[Ocfs2-devel] [PATCH] ocfs2: fix missing reset j_num_trans for sync

[Heming Zhao]

Sorry, please pause this patch review.
When I was investigating fstest generic failed case 347 361, I found
the wake_up() action should move out the 'if()' area. The correct way
is calling wake_up() unconditionally.

Thanks,
Heming

On 4/21/23 4:36 PM, Heming Zhao wrote:
> fstest generic case 266 272 281 trigger hanging issue when umount.
> 
> I use 266 to describe the root cause.
> 
> ```
>   49 _dmerror_unmount
>   50 _dmerror_mount
>   51
>   52 echo "Compare files"
>   53 md5sum $testdir/file1 | _filter_scratch
>   54 md5sum $testdir/file2 | _filter_scratch
>   55
>   56 echo "CoW and unmount"
>   57 sync
>   58 _dmerror_load_error_table
>   59 urk=$($XFS_IO_PROG -f -c "pwrite -S 0x63 -b $bufsize 0 $filesize" \
>   60     -c "fdatasync" $testdir/file2 2>&1)
>   61 echo $urk >> $seqres.full
>   62 echo "$urk" | grep -q "error" || _fail "pwrite did not fail"
>   63
>   64 echo "Clean up the mess"
>   65 _dmerror_unmount
> ```
> 
> After line 49 50 umount & mount ocfs2 dev, this case run md5sum to
> verify target file. line 57 run 'sync' before line 58 changes the dm
> target from dm-linear to dm-error. this case is hanging at line 65.
> 
> In 266, md5sum calls jbd2 trans pair: ocfs2_[start|commit]_trans to
> do journal job. but there is only ->j_num_trans+1 in ocfs2_start_trans,
> the ocfs2_commit_trans doesn't do reduction operation.
> 
> call flow:
> ```
> [md5sum]
>   vfs_read
>    ocfs2_file_read_iter
>     ocfs2_inode_lock_atime
>      ocfs2_update_inode_atime
>       + ocfs2_start_trans //atomic_inc j_num_trans
>       + ...
>       + ocfs2_commit_trans//no modify j_num_trans
> 
> sync //no modify j_num_trans
> 
> _dmerror_load_error_table //all write will return error after this line
> 
> _dmerror_unmount //found j_num_trans is not zero, run commit thread
>                 //but the underlying dev is dm-error, journaling IO
>                 //failed all the time and keep going to retry.
> ```
> 
> *** How to fix ***
> 
> kick commit thread in sync path, which can reset j_num_trans to 0.
> 
> Signed-off-by: Heming Zhao <heming.zhao at suse.com>
> ---
>   fs/ocfs2/super.c | 3 +++
>   1 file changed, 3 insertions(+)
> 
> diff --git a/fs/ocfs2/super.c b/fs/ocfs2/super.c
> index 0b0e6a132101..9929320bd967 100644
> --- a/fs/ocfs2/super.c
> +++ b/fs/ocfs2/super.c
> @@ -408,6 +408,9 @@ static int ocfs2_sync_fs(struct super_block *sb, int wait)
>   
>   	if (jbd2_journal_start_commit(osb->journal->j_journal,
>   				      &target)) {
> +		/* kick commit thread to reset journal->j_num_trans */
> +		wake_up(&osb->checkpoint_event);
> +
>   		if (wait)
>   			jbd2_log_wait_commit(osb->journal->j_journal,
>   					     target);
>