hi linus-
sock_writeable determines whether there is space in a socket's output
buffer. socket write_space callbacks use it to determine whether to wake
up those that are waiting for more output buffer space.
however, sock_writeable is not appropriate for TCP sockets. because the
RPC layer's write_space callback uses it for TCP sockets, the RPC layer
hammers on sock_sendmsg with dozens of write requests that are only a few
hundred bytes long when it is trying to send a large write RPC request.
this patch adds logic to the RPC layer's write_space callback that
properly handles TCP sockets.
patch reviewed by Trond and Alexey. patch forthcoming for 2.4.20-pre.
diff -drN -U2 01-lock_write/net/sunrpc/xprt.c 02-write_space/net/sunrpc/xprt.c
--- 01-lock_write/net/sunrpc/xprt.c Wed Aug 28 17:09:08 2002
+++ 02-write_space/net/sunrpc/xprt.c Wed Aug 28 17:18:14 2002
@@ -957,6 +957,8 @@
/*
- * The following 2 routines allow a task to sleep while socket memory is
- * low.
+ * Called when more output buffer space is available for this socket.
+ * We try not to wake our writers until they can make "significant"
+ * progress, otherwise we'll waste resources thrashing sock_sendmsg
+ * with a bunch of small requests.
*/
static void
@@ -972,6 +974,13 @@
/* Wait until we have enough socket memory */
- if (!sock_writeable(sk))
- return;
+ if (xprt->stream) {
+ /* from net/ipv4/tcp.c:tcp_write_space */
+ if (tcp_wspace(sk) < tcp_min_write_space(sk))
+ return;
+ } else {
+ /* from net/core/sock.c:sock_def_write_space */
+ if (!sock_writeable(sk))
+ return;
+ }
if (!test_and_clear_bit(SOCK_NOSPACE, &sock->flags))
--
corporate: <cel at netapp dot com>
personal: <chucklever at bigfoot dot com>