Re: [Ocfs2-devel] [PATCH v2] ocfs2: Fix handle refcount leak in two exception handling paths

[Joseph Qi]

On 9/10/21 1:48 AM, Wengang Wang wrote:
> 
> 
> On Sep 9, 2021, at 4:07 AM, Joseph Qi <joseph.qi@xxxxxxxxxxxxxxxxx<mailto:joseph.qi@xxxxxxxxxxxxxxxxx>> wrote:
> 
> Hi Wengang,
> 
> On 9/9/21 1:12 AM, Wengang Wang wrote:
> Hi,
> 
> Sorry for late involving, but this doesn’t look right to me.
> 
> On Sep 8, 2021, at 3:51 AM, Joseph Qi <joseph.qi@xxxxxxxxxxxxxxxxx<mailto:joseph.qi@xxxxxxxxxxxxxxxxx>> wrote:
> 
> 
> 
> On 9/8/21 6:20 PM, Chenyuan Mi wrote:
> The reference counting issue happens in two exception handling paths
> of ocfs2_replay_truncate_records(). When executing these two exception
> handling paths, the function forgets to decrease the refcount of handle
> increased by ocfs2_start_trans(), causing a refcount leak.
> 
> Fix this issue by using ocfs2_commit_trans() to decrease the refcount
> of handle in two handling paths.
> 
> Signed-off-by: Chenyuan Mi <cymi20@xxxxxxxxxxxx<mailto:cymi20@xxxxxxxxxxxx>>
> Signed-off-by: Xiyu Yang <xiyuyang19@xxxxxxxxxxxx<mailto:xiyuyang19@xxxxxxxxxxxx>>
> Signed-off-by: Xin Tan <tanxin.ctf@xxxxxxxxx<mailto:tanxin.ctf@xxxxxxxxx>>
> 
> Reviewed-by: Joseph Qi <joseph.qi@xxxxxxxxxxxxxxxxx<mailto:joseph.qi@xxxxxxxxxxxxxxxxx>>
> ---
> fs/ocfs2/alloc.c | 2 ++
> 1 file changed, 2 insertions(+)
> 
> diff --git a/fs/ocfs2/alloc.c b/fs/ocfs2/alloc.c
> index f1cc8258d34a..b05fde7edc3a 100644
> --- a/fs/ocfs2/alloc.c
> +++ b/fs/ocfs2/alloc.c
> @@ -5940,6 +5940,7 @@ static int ocfs2_replay_truncate_records(struct ocfs2_super *osb,
> status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh,
>  OCFS2_JOURNAL_ACCESS_WRITE);
> if (status < 0) {
> + ocfs2_commit_trans(osb, handle);
> mlog_errno(status);
> goto bail;
> }
> @@ -5964,6 +5965,7 @@ static int ocfs2_replay_truncate_records(struct ocfs2_super *osb,
>      data_alloc_bh, start_blk,
>      num_clusters);
> if (status < 0) {
> + ocfs2_commit_trans(osb, handle);
> 
> As a transaction, stuff expected to be in the same handle should be treated as atomic.
> Here the stuff includes the tl_bh and other metadata block which will be modified in ocfs2_free_clusters().
> Coming here, some of related meta blocks may be in the handle but others are not due to the error happened.
> If you do a commit, partial meta blocks are committed to log. — that breaks the atomic idea, it will cause FS inconsistency.
> So what’s reason you want to commit the meta block changes, which is not all of expected, in this handle to journal log?
> 
> Do you really see a hit on the failure? or just you detected the refcount leak by code review?
> 
> You may want to look at ocfs2_journal_dirty() for the error handling part.
> 
> 
> For the first error handling, since we don't call ocfs2_journal_dirty()
> yet, so won't be a problem.
> For the second error handling, I think we don't have a better way. Look
> at other callers of ocfs2_free_clusters(), we simply ignore the error
> code.
> Anyway, we should commit transaction if starts, otherwise journal will
> be abnormal.
> 
> I don't think so. If error happened, we should fail ocfs2, rather than do a partial committing.
> 

Umm... not exactly...
Take ocfs2_free_clusters() for example, when it fails in case of EIO or
ENOMEM, we can't just abort journal in such cases, because it is not so
serious, only a bit blocks still occupied and they will recovery during
the next mount. 
That's why we have "errors=continue" in most filesystems, we should always
consider the business continuity first.
Also you can look at ext4_free_blocks() for reference.

Thanks,
Joseph