Btrfs: avoid visiting all extent items when cloning a range
authorFilipe Manana <fdmanana@gmail.com>
Sat, 31 May 2014 01:31:05 +0000 (02:31 +0100)
committerChris Mason <clm@fb.com>
Tue, 10 Jun 2014 00:21:04 +0000 (17:21 -0700)
When cloning a range of a file, we were visiting all the extent items in
the btree that belong to our source inode. We don't need to visit those
extent items that don't overlap the range we are cloning, as doing so only
makes us waste time and do unnecessary btree navigations (btrfs_next_leaf)
for inodes that have a large number of file extent items in the btree.

Signed-off-by: Filipe David Borba Manana <fdmanana@gmail.com>
Signed-off-by: Chris Mason <clm@fb.com>
fs/btrfs/ioctl.c

index 38f2169b73a4e73dfbe36125ee79f799691d9595..f0b4237cedfc52193c7146fcc71e8ab305982ef2 100644 (file)
@@ -3020,7 +3020,7 @@ static int btrfs_clone(struct inode *src, struct inode *inode,
        /* clone data */
        key.objectid = btrfs_ino(src);
        key.type = BTRFS_EXTENT_DATA_KEY;
-       key.offset = 0;
+       key.offset = off;
 
        while (1) {
                /*
@@ -3032,6 +3032,17 @@ static int btrfs_clone(struct inode *src, struct inode *inode,
                                0, 0);
                if (ret < 0)
                        goto out;
+               /*
+                * First search, if no extent item that starts at offset off was
+                * found but the previous item is an extent item, it's possible
+                * it might overlap our target range, therefore process it.
+                */
+               if (key.offset == off && ret > 0 && path->slots[0] > 0) {
+                       btrfs_item_key_to_cpu(path->nodes[0], &key,
+                                             path->slots[0] - 1);
+                       if (key.type == BTRFS_EXTENT_DATA_KEY)
+                               path->slots[0]--;
+               }
 
                nritems = btrfs_header_nritems(path->nodes[0]);
 process_slot:
@@ -3081,10 +3092,16 @@ process_slot:
                                                                    extent);
                        }
 
-                       if (key.offset + datal <= off ||
-                           key.offset >= off + len - 1) {
+                       /*
+                        * The first search might have left us at an extent
+                        * item that ends before our target range's start, can
+                        * happen if we have holes and NO_HOLES feature enabled.
+                        */
+                       if (key.offset + datal <= off) {
                                path->slots[0]++;
                                goto process_slot;
+                       } else if (key.offset >= off + len) {
+                               break;
                        }
 
                        size = btrfs_item_size_nr(leaf, slot);
@@ -3291,6 +3308,8 @@ process_slot:
                                goto out;
                        }
                        ret = btrfs_end_transaction(trans, root);
+                       if (new_key.offset + datal >= destoff + len)
+                               break;
                }
                btrfs_release_path(path);
                key.offset++;
@@ -3298,7 +3317,6 @@ process_slot:
        ret = 0;
 
 out:
-       btrfs_release_path(path);
        btrfs_free_path(path);
        vfree(buf);
        return ret;