OSDN Git Service

bpf, sockmap: fix sock_map_ctx_update_elem race with exist/noexist
authorDaniel Borkmann <daniel@iogearbox.net>
Thu, 16 Aug 2018 19:49:10 +0000 (21:49 +0200)
committerAlexei Starovoitov <ast@kernel.org>
Thu, 16 Aug 2018 21:58:08 +0000 (14:58 -0700)
The current code in sock_map_ctx_update_elem() allows for BPF_EXIST
and BPF_NOEXIST map update flags. While on array-like maps this approach
is rather uncommon, e.g. bpf_fd_array_map_update_elem() and others
enforce map update flags to be BPF_ANY such that xchg() can be used
directly, the current implementation in sock map does not guarantee
that such operation with BPF_EXIST / BPF_NOEXIST is atomic.

The initial test does a READ_ONCE(stab->sock_map[i]) to fetch the
socket from the slot which is then tested for NULL / non-NULL. However
later after __sock_map_ctx_update_elem(), the actual update is done
through osock = xchg(&stab->sock_map[i], sock). Problem is that in
the meantime a different CPU could have updated / deleted a socket
on that specific slot and thus flag contraints won't hold anymore.

I've been thinking whether best would be to just break UAPI and do
an enforcement of BPF_ANY to check if someone actually complains,
however trouble is that already in BPF kselftest we use BPF_NOEXIST
for the map update, and therefore it might have been copied into
applications already. The fix to keep the current behavior intact
would be to add a map lock similar to the sock hash bucket lock only
for covering the whole map.

Fixes: 174a79ff9515 ("bpf: sockmap with sk redirect support")
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Acked-by: Song Liu <songliubraving@fb.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
kernel/bpf/sockmap.c

index 921cb6b..98e621a 100644 (file)
@@ -58,6 +58,7 @@ struct bpf_stab {
        struct bpf_map map;
        struct sock **sock_map;
        struct bpf_sock_progs progs;
+       raw_spinlock_t lock;
 };
 
 struct bucket {
@@ -89,9 +90,9 @@ enum smap_psock_state {
 
 struct smap_psock_map_entry {
        struct list_head list;
+       struct bpf_map *map;
        struct sock **entry;
        struct htab_elem __rcu *hash_link;
-       struct bpf_htab __rcu *htab;
 };
 
 struct smap_psock {
@@ -343,13 +344,18 @@ static void bpf_tcp_close(struct sock *sk, long timeout)
        e = psock_map_pop(sk, psock);
        while (e) {
                if (e->entry) {
-                       osk = cmpxchg(e->entry, sk, NULL);
+                       struct bpf_stab *stab = container_of(e->map, struct bpf_stab, map);
+
+                       raw_spin_lock_bh(&stab->lock);
+                       osk = *e->entry;
                        if (osk == sk) {
+                               *e->entry = NULL;
                                smap_release_sock(psock, sk);
                        }
+                       raw_spin_unlock_bh(&stab->lock);
                } else {
                        struct htab_elem *link = rcu_dereference(e->hash_link);
-                       struct bpf_htab *htab = rcu_dereference(e->htab);
+                       struct bpf_htab *htab = container_of(e->map, struct bpf_htab, map);
                        struct hlist_head *head;
                        struct htab_elem *l;
                        struct bucket *b;
@@ -1642,6 +1648,7 @@ static struct bpf_map *sock_map_alloc(union bpf_attr *attr)
                return ERR_PTR(-ENOMEM);
 
        bpf_map_init_from_attr(&stab->map, attr);
+       raw_spin_lock_init(&stab->lock);
 
        /* make sure page count doesn't overflow */
        cost = (u64) stab->map.max_entries * sizeof(struct sock *);
@@ -1716,14 +1723,15 @@ static void sock_map_free(struct bpf_map *map)
         * and a grace period expire to ensure psock is really safe to remove.
         */
        rcu_read_lock();
+       raw_spin_lock_bh(&stab->lock);
        for (i = 0; i < stab->map.max_entries; i++) {
                struct smap_psock *psock;
                struct sock *sock;
 
-               sock = xchg(&stab->sock_map[i], NULL);
+               sock = stab->sock_map[i];
                if (!sock)
                        continue;
-
+               stab->sock_map[i] = NULL;
                psock = smap_psock_sk(sock);
                /* This check handles a racing sock event that can get the
                 * sk_callback_lock before this case but after xchg happens
@@ -1735,6 +1743,7 @@ static void sock_map_free(struct bpf_map *map)
                        smap_release_sock(psock, sock);
                }
        }
+       raw_spin_unlock_bh(&stab->lock);
        rcu_read_unlock();
 
        sock_map_remove_complete(stab);
@@ -1778,14 +1787,16 @@ static int sock_map_delete_elem(struct bpf_map *map, void *key)
        if (k >= map->max_entries)
                return -EINVAL;
 
-       sock = xchg(&stab->sock_map[k], NULL);
+       raw_spin_lock_bh(&stab->lock);
+       sock = stab->sock_map[k];
+       stab->sock_map[k] = NULL;
+       raw_spin_unlock_bh(&stab->lock);
        if (!sock)
                return -EINVAL;
 
        psock = smap_psock_sk(sock);
        if (!psock)
-               goto out;
-
+               return 0;
        if (psock->bpf_parse) {
                write_lock_bh(&sock->sk_callback_lock);
                smap_stop_sock(psock, sock);
@@ -1793,7 +1804,6 @@ static int sock_map_delete_elem(struct bpf_map *map, void *key)
        }
        smap_list_map_remove(psock, &stab->sock_map[k]);
        smap_release_sock(psock, sock);
-out:
        return 0;
 }
 
@@ -1829,11 +1839,9 @@ out:
 static int __sock_map_ctx_update_elem(struct bpf_map *map,
                                      struct bpf_sock_progs *progs,
                                      struct sock *sock,
-                                     struct sock **map_link,
                                      void *key)
 {
        struct bpf_prog *verdict, *parse, *tx_msg;
-       struct smap_psock_map_entry *e = NULL;
        struct smap_psock *psock;
        bool new = false;
        int err = 0;
@@ -1906,14 +1914,6 @@ static int __sock_map_ctx_update_elem(struct bpf_map *map,
                new = true;
        }
 
-       if (map_link) {
-               e = kzalloc(sizeof(*e), GFP_ATOMIC | __GFP_NOWARN);
-               if (!e) {
-                       err = -ENOMEM;
-                       goto out_free;
-               }
-       }
-
        /* 3. At this point we have a reference to a valid psock that is
         * running. Attach any BPF programs needed.
         */
@@ -1935,17 +1935,6 @@ static int __sock_map_ctx_update_elem(struct bpf_map *map,
                write_unlock_bh(&sock->sk_callback_lock);
        }
 
-       /* 4. Place psock in sockmap for use and stop any programs on
-        * the old sock assuming its not the same sock we are replacing
-        * it with. Because we can only have a single set of programs if
-        * old_sock has a strp we can stop it.
-        */
-       if (map_link) {
-               e->entry = map_link;
-               spin_lock_bh(&psock->maps_lock);
-               list_add_tail(&e->list, &psock->maps);
-               spin_unlock_bh(&psock->maps_lock);
-       }
        return err;
 out_free:
        smap_release_sock(psock, sock);
@@ -1956,7 +1945,6 @@ out_progs:
        }
        if (tx_msg)
                bpf_prog_put(tx_msg);
-       kfree(e);
        return err;
 }
 
@@ -1966,36 +1954,57 @@ static int sock_map_ctx_update_elem(struct bpf_sock_ops_kern *skops,
 {
        struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
        struct bpf_sock_progs *progs = &stab->progs;
-       struct sock *osock, *sock;
+       struct sock *osock, *sock = skops->sk;
+       struct smap_psock_map_entry *e;
+       struct smap_psock *psock;
        u32 i = *(u32 *)key;
        int err;
 
        if (unlikely(flags > BPF_EXIST))
                return -EINVAL;
-
        if (unlikely(i >= stab->map.max_entries))
                return -E2BIG;
 
-       sock = READ_ONCE(stab->sock_map[i]);
-       if (flags == BPF_EXIST && !sock)
-               return -ENOENT;
-       else if (flags == BPF_NOEXIST && sock)
-               return -EEXIST;
+       e = kzalloc(sizeof(*e), GFP_ATOMIC | __GFP_NOWARN);
+       if (!e)
+               return -ENOMEM;
 
-       sock = skops->sk;
-       err = __sock_map_ctx_update_elem(map, progs, sock, &stab->sock_map[i],
-                                        key);
+       err = __sock_map_ctx_update_elem(map, progs, sock, key);
        if (err)
                goto out;
 
-       osock = xchg(&stab->sock_map[i], sock);
-       if (osock) {
-               struct smap_psock *opsock = smap_psock_sk(osock);
+       /* psock guaranteed to be present. */
+       psock = smap_psock_sk(sock);
+       raw_spin_lock_bh(&stab->lock);
+       osock = stab->sock_map[i];
+       if (osock && flags == BPF_NOEXIST) {
+               err = -EEXIST;
+               goto out_unlock;
+       }
+       if (!osock && flags == BPF_EXIST) {
+               err = -ENOENT;
+               goto out_unlock;
+       }
+
+       e->entry = &stab->sock_map[i];
+       e->map = map;
+       spin_lock_bh(&psock->maps_lock);
+       list_add_tail(&e->list, &psock->maps);
+       spin_unlock_bh(&psock->maps_lock);
 
-               smap_list_map_remove(opsock, &stab->sock_map[i]);
-               smap_release_sock(opsock, osock);
+       stab->sock_map[i] = sock;
+       if (osock) {
+               psock = smap_psock_sk(osock);
+               smap_list_map_remove(psock, &stab->sock_map[i]);
+               smap_release_sock(psock, osock);
        }
+       raw_spin_unlock_bh(&stab->lock);
+       return 0;
+out_unlock:
+       smap_release_sock(psock, sock);
+       raw_spin_unlock_bh(&stab->lock);
 out:
+       kfree(e);
        return err;
 }
 
@@ -2358,7 +2367,7 @@ static int sock_hash_ctx_update_elem(struct bpf_sock_ops_kern *skops,
        b = __select_bucket(htab, hash);
        head = &b->head;
 
-       err = __sock_map_ctx_update_elem(map, progs, sock, NULL, key);
+       err = __sock_map_ctx_update_elem(map, progs, sock, key);
        if (err)
                goto err;
 
@@ -2384,8 +2393,7 @@ static int sock_hash_ctx_update_elem(struct bpf_sock_ops_kern *skops,
        }
 
        rcu_assign_pointer(e->hash_link, l_new);
-       rcu_assign_pointer(e->htab,
-                          container_of(map, struct bpf_htab, map));
+       e->map = map;
        spin_lock_bh(&psock->maps_lock);
        list_add_tail(&e->list, &psock->maps);
        spin_unlock_bh(&psock->maps_lock);