On Mon, 16 Mar 2015 21:00:35 +0000 Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
> On Mon, Mar 16, 2015 at 03:43:19PM +1100, NeilBrown wrote:
> > Some of dentry_has_perm() is not rcu-safe, so if LOOKUP_RCU
> > is set in selinux_inode_follow_link(), give up with
> > -ECHILD.
> >
> > It is possible that dentry_has_perm could sometimes complete
> > in RCU more, in which case the flag could be propagated further
> > down the stack...
>
> It bloody well can. Expand it a bit and you'll see - the nastiness
> comes from avc_audit() doing
> return slow_avc_audit(ssid, tsid, tclass,
> requested, audited, denied, result,
> a, 0);
> and passing that 0 to slow_avc_audit(). Pass it MAY_NOT_BLOCK instead
> and it'll bugger off with -ECHILD in blocking case.
>
> Call chain is dentry_has_perm -> inode_has_perm -> avc_has_perm -> avc_audit.
> Expand those (including avc_audit()) and make slow_avc_audit() get
> flags & LOOKUP_RCU ? MAY_NOT_BLOCK : 0.
There is more to it than that.
avc_has_perm calls avc_has_perm_noaudit which does:
rcu_read_lock();
...
if (unlikely(!node)) {
node = avc_compute_av(ssid, tsid, tclass, avd);
} else ...
...
rcu_read_unlock();
and avc_compute_av() does
rcu_read_unlock();
security_compute_av(ssid, tsid, tclass, avd);
rcu_read_lock();
(yes: unlock, and then lock).
so avc_has_perm_noaudit needs to bail out of RCU-walk if node turns out to be
NULL.
So I either add another 'flags' arg to that, or replace the current one which
is unused .... or leave it as someone else's problem :-)
NeilBrown
Attachment:
pgpAbuFxYDeHE.pgp
Description: OpenPGP digital signature
