On Wed, Oct 17, 2018 at 11:29 AM Miklos Szeredi <miklos@xxxxxxxxxx> wrote:
>
> Cc: linux-unionfs@vger
>
> On Wed, Oct 17, 2018 at 10:00 AM, Stefan Agner <stefan@xxxxxxxx> wrote:
> > Hi,
> >
> > I noticed this warning since we moved to 4.18. It appears when
> > using Docker (which uses overlayfs). Is this a known issue?
> >
> > [ 543.235366] WARNING: possible recursive locking detected
> > [ 543.240747] 4.18.14 #1 Not tainted
> > [ 543.244195] --------------------------------------------
> > [ 543.249573] dockerd/522 is trying to acquire lock:
> > [ 543.254426] 86b0f89c (sb_writers#7){.+.+}, at: mnt_want_write+0x20/0x4c
> > [ 543.261152] but task is already holding lock:
> > [ 543.267053] 86b0f89c (sb_writers#7){.+.+}, at: mnt_want_write_file_path+0x24/0x54
> > [ 543.274641] other info that might help us debug this:
> > [ 543.281242] Possible unsafe locking scenario:
> > [ 543.287227] CPU0
> > [ 543.289706] ----
> > [ 543.292183] lock(sb_writers#7);
> > [ 543.295547] lock(sb_writers#7);
> > [ 543.298912] *** DEADLOCK ***
> > [ 543.306825] May be due to missing lock nesting notation
> > [ 543.315594] 2 locks held by dockerd/522:
> > [ 543.320487] #0: 86b0f89c (sb_writers#7){.+.+}, at: mnt_want_write_file_path+0x24/0x54
> > [ 543.330353] #1: fbe4681b (&ovl_i_mutex_key[depth]){+.+.}, at: chown_common+0xf8/0x1c0
This is a stable tree regression because upstream commit
a6795a585929 vfs: fix freeze protection in mnt_want_write_file() for overlayfs
SHOULD NOT have been applied to kernel <= 4.18
Miklos,
You must have confused "the algorithm" by including a "fix" commit in
rc1 pull request
without mentioning that it "Fixes" a commit in the same pull request.
It probably didn't help that the commit applied cleanly to the wrong
function :-/
Upstream commit a6795a585929:
@@ -441,10 +441,10 @@ int mnt_want_write_file(struct file *file)
{
int ret;
- sb_start_write(file->f_path.mnt->mnt_sb);
+ sb_start_write(file_inode(file)->i_sb);
Backported v4.18.14 commit 5e1002ab5c9b:
@@ -446,10 +446,10 @@ int mnt_want_write_file_path(struct file *file)
{
int ret;
- sb_start_write(file->f_path.mnt->mnt_sb);
+ sb_start_write(file_inode(file)->i_sb);
Ouch! Is there a way for git/quilt to raise a red flag in cases like this?
I could reproduce another lockdep warning on v4.18.14 with xfstest
overlay/030, which is fixed by reverting this commit.
Stefan,
Can you verify that reverting that commit solved the problem for you?
Thanks,
Amir.
