udp: fix data-race in udp_set_dev_scratch()
authorEric Dumazet <edumazet@google.com>
Thu, 24 Oct 2019 18:43:31 +0000 (11:43 -0700)
committerDavid S. Miller <davem@davemloft.net>
Mon, 28 Oct 2019 20:53:40 +0000 (13:53 -0700)
KCSAN reported a data-race in udp_set_dev_scratch() [1]

The issue here is that we must not write over skb fields
if skb is shared. A similar issue has been fixed in commit
89c22d8c3b27 ("net: Fix skb csum races when peeking")

While we are at it, use a helper only dealing with
udp_skb_scratch(skb)->csum_unnecessary, as this allows
udp_set_dev_scratch() to be called once and thus inlined.

[1]
BUG: KCSAN: data-race in udp_set_dev_scratch / udpv6_recvmsg

write to 0xffff888120278317 of 1 bytes by task 10411 on cpu 1:
 udp_set_dev_scratch+0xea/0x200 net/ipv4/udp.c:1308
 __first_packet_length+0x147/0x420 net/ipv4/udp.c:1556
 first_packet_length+0x68/0x2a0 net/ipv4/udp.c:1579
 udp_poll+0xea/0x110 net/ipv4/udp.c:2720
 sock_poll+0xed/0x250 net/socket.c:1256
 vfs_poll include/linux/poll.h:90 [inline]
 do_select+0x7d0/0x1020 fs/select.c:534
 core_sys_select+0x381/0x550 fs/select.c:677
 do_pselect.constprop.0+0x11d/0x160 fs/select.c:759
 __do_sys_pselect6 fs/select.c:784 [inline]
 __se_sys_pselect6 fs/select.c:769 [inline]
 __x64_sys_pselect6+0x12e/0x170 fs/select.c:769
 do_syscall_64+0xcc/0x370 arch/x86/entry/common.c:290
 entry_SYSCALL_64_after_hwframe+0x44/0xa9

read to 0xffff888120278317 of 1 bytes by task 10413 on cpu 0:
 udp_skb_csum_unnecessary include/net/udp.h:358 [inline]
 udpv6_recvmsg+0x43e/0xe90 net/ipv6/udp.c:310
 inet6_recvmsg+0xbb/0x240 net/ipv6/af_inet6.c:592
 sock_recvmsg_nosec+0x5c/0x70 net/socket.c:871
 ___sys_recvmsg+0x1a0/0x3e0 net/socket.c:2480
 do_recvmmsg+0x19a/0x5c0 net/socket.c:2601
 __sys_recvmmsg+0x1ef/0x200 net/socket.c:2680
 __do_sys_recvmmsg net/socket.c:2703 [inline]
 __se_sys_recvmmsg net/socket.c:2696 [inline]
 __x64_sys_recvmmsg+0x89/0xb0 net/socket.c:2696
 do_syscall_64+0xcc/0x370 arch/x86/entry/common.c:290
 entry_SYSCALL_64_after_hwframe+0x44/0xa9

Reported by Kernel Concurrency Sanitizer on:
CPU: 0 PID: 10413 Comm: syz-executor.0 Not tainted 5.4.0-rc3+ #0
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011

Fixes: 2276f58ac589 ("udp: use a separate rx queue for packet reception")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: syzbot <syzkaller@googlegroups.com>
Cc: Paolo Abeni <pabeni@redhat.com>
Reviewed-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/ipv4/udp.c

index 345a3d43f5a655e009e99c16bb19e047cdf003c6..d1ed160af202c054839387201abd3f13b55d00e9 100644 (file)
@@ -1316,6 +1316,20 @@ static void udp_set_dev_scratch(struct sk_buff *skb)
                scratch->_tsize_state |= UDP_SKB_IS_STATELESS;
 }
 
+static void udp_skb_csum_unnecessary_set(struct sk_buff *skb)
+{
+       /* We come here after udp_lib_checksum_complete() returned 0.
+        * This means that __skb_checksum_complete() might have
+        * set skb->csum_valid to 1.
+        * On 64bit platforms, we can set csum_unnecessary
+        * to true, but only if the skb is not shared.
+        */
+#if BITS_PER_LONG == 64
+       if (!skb_shared(skb))
+               udp_skb_scratch(skb)->csum_unnecessary = true;
+#endif
+}
+
 static int udp_skb_truesize(struct sk_buff *skb)
 {
        return udp_skb_scratch(skb)->_tsize_state & ~UDP_SKB_IS_STATELESS;
@@ -1550,10 +1564,7 @@ static struct sk_buff *__first_packet_length(struct sock *sk,
                        *total += skb->truesize;
                        kfree_skb(skb);
                } else {
-                       /* the csum related bits could be changed, refresh
-                        * the scratch area
-                        */
-                       udp_set_dev_scratch(skb);
+                       udp_skb_csum_unnecessary_set(skb);
                        break;
                }
        }