Switch mrt dumping to fd passing. This gives some speed up when extensive

dumping is done. Acctually mrt dumps were broken because of the fd passing.
The nice side effect is a much cleaner code, especially in the parent process.
OK henning@

9 files changed, 435 insertions, 491 deletions

diff --git a/usr.sbin/bgpd/bgpd.c b/usr.sbin/bgpd/bgpd.c
index b66996e6ce7..01ff6cd1a42 100644
--- a/usr.sbin/bgpd/bgpd.c
+++ b/usr.sbin/bgpd/bgpd.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: bgpd.c,v 1.97 2004/06/20 18:35:11 henning Exp $ */
+/*	$OpenBSD: bgpd.c,v 1.98 2004/07/03 17:19:59 claudio Exp $ */
 
 /*
  * Copyright (c) 2003, 2004 Henning Brauer <henning@openbsd.org>
@@ -83,11 +83,11 @@ usage(void)
 	exit(1);
 }
 
-#define POLL_MAX		8
 #define PFD_PIPE_SESSION	0
 #define PFD_PIPE_ROUTE		1
 #define PFD_SOCK_ROUTE		2
-#define PFD_MRT_START		3
+#define POLL_MAX		3
+#define MAX_TIMEOUT		3600
 
 int
 main(int argc, char *argv[])
@@ -99,13 +99,13 @@ main(int argc, char *argv[])
 	struct filter_head	*rules_l;
 	struct network		*net;
 	struct filter_rule	*r;
-	struct mrt		*(mrt[POLL_MAX]), *m;
+	struct mrt		*m;
 	struct listen_addr	*la;
 	struct pollfd		 pfd[POLL_MAX];
 	pid_t			 io_pid = 0, rde_pid = 0, pid;
 	char			*conffile;
 	int			 debug = 0;
-	int			 ch, i, j, nfds, timeout;
+	int			 ch, nfds, timeout;
 	int			 pipe_m2s[2];
 	int			 pipe_m2r[2];
 	int			 pipe_s2r[2];
@@ -236,6 +236,8 @@ main(int argc, char *argv[])
 		free(la);
 	}
 
+	mrt_reconfigure(&mrt_l);
+
 	while (quit == 0) {
 		pfd[PFD_PIPE_SESSION].fd = ibuf_se.fd;
 		pfd[PFD_PIPE_SESSION].events = POLLIN;
@@ -247,10 +249,12 @@ main(int argc, char *argv[])
 			pfd[PFD_PIPE_ROUTE].events |= POLLOUT;
 		pfd[PFD_SOCK_ROUTE].fd = rfd;
 		pfd[PFD_SOCK_ROUTE].events = POLLIN;
-		i = PFD_MRT_START;
-		i = mrt_select(&mrt_l, pfd, mrt, i, POLL_MAX, &timeout);
 
-		if ((nfds = poll(pfd, i, INFTIM)) == -1)
+		timeout = mrt_timeout(&mrt_l);
+		if (timeout > MAX_TIMEOUT)
+			timeout = MAX_TIMEOUT;
+		
+		if ((nfds = poll(pfd, POLL_MAX, timeout * 1000)) == -1)
 			if (errno != EINTR) {
 				log_warn("poll error");
 				quit = 1;
@@ -288,14 +292,6 @@ main(int argc, char *argv[])
 				quit = 1;
 		}
 
-		for (j = PFD_MRT_START; j < i && nfds > 0 ; j++) {
-			if (pfd[j].revents & POLLOUT) {
-				if (mrt_write(mrt[j]) < 0) {
-					log_warn("mrt write error");
-				}
-			}
-		}
-
 		if (reconfig) {
 			log_info("rereading config");
 			reconfigure(conffile, &conf, &mrt_l, &peer_l, rules_l);
@@ -329,7 +325,7 @@ main(int argc, char *argv[])
 		free(p);
 	}
 	while ((m = LIST_FIRST(&mrt_l)) != NULL) {
-		LIST_REMOVE(m, list);
+		LIST_REMOVE(m, entry);
 		free(m);
 	}
 
@@ -424,6 +420,8 @@ reconfigure(char *conffile, struct bgpd_config *conf, struct mrt_head *mrt_l,
 	    imsg_compose(&ibuf_rde, IMSG_RECONF_DONE, 0, NULL, 0) == -1)
 		return (-1);
 
+	/* mrt changes can be sent out of bound */
+	mrt_reconfigure(mrt_l);
 	return (0);
 }
 
@@ -431,6 +429,7 @@ int
 dispatch_imsg(struct imsgbuf *ibuf, int idx, struct mrt_head *mrt_l)
 {
 	struct imsg		 imsg;
+	struct mrt		 mrt;
 	int			 n;
 
 	if ((n = imsg_read(ibuf)) == -1)
@@ -449,10 +448,14 @@ dispatch_imsg(struct imsgbuf *ibuf, int idx, struct mrt_head *mrt_l)
 			break;
 
 		switch (imsg.hdr.type) {
-		case IMSG_MRT_MSG:
-		case IMSG_MRT_END:
-			if (mrt_queue(mrt_l, &imsg) == -1)
-				log_warnx("mrt_queue failed.");
+		case IMSG_MRT_CLOSE:
+			if (imsg.hdr.len != IMSG_HEADER_SIZE +
+			    sizeof(struct mrt)) {
+				log_warnx("wrong imsg len");
+				break;
+			}
+			memcpy(&mrt, imsg.data, sizeof(struct mrt));
+			mrt_close(mrt_get(mrt_l, &mrt));
 			break;
 		case IMSG_KROUTE_CHANGE:
 			if (idx != PFD_PIPE_ROUTE)
diff --git a/usr.sbin/bgpd/bgpd.h b/usr.sbin/bgpd/bgpd.h
index 0823e88199d..8b7f35a3a02 100644
--- a/usr.sbin/bgpd/bgpd.h
+++ b/usr.sbin/bgpd/bgpd.h
@@ -1,4 +1,4 @@
-/*	$OpenBSD: bgpd.h,v 1.132 2004/06/25 20:08:46 henning Exp $ */
+/*	$OpenBSD: bgpd.h,v 1.133 2004/07/03 17:19:59 claudio Exp $ */
 
 /*
  * Copyright (c) 2003, 2004 Henning Brauer <henning@openbsd.org>
@@ -264,9 +264,9 @@ enum imsg_type {
 	IMSG_UPDATE_ERR,
 	IMSG_SESSION_UP,
 	IMSG_SESSION_DOWN,
-	IMSG_MRT_REQ,
-	IMSG_MRT_MSG,
-	IMSG_MRT_END,
+	IMSG_MRT_OPEN,
+	IMSG_MRT_REOPEN,
+	IMSG_MRT_CLOSE,
 	IMSG_KROUTE_CHANGE,
 	IMSG_KROUTE_DELETE,
 	IMSG_NEXTHOP_ADD,
@@ -551,6 +551,7 @@ struct buf	*buf_open(ssize_t);
 int		 buf_add(struct buf *, void *, ssize_t);
 void		*buf_reserve(struct buf *, ssize_t);
 int		 buf_close(struct msgbuf *, struct buf *);
+int		 buf_write(int, struct buf *);
 void		 buf_free(struct buf *);
 void		 msgbuf_init(struct msgbuf *);
 void		 msgbuf_clear(struct msgbuf *);
diff --git a/usr.sbin/bgpd/buffer.c b/usr.sbin/bgpd/buffer.c
index a47ced7b15d..6887b793a59 100644
--- a/usr.sbin/bgpd/buffer.c
+++ b/usr.sbin/bgpd/buffer.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: buffer.c,v 1.21 2004/06/20 18:35:12 henning Exp $ */
+/*	$OpenBSD: buffer.c,v 1.22 2004/07/03 17:19:59 claudio Exp $ */
 
 /*
  * Copyright (c) 2003, 2004 Henning Brauer <henning@openbsd.org>
@@ -29,7 +29,6 @@
 
 #include "bgpd.h"
 
-int	buf_write(int, struct buf *);
 void	buf_enqueue(struct msgbuf *, struct buf *);
 void	buf_dequeue(struct msgbuf *, struct buf *);
 
diff --git a/usr.sbin/bgpd/mrt.c b/usr.sbin/bgpd/mrt.c
index 121800b6eef..e68fccc95d8 100644
--- a/usr.sbin/bgpd/mrt.c
+++ b/usr.sbin/bgpd/mrt.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: mrt.c,v 1.32 2004/06/22 20:28:58 claudio Exp $ */
+/*	$OpenBSD: mrt.c,v 1.33 2004/07/03 17:19:59 claudio Exp $ */
 
 /*
  * Copyright (c) 2003, 2004 Claudio Jeker <claudio@openbsd.org>
@@ -34,17 +34,17 @@
 
 static u_int16_t	mrt_attr_length(struct attr_flags *);
 static int		mrt_attr_dump(void *, u_int16_t, struct attr_flags *);
-static int		mrt_dump_entry(struct mrt_config *, struct prefix *,
+static int		mrt_dump_entry(struct mrt *, struct prefix *,
 			    u_int16_t, struct peer_config *);
 static int		mrt_dump_header(struct buf *, u_int16_t, u_int16_t,
 			    u_int32_t);
-static int		mrt_open(struct mrt *);
+static int		mrt_open(struct mrt *, time_t);
 
 #define DUMP_BYTE(x, b)							\
 	do {								\
 		u_char		t = (b);				\
-		if (imsg_add((x), &t, sizeof(t)) == -1) {		\
-			log_warnx("mrt_dump1: imsg_add error");		\
+		if (buf_add((x), &t, sizeof(t)) == -1) {		\
+			log_warnx("mrt_dump1: buf_add error");		\
 			return (-1);					\
 		}							\
 	} while (0)
@@ -53,8 +53,8 @@ static int		mrt_open(struct mrt *);
 	do {								\
 		u_int16_t	t;					\
 		t = htons((s));						\
-		if (imsg_add((x), &t, sizeof(t)) == -1) {		\
-			log_warnx("mrt_dump2: imsg_add error");		\
+		if (buf_add((x), &t, sizeof(t)) == -1) {		\
+			log_warnx("mrt_dump2: buf_add error");		\
 			return (-1);					\
 		}							\
 	} while (0)
@@ -63,8 +63,8 @@ static int		mrt_open(struct mrt *);
 	do {								\
 		u_int32_t	t;					\
 		t = htonl((l));						\
-		if (imsg_add((x), &t, sizeof(t)) == -1) {		\
-			log_warnx("mrt_dump3: imsg_add error");		\
+		if (buf_add((x), &t, sizeof(t)) == -1) {		\
+			log_warnx("mrt_dump3: buf_add error");		\
 			return (-1);					\
 		}							\
 	} while (0)
@@ -72,14 +72,14 @@ static int		mrt_open(struct mrt *);
 #define DUMP_NLONG(x, l)						\
 	do {								\
 		u_int32_t	t = (l);				\
-		if (imsg_add((x), &t, sizeof(t)) == -1) {		\
-			log_warnx("mrt_dump4: imsg_add error");		\
+		if (buf_add((x), &t, sizeof(t)) == -1) {		\
+			log_warnx("mrt_dump4: buf_add error");		\
 			return (-1);					\
 		}							\
 	} while (0)
 
 int
-mrt_dump_bgp_msg(struct mrt_config *mrt, void *pkg, u_int16_t pkglen,
+mrt_dump_bgp_msg(struct mrt *mrt, void *pkg, u_int16_t pkglen,
     struct peer_config *peer, struct bgpd_config *bgp)
 {
 	struct buf	*buf;
@@ -87,15 +87,14 @@ mrt_dump_bgp_msg(struct mrt_config *mrt, void *pkg, u_int16_t pkglen,
 
 	len = pkglen + MRT_BGP4MP_HEADER_SIZE;
 
-	if ((buf = imsg_create(mrt->ibuf, IMSG_MRT_MSG, mrt->id,
-	    len + MRT_HEADER_SIZE)) == NULL) {
-		log_warnx("mrt_dump_bgp_msg: imsg_open error");
+	if ((buf = buf_open(len + MRT_HEADER_SIZE)) == NULL) {
+		log_warnx("mrt_dump_bgp_msg: buf_open error");
 		return (-1);
 	}
 
 	if (mrt_dump_header(buf, MSG_PROTOCOL_BGP4MP, BGP4MP_MESSAGE,
 	    len) == -1) {
-		log_warnx("mrt_dump_bgp_msg: imsg_add error");
+		log_warnx("mrt_dump_bgp_msg: buf_add error");
 		return (-1);
 	}
 
@@ -106,21 +105,19 @@ mrt_dump_bgp_msg(struct mrt_config *mrt, void *pkg, u_int16_t pkglen,
 	DUMP_NLONG(buf, peer->local_addr.v4.s_addr);
 	DUMP_NLONG(buf, peer->remote_addr.v4.s_addr);
 
-	if (imsg_add(buf, pkg, pkglen) == -1) {
-		log_warnx("mrt_dump_bgp_msg: imsg_add error");
+	if (buf_add(buf, pkg, pkglen) == -1) {
+		log_warnx("mrt_dump_bgp_msg: buf_add error");
 		return (-1);
 	}
 
-	if ((imsg_close(mrt->ibuf, buf)) == -1) {
-		log_warnx("mrt_dump_bgp_msg: imsg_close error");
-		return (-1);
-	}
+	TAILQ_INSERT_TAIL(&mrt->bufs, buf, entry);
+	mrt->queued++;
 
 	return (len + MRT_HEADER_SIZE);
 }
 
 int
-mrt_dump_state(struct mrt_config *mrt, u_int16_t old_state, u_int16_t new_state,
+mrt_dump_state(struct mrt *mrt, u_int16_t old_state, u_int16_t new_state,
     struct peer_config *peer, struct bgpd_config *bgp)
 {
 	struct buf	*buf;
@@ -128,15 +125,14 @@ mrt_dump_state(struct mrt_config *mrt, u_int16_t old_state, u_int16_t new_state,
 
 	len = 4 + MRT_BGP4MP_HEADER_SIZE;
 
-	if ((buf = imsg_create(mrt->ibuf, IMSG_MRT_MSG, mrt->id,
-	    len + MRT_HEADER_SIZE)) == NULL) {
-		log_warnx("mrt_dump_bgp_state: imsg_open error");
+	if ((buf = buf_open(len + MRT_HEADER_SIZE)) == NULL) {
+		log_warnx("mrt_dump_bgp_state: buf_open error");
 		return (-1);
 	}
 
 	if (mrt_dump_header(buf, MSG_PROTOCOL_BGP4MP, BGP4MP_STATE_CHANGE,
 	    len) == -1) {
-		log_warnx("mrt_dump_bgp_state: imsg_add error");
+		log_warnx("mrt_dump_bgp_state: buf_add error");
 		return (-1);
 	}
 
@@ -150,10 +146,8 @@ mrt_dump_state(struct mrt_config *mrt, u_int16_t old_state, u_int16_t new_state,
 	DUMP_SHORT(buf, old_state);
 	DUMP_SHORT(buf, new_state);
 
-	if ((imsg_close(mrt->ibuf, buf)) == -1) {
-		log_warnx("mrt_dump_bgp_state: imsg_close error");
-		return (-1);
-	}
+	TAILQ_INSERT_TAIL(&mrt->bufs, buf, entry);
+	mrt->queued++;
 
 	return (len + MRT_HEADER_SIZE);
 }
@@ -232,7 +226,7 @@ mrt_attr_dump(void *p, u_int16_t len, struct attr_flags *a)
 }
 
 static int
-mrt_dump_entry(struct mrt_config *mrt, struct prefix *p, u_int16_t snum,
+mrt_dump_entry(struct mrt *mrt, struct prefix *p, u_int16_t snum,
     struct peer_config *peer)
 {
 	struct buf	*buf;
@@ -244,9 +238,8 @@ mrt_dump_entry(struct mrt_config *mrt, struct prefix *p, u_int16_t snum,
 	len = MRT_DUMP_HEADER_SIZE + attr_len;
 	pt_getaddr(p->prefix, &addr);
 
-	if ((buf = imsg_create(mrt->ibuf, IMSG_MRT_MSG, mrt->id,
-	    len + MRT_HEADER_SIZE)) == NULL) {
-		log_warnx("mrt_dump_entry: imsg_open error");
+	if ((buf = buf_open(len + MRT_HEADER_SIZE)) == NULL) {
+		log_warnx("mrt_dump_entry: buf_open error");
 		return (-1);
 	}
 
@@ -277,10 +270,8 @@ mrt_dump_entry(struct mrt_config *mrt, struct prefix *p, u_int16_t snum,
 		return (-1);
 	}
 
-	if ((imsg_close(mrt->ibuf, buf)) == -1) {
-		log_warnx("mrt_dump_bgp_state: imsg_close error");
-		return (-1);
-	}
+	TAILQ_INSERT_TAIL(&mrt->bufs, buf, entry);
+	mrt->queued++;
 
 	return (len + MRT_HEADER_SIZE);
 }
@@ -296,7 +287,7 @@ mrt_clear_seq(void)
 void
 mrt_dump_upcall(struct pt_entry *pt, void *ptr)
 {
-	struct mrt_config	*mrtbuf = ptr;
+	struct mrt		*mrtbuf = ptr;
 	struct prefix		*p;
 
 	/*
@@ -325,6 +316,39 @@ mrt_dump_header(struct buf *buf, u_int16_t type, u_int16_t subtype,
 	return (0);
 }
 
+int
+mrt_write(struct mrt *mrt)
+{
+	struct buf	*b;
+	int		 r = 0;
+
+	while ((b = TAILQ_FIRST(&mrt->bufs)) &&
+	    (r = buf_write(mrt->fd, b)) == 1) {
+		TAILQ_REMOVE(&mrt->bufs, b, entry);
+		mrt->queued--;
+		buf_free(b);
+	}
+	if (r == -1) {
+		log_warn("mrt dump write");
+		mrt_clean(mrt);
+		return (-1);
+	}
+	return (0);
+}
+
+void
+mrt_clean(struct mrt *mrt)
+{
+	struct buf	*b;
+
+	close(mrt->fd);
+	while ((b = TAILQ_FIRST(&mrt->bufs))) {
+		TAILQ_REMOVE(&mrt->bufs, b, entry);
+		buf_free(b);
+	}
+	mrt->queued = 0;
+}
+
 static struct imsgbuf	*mrt_imsgbuf[2];
 
 void
@@ -334,339 +358,147 @@ mrt_init(struct imsgbuf *rde, struct imsgbuf *se)
 	mrt_imsgbuf[1] = se;
 }
 
-static int
-mrt_open(struct mrt *mrt)
+int
+mrt_open(struct mrt *mrt, time_t now)
 {
-	time_t	now;
+	enum imsg_type	type;
+	int		i;
 
-	now = time(NULL);
-	if (strftime(mrt->file, sizeof(mrt->file), mrt->name,
-		    localtime(&now)) == 0) {
+	mrt_close(mrt);
+	if (strftime(MRT2MC(mrt)->file, sizeof(MRT2MC(mrt)->file),
+	    MRT2MC(mrt)->name, localtime(&now)) == 0) {
 		log_warnx("mrt_open: strftime conversion failed");
-		mrt->msgbuf.fd = -1;
-		return (0);
+		mrt->fd = -1;
+		return (-1);
 	}
 
-	mrt->msgbuf.fd = open(mrt->file,
+	mrt->fd = open(MRT2MC(mrt)->file,
 	    O_WRONLY|O_NONBLOCK|O_CREAT|O_TRUNC, 0644);
-	if (mrt->msgbuf.fd == -1) {
-		log_warnx("mrt_open %s: %s",
-		    mrt->file, strerror(errno));
-		return (0);
+	if (mrt->fd == -1) {
+		log_warn("mrt_open %s", MRT2MC(mrt)->file);
+		return (1);
 	}
-	return (1);
-}
 
-static int
-mrt_close(struct mrt *mrt)
-{
-	/*
-	 * close the mrt filedescriptor but first ensure that the last
-	 * mrt message was written correctly. If not mrt_write needs to do
-	 * that the next time called.
-	 * To ensure this we need to fiddle around with internal msgbuf stuff.
-	 */
-	if (msgbuf_unbounded(&mrt->msgbuf))
-		return (0);
+	if (MRT2MC(mrt)->state == MRT_STATE_OPEN)
+		type = IMSG_MRT_OPEN;
+	else
+		type = IMSG_MRT_REOPEN;
 
-	if (mrt->msgbuf.fd != -1) {
-		close(mrt->msgbuf.fd);
-		mrt->msgbuf.fd = -1;
-	}
+	i = mrt->type == MRT_TABLE_DUMP ? 0 : 1;
+
+	if (imsg_compose_fdpass(mrt_imsgbuf[i], type, mrt->fd,
+	    mrt, sizeof(struct mrt)) == -1)
+		log_warn("mrt_close");
 
 	return (1);
 }
 
 void
-mrt_abort(struct mrt *mrt)
+mrt_close(struct mrt *mrt)
 {
+	if (mrt == NULL)
+		return;
 	/*
-	 * something failed horribly. Stop all dumping and go back to start
-	 * position. Retry after MRT_MIN_RETRY or ReopenTimerInterval. Which-
-	 * ever is bigger.
+	 * this function is normaly called twice. First because of a imsg 
+	 * form the child to inform the parent to close the fd. The second time
+	 * it is called after reconfigure when the mrt file gets removed.
+	 * In that case the parent must inform the child to close and remove
+	 * this mrt dump descriptor.
 	 */
-	msgbuf_clear(&mrt->msgbuf);
-	mrt_close(mrt);
-	mrt->state = MRT_STATE_STOPPED;
-
-	if (MRT_MIN_RETRY > mrt->ReopenTimerInterval)
-		mrt->ReopenTimer = MRT_MIN_RETRY + time(NULL);
-	else
-		mrt->ReopenTimer = mrt->ReopenTimerInterval + time(NULL);
+	if (MRT2MC(mrt)->state == MRT_STATE_REMOVE)
+		if (imsg_compose(
+		    mrt_imsgbuf[mrt->type == MRT_TABLE_DUMP ? 0 : 1],
+		    IMSG_MRT_CLOSE, 0, mrt, sizeof(struct mrt)) == -1)
+			log_warn("mrt_close");
+
+	if (mrt->fd == -1)
+		return;
+	close(mrt->fd);
+	mrt->fd = -1;
 }
 
 int
-mrt_queue(struct mrt_head *mrtc, struct imsg *imsg)
+mrt_timeout(struct mrt_head *mrt)
 {
-	struct buf	*wbuf;
 	struct mrt	*m;
-	ssize_t		 len;
-	int		 n;
-
-	if (imsg->hdr.type != IMSG_MRT_MSG && imsg->hdr.type != IMSG_MRT_END)
-		return (-1);
-
-	LIST_FOREACH(m, mrtc, list) {
-		if (m->conf.id != imsg->hdr.peerid)
-			continue;
-		if (m->state != MRT_STATE_RUNNING &&
-		    m->state != MRT_STATE_REOPEN)
-			return (0);
-
-		if (imsg->hdr.type == IMSG_MRT_END) {
-			m->state = MRT_STATE_CLOSE;
-			return (0);
-		}
-
-		len = imsg->hdr.len - IMSG_HEADER_SIZE;
-		wbuf = buf_open(len);
-		if (wbuf == NULL)
-			return (-1);
-		if (buf_add(wbuf, imsg->data, len) == -1) {
-			buf_free(wbuf);
-			return (-1);
-		}
-		if ((n = buf_close(&m->msgbuf, wbuf)) < 0) {
-			buf_free(wbuf);
-			return (-1);
-		}
-		return (n);
-	}
-	return (0);
-}
-
-int
-mrt_write(struct mrt *mrt)
-{
-	int	r;
+	time_t		 now;
+	int		 timeout = MRT_MAX_TIMEOUT;
 
-	if (mrt->state == MRT_STATE_REOPEN ||
-	    mrt->state == MRT_STATE_REMOVE)
-		r = msgbuf_writebound(&mrt->msgbuf);
-	else
-		r = msgbuf_write(&mrt->msgbuf);
-
-	switch (r) {
-	case 1:
-		/* only msgbuf_writebound returns 1 */
-		break;
-	case 0:
-		if (mrt->state == MRT_STATE_CLOSE && mrt->msgbuf.queued == 0) {
-			if (mrt_close(mrt) != 1) {
-				log_warnx("mrt_write: mrt_close failed");
-				mrt_abort(mrt);
-				return (0);
+	now = time(NULL);
+	LIST_FOREACH(m, mrt, entry) {
+		if (MRT2MC(m)->state == MRT_STATE_RUNNING &&
+		    MRT2MC(m)->ReopenTimerInterval != 0) {
+			if (MRT2MC(m)->ReopenTimer <= now) {
+				mrt_open(m, now);
+				MRT2MC(m)->ReopenTimer =
+				    now + MRT2MC(m)->ReopenTimerInterval;
 			}
-			mrt->state = MRT_STATE_STOPPED;
-		}
-		return (0);
-	case -1:
-		log_warnx("mrt_write: msgbuf_write: %s",
-		    strerror(errno));
-		mrt_abort(mrt);
-		return (0);
-	case -2:
-		log_warnx("mrt_write: msgbuf_write: %s",
-		    "connection closed");
-		mrt_abort(mrt);
-		return (0);
-	default:
-		fatalx("mrt_write: unexpected retval from msgbuf_write");
-	}
-
-	if (mrt_close(mrt) != 1) {
-		log_warnx("mrt_write: mrt_close failed");
-		mrt_abort(mrt);
-		return (0);
-	}
-
-	switch (mrt->state) {
-	case MRT_STATE_REMOVE:
-		/*
-		 * Remove request: free all left buffers and
-		 * remove the descriptor.
-		 */
-		msgbuf_clear(&mrt->msgbuf);
-		LIST_REMOVE(mrt, list);
-		free(mrt);
-		return (0);
-	case MRT_STATE_REOPEN:
-		if (mrt_open(mrt) == 0) {
-			mrt_abort(mrt);
-			return (0);
-		} else {
-			if (mrt->ReopenTimerInterval != 0)
-				mrt->ReopenTimer = time(NULL) +
-				    mrt->ReopenTimerInterval;
-			mrt->state = MRT_STATE_RUNNING;
+			if (MRT2MC(m)->ReopenTimer - now < timeout)
+				timeout = MRT2MC(m)->ReopenTimer - now;
 		}
-		break;
-	default:
-		break;
 	}
-	return (1);
+	return (timeout > 0 ? timeout : 0);
 }
 
-int
-mrt_select(struct mrt_head *mc, struct pollfd *pfd, struct mrt **mrt,
-    int start, int size, int *timeout)
+void
+mrt_reconfigure(struct mrt_head *mrt)
 {
 	struct mrt	*m, *xm;
 	time_t		 now;
-	int		 t;
 
 	now = time(NULL);
-	for (m = LIST_FIRST(mc); m != NULL; m = xm) {
-		xm = LIST_NEXT(m, list);
-		if (m->state == MRT_STATE_TOREMOVE) {
-			imsg_compose(m->ibuf, IMSG_MRT_END, 0,
-			    &m->conf, sizeof(m->conf));
-			if (mrt_close(m) == 0) {
-				m->state = MRT_STATE_REMOVE;
-				m->ReopenTimer = 0;
-			} else {
-				msgbuf_clear(&m->msgbuf);
-				LIST_REMOVE(m, list);
-				free(m);
+	for (m = LIST_FIRST(mrt); m != NULL; m = xm) {
+		xm = LIST_NEXT(m, entry);
+		if (MRT2MC(m)->state == MRT_STATE_OPEN ||
+		    MRT2MC(m)->state == MRT_STATE_REOPEN) {
+			if (mrt_open(m, now) == -1)
 				continue;
-			}
-		}
-		if (m->state == MRT_STATE_OPEN) {
-			switch (m->conf.type) {
-			case MRT_TABLE_DUMP:
-				m->ibuf = mrt_imsgbuf[0];
-				break;
-			case MRT_ALL_IN:
-			case MRT_ALL_OUT:
-			case MRT_UPDATE_IN:
-			case MRT_UPDATE_OUT:
-				m->ibuf = mrt_imsgbuf[1];
-				break;
-			default:
-				continue;
-			}
-			if (mrt_open(m) == 0) {
-				mrt_abort(m);
-				t = m->ReopenTimer - now;
-				if (*timeout > t)
-					*timeout = t;
-				continue;
-			}
-			if (m->ReopenTimerInterval != 0)
-				m->ReopenTimer = now + m->ReopenTimerInterval;
-			m->state = MRT_STATE_RUNNING;
-			imsg_compose(m->ibuf, IMSG_MRT_REQ, 0,
-			    &m->conf, sizeof(m->conf));
-		}
-		if (m->state == MRT_STATE_REOPEN) {
-			if (mrt_close(m) == 0) {
-				m->state = MRT_STATE_REOPEN;
-				continue;
-			}
-			if (mrt_open(m) == 0) {
-				mrt_abort(m);
-				t = m->ReopenTimer - now;
-				if (*timeout > t)
-					*timeout = t;
-				continue;
-			}
-			if (m->ReopenTimerInterval != 0)
-				m->ReopenTimer = now + m->ReopenTimerInterval;
-			m->state = MRT_STATE_RUNNING;
-		}
-		if (m->ReopenTimer != 0) {
-			t = m->ReopenTimer - now;
-			if (t <= 0 && (m->state == MRT_STATE_RUNNING ||
-			    m->state == MRT_STATE_STOPPED)) {
-				if (m->state == MRT_STATE_RUNNING) {
-					/* reopen file */
-					if (mrt_close(m) == 0) {
-						m->state = MRT_STATE_REOPEN;
-						continue;
-					}
-				}
-				if (mrt_open(m) == 0) {
-					mrt_abort(m);
-					t = m->ReopenTimer - now;
-					if (*timeout > t)
-						*timeout = t;
-					continue;
-				}
-				if (m->conf.type == MRT_TABLE_DUMP &&
-				    m->state == MRT_STATE_STOPPED) {
-					imsg_compose(mrt_imsgbuf[0],
-					    IMSG_MRT_REQ, 0,
-					    &m->conf, sizeof(m->conf));
-				}
-
-				m->state = MRT_STATE_RUNNING;
-				if (m->ReopenTimerInterval != 0) {
-					m->ReopenTimer = now +
-					    m->ReopenTimerInterval;
-					if (*timeout > m->ReopenTimerInterval)
-						*timeout = t;
-				}
-			}
+			if (MRT2MC(m)->ReopenTimerInterval != 0)
+				MRT2MC(m)->ReopenTimer =
+				    now + MRT2MC(m)->ReopenTimerInterval;
+			MRT2MC(m)->state = MRT_STATE_RUNNING;
 		}
-		if (m->msgbuf.queued > 0) {
-			if (m->msgbuf.fd == -1 ||
-			    m->state == MRT_STATE_STOPPED) {
-				log_warnx("mrt_select: orphaned buffer");
-				mrt_abort(m);
-				continue;
-			}
-			if (start < size) {
-				pfd[start].fd = m->msgbuf.fd;
-				pfd[start].events = POLLOUT;
-				mrt[start++] = m;
-			}
+		if (MRT2MC(m)->state == MRT_STATE_REMOVE) {
+			mrt_close(m);
+			LIST_REMOVE(m, entry);
+			free(m);
+			continue;
 		}
 	}
-	return (start);
 }
 
-int
+void
 mrt_handler(struct mrt_head *mrt)
 {
 	struct mrt	*m;
 	time_t		 now;
 
 	now = time(NULL);
-	LIST_FOREACH(m, mrt, list) {
-		if (m->state == MRT_STATE_RUNNING)
-			m->state = MRT_STATE_REOPEN;
-		if (m->conf.type == MRT_TABLE_DUMP) {
-			if (m->state == MRT_STATE_STOPPED) {
-				if (mrt_open(m) == 0) {
-					mrt_abort(m);
-					break;
-				}
-				imsg_compose(mrt_imsgbuf[0], IMSG_MRT_REQ, 0,
-				    &m->conf, sizeof(m->conf));
-				m->state = MRT_STATE_RUNNING;
-			}
+	LIST_FOREACH(m, mrt, entry) {
+		if (MRT2MC(m)->state == MRT_STATE_RUNNING &&
+		    (MRT2MC(m)->ReopenTimerInterval != 0 ||
+		     m->type == MRT_TABLE_DUMP)) {
+			if (mrt_open(m, now) == -1)
+				continue;
+			MRT2MC(m)->ReopenTimer =
+			    now + MRT2MC(m)->ReopenTimerInterval;
 		}
-		if (m->ReopenTimerInterval != 0)
-			m->ReopenTimer = now + m->ReopenTimerInterval;
 	}
-	return (0);
 }
 
-static u_int32_t	 max_id = 1;
-
-static struct mrt *
-getconf(struct mrt_head *c, struct mrt *m)
+struct mrt *
+mrt_get(struct mrt_head *c, struct mrt *m)
 {
 	struct mrt	*t;
 
-	LIST_FOREACH(t, c, list) {
-		if (t->conf.type != m->conf.type)
+	LIST_FOREACH(t, c, entry) {
+		if (t->type != m->type)
 			continue;
-		if (t->conf.type == MRT_TABLE_DUMP)
-			return t;
-		if (t->conf.peer_id == m->conf.peer_id &&
-		    t->conf.group_id == m->conf.group_id)
-			return t;
+		if (t->type == MRT_TABLE_DUMP)
+			return (t);
+		if (t->peer_id == m->peer_id &&
+		    t->group_id == m->group_id)
+			return (t);
 	}
 	return (NULL);
 }
@@ -676,33 +508,35 @@ mrt_mergeconfig(struct mrt_head *xconf, struct mrt_head *nconf)
 {
 	struct mrt	*m, *xm;
 
-	LIST_FOREACH(m, nconf, list)
-		if ((xm = getconf(xconf, m)) == NULL) {
+	LIST_FOREACH(m, nconf, entry) {
+		if ((xm = mrt_get(xconf, m)) == NULL) {
 			/* NEW */
-			if ((xm = calloc(1, sizeof(struct mrt))) == NULL)
+			if ((xm = calloc(1, sizeof(struct mrt_config))) == NULL)
 				fatal("mrt_mergeconfig");
-			memcpy(xm, m, sizeof(struct mrt));
-			msgbuf_init(&xm->msgbuf);
-			xm->conf.id = max_id++;
-			xm->state = MRT_STATE_OPEN;
-			LIST_INSERT_HEAD(xconf, xm, list);
+			memcpy(xm, m, sizeof(struct mrt_config));
+			xm->fd = -1;
+			MRT2MC(xm)->state = MRT_STATE_OPEN;
+			LIST_INSERT_HEAD(xconf, xm, entry);
 		} else {
 			/* MERGE */
-			if (strlcpy(xm->name, m->name, sizeof(xm->name)) >=
-			    sizeof(xm->name))
+			if (strlcpy(MRT2MC(xm)->name, MRT2MC(xm)->name,
+			    sizeof(MRT2MC(xm)->name)) >=
+			    sizeof(MRT2MC(xm)->name))
 				fatalx("mrt_mergeconfig: strlcpy");
-			xm->ReopenTimerInterval = m->ReopenTimerInterval;
-			xm->state = MRT_STATE_REOPEN;
+			MRT2MC(xm)->ReopenTimerInterval =
+			    MRT2MC(m)->ReopenTimerInterval;
+			MRT2MC(xm)->state = MRT_STATE_REOPEN;
 		}
+	}
 
-	LIST_FOREACH(xm, xconf, list)
-		if (getconf(nconf, xm) == NULL)
+	LIST_FOREACH(xm, xconf, entry)
+		if (mrt_get(nconf, xm) == NULL)
 			/* REMOVE */
-			xm->state = MRT_STATE_TOREMOVE;
+			MRT2MC(xm)->state = MRT_STATE_REMOVE;
 
 	/* free config */
 	while ((m = LIST_FIRST(nconf)) != NULL) {
-		LIST_REMOVE(m, list);
+		LIST_REMOVE(m, entry);
 		free(m);
 	}
 
diff --git a/usr.sbin/bgpd/mrt.h b/usr.sbin/bgpd/mrt.h
index 8e281461a59..19c457f0edf 100644
--- a/usr.sbin/bgpd/mrt.h
+++ b/usr.sbin/bgpd/mrt.h
@@ -1,4 +1,4 @@
-/*	$OpenBSD: mrt.h,v 1.10 2004/02/25 19:48:18 claudio Exp $ */
+/*	$OpenBSD: mrt.h,v 1.11 2004/07/03 17:19:59 claudio Exp $ */
 
 /*
  * Copyright (c) 2003, 2004 Claudio Jeker <claudio@openbsd.org>
@@ -244,56 +244,52 @@ enum mrt_type {
 };
 
 enum mrt_state {
-	MRT_STATE_STOPPED,
 	MRT_STATE_RUNNING,
 	MRT_STATE_OPEN,
-	MRT_STATE_CLOSE,
 	MRT_STATE_REOPEN,
-	MRT_STATE_TOREMOVE,
 	MRT_STATE_REMOVE
 };
 
-LIST_HEAD(mrt_config_head, mrt_config);
-
-struct mrt_config {
-	enum mrt_type		 type;
-	u_int32_t		 id;
-	u_int32_t		 peer_id;
-	u_int32_t		 group_id;
-	struct imsgbuf		*ibuf;
-	LIST_ENTRY(mrt_config)	 list;			/* used in the SE */
+struct mrt {
+	enum mrt_type		type;
+	u_int32_t		peer_id;
+	u_int32_t		group_id;
+	u_int32_t		queued;
+	int			fd;
+	TAILQ_HEAD(, buf)	bufs;
+	LIST_ENTRY(mrt)		entry;
 };
 
-struct mrt {
-	struct mrt_config	 conf;
-	time_t			 ReopenTimer;
-	time_t			 ReopenTimerInterval;
-	enum mrt_state		 state;
-	struct msgbuf		 msgbuf;
-	struct imsgbuf		*ibuf;
-	char			 name[MRT_FILE_LEN];	/* base file name */
-	char			 file[MRT_FILE_LEN];	/* actual file name */
-	LIST_ENTRY(mrt)		 list;			/* used in the parent */
+struct mrt_config {
+	struct mrt		conf;
+	time_t			ReopenTimer;
+	time_t			ReopenTimerInterval;
+	enum mrt_state		state;
+	char			name[MRT_FILE_LEN];	/* base file name */
+	char			file[MRT_FILE_LEN];	/* actual file name */
 };
 
+#define	MRT2MC(x)	((struct mrt_config *)(x))
+#define	MRT_MAX_TIMEOUT	7200
 
 struct prefix;
 struct pt_entry;
 
 /* prototypes */
-int	mrt_dump_bgp_msg(struct mrt_config *, void *, u_int16_t,
-	    struct peer_config *, struct bgpd_config *);
-int	mrt_dump_state(struct mrt_config *, u_int16_t, u_int16_t,
-	    struct peer_config *, struct bgpd_config *);
-void	mrt_clear_seq(void);
-void	mrt_dump_upcall(struct pt_entry *, void *);
-void	mrt_init(struct imsgbuf *, struct imsgbuf *);
-void	mrt_abort(struct mrt *);
-int	mrt_queue(struct mrt_head *, struct imsg *);
-int	mrt_write(struct mrt *);
-int	mrt_select(struct mrt_head *, struct pollfd *, struct mrt **,
-	    int, int, int *);
-int	mrt_handler(struct mrt_head *);
-int	mrt_mergeconfig(struct mrt_head *, struct mrt_head *);
+int		 mrt_dump_bgp_msg(struct mrt *, void *, u_int16_t,
+		     struct peer_config *, struct bgpd_config *);
+int		 mrt_dump_state(struct mrt *, u_int16_t, u_int16_t,
+		     struct peer_config *, struct bgpd_config *);
+void		 mrt_clear_seq(void);
+void		 mrt_dump_upcall(struct pt_entry *, void *);
+int		 mrt_write(struct mrt *);
+void		 mrt_clean(struct mrt *);
+void		 mrt_init(struct imsgbuf *, struct imsgbuf *);
+void		 mrt_close(struct mrt *);
+int		 mrt_timeout(struct mrt_head *);
+void		 mrt_reconfigure(struct mrt_head *);
+void		 mrt_handler(struct mrt_head *);
+struct mrt	*mrt_get(struct mrt_head *, struct mrt *);
+int		 mrt_mergeconfig(struct mrt_head *, struct mrt_head *);
 
 #endif
diff --git a/usr.sbin/bgpd/parse.y b/usr.sbin/bgpd/parse.y
index a37537c017f..d970903d46c 100644
--- a/usr.sbin/bgpd/parse.y
+++ b/usr.sbin/bgpd/parse.y
@@ -1,4 +1,4 @@
-/*	$OpenBSD: parse.y,v 1.116 2004/06/23 00:11:27 claudio Exp $ */
+/*	$OpenBSD: parse.y,v 1.117 2004/07/03 17:19:59 claudio Exp $ */
 
 /*
  * Copyright (c) 2002, 2003, 2004 Henning Brauer <henning@openbsd.org>
@@ -1573,44 +1573,44 @@ add_mrtconfig(enum mrt_type type, char *name, time_t timeout, struct peer *p)
 {
 	struct mrt	*m, *n;
 
-	LIST_FOREACH(m, mrtconf, list) {
+	LIST_FOREACH(m, mrtconf, entry) {
 		if (p == NULL) {
-			if (m->conf.peer_id != 0 || m->conf.group_id != 0)
+			if (m->peer_id != 0 || m->group_id != 0)
 				continue;
 		} else {
-			if (m->conf.peer_id != p->conf.id ||
-			    m->conf.group_id != p->conf.groupid)
+			if (m->peer_id != p->conf.id ||
+			    m->group_id != p->conf.groupid)
 				continue;
 		}
-		if (m->conf.type == type) {
+		if (m->type == type) {
 			yyerror("only one mrtdump per type allowed.");
 			return (-1);
 		}
 	}
 
-	if ((n = calloc(1, sizeof(struct mrt))) == NULL)
+	if ((n = calloc(1, sizeof(struct mrt_config))) == NULL)
 		fatal("add_mrtconfig");
 
-	n->conf.type = type;
-	n->msgbuf.fd = -1;
-	if (strlcpy(n->name, name, sizeof(n->name)) >= sizeof(n->name)) {
+	n->type = type;
+	if (strlcpy(MRT2MC(n)->name, name, sizeof(MRT2MC(n)->name)) >=
+	    sizeof(MRT2MC(n)->name)) {
 		yyerror("filename \"%s\" too long: max %u",
-		    name, sizeof(n->name) - 1);
+		    name, sizeof(MRT2MC(n)->name) - 1);
 		free(n);
 		return (-1);
 	}
-	n->ReopenTimerInterval = timeout;
+	MRT2MC(n)->ReopenTimerInterval = timeout;
 	if (p != NULL) {
 		if (curgroup == p) {
-			n->conf.peer_id = 0;
-			n->conf.group_id = p->conf.id;
+			n->peer_id = 0;
+			n->group_id = p->conf.id;
 		} else {
-			n->conf.peer_id = p->conf.id;
-			n->conf.group_id = 0;
+			n->peer_id = p->conf.id;
+			n->group_id = 0;
 		}
 	}
 
-	LIST_INSERT_HEAD(mrtconf, n, list);
+	LIST_INSERT_HEAD(mrtconf, n, entry);
 
 	return (0);
 }
diff --git a/usr.sbin/bgpd/printconf.c b/usr.sbin/bgpd/printconf.c
index 95871b3b8c6..8f592d246b3 100644
--- a/usr.sbin/bgpd/printconf.c
+++ b/usr.sbin/bgpd/printconf.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: printconf.c,v 1.22 2004/06/20 18:35:12 henning Exp $	*/
+/*	$OpenBSD: printconf.c,v 1.23 2004/07/03 17:19:59 claudio Exp $	*/
 
 /*
  * Copyright (c) 2003, 2004 Henning Brauer <henning@openbsd.org>
@@ -353,16 +353,17 @@ print_mrt(u_int32_t pid, u_int32_t gid, const char *prep)
 	if (xmrt_l == NULL)
 		return;
 
-	LIST_FOREACH(m, xmrt_l, list)
-		if ((gid != 0 && m->conf.group_id == gid) ||
-		    (m->conf.peer_id == pid && m->conf.group_id == gid)) {
-			if (m->ReopenTimerInterval == 0)
+	LIST_FOREACH(m, xmrt_l, entry)
+		if ((gid != 0 && m->group_id == gid) ||
+		    (m->peer_id == pid && m->group_id == gid)) {
+			if (MRT2MC(m)->ReopenTimerInterval == 0)
 				printf("%sdump %s %s\n", prep,
-				    mrt_type(m->conf.type), m->name);
+				    mrt_type(m->type), MRT2MC(m)->name);
 			else
 				printf("%sdump %s %s %d\n", prep,
-				    mrt_type(m->conf.type),
-				    m->name, m->ReopenTimerInterval);
+				    mrt_type(m->type),
+				    MRT2MC(m)->name,
+				    MRT2MC(m)->ReopenTimerInterval);
 		}
 }
 
diff --git a/usr.sbin/bgpd/rde.c b/usr.sbin/bgpd/rde.c
index 1bc3ee8895c..0a0abd969a1 100644
--- a/usr.sbin/bgpd/rde.c
+++ b/usr.sbin/bgpd/rde.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: rde.c,v 1.121 2004/06/24 23:15:58 claudio Exp $ */
+/*	$OpenBSD: rde.c,v 1.122 2004/07/03 17:19:59 claudio Exp $ */
 
 /*
  * Copyright (c) 2003, 2004 Henning Brauer <henning@openbsd.org>
@@ -35,6 +35,7 @@
 
 #define	PFD_PIPE_MAIN		0
 #define PFD_PIPE_SESSION	1
+#define PFD_MRT_FILE		2
 
 void		 rde_sighdlr(int);
 void		 rde_dispatch_imsg_session(struct imsgbuf *);
@@ -84,6 +85,7 @@ struct rde_peer		 peerdynamic;
 struct filter_head	*rules_l, *newrules;
 struct imsgbuf		 ibuf_se;
 struct imsgbuf		 ibuf_main;
+struct mrt		*mrt;
 
 void
 rde_sighdlr(int sig)
@@ -107,10 +109,9 @@ rde_main(struct bgpd_config *config, struct network_head *net_l,
 {
 	pid_t			 pid;
 	struct passwd		*pw;
-	struct mrt		*m;
 	struct listen_addr	*la;
-	struct pollfd		 pfd[2];
-	int			 nfds;
+	struct pollfd		 pfd[3];
+	int			 nfds, i;
 
 	switch (pid = fork()) {
 	case -1:
@@ -153,10 +154,12 @@ rde_main(struct bgpd_config *config, struct network_head *net_l,
 	imsg_init(&ibuf_main, pipe_m2r[1]);
 
 	/* main mrt list and listener list are not used in the SE */
-	while ((m = LIST_FIRST(mrt_l)) != NULL) {
-		LIST_REMOVE(m, list);
-		free(m);
+	while ((mrt = LIST_FIRST(mrt_l)) != NULL) {
+		LIST_REMOVE(mrt, entry);
+		free(mrt);
 	}
+	mrt = NULL;
+
 	while ((la = TAILQ_FIRST(config->listen_addrs)) != NULL) {
 		TAILQ_REMOVE(config->listen_addrs, la, entry);
 		free(la);
@@ -184,7 +187,14 @@ rde_main(struct bgpd_config *config, struct network_head *net_l,
 		if (ibuf_se.w.queued > 0)
 			pfd[PFD_PIPE_SESSION].events |= POLLOUT;
 
-		if ((nfds = poll(pfd, 2, INFTIM)) == -1)
+		i = 2;
+		if (mrt && mrt->queued) {
+			pfd[PFD_MRT_FILE].fd = mrt->fd;
+			pfd[PFD_MRT_FILE].events = POLLOUT;
+			i++;
+		}
+				
+		if ((nfds = poll(pfd, i, INFTIM)) == -1)
 			if (errno != EINTR)
 				fatal("poll error");
 
@@ -207,6 +217,15 @@ rde_main(struct bgpd_config *config, struct network_head *net_l,
 			nfds--;
 			rde_dispatch_imsg_session(&ibuf_se);
 		}
+
+		if (nfds > 0 && pfd[PFD_MRT_FILE].revents & POLLOUT) {
+			if (mrt_write(mrt) == -1) {
+				free(mrt);
+				mrt = NULL;
+			} else if (mrt->queued == 0)
+				close(mrt->fd);
+		}
+
 		rde_update_queue_runner();
 	}
 
@@ -334,8 +353,8 @@ void
 rde_dispatch_imsg_parent(struct imsgbuf *ibuf)
 {
 	struct imsg		 imsg;
-	struct mrt_config	 mrt;
 	struct filter_rule	*r;
+	struct mrt		*xmrt;
 	int			 n;
 
 	if ((n = imsg_read(ibuf)) == -1)
@@ -400,18 +419,44 @@ rde_dispatch_imsg_parent(struct imsgbuf *ibuf)
 		case IMSG_NEXTHOP_UPDATE:
 			nexthop_update(imsg.data);
 			break;
-		case IMSG_MRT_REQ:
-			memcpy(&mrt, imsg.data, sizeof(mrt));
-			mrt.ibuf = &ibuf_main;
-			if (mrt.type == MRT_TABLE_DUMP) {
-				mrt_clear_seq();
-				pt_dump(mrt_dump_upcall, &mrt, AF_UNSPEC);
-				if (imsg_compose(&ibuf_main, IMSG_MRT_END,
-				    mrt.id, NULL, 0) == -1)
-					fatalx("imsg_compose error");
+		case IMSG_MRT_OPEN:
+		case IMSG_MRT_REOPEN:
+			if (imsg.hdr.len > IMSG_HEADER_SIZE +
+			    sizeof(struct mrt)) {
+				log_warnx("wrong imsg len");
+				break;
+			}
+
+			xmrt = calloc(1, sizeof(struct mrt));
+			if (xmrt == NULL)
+				fatal("rde_dispatch_imsg_parent");
+			memcpy(xmrt, imsg.data, sizeof(struct mrt));
+			TAILQ_INIT(&xmrt->bufs);
+
+			if ((xmrt->fd = imsg_get_fd(ibuf)) == -1)
+				log_warnx("expected to receive fd for mrt dump "
+				    "but didn't receive any");
+
+			/* tell parent to close fd */
+			if (imsg_compose(&ibuf_main, IMSG_MRT_CLOSE, 0,
+			    xmrt, sizeof(struct mrt)) == -1)
+				log_warn("rde_dispatch_imsg_parent: mrt close");
+
+			if (xmrt->type == MRT_TABLE_DUMP) {
+				/* do not dump if a other is still running */
+				if (mrt == NULL || mrt->queued == 0) {
+					free(mrt);
+					mrt = xmrt;
+					mrt_clear_seq();
+					pt_dump(mrt_dump_upcall, mrt,
+					    AF_UNSPEC);
+					break;
+				}
 			}
+			close(xmrt->fd);
+			free(xmrt);
 			break;
-		case IMSG_MRT_END:
+		case IMSG_MRT_CLOSE:
 			/* ignore end message because a dump is atomic */
 			break;
 		default:
@@ -850,7 +895,7 @@ rde_update_log(const char *message,
 
 	if (asprintf(&p, "%s/%u", log_addr(prefix), prefixlen) == -1)
 		p = NULL;
-	log_debug("neighbor %s (AS%u) %s %s/%u %s",
+	log_info("neighbor %s (AS%u) %s %s/%u %s",
 	    log_addr(&peer->conf.remote_addr), peer->conf.remote_as, message,
 	    p ? p : "out of memory", nexthop ? nexthop : "");
 
diff --git a/usr.sbin/bgpd/session.c b/usr.sbin/bgpd/session.c
index 00063c8aab8..91fa899164b 100644
--- a/usr.sbin/bgpd/session.c
+++ b/usr.sbin/bgpd/session.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: session.c,v 1.180 2004/06/22 07:58:19 alexander Exp $ */
+/*	$OpenBSD: session.c,v 1.181 2004/07/03 17:19:59 claudio Exp $ */
 
 /*
  * Copyright (c) 2003, 2004 Henning Brauer <henning@openbsd.org>
@@ -94,7 +94,7 @@ int			 csock = -1;
 struct imsgbuf		 ibuf_rde;
 struct imsgbuf		 ibuf_main;
 
-struct mrt_config_head	 mrt_l;
+struct mrt_head		 mrthead;
 
 void
 session_sighdlr(int sig)
@@ -160,16 +160,17 @@ session_main(struct bgpd_config *config, struct peer *cpeers,
     struct network_head *net_l, struct filter_head *rules,
     struct mrt_head *m_l, int pipe_m2s[2], int pipe_s2r[2])
 {
-	int			 nfds, i, j, timeout, idx_peers, idx_listeners;
+	int			 nfds, i, j, timeout;
+	int			 idx_peers, idx_listeners, idx_mrts;
 	pid_t			 pid;
 	time_t			 nextaction;
-	u_int			 pfd_elms = 0, peer_l_elms = 0, new_cnt;
-	u_int			 listener_cnt, peer_cnt, ctl_cnt;
+	u_int			 pfd_elms = 0, peer_l_elms = 0, mrt_l_elms = 0;
+	u_int			 listener_cnt, peer_cnt, ctl_cnt, mrt_cnt;
+	u_int			 new_cnt;
 	struct passwd		*pw;
 	struct peer		*p, **peer_l = NULL, *last, *next;
 	struct network		*net;
-	struct mrt		*m;
-	struct mrt_config	*mrt;
+	struct mrt		*m, **mrt_l = NULL;
 	struct filter_rule	*r;
 	struct pollfd		*pfd = NULL;
 	struct ctl_conn		*ctl_conn;
@@ -228,7 +229,7 @@ session_main(struct bgpd_config *config, struct peer *cpeers,
 	imsg_init(&ibuf_main, pipe_m2s[1]);
 	TAILQ_INIT(&ctl_conns);
 	csock = control_listen();
-	LIST_INIT(&mrt_l);
+	LIST_INIT(&mrthead);
 	peer_cnt = 0;
 	ctl_cnt = 0;
 
@@ -247,7 +248,7 @@ session_main(struct bgpd_config *config, struct peer *cpeers,
 
 	/* main mrt list is not used in the SE */
 	while ((m = LIST_FIRST(m_l)) != NULL) {
-		LIST_REMOVE(m, list);
+		LIST_REMOVE(m, entry);
 		free(m);
 	}
 
@@ -300,8 +301,28 @@ session_main(struct bgpd_config *config, struct peer *cpeers,
 			peer_l_elms = peer_cnt + PEER_L_RESERVE;
 		}
 
-		new_cnt =
-		    PFD_LISTENERS_START + listener_cnt + peer_cnt + ctl_cnt;
+		mrt_cnt = 0;
+		LIST_FOREACH(m, &mrthead, entry) {
+			if (m->queued)
+				mrt_cnt++;
+		}
+
+		if (mrt_cnt > mrt_l_elms ||
+		    mrt_cnt + 2 * PEER_L_RESERVE < mrt_l_elms) {
+			if ((newp = realloc(mrt_l, sizeof(struct mrt *) *
+			    (mrt_cnt + PEER_L_RESERVE))) == NULL) {
+				/* panic for now  */
+				log_warn("could not resize mrt_l from %u -> %u"
+				    " entries", mrt_l_elms,
+				    mrt_cnt + PEER_L_RESERVE);
+				fatalx("exiting");
+			}
+			mrt_l = newp;
+			mrt_l_elms = mrt_cnt + PEER_L_RESERVE;
+		}
+
+		new_cnt = PFD_LISTENERS_START + listener_cnt + peer_cnt +
+		    ctl_cnt + mrt_cnt;
 		if (new_cnt > pfd_elms ||
 		    (new_cnt + 2) * PFD_RESERVE < pfd_elms) {
 			if ((newp = realloc(pfd, sizeof(struct pollfd) *
@@ -391,6 +412,16 @@ session_main(struct bgpd_config *config, struct peer *cpeers,
 
 		idx_peers = i;
 
+		LIST_FOREACH(m, &mrthead, entry)
+			if (m->queued) {
+				pfd[i].fd = m->fd;
+				pfd[i].events = POLLOUT;
+				mrt_l[i - idx_peers] = m;
+				i++;
+			}
+
+		idx_mrts = i;
+
 		TAILQ_FOREACH(ctl_conn, &ctl_conns, entry) {
 			pfd[i].fd = ctl_conn->ibuf.fd;
 			pfd[i].events = POLLIN;
@@ -442,6 +473,12 @@ session_main(struct bgpd_config *config, struct peer *cpeers,
 			nfds -= session_dispatch_msg(&pfd[j],
 			    peer_l[j - idx_listeners]);
 
+		for (; nfds > 0 && j < idx_mrts; j++)
+			if (pfd[j].revents & POLLOUT) {
+				nfds--;
+				mrt_write(mrt_l[j - idx_peers]);
+			}
+
 		for (; nfds > 0 && j < i; j++)
 			nfds -= control_dispatch_msg(&pfd[j], &ctl_cnt);
 	}
@@ -453,9 +490,10 @@ session_main(struct bgpd_config *config, struct peer *cpeers,
 		free(p);
 	}
 
-	while ((mrt = LIST_FIRST(&mrt_l)) != NULL) {
-		LIST_REMOVE(mrt, list);
-		free(mrt);
+	while ((m = LIST_FIRST(&mrthead)) != NULL) {
+		mrt_clean(m);
+		LIST_REMOVE(m, entry);
+		free(m);
 	}
 
 	while ((la = TAILQ_FIRST(conf->listen_addrs)) != NULL) {
@@ -760,7 +798,7 @@ void
 change_state(struct peer *peer, enum session_state state,
     enum session_events event)
 {
-	struct mrt_config	*mrt;
+	struct mrt	*mrt;
 
 	switch (state) {
 	case STATE_IDLE:
@@ -818,7 +856,7 @@ change_state(struct peer *peer, enum session_state state,
 	}
 
 	log_statechange(peer, state, event);
-	LIST_FOREACH(mrt, &mrt_l, list) {
+	LIST_FOREACH(mrt, &mrthead, entry) {
 		if (mrt->type != MRT_ALL_IN && mrt->type != MRT_ALL_OUT)
 			continue;
 		if ((mrt->peer_id == 0 && mrt->group_id == 0) ||
@@ -1074,7 +1112,7 @@ session_open(struct peer *p)
 {
 	struct msg_open		 msg;
 	struct buf		*buf;
-	struct mrt_config	*mrt;
+	struct mrt		*mrt;
 	u_int16_t		 len;
 	int			 errs = 0;
 	u_int8_t		 op_type, op_len = 0, optparamlen = 0;
@@ -1154,7 +1192,7 @@ session_open(struct peer *p)
 	}
 
 	if (errs == 0) {
-		LIST_FOREACH(mrt, &mrt_l, list) {
+		LIST_FOREACH(mrt, &mrthead, entry) {
 			if (mrt->type != MRT_ALL_OUT)
 				continue;
 			if ((mrt->peer_id == 0 && mrt->group_id == 0) ||
@@ -1184,7 +1222,7 @@ session_keepalive(struct peer *peer)
 {
 	struct msg_header	 msg;
 	struct buf		*buf;
-	struct mrt_config	*mrt;
+	struct mrt		*mrt;
 	ssize_t			 len;
 	int			 errs = 0;
 
@@ -1208,7 +1246,7 @@ session_keepalive(struct peer *peer)
 		return;
 	}
 
-	LIST_FOREACH(mrt, &mrt_l, list) {
+	LIST_FOREACH(mrt, &mrthead, entry) {
 		if (mrt->type != MRT_ALL_OUT)
 			continue;
 		if ((mrt->peer_id == 0 && mrt->group_id == 0) ||
@@ -1234,7 +1272,7 @@ session_update(u_int32_t peerid, void *data, size_t datalen)
 	struct peer		*p;
 	struct msg_header	 msg;
 	struct buf		*buf;
-	struct mrt_config	*mrt;
+	struct mrt		*mrt;
 	ssize_t			 len;
 	int			 errs = 0;
 
@@ -1267,7 +1305,7 @@ session_update(u_int32_t peerid, void *data, size_t datalen)
 		return;
 	}
 
-	LIST_FOREACH(mrt, &mrt_l, list) {
+	LIST_FOREACH(mrt, &mrthead, entry) {
 		if (mrt->type != MRT_ALL_OUT && mrt->type != MRT_UPDATE_OUT)
 			continue;
 		if ((mrt->peer_id == 0 && mrt->group_id == 0) ||
@@ -1293,7 +1331,7 @@ session_notification(struct peer *peer, u_int8_t errcode, u_int8_t subcode,
 {
 	struct msg_header	 msg;
 	struct buf		*buf;
-	struct mrt_config	*mrt;
+	struct mrt		*mrt;
 	ssize_t			 len;
 	int			 errs = 0;
 
@@ -1322,7 +1360,7 @@ session_notification(struct peer *peer, u_int8_t errcode, u_int8_t subcode,
 		return;
 	}
 
-	LIST_FOREACH(mrt, &mrt_l, list) {
+	LIST_FOREACH(mrt, &mrthead, entry) {
 		if (mrt->type != MRT_ALL_OUT)
 			continue;
 		if ((mrt->peer_id == 0 && mrt->group_id == 0) ||
@@ -1481,7 +1519,7 @@ session_dispatch_msg(struct pollfd *pfd, struct peer *p)
 int
 parse_header(struct peer *peer, u_char *data, u_int16_t *len, u_int8_t *type)
 {
-	struct mrt_config	*mrt;
+	struct mrt		*mrt;
 	u_char			*p;
 	u_char			 one = 0xff;
 	int			 i;
@@ -1566,7 +1604,7 @@ parse_header(struct peer *peer, u_char *data, u_int16_t *len, u_int8_t *type)
 		    type, 1);
 		return (-1);
 	}
-	LIST_FOREACH(mrt, &mrt_l, list) {
+	LIST_FOREACH(mrt, &mrthead, entry) {
 		if (mrt->type != MRT_ALL_IN && (mrt->type != MRT_UPDATE_IN ||
 		    *type != UPDATE))
 			continue;
@@ -1971,8 +2009,8 @@ void
 session_dispatch_imsg(struct imsgbuf *ibuf, int idx, u_int *listener_cnt)
 {
 	struct imsg		 imsg;
-	struct mrt_config	 xmrt;
-	struct mrt_config	*mrt;
+	struct mrt		 xmrt;
+	struct mrt		*mrt;
 	struct peer_config	*pconf;
 	struct peer		*p, *next;
 	struct listen_addr	*la, *nla;
@@ -2118,25 +2156,52 @@ session_dispatch_imsg(struct imsgbuf *ibuf, int idx, u_int *listener_cnt)
 			pending_reconf = 0;
 			log_info("SE reconfigured");
 			break;
-		case IMSG_MRT_REQ:
-			if ((mrt = calloc(1, sizeof(struct mrt_config))) ==
-			    NULL)
-				fatal("session_dispatch_imsg");
-			memcpy(mrt, imsg.data, sizeof(struct mrt_config));
-			mrt->ibuf = &ibuf_main;
-			LIST_INSERT_HEAD(&mrt_l, mrt, list);
+		case IMSG_MRT_OPEN:
+		case IMSG_MRT_REOPEN:
+			if (imsg.hdr.len > IMSG_HEADER_SIZE +
+			    sizeof(struct mrt)) {
+				log_warnx("wrong imsg len");
+				break;
+			}
+
+			memcpy(&xmrt, imsg.data, sizeof(struct mrt));
+			if ((xmrt.fd = imsg_get_fd(ibuf)) == -1)
+				log_warnx("expected to receive fd for mrt dump "
+				    "but didn't receive any");
+
+			mrt = mrt_get(&mrthead, &xmrt);
+			if (mrt == NULL) {
+				/* new dump */
+				mrt = calloc(1, sizeof(struct mrt));
+				if (mrt == NULL)
+					fatal("session_dispatch_imsg");
+				memcpy(mrt, &xmrt, sizeof(struct mrt));
+				TAILQ_INIT(&mrt->bufs);
+				LIST_INSERT_HEAD(&mrthead, mrt, entry);
+			} else {
+				/* old dump reopened */
+				close(mrt->fd);
+				mrt->fd = xmrt.fd;
+			}
+
+			/* tell parent to close fd */
+			if (imsg_compose(&ibuf_main, IMSG_MRT_CLOSE, 0,
+			    &xmrt, sizeof(struct mrt)) == -1)
+				log_warn("session_dispatch_imsg: mrt close");
 			break;
-		case IMSG_MRT_END:
-			memcpy(&xmrt, imsg.data, sizeof(struct mrt_config));
-			LIST_FOREACH(mrt, &mrt_l, list) {
-				if (mrt->type != xmrt.type)
-					continue;
-				if (mrt->peer_id == xmrt.peer_id &&
-				    mrt->group_id == xmrt.group_id) {
-					LIST_REMOVE(mrt, list);
-					free(mrt);
-					break;
-				}
+		case IMSG_MRT_CLOSE:
+			if (imsg.hdr.len > IMSG_HEADER_SIZE +
+			    sizeof(struct mrt)) {
+				log_warnx("wrong imsg len");
+				break;
+			}
+
+			memcpy(&xmrt, imsg.data, sizeof(struct mrt));
+			mrt = mrt_get(&mrthead, &xmrt);
+			if (mrt != NULL) {
+				mrt_clean(mrt);
+				LIST_REMOVE(mrt, entry);
+				free(mrt);
 			}
 			break;
 		case IMSG_CTL_KROUTE: