ext4: make xattr inode reads faster

ext4_xattr_inode_read() currently reads each block sequentially while
waiting for io operation to complete before moving on to the next
block. This prevents request merging in block layer.

Add a ext4_bread_batch() function that starts reads for all blocks
then optionally waits for them to complete. A similar logic is used
in ext4_find_entry(), so update that code to use the new function.

Signed-off-by: Tahsin Erdogan <tahsin@google.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>

fs/ext4/ext4.h

@@ -2461,6 +2461,8 @@ extern void ext4_process_freed_data(struct super_block *sb, tid_t commit_tid);
 int ext4_inode_is_fast_symlink(struct inode *inode);
 struct buffer_head *ext4_getblk(handle_t *, struct inode *, ext4_lblk_t, int);
 struct buffer_head *ext4_bread(handle_t *, struct inode *, ext4_lblk_t, int);
+int ext4_bread_batch(struct inode *inode, ext4_lblk_t block, int bh_count,
+		     bool wait, struct buffer_head **bhs);
 int ext4_get_block_unwritten(struct inode *inode, sector_t iblock,
 			     struct buffer_head *bh_result, int create);
 int ext4_get_block(struct inode *inode, sector_t iblock,

fs/ext4/inode.c

@@ -1015,6 +1015,50 @@ struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
 	return ERR_PTR(-EIO);
 }
 
+/* Read a contiguous batch of blocks. */
+int ext4_bread_batch(struct inode *inode, ext4_lblk_t block, int bh_count,
+		     bool wait, struct buffer_head **bhs)
+{
+	int i, err;
+
+	for (i = 0; i < bh_count; i++) {
+		bhs[i] = ext4_getblk(NULL, inode, block + i, 0 /* map_flags */);
+		if (IS_ERR(bhs[i])) {
+			err = PTR_ERR(bhs[i]);
+			bh_count = i;
+			goto out_brelse;
+		}
+	}
+
+	for (i = 0; i < bh_count; i++)
+		/* Note that NULL bhs[i] is valid because of holes. */
+		if (bhs[i] && !buffer_uptodate(bhs[i]))
+			ll_rw_block(REQ_OP_READ, REQ_META | REQ_PRIO, 1,
+				    &bhs[i]);
+
+	if (!wait)
+		return 0;
+
+	for (i = 0; i < bh_count; i++)
+		if (bhs[i])
+			wait_on_buffer(bhs[i]);
+
+	for (i = 0; i < bh_count; i++) {
+		if (bhs[i] && !buffer_uptodate(bhs[i])) {
+			err = -EIO;
+			goto out_brelse;
+		}
+	}
+	return 0;
+
+out_brelse:
+	for (i = 0; i < bh_count; i++) {
+		brelse(bhs[i]);
+		bhs[i] = NULL;
+	}
+	return err;
+}
+
 int ext4_walk_page_buffers(handle_t *handle,
 			   struct buffer_head *head,
 			   unsigned from,

fs/ext4/namei.c

@@ -1342,13 +1342,12 @@ static struct buffer_head * ext4_find_entry (struct inode *dir,
 	struct super_block *sb;
 	struct buffer_head *bh_use[NAMEI_RA_SIZE];
 	struct buffer_head *bh, *ret = NULL;
-	ext4_lblk_t start, block, b;
+	ext4_lblk_t start, block;
 	const u8 *name = d_name->name;
-	int ra_max = 0;		/* Number of bh's in the readahead
+	size_t ra_max = 0;	/* Number of bh's in the readahead
 				   buffer, bh_use[] */
-	int ra_ptr = 0;		/* Current index into readahead
+	size_t ra_ptr = 0;	/* Current index into readahead
 				   buffer */
-	int num = 0;
 	ext4_lblk_t  nblocks;
 	int i, namelen, retval;
 	struct ext4_filename fname;
@@ -1411,31 +1410,17 @@ restart:
 		if (ra_ptr >= ra_max) {
 			/* Refill the readahead buffer */
 			ra_ptr = 0;
-			b = block;
+			if (block < start)
-			for (ra_max = 0; ra_max < NAMEI_RA_SIZE; ra_max++) {
+				ra_max = start - block;
-				/*
+			else
-				 * Terminate if we reach the end of the
+				ra_max = nblocks - block;
-				 * directory and must wrap, or if our
+			ra_max = min(ra_max, ARRAY_SIZE(bh_use));
-				 * search has finished at this block.
+			retval = ext4_bread_batch(dir, block, ra_max,
-				 */
+						  false /* wait */, bh_use);
-				if (b >= nblocks || (num && block == start)) {
+			if (retval) {
-					bh_use[ra_max] = NULL;
+				ret = ERR_PTR(retval);
-					break;
+				ra_max = 0;
-				}
+				goto cleanup_and_exit;
-				num++;
-				bh = ext4_getblk(NULL, dir, b++, 0);
-				if (IS_ERR(bh)) {
-					if (ra_max == 0) {
-						ret = bh;
-						goto cleanup_and_exit;
-					}
-					break;
-				}
-				bh_use[ra_max] = bh;
-				if (bh)
-					ll_rw_block(REQ_OP_READ,
-						    REQ_META | REQ_PRIO,
-						    1, &bh);
 			}
 		}
 		if ((bh = bh_use[ra_ptr++]) == NULL)

fs/ext4/xattr.c

@@ -317,28 +317,41 @@ static void ext4_xattr_inode_set_hash(struct inode *ea_inode, u32 hash)
  */
 static int ext4_xattr_inode_read(struct inode *ea_inode, void *buf, size_t size)
 {
-	unsigned long block = 0;
+	int blocksize = 1 << ea_inode->i_blkbits;
-	struct buffer_head *bh;
+	int bh_count = (size + blocksize - 1) >> ea_inode->i_blkbits;
-	int blocksize = ea_inode->i_sb->s_blocksize;
+	int tail_size = (size % blocksize) ?: blocksize;
-	size_t csize, copied = 0;
+	struct buffer_head *bhs_inline[8];
-	void *copy_pos = buf;
+	struct buffer_head **bhs = bhs_inline;
+	int i, ret;
 
-	while (copied < size) {
+	if (bh_count > ARRAY_SIZE(bhs_inline)) {
-		csize = (size - copied) > blocksize ? blocksize : size - copied;
+		bhs = kmalloc_array(bh_count, sizeof(*bhs), GFP_NOFS);
-		bh = ext4_bread(NULL, ea_inode, block, 0);
+		if (!bhs)
-		if (IS_ERR(bh))
+			return -ENOMEM;
-			return PTR_ERR(bh);
-		if (!bh)
-			return -EFSCORRUPTED;
-
-		memcpy(copy_pos, bh->b_data, csize);
-		brelse(bh);
-
-		copy_pos += csize;
-		block += 1;
-		copied += csize;
 	}
-	return 0;
+
+	ret = ext4_bread_batch(ea_inode, 0 /* block */, bh_count,
+			       true /* wait */, bhs);
+	if (ret)
+		goto free_bhs;
+
+	for (i = 0; i < bh_count; i++) {
+		/* There shouldn't be any holes in ea_inode. */
+		if (!bhs[i]) {
+			ret = -EFSCORRUPTED;
+			goto put_bhs;
+		}
+		memcpy((char *)buf + blocksize * i, bhs[i]->b_data,
+		       i < bh_count - 1 ? blocksize : tail_size);
+	}
+	ret = 0;
+put_bhs:
+	for (i = 0; i < bh_count; i++)
+		brelse(bhs[i]);
+free_bhs:
+	if (bhs != bhs_inline)
+		kfree(bhs);
+	return ret;
 }
 
 static int ext4_xattr_inode_iget(struct inode *parent, unsigned long ea_ino,