On Mon, Sep 10, 2018 at 4:30 PM Sowmini Varadhan
<sowmini.varad...@oracle.com> wrote:
>
> On (09/10/18 15:43), Santosh Shilimkar wrote:
> > On 9/10/2018 3:24 PM, Cong Wang wrote:
> > >When a rds sock is bound, it is inserted into the bind_hash_table
> > >which is protected by RCU. But when releasing rd sock, after it
> > >is removed from this hash table, it is freed immediately without
> > >respecting RCU grace period. This could cause some use-after-free
> > >as reported by syzbot.
> > >
> > Indeed.
> >
> > >Mark the rds sock as SOCK_RCU_FREE before inserting it into the
> > >bind_hash_table, so that it would be always freed after a RCU grace
> > >period.
>
> So I'm not sure I understand.
>
> Yes, Cong's fix may eliminate *some* of the syzbot failures, but the
> basic problem is not solved.
>
> To take one example of possible races (one that was discussed in
> https://www.spinics.net/lists/netdev/msg475074.html)
> rds_recv_incoming->rds_find_bound is being called in rds_send_worker
> context and the rds_find_bound code is
>
> 63 rs = rhashtable_lookup_fast(&bind_hash_table, &key, ht_parms);
> 64 if (rs && !sock_flag(rds_rs_to_sk(rs), SOCK_DEAD))
> 65 rds_sock_addref(rs);
> 66 else
> 67 rs = NULL;
> 68
>
> After we find an rs at line 63, how can we be sure that the entire
> logic of rds_release does not execute on another cpu, and free the rs,
> before we hit line 64 with the bad rs?
This is a different problem. The RCU grace period should be just
extended:
diff --git a/net/rds/bind.c b/net/rds/bind.c
index 3ab55784b637..fa7592d0760c 100644
--- a/net/rds/bind.c
+++ b/net/rds/bind.c
@@ -76,11 +76,13 @@ struct rds_sock *rds_find_bound(const struct
in6_addr *addr, __be16 port,
struct rds_sock *rs;
__rds_create_bind_key(key, addr, port, scope_id);
- rs = rhashtable_lookup_fast(&bind_hash_table, key, ht_parms);
+ rcu_read_lock();
+ rs = rhashtable_lookup(&bind_hash_table, key, ht_parms);
if (rs && !sock_flag(rds_rs_to_sk(rs), SOCK_DEAD))
rds_sock_addref(rs);
else
rs = NULL;
+ rcu_read_unlock();
rdsdebug("returning rs %p for %pI6c:%u\n", rs, addr,
ntohs(port));
>
> Normally synchronize_rcu() or synchronize_net() in rds_release() would
> ensure this. How do we ensure this with SOCK_RCU_FREE (or is the
> intention to just reduce *some* of the syzbot failures)?
Although sock release path is not a very hot path, but blocking
it isn't a good idea either, especially when you can use call_rcu(),
which has the same effect.
I don't see any reason we should prefer synchronize_rcu() here.
