Re: [PATCH v5 net-next 3/5] tcp: add TCP support for low latency receive poll.

On 28/05/2013 03:36, Eric Dumazet wrote:

On Mon, 2013-05-27 at 10:44 +0300, Eliezer Tamir wrote:

quoted

adds busy-poll support for TCP.

Really, this is a small changelog for such an addition :(

OK

How poll()/epoll() is supported ?

poll()/select() are done by the code added to fs/select.c in 2/5.
epoll() is not yet supported.

quoted

diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
index 9579e1a..4d82939 100644
--- a/net/ipv4/tcp_input.c
+++ b/net/ipv4/tcp_input.c

@@ -74,6 +74,7 @@
  #include <linux/ipsec.h>
  #include <asm/unaligned.h>
  #include <net/netdma.h>
+#include <net/ll_poll.h>

Not sure why this include is needed in this file ?

You added nothing else but this line.

This is a mistake, a remnant from an earlier version when sk_mark_ll() was where we copy data to the socket.
I will remove it.

quoted

  #include <net/netdma.h>
  #include <net/secure_seq.h>
  #include <net/tcp_memcontrol.h>
+#include <net/ll_poll.h>

  #include <linux/inet.h>
  #include <linux/ipv6.h>

@@ -2011,6 +2012,7 @@ process:
  	if (sk_filter(sk, skb))
  		goto discard_and_relse;

+	sk_mark_ll(sk, skb);
  	skb->dev = NULL;

  	bh_lock_sock_nested(sk);

How IPv6 is handled ?

IPv6 is currently not supported (it was not supported in any version of
this patch set, the POC code in fact was hard-codded for UDPv4/TCPv4).

If there is interest, I will add it, I think it will not be complicated.
However, I would prefer to wait with that for a second stage.

My main concern is that adding IPv6 will significantly increase my
 testing effort, which is already 90% of what I'm spending time on.

IMHO epoll() and a more robust support for select()/poll() should have
a higher priority, but I'm open to suggestions.

I would like to get what we have so far applied so more people can try
it, then work on all of the other things that we need.

Dave, I would like to hear your opinion on this, please.

-Eliezer