Btrfs: make fiemap not blow when you have lots of snapshots
authorJosef Bacik <jbacik@fb.com>
Wed, 10 Sep 2014 20:20:45 +0000 (16:20 -0400)
committerChris Mason <clm@fb.com>
Wed, 17 Sep 2014 20:38:24 +0000 (13:38 -0700)
We have been iterating all references for each extent we have in a file when we
do fiemap to see if it is shared.  This is fine when you have a few clones or a
few snapshots, but when you have 5k snapshots suddenly fiemap just sits there
and stares at you.  So add btrfs_check_shared which will use the backref walking
code but will short circuit as soon as it finds a root or inode that doesn't
match the one we currently have.  This makes fiemap on my testbox go from
looking at me blankly for a day to spitting out actual output in a reasonable
amount of time.  Thanks,

Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
fs/btrfs/backref.c
fs/btrfs/backref.h
fs/btrfs/extent_io.c

index 4de97926939eb3fb3220140dc3316f4bccb05bb1..6829dc5aa657aa8227436bed539a362bc32ee596 100644 (file)
@@ -25,6 +25,9 @@
 #include "delayed-ref.h"
 #include "locking.h"
 
+/* Just an arbitrary number so we can be sure this happened */
+#define BACKREF_FOUND_SHARED 6
+
 struct extent_inode_elem {
        u64 inum;
        u64 offset;
@@ -377,7 +380,8 @@ out:
 static int __resolve_indirect_refs(struct btrfs_fs_info *fs_info,
                                   struct btrfs_path *path, u64 time_seq,
                                   struct list_head *head,
-                                  const u64 *extent_item_pos, u64 total_refs)
+                                  const u64 *extent_item_pos, u64 total_refs,
+                                  u64 root_objectid)
 {
        int err;
        int ret = 0;
@@ -402,6 +406,10 @@ static int __resolve_indirect_refs(struct btrfs_fs_info *fs_info,
                        continue;
                if (ref->count == 0)
                        continue;
+               if (root_objectid && ref->root_id != root_objectid) {
+                       ret = BACKREF_FOUND_SHARED;
+                       goto out;
+               }
                err = __resolve_indirect_ref(fs_info, path, time_seq, ref,
                                             parents, extent_item_pos,
                                             total_refs);
@@ -561,7 +569,8 @@ static void __merge_refs(struct list_head *head, int mode)
  * smaller or equal that seq to the list
  */
 static int __add_delayed_refs(struct btrfs_delayed_ref_head *head, u64 seq,
-                             struct list_head *prefs, u64 *total_refs)
+                             struct list_head *prefs, u64 *total_refs,
+                             u64 inum)
 {
        struct btrfs_delayed_extent_op *extent_op = head->extent_op;
        struct rb_node *n = &head->node.rb_node;
@@ -625,6 +634,16 @@ static int __add_delayed_refs(struct btrfs_delayed_ref_head *head, u64 seq,
                        key.objectid = ref->objectid;
                        key.type = BTRFS_EXTENT_DATA_KEY;
                        key.offset = ref->offset;
+
+                       /*
+                        * Found a inum that doesn't match our known inum, we
+                        * know it's shared.
+                        */
+                       if (inum && ref->objectid != inum) {
+                               ret = BACKREF_FOUND_SHARED;
+                               break;
+                       }
+
                        ret = __add_prelim_ref(prefs, ref->root, &key, 0, 0,
                                               node->bytenr,
                                               node->ref_mod * sgn, GFP_ATOMIC);
@@ -659,7 +678,7 @@ static int __add_delayed_refs(struct btrfs_delayed_ref_head *head, u64 seq,
 static int __add_inline_refs(struct btrfs_fs_info *fs_info,
                             struct btrfs_path *path, u64 bytenr,
                             int *info_level, struct list_head *prefs,
-                            u64 *total_refs)
+                            u64 *total_refs, u64 inum)
 {
        int ret = 0;
        int slot;
@@ -744,6 +763,12 @@ static int __add_inline_refs(struct btrfs_fs_info *fs_info,
                                                                      dref);
                        key.type = BTRFS_EXTENT_DATA_KEY;
                        key.offset = btrfs_extent_data_ref_offset(leaf, dref);
+
+                       if (inum && key.objectid != inum) {
+                               ret = BACKREF_FOUND_SHARED;
+                               break;
+                       }
+
                        root = btrfs_extent_data_ref_root(leaf, dref);
                        ret = __add_prelim_ref(prefs, root, &key, 0, 0,
                                               bytenr, count, GFP_NOFS);
@@ -765,7 +790,7 @@ static int __add_inline_refs(struct btrfs_fs_info *fs_info,
  */
 static int __add_keyed_refs(struct btrfs_fs_info *fs_info,
                            struct btrfs_path *path, u64 bytenr,
-                           int info_level, struct list_head *prefs)
+                           int info_level, struct list_head *prefs, u64 inum)
 {
        struct btrfs_root *extent_root = fs_info->extent_root;
        int ret;
@@ -827,6 +852,12 @@ static int __add_keyed_refs(struct btrfs_fs_info *fs_info,
                                                                      dref);
                        key.type = BTRFS_EXTENT_DATA_KEY;
                        key.offset = btrfs_extent_data_ref_offset(leaf, dref);
+
+                       if (inum && key.objectid != inum) {
+                               ret = BACKREF_FOUND_SHARED;
+                               break;
+                       }
+
                        root = btrfs_extent_data_ref_root(leaf, dref);
                        ret = __add_prelim_ref(prefs, root, &key, 0, 0,
                                               bytenr, count, GFP_NOFS);
@@ -854,7 +885,8 @@ static int __add_keyed_refs(struct btrfs_fs_info *fs_info,
 static int find_parent_nodes(struct btrfs_trans_handle *trans,
                             struct btrfs_fs_info *fs_info, u64 bytenr,
                             u64 time_seq, struct ulist *refs,
-                            struct ulist *roots, const u64 *extent_item_pos)
+                            struct ulist *roots, const u64 *extent_item_pos,
+                            u64 root_objectid, u64 inum)
 {
        struct btrfs_key key;
        struct btrfs_path *path;
@@ -929,7 +961,8 @@ again:
                        }
                        spin_unlock(&delayed_refs->lock);
                        ret = __add_delayed_refs(head, time_seq,
-                                                &prefs_delayed, &total_refs);
+                                                &prefs_delayed, &total_refs,
+                                                inum);
                        mutex_unlock(&head->mutex);
                        if (ret)
                                goto out;
@@ -951,11 +984,11 @@ again:
                     key.type == BTRFS_METADATA_ITEM_KEY)) {
                        ret = __add_inline_refs(fs_info, path, bytenr,
                                                &info_level, &prefs,
-                                               &total_refs);
+                                               &total_refs, inum);
                        if (ret)
                                goto out;
                        ret = __add_keyed_refs(fs_info, path, bytenr,
-                                              info_level, &prefs);
+                                              info_level, &prefs, inum);
                        if (ret)
                                goto out;
                }
@@ -971,7 +1004,8 @@ again:
        __merge_refs(&prefs, 1);
 
        ret = __resolve_indirect_refs(fs_info, path, time_seq, &prefs,
-                                     extent_item_pos, total_refs);
+                                     extent_item_pos, total_refs,
+                                     root_objectid);
        if (ret)
                goto out;
 
@@ -981,6 +1015,11 @@ again:
                ref = list_first_entry(&prefs, struct __prelim_ref, list);
                WARN_ON(ref->count < 0);
                if (roots && ref->count && ref->root_id && ref->parent == 0) {
+                       if (root_objectid && ref->root_id != root_objectid) {
+                               ret = BACKREF_FOUND_SHARED;
+                               goto out;
+                       }
+
                        /* no parent == root of tree */
                        ret = ulist_add(roots, ref->root_id, 0, GFP_NOFS);
                        if (ret < 0)
@@ -1087,7 +1126,7 @@ static int btrfs_find_all_leafs(struct btrfs_trans_handle *trans,
                return -ENOMEM;
 
        ret = find_parent_nodes(trans, fs_info, bytenr,
-                               time_seq, *leafs, NULL, extent_item_pos);
+                               time_seq, *leafs, NULL, extent_item_pos, 0, 0);
        if (ret < 0 && ret != -ENOENT) {
                free_leaf_list(*leafs);
                return ret;
@@ -1130,7 +1169,7 @@ static int __btrfs_find_all_roots(struct btrfs_trans_handle *trans,
        ULIST_ITER_INIT(&uiter);
        while (1) {
                ret = find_parent_nodes(trans, fs_info, bytenr,
-                                       time_seq, tmp, *roots, NULL);
+                                       time_seq, tmp, *roots, NULL, 0, 0);
                if (ret < 0 && ret != -ENOENT) {
                        ulist_free(tmp);
                        ulist_free(*roots);
@@ -1161,6 +1200,54 @@ int btrfs_find_all_roots(struct btrfs_trans_handle *trans,
        return ret;
 }
 
+int btrfs_check_shared(struct btrfs_trans_handle *trans,
+                      struct btrfs_fs_info *fs_info, u64 root_objectid,
+                      u64 inum, u64 bytenr)
+{
+       struct ulist *tmp = NULL;
+       struct ulist *roots = NULL;
+       struct ulist_iterator uiter;
+       struct ulist_node *node;
+       struct seq_list elem = {};
+       int ret = 0;
+
+       tmp = ulist_alloc(GFP_NOFS);
+       roots = ulist_alloc(GFP_NOFS);
+       if (!tmp || !roots) {
+               ulist_free(tmp);
+               ulist_free(roots);
+               return -ENOMEM;
+       }
+
+       if (trans)
+               btrfs_get_tree_mod_seq(fs_info, &elem);
+       else
+               down_read(&fs_info->commit_root_sem);
+       ULIST_ITER_INIT(&uiter);
+       while (1) {
+               ret = find_parent_nodes(trans, fs_info, bytenr, elem.seq, tmp,
+                                       roots, NULL, root_objectid, inum);
+               if (ret == BACKREF_FOUND_SHARED) {
+                       ret = 1;
+                       break;
+               }
+               if (ret < 0 && ret != -ENOENT)
+                       break;
+               node = ulist_next(tmp, &uiter);
+               if (!node)
+                       break;
+               bytenr = node->val;
+               cond_resched();
+       }
+       if (trans)
+               btrfs_put_tree_mod_seq(fs_info, &elem);
+       else
+               up_read(&fs_info->commit_root_sem);
+       ulist_free(tmp);
+       ulist_free(roots);
+       return ret;
+}
+
 /*
  * this makes the path point to (inum INODE_ITEM ioff)
  */
index 86fc20fec28243f0f594a40a17da64a5d79da30f..2a1ac6bfc724637f3a80ac15c6ce6237dc7998f0 100644 (file)
@@ -71,6 +71,9 @@ int btrfs_find_one_extref(struct btrfs_root *root, u64 inode_objectid,
                          u64 start_off, struct btrfs_path *path,
                          struct btrfs_inode_extref **ret_extref,
                          u64 *found_off);
+int btrfs_check_shared(struct btrfs_trans_handle *trans,
+                      struct btrfs_fs_info *fs_info, u64 root_objectid,
+                      u64 inum, u64 bytenr);
 
 int __init btrfs_prelim_ref_init(void);
 void btrfs_prelim_ref_exit(void);
index d70feb54506678cd45855f300b690382ec57dfea..d5e71d4646dd65c6fbf83c23602403f0eefeddec 100644 (file)
@@ -4169,19 +4169,6 @@ static struct extent_map *get_extent_skip_holes(struct inode *inode,
        return NULL;
 }
 
-static noinline int count_ext_ref(u64 inum, u64 offset, u64 root_id, void *ctx)
-{
-       unsigned long cnt = *((unsigned long *)ctx);
-
-       cnt++;
-       *((unsigned long *)ctx) = cnt;
-
-       /* Now we're sure that the extent is shared. */
-       if (cnt > 1)
-               return 1;
-       return 0;
-}
-
 int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
                __u64 start, __u64 len, get_extent_t *get_extent)
 {
@@ -4198,6 +4185,7 @@ int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
        struct extent_map *em = NULL;
        struct extent_state *cached_state = NULL;
        struct btrfs_path *path;
+       struct btrfs_root *root = BTRFS_I(inode)->root;
        int end = 0;
        u64 em_start = 0;
        u64 em_len = 0;
@@ -4218,8 +4206,8 @@ int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
         * lookup the last file extent.  We're not using i_size here
         * because there might be preallocation past i_size
         */
-       ret = btrfs_lookup_file_extent(NULL, BTRFS_I(inode)->root,
-                                      path, btrfs_ino(inode), -1, 0);
+       ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(inode), -1,
+                                      0);
        if (ret < 0) {
                btrfs_free_path(path);
                return ret;
@@ -4312,25 +4300,27 @@ int extent_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
                } else if (em->block_start == EXTENT_MAP_DELALLOC) {
                        flags |= (FIEMAP_EXTENT_DELALLOC |
                                  FIEMAP_EXTENT_UNKNOWN);
-               } else {
-                       unsigned long ref_cnt = 0;
+               } else if (fieinfo->fi_extents_max) {
+                       u64 bytenr = em->block_start -
+                               (em->start - em->orig_start);
 
                        disko = em->block_start + offset_in_extent;
 
                        /*
                         * As btrfs supports shared space, this information
                         * can be exported to userspace tools via
-                        * flag FIEMAP_EXTENT_SHARED.
+                        * flag FIEMAP_EXTENT_SHARED.  If fi_extents_max == 0
+                        * then we're just getting a count and we can skip the
+                        * lookup stuff.
                         */
-                       ret = iterate_inodes_from_logical(
-                                       em->block_start,
-                                       BTRFS_I(inode)->root->fs_info,
-                                       path, count_ext_ref, &ref_cnt);
-                       if (ret < 0 && ret != -ENOENT)
+                       ret = btrfs_check_shared(NULL, root->fs_info,
+                                                root->objectid,
+                                                btrfs_ino(inode), bytenr);
+                       if (ret < 0)
                                goto out_free;
-
-                       if (ref_cnt > 1)
+                       if (ret)
                                flags |= FIEMAP_EXTENT_SHARED;
+                       ret = 0;
                }
                if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags))
                        flags |= FIEMAP_EXTENT_ENCODED;