net: silence KCSAN warnings around sk_add_backlog() calls
authorEric Dumazet <edumazet@google.com>
Wed, 9 Oct 2019 22:21:13 +0000 (15:21 -0700)
committerJakub Kicinski <jakub.kicinski@netronome.com>
Thu, 10 Oct 2019 04:42:59 +0000 (21:42 -0700)
sk_add_backlog() callers usually read sk->sk_rcvbuf without
owning the socket lock. This means sk_rcvbuf value can
be changed by other cpus, and KCSAN complains.

Add READ_ONCE() annotations to document the lockless nature
of these reads.

Note that writes over sk_rcvbuf should also use WRITE_ONCE(),
but this will be done in separate patches to ease stable
backports (if we decide this is relevant for stable trees).

BUG: KCSAN: data-race in tcp_add_backlog / tcp_recvmsg

write to 0xffff88812ab369f8 of 8 bytes by interrupt on cpu 1:
 __sk_add_backlog include/net/sock.h:902 [inline]
 sk_add_backlog include/net/sock.h:933 [inline]
 tcp_add_backlog+0x45a/0xcc0 net/ipv4/tcp_ipv4.c:1737
 tcp_v4_rcv+0x1aba/0x1bf0 net/ipv4/tcp_ipv4.c:1925
 ip_protocol_deliver_rcu+0x51/0x470 net/ipv4/ip_input.c:204
 ip_local_deliver_finish+0x110/0x140 net/ipv4/ip_input.c:231
 NF_HOOK include/linux/netfilter.h:305 [inline]
 NF_HOOK include/linux/netfilter.h:299 [inline]
 ip_local_deliver+0x133/0x210 net/ipv4/ip_input.c:252
 dst_input include/net/dst.h:442 [inline]
 ip_rcv_finish+0x121/0x160 net/ipv4/ip_input.c:413
 NF_HOOK include/linux/netfilter.h:305 [inline]
 NF_HOOK include/linux/netfilter.h:299 [inline]
 ip_rcv+0x18f/0x1a0 net/ipv4/ip_input.c:523
 __netif_receive_skb_one_core+0xa7/0xe0 net/core/dev.c:5004
 __netif_receive_skb+0x37/0xf0 net/core/dev.c:5118
 netif_receive_skb_internal+0x59/0x190 net/core/dev.c:5208
 napi_skb_finish net/core/dev.c:5671 [inline]
 napi_gro_receive+0x28f/0x330 net/core/dev.c:5704
 receive_buf+0x284/0x30b0 drivers/net/virtio_net.c:1061
 virtnet_receive drivers/net/virtio_net.c:1323 [inline]
 virtnet_poll+0x436/0x7d0 drivers/net/virtio_net.c:1428
 napi_poll net/core/dev.c:6352 [inline]
 net_rx_action+0x3ae/0xa50 net/core/dev.c:6418

read to 0xffff88812ab369f8 of 8 bytes by task 7271 on cpu 0:
 tcp_recvmsg+0x470/0x1a30 net/ipv4/tcp.c:2047
 inet_recvmsg+0xbb/0x250 net/ipv4/af_inet.c:838
 sock_recvmsg_nosec net/socket.c:871 [inline]
 sock_recvmsg net/socket.c:889 [inline]
 sock_recvmsg+0x92/0xb0 net/socket.c:885
 sock_read_iter+0x15f/0x1e0 net/socket.c:967
 call_read_iter include/linux/fs.h:1864 [inline]
 new_sync_read+0x389/0x4f0 fs/read_write.c:414
 __vfs_read+0xb1/0xc0 fs/read_write.c:427
 vfs_read fs/read_write.c:461 [inline]
 vfs_read+0x143/0x2c0 fs/read_write.c:446
 ksys_read+0xd5/0x1b0 fs/read_write.c:587
 __do_sys_read fs/read_write.c:597 [inline]
 __se_sys_read fs/read_write.c:595 [inline]
 __x64_sys_read+0x4c/0x60 fs/read_write.c:595
 do_syscall_64+0xcf/0x2f0 arch/x86/entry/common.c:296
 entry_SYSCALL_64_after_hwframe+0x44/0xa9

Reported by Kernel Concurrency Sanitizer on:
CPU: 0 PID: 7271 Comm: syz-fuzzer Not tainted 5.3.0+ #0
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011

Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: syzbot <syzkaller@googlegroups.com>
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
net/core/sock.c
net/ipv4/tcp_ipv4.c
net/llc/llc_conn.c
net/sctp/input.c
net/tipc/socket.c
net/x25/x25_dev.c

index 50647a10fdb7f050e963e2734f0d3555fa4bd7aa..1cf06934da50b98fccc849d396680cee46badb7d 100644 (file)
@@ -522,7 +522,7 @@ int __sk_receive_skb(struct sock *sk, struct sk_buff *skb,
                rc = sk_backlog_rcv(sk, skb);
 
                mutex_release(&sk->sk_lock.dep_map, 1, _RET_IP_);
-       } else if (sk_add_backlog(sk, skb, sk->sk_rcvbuf)) {
+       } else if (sk_add_backlog(sk, skb, READ_ONCE(sk->sk_rcvbuf))) {
                bh_unlock_sock(sk);
                atomic_inc(&sk->sk_drops);
                goto discard_and_relse;
index bf124b1742df864a3007d137ff31c8bfb2bee12a..492bf6a6b0237a677aae5d7ef365a5fc7356e238 100644 (file)
@@ -1644,7 +1644,7 @@ int tcp_v4_early_demux(struct sk_buff *skb)
 
 bool tcp_add_backlog(struct sock *sk, struct sk_buff *skb)
 {
-       u32 limit = sk->sk_rcvbuf + sk->sk_sndbuf;
+       u32 limit = READ_ONCE(sk->sk_rcvbuf) + READ_ONCE(sk->sk_sndbuf);
        struct skb_shared_info *shinfo;
        const struct tcphdr *th;
        struct tcphdr *thtail;
index a79b739eb223668aa837ac08bbcebc0ae5750a0a..7b620acaca9ec194e03b590c21fb004401dbc7ce 100644 (file)
@@ -813,7 +813,7 @@ void llc_conn_handler(struct llc_sap *sap, struct sk_buff *skb)
        else {
                dprintk("%s: adding to backlog...\n", __func__);
                llc_set_backlog_type(skb, LLC_PACKET);
-               if (sk_add_backlog(sk, skb, sk->sk_rcvbuf))
+               if (sk_add_backlog(sk, skb, READ_ONCE(sk->sk_rcvbuf)))
                        goto drop_unlock;
        }
 out:
index f2771375bfc06952e783108e28f850b44e479421..2277981559d0700669807dff4981e23e03be66ef 100644 (file)
@@ -322,7 +322,7 @@ int sctp_backlog_rcv(struct sock *sk, struct sk_buff *skb)
                bh_lock_sock(sk);
 
                if (sock_owned_by_user(sk) || !sctp_newsk_ready(sk)) {
-                       if (sk_add_backlog(sk, skb, sk->sk_rcvbuf))
+                       if (sk_add_backlog(sk, skb, READ_ONCE(sk->sk_rcvbuf)))
                                sctp_chunk_free(chunk);
                        else
                                backloged = 1;
@@ -337,7 +337,7 @@ int sctp_backlog_rcv(struct sock *sk, struct sk_buff *skb)
                        return 0;
        } else {
                if (!sctp_newsk_ready(sk)) {
-                       if (!sk_add_backlog(sk, skb, sk->sk_rcvbuf))
+                       if (!sk_add_backlog(sk, skb, READ_ONCE(sk->sk_rcvbuf)))
                                return 0;
                        sctp_chunk_free(chunk);
                } else {
@@ -364,7 +364,7 @@ static int sctp_add_backlog(struct sock *sk, struct sk_buff *skb)
        struct sctp_ep_common *rcvr = chunk->rcvr;
        int ret;
 
-       ret = sk_add_backlog(sk, skb, sk->sk_rcvbuf);
+       ret = sk_add_backlog(sk, skb, READ_ONCE(sk->sk_rcvbuf));
        if (!ret) {
                /* Hold the assoc/ep while hanging on the backlog queue.
                 * This way, we know structures we need will not disappear
index 3b9f8cc328f5c3aa2f39b5a8267bd206e714835f..7c736cfec57f8428cdc16500fa70845716ed9fae 100644 (file)
@@ -2119,13 +2119,13 @@ static unsigned int rcvbuf_limit(struct sock *sk, struct sk_buff *skb)
        struct tipc_msg *hdr = buf_msg(skb);
 
        if (unlikely(msg_in_group(hdr)))
-               return sk->sk_rcvbuf;
+               return READ_ONCE(sk->sk_rcvbuf);
 
        if (unlikely(!msg_connected(hdr)))
-               return sk->sk_rcvbuf << msg_importance(hdr);
+               return READ_ONCE(sk->sk_rcvbuf) << msg_importance(hdr);
 
        if (likely(tsk->peer_caps & TIPC_BLOCK_FLOWCTL))
-               return sk->sk_rcvbuf;
+               return READ_ONCE(sk->sk_rcvbuf);
 
        return FLOWCTL_MSG_LIM;
 }
index 5c111bc3c8ea5a8c079cd0d83733e7b253d47158..00e782335cb0740daa172f0c1835bcc3285ba46b 100644 (file)
@@ -55,7 +55,7 @@ static int x25_receive_data(struct sk_buff *skb, struct x25_neigh *nb)
                if (!sock_owned_by_user(sk)) {
                        queued = x25_process_rx_frame(sk, skb);
                } else {
-                       queued = !sk_add_backlog(sk, skb, sk->sk_rcvbuf);
+                       queued = !sk_add_backlog(sk, skb, READ_ONCE(sk->sk_rcvbuf));
                }
                bh_unlock_sock(sk);
                sock_put(sk);