Btrfs: fix wrong super generation mismatch when scrubbing supers
authorWang Shilong <wangsl.fnst@cn.fujitsu.com>
Wed, 4 Dec 2013 13:15:19 +0000 (21:15 +0800)
committerChris Mason <clm@fb.com>
Tue, 28 Jan 2014 21:19:52 +0000 (13:19 -0800)
We came a race condition when scrubbing superblocks, the story is:

In commiting transaction, we will update @last_trans_commited after
writting superblocks, if scrubber start after writting superblocks
and before updating @last_trans_commited, generation mismatch happens!

We fix this by checking @scrub_pause_req, and we won't start a srubber
until commiting transaction is finished.(after btrfs_scrub_continue()
finished.)

Reported-by: Sebastian Ochmann <ochmann@informatik.uni-bonn.de>
Signed-off-by: Wang Shilong <wangsl.fnst@cn.fujitsu.com>
Reviewed-by: Miao Xie <miaox@cn.fujitsu.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Chris Mason <clm@fb.com>
fs/btrfs/scrub.c

index e5481ae..6acb573 100644 (file)
@@ -256,6 +256,7 @@ static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
                            int mirror_num, u64 physical_for_dev_replace);
 static void copy_nocow_pages_worker(struct btrfs_work *work);
+static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
 
 
 static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
@@ -269,6 +270,16 @@ static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
        wake_up(&sctx->list_wait);
 }
 
+static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
+{
+       while (atomic_read(&fs_info->scrub_pause_req)) {
+               mutex_unlock(&fs_info->scrub_lock);
+               wait_event(fs_info->scrub_pause_wait,
+                  atomic_read(&fs_info->scrub_pause_req) == 0);
+               mutex_lock(&fs_info->scrub_lock);
+       }
+}
+
 /*
  * used for workers that require transaction commits (i.e., for the
  * NOCOW case)
@@ -2310,14 +2321,10 @@ static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
                btrfs_reada_wait(reada2);
 
        mutex_lock(&fs_info->scrub_lock);
-       while (atomic_read(&fs_info->scrub_pause_req)) {
-               mutex_unlock(&fs_info->scrub_lock);
-               wait_event(fs_info->scrub_pause_wait,
-                  atomic_read(&fs_info->scrub_pause_req) == 0);
-               mutex_lock(&fs_info->scrub_lock);
-       }
+       scrub_blocked_if_needed(fs_info);
        atomic_dec(&fs_info->scrubs_paused);
        mutex_unlock(&fs_info->scrub_lock);
+
        wake_up(&fs_info->scrub_pause_wait);
 
        /*
@@ -2357,15 +2364,12 @@ static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
                        atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
                        atomic_inc(&fs_info->scrubs_paused);
                        wake_up(&fs_info->scrub_pause_wait);
+
                        mutex_lock(&fs_info->scrub_lock);
-                       while (atomic_read(&fs_info->scrub_pause_req)) {
-                               mutex_unlock(&fs_info->scrub_lock);
-                               wait_event(fs_info->scrub_pause_wait,
-                                  atomic_read(&fs_info->scrub_pause_req) == 0);
-                               mutex_lock(&fs_info->scrub_lock);
-                       }
+                       scrub_blocked_if_needed(fs_info);
                        atomic_dec(&fs_info->scrubs_paused);
                        mutex_unlock(&fs_info->scrub_lock);
+
                        wake_up(&fs_info->scrub_pause_wait);
                }
 
@@ -2687,14 +2691,10 @@ int scrub_enumerate_chunks(struct scrub_ctx *sctx,
                           atomic_read(&sctx->workers_pending) == 0);
 
                mutex_lock(&fs_info->scrub_lock);
-               while (atomic_read(&fs_info->scrub_pause_req)) {
-                       mutex_unlock(&fs_info->scrub_lock);
-                       wait_event(fs_info->scrub_pause_wait,
-                          atomic_read(&fs_info->scrub_pause_req) == 0);
-                       mutex_lock(&fs_info->scrub_lock);
-               }
+               scrub_blocked_if_needed(fs_info);
                atomic_dec(&fs_info->scrubs_paused);
                mutex_unlock(&fs_info->scrub_lock);
+
                wake_up(&fs_info->scrub_pause_wait);
 
                btrfs_put_block_group(cache);
@@ -2906,7 +2906,13 @@ int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
        }
        sctx->readonly = readonly;
        dev->scrub_device = sctx;
+       mutex_unlock(&fs_info->fs_devices->device_list_mutex);
 
+       /*
+        * checking @scrub_pause_req here, we can avoid
+        * race between committing transaction and scrubbing.
+        */
+       scrub_blocked_if_needed(fs_info);
        atomic_inc(&fs_info->scrubs_running);
        mutex_unlock(&fs_info->scrub_lock);
 
@@ -2915,9 +2921,10 @@ int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
                 * by holding device list mutex, we can
                 * kick off writing super in log tree sync.
                 */
+               mutex_lock(&fs_info->fs_devices->device_list_mutex);
                ret = scrub_supers(sctx, dev);
+               mutex_unlock(&fs_info->fs_devices->device_list_mutex);
        }
-       mutex_unlock(&fs_info->fs_devices->device_list_mutex);
 
        if (!ret)
                ret = scrub_enumerate_chunks(sctx, dev, start, end,