drm/nouveau: greatly simplify mm, killing some bugs in the process

Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>

drivers/gpu/drm/nouveau/nouveau_mem.c

@@ -742,30 +742,24 @@ nouveau_vram_manager_debug(struct ttm_mem_type_manager *man, const char *prefix)
 {
 	struct nouveau_mm *mm = man->priv;
 	struct nouveau_mm_node *r;
-	u64 total = 0, ttotal[3] = {}, tused[3] = {}, tfree[3] = {};
+	u32 total = 0, free = 0;
-	int i;
 
 	mutex_lock(&mm->mutex);
 	list_for_each_entry(r, &mm->nodes, nl_entry) {
-		printk(KERN_DEBUG "%s %s-%d: 0x%010llx 0x%010llx\n",
+		printk(KERN_DEBUG "%s %d: 0x%010llx 0x%010llx\n",
-		       prefix, r->free ? "free" : "used", r->type,
+		       prefix, r->type, ((u64)r->offset << 12),
-		       ((u64)r->offset << 12),
 		       (((u64)r->offset + r->length) << 12));
+
 		total += r->length;
-		ttotal[r->type] += r->length;
+		if (!r->type)
-		if (r->free)
+			free += r->length;
-			tfree[r->type] += r->length;
-		else
-			tused[r->type] += r->length;
 	}
 	mutex_unlock(&mm->mutex);
 
-	printk(KERN_DEBUG "%s  total: 0x%010llx\n", prefix, total << 12);
+	printk(KERN_DEBUG "%s  total: 0x%010llx free: 0x%010llx\n",
-	for (i = 0; i < 3; i++) {
+	       prefix, (u64)total << 12, (u64)free << 12);
-		printk(KERN_DEBUG "%s type %d: 0x%010llx, "
+	printk(KERN_DEBUG "%s  block: 0x%08x\n",
-				  "used 0x%010llx, free 0x%010llx\n", prefix,
+	       prefix, mm->block_size << 12);
-		       i, ttotal[i] << 12, tused[i] << 12, tfree[i] << 12);
-	}
 }
 
 const struct ttm_mem_type_manager_func nouveau_vram_manager = {

drivers/gpu/drm/nouveau/nouveau_mm.c

@@ -48,175 +48,76 @@ region_split(struct nouveau_mm *rmm, struct nouveau_mm_node *a, u32 size)
 
 	b->offset = a->offset;
 	b->length = size;
-	b->free   = a->free;
 	b->type   = a->type;
 	a->offset += size;
 	a->length -= size;
 	list_add_tail(&b->nl_entry, &a->nl_entry);
-	if (b->free)
+	if (b->type == 0)
 		list_add_tail(&b->fl_entry, &a->fl_entry);
 	return b;
 }
 
-static struct nouveau_mm_node *
+#define node(root, dir) ((root)->nl_entry.dir == &rmm->nodes) ? NULL : \
-nouveau_mm_merge(struct nouveau_mm *rmm, struct nouveau_mm_node *this)
+	list_entry((root)->nl_entry.dir, struct nouveau_mm_node, nl_entry)
-{
-	struct nouveau_mm_node *prev, *next;
-
-	/* try to merge with free adjacent entries of same type */
-	prev = list_entry(this->nl_entry.prev, struct nouveau_mm_node, nl_entry);
-	if (this->nl_entry.prev != &rmm->nodes) {
-		if (prev->free && prev->type == this->type) {
-			prev->length += this->length;
-			region_put(rmm, this);
-			this = prev;
-		}
-	}
-
-	next = list_entry(this->nl_entry.next, struct nouveau_mm_node, nl_entry);
-	if (this->nl_entry.next != &rmm->nodes) {
-		if (next->free && next->type == this->type) {
-			next->offset  = this->offset;
-			next->length += this->length;
-			region_put(rmm, this);
-			this = next;
-		}
-	}
-
-	return this;
-}
 
 void
 nouveau_mm_put(struct nouveau_mm *rmm, struct nouveau_mm_node *this)
 {
-	u32 block_s, block_l;
+	struct nouveau_mm_node *prev = node(this, prev);
+	struct nouveau_mm_node *next = node(this, next);
 
-	this->free = true;
 	list_add(&this->fl_entry, &rmm->free);
-	this = nouveau_mm_merge(rmm, this);
-
-	/* any entirely free blocks now?  we'll want to remove typing
-	 * on them now so they can be use for any memory allocation
-	 */
-	block_s = roundup(this->offset, rmm->block_size);
-	if (block_s + rmm->block_size > this->offset + this->length)
-		return;
-
-	/* split off any still-typed region at the start */
-	if (block_s != this->offset) {
-		if (!region_split(rmm, this, block_s - this->offset))
-			return;
-	}
-
-	/* split off the soon-to-be-untyped block(s) */
-	block_l = rounddown(this->length, rmm->block_size);
-	if (block_l != this->length) {
-		this = region_split(rmm, this, block_l);
-		if (!this)
-			return;
-	}
-
-	/* mark as having no type, and retry merge with any adjacent
-	 * untyped blocks
-	 */
 	this->type = 0;
-	nouveau_mm_merge(rmm, this);
+
+	if (prev && prev->type == 0) {
+		prev->length += this->length;
+		region_put(rmm, this);
+		this = prev;
+	}
+
+	if (next && next->type == 0) {
+		next->offset  = this->offset;
+		next->length += this->length;
+		region_put(rmm, this);
+	}
 }
 
 int
 nouveau_mm_get(struct nouveau_mm *rmm, int type, u32 size, u32 size_nc,
 	       u32 align, struct nouveau_mm_node **pnode)
 {
-	struct nouveau_mm_node *this, *tmp, *next;
+	struct nouveau_mm_node *prev, *this, *next;
-	u32 splitoff, avail, alloc;
+	u32 min = size_nc ? size_nc : size;
+	u32 align_mask = align - 1;
+	u32 splitoff;
+	u32 s, e;
 
-	list_for_each_entry_safe(this, tmp, &rmm->free, fl_entry) {
+	list_for_each_entry(this, &rmm->free, fl_entry) {
-		next = list_entry(this->nl_entry.next, struct nouveau_mm_node, nl_entry);
+		e = this->offset + this->length;
-		if (this->nl_entry.next == &rmm->nodes)
+		s = this->offset;
-			next = NULL;
 
-		/* skip wrongly typed blocks */
+		prev = node(this, prev);
-		if (this->type && this->type != type)
+		if (prev && prev->type != type)
+			s = roundup(s, rmm->block_size);
+
+		next = node(this, next);
+		if (next && next->type != type)
+			e = rounddown(e, rmm->block_size);
+
+		s  = (s + align_mask) & ~align_mask;
+		e &= ~align_mask;
+		if (s > e || e - s < min)
 			continue;
 
-		/* account for alignment */
+		splitoff = s - this->offset;
-		splitoff = this->offset & (align - 1);
+		if (splitoff && !region_split(rmm, this, splitoff))
-		if (splitoff)
-			splitoff = align - splitoff;
-
-		if (this->length <= splitoff)
-			continue;
-
-		/* determine total memory available from this, and
-		 * the next block (if appropriate)
-		 */
-		avail = this->length;
-		if (next && next->free && (!next->type || next->type == type))
-			avail += next->length;
-
-		avail -= splitoff;
-
-		/* determine allocation size */
-		if (size_nc) {
-			alloc = min(avail, size);
-			alloc = rounddown(alloc, size_nc);
-			if (alloc == 0)
-				continue;
-		} else {
-			alloc = size;
-			if (avail < alloc)
-				continue;
-		}
-
-		/* untyped block, split off a chunk that's a multiple
-		 * of block_size and type it
-		 */
-		if (!this->type) {
-			u32 block = roundup(alloc + splitoff, rmm->block_size);
-			if (this->length < block)
-				continue;
-
-			this = region_split(rmm, this, block);
-			if (!this)
-				return -ENOMEM;
-
-			this->type = type;
-		}
-
-		/* stealing memory from adjacent block */
-		if (alloc > this->length) {
-			u32 amount = alloc - (this->length - splitoff);
-
-			if (!next->type) {
-				amount = roundup(amount, rmm->block_size);
-
-				next = region_split(rmm, next, amount);
-				if (!next)
-					return -ENOMEM;
-
-				next->type = type;
-			}
-
-			this->length += amount;
-			next->offset += amount;
-			next->length -= amount;
-			if (!next->length) {
-				list_del(&next->nl_entry);
-				list_del(&next->fl_entry);
-				kfree(next);
-			}
-		}
-
-		if (splitoff) {
-			if (!region_split(rmm, this, splitoff))
-				return -ENOMEM;
-		}
-
-		this = region_split(rmm, this, alloc);
-		if (this == NULL)
 			return -ENOMEM;
 
-		this->free = false;
+		this = region_split(rmm, this, min(size, e - s));
+		if (!this)
+			return -ENOMEM;
+
+		this->type = type;
 		list_del(&this->fl_entry);
 		*pnode = this;
 		return 0;
@@ -234,7 +135,6 @@ nouveau_mm_init(struct nouveau_mm **prmm, u32 offset, u32 length, u32 block)
 	heap = kzalloc(sizeof(*heap), GFP_KERNEL);
 	if (!heap)
 		return -ENOMEM;
-	heap->free = true;
 	heap->offset = roundup(offset, block);
 	heap->length = rounddown(offset + length, block) - heap->offset;
 

drivers/gpu/drm/nouveau/nouveau_mm.h

@@ -30,9 +30,7 @@ struct nouveau_mm_node {
 	struct list_head fl_entry;
 	struct list_head rl_entry;
 
-	bool free;
+	u8  type;
-	int  type;
-
 	u32 offset;
 	u32 length;
 };