net: reduce skb_warn_bad_offload() noise
authorEric Dumazet <edumazet@google.com>
Tue, 31 Jan 2017 18:20:32 +0000 (10:20 -0800)
committerDavid S. Miller <davem@davemloft.net>
Wed, 1 Feb 2017 17:02:48 +0000 (12:02 -0500)
Dmitry reported warnings occurring in __skb_gso_segment() [1]

All SKB_GSO_DODGY producers can allow user space to feed
packets that trigger the current check.

We could prevent them from doing so, rejecting packets, but
this might add regressions to existing programs.

It turns out our SKB_GSO_DODGY handlers properly set up checksum
information that is needed anyway when packets needs to be segmented.

By checking again skb_needs_check() after skb_mac_gso_segment(),
we should remove these pesky warnings, at a very minor cost.

With help from Willem de Bruijn

[1]
WARNING: CPU: 1 PID: 6768 at net/core/dev.c:2439 skb_warn_bad_offload+0x2af/0x390 net/core/dev.c:2434
lo: caps=(0x000000a2803b7c69, 0x0000000000000000) len=138 data_len=0 gso_size=15883 gso_type=4 ip_summed=0
Kernel panic - not syncing: panic_on_warn set ...

CPU: 1 PID: 6768 Comm: syz-executor1 Not tainted 4.9.0 #5
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011
 ffff8801c063ecd8 ffffffff82346bdf ffffffff00000001 1ffff100380c7d2e
 ffffed00380c7d26 0000000041b58ab3 ffffffff84b37e38 ffffffff823468f1
 ffffffff84820740 ffffffff84f289c0 dffffc0000000000 ffff8801c063ee20
Call Trace:
 [<ffffffff82346bdf>] __dump_stack lib/dump_stack.c:15 [inline]
 [<ffffffff82346bdf>] dump_stack+0x2ee/0x3ef lib/dump_stack.c:51
 [<ffffffff81827e34>] panic+0x1fb/0x412 kernel/panic.c:179
 [<ffffffff8141f704>] __warn+0x1c4/0x1e0 kernel/panic.c:542
 [<ffffffff8141f7e5>] warn_slowpath_fmt+0xc5/0x100 kernel/panic.c:565
 [<ffffffff8356cbaf>] skb_warn_bad_offload+0x2af/0x390 net/core/dev.c:2434
 [<ffffffff83585cd2>] __skb_gso_segment+0x482/0x780 net/core/dev.c:2706
 [<ffffffff83586f19>] skb_gso_segment include/linux/netdevice.h:3985 [inline]
 [<ffffffff83586f19>] validate_xmit_skb+0x5c9/0xc20 net/core/dev.c:2969
 [<ffffffff835892bb>] __dev_queue_xmit+0xe6b/0x1e70 net/core/dev.c:3383
 [<ffffffff8358a2d7>] dev_queue_xmit+0x17/0x20 net/core/dev.c:3424
 [<ffffffff83ad161d>] packet_snd net/packet/af_packet.c:2930 [inline]
 [<ffffffff83ad161d>] packet_sendmsg+0x32ed/0x4d30 net/packet/af_packet.c:2955
 [<ffffffff834f0aaa>] sock_sendmsg_nosec net/socket.c:621 [inline]
 [<ffffffff834f0aaa>] sock_sendmsg+0xca/0x110 net/socket.c:631
 [<ffffffff834f329a>] ___sys_sendmsg+0x8fa/0x9f0 net/socket.c:1954
 [<ffffffff834f5e58>] __sys_sendmsg+0x138/0x300 net/socket.c:1988
 [<ffffffff834f604d>] SYSC_sendmsg net/socket.c:1999 [inline]
 [<ffffffff834f604d>] SyS_sendmsg+0x2d/0x50 net/socket.c:1995
 [<ffffffff84371941>] entry_SYSCALL_64_fastpath+0x1f/0xc2

Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: Dmitry Vyukov <dvyukov@google.com>
Cc: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/core/dev.c

index e61528c..727b6fd 100644 (file)
@@ -2658,11 +2658,12 @@ static inline bool skb_needs_check(struct sk_buff *skb, bool tx_path)
 struct sk_buff *__skb_gso_segment(struct sk_buff *skb,
                                  netdev_features_t features, bool tx_path)
 {
+       struct sk_buff *segs;
+
        if (unlikely(skb_needs_check(skb, tx_path))) {
                int err;
 
-               skb_warn_bad_offload(skb);
-
+               /* We're going to init ->check field in TCP or UDP header */
                err = skb_cow_head(skb, 0);
                if (err < 0)
                        return ERR_PTR(err);
@@ -2690,7 +2691,12 @@ struct sk_buff *__skb_gso_segment(struct sk_buff *skb,
        skb_reset_mac_header(skb);
        skb_reset_mac_len(skb);
 
-       return skb_mac_gso_segment(skb, features);
+       segs = skb_mac_gso_segment(skb, features);
+
+       if (unlikely(skb_needs_check(skb, tx_path)))
+               skb_warn_bad_offload(skb);
+
+       return segs;
 }
 EXPORT_SYMBOL(__skb_gso_segment);