ttm: Make memory manager base alignment-aware.

3 files changed, 22 insertions, 6 deletions

diff --git a/linux-core/drmP.h b/linux-core/drmP.h
index 8044367f..b164e02c 100644
--- a/linux-core/drmP.h
+++ b/linux-core/drmP.h
@@ -1095,10 +1095,11 @@ extern void drm_ttm_mm_init(drm_device_t *dev, drm_ttm_mm_t *mm, unsigned long s
  * Basic memory manager support (drm_mm.c) 
  */
 
-extern drm_mm_node_t * drm_mm_get_block_locked(drm_mm_node_t * parent, unsigned long size);
+extern drm_mm_node_t * drm_mm_get_block_locked(drm_mm_node_t * parent, unsigned long size,
+					       unsigned alignment);
 extern void drm_mm_put_block_locked(drm_mm_t *mm, drm_mm_node_t *cur);
 extern drm_mm_node_t *drm_mm_search_free_locked(const drm_mm_t *mm, unsigned long size, 
-					 int best_match);
+						unsigned alignment, int best_match);
 extern int drm_mm_init(drm_mm_t *mm, unsigned long start, unsigned long size);
 
 /* Inline replacements for DRM_IOREMAP macros */
diff --git a/linux-core/drm_mm.c b/linux-core/drm_mm.c
index 05d316ee..82bbf93d 100644
--- a/linux-core/drm_mm.c
+++ b/linux-core/drm_mm.c
@@ -30,16 +30,26 @@
  * Generic simple memory manager implementation. Intended to be used as a base
  * class implementation for more advanced memory managers. The mm.mm_lock spinlock
  * needs to be explicitly taken by the client before calling any _locked function.
+ *
+ * Note that the algorithm used is quite simple and there might be substantial
+ * performance gains if a smarter free list is implemented. Currently it is just an
+ * unordered stack of free regions. Feel free to improve. 
+ *
+ * Aligned allocations can also see some improvement.
  */
 
 #include "drmP.h"
 
 drm_mm_node_t *drm_mm_get_block_locked(drm_mm_node_t * parent,
-				       unsigned long size)
+				       unsigned long size,
+				       unsigned alignment)
 {
 
 	drm_mm_node_t * child;
 
+	if (alignment) 
+		size += alignment - 1;
+
 	if (parent->size == size) {
 		list_del_init(&parent->fl_entry);
 		parent->free = FALSE;
@@ -112,7 +122,9 @@ void drm_mm_put_block_locked(drm_mm_t * mm, drm_mm_node_t * cur)
 }
 
 drm_mm_node_t *drm_mm_search_free_locked(const drm_mm_t * mm,
-					 unsigned long size, int best_match)
+					 unsigned long size, 
+					 unsigned alignment,
+					 int best_match)
 {
 	struct list_head *list;
 	const struct list_head *free_stack = &mm->root_node.fl_entry;
@@ -123,6 +135,9 @@ drm_mm_node_t *drm_mm_search_free_locked(const drm_mm_t * mm,
 	best = NULL;
 	best_size = ~0UL;
 
+	if (alignment) 
+		size += alignment - 1;
+
 	list_for_each(list, free_stack) {
 		entry = list_entry(list, drm_mm_node_t, fl_entry);
 		if (entry->size >= size) {
diff --git a/linux-core/drm_ttm.c b/linux-core/drm_ttm.c
index c8048d12..03c81bb2 100644
--- a/linux-core/drm_ttm.c
+++ b/linux-core/drm_ttm.c
@@ -885,7 +885,7 @@ static int drm_validate_ttm_region(drm_ttm_backend_list_t * entry,
 	while (!mm_node) {
 		mm_node =
 		    drm_mm_search_free_locked(&entry->mm->mm, num_pages,
-					      0);
+					      0, 0);
 		if (!mm_node) {
 			ret = drm_ttm_evict_lru_sl(entry, &have_fence, &cur_fence);
 			if (ret)
@@ -894,7 +894,7 @@ static int drm_validate_ttm_region(drm_ttm_backend_list_t * entry,
 	}
 
 	if (!entry->mm_node) {
-		mm_node = drm_mm_get_block_locked(mm_node, num_pages);
+		mm_node = drm_mm_get_block_locked(mm_node, num_pages, 0);
 		mm_node->private = mm_priv;
 		mm_priv->region = entry;
 		entry->mm_node = mm_node;