2 * Copyright (c) 2003, 2004 Matthew Dillon. All rights reserved.
3 * Copyright (c) 2003, 2004 Jeffrey M. Hsu. All rights reserved.
4 * Copyright (c) 2003 Jonathan Lemon. All rights reserved.
5 * Copyright (c) 2003, 2004 The DragonFly Project. All rights reserved.
7 * This code is derived from software contributed to The DragonFly Project
8 * by Jonathan Lemon, Jeffrey M. Hsu, and Matthew Dillon.
10 * Jonathan Lemon gave Jeffrey Hsu permission to combine his copyright
11 * into this one around July 8 2004.
13 * Redistribution and use in source and binary forms, with or without
14 * modification, are permitted provided that the following conditions
16 * 1. Redistributions of source code must retain the above copyright
17 * notice, this list of conditions and the following disclaimer.
18 * 2. Redistributions in binary form must reproduce the above copyright
19 * notice, this list of conditions and the following disclaimer in the
20 * documentation and/or other materials provided with the distribution.
21 * 3. Neither the name of The DragonFly Project nor the names of its
22 * contributors may be used to endorse or promote products derived
23 * from this software without specific, prior written permission.
25 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
26 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
27 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
28 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
29 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
30 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
31 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
32 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
33 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
34 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
35 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/kernel.h>
42 #include <sys/malloc.h>
43 #include <sys/msgport.h>
45 #include <sys/interrupt.h>
46 #include <sys/socket.h>
47 #include <sys/sysctl.h>
48 #include <sys/socketvar.h>
50 #include <net/if_var.h>
51 #include <net/netisr2.h>
52 #include <machine/cpufunc.h>
53 #include <machine/smp.h>
55 #include <sys/thread2.h>
56 #include <sys/msgport2.h>
57 #include <net/netmsg2.h>
58 #include <sys/mplock2.h>
60 #include <vm/vm_extern.h>
62 static void netmsg_service_port_init(lwkt_port_t);
63 static void netmsg_service_loop(void *arg);
64 static void netisr_hashfn0(struct mbuf **mp, int hoff);
65 static void netisr_nohashck(struct mbuf *, const struct pktinfo *);
67 struct netmsg_port_registration {
68 TAILQ_ENTRY(netmsg_port_registration) npr_entry;
72 struct netisr_rollup {
73 TAILQ_ENTRY(netisr_rollup) ru_entry;
79 struct netmsg_rollup {
80 struct netmsg_base base;
86 struct netmsg_barrier {
87 struct netmsg_base base;
88 volatile cpumask_t *br_cpumask;
89 volatile uint32_t br_done;
92 #define NETISR_BR_NOTDONE 0x1
93 #define NETISR_BR_WAITDONE 0x80000000
95 struct netisr_barrier {
96 struct netmsg_barrier *br_msgs[MAXCPU];
101 struct thread thread;
105 TAILQ_HEAD(, netisr_rollup) netrulist;
108 static struct netisr_data *netisr_data[MAXCPU];
110 static struct netisr netisrs[NETISR_MAX];
111 static TAILQ_HEAD(,netmsg_port_registration) netreglist;
113 /* Per-CPU thread to handle any protocol. */
114 struct thread *netisr_threads[MAXCPU];
116 lwkt_port netisr_afree_rport;
117 lwkt_port netisr_afree_free_so_rport;
118 lwkt_port netisr_adone_rport;
119 lwkt_port netisr_apanic_rport;
120 lwkt_port netisr_sync_port;
122 static int (*netmsg_fwd_port_fn)(lwkt_port_t, lwkt_msg_t);
124 SYSCTL_NODE(_net, OID_AUTO, netisr, CTLFLAG_RW, 0, "netisr");
126 static int netisr_rollup_limit = 32;
127 SYSCTL_INT(_net_netisr, OID_AUTO, rollup_limit, CTLFLAG_RW,
128 &netisr_rollup_limit, 0, "Message to process before rollup");
131 TUNABLE_INT("net.netisr.ncpus", &netisr_ncpus);
132 SYSCTL_INT(_net_netisr, OID_AUTO, ncpus, CTLFLAG_RD,
133 &netisr_ncpus, 0, "# of CPUs to handle network messages");
136 * netisr_afree_rport replymsg function, only used to handle async
137 * messages which the sender has abandoned to their fate.
140 netisr_autofree_reply(lwkt_port_t port, lwkt_msg_t msg)
142 kfree(msg, M_LWKTMSG);
146 netisr_autofree_free_so_reply(lwkt_port_t port, lwkt_msg_t msg)
148 sofree(((netmsg_t)msg)->base.nm_so);
149 kfree(msg, M_LWKTMSG);
153 * We need a custom putport function to handle the case where the
154 * message target is the current thread's message port. This case
155 * can occur when the TCP or UDP stack does a direct callback to NFS and NFS
156 * then turns around and executes a network operation synchronously.
158 * To prevent deadlocking, we must execute these self-referential messages
159 * synchronously, effectively turning the message into a glorified direct
160 * procedure call back into the protocol stack. The operation must be
161 * complete on return or we will deadlock, so panic if it isn't.
163 * However, the target function is under no obligation to immediately
164 * reply the message. It may forward it elsewhere.
167 netmsg_put_port(lwkt_port_t port, lwkt_msg_t lmsg)
169 netmsg_base_t nmsg = (void *)lmsg;
171 if ((lmsg->ms_flags & MSGF_SYNC) && port == &curthread->td_msgport) {
172 nmsg->nm_dispatch((netmsg_t)nmsg);
175 return(netmsg_fwd_port_fn(port, lmsg));
180 * UNIX DOMAIN sockets still have to run their uipc functions synchronously,
181 * because they depend on the user proc context for a number of things
182 * (like creds) which we have not yet incorporated into the message structure.
184 * However, we maintain or message/port abstraction. Having a special
185 * synchronous port which runs the commands synchronously gives us the
186 * ability to serialize operations in one place later on when we start
190 netmsg_sync_putport(lwkt_port_t port, lwkt_msg_t lmsg)
192 netmsg_base_t nmsg = (void *)lmsg;
194 KKASSERT((lmsg->ms_flags & MSGF_DONE) == 0);
196 lmsg->ms_target_port = port; /* required for abort */
197 nmsg->nm_dispatch((netmsg_t)nmsg);
206 if (netisr_ncpus <= 0) {
208 netisr_ncpus = ncpus2;
209 } else if (netisr_ncpus > ncpus) {
210 netisr_ncpus = ncpus;
212 if (netisr_ncpus > NETISR_CPUMAX)
213 netisr_ncpus = NETISR_CPUMAX;
215 TAILQ_INIT(&netreglist);
218 * Create default per-cpu threads for generic protocol handling.
220 for (i = 0; i < ncpus; ++i) {
221 struct netisr_data *nd;
223 nd = (void *)kmem_alloc3(&kernel_map, sizeof(*nd),
224 VM_SUBSYS_GD, KM_CPU(i));
225 memset(nd, 0, sizeof(*nd));
226 TAILQ_INIT(&nd->netrulist);
229 lwkt_create(netmsg_service_loop, NULL, &netisr_threads[i],
230 &nd->thread, TDF_NOSTART|TDF_FORCE_SPINPORT|TDF_FIXEDCPU,
232 netmsg_service_port_init(&netisr_threads[i]->td_msgport);
233 lwkt_schedule(netisr_threads[i]);
237 * The netisr_afree_rport is a special reply port which automatically
238 * frees the replied message. The netisr_adone_rport simply marks
239 * the message as being done. The netisr_apanic_rport panics if
240 * the message is replied to.
242 lwkt_initport_replyonly(&netisr_afree_rport, netisr_autofree_reply);
243 lwkt_initport_replyonly(&netisr_afree_free_so_rport,
244 netisr_autofree_free_so_reply);
245 lwkt_initport_replyonly_null(&netisr_adone_rport);
246 lwkt_initport_panic(&netisr_apanic_rport);
249 * The netisr_syncport is a special port which executes the message
250 * synchronously and waits for it if EASYNC is returned.
252 lwkt_initport_putonly(&netisr_sync_port, netmsg_sync_putport);
254 SYSINIT(netisr, SI_SUB_PRE_DRIVERS, SI_ORDER_FIRST, netisr_init, NULL);
257 * Finish initializing the message port for a netmsg service. This also
258 * registers the port for synchronous cleanup operations such as when an
259 * ifnet is being destroyed. There is no deregistration API yet.
262 netmsg_service_port_init(lwkt_port_t port)
264 struct netmsg_port_registration *reg;
267 * Override the putport function. Our custom function checks for
268 * self-references and executes such commands synchronously.
270 if (netmsg_fwd_port_fn == NULL)
271 netmsg_fwd_port_fn = port->mp_putport;
272 KKASSERT(netmsg_fwd_port_fn == port->mp_putport);
273 port->mp_putport = netmsg_put_port;
276 * Keep track of ports using the netmsg API so we can synchronize
277 * certain operations (such as freeing an ifnet structure) across all
280 reg = kmalloc(sizeof(*reg), M_TEMP, M_WAITOK|M_ZERO);
281 reg->npr_port = port;
282 TAILQ_INSERT_TAIL(&netreglist, reg, npr_entry);
286 * This function synchronizes the caller with all netmsg services. For
287 * example, if an interface is being removed we must make sure that all
288 * packets related to that interface complete processing before the structure
289 * can actually be freed. This sort of synchronization is an alternative to
290 * ref-counting the netif, removing the ref counting overhead in favor of
291 * placing additional overhead in the netif freeing sequence (where it is
295 netmsg_service_sync(void)
297 struct netmsg_port_registration *reg;
298 struct netmsg_base smsg;
300 netmsg_init(&smsg, NULL, &curthread->td_msgport, 0, netmsg_sync_handler);
302 TAILQ_FOREACH(reg, &netreglist, npr_entry) {
303 lwkt_domsg(reg->npr_port, &smsg.lmsg, 0);
308 * The netmsg function simply replies the message. API semantics require
309 * EASYNC to be returned if the netmsg function disposes of the message.
312 netmsg_sync_handler(netmsg_t msg)
314 lwkt_replymsg(&msg->lmsg, 0);
318 * Generic netmsg service loop. Some protocols may roll their own but all
319 * must do the basic command dispatch function call done here.
322 netmsg_service_loop(void *arg)
325 thread_t td = curthread;
327 struct netisr_data *nd = netisr_data[mycpuid];
329 td->td_type = TD_TYPE_NETISR;
331 while ((msg = lwkt_waitport(&td->td_msgport, 0))) {
332 struct netisr_rollup *ru;
335 * Run up to 512 pending netmsgs.
337 limit = netisr_rollup_limit;
339 KASSERT(msg->nm_dispatch != NULL,
340 ("netmsg_service isr %d badmsg",
341 msg->lmsg.u.ms_result));
343 * Don't match so_port, if the msg explicitly
344 * asks us to ignore its so_port.
346 if ((msg->lmsg.ms_flags & MSGF_IGNSOPORT) == 0 &&
348 msg->nm_so->so_port != &td->td_msgport) {
350 * Sockets undergoing connect or disconnect
351 * ops can change ports on us. Chase the
356 * This could be quite common for protocols
357 * which support asynchronous pru_connect,
358 * e.g. TCP, so kprintf socket port chasing
359 * could be too verbose for the console.
361 kprintf("%s: Warning, port changed so=%p\n",
362 __func__, msg->nm_so);
364 lwkt_forwardmsg(msg->nm_so->so_port,
368 * We are on the correct port, dispatch it.
371 nd->netlastfunc = msg->nm_dispatch;
373 msg->nm_dispatch((netmsg_t)msg);
377 } while ((msg = lwkt_getport(&td->td_msgport)) != NULL);
380 * Run all registered rollup functions for this cpu
381 * (e.g. tcp_willblock()).
383 TAILQ_FOREACH(ru, &nd->netrulist, ru_entry)
389 * Forward a packet to a netisr service function.
391 * If the packet has not been assigned to a protocol thread we call
392 * the port characterization function to assign it. The caller must
393 * clear M_HASH (or not have set it in the first place) if the caller
394 * wishes the packet to be recharacterized.
397 netisr_queue(int num, struct mbuf *m)
400 struct netmsg_packet *pmsg;
403 KASSERT((num > 0 && num <= NELEM(netisrs)),
404 ("Bad isr %d", num));
407 if (ni->ni_handler == NULL) {
408 kprintf("%s: Unregistered isr %d\n", __func__, num);
414 * Figure out which protocol thread to send to. This does not
415 * have to be perfect but performance will be really good if it
416 * is correct. Major protocol inputs such as ip_input() will
417 * re-characterize the packet as necessary.
419 if ((m->m_flags & M_HASH) == 0) {
420 ni->ni_hashfn(&m, 0);
423 if ((m->m_flags & M_HASH) == 0) {
424 kprintf("%s(%d): packet hash failed\n",
432 * Get the protocol port based on the packet hash, initialize
433 * the netmsg, and send it off.
435 port = netisr_hashport(m->m_pkthdr.hash);
436 pmsg = &m->m_hdr.mh_netmsg;
437 netmsg_init(&pmsg->base, NULL, &netisr_apanic_rport,
440 pmsg->base.lmsg.u.ms_result = num;
441 lwkt_sendmsg(port, &pmsg->base.lmsg);
447 * Run a netisr service function on the packet.
449 * The packet must have been correctly characterized!
452 netisr_handle(int num, struct mbuf *m)
455 struct netmsg_packet *pmsg;
459 * Get the protocol port based on the packet hash
461 KASSERT((m->m_flags & M_HASH), ("packet not characterized"));
462 port = netisr_hashport(m->m_pkthdr.hash);
463 KASSERT(&curthread->td_msgport == port, ("wrong msgport"));
465 KASSERT((num > 0 && num <= NELEM(netisrs)), ("bad isr %d", num));
467 if (ni->ni_handler == NULL) {
468 kprintf("%s: unregistered isr %d\n", __func__, num);
474 * Initialize the netmsg, and run the handler directly.
476 pmsg = &m->m_hdr.mh_netmsg;
477 netmsg_init(&pmsg->base, NULL, &netisr_apanic_rport,
480 pmsg->base.lmsg.u.ms_result = num;
481 ni->ni_handler((netmsg_t)&pmsg->base);
487 * Pre-characterization of a deeper portion of the packet for the
490 * The base of the ISR type (e.g. IP) that we want to characterize is
491 * at (hoff) relative to the beginning of the mbuf. This allows
492 * e.g. ether_characterize() to not have to adjust the m_data/m_len.
495 netisr_characterize(int num, struct mbuf **mp, int hoff)
506 if (num < 0 || num >= NETISR_MAX) {
507 if (num == NETISR_MAX) {
511 panic("Bad isr %d", num);
518 if (ni->ni_handler == NULL) {
519 kprintf("%s: Unregistered isr %d\n", __func__, num);
525 * Characterize the packet
527 if ((m->m_flags & M_HASH) == 0) {
528 ni->ni_hashfn(mp, hoff);
530 if (m && (m->m_flags & M_HASH) == 0) {
531 kprintf("%s(%d): packet hash failed\n",
538 netisr_register(int num, netisr_fn_t handler, netisr_hashfn_t hashfn)
542 KASSERT((num > 0 && num <= NELEM(netisrs)),
543 ("netisr_register: bad isr %d", num));
544 KKASSERT(handler != NULL);
547 hashfn = netisr_hashfn0;
551 ni->ni_handler = handler;
552 ni->ni_hashck = netisr_nohashck;
553 ni->ni_hashfn = hashfn;
554 netmsg_init(&ni->ni_netmsg, NULL, &netisr_adone_rport, 0, NULL);
558 netisr_register_hashcheck(int num, netisr_hashck_t hashck)
562 KASSERT((num > 0 && num <= NELEM(netisrs)),
563 ("netisr_register: bad isr %d", num));
566 ni->ni_hashck = hashck;
570 netisr_register_rollup_dispatch(netmsg_t nmsg)
572 struct netmsg_rollup *nm = (struct netmsg_rollup *)nmsg;
574 struct netisr_data *nd = netisr_data[cpuid];
575 struct netisr_rollup *new_ru, *ru;
577 new_ru = kmalloc(sizeof(*new_ru), M_TEMP, M_WAITOK|M_ZERO);
578 new_ru->ru_func = nm->func;
579 new_ru->ru_prio = nm->prio;
582 * Higher priority "rollup" appears first
584 TAILQ_FOREACH(ru, &nd->netrulist, ru_entry) {
585 if (ru->ru_prio < new_ru->ru_prio) {
586 TAILQ_INSERT_BEFORE(ru, new_ru, ru_entry);
590 TAILQ_INSERT_TAIL(&nd->netrulist, new_ru, ru_entry);
594 KKASSERT(nm->key != NULL);
595 new_ru->ru_key = nm->key;
597 netisr_forwardmsg_all(&nm->base, cpuid + 1);
600 struct netisr_rollup *
601 netisr_register_rollup(netisr_ru_t func, int prio)
603 struct netmsg_rollup nm;
605 netmsg_init(&nm.base, NULL, &curthread->td_msgport, MSGF_PRIORITY,
606 netisr_register_rollup_dispatch);
610 netisr_domsg_global(&nm.base);
612 KKASSERT(nm.key != NULL);
617 netisr_unregister_rollup_dispatch(netmsg_t nmsg)
619 struct netmsg_rollup *nm = (struct netmsg_rollup *)nmsg;
621 struct netisr_data *nd = netisr_data[cpuid];
622 struct netisr_rollup *ru;
624 TAILQ_FOREACH(ru, &nd->netrulist, ru_entry) {
625 if (ru->ru_key == nm->key)
629 panic("netisr: no rullup for %p", nm->key);
631 TAILQ_REMOVE(&nd->netrulist, ru, ru_entry);
634 netisr_forwardmsg_all(&nm->base, cpuid + 1);
638 netisr_unregister_rollup(struct netisr_rollup *key)
640 struct netmsg_rollup nm;
642 netmsg_init(&nm.base, NULL, &curthread->td_msgport, MSGF_PRIORITY,
643 netisr_unregister_rollup_dispatch);
645 netisr_domsg_global(&nm.base);
649 * Return a default protocol control message processing thread port
652 cpu0_ctlport(int cmd __unused, struct sockaddr *sa __unused,
653 void *extra __unused, int *cpuid)
656 return netisr_cpuport(*cpuid);
660 * This is a default netisr packet characterization function which
661 * sets M_HASH. If a netisr is registered with a NULL hashfn function
662 * this one is assigned.
664 * This function makes no attempt to validate the packet.
667 netisr_hashfn0(struct mbuf **mp, int hoff __unused)
674 * schednetisr() is used to call the netisr handler from the appropriate
675 * netisr thread for polling and other purposes.
677 * This function may be called from a hard interrupt or IPI and must be
678 * MP SAFE and non-blocking. We use a fixed per-cpu message instead of
679 * trying to allocate one. We must get ourselves onto the target cpu
680 * to safely check the MSGF_DONE bit on the message but since the message
681 * will be sent to that cpu anyway this does not add any extra work beyond
682 * what lwkt_sendmsg() would have already had to do to schedule the target
686 schednetisr_remote(void *data)
688 int num = (int)(intptr_t)data;
689 struct netisr *ni = &netisrs[num];
690 lwkt_port_t port = &netisr_threads[0]->td_msgport;
693 pmsg = &netisrs[num].ni_netmsg;
694 if (pmsg->lmsg.ms_flags & MSGF_DONE) {
695 netmsg_init(pmsg, NULL, &netisr_adone_rport, 0, ni->ni_handler);
696 pmsg->lmsg.u.ms_result = num;
697 lwkt_sendmsg(port, &pmsg->lmsg);
704 KASSERT((num > 0 && num <= NELEM(netisrs)),
705 ("schednetisr: bad isr %d", num));
706 KKASSERT(netisrs[num].ni_handler != NULL);
707 if (mycpu->gd_cpuid != 0) {
708 lwkt_send_ipiq(globaldata_find(0),
709 schednetisr_remote, (void *)(intptr_t)num);
712 schednetisr_remote((void *)(intptr_t)num);
718 netisr_barrier_dispatch(netmsg_t nmsg)
720 struct netmsg_barrier *msg = (struct netmsg_barrier *)nmsg;
722 ATOMIC_CPUMASK_NANDBIT(*msg->br_cpumask, mycpu->gd_cpuid);
723 if (CPUMASK_TESTZERO(*msg->br_cpumask))
724 wakeup(msg->br_cpumask);
727 uint32_t done = msg->br_done;
730 if ((done & NETISR_BR_NOTDONE) == 0)
733 tsleep_interlock(&msg->br_done, 0);
734 if (atomic_cmpset_int(&msg->br_done,
735 done, done | NETISR_BR_WAITDONE))
736 tsleep(&msg->br_done, PINTERLOCKED, "nbrdsp", 0);
739 lwkt_replymsg(&nmsg->lmsg, 0);
742 struct netisr_barrier *
743 netisr_barrier_create(void)
745 struct netisr_barrier *br;
747 br = kmalloc(sizeof(*br), M_LWKTMSG, M_WAITOK | M_ZERO);
752 netisr_barrier_set(struct netisr_barrier *br)
754 volatile cpumask_t other_cpumask;
758 KKASSERT(!br->br_isset);
760 other_cpumask = mycpu->gd_other_cpus;
761 CPUMASK_ANDMASK(other_cpumask, smp_active_mask);
764 for (i = 0; i < ncpus; ++i) {
765 struct netmsg_barrier *msg;
770 msg = kmalloc(sizeof(struct netmsg_barrier),
771 M_LWKTMSG, M_WAITOK);
774 * Don't use priority message here; mainly to keep
775 * it ordered w/ the previous data packets sent by
778 netmsg_init(&msg->base, NULL, &netisr_afree_rport, 0,
779 netisr_barrier_dispatch);
780 msg->br_cpumask = &other_cpumask;
781 msg->br_done = NETISR_BR_NOTDONE;
783 KKASSERT(br->br_msgs[i] == NULL);
784 br->br_msgs[i] = msg;
787 for (i = 0; i < ncpus; ++i) {
790 lwkt_sendmsg(netisr_cpuport(i), &br->br_msgs[i]->base.lmsg);
793 while (CPUMASK_TESTNZERO(other_cpumask)) {
794 tsleep_interlock(&other_cpumask, 0);
795 if (CPUMASK_TESTNZERO(other_cpumask))
796 tsleep(&other_cpumask, PINTERLOCKED, "nbrset", 0);
802 netisr_barrier_rem(struct netisr_barrier *br)
807 KKASSERT(br->br_isset);
810 for (i = 0; i < ncpus; ++i) {
811 struct netmsg_barrier *msg = br->br_msgs[i];
814 msg = br->br_msgs[i];
815 br->br_msgs[i] = NULL;
820 done = atomic_swap_int(&msg->br_done, 0);
821 if (done & NETISR_BR_WAITDONE)
822 wakeup(&msg->br_done);
828 netisr_nohashck(struct mbuf *m, const struct pktinfo *pi __unused)
830 m->m_flags &= ~M_HASH;
834 netisr_hashcheck(int num, struct mbuf *m, const struct pktinfo *pi)
838 if (num < 0 || num >= NETISR_MAX)
839 panic("Bad isr %d", num);
845 if (ni->ni_handler == NULL)
846 panic("Unregistered isr %d", num);
848 ni->ni_hashck(m, pi);