On Sep 6, 2021, at 4:42 AM, Joseph Qi <joseph.qi@linux.alibaba.com> wrote:

On 9/3/21 9:26 AM, Wengang Wang wrote:

ocfs2_data_convert_worker() is currently dropping any cached acl info
for FILE before down-converting meta lock. It should also drop for DIRECTORY.
Otherwise the second acl lookup returns the cached one (from VFS layer) which
could be already stale.

The problem we are seeing is that the acl changes on one node doesn't get
refreshed on other nodes in the following case:

 Node 1                    Node 2
--------------            ----------------
getfacl dir1

  getfacl dir1    <-- this is OK

setfacl -m u:user1:rwX dir1
getfacl dir1   <-- see the change for user1

  getfacl dir1    <-- can't see change for user1

Signed-off-by: Wengang Wang <wen.gang.wang@oracle.com>

Looks sane to me.
Reviewed-by: Joseph Qi <joseph.qi@linux.alibaba.com>

---
fs/ocfs2/dlmglue.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/fs/ocfs2/dlmglue.c b/fs/ocfs2/dlmglue.c
index 50a863fc1779..207ec61569ea 100644
--- a/fs/ocfs2/dlmglue.c
+++ b/fs/ocfs2/dlmglue.c
@@ -3933,7 +3933,7 @@ static int ocfs2_data_convert_worker(struct ocfs2_lock_res *lockres,
oi = OCFS2_I(inode);
oi->ip_dir_lock_gen++;
mlog(0, "generation: %u\n", oi->ip_dir_lock_gen);
- goto out;
+ goto out_forget;
}

if (!S_ISREG(inode->i_mode))
@@ -3964,6 +3964,7 @@ static int ocfs2_data_convert_worker(struct ocfs2_lock_res *lockres,
filemap_fdatawait(mapping);
}

+out_forget:
forget_all_cached_acls(inode);

out: