Use the zone allocator to allocate inpcbs and tcpcbs. Each protocol creates

its own zone; this is used particularly by TCP which allocates both inpcb and
tcpcb in a single allocation.  (Some hackery ensures that the tcpcb is
reasonably aligned.)  Also keep track of the number of pcbs of each type
allocated, and keep a generation count (instance version number) for future
use.

7 files changed, 140 insertions, 33 deletions

diff --git a/sys/netinet/in_pcb.c b/sys/netinet/in_pcb.c
index ede159d..59bb5d9 100644
--- a/sys/netinet/in_pcb.c
+++ b/sys/netinet/in_pcb.c
@@ -31,7 +31,7 @@
  * SUCH DAMAGE.
  *
  *	@(#)in_pcb.c	8.4 (Berkeley) 5/24/95
- *	$Id: in_pcb.c,v 1.38 1998/01/27 09:15:03 davidg Exp $
+ *	$Id: in_pcb.c,v 1.39 1998/03/01 19:39:26 guido Exp $
  */
 
 #include <sys/param.h>
@@ -44,6 +44,7 @@
 #include <sys/proc.h>
 #include <sys/kernel.h>
 #include <sys/sysctl.h>
+#include <vm/vm_zone.h>		/* for zalloci, zfreei prototypes */
 
 #include <net/if.h>
 #include <net/route.h>
@@ -125,14 +126,15 @@ in_pcballoc(so, pcbinfo, p)
 {
 	register struct inpcb *inp;
 
-	MALLOC(inp, struct inpcb *, sizeof(*inp), M_PCB, 
-	       p ? M_WAITOK : M_NOWAIT);
+	inp = zalloci(pcbinfo->ipi_zone);
 	if (inp == NULL)
 		return (ENOBUFS);
 	bzero((caddr_t)inp, sizeof(*inp));
+	inp->inp_gencnt = ++pcbinfo->ipi_gencnt;
 	inp->inp_pcbinfo = pcbinfo;
 	inp->inp_socket = so;
 	LIST_INSERT_HEAD(pcbinfo->listhead, inp, inp_list);
+	pcbinfo->ipi_count++;
 	so->so_pcb = (caddr_t)inp;
 	return (0);
 }
@@ -464,7 +466,9 @@ in_pcbdetach(inp)
 	struct inpcb *inp;
 {
 	struct socket *so = inp->inp_socket;
+	struct inpcbinfo *ipi = inp->inp_pcbinfo;
 
+	inp->inp_gencnt = ++ipi->ipi_gencnt;
 	in_pcbremlists(inp);
 	so->so_pcb = 0;
 	sofree(so);
@@ -473,7 +477,7 @@ in_pcbdetach(inp)
 	if (inp->inp_route.ro_rt)
 		rtfree(inp->inp_route.ro_rt);
 	ip_freemoptions(inp->inp_moptions);
-	FREE(inp, M_PCB);
+	zfreei(ipi->ipi_zone, inp);
 }
 
 /*
@@ -879,4 +883,5 @@ in_pcbremlists(inp)
 		}
 	}
 	LIST_REMOVE(inp, inp_list);
+	inp->inp_pcbinfo->ipi_count--;
 }
diff --git a/sys/netinet/in_pcb.h b/sys/netinet/in_pcb.h
index 27f3586..6b446ad 100644
--- a/sys/netinet/in_pcb.h
+++ b/sys/netinet/in_pcb.h
@@ -31,7 +31,7 @@
  * SUCH DAMAGE.
  *
  *	@(#)in_pcb.h	8.1 (Berkeley) 6/10/93
- * $Id: in_pcb.h,v 1.22 1997/08/16 19:15:36 wollman Exp $
+ * $Id: in_pcb.h,v 1.23 1998/01/27 09:15:04 davidg Exp $
  */
 
 #ifndef _NETINET_IN_PCB_H_
@@ -49,6 +49,12 @@
 LIST_HEAD(inpcbhead, inpcb);
 LIST_HEAD(inpcbporthead, inpcbport);
 
+/*
+ * NB: the zone allocator is type-stable EXCEPT FOR THE FIRST TWO LONGS
+ * of the structure.  Therefore, it is important that the members in
+ * that position not contain any information which is required to be
+ * stable.
+ */
 struct inpcb {
 	LIST_ENTRY(inpcb) inp_hash;	/* hash list */
 	struct	in_addr inp_faddr;	/* foreign host table entry */
@@ -69,7 +75,14 @@ struct inpcb {
 	struct	ip_moptions *inp_moptions; /* IP multicast options */
 	LIST_ENTRY(inpcb) inp_portlist;	/* list for this PCB's local port */
 	struct	inpcbport *inp_phd;	/* head of list for this PCB's local port */
+	u_quad_t inp_gencnt;	/* generation count of this instance */
 };
+/*
+ * The range of the generation count, as used in this implementation,
+ * is 9e19.  We would have to create 300 billion connections per
+ * second for this number to roll over in a year.  This seems sufficiently
+ * unlikely that we simply don't concern ourselves with that possibility.
+ */
 
 struct inpcbport {
 	LIST_ENTRY(inpcbport) phd_hash;
@@ -77,15 +90,18 @@ struct inpcbport {
 	u_short phd_port;
 };
 
-struct inpcbinfo {
-	struct inpcbhead *hashbase;
-	unsigned long hashmask;
-	struct inpcbporthead *porthashbase;
-	unsigned long porthashmask;
-	struct inpcbhead *listhead;
-	unsigned short lastport;
-	unsigned short lastlow;
-	unsigned short lasthi;
+struct inpcbinfo {		/* XXX documentation, prefixes */
+	struct	inpcbhead *hashbase;
+	u_long	hashmask;
+	struct	inpcbporthead *porthashbase;
+	u_long	porthashmask;
+	struct	inpcbhead *listhead;
+	u_short	lastport;
+	u_short	lastlow;
+	u_short	lasthi;
+	struct	vm_zone *ipi_zone; /* zone to allocate pcbs from */
+	u_int	ipi_count;	/* number of pcbs in this list */
+	u_quad_t ipi_gencnt;	/* current generation count */
 };
 
 #define INP_PCBHASH(faddr, lport, fport, mask) \
@@ -131,6 +147,7 @@ void	in_pcbnotify __P((struct inpcbhead *, struct sockaddr *,
 void	in_pcbrehash __P((struct inpcb *));
 int	in_setpeeraddr __P((struct socket *so, struct sockaddr **nam));
 int	in_setsockaddr __P((struct socket *so, struct sockaddr **nam));
+
 #endif
 
 #endif
diff --git a/sys/netinet/ip_divert.c b/sys/netinet/ip_divert.c
index dc9b049..4eddf3a 100644
--- a/sys/netinet/ip_divert.c
+++ b/sys/netinet/ip_divert.c
@@ -30,7 +30,7 @@
  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  * SUCH DAMAGE.
  *
- *	$Id: ip_divert.c,v 1.19 1998/02/04 22:33:06 eivind Exp $
+ *	$Id: ip_divert.c,v 1.20 1998/02/06 12:13:50 eivind Exp $
  */
 
 #include "opt_inet.h"
@@ -47,6 +47,7 @@
 #include <sys/socketvar.h>
 #include <sys/systm.h>
 #include <sys/proc.h>
+#include <vm/vm_zone.h>
 
 #include <net/if.h>
 #include <net/route.h>
@@ -116,6 +117,8 @@ div_init(void)
 	 */
 	divcbinfo.hashbase = hashinit(1, M_PCB, &divcbinfo.hashmask);
 	divcbinfo.porthashbase = hashinit(1, M_PCB, &divcbinfo.porthashmask);
+	divcbinfo.ipi_zone = zinit("divcb", sizeof(struct inpcb),
+				   nmbclusters/4, ZONE_INTERRUPT, 0);
 }
 
 /*
diff --git a/sys/netinet/raw_ip.c b/sys/netinet/raw_ip.c
index 95f364e..a8103ff 100644
--- a/sys/netinet/raw_ip.c
+++ b/sys/netinet/raw_ip.c
@@ -31,7 +31,7 @@
  * SUCH DAMAGE.
  *
  *	@(#)raw_ip.c	8.7 (Berkeley) 5/15/95
- *	$Id: raw_ip.c,v 1.50 1997/12/18 09:13:39 davidg Exp $
+ *	$Id: raw_ip.c,v 1.51 1998/01/27 09:15:07 davidg Exp $
  */
 
 #include <sys/param.h>
@@ -44,6 +44,7 @@
 #include <sys/socket.h>
 #include <sys/socketvar.h>
 #include <sys/sysctl.h>
+#include <vm/vm_zone.h>
 
 #include <net/if.h>
 #include <net/route.h>
@@ -94,6 +95,8 @@ rip_init()
 	 */
 	ripcbinfo.hashbase = hashinit(1, M_PCB, &ripcbinfo.hashmask);
 	ripcbinfo.porthashbase = hashinit(1, M_PCB, &ripcbinfo.porthashmask);
+	ripcbinfo.ipi_zone = zinit("ripcb", sizeof(struct inpcb),
+				   nmbclusters/4, ZONE_INTERRUPT, 0);
 }
 
 static struct	sockaddr_in ripsrc = { sizeof(ripsrc), AF_INET };
diff --git a/sys/netinet/tcp_subr.c b/sys/netinet/tcp_subr.c
index b0bdac2..bf424d5 100644
--- a/sys/netinet/tcp_subr.c
+++ b/sys/netinet/tcp_subr.c
@@ -31,7 +31,7 @@
  * SUCH DAMAGE.
  *
  *	@(#)tcp_subr.c	8.2 (Berkeley) 5/24/95
- *	$Id: tcp_subr.c,v 1.41 1998/01/25 04:23:32 eivind Exp $
+ *	$Id: tcp_subr.c,v 1.42 1998/01/27 09:15:10 davidg Exp $
  */
 
 #include "opt_compat.h"
@@ -46,6 +46,7 @@
 #include <sys/socket.h>
 #include <sys/socketvar.h>
 #include <sys/protosw.h>
+#include <vm/vm_zone.h>
 
 #include <net/route.h>
 #include <net/if.h>
@@ -94,6 +95,26 @@ static void	tcp_notify __P((struct inpcb *, int));
 #endif
 
 /*
+ * This is the actual shape of what we allocate using the zone
+ * allocator.  Doing it this way allows us to protect both structures
+ * using the same generation count, and also eliminates the overhead
+ * of allocating tcpcbs separately.  By hiding the structure here,
+ * we avoid changing most of the rest of the code (although it needs
+ * to be changed, eventually, for greater efficiency).
+ */
+#define ALIGNMENT	32
+#define	ALIGNM1		(ALIGNMENT-1)
+struct	inp_tp {
+	union {
+		struct	inpcb inp;
+		char	align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1];
+	} inp_tp_u;
+	struct	tcpcb tcb;
+};
+#undef ALIGNMENT
+#undef ALIGNM1
+
+/*
  * Tcp initialization
  */
 void
@@ -107,6 +128,23 @@ tcp_init()
 	tcbinfo.listhead = &tcb;
 	tcbinfo.hashbase = hashinit(TCBHASHSIZE, M_PCB, &tcbinfo.hashmask);
 	tcbinfo.porthashbase = hashinit(TCBHASHSIZE, M_PCB, &tcbinfo.porthashmask);
+	/* For the moment, we just worry about putting inpcbs here. */
+	/*
+	 * Rationale for a maximum of `nmbclusters':
+	 * 	1) It's a convenient value, sized by config, based on
+	 *	   parameters already known to be tweakable as needed
+	 *	   for network-intensive systems.
+	 *	2) Under the Old World Order, when pcbs were stored in
+	 *	   mbufs, it was of course impossible to have more
+	 *	   pcbs than mbufs.
+	 *	3) The zone allocator doesn't allocate physical memory
+	 *	   for this many pcbs; it just sizes the virtual
+	 *	   address space appropriately.  Thus, even for very large
+	 *	   values of nmbclusters, we don't actually take up much
+	 *	   memory unless required.
+	 */
+	tcbinfo.ipi_zone = zinit("tcpcb", sizeof(struct inp_tp), nmbclusters,
+				 ZONE_INTERRUPT, 0);
 	if (max_protohdr < sizeof(struct tcpiphdr))
 		max_protohdr = sizeof(struct tcpiphdr);
 	if (max_linkhdr + sizeof(struct tcpiphdr) > MHLEN)
@@ -246,17 +284,18 @@ tcp_respond(tp, ti, m, ack, seq, flags)
 /*
  * Create a new TCP control block, making an
  * empty reassembly queue and hooking it to the argument
- * protocol control block.
+ * protocol control block.  The `inp' parameter must have
+ * come from the zone allocator set up in tcp_init().
  */
 struct tcpcb *
 tcp_newtcpcb(inp)
 	struct inpcb *inp;
 {
+	struct	inp_tp *it;
 	register struct tcpcb *tp;
 
-	tp = malloc(sizeof(*tp), M_PCB, M_NOWAIT);
-	if (tp == NULL)
-		return ((struct tcpcb *)0);
+	it = (struct inp_tp *)inp;
+	tp = &it->tcb;
 	bzero((char *) tp, sizeof(struct tcpcb));
 	tp->seg_next = tp->seg_prev = (struct tcpiphdr *)tp;
 	tp->t_maxseg = tp->t_maxopd = tcp_mssdflt;
@@ -265,7 +304,7 @@ tcp_newtcpcb(inp)
 		tp->t_flags = (TF_REQ_SCALE|TF_REQ_TSTMP);
 	if (tcp_do_rfc1644)
 		tp->t_flags |= TF_REQ_CC;
-	tp->t_inpcb = inp;
+	tp->t_inpcb = inp;	/* XXX */
 	/*
 	 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
 	 * rtt estimate.  Set rttvar so that srtt + 4 * rttvar gives
@@ -279,7 +318,7 @@ tcp_newtcpcb(inp)
 	tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
 	inp->inp_ip_ttl = ip_defttl;
 	inp->inp_ppcb = (caddr_t)tp;
-	return (tp);
+	return (tp);		/* XXX */
 }
 
 /*
@@ -418,7 +457,6 @@ tcp_close(tp)
 	if (tp->t_template)
 		(void) m_free(dtom(tp->t_template));
 	inp->inp_ppcb = NULL;
-	free(tp, M_PCB);
 	soisdisconnected(so);
 	in_pcbdetach(inp);
 	tcpstat.tcps_closed++;
diff --git a/sys/netinet/tcp_timewait.c b/sys/netinet/tcp_timewait.c
index b0bdac2..bf424d5 100644
--- a/sys/netinet/tcp_timewait.c
+++ b/sys/netinet/tcp_timewait.c
@@ -31,7 +31,7 @@
  * SUCH DAMAGE.
  *
  *	@(#)tcp_subr.c	8.2 (Berkeley) 5/24/95
- *	$Id: tcp_subr.c,v 1.41 1998/01/25 04:23:32 eivind Exp $
+ *	$Id: tcp_subr.c,v 1.42 1998/01/27 09:15:10 davidg Exp $
  */
 
 #include "opt_compat.h"
@@ -46,6 +46,7 @@
 #include <sys/socket.h>
 #include <sys/socketvar.h>
 #include <sys/protosw.h>
+#include <vm/vm_zone.h>
 
 #include <net/route.h>
 #include <net/if.h>
@@ -94,6 +95,26 @@ static void	tcp_notify __P((struct inpcb *, int));
 #endif
 
 /*
+ * This is the actual shape of what we allocate using the zone
+ * allocator.  Doing it this way allows us to protect both structures
+ * using the same generation count, and also eliminates the overhead
+ * of allocating tcpcbs separately.  By hiding the structure here,
+ * we avoid changing most of the rest of the code (although it needs
+ * to be changed, eventually, for greater efficiency).
+ */
+#define ALIGNMENT	32
+#define	ALIGNM1		(ALIGNMENT-1)
+struct	inp_tp {
+	union {
+		struct	inpcb inp;
+		char	align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1];
+	} inp_tp_u;
+	struct	tcpcb tcb;
+};
+#undef ALIGNMENT
+#undef ALIGNM1
+
+/*
  * Tcp initialization
  */
 void
@@ -107,6 +128,23 @@ tcp_init()
 	tcbinfo.listhead = &tcb;
 	tcbinfo.hashbase = hashinit(TCBHASHSIZE, M_PCB, &tcbinfo.hashmask);
 	tcbinfo.porthashbase = hashinit(TCBHASHSIZE, M_PCB, &tcbinfo.porthashmask);
+	/* For the moment, we just worry about putting inpcbs here. */
+	/*
+	 * Rationale for a maximum of `nmbclusters':
+	 * 	1) It's a convenient value, sized by config, based on
+	 *	   parameters already known to be tweakable as needed
+	 *	   for network-intensive systems.
+	 *	2) Under the Old World Order, when pcbs were stored in
+	 *	   mbufs, it was of course impossible to have more
+	 *	   pcbs than mbufs.
+	 *	3) The zone allocator doesn't allocate physical memory
+	 *	   for this many pcbs; it just sizes the virtual
+	 *	   address space appropriately.  Thus, even for very large
+	 *	   values of nmbclusters, we don't actually take up much
+	 *	   memory unless required.
+	 */
+	tcbinfo.ipi_zone = zinit("tcpcb", sizeof(struct inp_tp), nmbclusters,
+				 ZONE_INTERRUPT, 0);
 	if (max_protohdr < sizeof(struct tcpiphdr))
 		max_protohdr = sizeof(struct tcpiphdr);
 	if (max_linkhdr + sizeof(struct tcpiphdr) > MHLEN)
@@ -246,17 +284,18 @@ tcp_respond(tp, ti, m, ack, seq, flags)
 /*
  * Create a new TCP control block, making an
  * empty reassembly queue and hooking it to the argument
- * protocol control block.
+ * protocol control block.  The `inp' parameter must have
+ * come from the zone allocator set up in tcp_init().
  */
 struct tcpcb *
 tcp_newtcpcb(inp)
 	struct inpcb *inp;
 {
+	struct	inp_tp *it;
 	register struct tcpcb *tp;
 
-	tp = malloc(sizeof(*tp), M_PCB, M_NOWAIT);
-	if (tp == NULL)
-		return ((struct tcpcb *)0);
+	it = (struct inp_tp *)inp;
+	tp = &it->tcb;
 	bzero((char *) tp, sizeof(struct tcpcb));
 	tp->seg_next = tp->seg_prev = (struct tcpiphdr *)tp;
 	tp->t_maxseg = tp->t_maxopd = tcp_mssdflt;
@@ -265,7 +304,7 @@ tcp_newtcpcb(inp)
 		tp->t_flags = (TF_REQ_SCALE|TF_REQ_TSTMP);
 	if (tcp_do_rfc1644)
 		tp->t_flags |= TF_REQ_CC;
-	tp->t_inpcb = inp;
+	tp->t_inpcb = inp;	/* XXX */
 	/*
 	 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
 	 * rtt estimate.  Set rttvar so that srtt + 4 * rttvar gives
@@ -279,7 +318,7 @@ tcp_newtcpcb(inp)
 	tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
 	inp->inp_ip_ttl = ip_defttl;
 	inp->inp_ppcb = (caddr_t)tp;
-	return (tp);
+	return (tp);		/* XXX */
 }
 
 /*
@@ -418,7 +457,6 @@ tcp_close(tp)
 	if (tp->t_template)
 		(void) m_free(dtom(tp->t_template));
 	inp->inp_ppcb = NULL;
-	free(tp, M_PCB);
 	soisdisconnected(so);
 	in_pcbdetach(inp);
 	tcpstat.tcps_closed++;
diff --git a/sys/netinet/udp_usrreq.c b/sys/netinet/udp_usrreq.c
index 773e42d..f9dd1a7 100644
--- a/sys/netinet/udp_usrreq.c
+++ b/sys/netinet/udp_usrreq.c
@@ -31,7 +31,7 @@
  * SUCH DAMAGE.
  *
  *	@(#)udp_usrreq.c	8.6 (Berkeley) 5/23/95
- *	$Id: udp_usrreq.c,v 1.43 1998/01/25 17:25:41 steve Exp $
+ *	$Id: udp_usrreq.c,v 1.44 1998/01/27 09:15:13 davidg Exp $
  */
 
 #include <sys/param.h>
@@ -44,6 +44,7 @@
 #include <sys/socketvar.h>
 #include <sys/sysctl.h>
 #include <sys/syslog.h>
+#include <vm/vm_zone.h>
 
 #include <net/if.h>
 #include <net/route.h>
@@ -98,6 +99,8 @@ udp_init()
 	udbinfo.listhead = &udb;
 	udbinfo.hashbase = hashinit(UDBHASHSIZE, M_PCB, &udbinfo.hashmask);
 	udbinfo.porthashbase = hashinit(UDBHASHSIZE, M_PCB, &udbinfo.porthashmask);
+	udbinfo.ipi_zone = zinit("udpcb", sizeof(struct inpcb), nmbclusters,
+			     ZONE_INTERRUPT, 0);
 }
 
 void