drm/amdgpu: flip frag_ptes and update_pts

We can add the fragment params before we split the update for the page tables.
That should save a few CPU cycles for larger updates.

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c

@@ -701,83 +701,6 @@ error_free:
 	return r;
 }
 
-/**
- * amdgpu_vm_frag_ptes - add fragment information to PTEs
- *
- * @params: see amdgpu_pte_update_params definition
- * @pe_start: first PTE to handle
- * @pe_end: last PTE to handle
- * @addr: addr those PTEs should point to
- * @flags: hw mapping flags
- */
-static void amdgpu_vm_frag_ptes(struct amdgpu_pte_update_params	*params,
-				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.
-	 */
-
-	/* SI and newer are optimized for 64KB */
-	uint64_t frag_flags = AMDGPU_PTE_FRAG(AMDGPU_LOG2_PAGES_PER_FRAG);
-	uint64_t frag_align = 0x80;
-
-	uint64_t frag_start = ALIGN(pe_start, frag_align);
-	uint64_t frag_end = pe_end & ~(frag_align - 1);
-
-	unsigned count;
-
-	/* Abort early if there isn't anything to do */
-	if (pe_start == pe_end)
-		return;
-
-	/* system pages are non continuously */
-	if (params->src || params->pages_addr ||
-		!(flags & AMDGPU_PTE_VALID) || (frag_start >= frag_end)) {
-
-		count = (pe_end - pe_start) / 8;
-		amdgpu_vm_update_pages(params, pe_start, addr, count,
-				       AMDGPU_GPU_PAGE_SIZE, flags);
-		return;
-	}
-
-	/* handle the 4K area at the beginning */
-	if (pe_start != frag_start) {
-		count = (frag_start - pe_start) / 8;
-		amdgpu_vm_update_pages(params, pe_start, addr, count,
-				       AMDGPU_GPU_PAGE_SIZE, flags);
-		addr += AMDGPU_GPU_PAGE_SIZE * count;
-	}
-
-	/* handle the area in the middle */
-	count = (frag_end - frag_start) / 8;
-	amdgpu_vm_update_pages(params, frag_start, addr, count,
-			       AMDGPU_GPU_PAGE_SIZE, flags | frag_flags);
-
-	/* handle the 4K area at the end */
-	if (frag_end != pe_end) {
-		addr += AMDGPU_GPU_PAGE_SIZE * count;
-		count = (pe_end - frag_end) / 8;
-		amdgpu_vm_update_pages(params, frag_end, addr, count,
-				       AMDGPU_GPU_PAGE_SIZE, flags);
-	}
-}
-
 /**
  * amdgpu_vm_update_ptes - make sure that page tables are valid
  *
@@ -797,7 +720,7 @@ static void amdgpu_vm_update_ptes(struct amdgpu_pte_update_params *params,
 {
 	const uint64_t mask = AMDGPU_VM_PTE_COUNT - 1;
 
-	uint64_t cur_pe_start, cur_pe_end, cur_dst;
+	uint64_t cur_pe_start, cur_nptes, cur_dst;
 	uint64_t addr; /* next GPU address to be updated */
 	uint64_t pt_idx;
 	struct amdgpu_bo *pt;
@@ -816,7 +739,7 @@ static void amdgpu_vm_update_ptes(struct amdgpu_pte_update_params *params,
 
 	cur_pe_start = amdgpu_bo_gpu_offset(pt);
 	cur_pe_start += (addr & mask) * 8;
-	cur_pe_end = cur_pe_start + 8 * nptes;
+	cur_nptes = nptes;
 	cur_dst = dst;
 
 	/* for next ptb*/
@@ -836,18 +759,19 @@ static void amdgpu_vm_update_ptes(struct amdgpu_pte_update_params *params,
 		next_pe_start = amdgpu_bo_gpu_offset(pt);
 		next_pe_start += (addr & mask) * 8;
 
-		if (cur_pe_end == next_pe_start) {
+		if ((cur_pe_start + 8 * cur_nptes) == next_pe_start) {
 			/* The next ptb is consecutive to current ptb.
-			 * Don't call amdgpu_vm_frag_ptes now.
+			 * Don't call amdgpu_vm_update_pages now.
 			 * Will update two ptbs together in future.
 			*/
-			cur_pe_end += 8 * nptes;
+			cur_nptes += nptes;
 		} else {
-			amdgpu_vm_frag_ptes(params, cur_pe_start, cur_pe_end,
-					    cur_dst, flags);
+			amdgpu_vm_update_pages(params, cur_pe_start, cur_dst,
+					       cur_nptes, AMDGPU_GPU_PAGE_SIZE,
+					       flags);
 
 			cur_pe_start = next_pe_start;
-			cur_pe_end = next_pe_start + 8 * nptes;
+			cur_nptes = nptes;
 			cur_dst = dst;
 		}
 
@@ -856,7 +780,75 @@ static void amdgpu_vm_update_ptes(struct amdgpu_pte_update_params *params,
 		dst += nptes * AMDGPU_GPU_PAGE_SIZE;
 	}
 
-	amdgpu_vm_frag_ptes(params, cur_pe_start, cur_pe_end, cur_dst, flags);
+	amdgpu_vm_update_pages(params, cur_pe_start, cur_dst, cur_nptes,
+			       AMDGPU_GPU_PAGE_SIZE, flags);
+}
+
+/*
+ * amdgpu_vm_frag_ptes - add fragment information to PTEs
+ *
+ * @params: see amdgpu_pte_update_params definition
+ * @vm: requested vm
+ * @start: first PTE to handle
+ * @end: last PTE to handle
+ * @dst: addr those PTEs should point to
+ * @flags: hw mapping flags
+ */
+static void amdgpu_vm_frag_ptes(struct amdgpu_pte_update_params	*params,
+				struct amdgpu_vm *vm,
+				uint64_t start, uint64_t end,
+				uint64_t dst, 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.
+	 */
+
+	/* 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;
+
+	uint64_t frag_start = ALIGN(start, frag_align);
+	uint64_t frag_end = end & ~(frag_align - 1);
+
+	/* system pages are non continuously */
+	if (params->src || params->pages_addr || !(flags & AMDGPU_PTE_VALID) ||
+	    (frag_start >= frag_end)) {
+
+		amdgpu_vm_update_ptes(params, vm, start, end, dst, flags);
+		return;
+	}
+
+	/* handle the 4K area at the beginning */
+	if (start != frag_start) {
+		amdgpu_vm_update_ptes(params, vm, start, frag_start,
+				      dst, flags);
+		dst += (frag_start - start) * AMDGPU_GPU_PAGE_SIZE;
+	}
+
+	/* handle the area in the middle */
+	amdgpu_vm_update_ptes(params, vm, frag_start, frag_end, dst,
+			      flags | frag_flags);
+
+	/* handle the 4K area at the end */
+	if (frag_end != end) {
+		dst += (frag_end - frag_start) * AMDGPU_GPU_PAGE_SIZE;
+		amdgpu_vm_update_ptes(params, vm, frag_end, end, dst, flags);
+	}
 }
 
 /**
@@ -953,7 +945,7 @@ static int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev,
 	if (r)
 		goto error_free;
 
-	amdgpu_vm_update_ptes(&params, vm, start, last + 1, addr, flags);
+	amdgpu_vm_frag_ptes(&params, vm, start, last + 1, addr, flags);
 
 	amdgpu_ring_pad_ib(ring, params.ib);
 	WARN_ON(params.ib->length_dw > ndw);