[v3] jbd2: Fix potential data lost in recovering journal raced with synchronizing fs bdev

Message ID 20230919012525.1783108-1-chengzhihao1@huawei.com
State New
Headers
Series [v3] jbd2: Fix potential data lost in recovering journal raced with synchronizing fs bdev |

Commit Message

Zhihao Cheng Sept. 19, 2023, 1:25 a.m. UTC
  JBD2 makes sure journal data is fallen on fs device by sync_blockdev(),
however, other process could intercept the EIO information from bdev's
mapping, which leads journal recovering successful even EIO occurs during
data written back to fs device.

We found this problem in our product, iscsi + multipath is chosen for block
device of ext4. Unstable network may trigger kpartx to rescan partitions in
device mapper layer. Detailed process is shown as following:

  mount          kpartx          irq
jbd2_journal_recover
 do_one_pass
  memcpy(nbh->b_data, obh->b_data) // copy data to fs dev from journal
  mark_buffer_dirty // mark bh dirty
         vfs_read
	  generic_file_read_iter // dio
	   filemap_write_and_wait_range
	    __filemap_fdatawrite_range
	     do_writepages
	      block_write_full_folio
	       submit_bh_wbc
	            >>  EIO occurs in disk  <<
	                     end_buffer_async_write
			      mark_buffer_write_io_error
			       mapping_set_error
			        set_bit(AS_EIO, &mapping->flags) // set!
	    filemap_check_errors
	     test_and_clear_bit(AS_EIO, &mapping->flags) // clear!
 err2 = sync_blockdev
  filemap_write_and_wait
   filemap_check_errors
    test_and_clear_bit(AS_EIO, &mapping->flags) // false
 err2 = 0

Filesystem is mounted successfully even data from journal is failed written
into disk, and ext4/ocfs2 could become corrupted.

Fix it by comparing the wb_err state in fs block device before recovering
and after recovering.

Fetch a reproducer in [Link].

Link: https://bugzilla.kernel.org/show_bug.cgi?id=217888
Cc: stable@vger.kernel.org
Signed-off-by: Zhihao Cheng <chengzhihao1@huawei.com>
Signed-off-by: Zhang Yi <yi.zhang@huawei.com>
---
 v1->v3: Initialize wb_err. Untialized wb_err could be same with
	 mapping->wb_err(eg. EIO without ERRSEQ_SEEN). When EIO
	 occurs again, mapping->wb_err won't be changed, and wb_err
	 is still same with mapping->wb_err.
 fs/jbd2/recovery.c | 8 ++++++++
 1 file changed, 8 insertions(+)
  

Comments

Jan Kara Sept. 19, 2023, 10:14 a.m. UTC | #1
On Tue 19-09-23 09:25:25, Zhihao Cheng wrote:
> JBD2 makes sure journal data is fallen on fs device by sync_blockdev(),
> however, other process could intercept the EIO information from bdev's
> mapping, which leads journal recovering successful even EIO occurs during
> data written back to fs device.
> 
> We found this problem in our product, iscsi + multipath is chosen for block
> device of ext4. Unstable network may trigger kpartx to rescan partitions in
> device mapper layer. Detailed process is shown as following:
> 
>   mount          kpartx          irq
> jbd2_journal_recover
>  do_one_pass
>   memcpy(nbh->b_data, obh->b_data) // copy data to fs dev from journal
>   mark_buffer_dirty // mark bh dirty
>          vfs_read
> 	  generic_file_read_iter // dio
> 	   filemap_write_and_wait_range
> 	    __filemap_fdatawrite_range
> 	     do_writepages
> 	      block_write_full_folio
> 	       submit_bh_wbc
> 	            >>  EIO occurs in disk  <<
> 	                     end_buffer_async_write
> 			      mark_buffer_write_io_error
> 			       mapping_set_error
> 			        set_bit(AS_EIO, &mapping->flags) // set!
> 	    filemap_check_errors
> 	     test_and_clear_bit(AS_EIO, &mapping->flags) // clear!
>  err2 = sync_blockdev
>   filemap_write_and_wait
>    filemap_check_errors
>     test_and_clear_bit(AS_EIO, &mapping->flags) // false
>  err2 = 0
> 
> Filesystem is mounted successfully even data from journal is failed written
> into disk, and ext4/ocfs2 could become corrupted.
> 
> Fix it by comparing the wb_err state in fs block device before recovering
> and after recovering.
> 
> Fetch a reproducer in [Link].
> 
> Link: https://bugzilla.kernel.org/show_bug.cgi?id=217888
> Cc: stable@vger.kernel.org
> Signed-off-by: Zhihao Cheng <chengzhihao1@huawei.com>
> Signed-off-by: Zhang Yi <yi.zhang@huawei.com>
> ---
>  v1->v3: Initialize wb_err. Untialized wb_err could be same with
> 	 mapping->wb_err(eg. EIO without ERRSEQ_SEEN). When EIO
> 	 occurs again, mapping->wb_err won't be changed, and wb_err
> 	 is still same with mapping->wb_err.

Yeah, good catch. The patch still looks good to me. Feel free to add:

Reviewed-by: Jan Kara <jack@suse.cz>

								Honza


>  fs/jbd2/recovery.c | 8 ++++++++
>  1 file changed, 8 insertions(+)
> 
> diff --git a/fs/jbd2/recovery.c b/fs/jbd2/recovery.c
> index c269a7d29a46..5b771a3d8d9a 100644
> --- a/fs/jbd2/recovery.c
> +++ b/fs/jbd2/recovery.c
> @@ -289,6 +289,8 @@ int jbd2_journal_recover(journal_t *journal)
>  	journal_superblock_t *	sb;
>  
>  	struct recovery_info	info;
> +	errseq_t		wb_err;
> +	struct address_space	*mapping;
>  
>  	memset(&info, 0, sizeof(info));
>  	sb = journal->j_superblock;
> @@ -306,6 +308,9 @@ int jbd2_journal_recover(journal_t *journal)
>  		return 0;
>  	}
>  
> +	wb_err = 0;
> +	mapping = journal->j_fs_dev->bd_inode->i_mapping;
> +	errseq_check_and_advance(&mapping->wb_err, &wb_err);
>  	err = do_one_pass(journal, &info, PASS_SCAN);
>  	if (!err)
>  		err = do_one_pass(journal, &info, PASS_REVOKE);
> @@ -327,6 +332,9 @@ int jbd2_journal_recover(journal_t *journal)
>  
>  	jbd2_journal_clear_revoke(journal);
>  	err2 = sync_blockdev(journal->j_fs_dev);
> +	if (!err)
> +		err = err2;
> +	err2 = errseq_check_and_advance(&mapping->wb_err, &wb_err);
>  	if (!err)
>  		err = err2;
>  	/* Make sure all replayed data is on permanent storage */
> -- 
> 2.39.2
>
  
Theodore Ts'o Oct. 6, 2023, 6:06 p.m. UTC | #2
On Tue, 19 Sep 2023 09:25:25 +0800, Zhihao Cheng wrote:
> JBD2 makes sure journal data is fallen on fs device by sync_blockdev(),
> however, other process could intercept the EIO information from bdev's
> mapping, which leads journal recovering successful even EIO occurs during
> data written back to fs device.
> 
> We found this problem in our product, iscsi + multipath is chosen for block
> device of ext4. Unstable network may trigger kpartx to rescan partitions in
> device mapper layer. Detailed process is shown as following:
> 
> [...]

Applied, thanks!

[1/1] jbd2: Fix potential data lost in recovering journal raced with synchronizing fs bdev
      commit: 61187fce8600e8ef90e601be84f9d0f3222c1206

Best regards,
  

Patch

diff --git a/fs/jbd2/recovery.c b/fs/jbd2/recovery.c
index c269a7d29a46..5b771a3d8d9a 100644
--- a/fs/jbd2/recovery.c
+++ b/fs/jbd2/recovery.c
@@ -289,6 +289,8 @@  int jbd2_journal_recover(journal_t *journal)
 	journal_superblock_t *	sb;
 
 	struct recovery_info	info;
+	errseq_t		wb_err;
+	struct address_space	*mapping;
 
 	memset(&info, 0, sizeof(info));
 	sb = journal->j_superblock;
@@ -306,6 +308,9 @@  int jbd2_journal_recover(journal_t *journal)
 		return 0;
 	}
 
+	wb_err = 0;
+	mapping = journal->j_fs_dev->bd_inode->i_mapping;
+	errseq_check_and_advance(&mapping->wb_err, &wb_err);
 	err = do_one_pass(journal, &info, PASS_SCAN);
 	if (!err)
 		err = do_one_pass(journal, &info, PASS_REVOKE);
@@ -327,6 +332,9 @@  int jbd2_journal_recover(journal_t *journal)
 
 	jbd2_journal_clear_revoke(journal);
 	err2 = sync_blockdev(journal->j_fs_dev);
+	if (!err)
+		err = err2;
+	err2 = errseq_check_and_advance(&mapping->wb_err, &wb_err);
 	if (!err)
 		err = err2;
 	/* Make sure all replayed data is on permanent storage */