On Tue, Jan 31, 2023 at 06:57 PM +01, Eric Dumazet wrote:

On Tue, Jan 31, 2023 at 6:46 PM Jakub Sitnicki [off-list ref] wrote:

quoted

While UDP_GRO cmsg interface lacks documentation, the selftests added in
commit 3327a9c46352 ("selftests: add functionals test for UDP GRO") suggest
that the user-space should allocate CMSG_SPACE for an u16 value and
interpret the returned bytes as such:

static int recv_msg(int fd, char *buf, int len, int *gso_size)
{
        char control[CMSG_SPACE(sizeof(uint16_t))] = {0};
        ...
                        if (cmsg->cmsg_level == SOL_UDP
                            && cmsg->cmsg_type == UDP_GRO) {
                                gsosizeptr = (uint16_t *) CMSG_DATA(cmsg);
                                *gso_size = *gsosizeptr;
                                break;
                        }
        ...
}

Today user-space will receive 4 bytes of data with an UDP_GRO cmsg, because
the kernel packs an int into the cmsg data, as we can confirm with strace:

  recvmsg(8, {msg_name=...,
              msg_iov=[{iov_base="\0\0..."..., iov_len=96000}],
              msg_iovlen=1,
              msg_control=[{cmsg_len=20,         <-- sizeof(cmsghdr) + 4
                            cmsg_level=SOL_UDP,
                            cmsg_type=0x68}],    <-- UDP_GRO
                            msg_controllen=24,
                            msg_flags=0}, 0) = 11200

This means that either UDP_GRO selftests are broken on big endian, or this
is a programming error. Assume the latter and pass only the needed 2 bytes
of data with the cmsg.

Fixing it like that has an added advantage that the cmsg becomes compatible
with what is expected by UDP_SEGMENT cmsg. It becomes possible to reuse the
cmsg when GSO packets are received on one socket and sent out of another.

Fixes: bcd1665e3569 ("udp: add support for UDP_GRO cmsg")
Signed-off-by: Jakub Sitnicki <jakub@cloudflare.com>
---
 include/linux/udp.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/linux/udp.h b/include/linux/udp.h
index a2892e151644..44bb8d699248 100644
--- a/include/linux/udp.h
+++ b/include/linux/udp.h

@@ -125,7 +125,7 @@ static inline bool udp_get_no_check6_rx(struct sock *sk)
 static inline void udp_cmsg_recv(struct msghdr *msg, struct sock *sk,
                                 struct sk_buff *skb)
 {
-       int gso_size;
+       __u16 gso_size;

        if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4) {
                gso_size = skb_shinfo(skb)->gso_size;
--

2.39.1

This would break some applications.

I think the test can be fixed instead, this seems less risky.

Thanks for guidance. Will fix the test.