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/netisr.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_sync_func(netmsg_t msg);
61 static void netmsg_service_loop(void *arg);
62 static void cpu0_cpufn(struct mbuf **mp, int hoff);
64 struct netmsg_port_registration {
65 TAILQ_ENTRY(netmsg_port_registration) npr_entry;
69 struct netmsg_rollup {
70 TAILQ_ENTRY(netmsg_rollup) ru_entry;
74 struct netmsg_barrier {
75 struct netmsg_base base;
76 volatile cpumask_t *br_cpumask;
80 struct netisr_barrier {
81 struct netmsg_barrier *br_msgs[MAXCPU];
85 static struct netisr netisrs[NETISR_MAX];
86 static TAILQ_HEAD(,netmsg_port_registration) netreglist;
87 static TAILQ_HEAD(,netmsg_rollup) netrulist;
89 /* Per-CPU thread to handle any protocol. */
90 static struct thread netisr_cpu[MAXCPU];
91 lwkt_port netisr_afree_rport;
92 lwkt_port netisr_afree_free_so_rport;
93 lwkt_port netisr_adone_rport;
94 lwkt_port netisr_apanic_rport;
95 lwkt_port netisr_sync_port;
97 static int (*netmsg_fwd_port_fn)(lwkt_port_t, lwkt_msg_t);
99 SYSCTL_NODE(_net, OID_AUTO, netisr, CTLFLAG_RW, 0, "netisr");
102 * netisr_afree_rport replymsg function, only used to handle async
103 * messages which the sender has abandoned to their fate.
106 netisr_autofree_reply(lwkt_port_t port, lwkt_msg_t msg)
108 kfree(msg, M_LWKTMSG);
112 netisr_autofree_free_so_reply(lwkt_port_t port, lwkt_msg_t msg)
114 sofree(((netmsg_t)msg)->base.nm_so);
115 kfree(msg, M_LWKTMSG);
119 * We need a custom putport function to handle the case where the
120 * message target is the current thread's message port. This case
121 * can occur when the TCP or UDP stack does a direct callback to NFS and NFS
122 * then turns around and executes a network operation synchronously.
124 * To prevent deadlocking, we must execute these self-referential messages
125 * synchronously, effectively turning the message into a glorified direct
126 * procedure call back into the protocol stack. The operation must be
127 * complete on return or we will deadlock, so panic if it isn't.
129 * However, the target function is under no obligation to immediately
130 * reply the message. It may forward it elsewhere.
133 netmsg_put_port(lwkt_port_t port, lwkt_msg_t lmsg)
135 netmsg_base_t nmsg = (void *)lmsg;
137 if ((lmsg->ms_flags & MSGF_SYNC) && port == &curthread->td_msgport) {
138 nmsg->nm_dispatch((netmsg_t)nmsg);
141 return(netmsg_fwd_port_fn(port, lmsg));
146 * UNIX DOMAIN sockets still have to run their uipc functions synchronously,
147 * because they depend on the user proc context for a number of things
148 * (like creds) which we have not yet incorporated into the message structure.
150 * However, we maintain or message/port abstraction. Having a special
151 * synchronous port which runs the commands synchronously gives us the
152 * ability to serialize operations in one place later on when we start
156 netmsg_sync_putport(lwkt_port_t port, lwkt_msg_t lmsg)
158 netmsg_base_t nmsg = (void *)lmsg;
160 KKASSERT((lmsg->ms_flags & MSGF_DONE) == 0);
162 lmsg->ms_target_port = port; /* required for abort */
163 nmsg->nm_dispatch((netmsg_t)nmsg);
172 TAILQ_INIT(&netreglist);
173 TAILQ_INIT(&netrulist);
176 * Create default per-cpu threads for generic protocol handling.
178 for (i = 0; i < ncpus; ++i) {
179 lwkt_create(netmsg_service_loop, NULL, NULL,
180 &netisr_cpu[i], TDF_STOPREQ, i,
182 netmsg_service_port_init(&netisr_cpu[i].td_msgport);
183 lwkt_schedule(&netisr_cpu[i]);
187 * The netisr_afree_rport is a special reply port which automatically
188 * frees the replied message. The netisr_adone_rport simply marks
189 * the message as being done. The netisr_apanic_rport panics if
190 * the message is replied to.
192 lwkt_initport_replyonly(&netisr_afree_rport, netisr_autofree_reply);
193 lwkt_initport_replyonly(&netisr_afree_free_so_rport,
194 netisr_autofree_free_so_reply);
195 lwkt_initport_replyonly_null(&netisr_adone_rport);
196 lwkt_initport_panic(&netisr_apanic_rport);
199 * The netisr_syncport is a special port which executes the message
200 * synchronously and waits for it if EASYNC is returned.
202 lwkt_initport_putonly(&netisr_sync_port, netmsg_sync_putport);
205 SYSINIT(netisr, SI_SUB_PRE_DRIVERS, SI_ORDER_FIRST, netisr_init, NULL);
208 * Finish initializing the message port for a netmsg service. This also
209 * registers the port for synchronous cleanup operations such as when an
210 * ifnet is being destroyed. There is no deregistration API yet.
213 netmsg_service_port_init(lwkt_port_t port)
215 struct netmsg_port_registration *reg;
218 * Override the putport function. Our custom function checks for
219 * self-references and executes such commands synchronously.
221 if (netmsg_fwd_port_fn == NULL)
222 netmsg_fwd_port_fn = port->mp_putport;
223 KKASSERT(netmsg_fwd_port_fn == port->mp_putport);
224 port->mp_putport = netmsg_put_port;
227 * Keep track of ports using the netmsg API so we can synchronize
228 * certain operations (such as freeing an ifnet structure) across all
231 reg = kmalloc(sizeof(*reg), M_TEMP, M_WAITOK|M_ZERO);
232 reg->npr_port = port;
233 TAILQ_INSERT_TAIL(&netreglist, reg, npr_entry);
237 * This function synchronizes the caller with all netmsg services. For
238 * example, if an interface is being removed we must make sure that all
239 * packets related to that interface complete processing before the structure
240 * can actually be freed. This sort of synchronization is an alternative to
241 * ref-counting the netif, removing the ref counting overhead in favor of
242 * placing additional overhead in the netif freeing sequence (where it is
246 netmsg_service_sync(void)
248 struct netmsg_port_registration *reg;
249 struct netmsg_base smsg;
251 netmsg_init(&smsg, NULL, &curthread->td_msgport, 0, netmsg_sync_func);
253 TAILQ_FOREACH(reg, &netreglist, npr_entry) {
254 lwkt_domsg(reg->npr_port, &smsg.lmsg, 0);
259 * The netmsg function simply replies the message. API semantics require
260 * EASYNC to be returned if the netmsg function disposes of the message.
263 netmsg_sync_func(netmsg_t msg)
265 lwkt_replymsg(&msg->lmsg, 0);
269 * Generic netmsg service loop. Some protocols may roll their own but all
270 * must do the basic command dispatch function call done here.
273 netmsg_service_loop(void *arg)
275 struct netmsg_rollup *ru;
277 thread_t td = curthread;;
280 while ((msg = lwkt_waitport(&td->td_msgport, 0))) {
282 * Run up to 512 pending netmsgs.
286 KASSERT(msg->nm_dispatch != NULL,
287 ("netmsg_service isr %d badmsg\n",
288 msg->lmsg.u.ms_result));
290 msg->nm_so->so_port != &td->td_msgport) {
292 * Sockets undergoing connect or disconnect
293 * ops can change ports on us. Chase the
296 kprintf("netmsg_service_loop: Warning, "
297 "port changed so=%p\n", msg->nm_so);
298 lwkt_forwardmsg(msg->nm_so->so_port,
302 * We are on the correct port, dispatch it.
304 msg->nm_dispatch((netmsg_t)msg);
308 } while ((msg = lwkt_getport(&td->td_msgport)) != NULL);
311 * Run all registered rollup functions for this cpu
312 * (e.g. tcp_willblock()).
314 TAILQ_FOREACH(ru, &netrulist, ru_entry)
320 * Forward a packet to a netisr service function.
322 * If the packet has not been assigned to a protocol thread we call
323 * the port characterization function to assign it. The caller must
324 * clear M_HASH (or not have set it in the first place) if the caller
325 * wishes the packet to be recharacterized.
328 netisr_queue(int num, struct mbuf *m)
331 struct netmsg_packet *pmsg;
334 KASSERT((num > 0 && num <= NELEM(netisrs)),
335 ("Bad isr %d", num));
338 if (ni->ni_handler == NULL) {
339 kprintf("Unregistered isr %d\n", num);
345 * Figure out which protocol thread to send to. This does not
346 * have to be perfect but performance will be really good if it
347 * is correct. Major protocol inputs such as ip_input() will
348 * re-characterize the packet as necessary.
350 if ((m->m_flags & M_HASH) == 0) {
356 if ((m->m_flags & M_HASH) == 0) {
357 kprintf("netisr_queue(%d): packet hash failed\n", num);
364 * Get the protocol port based on the packet hash, initialize
365 * the netmsg, and send it off.
367 port = cpu_portfn(m->m_pkthdr.hash);
368 pmsg = &m->m_hdr.mh_netmsg;
369 netmsg_init(&pmsg->base, NULL, &netisr_apanic_rport,
372 pmsg->base.lmsg.u.ms_result = num;
373 lwkt_sendmsg(port, &pmsg->base.lmsg);
379 * Pre-characterization of a deeper portion of the packet for the
382 * The base of the ISR type (e.g. IP) that we want to characterize is
383 * at (hoff) relative to the beginning of the mbuf. This allows
384 * e.g. ether_input_chain() to not have to adjust the m_data/m_len.
387 netisr_characterize(int num, struct mbuf **mp, int hoff)
398 if (num < 0 || num >= NETISR_MAX) {
399 if (num == NETISR_MAX) {
400 m->m_flags |= M_HASH;
401 m->m_pkthdr.hash = 0;
404 panic("Bad isr %d", num);
411 if (ni->ni_handler == NULL) {
412 kprintf("Unregistered isr %d\n", num);
418 * Characterize the packet
420 if ((m->m_flags & M_HASH) == 0) {
421 ni->ni_cpufn(mp, hoff);
423 if (m && (m->m_flags & M_HASH) == 0)
424 kprintf("netisr_queue(%d): packet hash failed\n", num);
429 netisr_register(int num, netisr_fn_t handler, netisr_cpufn_t cpufn)
433 KASSERT((num > 0 && num <= NELEM(netisrs)),
434 ("netisr_register: bad isr %d", num));
435 KKASSERT(handler != NULL);
442 ni->ni_handler = handler;
443 ni->ni_cpufn = cpufn;
444 netmsg_init(&ni->ni_netmsg, NULL, &netisr_adone_rport, 0, NULL);
448 netisr_register_rollup(netisr_ru_t ru_func)
450 struct netmsg_rollup *ru;
452 ru = kmalloc(sizeof(*ru), M_TEMP, M_WAITOK|M_ZERO);
453 ru->ru_func = ru_func;
454 TAILQ_INSERT_TAIL(&netrulist, ru, ru_entry);
458 * Return the message port for the general protocol message servicing
459 * thread for a particular cpu.
464 KKASSERT(cpu >= 0 && cpu < ncpus);
465 return (&netisr_cpu[cpu].td_msgport);
469 * Return the current cpu's network protocol thread.
474 return(cpu_portfn(mycpu->gd_cpuid));
478 * Return a default protocol control message processing thread port
481 cpu0_ctlport(int cmd __unused, struct sockaddr *sa __unused,
482 void *extra __unused)
484 return (&netisr_cpu[0].td_msgport);
488 * This is a default netisr packet characterization function which
489 * sets M_HASH. If a netisr is registered with a NULL cpufn function
490 * this one is assigned.
492 * This function makes no attempt to validate the packet.
495 cpu0_cpufn(struct mbuf **mp, int hoff __unused)
497 struct mbuf *m = *mp;
499 m->m_flags |= M_HASH;
500 m->m_pkthdr.hash = 0;
504 * schednetisr() is used to call the netisr handler from the appropriate
505 * netisr thread for polling and other purposes.
507 * This function may be called from a hard interrupt or IPI and must be
508 * MP SAFE and non-blocking. We use a fixed per-cpu message instead of
509 * trying to allocate one. We must get ourselves onto the target cpu
510 * to safely check the MSGF_DONE bit on the message but since the message
511 * will be sent to that cpu anyway this does not add any extra work beyond
512 * what lwkt_sendmsg() would have already had to do to schedule the target
516 schednetisr_remote(void *data)
518 int num = (int)(intptr_t)data;
519 struct netisr *ni = &netisrs[num];
520 lwkt_port_t port = &netisr_cpu[0].td_msgport;
523 pmsg = &netisrs[num].ni_netmsg;
524 if (pmsg->lmsg.ms_flags & MSGF_DONE) {
525 netmsg_init(pmsg, NULL, &netisr_adone_rport, 0, ni->ni_handler);
526 pmsg->lmsg.u.ms_result = num;
527 lwkt_sendmsg(port, &pmsg->lmsg);
534 KASSERT((num > 0 && num <= NELEM(netisrs)),
535 ("schednetisr: bad isr %d", num));
536 KKASSERT(netisrs[num].ni_handler != NULL);
538 if (mycpu->gd_cpuid != 0) {
539 lwkt_send_ipiq(globaldata_find(0),
540 schednetisr_remote, (void *)(intptr_t)num);
543 schednetisr_remote((void *)(intptr_t)num);
548 schednetisr_remote((void *)(intptr_t)num);
556 netisr_barrier_dispatch(netmsg_t nmsg)
558 struct netmsg_barrier *msg = (struct netmsg_barrier *)nmsg;
560 atomic_clear_cpumask(msg->br_cpumask, mycpu->gd_cpumask);
561 if (*msg->br_cpumask == 0)
562 wakeup(msg->br_cpumask);
564 tsleep_interlock(&msg->br_done, 0);
566 tsleep(&msg->br_done, PINTERLOCKED, "nbrdsp", 0);
568 lwkt_replymsg(&nmsg->lmsg, 0);
573 struct netisr_barrier *
574 netisr_barrier_create(void)
576 struct netisr_barrier *br;
578 br = kmalloc(sizeof(*br), M_LWKTMSG, M_WAITOK | M_ZERO);
583 netisr_barrier_set(struct netisr_barrier *br)
586 volatile cpumask_t other_cpumask;
589 KKASSERT(&curthread->td_msgport == cpu_portfn(0));
590 KKASSERT(!br->br_isset);
592 other_cpumask = mycpu->gd_other_cpus & smp_active_mask;
595 for (i = 0; i < ncpus; ++i) {
596 struct netmsg_barrier *msg;
601 msg = kmalloc(sizeof(struct netmsg_barrier),
602 M_LWKTMSG, M_WAITOK);
603 netmsg_init(&msg->base, NULL, &netisr_afree_rport,
604 MSGF_PRIORITY, netisr_barrier_dispatch);
605 msg->br_cpumask = &other_cpumask;
608 KKASSERT(br->br_msgs[i] == NULL);
609 br->br_msgs[i] = msg;
612 for (i = 0; i < ncpus; ++i) {
615 lwkt_sendmsg(cpu_portfn(i), &br->br_msgs[i]->base.lmsg);
618 while (other_cpumask != 0) {
619 tsleep_interlock(&other_cpumask, 0);
620 if (other_cpumask != 0)
621 tsleep(&other_cpumask, PINTERLOCKED, "nbrset", 0);
628 netisr_barrier_rem(struct netisr_barrier *br)
633 KKASSERT(&curthread->td_msgport == cpu_portfn(0));
634 KKASSERT(br->br_isset);
637 for (i = 0; i < ncpus; ++i) {
638 struct netmsg_barrier *msg = br->br_msgs[i];
640 msg = br->br_msgs[i];
641 br->br_msgs[i] = NULL;
647 wakeup(&msg->br_done);