ext4: add support for extent pre-caching

Add a new fiemap flag which forces the all of the extents in an inode
to be cached in the extent_status tree. This is critically important
when using AIO to a preallocated file, since if we need to read in
blocks from the extent tree, the io_submit(2) system call becomes
synchronous, and the AIO is no longer "A", which is bad.

In addition, for most files which have an external leaf tree block,
the cost of caching the information in the extent status tree will be
less than caching the entire 4k block in the buffer cache. So it is
generally a win to keep the extent information cached.

Signed-off-by: "Theodore Ts'o" <[e-mail]>

fs/ext4/ext4.h

···
 #define EXT4_GET_BLOCKS_NO_PUT_HOLE		0x0200
 
 /*
- * The bit position of this flag must not overlap with any of the
- * EXT4_GET_BLOCKS_*.  It is used by ext4_ext_find_extent(),
+ * The bit position of these flags must not overlap with any of the
+ * EXT4_GET_BLOCKS_*.  They are used by ext4_ext_find_extent(),
  * read_extent_tree_block(), ext4_split_extent_at(),
- * ext4_ext_insert_extent(), and ext4_ext_create_new_leaf() to
- * indicate that the we shouldn't be caching the extents when reading
- * from the extent tree while a truncate or punch hole operation
- * is in progress.
+ * ext4_ext_insert_extent(), and ext4_ext_create_new_leaf().
+ * EXT4_EX_NOCACHE is used to indicate that the we shouldn't be
+ * caching the extents when reading from the extent tree while a
+ * truncate or punch hole operation is in progress.
  */
 #define EXT4_EX_NOCACHE				0x0400
+#define EXT4_EX_FORCE_CACHE			0x0800
 
 /*
  * Flags used by ext4_free_blocks
···
 #define EXT4_IOC_MOVE_EXT		_IOWR('f', 15, struct move_extent)
 #define EXT4_IOC_RESIZE_FS		_IOW('f', 16, __u64)
 #define EXT4_IOC_SWAP_BOOT		_IO('f', 17)
+#define EXT4_IOC_PRECACHE_EXTENTS	_IO('f', 18)
 
 #if defined(__KERNEL__) && defined(CONFIG_COMPAT)
 /*
···
 					   nolocking */
 	EXT4_STATE_MAY_INLINE_DATA,	/* may have in-inode data */
 	EXT4_STATE_ORDERED_MODE,	/* data=ordered mode */
+	EXT4_STATE_EXT_PRECACHED,	/* extents have been precached */
 };
 
 #define EXT4_INODE_BIT_FNS(name, field, offset)				\
···
 extern int ext4_find_delalloc_cluster(struct inode *inode, ext4_lblk_t lblk);
 extern int ext4_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
 			__u64 start, __u64 len);
-
+extern int ext4_ext_precache(struct inode *inode);
 
 /* move_extent.c */
 extern void ext4_double_down_write_data_sem(struct inode *first,

fs/ext4/extents.c

···
 		if (err < 0)
 			goto errout;
 	}
-	if (buffer_verified(bh))
+	if (buffer_verified(bh) && !(flags & EXT4_EX_FORCE_CACHE))
 		return bh;
 	err = __ext4_ext_check(function, line, inode,
 			       ext_block_hdr(bh), depth, pblk);
···
 #define read_extent_tree_block(inode, pblk, depth, flags)		\
 	__read_extent_tree_block(__func__, __LINE__, (inode), (pblk),   \
 				 (depth), (flags))
+
+/*
+ * This function is called to cache a file's extent information in the
+ * extent status tree
+ */
+int ext4_ext_precache(struct inode *inode)
+{
+	struct ext4_inode_info *ei = EXT4_I(inode);
+	struct ext4_ext_path *path = NULL;
+	struct buffer_head *bh;
+	int i = 0, depth, ret = 0;
+
+	if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
+		return 0;	/* not an extent-mapped inode */
+
+	down_read(&ei->i_data_sem);
+	depth = ext_depth(inode);
+
+	path = kzalloc(sizeof(struct ext4_ext_path) * (depth + 1),
+		       GFP_NOFS);
+	if (path == NULL) {
+		up_read(&ei->i_data_sem);
+		return -ENOMEM;
+	}
+
+	/* Don't cache anything if there are no external extent blocks */
+	if (depth == 0)
+		goto out;
+	path[0].p_hdr = ext_inode_hdr(inode);
+	ret = ext4_ext_check(inode, path[0].p_hdr, depth, 0);
+	if (ret)
+		goto out;
+	path[0].p_idx = EXT_FIRST_INDEX(path[0].p_hdr);
+	while (i >= 0) {
+		/*
+		 * If this is a leaf block or we've reached the end of
+		 * the index block, go up
+		 */
+		if ((i == depth) ||
+		    path[i].p_idx > EXT_LAST_INDEX(path[i].p_hdr)) {
+			brelse(path[i].p_bh);
+			path[i].p_bh = NULL;
+			i--;
+			continue;
+		}
+		bh = read_extent_tree_block(inode,
+					    ext4_idx_pblock(path[i].p_idx++),
+					    depth - i - 1,
+					    EXT4_EX_FORCE_CACHE);
+		if (IS_ERR(bh)) {
+			ret = PTR_ERR(bh);
+			break;
+		}
+		i++;
+		path[i].p_bh = bh;
+		path[i].p_hdr = ext_block_hdr(bh);
+		path[i].p_idx = EXT_FIRST_INDEX(path[i].p_hdr);
+	}
+	ext4_set_inode_state(inode, EXT4_STATE_EXT_PRECACHED);
+out:
+	up_read(&ei->i_data_sem);
+	ext4_ext_drop_refs(path);
+	kfree(path);
+	return ret;
+}
 
 #ifdef EXT_DEBUG
 static void ext4_ext_show_path(struct inode *inode, struct ext4_ext_path *path)
···
 		error = ext4_inline_data_fiemap(inode, fieinfo, &has_inline);
 
 		if (has_inline)
+			return error;
+	}
+
+	if (fieinfo->fi_flags & FIEMAP_FLAG_CACHE) {
+		error = ext4_ext_precache(inode);
+		if (error)
 			return error;
 	}
 

fs/ext4/extents_status.c

···
 	write_lock(&EXT4_I(inode)->i_es_lock);
 
 	es = __es_tree_search(&EXT4_I(inode)->i_es_tree.root, lblk);
-	if (es && ((es->es_lblk <= lblk) || (es->es_lblk <= end)))
-		goto out;
-
-	__es_insert_extent(inode, &newes);
-out:
+	if (!es || es->es_lblk > end)
+		__es_insert_extent(inode, &newes);
 	write_unlock(&EXT4_I(inode)->i_es_lock);
 }
 
···
 	eia = list_entry(a, struct ext4_inode_info, i_es_lru);
 	eib = list_entry(b, struct ext4_inode_info, i_es_lru);
 
+	if (ext4_test_inode_state(&eia->vfs_inode, EXT4_STATE_EXT_PRECACHED) &&
+	    !ext4_test_inode_state(&eib->vfs_inode, EXT4_STATE_EXT_PRECACHED))
+		return 1;
+	if (!ext4_test_inode_state(&eia->vfs_inode, EXT4_STATE_EXT_PRECACHED) &&
+	    ext4_test_inode_state(&eib->vfs_inode, EXT4_STATE_EXT_PRECACHED))
+		return -1;
 	if (eia->i_touch_when == eib->i_touch_when)
 		return 0;
 	if (time_after(eia->i_touch_when, eib->i_touch_when))
···
 {
 	struct ext4_inode_info *ei;
 	struct list_head *cur, *tmp;
-	LIST_HEAD(skiped);
+	LIST_HEAD(skipped);
 	int ret, nr_shrunk = 0;
+	int retried = 0, skip_precached = 1, nr_skipped = 0;
 
 	spin_lock(&sbi->s_es_lru_lock);
 
-	/*
-	 * If the inode that is at the head of LRU list is newer than
-	 * last_sorted time, that means that we need to sort this list.
-	 */
-	ei = list_first_entry(&sbi->s_es_lru, struct ext4_inode_info, i_es_lru);
-	if (sbi->s_es_last_sorted < ei->i_touch_when) {
-		list_sort(NULL, &sbi->s_es_lru, ext4_inode_touch_time_cmp);
-		sbi->s_es_last_sorted = jiffies;
-	}
-
+retry:
 	list_for_each_safe(cur, tmp, &sbi->s_es_lru) {
 		/*
 		 * If we have already reclaimed all extents from extent
···
 
 		ei = list_entry(cur, struct ext4_inode_info, i_es_lru);
 
-		/* Skip the inode that is newer than the last_sorted time */
-		if (sbi->s_es_last_sorted < ei->i_touch_when) {
-			list_move_tail(cur, &skiped);
+		/*
+		 * Skip the inode that is newer than the last_sorted
+		 * time.  Normally we try hard to avoid shrinking
+		 * precached inodes, but we will as a last resort.
+		 */
+		if ((sbi->s_es_last_sorted < ei->i_touch_when) ||
+		    (skip_precached && ext4_test_inode_state(&ei->vfs_inode,
+						EXT4_STATE_EXT_PRECACHED))) {
+			nr_skipped++;
+			list_move_tail(cur, &skipped);
 			continue;
 		}
 
···
 	}
 
 	/* Move the newer inodes into the tail of the LRU list. */
-	list_splice_tail(&skiped, &sbi->s_es_lru);
+	list_splice_tail(&skipped, &sbi->s_es_lru);
+	INIT_LIST_HEAD(&skipped);
+
+	/*
+	 * If we skipped any inodes, and we weren't able to make any
+	 * forward progress, sort the list and try again.
+	 */
+	if ((nr_shrunk == 0) && nr_skipped && !retried) {
+		retried++;
+		list_sort(NULL, &sbi->s_es_lru, ext4_inode_touch_time_cmp);
+		sbi->s_es_last_sorted = jiffies;
+		ei = list_first_entry(&sbi->s_es_lru, struct ext4_inode_info,
+				      i_es_lru);
+		/*
+		 * If there are no non-precached inodes left on the
+		 * list, start releasing precached extents.
+		 */
+		if (ext4_test_inode_state(&ei->vfs_inode,
+					  EXT4_STATE_EXT_PRECACHED))
+			skip_precached = 0;
+		goto retry;
+	}
+
 	spin_unlock(&sbi->s_es_lru_lock);
 
 	if (locked_ei && nr_shrunk == 0)
-		nr_shrunk = __es_try_to_reclaim_extents(ei, nr_to_scan);
+		nr_shrunk = __es_try_to_reclaim_extents(locked_ei, nr_to_scan);
 
 	return nr_shrunk;
 }
···
 	struct rb_node *node;
 	struct extent_status *es;
 	int nr_shrunk = 0;
+	static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
+				      DEFAULT_RATELIMIT_BURST);
 
 	if (ei->i_es_lru_nr == 0)
 		return 0;
+
+	if (ext4_test_inode_state(inode, EXT4_STATE_EXT_PRECACHED) &&
+	    __ratelimit(&_rs))
+		ext4_warning(inode->i_sb, "forced shrink of precached extents");
 
 	node = rb_first(&tree->root);
 	while (node != NULL) {

fs/ext4/ioctl.c

···
 
 		return 0;
 	}
+	case EXT4_IOC_PRECACHE_EXTENTS:
+		return ext4_ext_precache(inode);
 
 	default:
 		return -ENOTTY;
···
 	case EXT4_IOC_MOVE_EXT:
 	case FITRIM:
 	case EXT4_IOC_RESIZE_FS:
+	case EXT4_IOC_PRECACHE_EXTENTS:
 		break;
 	default:
 		return -ENOIOCTLCMD;

include/uapi/linux/fiemap.h

···
 
 #define FIEMAP_FLAG_SYNC	0x00000001 /* sync file data before map */
 #define FIEMAP_FLAG_XATTR	0x00000002 /* map extended attribute tree */
+#define FIEMAP_FLAG_CACHE	0x00000004 /* request caching of the extents */
 
 #define FIEMAP_FLAGS_COMPAT	(FIEMAP_FLAG_SYNC | FIEMAP_FLAG_XATTR)