drm/radeon: add large PTE support for NI, SI and CIK v5

This patch implements support for VRAM page table entry compression.
PTE construction is enhanced to identify physically contiguous page
ranges and mark them in the PTE fragment field. L1/L2 TLB support is
enabled for 64KB (SI/CIK) and 256KB (NI) PTE fragments, significantly
improving TLB utilization for VRAM allocations.

Linear store bandwidth is improved from 60GB/s to 125GB/s on Pitcairn.
Unigine Heaven 3.0 sees an average improvement from 24.7 to 27.7 FPS
on default settings at 1920x1200 resolution with vsync disabled.

See main comment in radeon_vm.c for a technical description.

v2 (chk): rebased and simplified.
v3 (chk): add missing hw setup
v4 (chk): rebased on current drm-fixes-3.15
v5 (chk): fix comments and commit text
Signed-off-by: Jay Cornwall <jay@jcornwall.me>
Signed-off-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drivers/gpu/drm/radeon/cik.c

@@ -5328,6 +5328,7 @@ static int cik_pcie_gart_enable(struct radeon_device *rdev)
 	WREG32(MC_VM_MX_L1_TLB_CNTL,
 	       (0xA << 7) |
 	       ENABLE_L1_TLB |
+	       ENABLE_L1_FRAGMENT_PROCESSING |
 	       SYSTEM_ACCESS_MODE_NOT_IN_SYS |
 	       ENABLE_ADVANCED_DRIVER_MODEL |
 	       SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU);
@@ -5340,7 +5341,8 @@ static int cik_pcie_gart_enable(struct radeon_device *rdev)
 	       CONTEXT1_IDENTITY_ACCESS_MODE(1));
 	WREG32(VM_L2_CNTL2, INVALIDATE_ALL_L1_TLBS | INVALIDATE_L2_CACHE);
 	WREG32(VM_L2_CNTL3, L2_CACHE_BIGK_ASSOCIATIVITY |
-	       L2_CACHE_BIGK_FRAGMENT_SIZE(6));
+	       BANK_SELECT(4) |
+	       L2_CACHE_BIGK_FRAGMENT_SIZE(4));
 	/* setup context0 */
 	WREG32(VM_CONTEXT0_PAGE_TABLE_START_ADDR, rdev->mc.gtt_start >> 12);
 	WREG32(VM_CONTEXT0_PAGE_TABLE_END_ADDR, rdev->mc.gtt_end >> 12);

drivers/gpu/drm/radeon/ni.c

@@ -1228,12 +1228,14 @@ static int cayman_pcie_gart_enable(struct radeon_device *rdev)
 	       SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU);
 	/* Setup L2 cache */
 	WREG32(VM_L2_CNTL, ENABLE_L2_CACHE |
+	       ENABLE_L2_FRAGMENT_PROCESSING |
 	       ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE |
 	       ENABLE_L2_PDE0_CACHE_LRU_UPDATE_BY_WRITE |
 	       EFFECTIVE_L2_QUEUE_SIZE(7) |
 	       CONTEXT1_IDENTITY_ACCESS_MODE(1));
 	WREG32(VM_L2_CNTL2, INVALIDATE_ALL_L1_TLBS | INVALIDATE_L2_CACHE);
 	WREG32(VM_L2_CNTL3, L2_CACHE_BIGK_ASSOCIATIVITY |
+	       BANK_SELECT(6) |
 	       L2_CACHE_BIGK_FRAGMENT_SIZE(6));
 	/* setup context0 */
 	WREG32(VM_CONTEXT0_PAGE_TABLE_START_ADDR, rdev->mc.gtt_start >> 12);

drivers/gpu/drm/radeon/radeon.h

@@ -848,6 +848,11 @@ struct radeon_mec {
 #define R600_PTE_READABLE	(1 << 5)
 #define R600_PTE_WRITEABLE	(1 << 6)
 
+/* PTE (Page Table Entry) fragment field for different page sizes */
+#define R600_PTE_FRAG_4KB	(0 << 7)
+#define R600_PTE_FRAG_64KB	(4 << 7)
+#define R600_PTE_FRAG_256KB	(6 << 7)
+
 struct radeon_vm_pt {
 	struct radeon_bo		*bo;
 	uint64_t			addr;

drivers/gpu/drm/radeon/radeon_vm.c

@@ -657,6 +657,84 @@ int radeon_vm_update_page_directory(struct radeon_device *rdev,
 	return 0;
 }
 
+/**
+ * radeon_vm_frag_ptes - add fragment information to PTEs
+ *
+ * @rdev: radeon_device pointer
+ * @ib: IB for the update
+ * @pe_start: first PTE to handle
+ * @pe_end: last PTE to handle
+ * @addr: addr those PTEs should point to
+ * @flags: hw mapping flags
+ *
+ * Global and local mutex must be locked!
+ */
+static void radeon_vm_frag_ptes(struct radeon_device *rdev,
+				struct radeon_ib *ib,
+				uint64_t pe_start, uint64_t pe_end,
+				uint64_t addr, uint32_t flags)
+{
+	/**
+	 * The MC L1 TLB supports variable sized pages, based on a fragment
+	 * field in the PTE. When this field is set to a non-zero value, page
+	 * granularity is increased from 4KB to (1 << (12 + frag)). The PTE
+	 * flags are considered valid for all PTEs within the fragment range
+	 * and corresponding mappings are assumed to be physically contiguous.
+	 *
+	 * The L1 TLB can store a single PTE for the whole fragment,
+	 * significantly increasing the space available for translation
+	 * caching. This leads to large improvements in throughput when the
+	 * TLB is under pressure.
+	 *
+	 * The L2 TLB distributes small and large fragments into two
+	 * asymmetric partitions. The large fragment cache is significantly
+	 * larger. Thus, we try to use large fragments wherever possible.
+	 * Userspace can support this by aligning virtual base address and
+	 * allocation size to the fragment size.
+	 */
+
+	/* NI is optimized for 256KB fragments, SI and newer for 64KB */
+	uint64_t frag_flags = rdev->family == CHIP_CAYMAN ?
+			R600_PTE_FRAG_256KB : R600_PTE_FRAG_64KB;
+	uint64_t frag_align = rdev->family == CHIP_CAYMAN ? 0x200 : 0x80;
+
+	uint64_t frag_start = ALIGN(pe_start, frag_align);
+	uint64_t frag_end = pe_end & ~(frag_align - 1);
+
+	unsigned count;
+
+	/* system pages are non continuously */
+	if ((flags & R600_PTE_SYSTEM) || !(flags & R600_PTE_VALID) ||
+	    (frag_start >= frag_end)) {
+
+		count = (pe_end - pe_start) / 8;
+		radeon_asic_vm_set_page(rdev, ib, pe_start, addr, count,
+					RADEON_GPU_PAGE_SIZE, flags);
+		return;
+	}
+
+	/* handle the 4K area at the beginning */
+	if (pe_start != frag_start) {
+		count = (frag_start - pe_start) / 8;
+		radeon_asic_vm_set_page(rdev, ib, pe_start, addr, count,
+					RADEON_GPU_PAGE_SIZE, flags);
+		addr += RADEON_GPU_PAGE_SIZE * count;
+	}
+
+	/* handle the area in the middle */
+	count = (frag_end - frag_start) / 8;
+	radeon_asic_vm_set_page(rdev, ib, frag_start, addr, count,
+				RADEON_GPU_PAGE_SIZE, flags | frag_flags);
+
+	/* handle the 4K area at the end */
+	if (frag_end != pe_end) {
+		addr += RADEON_GPU_PAGE_SIZE * count;
+		count = (pe_end - frag_end) / 8;
+		radeon_asic_vm_set_page(rdev, ib, frag_end, addr, count,
+					RADEON_GPU_PAGE_SIZE, flags);
+	}
+}
+
 /**
  * radeon_vm_update_ptes - make sure that page tables are valid
  *
@@ -703,10 +781,9 @@ static void radeon_vm_update_ptes(struct radeon_device *rdev,
 		if ((last_pte + 8 * count) != pte) {
 
 			if (count) {
-				radeon_asic_vm_set_page(rdev, ib, last_pte,
-							last_dst, count,
-							RADEON_GPU_PAGE_SIZE,
-							flags);
+				radeon_vm_frag_ptes(rdev, ib, last_pte,
+						    last_pte + 8 * count,
+						    last_dst, flags);
 			}
 
 			count = nptes;
@@ -721,9 +798,9 @@ static void radeon_vm_update_ptes(struct radeon_device *rdev,
 	}
 
 	if (count) {
-		radeon_asic_vm_set_page(rdev, ib, last_pte,
-					last_dst, count,
-					RADEON_GPU_PAGE_SIZE, flags);
+		radeon_vm_frag_ptes(rdev, ib, last_pte,
+				    last_pte + 8 * count,
+				    last_dst, flags);
 	}
 }
 

drivers/gpu/drm/radeon/si.c

@@ -4044,18 +4044,21 @@ static int si_pcie_gart_enable(struct radeon_device *rdev)
 	WREG32(MC_VM_MX_L1_TLB_CNTL,
 	       (0xA << 7) |
 	       ENABLE_L1_TLB |
+	       ENABLE_L1_FRAGMENT_PROCESSING |
 	       SYSTEM_ACCESS_MODE_NOT_IN_SYS |
 	       ENABLE_ADVANCED_DRIVER_MODEL |
 	       SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU);
 	/* Setup L2 cache */
 	WREG32(VM_L2_CNTL, ENABLE_L2_CACHE |
+	       ENABLE_L2_FRAGMENT_PROCESSING |
 	       ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE |
 	       ENABLE_L2_PDE0_CACHE_LRU_UPDATE_BY_WRITE |
 	       EFFECTIVE_L2_QUEUE_SIZE(7) |
 	       CONTEXT1_IDENTITY_ACCESS_MODE(1));
 	WREG32(VM_L2_CNTL2, INVALIDATE_ALL_L1_TLBS | INVALIDATE_L2_CACHE);
 	WREG32(VM_L2_CNTL3, L2_CACHE_BIGK_ASSOCIATIVITY |
-	       L2_CACHE_BIGK_FRAGMENT_SIZE(0));
+	       BANK_SELECT(4) |
+	       L2_CACHE_BIGK_FRAGMENT_SIZE(4));
 	/* setup context0 */
 	WREG32(VM_CONTEXT0_PAGE_TABLE_START_ADDR, rdev->mc.gtt_start >> 12);
 	WREG32(VM_CONTEXT0_PAGE_TABLE_END_ADDR, rdev->mc.gtt_end >> 12);