Hi Dave,
I've attached a small patch that (hopefully) will provide a little
performance boost for tcp. In tcp_do_sendmsg when the test is made to
see whether additional memory can be allocated for an skb, it's possible
that the extra amount would put the socket over its write limit. Then
the next skb to be allocated might have to wait for the memory limit to
fall.
But if the additional amount requested stays just below the socket's
memory limit, then the next skb won't have to wait, and performance
should be marginally better. I'm not sure whether this will happen often
enough to be worthwhile, but the extra code is minimal, so it's worth a
try.
The patch also makes a minor change to sock_kmalloc, to decrement the
memory limit if the allocation fails.
Regards,
Bill
--------------514C9EA8A7FCFDB0073F43F2
Content-Type: text/plain; charset=us-ascii; name="net_tcp103-patch"
Content-Transfer-Encoding: 7bit
Content-Disposition: inline; filename="net_tcp103-patch"
--- linux-2.1.103/net/ipv4/tcp.c.old Tue May 5 11:24:34 1998
+++ linux-2.1.103/net/ipv4/tcp.c Thu May 28 15:49:43 1998
@@ -804,22 +804,33 @@
if(copy > seglen)
copy = seglen;
- tmp = MAX_HEADER + sk->prot->max_header;
+ /*
+ * Calculate the minimum skb size to be allocated.
+ */
+ tmp = MAX_HEADER + sk->prot->max_header + copy;
queue_it = 0;
+ /*
+ * See if we can allocate more, but don't use up
+ * all of the space if it wouldn't otherwise.
+ */
if (copy < min(mss_now, tp->max_window >> 1) &&
!(flags & MSG_OOB)) {
- tmp += min(mss_now, tp->max_window);
+ int more = min(mss_now, tp->max_window) - copy;
+ int space = sock_wspace(sk);
+ if (tmp < space && more >= space - tmp) {
+printk("tcp_do_sendmsg: limiting extra %d to %d\n", more, space - tmp - 1);
+ more = space - tmp - 1;
+ }
+ tmp += more;
/* What is happening here is that we want to
- * tack on later members of the users iovec
+ * tack on later members of the user's iovec
* if possible into a single frame. When we
* leave this loop our caller checks to see if
* we can send queued frames onto the wire.
* See tcp_v[46]_sendmsg() for this.
*/
- queue_it = 1;
- } else {
- tmp += copy;
+ queue_it = (more > 0);
}
skb = sock_wmalloc(sk, tmp, 0, GFP_KERNEL);
--- linux-2.1.103/net/core/sock.c.old Sun May 17 12:19:41 1998
+++ linux-2.1.103/net/core/sock.c Thu May 28 13:51:33 1998
@@ -564,15 +564,18 @@
void *sock_kmalloc(struct sock *sk, int size, int priority)
{
- void *mem = NULL;
if (atomic_read(&sk->omem_alloc)+size < sysctl_optmem_max) {
+ void *mem;
/* First do the add, to avoid the race if kmalloc
* might sleep.
*/
atomic_add(size, &sk->omem_alloc);
mem = kmalloc(size, priority);
+ if (mem)
+ return mem;
+ atomic_sub(size, &sk->omem_alloc);
}
- return mem;
+ return NULL;
}
void sock_kfree_s(struct sock *sk, void *mem, int size)
--------------514C9EA8A7FCFDB0073F43F2--
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu