tcp: fix refcnt leak with ebpf congestion control
authorSabrina Dubroca <sd@queasysnail.net>
Fri, 25 Aug 2017 11:10:12 +0000 (13:10 +0200)
committerDavid S. Miller <davem@davemloft.net>
Sat, 26 Aug 2017 00:16:27 +0000 (17:16 -0700)
There are a few bugs around refcnt handling in the new BPF congestion
control setsockopt:

 - The new ca is assigned to icsk->icsk_ca_ops even in the case where we
   cannot get a reference on it. This would lead to a use after free,
   since that ca is going away soon.

 - Changing the congestion control case doesn't release the refcnt on
   the previous ca.

 - In the reinit case, we first leak a reference on the old ca, then we
   call tcp_reinit_congestion_control on the ca that we have just
   assigned, leading to deinitializing the wrong ca (->release of the
   new ca on the old ca's data) and releasing the refcount on the ca
   that we actually want to use.

This is visible by building (for example) BIC as a module and setting
net.ipv4.tcp_congestion_control=bic, and using tcp_cong_kern.c from
samples/bpf.

This patch fixes the refcount issues, and moves reinit back into tcp
core to avoid passing a ca pointer back to BPF.

Fixes: 91b5b21c7c16 ("bpf: Add support for changing congestion control")
Signed-off-by: Sabrina Dubroca <sd@queasysnail.net>
Acked-by: Lawrence Brakmo <brakmo@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/net/tcp.h
net/core/filter.c
net/ipv4/tcp.c
net/ipv4/tcp_cong.c

index ada65e767b28dfcabb662a7b08f65c6fc04f5b73..f642a39f9eeeeb3a1bbff48fd467c3a3acb96160 100644 (file)
@@ -1004,9 +1004,7 @@ void tcp_get_default_congestion_control(char *name);
 void tcp_get_available_congestion_control(char *buf, size_t len);
 void tcp_get_allowed_congestion_control(char *buf, size_t len);
 int tcp_set_allowed_congestion_control(char *allowed);
-int tcp_set_congestion_control(struct sock *sk, const char *name, bool load);
-void tcp_reinit_congestion_control(struct sock *sk,
-                                  const struct tcp_congestion_ops *ca);
+int tcp_set_congestion_control(struct sock *sk, const char *name, bool load, bool reinit);
 u32 tcp_slow_start(struct tcp_sock *tp, u32 acked);
 void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w, u32 acked);
 
index 8eb81e5fae0804266bdec45a3865c237737681e2..169974998c7692b063947cb925fede167f2fb817 100644 (file)
@@ -2836,15 +2836,12 @@ BPF_CALL_5(bpf_setsockopt, struct bpf_sock_ops_kern *, bpf_sock,
                   sk->sk_prot->setsockopt == tcp_setsockopt) {
                if (optname == TCP_CONGESTION) {
                        char name[TCP_CA_NAME_MAX];
+                       bool reinit = bpf_sock->op > BPF_SOCK_OPS_NEEDS_ECN;
 
                        strncpy(name, optval, min_t(long, optlen,
                                                    TCP_CA_NAME_MAX-1));
                        name[TCP_CA_NAME_MAX-1] = 0;
-                       ret = tcp_set_congestion_control(sk, name, false);
-                       if (!ret && bpf_sock->op > BPF_SOCK_OPS_NEEDS_ECN)
-                               /* replacing an existing ca */
-                               tcp_reinit_congestion_control(sk,
-                                       inet_csk(sk)->icsk_ca_ops);
+                       ret = tcp_set_congestion_control(sk, name, false, reinit);
                } else {
                        struct tcp_sock *tp = tcp_sk(sk);
 
index 71ce33decd971feee72f225b2bf7ccaf2f5f456f..a3e91b552edce4edee0d3b9ee5e07105946d2dd9 100644 (file)
@@ -2481,7 +2481,7 @@ static int do_tcp_setsockopt(struct sock *sk, int level,
                name[val] = 0;
 
                lock_sock(sk);
-               err = tcp_set_congestion_control(sk, name, true);
+               err = tcp_set_congestion_control(sk, name, true, true);
                release_sock(sk);
                return err;
        }
index fde983f6376be98247b9f2ff3d0307c2502e6392..421ea1b918da5bc4a3974531539cd67266f70798 100644 (file)
@@ -189,8 +189,8 @@ void tcp_init_congestion_control(struct sock *sk)
                INET_ECN_dontxmit(sk);
 }
 
-void tcp_reinit_congestion_control(struct sock *sk,
-                                  const struct tcp_congestion_ops *ca)
+static void tcp_reinit_congestion_control(struct sock *sk,
+                                         const struct tcp_congestion_ops *ca)
 {
        struct inet_connection_sock *icsk = inet_csk(sk);
 
@@ -338,7 +338,7 @@ out:
  * tcp_reinit_congestion_control (if the current congestion control was
  * already initialized.
  */
-int tcp_set_congestion_control(struct sock *sk, const char *name, bool load)
+int tcp_set_congestion_control(struct sock *sk, const char *name, bool load, bool reinit)
 {
        struct inet_connection_sock *icsk = inet_csk(sk);
        const struct tcp_congestion_ops *ca;
@@ -360,9 +360,18 @@ int tcp_set_congestion_control(struct sock *sk, const char *name, bool load)
        if (!ca) {
                err = -ENOENT;
        } else if (!load) {
-               icsk->icsk_ca_ops = ca;
-               if (!try_module_get(ca->owner))
+               const struct tcp_congestion_ops *old_ca = icsk->icsk_ca_ops;
+
+               if (try_module_get(ca->owner)) {
+                       if (reinit) {
+                               tcp_reinit_congestion_control(sk, ca);
+                       } else {
+                               icsk->icsk_ca_ops = ca;
+                               module_put(old_ca->owner);
+                       }
+               } else {
                        err = -EBUSY;
+               }
        } else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) ||
                     ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN))) {
                err = -EPERM;