[Intel-gfx] [PATCH] drm/i915: Batch copy_from_user for relocation processing

Chris Wilson chris at chris-wilson.co.uk
Sat Mar 24 01:20:24 CET 2012


Originally the code tried to allocate a large enough array to perform
the copy using vmalloc, performance wasn't great and throughput was
improved by processing each individual relocation entry separately.
This too is not as efficient as one would desire. A compromise is then
to allocate a single page (since that is relatively cheap) and process
the relocations in page size batches.

x11perf -copywinwin10:	n450/pnv	i3-330m		i5-2520m (cpu)
               Before: 	  249000	 785000		 1280000 (80%)
                After:	  264000	 896000		 1280000 (65%)

Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
---
 drivers/gpu/drm/i915/i915_gem_execbuffer.c |   92 +++++++++++++++++++++++-----
 1 files changed, 77 insertions(+), 15 deletions(-)

diff --git a/drivers/gpu/drm/i915/i915_gem_execbuffer.c b/drivers/gpu/drm/i915/i915_gem_execbuffer.c
index 2d5d41b..1da04db 100644
--- a/drivers/gpu/drm/i915/i915_gem_execbuffer.c
+++ b/drivers/gpu/drm/i915/i915_gem_execbuffer.c
@@ -407,28 +407,90 @@ i915_gem_execbuffer_relocate_object(struct drm_i915_gem_object *obj,
 {
 	struct drm_i915_gem_relocation_entry __user *user_relocs;
 	struct drm_i915_gem_exec_object2 *entry = obj->exec_entry;
-	int i, ret;
+	int i, ret = 0;
+
+	if (entry->relocation_count == 0)
+		return 0;
+
+#define N_RELOC_PER_PAGE \
+	(PAGE_SIZE / sizeof(struct drm_i915_gem_relocation_entry))
 
 	user_relocs = (void __user *)(uintptr_t)entry->relocs_ptr;
-	for (i = 0; i < entry->relocation_count; i++) {
-		struct drm_i915_gem_relocation_entry reloc;
 
-		if (__copy_from_user_inatomic(&reloc,
-					      user_relocs+i,
-					      sizeof(reloc)))
-			return -EFAULT;
+	if (entry->relocation_count < N_RELOC_PER_PAGE / 2) {
+fallback:
+		for (i = 0; i < entry->relocation_count; i++) {
+			struct drm_i915_gem_relocation_entry reloc;
+			u64 offset;
 
-		ret = i915_gem_execbuffer_relocate_entry(obj, eb, &reloc);
-		if (ret)
-			return ret;
+			if (__copy_from_user_inatomic(&reloc,
+						      user_relocs+i,
+						      sizeof(reloc)))
+				return -EFAULT;
 
-		if (__copy_to_user_inatomic(&user_relocs[i].presumed_offset,
-					    &reloc.presumed_offset,
-					    sizeof(reloc.presumed_offset)))
-			return -EFAULT;
+			offset = reloc.presumed_offset;
+
+			ret = i915_gem_execbuffer_relocate_entry(obj, eb, &reloc);
+			if (ret)
+				return ret;
+
+			if (reloc.presumed_offset != offset &&
+			    __copy_to_user_inatomic(&user_relocs[i].presumed_offset,
+						    &reloc.presumed_offset,
+						    sizeof(reloc.presumed_offset)))
+				return -EFAULT;
+		}
+	} else {
+		unsigned long page;
+		int remain;
+
+		page = __get_free_page(GFP_TEMPORARY);
+		if (unlikely(page == 0))
+			goto fallback;
+
+		i = 0;
+		remain = entry->relocation_count;
+		do {
+			struct drm_i915_gem_relocation_entry *r =
+				(struct drm_i915_gem_relocation_entry *)page;
+			int count = remain;
+			if (count > N_RELOC_PER_PAGE)
+				count = N_RELOC_PER_PAGE;
+			remain -= count;
+
+			if (__copy_from_user_inatomic(r, user_relocs+i,
+						      count*sizeof(r[0]))) {
+				ret = -EFAULT;
+				goto err;
+			}
+
+			do {
+				u64 offset = r->presumed_offset;
+
+				ret = i915_gem_execbuffer_relocate_entry(obj, eb, r);
+				if (ret)
+					goto err;
+
+				if (r->presumed_offset != offset &&
+				    __copy_to_user_inatomic(&user_relocs[i].presumed_offset,
+							    &r->presumed_offset,
+							    sizeof(r->presumed_offset))) {
+					ret = -EFAULT;
+					goto err;
+				}
+
+				i++;
+				r++;
+			} while (--count);
+		} while (remain);
+
+		ret = 0;
+err:
+		free_page(page);
 	}
 
-	return 0;
+	return ret;
+#undef N_RELOC_PER_PAGE
 }
 
 static int
-- 
1.7.9.1




More information about the Intel-gfx mailing list