this is only a work in progress, we can perfect afterwards, but it is time

to get some experience with these ideas.
add sbcheckreserve() api; called by accepting sockets.  if over 95% of
mbuf clusters are busy, consider this a resource starvation just like the
other reasons for accept failing.  also, if over 50% of mbuf clusters are
busy, shrink recv & send sockbuf reserves to "the minimum".

2 files changed, 33 insertions, 3 deletions

diff --git a/sys/kern/uipc_socket2.c b/sys/kern/uipc_socket2.c
index 0fa88f6c008..eb831138219 100644
--- a/sys/kern/uipc_socket2.c
+++ b/sys/kern/uipc_socket2.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: uipc_socket2.c,v 1.36 2004/04/01 23:56:05 tedu Exp $	*/
+/*	$OpenBSD: uipc_socket2.c,v 1.37 2004/04/19 22:38:39 deraadt Exp $	*/
 /*	$NetBSD: uipc_socket2.c,v 1.11 1996/02/04 02:17:55 christos Exp $	*/
 
 /*
@@ -154,9 +154,13 @@ sonewconn(struct socket *head, int connstatus)
 {
 	struct socket *so;
 	int soqueue = connstatus ? 1 : 0;
+	extern u_long unpst_sendspace, unpst_recvspace;
+	u_long snd_sb_hiwat, rcv_sb_hiwat;
 
 	splassert(IPL_SOFTNET);
 
+	if (mclpool.pr_nout > mclpool.pr_hardlimit * 95 / 100)
+		return ((struct socket *)0);
 	if (head->so_qlen + head->so_q0len > head->so_qlimit * 3)
 		return ((struct socket *)0);
 	so = pool_get(&socket_pool, PR_NOWAIT);
@@ -176,7 +180,19 @@ sonewconn(struct socket *head, int connstatus)
 	so->so_rgid = head->so_rgid;
 	so->so_siguid = head->so_siguid;
 	so->so_sigeuid = head->so_sigeuid;
-	(void) soreserve(so, head->so_snd.sb_hiwat, head->so_rcv.sb_hiwat);
+
+	/*
+	 * If we are tight on mbuf clusters, create the new socket
+	 * with the minimum.  Sorry, you lose.
+	 */
+	snd_sb_hiwat = head->so_snd.sb_hiwat;
+	if (sbcheckreserve(snd_sb_hiwat, unpst_sendspace))
+		snd_sb_hiwat = unpst_sendspace;		/* and udp? */
+	rcv_sb_hiwat = head->so_rcv.sb_hiwat;
+	if (sbcheckreserve(rcv_sb_hiwat, unpst_recvspace))
+		rcv_sb_hiwat = unpst_recvspace;		/* and udp? */
+
+	(void) soreserve(so, snd_sb_hiwat, rcv_sb_hiwat);
 	soqinsque(head, so, soqueue);
 	if ((*so->so_proto->pr_usrreq)(so, PRU_ATTACH,
 	    (struct mbuf *)0, (struct mbuf *)0, (struct mbuf *)0)) {
@@ -395,6 +411,19 @@ sbreserve(sb, cc)
 }
 
 /*
+ * If over 50% of mbuf clusters in use, do not accept any
+ * greater than normal request.
+ */
+int
+sbcheckreserve(u_long cnt, u_long defcnt)
+{
+	if (cnt > defcnt &&
+	    mclpool.pr_nout> mclpool.pr_hardlimit / 2)
+		return (ENOBUFS);
+	return (0);
+}
+
+/*
  * Free mbufs held by a socket, and reserved mbuf space.
  */
 void
diff --git a/sys/sys/socketvar.h b/sys/sys/socketvar.h
index 7e3f71af05d..3178750c14e 100644
--- a/sys/sys/socketvar.h
+++ b/sys/sys/socketvar.h
@@ -1,4 +1,4 @@
-/*	$OpenBSD: socketvar.h,v 1.33 2004/04/12 23:58:10 tedu Exp $	*/
+/*	$OpenBSD: socketvar.h,v 1.34 2004/04/19 22:38:37 deraadt Exp $	*/
 /*	$NetBSD: socketvar.h,v 1.18 1996/02/09 18:25:38 christos Exp $	*/
 
 /*-
@@ -259,6 +259,7 @@ void	sbdroprecord(struct sockbuf *sb);
 void	sbflush(struct sockbuf *sb);
 void	sbinsertoob(struct sockbuf *sb, struct mbuf *m0);
 void	sbrelease(struct sockbuf *sb);
+int	sbcheckreserve(u_long cnt, u_long defcnt);
 int	sbreserve(struct sockbuf *sb, u_long cc);
 int	sbwait(struct sockbuf *sb);
 int	sb_lock(struct sockbuf *sb);