Re: [PATCH v3 6/8] udp: Unify udp_sock_handler_splice() with udp_sock_handler()

[Stefano Brivio <sbrivio@redhat.com>]

On Wed,  4 Jan 2023 16:44:24 +1100
David Gibson <david@gibson.dropbear.id.au> wrote:

> These two functions now have a very similar structure, and their first
> part (calling recvmmsg()) is functionally identical.  So, merge the two
> functions into one.
> 
> This does have the side effect of meaning we no longer receive multiple
> packets at once for splice (we already didn't for tap).  This does hurt
> throughput for small spliced packets, but improves it for large spliced
> packets and tap packets.
> 
> Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
> ---
>  udp.c | 95 ++++++++++++++++++++++-------------------------------------
>  1 file changed, 35 insertions(+), 60 deletions(-)
> 
> diff --git a/udp.c b/udp.c
> index f6c07a5..32ca83e 100644
> --- a/udp.c
> +++ b/udp.c
> @@ -590,52 +590,6 @@ static void udp_splice_sendfrom(const struct ctx *c, unsigned start, unsigned n,
>  	sendmmsg(s, mmh_send + start, n, MSG_NOSIGNAL);
>  }
>  
> -/**
> - * udp_sock_handler_splice() - Handler for socket mapped to "spliced" connection
> - * @c:		Execution context
> - * @ref:	epoll reference
> - * @events:	epoll events bitmap
> - * @now:	Current timestamp
> - */
> -static void udp_sock_handler_splice(const struct ctx *c, union epoll_ref ref,
> -				    uint32_t events, const struct timespec *now)
> -{
> -	in_port_t dst = ref.r.p.udp.udp.port;
> -	int v6 = ref.r.p.udp.udp.v6, n, i, m;
> -	struct mmsghdr *mmh_recv;
> -
> -	if (!(events & EPOLLIN))
> -		return;
> -
> -	if (v6)
> -		mmh_recv = udp6_l2_mh_sock;
> -	else
> -		mmh_recv = udp4_l2_mh_sock;
> -
> -	n = recvmmsg(ref.r.s, mmh_recv, UDP_MAX_FRAMES, 0, NULL);
> -
> -	if (n <= 0)
> -		return;
> -
> -	if (v6)
> -		udp6_localname.sin6_port = htons(dst);
> -	else
> -		udp4_localname.sin_port = htons(dst);
> -
> -	for (i = 0; i < n; i += m) {
> -		in_port_t src = sa_port(v6, mmh_recv[i].msg_hdr.msg_name);
> -
> -		for (m = 1; i + m < n; m++) {
> -			void *mname = mmh_recv[i + m].msg_hdr.msg_name;
> -			if (sa_port(v6, mname) != src)
> -				break;
> -		}
> -
> -		udp_splice_sendfrom(c, i, m, src, dst, v6, ref.r.p.udp.udp.ns,
> -				    ref.r.p.udp.udp.orig, now);
> -	}
> -}
> -
>  /**
>   * udp_update_hdr4() - Update headers for one IPv4 datagram
>   * @c:		Execution context
> @@ -944,32 +898,53 @@ void udp_sock_handler(const struct ctx *c, union epoll_ref ref, uint32_t events,
>  		      const struct timespec *now)
>  {
>  	/* For not entirely clear reasons (data locality?) pasta gets
> -	 * better throughput if we receive the datagrams one at a
> -	 * time.
> +	 * better throughput if we receive tap datagrams one at a
> +	 * atime.  For small splice datagrams throughput is slightly
> +	 * better if we do batch, but it's slightly worse for large
> +	 * splice datagrams.  Since we don't know before we receive
> +	 * whether we'll use tap or splice, always go one at a time
> +	 * for pasta mode.
>  	 */
>  	ssize_t n = (c->mode == MODE_PASST ? UDP_MAX_FRAMES : 1);
>  	in_port_t dstport = ref.r.p.udp.udp.port;
>  	bool v6 = ref.r.p.udp.udp.v6;
> -	struct mmsghdr *sock_mmh;
> +	struct mmsghdr *mmh_recv;
> +	unsigned int i, m;
> +	ssize_t n;

This doesn't build, you're redefining 'n' after the new version of 4/8.

I could drop this on merge (the rest of the series would be ready to
merge) but as you usually prefer to respin anyway, I'll wait for that.

-- 
Stefano