@@ -3016,7 +3016,8 @@ static inline int btrfs_next_item(struct btrfs_root *root, struct btrfs_path *p)
int btrfs_leaf_free_space(struct extent_buffer *leaf);
int __must_check btrfs_drop_snapshot(struct btrfs_root *root,
struct btrfs_block_rsv *block_rsv,
- int update_ref, int for_reloc);
+ int update_ref, int for_reloc,
+ int pause_qgroup);
int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
struct btrfs_root *root,
struct extent_buffer *node,
@@ -1747,6 +1747,8 @@ static int cleaner_kthread(void *arg)
if (kthread_should_stop())
return 0;
if (!again) {
+ if (btrfs_quota_is_paused(fs_info))
+ btrfs_quota_resume(fs_info);
set_current_state(TASK_INTERRUPTIBLE);
schedule();
__set_current_state(TASK_RUNNING);
@@ -9236,7 +9236,7 @@ static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
*/
int btrfs_drop_snapshot(struct btrfs_root *root,
struct btrfs_block_rsv *block_rsv, int update_ref,
- int for_reloc)
+ int for_reloc, int pause_qgroup)
{
struct btrfs_fs_info *fs_info = root->fs_info;
struct btrfs_path *path;
@@ -9343,6 +9343,17 @@ int btrfs_drop_snapshot(struct btrfs_root *root,
}
}
+ /*
+ * If this subvolume is shared, and pretty high, we tend to pause
+ * qgroup to avoid qgroup extent records flood due to sudden subtree
+ * owner change happened in one transaction.
+ */
+ if (test_bit(BTRFS_FS_QUOTA_ENABLED, &fs_info->flags) &&
+ pause_qgroup && (root->root_key.offset ||
+ btrfs_root_last_snapshot(&root->root_item)) &&
+ btrfs_header_level(root->node))
+ btrfs_quota_pause(trans);
+
wc->restarted = test_bit(BTRFS_ROOT_DEAD_TREE, &root->state);
wc->level = level;
wc->shared_level = -1;
@@ -2190,14 +2190,14 @@ static int clean_dirty_subvols(struct reloc_control *rc)
root->reloc_root = NULL;
if (reloc_root) {
- ret2 = btrfs_drop_snapshot(reloc_root, NULL, 0, 1);
+ ret2 = btrfs_drop_snapshot(reloc_root, NULL, 0, 1, 0);
if (ret2 < 0 && !ret)
ret = ret2;
}
btrfs_put_fs_root(root);
} else {
/* Orphan reloc tree, just clean it up */
- ret2 = btrfs_drop_snapshot(root, NULL, 0, 1);
+ ret2 = btrfs_drop_snapshot(root, NULL, 0, 1, 0);
if (ret2 < 0 && !ret)
ret = ret2;
}
@@ -2317,9 +2317,9 @@ int btrfs_clean_one_deleted_snapshot(struct btrfs_fs_info *fs_info)
if (btrfs_header_backref_rev(root->node) <
BTRFS_MIXED_BACKREF_REV)
- ret = btrfs_drop_snapshot(root, NULL, 0, 0);
+ ret = btrfs_drop_snapshot(root, NULL, 0, 0, 1);
else
- ret = btrfs_drop_snapshot(root, NULL, 1, 0);
+ ret = btrfs_drop_snapshot(root, NULL, 1, 0, 1);
return (ret < 0) ? 0 : 1;
}
Btrfs qgroup has one big performance overhead for certain snapshot drop. Current btrfs_drop_snapshot() will try its best to find the highest shared node, and just drop one ref of that common node. This behavior is good for minimize extent tree modification, but a disaster for qgroup. Example: Root 300 Root 301 A B | \ / | | X | | / \ | C D / \ / \ E F G H / \ / \ / \ / \ I J K L M N O P In above case, if we're dropping root 301, btrfs_drop_snapshot() will only drop one ref for tree block C and D. But for qgroup, tree blocks E~P also have their owner changed, from 300, 301 to 300 only. Currently we use btrfs_qgroup_trace_subtree() to manually re-dirty tree block E~P. And since such ref change happens in one transaction for each ref drop, we can't split the overhead to different transactions. This could cause qgroup extent record flood, hugely damage performance or even cause OOM for too many qgroup extent records. This patch will try to solve it in a different method, since we can't split the overhead into different transactions, instead of doing such heavy work, we just pause qgroup at the very beginning of btrfs_drop_snapshot(). So later owner change won't trigger qgroup subtree trace, and after all subvolumes get removed, we just trigger a qgroup rescan to rebuild qgroup accounting. Also, to co-operate previous qgroup balance optimization, we don't need to pause qgroup to balance, thus introduce a new parameter for btrfs_drop_snapshot() to inform caller that we have other optimization to handle balance. Signed-off-by: Qu Wenruo <wqu@suse.com> --- fs/btrfs/ctree.h | 3 ++- fs/btrfs/disk-io.c | 2 ++ fs/btrfs/extent-tree.c | 13 ++++++++++++- fs/btrfs/relocation.c | 4 ++-- fs/btrfs/transaction.c | 4 ++-- 5 files changed, 20 insertions(+), 6 deletions(-)