Btrfs: fix deadlock on metadata reservation when evicting a inode
authorMiao Xie <miaox@cn.fujitsu.com>
Fri, 18 Nov 2011 09:43:00 +0000 (17:43 +0800)
committerDavid Sterba <dsterba@suse.cz>
Wed, 30 Nov 2011 17:46:03 +0000 (18:46 +0100)
When I ran the xfstests, I found the test tasks was blocked on meta-data
reservation.

By debugging, I found the reason of this bug:
   start transaction
        |
v
   reserve meta-data space
|
v
   flush delay allocation -> iput inode -> evict inode
^ |
| v
   wait for delay allocation flush <- reserve meta-data space

And besides that, the flush on evicting inode will block the thread, which
is reclaiming the memory, and make oom happen easily.

Fix this bug by skipping the flush step when evicting inode.

Signed-off-by: Miao Xie <miaox@cn.fujitsu.com>
fs/btrfs/ctree.h
fs/btrfs/extent-tree.c
fs/btrfs/inode.c

index 04a5dfcee5a1fd278035bdbc36e86be4b2dd0a45..50634abef9b4a51336e95b8bc2aafb4edb7844a0 100644 (file)
@@ -2369,6 +2369,9 @@ int btrfs_block_rsv_check(struct btrfs_root *root,
 int btrfs_block_rsv_refill(struct btrfs_root *root,
                          struct btrfs_block_rsv *block_rsv,
                          u64 min_reserved);
+int btrfs_block_rsv_refill_noflush(struct btrfs_root *root,
+                                  struct btrfs_block_rsv *block_rsv,
+                                  u64 min_reserved);
 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
                            struct btrfs_block_rsv *dst_rsv,
                            u64 num_bytes);
index 5d86877f10e1ca90074a71ef50c9d051730986c3..b7e5f6898d07c00cadef581ccefb967acc13dcae 100644 (file)
@@ -3887,9 +3887,9 @@ int btrfs_block_rsv_check(struct btrfs_root *root,
        return ret;
 }
 
-int btrfs_block_rsv_refill(struct btrfs_root *root,
-                         struct btrfs_block_rsv *block_rsv,
-                         u64 min_reserved)
+static inline int __btrfs_block_rsv_refill(struct btrfs_root *root,
+                                          struct btrfs_block_rsv *block_rsv,
+                                          u64 min_reserved, int flush)
 {
        u64 num_bytes = 0;
        int ret = -ENOSPC;
@@ -3908,7 +3908,7 @@ int btrfs_block_rsv_refill(struct btrfs_root *root,
        if (!ret)
                return 0;
 
-       ret = reserve_metadata_bytes(root, block_rsv, num_bytes, 1);
+       ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
        if (!ret) {
                block_rsv_add_bytes(block_rsv, num_bytes, 0);
                return 0;
@@ -3917,6 +3917,20 @@ int btrfs_block_rsv_refill(struct btrfs_root *root,
        return ret;
 }
 
+int btrfs_block_rsv_refill(struct btrfs_root *root,
+                          struct btrfs_block_rsv *block_rsv,
+                          u64 min_reserved)
+{
+       return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 1);
+}
+
+int btrfs_block_rsv_refill_noflush(struct btrfs_root *root,
+                                  struct btrfs_block_rsv *block_rsv,
+                                  u64 min_reserved)
+{
+       return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 0);
+}
+
 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
                            struct btrfs_block_rsv *dst_rsv,
                            u64 num_bytes)
index 8ad26b135a1c0fa7f88890464d6c63b5accd0d7e..c5ccec23984c7abbcf340c359eeace1e5a447502 100644 (file)
@@ -3490,7 +3490,7 @@ void btrfs_evict_inode(struct inode *inode)
         * doing the truncate.
         */
        while (1) {
-               ret = btrfs_block_rsv_refill(root, rsv, min_size);
+               ret = btrfs_block_rsv_refill_noflush(root, rsv, min_size);
 
                /*
                 * Try and steal from the global reserve since we will