On Fri, 10 May 2024 07:26:13 +0800 Hillf Danton <hdanton@xxxxxxxx> wrote:
> On Thu, 9 May 2024 17:52:21 +0300 Amir Goldstein <amir73il@xxxxxxxxx>
> > On Thu, May 9, 2024 at 1:49 PM Hillf Danton <hdanton@xxxxxxxx> wrote:
> > >
> > > The correct locking order is
> > >
> > > sb_writers
> >
> > This is sb of overlayfs
> >
> > > inode lock
> >
> > This is real inode
> >
> WRT sb_writers the order
>
> lock inode parent
> lock inode kid
>
> becomes
> lock inode kid
> sb_writers
> lock inode parent
>
> given call trace
>
> > -> #2 (sb_writers#4){.+.+}-{0:0}:
> > lock_acquire+0x1ed/0x550 kernel/locking/lockdep.c:5754
> > percpu_down_read include/linux/percpu-rwsem.h:51 [inline]
> > __sb_start_write include/linux/fs.h:1664 [inline]
> > sb_start_write+0x4d/0x1c0 include/linux/fs.h:1800
> > mnt_want_write+0x3f/0x90 fs/namespace.c:409
> > ovl_create_object+0x13b/0x370 fs/overlayfs/dir.c:629
> > lookup_open fs/namei.c:3497 [inline]
> > open_last_lookups fs/namei.c:3566 [inline]
>
> and code snippet [1]
>
> if (open_flag & O_CREAT)
> inode_lock(dir->d_inode);
> else
> inode_lock_shared(dir->d_inode);
> dentry = lookup_open(nd, file, op, got_write);
>
> [1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/fs/namei.c?id=dccb07f2914c#n3566
JFYI simply cutting off mnt_want_write() in ovl_create_object() survived
the syzpot repro [2], so acquiring sb_writers with inode locked at least
in the lookup path makes trouble.
[2] https://lore.kernel.org/lkml/000000000000975906061817416b@xxxxxxxxxx/
