[RFC PATCH net] tcp: allow to use TCP Fastopen with MSG_ZEROCOPY

[RFC PATCH net] tcp: allow to use TCP Fastopen with MSG_ZEROCOPY

[Alexey Kodanev]

With TCP Fastopen we can have the following cases, which could also
use MSG_ZEROCOPY flag with send() and sendto():

* sendto() + MSG_FASTOPEN flag, sk state can be in TCP_CLOSE at
  the start of tcp_sendmsg()

* set socket option TCP_FASTOPEN_CONNECT, then connect()
  and send(), sk state in TCP_SYN_SENT

Currently, both cases with tcp_sendmsg() and MSG_ZEROCOPY flag results
to EINVAL error, because of the check for TCP_ESTABLISHED sk state in
the beginning of tcp_sendmsg().

Both conditions require two more checks there: !tp->fastopen_connect
and !(flags & MSG_FASTOPEN). It looks like we could remove the original
check altogether for this unlikely event instead. That way tcp_sendmsg()
without TFO should fail with EPIPE on sk_stream_wait_connect(), as
before the introduction of MSG_ZEROCOPY there. And work smoothly for
the TFO cases.

Fixes: f214f915e7db ("tcp: enable MSG_ZEROCOPY")
Signed-off-by: Alexey Kodanev <alexey.kodanev@oracle.com>
---

Is there something that I've overlooked and we can't use it here, and
we should handle this type of error, while using sendto() + TFO,
in userspace?

 net/ipv4/tcp.c | 5 -----
 1 file changed, 5 deletions(-)

diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
index 9225610..768f02c 100644
--- a/net/ipv4/tcp.c
+++ b/net/ipv4/tcp.c
@@ -1193,11 +1193,6 @@ int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
 	flags = msg->msg_flags;
 
 	if (flags & MSG_ZEROCOPY && size) {
-		if (sk->sk_state != TCP_ESTABLISHED) {
-			err = -EINVAL;
-			goto out_err;
-		}
-
 		skb = tcp_write_queue_tail(sk);
 		uarg = sock_zerocopy_realloc(sk, size, skb_zcopy(skb));
 		if (!uarg) {
-- 
1.8.3.1

Re: [RFC PATCH net] tcp: allow to use TCP Fastopen with MSG_ZEROCOPY

[Willem de Bruijn]

On Tue, Apr 3, 2018 at 2:43 PM, Alexey Kodanev
<alexey.kodanev@oracle.com> wrote:
> With TCP Fastopen we can have the following cases, which could also
> use MSG_ZEROCOPY flag with send() and sendto():
>
> * sendto() + MSG_FASTOPEN flag, sk state can be in TCP_CLOSE at
>   the start of tcp_sendmsg()
>
> * set socket option TCP_FASTOPEN_CONNECT, then connect()
>   and send(), sk state in TCP_SYN_SENT
>
> Currently, both cases with tcp_sendmsg() and MSG_ZEROCOPY flag results
> to EINVAL error, because of the check for TCP_ESTABLISHED sk state in
> the beginning of tcp_sendmsg().
>
> Both conditions require two more checks there: !tp->fastopen_connect
> and !(flags & MSG_FASTOPEN).  It looks like we could remove the original
> check altogether for this unlikely event instead. That way tcp_sendmsg()
> without TFO should fail with EPIPE on sk_stream_wait_connect(), as
> before the introduction of MSG_ZEROCOPY there. And work smoothly for
> the TFO cases.
>
> Fixes: f214f915e7db ("tcp: enable MSG_ZEROCOPY")

This patch adds MSG_ZEROCOPY support for TFO sockets. It is not
a fix that needs to go to stable.

> Signed-off-by: Alexey Kodanev <alexey.kodanev@oracle.com>
> ---
>
> Is there something that I've overlooked and we can't use it here, and
> we should handle this type of error, while using sendto() + TFO,
> in userspace?
>
>  net/ipv4/tcp.c | 5 -----
>  1 file changed, 5 deletions(-)
>
> diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
> index 9225610..768f02c 100644
> --- a/net/ipv4/tcp.c
> +++ b/net/ipv4/tcp.c
> @@ -1193,11 +1193,6 @@ int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
>         flags = msg->msg_flags;
>
>         if (flags & MSG_ZEROCOPY && size) {
> -               if (sk->sk_state != TCP_ESTABLISHED) {
> -                       err = -EINVAL;
> -                       goto out_err;
> -               }
> -
>                 skb = tcp_write_queue_tail(sk);
>                 uarg = sock_zerocopy_realloc(sk, size, skb_zcopy(skb));
>                 if (!uarg) {
> --
> 1.8.3.1
>