OSDN Git Service

btrfs: Handle delalloc error correctly to avoid ordered extent hang
authorQu Wenruo <quwenruo@cn.fujitsu.com>
Wed, 8 Mar 2017 02:25:52 +0000 (10:25 +0800)
committerFilipe Manana <fdmanana@suse.com>
Wed, 26 Apr 2017 15:27:21 +0000 (16:27 +0100)
[BUG]
If run_delalloc_range() returns error and there is already some ordered
extents created, btrfs will be hanged with the following backtrace:

Call Trace:
 __schedule+0x2d4/0xae0
 schedule+0x3d/0x90
 btrfs_start_ordered_extent+0x160/0x200 [btrfs]
 ? wake_atomic_t_function+0x60/0x60
 btrfs_run_ordered_extent_work+0x25/0x40 [btrfs]
 btrfs_scrubparity_helper+0x1c1/0x620 [btrfs]
 btrfs_flush_delalloc_helper+0xe/0x10 [btrfs]
 process_one_work+0x2af/0x720
 ? process_one_work+0x22b/0x720
 worker_thread+0x4b/0x4f0
 kthread+0x10f/0x150
 ? process_one_work+0x720/0x720
 ? kthread_create_on_node+0x40/0x40
 ret_from_fork+0x2e/0x40

[CAUSE]

|<------------------ delalloc range --------------------------->|
| OE 1 | OE 2 | ... | OE n |
|<>|                       |<---------- cleanup range --------->|
 ||
 \_=> First page handled by end_extent_writepage() in __extent_writepage()

The problem is caused by error handler of run_delalloc_range(), which
doesn't handle any created ordered extents, leaving them waiting on
btrfs_finish_ordered_io() to finish.

However after run_delalloc_range() returns error, __extent_writepage()
won't submit bio, so btrfs_writepage_end_io_hook() won't be triggered
except the first page, and btrfs_finish_ordered_io() won't be triggered
for created ordered extents either.

So OE 2~n will hang forever, and if OE 1 is larger than one page, it
will also hang.

[FIX]
Introduce btrfs_cleanup_ordered_extents() function to cleanup created
ordered extents and finish them manually.

The function is based on existing
btrfs_endio_direct_write_update_ordered() function, and modify it to
act just like btrfs_writepage_endio_hook() but handles specified range
other than one page.

After fix, delalloc error will be handled like:

|<------------------ delalloc range --------------------------->|
| OE 1 | OE 2 | ... | OE n |
|<>|<--------  ----------->|<------ old error handler --------->|
 ||          ||
 ||          \_=> Cleaned up by cleanup_ordered_extents()
 \_=> First page handled by end_extent_writepage() in __extent_writepage()

Signed-off-by: Qu Wenruo <quwenruo@cn.fujitsu.com>
Reviewed-by: Filipe Manana <fdmanana@suse.com>
Reviewed-by: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: Filipe Manana <fdmanana@suse.com>
fs/btrfs/inode.c

index 844bb89..b8b2a3c 100644 (file)
@@ -115,6 +115,31 @@ static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
                                       u64 ram_bytes, int compress_type,
                                       int type);
 
+static void __endio_write_update_ordered(struct inode *inode,
+                                        const u64 offset, const u64 bytes,
+                                        const bool uptodate);
+
+/*
+ * Cleanup all submitted ordered extents in specified range to handle errors
+ * from the fill_dellaloc() callback.
+ *
+ * NOTE: caller must ensure that when an error happens, it can not call
+ * extent_clear_unlock_delalloc() to clear both the bits EXTENT_DO_ACCOUNTING
+ * and EXTENT_DELALLOC simultaneously, because that causes the reserved metadata
+ * to be released, which we want to happen only when finishing the ordered
+ * extent (btrfs_finish_ordered_io()). Also note that the caller of the
+ * fill_delalloc() callback already does proper cleanup for the first page of
+ * the range, that is, it invokes the callback writepage_end_io_hook() for the
+ * range of the first page.
+ */
+static inline void btrfs_cleanup_ordered_extents(struct inode *inode,
+                                                const u64 offset,
+                                                const u64 bytes)
+{
+       return __endio_write_update_ordered(inode, offset + PAGE_SIZE,
+                                           bytes - PAGE_SIZE, false);
+}
+
 static int btrfs_dirty_inode(struct inode *inode);
 
 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
@@ -1536,6 +1561,8 @@ static int run_delalloc_range(struct inode *inode, struct page *locked_page,
                ret = cow_file_range_async(inode, locked_page, start, end,
                                           page_started, nr_written);
        }
+       if (ret)
+               btrfs_cleanup_ordered_extents(inode, start, end - start + 1);
        return ret;
 }
 
@@ -8154,17 +8181,26 @@ static void btrfs_endio_direct_read(struct bio *bio)
        bio_put(bio);
 }
 
-static void btrfs_endio_direct_write_update_ordered(struct inode *inode,
-                                                   const u64 offset,
-                                                   const u64 bytes,
-                                                   const int uptodate)
+static void __endio_write_update_ordered(struct inode *inode,
+                                        const u64 offset, const u64 bytes,
+                                        const bool uptodate)
 {
        struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
        struct btrfs_ordered_extent *ordered = NULL;
+       struct btrfs_workqueue *wq;
+       btrfs_work_func_t func;
        u64 ordered_offset = offset;
        u64 ordered_bytes = bytes;
        int ret;
 
+       if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
+               wq = fs_info->endio_freespace_worker;
+               func = btrfs_freespace_write_helper;
+       } else {
+               wq = fs_info->endio_write_workers;
+               func = btrfs_endio_write_helper;
+       }
+
 again:
        ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
                                                   &ordered_offset,
@@ -8173,9 +8209,8 @@ again:
        if (!ret)
                goto out_test;
 
-       btrfs_init_work(&ordered->work, btrfs_endio_write_helper,
-                       finish_ordered_fn, NULL, NULL);
-       btrfs_queue_work(fs_info->endio_write_workers, &ordered->work);
+       btrfs_init_work(&ordered->work, func, finish_ordered_fn, NULL, NULL);
+       btrfs_queue_work(wq, &ordered->work);
 out_test:
        /*
         * our bio might span multiple ordered extents.  If we haven't
@@ -8193,10 +8228,8 @@ static void btrfs_endio_direct_write(struct bio *bio)
        struct btrfs_dio_private *dip = bio->bi_private;
        struct bio *dio_bio = dip->dio_bio;
 
-       btrfs_endio_direct_write_update_ordered(dip->inode,
-                                               dip->logical_offset,
-                                               dip->bytes,
-                                               !bio->bi_error);
+       __endio_write_update_ordered(dip->inode, dip->logical_offset,
+                                    dip->bytes, !bio->bi_error);
 
        kfree(dip);
 
@@ -8557,10 +8590,10 @@ free_ordered:
                io_bio = NULL;
        } else {
                if (write)
-                       btrfs_endio_direct_write_update_ordered(inode,
+                       __endio_write_update_ordered(inode,
                                                file_offset,
                                                dio_bio->bi_iter.bi_size,
-                                               0);
+                                               false);
                else
                        unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
                              file_offset + dio_bio->bi_iter.bi_size - 1);
@@ -8695,11 +8728,11 @@ static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
                         */
                        if (dio_data.unsubmitted_oe_range_start <
                            dio_data.unsubmitted_oe_range_end)
-                               btrfs_endio_direct_write_update_ordered(inode,
+                               __endio_write_update_ordered(inode,
                                        dio_data.unsubmitted_oe_range_start,
                                        dio_data.unsubmitted_oe_range_end -
                                        dio_data.unsubmitted_oe_range_start,
-                                       0);
+                                       false);
                } else if (ret >= 0 && (size_t)ret < count)
                        btrfs_delalloc_release_space(inode, offset,
                                                     count - (size_t)ret);