net: annotate data-races around sk->sk_{rcv|snd}timeo
authorEric Dumazet <edumazet@google.com>
Fri, 28 Jul 2023 15:03:12 +0000 (15:03 +0000)
committerDavid S. Miller <davem@davemloft.net>
Sat, 29 Jul 2023 17:13:41 +0000 (18:13 +0100)
sk_getsockopt() runs without locks, we must add annotations
to sk->sk_rcvtimeo and sk->sk_sndtimeo.

In the future we might allow fetching these fields before
we lock the socket in TCP fast path.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/core/sock.c
net/sched/em_meta.c

index 08e6050..264c99c 100644 (file)
@@ -429,6 +429,7 @@ static int sock_set_timeout(long *timeo_p, sockptr_t optval, int optlen,
 {
        struct __kernel_sock_timeval tv;
        int err = sock_copy_user_timeval(&tv, optval, optlen, old_timeval);
+       long val;
 
        if (err)
                return err;
@@ -439,7 +440,7 @@ static int sock_set_timeout(long *timeo_p, sockptr_t optval, int optlen,
        if (tv.tv_sec < 0) {
                static int warned __read_mostly;
 
-               *timeo_p = 0;
+               WRITE_ONCE(*timeo_p, 0);
                if (warned < 10 && net_ratelimit()) {
                        warned++;
                        pr_info("%s: `%s' (pid %d) tries to set negative timeout\n",
@@ -447,11 +448,12 @@ static int sock_set_timeout(long *timeo_p, sockptr_t optval, int optlen,
                }
                return 0;
        }
-       *timeo_p = MAX_SCHEDULE_TIMEOUT;
-       if (tv.tv_sec == 0 && tv.tv_usec == 0)
-               return 0;
-       if (tv.tv_sec < (MAX_SCHEDULE_TIMEOUT / HZ - 1))
-               *timeo_p = tv.tv_sec * HZ + DIV_ROUND_UP((unsigned long)tv.tv_usec, USEC_PER_SEC / HZ);
+       val = MAX_SCHEDULE_TIMEOUT;
+       if ((tv.tv_sec || tv.tv_usec) &&
+           (tv.tv_sec < (MAX_SCHEDULE_TIMEOUT / HZ - 1)))
+               val = tv.tv_sec * HZ + DIV_ROUND_UP((unsigned long)tv.tv_usec,
+                                                   USEC_PER_SEC / HZ);
+       WRITE_ONCE(*timeo_p, val);
        return 0;
 }
 
@@ -813,9 +815,9 @@ void sock_set_sndtimeo(struct sock *sk, s64 secs)
 {
        lock_sock(sk);
        if (secs && secs < MAX_SCHEDULE_TIMEOUT / HZ - 1)
-               sk->sk_sndtimeo = secs * HZ;
+               WRITE_ONCE(sk->sk_sndtimeo, secs * HZ);
        else
-               sk->sk_sndtimeo = MAX_SCHEDULE_TIMEOUT;
+               WRITE_ONCE(sk->sk_sndtimeo, MAX_SCHEDULE_TIMEOUT);
        release_sock(sk);
 }
 EXPORT_SYMBOL(sock_set_sndtimeo);
@@ -1721,12 +1723,14 @@ int sk_getsockopt(struct sock *sk, int level, int optname,
 
        case SO_RCVTIMEO_OLD:
        case SO_RCVTIMEO_NEW:
-               lv = sock_get_timeout(sk->sk_rcvtimeo, &v, SO_RCVTIMEO_OLD == optname);
+               lv = sock_get_timeout(READ_ONCE(sk->sk_rcvtimeo), &v,
+                                     SO_RCVTIMEO_OLD == optname);
                break;
 
        case SO_SNDTIMEO_OLD:
        case SO_SNDTIMEO_NEW:
-               lv = sock_get_timeout(sk->sk_sndtimeo, &v, SO_SNDTIMEO_OLD == optname);
+               lv = sock_get_timeout(READ_ONCE(sk->sk_sndtimeo), &v,
+                                     SO_SNDTIMEO_OLD == optname);
                break;
 
        case SO_RCVLOWAT:
index af85a73..6fdba06 100644 (file)
@@ -568,7 +568,7 @@ META_COLLECTOR(int_sk_rcvtimeo)
                *err = -1;
                return;
        }
-       dst->value = sk->sk_rcvtimeo / HZ;
+       dst->value = READ_ONCE(sk->sk_rcvtimeo) / HZ;
 }
 
 META_COLLECTOR(int_sk_sndtimeo)
@@ -579,7 +579,7 @@ META_COLLECTOR(int_sk_sndtimeo)
                *err = -1;
                return;
        }
-       dst->value = sk->sk_sndtimeo / HZ;
+       dst->value = READ_ONCE(sk->sk_sndtimeo) / HZ;
 }
 
 META_COLLECTOR(int_sk_sendmsg_off)