Re: [syzbot] [gfs2?] WARNING: suspicious RCU usage in gfs2_permission

[Andreas Gruenbacher]

Al,

On Wed, Oct 25, 2023 at 5:29 AM Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
> On Fri, Oct 20, 2023 at 12:10:38AM -0700, syzbot wrote:
> > syzbot has bisected this issue to:
> >
> > commit 0abd1557e21c617bd13fc18f7725fc6363c05913
> > Author: Al Viro <viro@xxxxxxxxxxxxxxxxxx>
> > Date:   Mon Oct 2 02:33:44 2023 +0000
> >
> >     gfs2: fix an oops in gfs2_permission
> >
> > bisection log:  https://syzkaller.appspot.com/x/bisect.txt?x=10b21c33680000
> > start commit:   2dac75696c6d Add linux-next specific files for 20231018
> > git tree:       linux-next
> > final oops:     https://syzkaller.appspot.com/x/report.txt?x=12b21c33680000
> > console output: https://syzkaller.appspot.com/x/log.txt?x=14b21c33680000
> > kernel config:  https://syzkaller.appspot.com/x/.config?x=6f8545e1ef7a2b66
> > dashboard link: https://syzkaller.appspot.com/bug?extid=3e5130844b0c0e2b4948
> > syz repro:      https://syzkaller.appspot.com/x/repro.syz?x=101c8d09680000
> > C reproducer:   https://syzkaller.appspot.com/x/repro.c?x=11a07475680000
> >
> > Reported-by: syzbot+3e5130844b0c0e2b4948@xxxxxxxxxxxxxxxxxxxxxxxxx
> > Fixes: 0abd1557e21c ("gfs2: fix an oops in gfs2_permission")
> >
> > For information about bisection process see: https://goo.gl/tpsmEJ#bisection
>
> Complaints about rcu_dereference() outside of rcu_read_lock().
>
> We could replace that line with
>         if (mask & MAY_NOT_BLOCK)
>                 gl = rcu_dereference(ip->i_gl);
>         else
>                 gl = ip->i_gl;
> or by any equivalent way to tell lockdep it ought to STFU.

the following should do then, right?

    gl = rcu_dereference_check(ip->i_gl, !(mask & MAY_NOT_BLOCK));

> BTW, the amount of rcu_dereference_protected(..., true) is somewhat depressing...
>
> Probably need to turn
>                 ip->i_gl = NULL;
> in the end of gfs2_evict_inode() into rcu_assign_pointer(ip->i_gl, NULL);

That's what commit 0abd1557e21c6 does already so there's nothing to fix, right?

> and transpose it with the previous line -
>                 gfs2_glock_put_eventually(ip->i_gl);
>
> I don't think it really matters in this case, though - destruction of the object
> it used to point to is delayed in all cases.  Matter of taste (and lockdep
> false positives)...

I don't understand. What would lockdep complain about here?

Thanks,
Andreas