Increase the default buffer size for AF_UNIX from 8192 to 32768.

Using 8k for socketpairs was always on the low end side. Also this
avoid a fatal error in sshd that can be triggered when the network
stack is pushed hard enough to consume most of the allowed memory.
By increasing the default buffer size a bit the error in sshd is
avoided which is good enough for now.

Long term a better solution for sonewconn() and especially sbchecklowmem()
needs to be found. m_pool_used() returns not the right information for
them.

OK deraadt@ otto@

diff --git a/sys/kern/uipc_usrreq.c b/sys/kern/uipc_usrreq.c
index 1b5117b..33e2731 100644 (file)
--- a/sys/kern/uipc_usrreq.c
+++ b/sys/kern/uipc_usrreq.c
@@ -1,4 +1,4 @@
-/*     $OpenBSD: uipc_usrreq.c,v 1.209 2024/08/06 20:13:58 mvs Exp $   */
+/*     $OpenBSD: uipc_usrreq.c,v 1.210 2024/09/22 08:40:37 claudio Exp $       */
 /*     $NetBSD: uipc_usrreq.c,v 1.18 1996/02/09 19:00:50 christos Exp $        */
 
 /*
@@ -234,13 +234,13 @@ uipc_setaddr(const struct unpcb *unp, struct mbuf *nam)
  * and don't really want to reserve the sendspace.  Their recvspace should
  * be large enough for at least one max-size datagram plus address.
  */
-#define        PIPSIZ  8192
+#define        PIPSIZ  32768
 u_int  unpst_sendspace = PIPSIZ;       /* [a] */
 u_int  unpst_recvspace = PIPSIZ;       /* [a] */
 u_int  unpsq_sendspace = PIPSIZ;       /* [a] */
 u_int  unpsq_recvspace = PIPSIZ;       /* [a] */
-u_int  unpdg_sendspace = 2*1024;       /* [a] really max datagram size */
-u_int  unpdg_recvspace = 16*1024;      /* [a] */
+u_int  unpdg_sendspace = 8192;         /* [a] really max datagram size */
+u_int  unpdg_recvspace = PIPSIZ;       /* [a] */
 
 const struct sysctl_bounded_args unpstctl_vars[] = {
        { UNPCTL_RECVSPACE, &unpst_recvspace, 0, SB_MAX },