Btrfs: Block sized tree extents and extent deletion

Signed-off-by: Chris Mason <chris.mason@oracle.com>

fs/btrfs/ctree.h

@@ -28,21 +28,12 @@ struct header {
 
 struct tree_buffer;
 
-struct alloc_extent {
-	u64 blocknr;
-	u64 num_blocks;
-	u64 num_used;
-} __attribute__ ((__packed__));
-
 struct ctree_root {
 	struct tree_buffer *node;
 	struct ctree_root *extent_root;
-	struct alloc_extent *alloc_extent;
-	struct alloc_extent *reserve_extent;
+	struct key current_insert;
 	int fp;
 	struct radix_tree_root cache_radix;
-	struct alloc_extent ai1;
-	struct alloc_extent ai2;
 };
 
 struct ctree_root_info {
@@ -52,8 +43,6 @@ struct ctree_root_info {
 	u64 tree_root; /* the tree root */
 	u32 csum;
 	u32 ham;
-	struct alloc_extent alloc_extent;
-	struct alloc_extent reserve_extent;
 	u64 snapuuid[2]; /* root specific uuid */
 } __attribute__ ((__packed__));
 

fs/btrfs/disk-io.c

@@ -12,33 +12,13 @@
 
 static int allocated_blocks = 0;
 
-static int get_free_block(struct ctree_root *root, u64 *block)
+static int check_tree_block(struct ctree_root *root, struct tree_buffer *buf)
 {
-	struct stat st;
-	int ret = 0;
-
-	if (root->alloc_extent->num_used >= root->alloc_extent->num_blocks)
-		return -1;
-
-	*block = root->alloc_extent->blocknr + root->alloc_extent->num_used;
-	root->alloc_extent->num_used += 1;
-	if (root->alloc_extent->num_used >= root->alloc_extent->num_blocks) {
-		struct alloc_extent *ae = root->alloc_extent;
-		root->alloc_extent = root->reserve_extent;
-		root->reserve_extent = ae;
-		ae->num_blocks = 0;
-	}
-	st.st_size = 0;
-	ret = fstat(root->fp, &st);
-	if (st.st_size < (*block + 1) * CTREE_BLOCKSIZE) {
-		ret = ftruncate(root->fp,
-				(*block + 1) * CTREE_BLOCKSIZE);
-		if (ret) {
-			perror("ftruncate");
-			exit(1);
-		}
-	}
-	return ret;
+	if (buf->blocknr != buf->node.header.blocknr)
+		BUG();
+	if (root->node && buf->node.header.parentid != root->node->node.header.parentid)
+		BUG();
+	return 0;
 }
 
 struct tree_buffer *alloc_tree_block(struct ctree_root *root, u64 blocknr)
@@ -61,22 +41,23 @@ struct tree_buffer *alloc_tree_block(struct ctree_root *root, u64 blocknr)
 	return buf;
 }
 
-struct tree_buffer *alloc_free_block(struct ctree_root *root)
+struct tree_buffer *find_tree_block(struct ctree_root *root, u64 blocknr)
 {
-	u64 free_block;
-	int ret;
-	struct tree_buffer * buf;
-	ret = get_free_block(root, &free_block);
-	if (ret) {
-		BUG();
-		return NULL;
+	struct tree_buffer *buf;
+	buf = radix_tree_lookup(&root->cache_radix, blocknr);
+	if (buf) {
+		buf->count++;
+	} else {
+		buf = alloc_tree_block(root, blocknr);
+		if (!buf) {
+			BUG();
+			return NULL;
+		}
 	}
-	buf = alloc_tree_block(root, free_block);
-	if (!buf)
-		BUG();
 	return buf;
 }
 
+
 struct tree_buffer *read_tree_block(struct ctree_root *root, u64 blocknr)
 {
 	loff_t offset = blocknr * CTREE_BLOCKSIZE;
@@ -86,20 +67,17 @@ struct tree_buffer *read_tree_block(struct ctree_root *root, u64 blocknr)
 	buf = radix_tree_lookup(&root->cache_radix, blocknr);
 	if (buf) {
 		buf->count++;
-		goto test;
-	}
-	buf = alloc_tree_block(root, blocknr);
-	if (!buf)
-		return NULL;
-	ret = pread(root->fp, &buf->node, CTREE_BLOCKSIZE, offset);
-	if (ret != CTREE_BLOCKSIZE) {
-		free(buf);
-		return NULL;
+	} else {
+		buf = alloc_tree_block(root, blocknr);
+		if (!buf)
+			return NULL;
+		ret = pread(root->fp, &buf->node, CTREE_BLOCKSIZE, offset);
+		if (ret != CTREE_BLOCKSIZE) {
+			free(buf);
+			return NULL;
+		}
 	}
-test:
-	if (buf->blocknr != buf->node.header.blocknr)
-		BUG();
-	if (root->node && buf->node.header.parentid != root->node->node.header.parentid)
+	if (check_tree_block(root, buf))
 		BUG();
 	return buf;
 }
@@ -121,17 +99,10 @@ int write_tree_block(struct ctree_root *root, struct tree_buffer *buf)
 static int __setup_root(struct ctree_root *root, struct ctree_root *extent_root,
 			struct ctree_root_info *info, int fp)
 {
-	INIT_RADIX_TREE(&root->cache_radix, GFP_KERNEL);
 	root->fp = fp;
 	root->node = NULL;
 	root->node = read_tree_block(root, info->tree_root);
 	root->extent_root = extent_root;
-	memcpy(&root->ai1, &info->alloc_extent, sizeof(info->alloc_extent));
-	memcpy(&root->ai2, &info->reserve_extent, sizeof(info->reserve_extent));
-	root->alloc_extent = &root->ai1;
-	root->reserve_extent = &root->ai2;
-	printf("setup done reading root %p, used %lu available %lu\n", root, root->alloc_extent->num_used, root->alloc_extent->num_blocks);
-	printf("setup done reading root %p, reserve used %lu available %lu\n", root, root->reserve_extent->num_used, root->reserve_extent->num_blocks);
 	return 0;
 }
 
@@ -147,6 +118,8 @@ struct ctree_root *open_ctree(char *filename, struct ctree_super_block *super)
 		free(root);
 		return NULL;
 	}
+	INIT_RADIX_TREE(&root->cache_radix, GFP_KERNEL);
+	INIT_RADIX_TREE(&extent_root->cache_radix, GFP_KERNEL);
 	ret = pread(fp, super, sizeof(struct ctree_super_block),
 		     CTREE_SUPER_INFO_OFFSET(CTREE_BLOCKSIZE));
 	if (ret == 0 || super->root_info.tree_root == 0) {
@@ -168,8 +141,6 @@ struct ctree_root *open_ctree(char *filename, struct ctree_super_block *super)
 static int __update_root(struct ctree_root *root, struct ctree_root_info *info)
 {
 	info->tree_root = root->node->blocknr;
-	memcpy(&info->alloc_extent, root->alloc_extent, sizeof(struct alloc_extent));
-	memcpy(&info->reserve_extent, root->reserve_extent, sizeof(struct alloc_extent));
 	return 0;
 }
 
@@ -201,6 +172,7 @@ int close_ctree(struct ctree_root *root)
 void tree_block_release(struct ctree_root *root, struct tree_buffer *buf)
 {
 	buf->count--;
+	write_tree_block(root, buf);
 	if (buf->count < 0)
 		BUG();
 	if (buf->count == 0) {

fs/btrfs/disk-io.h

@@ -11,11 +11,11 @@ struct tree_buffer {
 };
 
 struct tree_buffer *read_tree_block(struct ctree_root *root, u64 blocknr);
+struct tree_buffer *find_tree_block(struct ctree_root *root, u64 blocknr);
 int write_tree_block(struct ctree_root *root, struct tree_buffer *buf);
 struct ctree_root *open_ctree(char *filename, struct ctree_super_block *s);
 int close_ctree(struct ctree_root *root);
 void tree_block_release(struct ctree_root *root, struct tree_buffer *buf);
-struct tree_buffer *alloc_free_block(struct ctree_root *root);
 int write_ctree_super(struct ctree_root *root, struct ctree_super_block *s);
 int mkfs(int fd);
 

fs/btrfs/mkfs.c

@@ -23,17 +23,10 @@ int mkfs(int fd)
 	info[0].blocknr = 16;
 	info[0].objectid = 1;
 	info[0].tree_root = 17;
-	info[0].alloc_extent.blocknr = 0;
-	info[0].alloc_extent.num_blocks = 64;
-	/* 0-17 are used (inclusive) */
-	info[0].alloc_extent.num_used = 18;
 
 	info[1].blocknr = 16;
 	info[1].objectid = 2;
-	info[1].tree_root = 64;
-	info[1].alloc_extent.blocknr = 64;
-	info[1].alloc_extent.num_blocks = 64;
-	info[1].alloc_extent.num_used = 1;
+	info[1].tree_root = 18;
 	ret = pwrite(fd, info, sizeof(info),
 		     CTREE_SUPER_INFO_OFFSET(CTREE_BLOCKSIZE));
 	if (ret != sizeof(info))
@@ -48,24 +41,36 @@ int mkfs(int fd)
 		return -1;
 
 	empty_leaf.header.parentid = 2;
-	empty_leaf.header.blocknr = 64;
-	empty_leaf.header.nritems = 2;
+	empty_leaf.header.blocknr = 18;
+	empty_leaf.header.nritems = 3;
+
+	/* item1, reserve blocks 0-16 */
 	item.key.objectid = 0;
-	item.key.offset = 64;
+	item.key.offset = 17;
 	item.key.flags = 0;
 	item.offset = LEAF_DATA_SIZE - sizeof(struct extent_item);
 	item.size = sizeof(struct extent_item);
 	extent_item.refs = 1;
-	extent_item.owner = 1;
+	extent_item.owner = 0;
 	memcpy(empty_leaf.items, &item, sizeof(item));
 	memcpy(empty_leaf.data + item.offset, &extent_item, item.size);
-	item.key.objectid = 64;
-	item.key.offset = 64;
+
+	/* item2, give block 17 to the root */
+	item.key.objectid = 17;
+	item.key.offset = 1;
 	item.offset = LEAF_DATA_SIZE - sizeof(struct extent_item) * 2;
-	extent_item.owner = 2;
+	extent_item.owner = 1;
 	memcpy(empty_leaf.items + 1, &item, sizeof(item));
 	memcpy(empty_leaf.data + item.offset, &extent_item, item.size);
-	ret = pwrite(fd, &empty_leaf, sizeof(empty_leaf), 64 * CTREE_BLOCKSIZE);
+
+	/* item3, give block 18 for the extent root */
+	item.key.objectid = 18;
+	item.key.offset = 1;
+	item.offset = LEAF_DATA_SIZE - sizeof(struct extent_item) * 3;
+	extent_item.owner = 2;
+	memcpy(empty_leaf.items + 2, &item, sizeof(item));
+	memcpy(empty_leaf.data + item.offset, &extent_item, item.size);
+	ret = pwrite(fd, &empty_leaf, sizeof(empty_leaf), 18 * CTREE_BLOCKSIZE);
 	if (ret != sizeof(empty_leaf))
 		return -1;
 	return 0;