Date: Fri, 2 Feb 2001 01:07:25 +0000 From: Tony Finch <dot@dotat.at> To: Garrett Wollman <wollman@khavrinen.lcs.mit.edu> Cc: freebsd-net@FreeBSD.ORG Subject: proper uncorking for TCP_NOPUSH, was Re: sendfile() Message-ID: <20010202010725.Z70673@hand.dotat.at> In-Reply-To: <20010201230644.W70673@hand.dotat.at> References: <1217774688.20010201133139@163.net> <20010201023825.A71975@xor.obsecurity.org> <20010201180010.Q70673@hand.dotat.at> <200102012010.PAA79234@khavrinen.lcs.mit.edu> <20010201230644.W70673@hand.dotat.at>
index | next in thread | previous in thread | raw e-mail
Tony Finch <dot@dotat.at> wrote:
>Garrett Wollman <wollman@khavrinen.lcs.mit.edu> wrote:
>>Tony Finch <dot@dotat.at> wrote:
>>>
>>> For this reason turning off TCP_CORK pushes out queued data, but
>>> this isn't the case for TCP_NOPUSH.
>>
>>This is a long-standing bug. You are welcome to fix it.
>
>I've been looking at this and it seems to me to be simply a case of
>calling tcp_output() from tcp_ctloutput() if TF_NOPUSH is cleared.
OK, apart from my previous patch containing dumb cut&paste errors it
works :-)
I have a program (below) which forks; the parent listens for a
connection and then reads from it reporting the size of the reads; the
child connects to the parent, sets TCP_NOPUSH, writes 10000 single
bytes, unsets TCP_NOPUSH, sleeps for 10 seconds, then exits.
On an upatched FreeBSD the following happens. Note that the
single-byte writes get coalesced nicely. Note also that there is a six
second delay between the writer finishing and the readere receiving
the data.
:; ./a.out
981075049.891658 reader init
981075050.898416 writer init
981075050.899179 reader go
981075050.902703 writer go
981075050.939887 write finished
981075055.898738 read 8192
981075055.899084 read 1808
981075060.948514 writer exit
981075060.949197 read 0
981075060.949237 reader exit
On a patched FreeBSD this happens. The reader gets the data as soon as
the writer turns off TCP_NOPUSH :-)
:; ./a.out
981075103.647630 reader init
981075104.654745 writer init
981075104.655310 reader go
981075104.655433 writer go
981075104.690353 write finished
981075104.690863 read 8192
981075104.691037 read 1808
981075114.695098 writer exit
981075114.695640 read 0
981075114.695761 reader exit
Proper patch below.
Tony.
--
f.a.n.finch fanf@covalent.net dot@dotat.at
"Dead! And yet there he stands!"
Index: tcp_usrreq.c
===================================================================
RCS file: /home/ncvs/src/sys/netinet/tcp_usrreq.c,v
retrieving revision 1.51
diff -u -r1.51 tcp_usrreq.c
--- tcp_usrreq.c 2000/01/09 19:17:28 1.51
+++ tcp_usrreq.c 2001/02/02 00:21:01
@@ -896,7 +896,6 @@
switch (sopt->sopt_name) {
case TCP_NODELAY:
case TCP_NOOPT:
- case TCP_NOPUSH:
error = sooptcopyin(sopt, &optval, sizeof optval,
sizeof optval);
if (error)
@@ -909,9 +908,6 @@
case TCP_NOOPT:
opt = TF_NOOPT;
break;
- case TCP_NOPUSH:
- opt = TF_NOPUSH;
- break;
default:
opt = 0; /* dead code to fool gcc */
break;
@@ -921,6 +917,20 @@
tp->t_flags |= opt;
else
tp->t_flags &= ~opt;
+ break;
+
+ case TCP_NOPUSH:
+ error = sooptcopyin(sopt, &optval, sizeof optval,
+ sizeof optval);
+ if (error)
+ break;
+
+ if (optval)
+ tp->t_flags |= TF_NOPUSH;
+ else {
+ tp->t_flags &= ~TF_NOPUSH;
+ error = tcp_output(tp);
+ }
break;
case TCP_MAXSEG:
#include <sys/types.h>
#include <sys/socket.h>
#include <sys/time.h>
#include <netinet/in.h>
#include <netinet/tcp.h>
#include <err.h>
#include <stdarg.h>
#include <stdio.h>
#include <string.h>
#include <unistd.h>
#define PORT_FOO 55555
static void
report(const char *fmt, ...) {
struct timeval tv;
va_list ap;
char buf[80];
va_start(ap, fmt);
if(gettimeofday(&tv, NULL)) err(1, "gettimeofday");
vsnprintf(buf, sizeof(buf), fmt, ap);
printf("%ld.%06ld %s\n", tv.tv_sec, tv.tv_usec, buf);
va_end(ap);
}
int
main(void) {
int i, r, s;
struct sockaddr_in addr;
char buf[8192];
setbuf(stdout, NULL);
bzero(&addr, sizeof(addr));
addr.sin_family = AF_INET;
addr.sin_port = htons(PORT_FOO);
addr.sin_addr.s_addr = htonl(INADDR_LOOPBACK);
switch(fork()) {
case(-1):
err(1, "fork");
case(0):
/* child */
sleep(1);
report("writer init");
s = socket(PF_INET, SOCK_STREAM, 0);
if(s < 0) err(1, "socket");
i = 1; r = setsockopt(s, IPPROTO_TCP, TCP_NOPUSH, &i, sizeof(i));
if(r < 0) err(1, "setsockopt");
r = connect(s, (struct sockaddr *)&addr, sizeof(addr));
if(r < 0) err(1, "connect");
report("writer go");
*buf = 0;
for(i = 0; i < 10000; i++) {
r = write(s, buf, 1);
if(r < 0) err(1, "write");
}
report("write finished");
i = 0; r = setsockopt(s, IPPROTO_TCP, TCP_NOPUSH, &i, sizeof(i));
if(r < 0) err(1, "setsockopt");
sleep(10);
report("writer exit");
exit(0);
default:
/* parent */
report("reader init");
s = socket(PF_INET, SOCK_STREAM, 0);
if(s < 0) err(1, "socket");
i = 1; r = setsockopt(s, SOL_SOCKET, SO_REUSEADDR, &i, sizeof(i));
if(r < 0) err(1, "setsockopt");
r = bind(s, (struct sockaddr *)&addr, sizeof(addr));
if(r < 0) err(1, "bind");
r = listen(s, SOMAXCONN);
if(r < 0) err(1, "listen");
s = accept(s, (struct sockaddr *)&addr, &i);
if(s < 0) err(1, "accept");
report("reader go");
do {
r = read(s, buf, sizeof(buf));
if(r < 0) err(1, "read");
report("read %d", r);
} while(r);
report("reader exit");
exit(0);
}
return(0);
}
To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-net" in the body of the message
home |
help
Want to link to this message? Use this
URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20010202010725.Z70673>