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 static void netmsg_service_loop(void *arg);
61 static void netisr_hashfn0(struct mbuf **mp, int hoff);
62 static void netisr_nohashck(struct mbuf *, const struct pktinfo *);
64 struct netmsg_port_registration {
65 TAILQ_ENTRY(netmsg_port_registration) npr_entry;
69 struct netmsg_rollup {
70 TAILQ_ENTRY(netmsg_rollup) ru_entry;
75 struct netmsg_barrier {
76 struct netmsg_base base;
77 volatile cpumask_t *br_cpumask;
78 volatile uint32_t br_done;
81 #define NETISR_BR_NOTDONE 0x1
82 #define NETISR_BR_WAITDONE 0x80000000
84 struct netisr_barrier {
85 struct netmsg_barrier *br_msgs[MAXCPU];
89 void *netlastfunc[MAXCPU];
90 static struct netisr netisrs[NETISR_MAX];
91 static TAILQ_HEAD(,netmsg_port_registration) netreglist;
92 static TAILQ_HEAD(,netmsg_rollup) netrulist;
94 /* Per-CPU thread to handle any protocol. */
95 struct thread *netisr_cpu[MAXCPU];
96 lwkt_port netisr_afree_rport;
97 lwkt_port netisr_afree_free_so_rport;
98 lwkt_port netisr_adone_rport;
99 lwkt_port netisr_apanic_rport;
100 lwkt_port netisr_sync_port;
102 static int (*netmsg_fwd_port_fn)(lwkt_port_t, lwkt_msg_t);
104 SYSCTL_NODE(_net, OID_AUTO, netisr, CTLFLAG_RW, 0, "netisr");
105 static int netisr_rollup_limit = 32;
106 SYSCTL_INT(_net_netisr, OID_AUTO, rollup_limit, CTLFLAG_RW,
107 &netisr_rollup_limit, 0, "Message to process before rollup");
111 * netisr_afree_rport replymsg function, only used to handle async
112 * messages which the sender has abandoned to their fate.
115 netisr_autofree_reply(lwkt_port_t port, lwkt_msg_t msg)
117 kfree(msg, M_LWKTMSG);
121 netisr_autofree_free_so_reply(lwkt_port_t port, lwkt_msg_t msg)
123 sofree(((netmsg_t)msg)->base.nm_so);
124 kfree(msg, M_LWKTMSG);
128 * We need a custom putport function to handle the case where the
129 * message target is the current thread's message port. This case
130 * can occur when the TCP or UDP stack does a direct callback to NFS and NFS
131 * then turns around and executes a network operation synchronously.
133 * To prevent deadlocking, we must execute these self-referential messages
134 * synchronously, effectively turning the message into a glorified direct
135 * procedure call back into the protocol stack. The operation must be
136 * complete on return or we will deadlock, so panic if it isn't.
138 * However, the target function is under no obligation to immediately
139 * reply the message. It may forward it elsewhere.
142 netmsg_put_port(lwkt_port_t port, lwkt_msg_t lmsg)
144 netmsg_base_t nmsg = (void *)lmsg;
146 if ((lmsg->ms_flags & MSGF_SYNC) && port == &curthread->td_msgport) {
147 nmsg->nm_dispatch((netmsg_t)nmsg);
150 return(netmsg_fwd_port_fn(port, lmsg));
155 * UNIX DOMAIN sockets still have to run their uipc functions synchronously,
156 * because they depend on the user proc context for a number of things
157 * (like creds) which we have not yet incorporated into the message structure.
159 * However, we maintain or message/port abstraction. Having a special
160 * synchronous port which runs the commands synchronously gives us the
161 * ability to serialize operations in one place later on when we start
165 netmsg_sync_putport(lwkt_port_t port, lwkt_msg_t lmsg)
167 netmsg_base_t nmsg = (void *)lmsg;
169 KKASSERT((lmsg->ms_flags & MSGF_DONE) == 0);
171 lmsg->ms_target_port = port; /* required for abort */
172 nmsg->nm_dispatch((netmsg_t)nmsg);
181 TAILQ_INIT(&netreglist);
182 TAILQ_INIT(&netrulist);
185 * Create default per-cpu threads for generic protocol handling.
187 for (i = 0; i < ncpus; ++i) {
188 lwkt_create(netmsg_service_loop, NULL, &netisr_cpu[i],
190 TDF_NOSTART|TDF_FORCE_SPINPORT|TDF_FIXEDCPU,
191 i, "netisr_cpu %d", i);
192 netmsg_service_port_init(&netisr_cpu[i]->td_msgport);
193 lwkt_schedule(netisr_cpu[i]);
197 * The netisr_afree_rport is a special reply port which automatically
198 * frees the replied message. The netisr_adone_rport simply marks
199 * the message as being done. The netisr_apanic_rport panics if
200 * the message is replied to.
202 lwkt_initport_replyonly(&netisr_afree_rport, netisr_autofree_reply);
203 lwkt_initport_replyonly(&netisr_afree_free_so_rport,
204 netisr_autofree_free_so_reply);
205 lwkt_initport_replyonly_null(&netisr_adone_rport);
206 lwkt_initport_panic(&netisr_apanic_rport);
209 * The netisr_syncport is a special port which executes the message
210 * synchronously and waits for it if EASYNC is returned.
212 lwkt_initport_putonly(&netisr_sync_port, netmsg_sync_putport);
215 SYSINIT(netisr, SI_SUB_PRE_DRIVERS, SI_ORDER_FIRST, netisr_init, NULL);
218 * Finish initializing the message port for a netmsg service. This also
219 * registers the port for synchronous cleanup operations such as when an
220 * ifnet is being destroyed. There is no deregistration API yet.
223 netmsg_service_port_init(lwkt_port_t port)
225 struct netmsg_port_registration *reg;
228 * Override the putport function. Our custom function checks for
229 * self-references and executes such commands synchronously.
231 if (netmsg_fwd_port_fn == NULL)
232 netmsg_fwd_port_fn = port->mp_putport;
233 KKASSERT(netmsg_fwd_port_fn == port->mp_putport);
234 port->mp_putport = netmsg_put_port;
237 * Keep track of ports using the netmsg API so we can synchronize
238 * certain operations (such as freeing an ifnet structure) across all
241 reg = kmalloc(sizeof(*reg), M_TEMP, M_WAITOK|M_ZERO);
242 reg->npr_port = port;
243 TAILQ_INSERT_TAIL(&netreglist, reg, npr_entry);
247 * This function synchronizes the caller with all netmsg services. For
248 * example, if an interface is being removed we must make sure that all
249 * packets related to that interface complete processing before the structure
250 * can actually be freed. This sort of synchronization is an alternative to
251 * ref-counting the netif, removing the ref counting overhead in favor of
252 * placing additional overhead in the netif freeing sequence (where it is
256 netmsg_service_sync(void)
258 struct netmsg_port_registration *reg;
259 struct netmsg_base smsg;
261 netmsg_init(&smsg, NULL, &curthread->td_msgport, 0, netmsg_sync_handler);
263 TAILQ_FOREACH(reg, &netreglist, npr_entry) {
264 lwkt_domsg(reg->npr_port, &smsg.lmsg, 0);
269 * The netmsg function simply replies the message. API semantics require
270 * EASYNC to be returned if the netmsg function disposes of the message.
273 netmsg_sync_handler(netmsg_t msg)
275 lwkt_replymsg(&msg->lmsg, 0);
279 * Generic netmsg service loop. Some protocols may roll their own but all
280 * must do the basic command dispatch function call done here.
283 netmsg_service_loop(void *arg)
285 struct netmsg_rollup *ru;
287 thread_t td = curthread;
290 td->td_type = TD_TYPE_NETISR;
292 while ((msg = lwkt_waitport(&td->td_msgport, 0))) {
294 * Run up to 512 pending netmsgs.
296 limit = netisr_rollup_limit;
298 KASSERT(msg->nm_dispatch != NULL,
299 ("netmsg_service isr %d badmsg",
300 msg->lmsg.u.ms_result));
302 * Don't match so_port, if the msg explicitly
303 * asks us to ignore its so_port.
305 if ((msg->lmsg.ms_flags & MSGF_IGNSOPORT) == 0 &&
307 msg->nm_so->so_port != &td->td_msgport) {
309 * Sockets undergoing connect or disconnect
310 * ops can change ports on us. Chase the
315 * This could be quite common for protocols
316 * which support asynchronous pru_connect,
317 * e.g. TCP, so kprintf socket port chasing
318 * could be too verbose for the console.
320 kprintf("%s: Warning, port changed so=%p\n",
321 __func__, msg->nm_so);
323 lwkt_forwardmsg(msg->nm_so->so_port,
327 * We are on the correct port, dispatch it.
329 netlastfunc[mycpuid] = msg->nm_dispatch;
330 msg->nm_dispatch((netmsg_t)msg);
334 } while ((msg = lwkt_getport(&td->td_msgport)) != NULL);
337 * Run all registered rollup functions for this cpu
338 * (e.g. tcp_willblock()).
340 TAILQ_FOREACH(ru, &netrulist, ru_entry)
346 * Forward a packet to a netisr service function.
348 * If the packet has not been assigned to a protocol thread we call
349 * the port characterization function to assign it. The caller must
350 * clear M_HASH (or not have set it in the first place) if the caller
351 * wishes the packet to be recharacterized.
354 netisr_queue(int num, struct mbuf *m)
357 struct netmsg_packet *pmsg;
360 KASSERT((num > 0 && num <= NELEM(netisrs)),
361 ("Bad isr %d", num));
364 if (ni->ni_handler == NULL) {
365 kprintf("%s: Unregistered isr %d\n", __func__, num);
371 * Figure out which protocol thread to send to. This does not
372 * have to be perfect but performance will be really good if it
373 * is correct. Major protocol inputs such as ip_input() will
374 * re-characterize the packet as necessary.
376 if ((m->m_flags & M_HASH) == 0) {
377 ni->ni_hashfn(&m, 0);
380 if ((m->m_flags & M_HASH) == 0) {
381 kprintf("%s(%d): packet hash failed\n",
389 * Get the protocol port based on the packet hash, initialize
390 * the netmsg, and send it off.
392 port = netisr_hashport(m->m_pkthdr.hash);
393 pmsg = &m->m_hdr.mh_netmsg;
394 netmsg_init(&pmsg->base, NULL, &netisr_apanic_rport,
397 pmsg->base.lmsg.u.ms_result = num;
398 lwkt_sendmsg(port, &pmsg->base.lmsg);
404 * Run a netisr service function on the packet.
406 * The packet must have been correctly characterized!
409 netisr_handle(int num, struct mbuf *m)
412 struct netmsg_packet *pmsg;
416 * Get the protocol port based on the packet hash
418 KASSERT((m->m_flags & M_HASH), ("packet not characterized"));
419 port = netisr_hashport(m->m_pkthdr.hash);
420 KASSERT(&curthread->td_msgport == port, ("wrong msgport"));
422 KASSERT((num > 0 && num <= NELEM(netisrs)), ("bad isr %d", num));
424 if (ni->ni_handler == NULL) {
425 kprintf("%s: unregistered isr %d\n", __func__, num);
431 * Initialize the netmsg, and run the handler directly.
433 pmsg = &m->m_hdr.mh_netmsg;
434 netmsg_init(&pmsg->base, NULL, &netisr_apanic_rport,
437 pmsg->base.lmsg.u.ms_result = num;
438 ni->ni_handler((netmsg_t)&pmsg->base);
444 * Pre-characterization of a deeper portion of the packet for the
447 * The base of the ISR type (e.g. IP) that we want to characterize is
448 * at (hoff) relative to the beginning of the mbuf. This allows
449 * e.g. ether_characterize() to not have to adjust the m_data/m_len.
452 netisr_characterize(int num, struct mbuf **mp, int hoff)
463 if (num < 0 || num >= NETISR_MAX) {
464 if (num == NETISR_MAX) {
465 m->m_flags |= M_HASH;
466 m->m_pkthdr.hash = 0;
469 panic("Bad isr %d", num);
476 if (ni->ni_handler == NULL) {
477 kprintf("%s: Unregistered isr %d\n", __func__, num);
483 * Characterize the packet
485 if ((m->m_flags & M_HASH) == 0) {
486 ni->ni_hashfn(mp, hoff);
488 if (m && (m->m_flags & M_HASH) == 0) {
489 kprintf("%s(%d): packet hash failed\n",
496 netisr_register(int num, netisr_fn_t handler, netisr_hashfn_t hashfn)
500 KASSERT((num > 0 && num <= NELEM(netisrs)),
501 ("netisr_register: bad isr %d", num));
502 KKASSERT(handler != NULL);
505 hashfn = netisr_hashfn0;
509 ni->ni_handler = handler;
510 ni->ni_hashck = netisr_nohashck;
511 ni->ni_hashfn = hashfn;
512 netmsg_init(&ni->ni_netmsg, NULL, &netisr_adone_rport, 0, NULL);
516 netisr_register_hashcheck(int num, netisr_hashck_t hashck)
520 KASSERT((num > 0 && num <= NELEM(netisrs)),
521 ("netisr_register: bad isr %d", num));
524 ni->ni_hashck = hashck;
528 netisr_register_rollup(netisr_ru_t ru_func, int prio)
530 struct netmsg_rollup *new_ru, *ru;
532 new_ru = kmalloc(sizeof(*new_ru), M_TEMP, M_WAITOK|M_ZERO);
533 new_ru->ru_func = ru_func;
534 new_ru->ru_prio = prio;
537 * Higher priority "rollup" appears first
539 TAILQ_FOREACH(ru, &netrulist, ru_entry) {
540 if (ru->ru_prio < new_ru->ru_prio) {
541 TAILQ_INSERT_BEFORE(ru, new_ru, ru_entry);
545 TAILQ_INSERT_TAIL(&netrulist, new_ru, ru_entry);
549 * Return a default protocol control message processing thread port
552 cpu0_ctlport(int cmd __unused, struct sockaddr *sa __unused,
553 void *extra __unused, int *cpuid)
556 return netisr_cpuport(*cpuid);
560 * This is a default netisr packet characterization function which
561 * sets M_HASH. If a netisr is registered with a NULL hashfn function
562 * this one is assigned.
564 * This function makes no attempt to validate the packet.
567 netisr_hashfn0(struct mbuf **mp, int hoff __unused)
569 struct mbuf *m = *mp;
571 m->m_flags |= M_HASH;
572 m->m_pkthdr.hash = 0;
576 * schednetisr() is used to call the netisr handler from the appropriate
577 * netisr thread for polling and other purposes.
579 * This function may be called from a hard interrupt or IPI and must be
580 * MP SAFE and non-blocking. We use a fixed per-cpu message instead of
581 * trying to allocate one. We must get ourselves onto the target cpu
582 * to safely check the MSGF_DONE bit on the message but since the message
583 * will be sent to that cpu anyway this does not add any extra work beyond
584 * what lwkt_sendmsg() would have already had to do to schedule the target
588 schednetisr_remote(void *data)
590 int num = (int)(intptr_t)data;
591 struct netisr *ni = &netisrs[num];
592 lwkt_port_t port = &netisr_cpu[0]->td_msgport;
595 pmsg = &netisrs[num].ni_netmsg;
596 if (pmsg->lmsg.ms_flags & MSGF_DONE) {
597 netmsg_init(pmsg, NULL, &netisr_adone_rport, 0, ni->ni_handler);
598 pmsg->lmsg.u.ms_result = num;
599 lwkt_sendmsg(port, &pmsg->lmsg);
606 KASSERT((num > 0 && num <= NELEM(netisrs)),
607 ("schednetisr: bad isr %d", num));
608 KKASSERT(netisrs[num].ni_handler != NULL);
609 if (mycpu->gd_cpuid != 0) {
610 lwkt_send_ipiq(globaldata_find(0),
611 schednetisr_remote, (void *)(intptr_t)num);
614 schednetisr_remote((void *)(intptr_t)num);
620 netisr_barrier_dispatch(netmsg_t nmsg)
622 struct netmsg_barrier *msg = (struct netmsg_barrier *)nmsg;
624 ATOMIC_CPUMASK_NANDBIT(*msg->br_cpumask, mycpu->gd_cpuid);
625 if (CPUMASK_TESTZERO(*msg->br_cpumask))
626 wakeup(msg->br_cpumask);
629 uint32_t done = msg->br_done;
632 if ((done & NETISR_BR_NOTDONE) == 0)
635 tsleep_interlock(&msg->br_done, 0);
636 if (atomic_cmpset_int(&msg->br_done,
637 done, done | NETISR_BR_WAITDONE))
638 tsleep(&msg->br_done, PINTERLOCKED, "nbrdsp", 0);
641 lwkt_replymsg(&nmsg->lmsg, 0);
644 struct netisr_barrier *
645 netisr_barrier_create(void)
647 struct netisr_barrier *br;
649 br = kmalloc(sizeof(*br), M_LWKTMSG, M_WAITOK | M_ZERO);
654 netisr_barrier_set(struct netisr_barrier *br)
656 volatile cpumask_t other_cpumask;
660 KKASSERT(!br->br_isset);
662 other_cpumask = mycpu->gd_other_cpus;
663 CPUMASK_ANDMASK(other_cpumask, smp_active_mask);
666 for (i = 0; i < ncpus; ++i) {
667 struct netmsg_barrier *msg;
672 msg = kmalloc(sizeof(struct netmsg_barrier),
673 M_LWKTMSG, M_WAITOK);
676 * Don't use priority message here; mainly to keep
677 * it ordered w/ the previous data packets sent by
680 netmsg_init(&msg->base, NULL, &netisr_afree_rport, 0,
681 netisr_barrier_dispatch);
682 msg->br_cpumask = &other_cpumask;
683 msg->br_done = NETISR_BR_NOTDONE;
685 KKASSERT(br->br_msgs[i] == NULL);
686 br->br_msgs[i] = msg;
689 for (i = 0; i < ncpus; ++i) {
692 lwkt_sendmsg(netisr_cpuport(i), &br->br_msgs[i]->base.lmsg);
695 while (CPUMASK_TESTNZERO(other_cpumask)) {
696 tsleep_interlock(&other_cpumask, 0);
697 if (CPUMASK_TESTNZERO(other_cpumask))
698 tsleep(&other_cpumask, PINTERLOCKED, "nbrset", 0);
704 netisr_barrier_rem(struct netisr_barrier *br)
709 KKASSERT(br->br_isset);
712 for (i = 0; i < ncpus; ++i) {
713 struct netmsg_barrier *msg = br->br_msgs[i];
716 msg = br->br_msgs[i];
717 br->br_msgs[i] = NULL;
722 done = atomic_swap_int(&msg->br_done, 0);
723 if (done & NETISR_BR_WAITDONE)
724 wakeup(&msg->br_done);
730 netisr_nohashck(struct mbuf *m, const struct pktinfo *pi __unused)
732 m->m_flags &= ~M_HASH;
736 netisr_hashcheck(int num, struct mbuf *m, const struct pktinfo *pi)
740 if (num < 0 || num >= NETISR_MAX)
741 panic("Bad isr %d", num);
747 if (ni->ni_handler == NULL)
748 panic("Unregistered isr %d", num);
750 ni->ni_hashck(m, pi);