Skip to contentSkip to conversation

habanalabs: flush only at the end of the map/unmap

PATCH New Oded Gabbay 1 participant 1 message Dec 14, 2019 Char and misc drivers

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 habanalabs: flush only at the end of the map/unmap

From: Pawel Piskorski <[email protected]>

Optimize hl_mmu_map and hl_mmu_unmap by not calling flush(ctx) within per-page loop.

---
 drivers/misc/habanalabs/goya/goya.c  | 20 +++++++------
 drivers/misc/habanalabs/habanalabs.h |  6 ++--
 drivers/misc/habanalabs/memory.c     |  9 ++++--
 drivers/misc/habanalabs/mmu.c        | 42 ++++++++++++++++++----------
 4 files changed, 50 insertions(+), 27 deletions(-)
drivers/misc/habanalabs/goya/goya.c +12 -8
index c8d16aa4382c..08d4bef86e23 100644
@@ -4737,7 +4737,8 @@ static int goya_mmu_add_mappings_for_device_cpu(struct hl_device *hdev)
 
 	for (off = 0 ; off < CPU_FW_IMAGE_SIZE ; off += PAGE_SIZE_2MB) {
 		rc = hl_mmu_map(hdev->kernel_ctx, prop->dram_base_address + off,
-				prop->dram_base_address + off, PAGE_SIZE_2MB);
+				prop->dram_base_address + off, PAGE_SIZE_2MB,
+				(off + PAGE_SIZE_2MB) == CPU_FW_IMAGE_SIZE);
 		if (rc) {
 			dev_err(hdev->dev, "Map failed for address 0x%llx\n",
 				prop->dram_base_address + off);
@@ -4747,7 +4748,7 @@ static int goya_mmu_add_mappings_for_device_cpu(struct hl_device *hdev)
 
 	if (!(hdev->cpu_accessible_dma_address & (PAGE_SIZE_2MB - 1))) {
 		rc = hl_mmu_map(hdev->kernel_ctx, VA_CPU_ACCESSIBLE_MEM_ADDR,
-			hdev->cpu_accessible_dma_address, PAGE_SIZE_2MB);
+			hdev->cpu_accessible_dma_address, PAGE_SIZE_2MB, true);
 
 		if (rc) {
 			dev_err(hdev->dev,
@@ -4760,7 +4761,7 @@ static int goya_mmu_add_mappings_for_device_cpu(struct hl_device *hdev)
 			rc = hl_mmu_map(hdev->kernel_ctx,
 				VA_CPU_ACCESSIBLE_MEM_ADDR + cpu_off,
 				hdev->cpu_accessible_dma_address + cpu_off,
-				PAGE_SIZE_4KB);
+				PAGE_SIZE_4KB, true);
 			if (rc) {
 				dev_err(hdev->dev,
 					"Map failed for CPU accessible memory\n");
@@ -4786,14 +4787,15 @@ static int goya_mmu_add_mappings_for_device_cpu(struct hl_device *hdev)
 	for (; cpu_off >= 0 ; cpu_off -= PAGE_SIZE_4KB)
 		if (hl_mmu_unmap(hdev->kernel_ctx,
 				VA_CPU_ACCESSIBLE_MEM_ADDR + cpu_off,
-				PAGE_SIZE_4KB))
+				PAGE_SIZE_4KB, true))
 			dev_warn_ratelimited(hdev->dev,
 				"failed to unmap address 0x%llx\n",
 				VA_CPU_ACCESSIBLE_MEM_ADDR + cpu_off);
 unmap:
 	for (; off >= 0 ; off -= PAGE_SIZE_2MB)
 		if (hl_mmu_unmap(hdev->kernel_ctx,
-				prop->dram_base_address + off, PAGE_SIZE_2MB))
+				prop->dram_base_address + off, PAGE_SIZE_2MB,
+				true))
 			dev_warn_ratelimited(hdev->dev,
 				"failed to unmap address 0x%llx\n",
 				prop->dram_base_address + off);
@@ -4818,14 +4820,15 @@ void goya_mmu_remove_device_cpu_mappings(struct hl_device *hdev)
 
 	if (!(hdev->cpu_accessible_dma_address & (PAGE_SIZE_2MB - 1))) {
 		if (hl_mmu_unmap(hdev->kernel_ctx, VA_CPU_ACCESSIBLE_MEM_ADDR,
-				PAGE_SIZE_2MB))
+				PAGE_SIZE_2MB, true))
 			dev_warn(hdev->dev,
 				"Failed to unmap CPU accessible memory\n");
 	} else {
 		for (cpu_off = 0 ; cpu_off < SZ_2M ; cpu_off += PAGE_SIZE_4KB)
 			if (hl_mmu_unmap(hdev->kernel_ctx,
 					VA_CPU_ACCESSIBLE_MEM_ADDR + cpu_off,
-					PAGE_SIZE_4KB))
+					PAGE_SIZE_4KB,
+					(cpu_off + PAGE_SIZE_4KB) >= SZ_2M))
 				dev_warn_ratelimited(hdev->dev,
 					"failed to unmap address 0x%llx\n",
 					VA_CPU_ACCESSIBLE_MEM_ADDR + cpu_off);
@@ -4833,7 +4836,8 @@ void goya_mmu_remove_device_cpu_mappings(struct hl_device *hdev)
 
 	for (off = 0 ; off < CPU_FW_IMAGE_SIZE ; off += PAGE_SIZE_2MB)
 		if (hl_mmu_unmap(hdev->kernel_ctx,
-				prop->dram_base_address + off, PAGE_SIZE_2MB))
+				prop->dram_base_address + off, PAGE_SIZE_2MB,
+				(off + PAGE_SIZE_2MB) >= CPU_FW_IMAGE_SIZE))
 			dev_warn_ratelimited(hdev->dev,
 					"Failed to unmap address 0x%llx\n",
 					prop->dram_base_address + off);
drivers/misc/habanalabs/habanalabs.h +4 -2
index 00c949f4ccd1..df34227dea31 100644
@@ -1573,8 +1573,10 @@ int hl_mmu_init(struct hl_device *hdev);
 void hl_mmu_fini(struct hl_device *hdev);
 int hl_mmu_ctx_init(struct hl_ctx *ctx);
 void hl_mmu_ctx_fini(struct hl_ctx *ctx);
-int hl_mmu_map(struct hl_ctx *ctx, u64 virt_addr, u64 phys_addr, u32 page_size);
-int hl_mmu_unmap(struct hl_ctx *ctx, u64 virt_addr, u32 page_size);
+int hl_mmu_map(struct hl_ctx *ctx, u64 virt_addr, u64 phys_addr,
+		u32 page_size, bool flush_pte);
+int hl_mmu_unmap(struct hl_ctx *ctx, u64 virt_addr, u32 page_size,
+		bool flush_pte);
 void hl_mmu_swap_out(struct hl_ctx *ctx);
 void hl_mmu_swap_in(struct hl_ctx *ctx);
 
drivers/misc/habanalabs/memory.c +6 -3
index 6c72cb4eff54..b612b1ad0aac 100644
@@ -747,7 +747,8 @@ static int map_phys_pg_pack(struct hl_ctx *ctx, u64 vaddr,
 	for (i = 0 ; i < phys_pg_pack->npages ; i++) {
 		paddr = phys_pg_pack->pages[i];
 
-		rc = hl_mmu_map(ctx, next_vaddr, paddr, page_size);
+		rc = hl_mmu_map(ctx, next_vaddr, paddr, page_size,
+				(i + 1) == phys_pg_pack->npages);
 		if (rc) {
 			dev_err(hdev->dev,
 				"map failed for handle %u, npages: %llu, mapped: %llu",
@@ -765,7 +766,8 @@ static int map_phys_pg_pack(struct hl_ctx *ctx, u64 vaddr,
 err:
 	next_vaddr = vaddr;
 	for (i = 0 ; i < mapped_pg_cnt ; i++) {
-		if (hl_mmu_unmap(ctx, next_vaddr, page_size))
+		if (hl_mmu_unmap(ctx, next_vaddr, page_size,
+					(i + 1) == mapped_pg_cnt))
 			dev_warn_ratelimited(hdev->dev,
 				"failed to unmap handle %u, va: 0x%llx, pa: 0x%llx, page size: %u\n",
 					phys_pg_pack->handle, next_vaddr,
@@ -794,7 +796,8 @@ static void unmap_phys_pg_pack(struct hl_ctx *ctx, u64 vaddr,
 	next_vaddr = vaddr;
 
 	for (i = 0 ; i < phys_pg_pack->npages ; i++, next_vaddr += page_size) {
-		if (hl_mmu_unmap(ctx, next_vaddr, page_size))
+		if (hl_mmu_unmap(ctx, next_vaddr, page_size,
+				       (i + 1) == phys_pg_pack->npages))
 			dev_warn_ratelimited(hdev->dev,
 			"unmap failed for vaddr: 0x%llx\n", next_vaddr);
 
drivers/misc/habanalabs/mmu.c +28 -14

124 lines not on this page · raw message

Signature · 1 line
2.17.1