Btrfs: fix race between start dirty bg cache writeout and bg deletion
authorFilipe Manana <fdmanana@suse.com>
Sat, 25 Apr 2015 17:29:16 +0000 (18:29 +0100)
committerChris Mason <clm@fb.com>
Sun, 26 Apr 2015 13:26:37 +0000 (06:26 -0700)
While running xfstests I ran into the following:

[20892.242791] ------------[ cut here ]------------
[20892.243776] WARNING: CPU: 0 PID: 13299 at fs/btrfs/super.c:260 __btrfs_abort_transaction+0x52/0x114 [btrfs]()
[20892.245874] BTRFS: Transaction aborted (error -2)
[20892.247329] Modules linked in: btrfs dm_snapshot dm_bufio dm_flakey dm_mod crc32c_generic xor raid6_pq nfsd auth_rpcgss oid_registry nfs_acl nfs lockd grace fscache sunrpc loop fuse$
[20892.258488] CPU: 0 PID: 13299 Comm: fsstress Tainted: G        W       4.0.0-rc5-btrfs-next-9+ #2
[20892.262011] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.7.5-0-ge51488c-20140602_164612-nilsson.home.kraxel.org 04/01/2014
[20892.264738]  0000000000000009 ffff880427f8bc18 ffffffff8142fa46 ffffffff8108b6a2
[20892.266244]  ffff880427f8bc68 ffff880427f8bc58 ffffffff81045ea5 ffff880427f8bc48
[20892.267761]  ffffffffa0509a6d 00000000fffffffe ffff8803545d6f40 ffffffffa05a15a0
[20892.269378] Call Trace:
[20892.269915]  [<ffffffff8142fa46>] dump_stack+0x4f/0x7b
[20892.271097]  [<ffffffff8108b6a2>] ? console_unlock+0x361/0x3ad
[20892.272173]  [<ffffffff81045ea5>] warn_slowpath_common+0xa1/0xbb
[20892.273386]  [<ffffffffa0509a6d>] ? __btrfs_abort_transaction+0x52/0x114 [btrfs]
[20892.274857]  [<ffffffff81045f05>] warn_slowpath_fmt+0x46/0x48
[20892.275851]  [<ffffffffa0509a6d>] __btrfs_abort_transaction+0x52/0x114 [btrfs]
[20892.277341]  [<ffffffffa0515e10>] write_one_cache_group+0x68/0xaf [btrfs]
[20892.278628]  [<ffffffffa052088a>] btrfs_start_dirty_block_groups+0x18d/0x29b [btrfs]
[20892.280191]  [<ffffffffa052f077>] btrfs_commit_transaction+0x130/0x9c9 [btrfs]
[20892.281781]  [<ffffffff8107d33d>] ? trace_hardirqs_on+0xd/0xf
[20892.282873]  [<ffffffffa054163b>] btrfs_sync_file+0x313/0x387 [btrfs]
[20892.284111]  [<ffffffff8117acad>] vfs_fsync_range+0x95/0xa4
[20892.285203]  [<ffffffff810e603f>] ? time_hardirqs_on+0x15/0x28
[20892.286290]  [<ffffffff8123960b>] ? trace_hardirqs_on_thunk+0x3a/0x3f
[20892.287469]  [<ffffffff8117acd8>] vfs_fsync+0x1c/0x1e
[20892.288412]  [<ffffffff8117ae54>] do_fsync+0x34/0x4e
[20892.289348]  [<ffffffff8117b07c>] SyS_fsync+0x10/0x14
[20892.290255]  [<ffffffff81435b32>] system_call_fastpath+0x12/0x17
[20892.291316] ---[ end trace 597f77e664245373 ]---
[20892.293955] BTRFS: error (device sdg) in write_one_cache_group:3184: errno=-2 No such entry
[20892.297390] BTRFS info (device sdg): forced readonly

This happens because in btrfs_start_dirty_block_groups() we splice the
transaction's list of dirty block groups into a local list and then we
keep extracting the first element of the list without holding the
cache_write_mutex mutex. This means that before we acquire that mutex
the first block group on the list might be removed by a conurrent task
running btrfs_remove_block_group(). So make sure we extract the first
element (and test the list emptyness) while holding that mutex.

Fixes: 1bbc621ef284 ("Btrfs: allow block group cache writeout
                      outside critical section in commit")

Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
fs/btrfs/extent-tree.c

index 1eef4ee01d1a3c7fd78e4546dd1bd2570f9c7783..fc0db9887c0ed3458c42b8c9fdeb853d2ee39d43 100644 (file)
@@ -3408,17 +3408,14 @@ int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans,
        int loops = 0;
 
        spin_lock(&cur_trans->dirty_bgs_lock);
-       if (!list_empty(&cur_trans->dirty_bgs)) {
-               list_splice_init(&cur_trans->dirty_bgs, &dirty);
+       if (list_empty(&cur_trans->dirty_bgs)) {
+               spin_unlock(&cur_trans->dirty_bgs_lock);
+               return 0;
        }
+       list_splice_init(&cur_trans->dirty_bgs, &dirty);
        spin_unlock(&cur_trans->dirty_bgs_lock);
 
 again:
-       if (list_empty(&dirty)) {
-               btrfs_free_path(path);
-               return 0;
-       }
-
        /*
         * make sure all the block groups on our dirty list actually
         * exist
@@ -3431,18 +3428,16 @@ again:
                        return -ENOMEM;
        }
 
+       /*
+        * cache_write_mutex is here only to save us from balance or automatic
+        * removal of empty block groups deleting this block group while we are
+        * writing out the cache
+        */
+       mutex_lock(&trans->transaction->cache_write_mutex);
        while (!list_empty(&dirty)) {
                cache = list_first_entry(&dirty,
                                         struct btrfs_block_group_cache,
                                         dirty_list);
-
-               /*
-                * cache_write_mutex is here only to save us from balance
-                * deleting this block group while we are writing out the
-                * cache
-                */
-               mutex_lock(&trans->transaction->cache_write_mutex);
-
                /*
                 * this can happen if something re-dirties a block
                 * group that is already under IO.  Just wait for it to
@@ -3495,7 +3490,6 @@ again:
                }
                if (!ret)
                        ret = write_one_cache_group(trans, root, path, cache);
-               mutex_unlock(&trans->transaction->cache_write_mutex);
 
                /* if its not on the io list, we need to put the block group */
                if (should_put)
@@ -3503,7 +3497,16 @@ again:
 
                if (ret)
                        break;
+
+               /*
+                * Avoid blocking other tasks for too long. It might even save
+                * us from writing caches for block groups that are going to be
+                * removed.
+                */
+               mutex_unlock(&trans->transaction->cache_write_mutex);
+               mutex_lock(&trans->transaction->cache_write_mutex);
        }
+       mutex_unlock(&trans->transaction->cache_write_mutex);
 
        /*
         * go through delayed refs for all the stuff we've just kicked off
@@ -3514,8 +3517,15 @@ again:
                loops++;
                spin_lock(&cur_trans->dirty_bgs_lock);
                list_splice_init(&cur_trans->dirty_bgs, &dirty);
+               /*
+                * dirty_bgs_lock protects us from concurrent block group
+                * deletes too (not just cache_write_mutex).
+                */
+               if (!list_empty(&dirty)) {
+                       spin_unlock(&cur_trans->dirty_bgs_lock);
+                       goto again;
+               }
                spin_unlock(&cur_trans->dirty_bgs_lock);
-               goto again;
        }
 
        btrfs_free_path(path);