On 08/12, Daniel Borkmann wrote:
> On 8/9/19 6:10 PM, Stanislav Fomichev wrote:
> > Add new helper bpf_sk_storage_clone which optionally clones sk storage
> > and call it from sk_clone_lock.
> >
> > Cc: Martin KaFai Lau <ka...@fb.com>
> > Cc: Yonghong Song <y...@fb.com>
> > Signed-off-by: Stanislav Fomichev <s...@google.com>
> [...]
> > +int bpf_sk_storage_clone(const struct sock *sk, struct sock *newsk)
> > +{
> > + struct bpf_sk_storage *new_sk_storage = NULL;
> > + struct bpf_sk_storage *sk_storage;
> > + struct bpf_sk_storage_elem *selem;
> > + int ret;
> > +
> > + RCU_INIT_POINTER(newsk->sk_bpf_storage, NULL);
> > +
> > + rcu_read_lock();
> > + sk_storage = rcu_dereference(sk->sk_bpf_storage);
> > +
> > + if (!sk_storage || hlist_empty(&sk_storage->list))
> > + goto out;
> > +
> > + hlist_for_each_entry_rcu(selem, &sk_storage->list, snode) {
> > + struct bpf_sk_storage_elem *copy_selem;
> > + struct bpf_sk_storage_map *smap;
> > + struct bpf_map *map;
> > + int refold;
> > +
> > + smap = rcu_dereference(SDATA(selem)->smap);
> > + if (!(smap->map.map_flags & BPF_F_CLONE))
> > + continue;
> > +
> > + map = bpf_map_inc_not_zero(&smap->map, false);
> > + if (IS_ERR(map))
> > + continue;
> > +
> > + copy_selem = bpf_sk_storage_clone_elem(newsk, smap, selem);
> > + if (!copy_selem) {
> > + ret = -ENOMEM;
> > + bpf_map_put(map);
> > + goto err;
> > + }
> > +
> > + if (new_sk_storage) {
> > + selem_link_map(smap, copy_selem);
> > + __selem_link_sk(new_sk_storage, copy_selem);
> > + } else {
> > + ret = sk_storage_alloc(newsk, smap, copy_selem);
> > + if (ret) {
> > + kfree(copy_selem);
> > + atomic_sub(smap->elem_size,
> > + &newsk->sk_omem_alloc);
> > + bpf_map_put(map);
> > + goto err;
> > + }
> > +
> > + new_sk_storage =
> > rcu_dereference(copy_selem->sk_storage);
> > + }
> > + bpf_map_put(map);
>
> The map get/put combination /under/ RCU read lock seems a bit odd to me, could
> you exactly describe the race that this would be preventing?
There is a race between sk storage release and sk storage clone.
bpf_sk_storage_map_free uses synchronize_rcu to wait for all existing
users to finish and the new ones are prevented via map's refcnt being
zero; we need to do something like that for the clone.
Martin suggested to use bpf_map_inc_not_zero/bpf_map_put.
If I read everythin correctly, I think without map_inc/map_put we
get the following race:
CPU0 CPU1
bpf_map_put
bpf_sk_storage_map_free(smap)
synchronize_rcu
// no more users via bpf or
// syscall, but clone
// can still happen
for each (bucket)
selem_unlink
selem_unlink_map(smap)
// adding anything at
// this point to the
// bucket will leak
rcu_read_lock
tcp_v4_rcv
tcp_v4_do_rcv
// sk is lockless TCP_LISTEN
tcp_v4_cookie_check
tcp_v4_syn_recv_sock
bpf_sk_storage_clone
rcu_dereference(sk->sk_bpf_storage)
selem_link_map(smap, copy)
// adding new element to the
// map -> leak
rcu_read_unlock
selem_unlink_sk
sk->sk_bpf_storage = NULL
synchronize_rcu
