
 	int r;
 
@@ -1048,12 +1052,17 @@ static int amdgpu_vm_update_level(struct amdgpu_device *adev,
 
 	/* walk over the address space and update the directory */
 	for (pt_idx = 0; pt_idx <= parent->last_entry_used; ++pt_idx) {
-		struct amdgpu_bo *bo = parent->entries[pt_idx].base.bo;
+		struct amdgpu_vm_pt *entry = &parent->entries[pt_idx];
+		struct amdgpu_bo *bo = entry->base.bo;
 		uint64_t pde, pt;
 
 		if (bo == NULL)
 			continue;
 
+		spin_lock(&vm->status_lock);
+		list_del_init(&entry->base.vm_status);
+		spin_unlock(&vm->status_lock);
+
 		pt = amdgpu_bo_gpu_offset(bo);
 		pt = amdgpu_gart_get_vm_pde(adev, pt);
 		if ((parent->entries[pt_idx].addr & AMDGPU_PDE_PTE) ||
@@ -1063,6 +1072,7 @@ static int amdgpu_vm_update_level(struct amdgpu_device *adev,
 		parent->entries[pt_idx].addr = pt | AMDGPU_PTE_VALID;
 
 		pde = pd_addr + pt_idx * 8;
+		incr = amdgpu_bo_size(bo);
 		if (((last_pde + 8 * count) != pde) ||
 		    ((last_pt + incr * count) != pt) ||
 		    (count == AMDGPU_VM_MAX_UPDATE_SIZE)) {
@@ -1123,20 +1133,6 @@ static int amdgpu_vm_update_level(struct amdgpu_device *adev,
 			dma_fence_put(fence);
 		}
 	}
-	/*
-	 * Recurse into the subdirectories. This recursion is harmless because
-	 * we only have a maximum of 5 layers.
-	 */
-	for (pt_idx = 0; pt_idx <= parent->last_entry_used; ++pt_idx) {
-		struct amdgpu_vm_pt *entry = &parent->entries[pt_idx];
-
-		if (!entry->base.bo)
-			continue;
-
-		r = amdgpu_vm_update_level(adev, vm, entry, level + 1);
-		if (r)
-			return r;
-	}
 
 	return 0;
 
@@ -1152,7 +1148,8 @@ static int amdgpu_vm_update_level(struct amdgpu_device *adev,
  *
  * Mark all PD level as invalid after an error.
  */
-static void amdgpu_vm_invalidate_level(struct amdgpu_vm_pt *parent)
+static void amdgpu_vm_invalidate_level(struct amdgpu_vm *vm,
+				       struct amdgpu_vm_pt *parent)
 {
 	unsigned pt_idx;
 
@@ -1167,7 +1164,10 @@ static void amdgpu_vm_invalidate_level(struct amdgpu_vm_pt *parent)
 			continue;
 
 		entry->addr = ~0ULL;
-		amdgpu_vm_invalidate_level(entry);
+		spin_lock(&vm->status_lock);
+		list_move(&entry->base.vm_status, &vm->relocated);
+		spin_unlock(&vm->status_lock);
+		amdgpu_vm_invalidate_level(vm, entry);
 	}
 }
 
@@ -1185,9 +1185,31 @@ int amdgpu_vm_update_directories(struct amdgpu_device *adev,
 {
 	int r;
 
-	r = amdgpu_vm_update_level(adev, vm, &vm->root, 0);
-	if (r)
-		amdgpu_vm_invalidate_level(&vm->root);
+	spin_lock(&vm->status_lock);
+	while (!list_empty(&vm->relocated)) {
+		struct amdgpu_vm_bo_base *bo_base, *parent;
+		struct amdgpu_vm_pt *pt;
+		struct amdgpu_bo *bo;
+
+		bo_base = list_first_entry(&vm->relocated,
+					   struct amdgpu_vm_bo_base,
+					   vm_status);
+		spin_unlock(&vm->status_lock);
+
+		bo = bo_base->bo->parent;
+		parent = list_first_entry(&bo->va, struct amdgpu_vm_bo_base,
+					  bo_list);
+		pt = container_of(parent, struct amdgpu_vm_pt, base);
+
+		r = amdgpu_vm_update_level(adev, vm, pt);
+		if (r) {
+			amdgpu_vm_invalidate_level(vm, &vm->root);
+			break;
+		}
+
+		spin_lock(&vm->status_lock);
+	}
+	spin_unlock(&vm->status_lock);
 
 	if (vm->use_cpu_for_update) {
 		/* Flush HDP */
@@ -1575,7 +1597,7 @@ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
 
 error_free:
 	amdgpu_job_free(job);
-	amdgpu_vm_invalidate_level(&vm->root);
+	amdgpu_vm_invalidate_level(vm, &vm->root);
 	return r;
 }
 
@@ -2374,9 +2396,13 @@ void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
 			continue;
 		}
 
-		/* Don't add page tables to the moved state */
-		if (bo->tbo.type == ttm_bo_type_kernel)
+		if (bo->tbo.type == ttm_bo_type_kernel) {
+			spin_lock(&bo_base->vm->status_lock);
+			if (list_empty(&bo_base->vm_status))
+				list_add(&bo_base->vm_status, &vm->relocated);
+			spin_unlock(&bo_base->vm->status_lock);
 			continue;
+		}
 
 		spin_lock(&bo_base->vm->status_lock);
 		if (list_empty(&bo_base->vm_status))
@@ -2451,6 +2477,7 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 		vm->reserved_vmid[i] = NULL;
 	spin_lock_init(&vm->status_lock);
 	INIT_LIST_HEAD(&vm->evicted);
+	INIT_LIST_HEAD(&vm->relocated);
 	INIT_LIST_HEAD(&vm->moved);
 	INIT_LIST_HEAD(&vm->cleared);
 	INIT_LIST_HEAD(&vm->freed);
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +3 -0
index a649950..bc1c995 100644
@@ -131,6 +131,9 @@ struct amdgpu_vm {
 	/* BOs who needs a validation */
 	struct list_head	evicted;
 
+	/* PT BOs which relocated and their parent need an update */
+	struct list_head	relocated;
+
 	/* BOs moved, but not yet updated in the PT */
 	struct list_head	moved;
 
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): 1502376647-4519-8-git-send-email-deathsimple-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

Reviewed-by Alex Deucher
#14 lore raw

RE: [PATCH 8/8] drm/amdgpu: rework page directory filling

Quoted 7 lines

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Christian König
Sent: Thursday, August 10, 2017 10:51 AM
To: [email protected]
Subject: [PATCH 8/8] drm/amdgpu: rework page directory filling

Quoted 5 lines

From: Christian König <[email protected]>

Keep track off relocated PDs/PTs instead of walking and checking all PDs.

Some comments on patch 4. The rest are:

Quoted 227 lines
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 85
++++++++++++++++++++++------------
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h |  3 ++
 2 files changed, 59 insertions(+), 29 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +8 -3
index 8e2ce91..8056e4a 100644
@@ -204,7 +204,10 @@ int amdgpu_vm_validate_pt_bos(struct
 amdgpu_device *adev, struct amdgpu_vm *vm,
 		}
 
 		spin_lock(&vm->status_lock);
-		list_del_init(&bo_base->vm_status);
+		if (bo->parent)
+			list_move(&bo_base->vm_status, &vm->relocated);
+		else
+			list_del_init(&bo_base->vm_status);
 	}
 	spin_unlock(&vm->status_lock);
@@ -303,8 +306,10 @@ static int amdgpu_vm_alloc_levels(struct
 amdgpu_device *adev,
 			entry->base.vm = vm;
 			entry->base.bo = pt;
 			list_add_tail(&entry->base.bo_list, &pt->va);
-			INIT_LIST_HEAD(&entry->base.vm_status);
-			entry->addr = 0;
+			spin_lock(&vm->status_lock);
+			list_add(&entry->base.vm_status, &vm->relocated);
+			spin_unlock(&vm->status_lock);
+			entry->addr = ~0ULL;
 		}
 
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +8 -4
 		if (level < adev->vm_manager.num_level) {
@@ -989,18 +994,17 @@ static int amdgpu_vm_wait_pd(struct
 amdgpu_device *adev, struct amdgpu_vm *vm,
  */
 static int amdgpu_vm_update_level(struct amdgpu_device *adev,
 				  struct amdgpu_vm *vm,
-				  struct amdgpu_vm_pt *parent,
-				  unsigned level)
+				  struct amdgpu_vm_pt *parent)
 {
 	struct amdgpu_bo *shadow;
 	struct amdgpu_ring *ring = NULL;
 	uint64_t pd_addr, shadow_addr = 0;
-	uint32_t incr = amdgpu_vm_bo_size(adev, level + 1);
 	uint64_t last_pde = ~0, last_pt = ~0, last_shadow = ~0;
 	unsigned count = 0, pt_idx, ndw = 0;
 	struct amdgpu_job *job;
 	struct amdgpu_pte_update_params params;
 	struct dma_fence *fence = NULL;
+	uint32_t incr;
 
 	int r;
@@ -1048,12 +1052,17 @@ static int amdgpu_vm_update_level(struct
 amdgpu_device *adev,
 
 	/* walk over the address space and update the directory */
 	for (pt_idx = 0; pt_idx <= parent->last_entry_used; ++pt_idx) {
-		struct amdgpu_bo *bo = parent->entries[pt_idx].base.bo;
+		struct amdgpu_vm_pt *entry = &parent->entries[pt_idx];
+		struct amdgpu_bo *bo = entry->base.bo;
 		uint64_t pde, pt;
 
 		if (bo == NULL)
 			continue;
 
+		spin_lock(&vm->status_lock);
+		list_del_init(&entry->base.vm_status);
+		spin_unlock(&vm->status_lock);
+
 		pt = amdgpu_bo_gpu_offset(bo);
 		pt = amdgpu_gart_get_vm_pde(adev, pt);
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +1 -0
 		if ((parent->entries[pt_idx].addr & AMDGPU_PDE_PTE) ||
@@ -1063,6 +1072,7 @@ static int amdgpu_vm_update_level(struct
 amdgpu_device *adev,
 		parent->entries[pt_idx].addr = pt | AMDGPU_PTE_VALID;
 
 		pde = pd_addr + pt_idx * 8;
+		incr = amdgpu_bo_size(bo);
 		if (((last_pde + 8 * count) != pde) ||
 		    ((last_pt + incr * count) != pt) ||
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +31 -19
 		    (count == AMDGPU_VM_MAX_UPDATE_SIZE)) {
@@ -1123,20 +1133,6 @@ static int amdgpu_vm_update_level(struct
 amdgpu_device *adev,
 			dma_fence_put(fence);
 		}
 	}
-	/*
-	 * Recurse into the subdirectories. This recursion is harmless because
-	 * we only have a maximum of 5 layers.
-	 */
-	for (pt_idx = 0; pt_idx <= parent->last_entry_used; ++pt_idx) {
-		struct amdgpu_vm_pt *entry = &parent->entries[pt_idx];
-
-		if (!entry->base.bo)
-			continue;
-
-		r = amdgpu_vm_update_level(adev, vm, entry, level + 1);
-		if (r)
-			return r;
-	}
 
 	return 0;
@@ -1152,7 +1148,8 @@ static int amdgpu_vm_update_level(struct
 amdgpu_device *adev,
  *
  * Mark all PD level as invalid after an error.
  */
-static void amdgpu_vm_invalidate_level(struct amdgpu_vm_pt *parent)
+static void amdgpu_vm_invalidate_level(struct amdgpu_vm *vm,
+				       struct amdgpu_vm_pt *parent)
 {
 	unsigned pt_idx;
@@ -1167,7 +1164,10 @@ static void amdgpu_vm_invalidate_level(struct
 amdgpu_vm_pt *parent)
 			continue;
 
 		entry->addr = ~0ULL;
-		amdgpu_vm_invalidate_level(entry);
+		spin_lock(&vm->status_lock);
+		list_move(&entry->base.vm_status, &vm->relocated);
+		spin_unlock(&vm->status_lock);
+		amdgpu_vm_invalidate_level(vm, entry);
 	}
 }
@@ -1185,9 +1185,31 @@ int amdgpu_vm_update_directories(struct
 amdgpu_device *adev,
 {
 	int r;
 
-	r = amdgpu_vm_update_level(adev, vm, &vm->root, 0);
-	if (r)
-		amdgpu_vm_invalidate_level(&vm->root);
+	spin_lock(&vm->status_lock);
+	while (!list_empty(&vm->relocated)) {
+		struct amdgpu_vm_bo_base *bo_base, *parent;
+		struct amdgpu_vm_pt *pt;
+		struct amdgpu_bo *bo;
+
+		bo_base = list_first_entry(&vm->relocated,
+					   struct amdgpu_vm_bo_base,
+					   vm_status);
+		spin_unlock(&vm->status_lock);
+
+		bo = bo_base->bo->parent;
+		parent = list_first_entry(&bo->va, struct
 amdgpu_vm_bo_base,
+					  bo_list);
+		pt = container_of(parent, struct amdgpu_vm_pt, base);
+
+		r = amdgpu_vm_update_level(adev, vm, pt);
+		if (r) {
+			amdgpu_vm_invalidate_level(vm, &vm->root);
+			break;
+		}
+
+		spin_lock(&vm->status_lock);
+	}
+	spin_unlock(&vm->status_lock);
 
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +7 -3
 	if (vm->use_cpu_for_update) {
 		/* Flush HDP */
@@ -1575,7 +1597,7 @@ static int amdgpu_vm_bo_update_mapping(struct
 amdgpu_device *adev,
 
 error_free:
 	amdgpu_job_free(job);
-	amdgpu_vm_invalidate_level(&vm->root);
+	amdgpu_vm_invalidate_level(vm, &vm->root);
 	return r;
 }
@@ -2374,9 +2396,13 @@ void amdgpu_vm_bo_invalidate(struct
 amdgpu_device *adev,
 			continue;
 		}
 
-		/* Don't add page tables to the moved state */
-		if (bo->tbo.type == ttm_bo_type_kernel)
+		if (bo->tbo.type == ttm_bo_type_kernel) {
+			spin_lock(&bo_base->vm->status_lock);
+			if (list_empty(&bo_base->vm_status))
+				list_add(&bo_base->vm_status, &vm-
 >relocated);
+			spin_unlock(&bo_base->vm->status_lock);
 			continue;
+		}
 
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +1 -0
 		spin_lock(&bo_base->vm->status_lock);
 		if (list_empty(&bo_base->vm_status))
@@ -2451,6 +2477,7 @@ int amdgpu_vm_init(struct amdgpu_device *adev,
 struct amdgpu_vm *vm,
 		vm->reserved_vmid[i] = NULL;
 	spin_lock_init(&vm->status_lock);
 	INIT_LIST_HEAD(&vm->evicted);
+	INIT_LIST_HEAD(&vm->relocated);
 	INIT_LIST_HEAD(&vm->moved);
 	INIT_LIST_HEAD(&vm->cleared);
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -0
 	INIT_LIST_HEAD(&vm->freed);
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +3 -1
index a649950..bc1c995 100644
@@ -131,6 +131,9 @@ struct amdgpu_vm {
 	/* BOs who needs a validation */
 	struct list_head	evicted;
 
+	/* PT BOs which relocated and their parent need an update */
+	struct list_head	relocated;
+
 	/* BOs moved, but not yet updated in the PT */
 	struct list_head	moved;
 
--

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

#15 lore raw

Re: [PATCH 8/8] drm/amdgpu: rework page directory filling

In reply to a message not in the archive
Quoted 9 lines · Christian König
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +2 -3
@@ -1185,9 +1185,31 @@ int amdgpu_vm_update_directories(struct amdgpu_device *adev,
 {
 	int r;
 
-	r = amdgpu_vm_update_level(adev, vm, &vm->root, 0);
-	if (r)
-		amdgpu_vm_invalidate_level(&vm->root);
+	spin_lock(&vm->status_lock);
+	while (!list_empty(&vm->relocated)) {

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_update_directories(struct amdgpu_device *adev, · line 1199

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +10 -0
+		struct amdgpu_vm_bo_base *bo_base, *parent;
+		struct amdgpu_vm_pt *pt;
+		struct amdgpu_bo *bo;
+
+		bo_base = list_first_entry(&vm->relocated,
+					   struct amdgpu_vm_bo_base,
+					   vm_status);
+		spin_unlock(&vm->status_lock);
+
+		bo = bo_base->bo->parent;

For the root page directory I think parent bo->parent will be NULL. Can the root page directory be relocated. If yes, this needs to be handled as a special case here.

Regards,
Felix

Quoted 6 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +6 -0
+		parent = list_first_entry(&bo->va, struct amdgpu_vm_bo_base,
+					  bo_list);
+		pt = container_of(parent, struct amdgpu_vm_pt, base);
+
+		r = amdgpu_vm_update_level(adev, vm, pt);
+		if (r) {

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_update_directories(struct amdgpu_device *adev, · line 1215

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +7 -0
+			amdgpu_vm_invalidate_level(vm, &vm->root);
+			break;
+		}
+
+		spin_lock(&vm->status_lock);
+	}
+	spin_unlock(&vm->status_lock);
 
 	if (vm->use_cpu_for_update) {
 		/* Flush HDP */

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