Hi,

For a while we are running network without kernel lock, but with a
network lock.  The latter is an exclusive sleeping rwlock.

It is possible to run the forwarding path in parallel on multiple
cores.  I use ix(4) interfaces which provide one input queue for
each CPU.  For that we have to start multiple softnet tasks and
replace the exclusive lock with a shared lock.  This works for IP
and IPv6 input and forwarding, but not for higher protocols.

So I implement a queue between IP and higher layers.  We had that
before when we were using netlock for IP and kernel lock for TCP.
Now we have shared lock for IP and exclusive lock for TCP.  By using
a queue, we can upgrade the lock once for multiple packets.

As you can see here, forwardings performance doubles from 4.5x10^9
to 9x10^9 .  Left column is current, right column is with my diff.
The other dots at 2x10^9 are with socket splicing which is not
affected.
http://bluhm.genua.de/perform/results/2021-04-21T10%3A50%3A37Z/gnuplot/forward.png

Here are all numbers with various network tests.
http://bluhm.genua.de/perform/results/2021-04-21T10%3A50%3A37Z/perform.html
TCP performance gets less deterministic due to the addition queue.

Kernel stack flame graph looks like this.  Machine uses 4 CPU.
http://bluhm.genua.de/files/kstack-multiqueue-forward.svg

Note the kernel lock around nd6_resolve().  I hat to put it there
as I have seen an MP related crash there.  This can be fixed
independently of this diff.

We need more MP preassure to find such bugs and races.  I think now
is a good time to give this diff broader testing and commit it.
You need interfaces with multiple queues to see a difference.

ok?

bluhm

Index: net/if.c
===================================================================
RCS file: /data/mirror/openbsd/cvs/src/sys/net/if.c,v
retrieving revision 1.640
diff -u -p -r1.640 if.c
--- net/if.c    26 Mar 2021 22:41:06 -0000      1.640
+++ net/if.c    21 Apr 2021 12:52:08 -0000
@@ -238,7 +238,7 @@ int ifq_congestion;
 
 int             netisr;
 
-#define        NET_TASKQ       1
+#define        NET_TASKQ       4
 struct taskq   *nettqmp[NET_TASKQ];
 
 struct task if_input_task_locked = TASK_INITIALIZER(if_netisr, NULL);
@@ -834,10 +834,10 @@ if_input_process(struct ifnet *ifp, stru
         * to PF globals, pipex globals, unicast and multicast addresses
         * lists and the socket layer.
         */
-       NET_LOCK();
+       NET_RLOCK_IN_SOFTNET();
        while ((m = ml_dequeue(ml)) != NULL)
                (*ifp->if_input)(ifp, m);
-       NET_UNLOCK();
+       NET_RUNLOCK_IN_SOFTNET();
 }
 
 void
@@ -895,6 +895,12 @@ if_netisr(void *unused)
                        KERNEL_UNLOCK();
                }
 #endif
+               if (n & (1 << NETISR_IP))
+                       ipintr();
+#ifdef INET6
+               if (n & (1 << NETISR_IPV6))
+                       ip6intr();
+#endif
 #if NPPP > 0
                if (n & (1 << NETISR_PPP)) {
                        KERNEL_LOCK();
@@ -3316,12 +3322,15 @@ unhandled_af(int af)
  * globals aren't ready to be accessed by multiple threads in
  * parallel.
  */
-int             nettaskqs = NET_TASKQ;
+int             nettaskqs;
 
 struct taskq *
 net_tq(unsigned int ifindex)
 {
        struct taskq *t = NULL;
+
+       if (nettaskqs == 0)
+               nettaskqs = min(NET_TASKQ, ncpus);
 
        t = nettqmp[ifindex % nettaskqs];
 
Index: net/if_ethersubr.c
===================================================================
RCS file: /data/mirror/openbsd/cvs/src/sys/net/if_ethersubr.c,v
retrieving revision 1.274
diff -u -p -r1.274 if_ethersubr.c
--- net/if_ethersubr.c  7 Mar 2021 06:02:32 -0000       1.274
+++ net/if_ethersubr.c  21 Apr 2021 12:54:49 -0000
@@ -245,7 +245,10 @@ ether_resolve(struct ifnet *ifp, struct 
                break;
 #ifdef INET6
        case AF_INET6:
+               KERNEL_LOCK();
+               /* XXXSMP there is a MP race in nd6_resolve() */
                error = nd6_resolve(ifp, rt, m, dst, eh->ether_dhost);
+               KERNEL_UNLOCK();
                if (error)
                        return (error);
                eh->ether_type = htons(ETHERTYPE_IPV6);
Index: net/ifq.c
===================================================================
RCS file: /data/mirror/openbsd/cvs/src/sys/net/ifq.c,v
retrieving revision 1.43
diff -u -p -r1.43 ifq.c
--- net/ifq.c   20 Feb 2021 04:37:26 -0000      1.43
+++ net/ifq.c   21 Apr 2021 13:12:44 -0000
@@ -243,7 +243,7 @@ void
 ifq_init(struct ifqueue *ifq, struct ifnet *ifp, unsigned int idx)
 {
        ifq->ifq_if = ifp;
-       ifq->ifq_softnet = net_tq(ifp->if_index); /* + idx */
+       ifq->ifq_softnet = net_tq(ifp->if_index + idx);
        ifq->ifq_softc = NULL;
 
        mtx_init(&ifq->ifq_mtx, IPL_NET);
@@ -617,7 +617,7 @@ void
 ifiq_init(struct ifiqueue *ifiq, struct ifnet *ifp, unsigned int idx)
 {
        ifiq->ifiq_if = ifp;
-       ifiq->ifiq_softnet = net_tq(ifp->if_index); /* + idx */
+       ifiq->ifiq_softnet = net_tq(ifp->if_index + idx);
        ifiq->ifiq_softc = NULL;
 
        mtx_init(&ifiq->ifiq_mtx, IPL_NET);
Index: net/netisr.h
===================================================================
RCS file: /data/mirror/openbsd/cvs/src/sys/net/netisr.h,v
retrieving revision 1.55
diff -u -p -r1.55 netisr.h
--- net/netisr.h        5 Jan 2021 20:43:36 -0000       1.55
+++ net/netisr.h        21 Apr 2021 12:52:08 -0000
@@ -41,8 +41,10 @@
  * interrupt used for scheduling the network code to calls
  * on the lowest level routine of each protocol.
  */
+#define        NETISR_IP       2               /* same as AF_INET */
 #define        NETISR_PFSYNC   5               /* for pfsync "immediate" tx */
 #define        NETISR_ARP      18              /* same as AF_LINK */
+#define        NETISR_IPV6     24              /* same as AF_INET6 */
 #define        NETISR_PPP      28              /* for PPP processing */
 #define        NETISR_BRIDGE   29              /* for bridge processing */
 #define        NETISR_SWITCH   31              /* for switch dataplane */
@@ -57,6 +59,8 @@ extern int    netisr;                 /* scheduling bits 
 extern struct task if_input_task_locked;
 
 void   arpintr(void);
+void   ipintr(void);
+void   ip6intr(void);
 void   pppintr(void);
 void   bridgeintr(void);
 void   switchintr(void);
Index: netinet/ip_input.c
===================================================================
RCS file: /data/mirror/openbsd/cvs/src/sys/netinet/ip_input.c,v
retrieving revision 1.356
diff -u -p -r1.356 ip_input.c
--- netinet/ip_input.c  30 Mar 2021 08:37:10 -0000      1.356
+++ netinet/ip_input.c  21 Apr 2021 12:52:08 -0000
@@ -131,6 +131,8 @@ const struct sysctl_bounded_args ipctl_v
        { IPCTL_ARPDOWN, &arpt_down, 0, INT_MAX },
 };
 
+struct niqueue ipintrq = NIQUEUE_INITIALIZER(IPQ_MAXLEN, NETISR_IP);
+
 struct pool ipqent_pool;
 struct pool ipq_pool;
 
@@ -144,6 +146,7 @@ static struct mbuf_queue    ipsendraw_mq;
 extern struct niqueue          arpinq;
 
 int    ip_ours(struct mbuf **, int *, int, int);
+int    ip_local(struct mbuf **, int *, int, int);
 int    ip_dooptions(struct mbuf *, struct ifnet *);
 int    in_ouraddr(struct mbuf *, struct ifnet *, struct rtentry **);
 
@@ -230,6 +233,43 @@ ip_init(void)
 }
 
 /*
+ * Enqueue packet for local delivery.  Queuing is used as a boundary
+ * between the network layer (input/forward path) running with shared
+ * NET_RLOCK_IN_SOFTNET() and the transport layer needing it exclusively.
+ */
+int
+ip_ours(struct mbuf **mp, int *offp, int nxt, int af)
+{
+       /* We are already in a IPv4/IPv6 local deliver loop. */
+       if (af != AF_UNSPEC)
+               return ip_local(mp, offp, nxt, af);
+
+       niq_enqueue(&ipintrq, *mp);
+       *mp = NULL;
+       return IPPROTO_DONE;
+}
+
+/*
+ * Dequeue and process locally delivered packets.
+ */
+void
+ipintr(void)
+{
+       struct mbuf *m;
+       int off, nxt;
+
+       while ((m = niq_dequeue(&ipintrq)) != NULL) {
+#ifdef DIAGNOSTIC
+               if ((m->m_flags & M_PKTHDR) == 0)
+                       panic("ipintr no HDR");
+#endif
+               off = 0;
+               nxt = ip_local(&m, &off, IPPROTO_IPV4, AF_UNSPEC);
+               KASSERT(nxt == IPPROTO_DONE);
+       }
+}
+
+/*
  * IPv4 input routine.
  *
  * Checksum and byte swap header.  Process options. Forward or deliver.
@@ -488,7 +528,7 @@ ip_input_if(struct mbuf **mp, int *offp,
  * If fragmented try to reassemble.  Pass to next level.
  */
 int
-ip_ours(struct mbuf **mp, int *offp, int nxt, int af)
+ip_local(struct mbuf **mp, int *offp, int nxt, int af)
 {
        struct mbuf *m = *mp;
        struct ip *ip = mtod(m, struct ip *);
@@ -1639,7 +1679,8 @@ ip_sysctl(int *name, u_int namelen, void
                    newlen));
 #endif
        case IPCTL_IFQUEUE:
-               return (EOPNOTSUPP);
+               return (sysctl_niq(name + 1, namelen - 1,
+                   oldp, oldlenp, newp, newlen, &ipintrq));
        case IPCTL_ARPQUEUE:
                return (sysctl_niq(name + 1, namelen - 1,
                    oldp, oldlenp, newp, newlen, &arpinq));
Index: netinet/ip_var.h
===================================================================
RCS file: /data/mirror/openbsd/cvs/src/sys/netinet/ip_var.h,v
retrieving revision 1.88
diff -u -p -r1.88 ip_var.h
--- netinet/ip_var.h    30 Mar 2021 08:37:11 -0000      1.88
+++ netinet/ip_var.h    21 Apr 2021 12:52:08 -0000
@@ -248,7 +248,6 @@ void         ip_stripoptions(struct mbuf *);
 int     ip_sysctl(int *, u_int, void *, size_t *, void *, size_t);
 void    ip_savecontrol(struct inpcb *, struct mbuf **, struct ip *,
            struct mbuf *);
-void    ipintr(void);
 int     ip_input_if(struct mbuf **, int *, int, int, struct ifnet *);
 int     ip_deliver(struct mbuf **, int *, int, int);
 void    ip_forward(struct mbuf *, struct ifnet *, struct rtentry *, int);
Index: netinet6/ip6_input.c
===================================================================
RCS file: /data/mirror/openbsd/cvs/src/sys/netinet6/ip6_input.c,v
retrieving revision 1.232
diff -u -p -r1.232 ip6_input.c
--- netinet6/ip6_input.c        10 Mar 2021 10:21:49 -0000      1.232
+++ netinet6/ip6_input.c        21 Apr 2021 12:52:08 -0000
@@ -115,11 +115,14 @@
 #include <netinet/ip_carp.h>
 #endif
 
+struct niqueue ip6intrq = NIQUEUE_INITIALIZER(IPQ_MAXLEN, NETISR_IPV6);
+
 struct cpumem *ip6counters;
 
 uint8_t ip6_soiikey[IP6_SOIIKEY_LEN];
 
 int ip6_ours(struct mbuf **, int *, int, int);
+int ip6_local(struct mbuf **, int *, int, int);
 int ip6_check_rh0hdr(struct mbuf *, int *);
 int ip6_hbhchcheck(struct mbuf *, int *, int *, int *);
 int ip6_hopopts_input(u_int32_t *, u_int32_t *, struct mbuf **, int *);
@@ -162,6 +165,43 @@ ip6_init(void)
        ip6counters = counters_alloc(ip6s_ncounters);
 }
 
+/*
+ * Enqueue packet for local delivery.  Queuing is used as a boundary
+ * between the network layer (input/forward path) running with shared
+ * NET_RLOCK_IN_SOFTNET() and the transport layer needing it exclusively.
+ */
+int
+ip6_ours(struct mbuf **mp, int *offp, int nxt, int af)
+{
+       /* We are already in a IPv4/IPv6 local deliver loop. */
+       if (af != AF_UNSPEC)
+               return ip6_local(mp, offp, nxt, af);
+
+       niq_enqueue(&ip6intrq, *mp);
+       *mp = NULL;
+       return IPPROTO_DONE;
+}
+
+/*
+ * Dequeue and process locally delivered packets.
+ */
+void
+ip6intr(void)
+{
+       struct mbuf *m;
+       int off, nxt;
+
+       while ((m = niq_dequeue(&ip6intrq)) != NULL) {
+#ifdef DIAGNOSTIC
+               if ((m->m_flags & M_PKTHDR) == 0)
+                       panic("ip6intr no HDR");
+#endif
+               off = 0;
+               nxt = ip6_local(&m, &off, IPPROTO_IPV6, AF_UNSPEC);
+               KASSERT(nxt == IPPROTO_DONE);
+       }
+}
+
 void
 ipv6_input(struct ifnet *ifp, struct mbuf *m)
 {
@@ -526,7 +566,7 @@ ip6_input_if(struct mbuf **mp, int *offp
 }
 
 int
-ip6_ours(struct mbuf **mp, int *offp, int nxt, int af)
+ip6_local(struct mbuf **mp, int *offp, int nxt, int af)
 {
        if (ip6_hbhchcheck(*mp, offp, &nxt, NULL))
                return IPPROTO_DONE;
@@ -1452,7 +1492,8 @@ ip6_sysctl(int *name, u_int namelen, voi
                NET_UNLOCK();
                return (error);
        case IPV6CTL_IFQUEUE:
-               return (EOPNOTSUPP);
+               return (sysctl_niq(name + 1, namelen - 1,
+                   oldp, oldlenp, newp, newlen, &ip6intrq));
        case IPV6CTL_SOIIKEY:
                return (ip6_sysctl_soiikey(oldp, oldlenp, newp, newlen));
        default:

Reply via email to