Skip to contentSkip to conversation

drm/amdgpu: increase fragmentation size for Vega10 v2

PATCH In review Christian König 2 participants 5 messages Jul 19, 2017 – Aug 4, 2017 3 messages not in the archive

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

PATCH 1/2 drm/amdgpu: increase fragmentation size for Vega10 v2

From: Christian König <[email protected]>

The fragment bits work differently for Vega10 compared to previous generations.

Increase the fragment size to 2MB for now to better handle that.

v2: handle the hardware setup as well

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c  | 5 +++--
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c   | 5 +++--
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h   | 4 +++-
 drivers/gpu/drm/amd/amdgpu/gfxhub_v1_0.c | 4 +++-
 drivers/gpu/drm/amd/amdgpu/mmhub_v1_0.c  | 4 +++-
 5 files changed, 15 insertions(+), 7 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +3 -2
index 7a8da32..fc77844 100644
@@ -588,8 +588,9 @@ static int amdgpu_info_ioctl(struct drm_device *dev, void *data, struct drm_file
 		dev_info.virtual_address_offset = AMDGPU_VA_RESERVED_SIZE;
 		dev_info.virtual_address_max = (uint64_t)adev->vm_manager.max_pfn * AMDGPU_GPU_PAGE_SIZE;
 		dev_info.virtual_address_alignment = max((int)PAGE_SIZE, AMDGPU_GPU_PAGE_SIZE);
-		dev_info.pte_fragment_size = (1 << AMDGPU_LOG2_PAGES_PER_FRAG) *
-					     AMDGPU_GPU_PAGE_SIZE;
+		dev_info.pte_fragment_size =
+			(1 << AMDGPU_LOG2_PAGES_PER_FRAG(adev)) *
+			AMDGPU_GPU_PAGE_SIZE;
 		dev_info.gart_page_size = AMDGPU_GPU_PAGE_SIZE;
 
 		dev_info.cu_active_number = adev->gfx.cu_info.number;
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +3 -2
index 55d1c7f..a3dbebe 100644
@@ -1381,8 +1381,9 @@ static int amdgpu_vm_frag_ptes(struct amdgpu_pte_update_params	*params,
 	 */
 
 	/* SI and newer are optimized for 64KB */
-	uint64_t frag_flags = AMDGPU_PTE_FRAG(AMDGPU_LOG2_PAGES_PER_FRAG);
-	uint64_t frag_align = 1 << AMDGPU_LOG2_PAGES_PER_FRAG;
+	unsigned pages_per_frag = AMDGPU_LOG2_PAGES_PER_FRAG(params->adev);
+	uint64_t frag_flags = AMDGPU_PTE_FRAG(pages_per_frag);
+	uint64_t frag_align = 1 << pages_per_frag;
 
 	uint64_t frag_start = ALIGN(start, frag_align);
 	uint64_t frag_end = end & ~(frag_align - 1);
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +3 -1
index 3441ec5..c4f5d1f 100644
@@ -51,7 +51,9 @@ struct amdgpu_bo_list_entry;
 #define AMDGPU_VM_PTB_ALIGN_SIZE   32768
 
 /* LOG2 number of continuous pages for the fragment field */
-#define AMDGPU_LOG2_PAGES_PER_FRAG 4
+#define AMDGPU_LOG2_PAGES_PER_FRAG(adev) \
+	((adev)->asic_type < CHIP_VEGA10 ? 4 : \
+	 (adev)->vm_manager.block_size)
 
 #define AMDGPU_PTE_VALID	(1ULL << 0)
 #define AMDGPU_PTE_SYSTEM	(1ULL << 1)
drivers/gpu/drm/amd/amdgpu/gfxhub_v1_0.c +3 -1
index 008ad3d..408723e 100644
@@ -129,7 +129,7 @@ static void gfxhub_v1_0_init_cache_regs(struct amdgpu_device *adev)
 	/* Setup L2 cache */
 	tmp = RREG32_SOC15(GC, 0, mmVM_L2_CNTL);
 	tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_CACHE, 1);
-	tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_FRAGMENT_PROCESSING, 0);
+	tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_FRAGMENT_PROCESSING, 1);
 	/* XXX for emulation, Refer to closed source code.*/
 	tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, L2_PDE0_CACHE_TAG_GENERATION_MODE,
 			    0);
@@ -144,6 +144,8 @@ static void gfxhub_v1_0_init_cache_regs(struct amdgpu_device *adev)
 	WREG32_SOC15(GC, 0, mmVM_L2_CNTL2, tmp);
 
 	tmp = mmVM_L2_CNTL3_DEFAULT;
+	tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, BANK_SELECT, 12);
+	tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, L2_CACHE_BIGK_FRAGMENT_SIZE, 9);
 	WREG32_SOC15(GC, 0, mmVM_L2_CNTL3, tmp);
 
 	tmp = mmVM_L2_CNTL4_DEFAULT;
drivers/gpu/drm/amd/amdgpu/mmhub_v1_0.c +3 -1
index 96f1628..ad8def3 100644
@@ -143,7 +143,7 @@ static void mmhub_v1_0_init_cache_regs(struct amdgpu_device *adev)
 	/* Setup L2 cache */
 	tmp = RREG32_SOC15(MMHUB, 0, mmVM_L2_CNTL);
 	tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_CACHE, 1);
-	tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_FRAGMENT_PROCESSING, 0);
+	tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_FRAGMENT_PROCESSING, 1);
 	/* XXX for emulation, Refer to closed source code.*/
 	tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, L2_PDE0_CACHE_TAG_GENERATION_MODE,
 			    0);
@@ -158,6 +158,8 @@ static void mmhub_v1_0_init_cache_regs(struct amdgpu_device *adev)
 	WREG32_SOC15(MMHUB, 0, mmVM_L2_CNTL2, tmp);
 
 	tmp = mmVM_L2_CNTL3_DEFAULT;
+	tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, BANK_SELECT, 12);
+	tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, L2_CACHE_BIGK_FRAGMENT_SIZE, 9);
 	WREG32_SOC15(MMHUB, 0, mmVM_L2_CNTL3, tmp);
 
 	tmp = mmVM_L2_CNTL4_DEFAULT;
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): 1500478012-20842-1-git-send-email-deathsimple-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

#2 lore raw

PATCH 2/2 drm/amdgpu: enable huge page handling in the VM v5

From: Christian König <[email protected]>

The hardware can use huge pages to map 2MB of address space with only one PDE.

v2: few cleanups and rebased
v3: skip PT updates if we are using the PDE
v4: rebased, added support for CPU based updates
v5: fix CPU based updates once more

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 119 +++++++++++++++++++++++++++------
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h |   4 ++
 2 files changed, 103 insertions(+), 20 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +99 -20
index a3dbebe..950777e 100644
@@ -351,6 +351,7 @@ 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) {
@@ -1116,7 +1117,8 @@ 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)
+		if (parent->entries[pt_idx].addr == pt ||
+		    parent->entries[pt_idx].huge_page)
 			continue;
 
 		parent->entries[pt_idx].addr = pt;
@@ -1257,29 +1259,95 @@ int amdgpu_vm_update_directories(struct amdgpu_device *adev,
 }
 
 /**
- * amdgpu_vm_find_pt - find the page table for an address
+ * amdgpu_vm_find_entry - find the entry for an address
  *
  * @p: see amdgpu_pte_update_params definition
  * @addr: virtual address in question
+ * @entry: resulting entry or NULL
+ * @parent: parent entry
  *
- * Find the page table BO for a virtual address, return NULL when none found.
+ * Find the vm_pt entry and it's parent for the given address.
  */
-static struct amdgpu_bo *amdgpu_vm_get_pt(struct amdgpu_pte_update_params *p,
-					  uint64_t addr)
+void amdgpu_vm_get_entry(struct amdgpu_pte_update_params *p, uint64_t addr,
+			 struct amdgpu_vm_pt **entry,
+			 struct amdgpu_vm_pt **parent)
 {
-	struct amdgpu_vm_pt *entry = &p->vm->root;
 	unsigned idx, level = p->adev->vm_manager.num_level;
 
-	while (entry->entries) {
+	*parent = NULL;
+	*entry = &p->vm->root;
+	while ((*entry)->entries) {
 		idx = addr >> (p->adev->vm_manager.block_size * level--);
-		idx %= amdgpu_bo_size(entry->bo) / 8;
-		entry = &entry->entries[idx];
+		idx %= amdgpu_bo_size((*entry)->bo) / 8;
+		*parent = *entry;
+		*entry = &(*entry)->entries[idx];
 	}
 
 	if (level)
-		return NULL;
+		*entry = NULL;
+}
+
+/**
+ * amdgpu_vm_handle_huge_pages - handle updating the PD with huge pages
+ *
+ * @p: see amdgpu_pte_update_params definition
+ * @entry: vm_pt entry to check
+ * @parent: parent entry
+ * @nptes: number of PTEs updated with this operation
+ * @dst: destination address where the PTEs should point to
+ * @flags: access flags fro the PTEs
+ *
+ * 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)
+{
+	bool use_cpu_update = (p->func == amdgpu_vm_cpu_set_ptes);
+	uint64_t pd_addr, pde;
+	int r;
 
-	return entry->bo;
+	/* In the case of a mixed PT the PDE must point to it*/
+	if (p->adev->asic_type < CHIP_VEGA10 ||
+	    nptes != AMDGPU_VM_PTE_COUNT(p->adev) ||
+	    p->func == amdgpu_vm_do_copy_ptes ||
+	    !(flags & AMDGPU_PTE_VALID)) {
+
+		dst = amdgpu_bo_gpu_offset(entry->bo);
+		dst = amdgpu_gart_get_vm_pde(p->adev, dst);
+		flags = AMDGPU_PTE_VALID;
+	} else {
+		flags |= AMDGPU_PDE_PTE;
+	}
+
+	if (entry->addr == dst &&
+	    entry->huge_page == !!(flags & AMDGPU_PDE_PTE))
+		return 0;
+
+	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;
+
+		pde = pd_addr + (entry - parent->entries) * 8;
+		amdgpu_vm_cpu_set_ptes(p, pde, dst, 1, 0, flags);
+	} else {
+		if (parent->bo->shadow) {
+			pd_addr = amdgpu_bo_gpu_offset(parent->bo->shadow);
+			pde = pd_addr + (entry - parent->entries) * 8;
+			amdgpu_vm_do_set_ptes(p, pde, dst, 1, 0, flags);
+		}
+		pd_addr = amdgpu_bo_gpu_offset(parent->bo);
+		pde = pd_addr + (entry - parent->entries) * 8;
+		amdgpu_vm_do_set_ptes(p, pde, dst, 1, 0, flags);
+	}
+
+	return 0;
 }
 
 /**
@@ -1306,22 +1374,31 @@ 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) {
-		pt = amdgpu_vm_get_pt(params, addr);
-		if (!pt) {
-			pr_err("PT not found, aborting update_ptes\n");
-			return -EINVAL;
-		}
+	for (addr = start; addr < end; addr += nptes,
+	     dst += nptes * AMDGPU_GPU_PAGE_SIZE) {
+		struct amdgpu_vm_pt *entry, *parent;
+
+		amdgpu_vm_get_entry(params, addr, &entry, &parent);
+		if (!entry)
+			return -ENOENT;
 
 		if ((addr & ~mask) == (end & ~mask))
 			nptes = end - addr;
 		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;
 
+		if (entry->huge_page)
+			continue;
+
+		pt = entry->bo;
 		if (use_cpu_update) {
 			pe_start = (unsigned long)pt->kptr;
 		} else {
@@ -1337,8 +1414,6 @@ static int amdgpu_vm_update_ptes(struct amdgpu_pte_update_params *params,
 		pe_start += (addr & mask) * 8;
 		params->func(params, pe_start, dst, nptes,
 			     AMDGPU_GPU_PAGE_SIZE, flags);
-
-		dst += nptes * AMDGPU_GPU_PAGE_SIZE;
 	}
 
 	return 0;
@@ -1490,6 +1565,9 @@ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
 	/* padding, etc. */
 	ndw = 64;
 
+	/* one PDE write for each huge page */
+	ndw += ((nptes >> adev->vm_manager.block_size) + 1) * 7;
+
 	if (src) {
 		/* only copy commands needed */
 		ndw += ncmds * 7;
@@ -1569,6 +1647,7 @@ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
 
 error_free:
 	amdgpu_job_free(job);
+	amdgpu_vm_invalidate_level(&vm->root);
 	return r;
 }
 
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +4 -0
index c4f5d1f..34d9174 100644
@@ -70,6 +70,9 @@ struct amdgpu_bo_list_entry;
 /* TILED for VEGA10, reserved for older ASICs  */
 #define AMDGPU_PTE_PRT		(1ULL << 51)
 
+/* PDE is handled as PTE for VEGA10 */
+#define AMDGPU_PDE_PTE		(1ULL << 54)
+
 /* VEGA10 only */
 #define AMDGPU_PTE_MTYPE(a)    ((uint64_t)a << 57)
 #define AMDGPU_PTE_MTYPE_MASK	AMDGPU_PTE_MTYPE(3ULL)
@@ -100,6 +103,7 @@ 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): 1500478012-20842-2-git-send-email-deathsimple-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

#4 lore raw

Re: [PATCH 2/2] drm/amdgpu: enable huge page handling in the VM v5

Two nit-picks inline. Otherwise this series is Reviewed-and-tested-by:
Felix Kuehling <[email protected]>

That said, I was not able to measure any performance improvement on Vega10 using mixbench yet. I checked that the same test is sensitive to TLB optimizations on Fiji. So probably we're still missing something.

Regards,
Felix

Quoted 154 lines · Christian König

From: Christian König <[email protected]>

The hardware can use huge pages to map 2MB of address space with only one PDE.

v2: few cleanups and rebased
v3: skip PT updates if we are using the PDE
v4: rebased, added support for CPU based updates
v5: fix CPU based updates once more

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 119 +++++++++++++++++++++++++++------
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h |   4 ++
 2 files changed, 103 insertions(+), 20 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +80 -15
index a3dbebe..950777e 100644
@@ -351,6 +351,7 @@ 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) {
@@ -1116,7 +1117,8 @@ 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)
+		if (parent->entries[pt_idx].addr == pt ||
+		    parent->entries[pt_idx].huge_page)
 			continue;
 
 		parent->entries[pt_idx].addr = pt;
@@ -1257,29 +1259,95 @@ int amdgpu_vm_update_directories(struct amdgpu_device *adev,
 }
 
 /**
- * amdgpu_vm_find_pt - find the page table for an address
+ * amdgpu_vm_find_entry - find the entry for an address
  *
  * @p: see amdgpu_pte_update_params definition
  * @addr: virtual address in question
+ * @entry: resulting entry or NULL
+ * @parent: parent entry
  *
- * Find the page table BO for a virtual address, return NULL when none found.
+ * Find the vm_pt entry and it's parent for the given address.
  */
-static struct amdgpu_bo *amdgpu_vm_get_pt(struct amdgpu_pte_update_params *p,
-					  uint64_t addr)
+void amdgpu_vm_get_entry(struct amdgpu_pte_update_params *p, uint64_t addr,
+			 struct amdgpu_vm_pt **entry,
+			 struct amdgpu_vm_pt **parent)
 {
-	struct amdgpu_vm_pt *entry = &p->vm->root;
 	unsigned idx, level = p->adev->vm_manager.num_level;
 
-	while (entry->entries) {
+	*parent = NULL;
+	*entry = &p->vm->root;
+	while ((*entry)->entries) {
 		idx = addr >> (p->adev->vm_manager.block_size * level--);
-		idx %= amdgpu_bo_size(entry->bo) / 8;
-		entry = &entry->entries[idx];
+		idx %= amdgpu_bo_size((*entry)->bo) / 8;
+		*parent = *entry;
+		*entry = &(*entry)->entries[idx];
 	}
 
 	if (level)
-		return NULL;
+		*entry = NULL;
+}
+
+/**
+ * amdgpu_vm_handle_huge_pages - handle updating the PD with huge pages
+ *
+ * @p: see amdgpu_pte_update_params definition
+ * @entry: vm_pt entry to check
+ * @parent: parent entry
+ * @nptes: number of PTEs updated with this operation
+ * @dst: destination address where the PTEs should point to
+ * @flags: access flags fro the PTEs
+ *
+ * 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)
+{
+	bool use_cpu_update = (p->func == amdgpu_vm_cpu_set_ptes);
+	uint64_t pd_addr, pde;
+	int r;
 
-	return entry->bo;
+	/* In the case of a mixed PT the PDE must point to it*/
+	if (p->adev->asic_type < CHIP_VEGA10 ||
+	    nptes != AMDGPU_VM_PTE_COUNT(p->adev) ||
+	    p->func == amdgpu_vm_do_copy_ptes ||
+	    !(flags & AMDGPU_PTE_VALID)) {
+
+		dst = amdgpu_bo_gpu_offset(entry->bo);
+		dst = amdgpu_gart_get_vm_pde(p->adev, dst);
+		flags = AMDGPU_PTE_VALID;
+	} else {
+		flags |= AMDGPU_PDE_PTE;
+	}
+
+	if (entry->addr == dst &&
+	    entry->huge_page == !!(flags & AMDGPU_PDE_PTE))
+		return 0;
+
+	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;
+
+		pde = pd_addr + (entry - parent->entries) * 8;
+		amdgpu_vm_cpu_set_ptes(p, pde, dst, 1, 0, flags);
+	} else {
+		if (parent->bo->shadow) {
+			pd_addr = amdgpu_bo_gpu_offset(parent->bo->shadow);
+			pde = pd_addr + (entry - parent->entries) * 8;
+			amdgpu_vm_do_set_ptes(p, pde, dst, 1, 0, flags);
+		}
+		pd_addr = amdgpu_bo_gpu_offset(parent->bo);
+		pde = pd_addr + (entry - parent->entries) * 8;
+		amdgpu_vm_do_set_ptes(p, pde, dst, 1, 0, flags);
+	}
+
+	return 0;
 }
 
 /**
@@ -1306,22 +1374,31 @@ 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) {
-		pt = amdgpu_vm_get_pt(params, addr);
-		if (!pt) {

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c static int amdgpu_vm_update_ptes(struct amdgpu_pte_update_params *params, · line 1386

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +7 -3
-			pr_err("PT not found, aborting update_ptes\n");
-			return -EINVAL;
-		}
+	for (addr = start; addr < end; addr += nptes,
+	     dst += nptes * AMDGPU_GPU_PAGE_SIZE) {
+		struct amdgpu_vm_pt *entry, *parent;
+
+		amdgpu_vm_get_entry(params, addr, &entry, &parent);
+		if (!entry)
+			return -ENOENT;

ENOENT means "no such file or directory". Not sure that makes sense outside of file system code.

Quoted 28 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +8 -2
 		if ((addr & ~mask) == (end & ~mask))
 			nptes = end - addr;
 		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;
 
+		if (entry->huge_page)
+			continue;
+
+		pt = entry->bo;
 		if (use_cpu_update) {
 			pe_start = (unsigned long)pt->kptr;
 		} else {
@@ -1337,8 +1414,6 @@ static int amdgpu_vm_update_ptes(struct amdgpu_pte_update_params *params,
 		pe_start += (addr & mask) * 8;
 		params->func(params, pe_start, dst, nptes,
 			     AMDGPU_GPU_PAGE_SIZE, flags);
-
-		dst += nptes * AMDGPU_GPU_PAGE_SIZE;
 	}
 
 	return 0;

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev, · line 1569

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +2 -0
@@ -1490,6 +1565,9 @@ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
 	/* padding, etc. */
 	ndw = 64;
 
+	/* one PDE write for each huge page */
+	ndw += ((nptes >> adev->vm_manager.block_size) + 1) * 7;

I think page directory updates only need 6 dwords. That's what update_level uses anyway.

Quoted 27 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +2 -0
+
 	if (src) {
 		/* only copy commands needed */
 		ndw += ncmds * 7;
@@ -1569,6 +1647,7 @@ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
 
 error_free:
 	amdgpu_job_free(job);
+	amdgpu_vm_invalidate_level(&vm->root);
 	return r;
 }
 
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +3 -0
index c4f5d1f..34d9174 100644
@@ -70,6 +70,9 @@ struct amdgpu_bo_list_entry;
 /* TILED for VEGA10, reserved for older ASICs  */
 #define AMDGPU_PTE_PRT		(1ULL << 51)
 
+/* PDE is handled as PTE for VEGA10 */
+#define AMDGPU_PDE_PTE		(1ULL << 54)
+
 /* VEGA10 only */
 #define AMDGPU_PTE_MTYPE(a)    ((uint64_t)a << 57)
 #define AMDGPU_PTE_MTYPE_MASK	AMDGPU_PTE_MTYPE(3ULL)

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h struct amdgpu_bo_list_entry; · line 109

Quoted 7 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +1 -0
@@ -100,6 +103,7 @@ 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;

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

#5 lore raw
In reply to a message not in the archive
Quoted 21 lines · Christian König

From: Christian König <[email protected]>

The fragment bits work differently for Vega10 compared to previous generations.

Increase the fragment size to 2MB for now to better handle that.

v2: handle the hardware setup as well

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c  | 5 +++--
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c   | 5 +++--
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h   | 4 +++-
 drivers/gpu/drm/amd/amdgpu/gfxhub_v1_0.c | 4 +++-
 drivers/gpu/drm/amd/amdgpu/mmhub_v1_0.c  | 4 +++-
 5 files changed, 15 insertions(+), 7 deletions(-)

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h struct amdgpu_bo_list_entry; · line 56

Quoted 7 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +3 -1
@@ -51,7 +51,9 @@ struct amdgpu_bo_list_entry;
 #define AMDGPU_VM_PTB_ALIGN_SIZE   32768
 
 /* LOG2 number of continuous pages for the fragment field */
-#define AMDGPU_LOG2_PAGES_PER_FRAG 4
+#define AMDGPU_LOG2_PAGES_PER_FRAG(adev) \
+	((adev)->asic_type < CHIP_VEGA10 ? 4 : \
+	 (adev)->vm_manager.block_size)

With the hardware configuration changed for 256KB BigK TLBs, should this be changed to 256KB fragments on GFX9 now to get another small TLB boost for allocations smaller than 2MB?

Regards,
Felix

_______________________________________________
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): 9a488f46-5ac2-a9e1-a998-6a03cc11c83e-5C7GfCeVMHo@public.gmane.org

#7 lore raw

Am 03.08.2017 um 23:01 schrieb Felix Kuehling:

Quoted 24 lines · Christian König
Quoted 23 lines

From: Christian König <[email protected]>

The fragment bits work differently for Vega10 compared to previous generations.

Increase the fragment size to 2MB for now to better handle that.

v2: handle the hardware setup as well

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c  | 5 +++--
  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c   | 5 +++--
  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h   | 4 +++-
  drivers/gpu/drm/amd/amdgpu/gfxhub_v1_0.c | 4 +++-
  drivers/gpu/drm/amd/amdgpu/mmhub_v1_0.c  | 4 +++-
  5 files changed, 15 insertions(+), 7 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +0 -0
@@ -51,7 +51,9 @@ struct amdgpu_bo_list_entry;
  #define AMDGPU_VM_PTB_ALIGN_SIZE   32768
 
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h struct amdgpu_bo_list_entry; · line 56

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +3 -1
  /* LOG2 number of continuous pages for the fragment field */
-#define AMDGPU_LOG2_PAGES_PER_FRAG 4
+#define AMDGPU_LOG2_PAGES_PER_FRAG(adev) \
+	((adev)->asic_type < CHIP_VEGA10 ? 4 : \
+	 (adev)->vm_manager.block_size)

With the hardware configuration changed for 256KB BigK TLBs, should this be changed to 256KB fragments on GFX9 now to get another small TLB boost for allocations smaller than 2MB?

As far as I understand it the answer is no. The fragment field on Vega10 only affects the L1, not the L2 any more.

And we need to report 2MB as fragment size to userspace to get the BigK effect on the TLB.

Well what we could do is changing the fragment generation a bit make the L1 be used more efficient.

I've tried that for gfx8 once, but that unfortunately backfired because it affected the L2 as well, but for gfx9 that should now work fine.

Christian.

Quoted 9 lines

Regards,
Felix

_______________________________________________
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