On Fri, Mar 15, 2019 at 12:18:13PM +0000, Al Viro wrote:
> On Fri, Mar 15, 2019 at 12:50:02PM +0100, Dominik Brodowski wrote:
> > On Fri, Mar 15, 2019 at 11:44:45AM +0000, David Howells wrote:
> > > Dominik Brodowski <linux@xxxxxxxxxxxxxxxxxxxx> wrote:
> > >
> > > > [ 0.839322] RIP: 0010:sysfs_init_fs_context+0x82/0xd0
> > >
> > > Could you load your kernel into gdb and then do:
> > >
> > > i li *sysfs_init_fs_context+0x82
> >
> > Doesn't seem necessary as per my mail to Al a few seconds ago:
> > kobj_ns_grab_current(KOBJ_NS_TYPE_NET) returns NULL, so
> >
> > fc->user_ns = get_user_ns(netns->user_ns);
> >
> > will try to dereference a null pointer.
> >
> > Thanks,
> > Dominik
> >
>
> Charming... It's netns being turned off, and the thing we'd missed
> is that kobj_ns_current_may_mount() becomes constant true in that
> setup. IOW, ns_capable(..., CAP_SYS_ADMIN) is suppressed entirely
> in that case (well, aside of what may_mount() has done in
> do_mount/sys_fsmount).
>
> So what we need is making that "change fc->user_ns" conditional
> on netns != NULL, as in
>
> diff --git a/fs/sysfs/mount.c b/fs/sysfs/mount.c
> index 4cb21b558a85..1b56686ab178 100644
> --- a/fs/sysfs/mount.c
> +++ b/fs/sysfs/mount.c
> @@ -71,9 +71,11 @@ static int sysfs_init_fs_context(struct fs_context *fc)
> kfc->magic = SYSFS_MAGIC;
> fc->fs_private = kfc;
> fc->ops = &sysfs_fs_context_ops;
> - if (fc->user_ns)
> - put_user_ns(fc->user_ns);
> - fc->user_ns = get_user_ns(netns->user_ns);
> + if (netns) {
> + if (fc->user_ns)
> + put_user_ns(fc->user_ns);
> + fc->user_ns = get_user_ns(netns->user_ns);
> + }
> fc->global = true;
> return 0;
> }
That fixes the issue at hand, thanks (verified both on top of the
problematic commit and on top of mainline).
Thanks,
Dominik
