Re: TCP_DEFER_ACCEPT is missing counter update

[Willy Tarreau <w@1wt.eu>]

Hi,

On Fri, Oct 16, 2009 at 12:40:11PM +0200, Eric Dumazet wrote:
> Julian Anastasov a écrit :
> > 	Hello,
> > 
> > On Fri, 16 Oct 2009, Willy Tarreau wrote:
> > 
> >>> 	This will need little change in inet_csk_reqsk_queue_prune()
> >>> but it saves SYN-ACK traffic during deferring period in the
> >>> common case when client sends ACK. If such compromise is
> >>> acceptable I can prepare and test some patch.
> >> I would personally like this a lot ! This will satisfy people who
> >> expect it to establish at the end of the "TCP_DEFER_ACCEPT delay"
> >> as can be interpreted from the man page, will reduce the number of
> >> useless SYN-ACKs that annoy other people while still making no
> >> visible change for anyone who would rely on the current behaviour.
> > 
> > 	OK, I don't have much time now, this is what I'm
> > going to test later today and later can provide proper comments:
> > 
> > Signed-off-by: Julian Anastasov <ja@ssi.bg>
> 
> I tested both patches and they perform very well, thank you !
> 
> For the minimum 1 sec value, tcpdump looks like :
> 12:32:03.850456 IP 127.0.0.1.20000 > 127.0.0.1.2222: S 1879889239:1879889239(0) win 32792 <mss 16396,nop,nop,timestamp 952803 0,nop,wscale 6>
> 12:32:03.850463 IP 127.0.0.1.2222 > 127.0.0.1.20000: S 1890330616:1890330616(0) ack 1879889240 win 32768 <mss 16396,nop,nop,timestamp 952803 952803,nop,wscale 6>
> 12:32:03.850469 IP 127.0.0.1.20000 > 127.0.0.1.2222: . ack 1 win 513 <nop,nop,timestamp 952803 952803>
> 
> 12:32:06.849989 IP 127.0.0.1.2222 > 127.0.0.1.20000: S 1890330616:1890330616(0) ack 1879889240 win 32768 <mss 16396,nop,nop,timestamp 955803 952803,nop,wscale 6>
> 12:32:06.849996 IP 127.0.0.1.20000 > 127.0.0.1.2222: . ack 1 win 513 <nop,nop,timestamp 955803 955803>
> 
> So listening application gets the accept() 3 seconds after initial SYN

Excellent! Nice work guys.

> # ss -emoian | grep SYN-RECV
> SYN-RECV   0      0                 127.0.0.1:2222             127.0.0.1:20000  timer:(on,24sec,4) ino:0 sk:f6f0ec80
> 
> I wonder if tcp_diag should be extented a bit to reflect fact that the ACK was received from client
> (ie forward the inet_rsk(req)->acked information to idiag_rqueue)

I personally have no opinion on this point.

Thanks!
Willy