Change 54404 by rwatson@rwatson_paprika on 2004/06/08 06:46:06

	Experimentally modify rt_dispatch() to queue routing socket
	messages to a netisr rather than delivering inline: this
	prevents a recursive call back into the socket code from the
	routing code, which can result in lock order problems.  If
	a sockaddr is passed in, stash the family in an m_tag for
	later use before injection into the raw socket code.

Affected files ...

... //depot/user/rwatson/netperf/sys/net/netisr.h#6 edit
... //depot/user/rwatson/netperf/sys/net/rtsock.c#14 edit
... //depot/user/rwatson/netperf/sys/sys/mbuf.h#13 edit

Differences ...

==== //depot/user/rwatson/netperf/sys/net/netisr.h#6 (text+ko) ====

@@ -50,6 +50,7 @@
  */
 #define	NETISR_POLL	0		/* polling callback, must be first */
 #define	NETISR_IP	2		/* same as AF_INET */
+#define	NETISR_ROUTE	14		/* routing socket */
 #define	NETISR_AARP	15		/* Appletalk ARP */
 #define	NETISR_ATALK2	16		/* Appletalk phase 2 */
 #define	NETISR_ATALK1	17		/* Appletalk phase 1 */

==== //depot/user/rwatson/netperf/sys/net/rtsock.c#14 (text+ko) ====

@@ -45,6 +45,7 @@
 #include <sys/systm.h>
 
 #include <net/if.h>
+#include <net/netisr.h>
 #include <net/raw_cb.h>
 #include <net/route.h>
 
@@ -79,6 +80,8 @@
 #define	RTSOCK_UNLOCK()	mtx_unlock(&rtsock_mtx)
 #define	RTSOCK_LOCK_ASSERT()	mtx_assert(&rtsock_mtx, MA_OWNED)
 
+static struct	ifqueue rtsintrq;
+
 struct walkarg {
 	int	w_tmemsize;
 	int	w_op, w_arg;
@@ -86,6 +89,7 @@
 	struct sysctl_req *w_req;
 };
 
+static void	rts_input(struct mbuf *m);
 static struct mbuf *rt_msg1(int type, struct rt_addrinfo *rtinfo);
 static int	rt_msg2(int type, struct rt_addrinfo *rtinfo,
 			caddr_t cp, struct walkarg *w);
@@ -101,6 +105,35 @@
 			struct rt_metrics *out);
 static void	rt_dispatch(struct mbuf *, const struct sockaddr *);
 
+static void
+rts_init(void)
+{
+
+	rtsintrq.ifq_maxlen = IFQ_MAXLEN;
+	mtx_init(&rtsintrq.ifq_mtx, "rts_inq", NULL, MTX_DEF);
+	netisr_register(NETISR_ROUTE, rts_input, &rtsintrq, NETISR_MPSAFE);
+}
+SYSINIT(rtsock, SI_SUB_PROTO_DOMAIN, SI_ORDER_THIRD, rts_init, 0)
+
+static void
+rts_input(struct mbuf *m)
+{
+	struct sockproto route_proto;
+	unsigned short *family;
+	struct m_tag *tag;
+
+	route_proto.sp_family = PF_ROUTE;
+	tag = m_tag_find(m, PACKET_TAG_RTSOCKFAM, NULL);
+	if (tag != NULL) {
+		family = (unsigned short *)(tag + 1);
+		route_proto.sp_protocol = *family;
+		m_tag_delete(m, tag);
+	} else
+		route_proto.sp_protocol = 0;
+
+	raw_input(m, &route_proto, &route_src, &route_dst);
+}
+
 /*
  * It really doesn't make any sense at all for this code to share much
  * with raw_usrreq.c, since its functionality is so restricted.  XXX
@@ -927,11 +960,26 @@
 static void
 rt_dispatch(struct mbuf *m, const struct sockaddr *sa)
 {
-	struct sockproto route_proto;
+	unsigned short *family;
+	struct m_tag *tag;
 
-	route_proto.sp_family = PF_ROUTE;
-	route_proto.sp_protocol = sa ?  sa->sa_family : 0;
-	raw_input(m, &route_proto, &route_src, &route_dst);
+	/*
+	 * Preserve the family from the sockaddr, if any, in an m_tag for
+	 * use when injecting the mbuf into the routing socket buffer from
+	 * the netisr.
+	 */
+	if (sa != NULL) {
+		tag = m_tag_get(PACKET_TAG_RTSOCKFAM, sizeof(unsigned short),
+		    M_NOWAIT);
+		if (tag == NULL) {
+			m_freem(m);
+			return;
+		}
+		family = (unsigned short *)(tag + 1);
+		*family = sa ? sa->sa_family : 0;
+		m_tag_prepend(m, tag);
+	}
+	netisr_queue(NETISR_ROUTE, m);
 }
 
 /*

==== //depot/user/rwatson/netperf/sys/sys/mbuf.h#13 (text+ko) ====

@@ -631,6 +631,7 @@
 #define	PACKET_TAG_PF_FRAGCACHE			22 /* PF fragment cached */
 #define	PACKET_TAG_PF_QID			23 /* PF ALTQ queue id */
 #define	PACKET_TAG_PF_TAG			24 /* PF tagged */
+#define	PACKET_TAG_RTSOCKFAM			25 /* rtsock sa family */
 
 /* Packet tag routines. */
 struct	m_tag	*m_tag_alloc(u_int32_t, int, int, int);