Re: refcount_t: underflow; use-after-free with CIFS umount after scsi-misc commit ef2cc88e2a205b8a11a19e78db63a70d3728cdf5

[Steve French]

On Sun, Dec 8, 2019 at 8:23 PM Linus Torvalds
<torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>
> On Sun, Dec 8, 2019 at 5:49 PM Arthur Marsh
> <arthur.marsh@xxxxxxxxxxxxxxxx> wrote:
> >
> > This still happens with 5.5.0-rc1:
>
> Does it happen 100% of the time?

I can reproduce it (although it was a little more difficult since WiFi doesn't
work on RC1 on some of my hardware - due to the 802.11 driver regression oops.
I was able to reproduce it to Samba localhost).


> Your bisection result looks pretty nonsensical - not that it's
> impossible (anything is possible), but it really doesn't look very
> likely. Which makes me think maybe it's slightly timing-sensitive or
> something?

The bisection result is implausible.  I just did some experiments and
it looks far more likely is that it is related to commit
72e73c78c446e ("cifs: close the shared root handle on tree disconnect")
so added Ronnie to the cc.  That patch added a call (at unmount time)
to close_shroot.
The idea of that patch made sense - although tree disconnect (and then
logoff of the session)
will indirectly free any open handles on the server for that session,
it is a little
cleaner to close the cached root SMB3 file handle explicitly.

void close_shroot(struct cached_fid *cfid)
{
        mutex_lock(&cfid->fid_mutex);
        kref_put(&cfid->refcount, smb2_close_cached_fid);
        mutex_unlock(&cfid->fid_mutex);
}


Taking out the one line change in the patch from last week that calls
close_shroot from
umount (SMB2_tdis, ie tree_disconnect) I don't see the problem so far
more likely
that it is related to that commit.   The problem seems to be related
to servers which
don't support directory leases.  Will spin up a patch to fix this if
Ronnie hasn't already fixed it