xref: /dragonfly/sys/net/netisr.c (revision 0720b42f)
1 /*
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.
6  *
7  * This code is derived from software contributed to The DragonFly Project
8  * by Jonathan Lemon, Jeffrey M. Hsu, and Matthew Dillon.
9  *
10  * Jonathan Lemon gave Jeffrey Hsu permission to combine his copyright
11  * into this one around July 8 2004.
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
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.
24  *
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
36  * SUCH DAMAGE.
37  */
38 
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>
44 #include <sys/proc.h>
45 #include <sys/interrupt.h>
46 #include <sys/socket.h>
47 #include <sys/sysctl.h>
48 #include <sys/socketvar.h>
49 #include <net/if.h>
50 #include <net/if_var.h>
51 #include <net/netisr2.h>
52 #include <machine/cpufunc.h>
53 #include <machine/smp.h>
54 
55 #include <sys/thread2.h>
56 #include <sys/msgport2.h>
57 #include <net/netmsg2.h>
58 #include <sys/mplock2.h>
59 
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 *);
63 
64 struct netmsg_port_registration {
65 	TAILQ_ENTRY(netmsg_port_registration) npr_entry;
66 	lwkt_port_t	npr_port;
67 };
68 
69 struct netmsg_rollup {
70 	TAILQ_ENTRY(netmsg_rollup) ru_entry;
71 	netisr_ru_t	ru_func;
72 	int		ru_prio;
73 };
74 
75 struct netmsg_barrier {
76 	struct netmsg_base	base;
77 	volatile cpumask_t	*br_cpumask;
78 	volatile uint32_t	br_done;
79 };
80 
81 #define NETISR_BR_NOTDONE	0x1
82 #define NETISR_BR_WAITDONE	0x80000000
83 
84 struct netisr_barrier {
85 	struct netmsg_barrier	*br_msgs[MAXCPU];
86 	int			br_isset;
87 };
88 
89 static struct netisr netisrs[NETISR_MAX];
90 static TAILQ_HEAD(,netmsg_port_registration) netreglist;
91 static TAILQ_HEAD(,netmsg_rollup) netrulist;
92 
93 /* Per-CPU thread to handle any protocol.  */
94 struct thread netisr_cpu[MAXCPU];
95 lwkt_port netisr_afree_rport;
96 lwkt_port netisr_afree_free_so_rport;
97 lwkt_port netisr_adone_rport;
98 lwkt_port netisr_apanic_rport;
99 lwkt_port netisr_sync_port;
100 
101 static int (*netmsg_fwd_port_fn)(lwkt_port_t, lwkt_msg_t);
102 
103 SYSCTL_NODE(_net, OID_AUTO, netisr, CTLFLAG_RW, 0, "netisr");
104 static int netisr_rollup_limit = 32;
105 SYSCTL_INT(_net_netisr, OID_AUTO, rollup_limit, CTLFLAG_RW,
106 	&netisr_rollup_limit, 0, "Message to process before rollup");
107 
108 
109 /*
110  * netisr_afree_rport replymsg function, only used to handle async
111  * messages which the sender has abandoned to their fate.
112  */
113 static void
114 netisr_autofree_reply(lwkt_port_t port, lwkt_msg_t msg)
115 {
116 	kfree(msg, M_LWKTMSG);
117 }
118 
119 static void
120 netisr_autofree_free_so_reply(lwkt_port_t port, lwkt_msg_t msg)
121 {
122 	sofree(((netmsg_t)msg)->base.nm_so);
123 	kfree(msg, M_LWKTMSG);
124 }
125 
126 /*
127  * We need a custom putport function to handle the case where the
128  * message target is the current thread's message port.  This case
129  * can occur when the TCP or UDP stack does a direct callback to NFS and NFS
130  * then turns around and executes a network operation synchronously.
131  *
132  * To prevent deadlocking, we must execute these self-referential messages
133  * synchronously, effectively turning the message into a glorified direct
134  * procedure call back into the protocol stack.  The operation must be
135  * complete on return or we will deadlock, so panic if it isn't.
136  *
137  * However, the target function is under no obligation to immediately
138  * reply the message.  It may forward it elsewhere.
139  */
140 static int
141 netmsg_put_port(lwkt_port_t port, lwkt_msg_t lmsg)
142 {
143 	netmsg_base_t nmsg = (void *)lmsg;
144 
145 	if ((lmsg->ms_flags & MSGF_SYNC) && port == &curthread->td_msgport) {
146 		nmsg->nm_dispatch((netmsg_t)nmsg);
147 		return(EASYNC);
148 	} else {
149 		return(netmsg_fwd_port_fn(port, lmsg));
150 	}
151 }
152 
153 /*
154  * UNIX DOMAIN sockets still have to run their uipc functions synchronously,
155  * because they depend on the user proc context for a number of things
156  * (like creds) which we have not yet incorporated into the message structure.
157  *
158  * However, we maintain or message/port abstraction.  Having a special
159  * synchronous port which runs the commands synchronously gives us the
160  * ability to serialize operations in one place later on when we start
161  * removing the BGL.
162  */
163 static int
164 netmsg_sync_putport(lwkt_port_t port, lwkt_msg_t lmsg)
165 {
166 	netmsg_base_t nmsg = (void *)lmsg;
167 
168 	KKASSERT((lmsg->ms_flags & MSGF_DONE) == 0);
169 
170 	lmsg->ms_target_port = port;	/* required for abort */
171 	nmsg->nm_dispatch((netmsg_t)nmsg);
172 	return(EASYNC);
173 }
174 
175 static void
176 netisr_init(void)
177 {
178 	int i;
179 
180 	TAILQ_INIT(&netreglist);
181 	TAILQ_INIT(&netrulist);
182 
183 	/*
184 	 * Create default per-cpu threads for generic protocol handling.
185 	 */
186 	for (i = 0; i < ncpus; ++i) {
187 		lwkt_create(netmsg_service_loop, NULL, NULL,
188 			    &netisr_cpu[i],
189 			    TDF_NOSTART|TDF_FORCE_SPINPORT|TDF_FIXEDCPU,
190 			    i, "netisr_cpu %d", i);
191 		netmsg_service_port_init(&netisr_cpu[i].td_msgport);
192 		lwkt_schedule(&netisr_cpu[i]);
193 	}
194 
195 	/*
196 	 * The netisr_afree_rport is a special reply port which automatically
197 	 * frees the replied message.  The netisr_adone_rport simply marks
198 	 * the message as being done.  The netisr_apanic_rport panics if
199 	 * the message is replied to.
200 	 */
201 	lwkt_initport_replyonly(&netisr_afree_rport, netisr_autofree_reply);
202 	lwkt_initport_replyonly(&netisr_afree_free_so_rport,
203 				netisr_autofree_free_so_reply);
204 	lwkt_initport_replyonly_null(&netisr_adone_rport);
205 	lwkt_initport_panic(&netisr_apanic_rport);
206 
207 	/*
208 	 * The netisr_syncport is a special port which executes the message
209 	 * synchronously and waits for it if EASYNC is returned.
210 	 */
211 	lwkt_initport_putonly(&netisr_sync_port, netmsg_sync_putport);
212 }
213 
214 SYSINIT(netisr, SI_SUB_PRE_DRIVERS, SI_ORDER_FIRST, netisr_init, NULL);
215 
216 /*
217  * Finish initializing the message port for a netmsg service.  This also
218  * registers the port for synchronous cleanup operations such as when an
219  * ifnet is being destroyed.  There is no deregistration API yet.
220  */
221 void
222 netmsg_service_port_init(lwkt_port_t port)
223 {
224 	struct netmsg_port_registration *reg;
225 
226 	/*
227 	 * Override the putport function.  Our custom function checks for
228 	 * self-references and executes such commands synchronously.
229 	 */
230 	if (netmsg_fwd_port_fn == NULL)
231 		netmsg_fwd_port_fn = port->mp_putport;
232 	KKASSERT(netmsg_fwd_port_fn == port->mp_putport);
233 	port->mp_putport = netmsg_put_port;
234 
235 	/*
236 	 * Keep track of ports using the netmsg API so we can synchronize
237 	 * certain operations (such as freeing an ifnet structure) across all
238 	 * consumers.
239 	 */
240 	reg = kmalloc(sizeof(*reg), M_TEMP, M_WAITOK|M_ZERO);
241 	reg->npr_port = port;
242 	TAILQ_INSERT_TAIL(&netreglist, reg, npr_entry);
243 }
244 
245 /*
246  * This function synchronizes the caller with all netmsg services.  For
247  * example, if an interface is being removed we must make sure that all
248  * packets related to that interface complete processing before the structure
249  * can actually be freed.  This sort of synchronization is an alternative to
250  * ref-counting the netif, removing the ref counting overhead in favor of
251  * placing additional overhead in the netif freeing sequence (where it is
252  * inconsequential).
253  */
254 void
255 netmsg_service_sync(void)
256 {
257 	struct netmsg_port_registration *reg;
258 	struct netmsg_base smsg;
259 
260 	netmsg_init(&smsg, NULL, &curthread->td_msgport, 0, netmsg_sync_handler);
261 
262 	TAILQ_FOREACH(reg, &netreglist, npr_entry) {
263 		lwkt_domsg(reg->npr_port, &smsg.lmsg, 0);
264 	}
265 }
266 
267 /*
268  * The netmsg function simply replies the message.  API semantics require
269  * EASYNC to be returned if the netmsg function disposes of the message.
270  */
271 void
272 netmsg_sync_handler(netmsg_t msg)
273 {
274 	lwkt_replymsg(&msg->lmsg, 0);
275 }
276 
277 /*
278  * Generic netmsg service loop.  Some protocols may roll their own but all
279  * must do the basic command dispatch function call done here.
280  */
281 static void
282 netmsg_service_loop(void *arg)
283 {
284 	struct netmsg_rollup *ru;
285 	netmsg_base_t msg;
286 	thread_t td = curthread;
287 	int limit;
288 
289 	td->td_type = TD_TYPE_NETISR;
290 
291 	while ((msg = lwkt_waitport(&td->td_msgport, 0))) {
292 		/*
293 		 * Run up to 512 pending netmsgs.
294 		 */
295 		limit = netisr_rollup_limit;
296 		do {
297 			KASSERT(msg->nm_dispatch != NULL,
298 				("netmsg_service isr %d badmsg",
299 				msg->lmsg.u.ms_result));
300 			/*
301 			 * Don't match so_port, if the msg explicitly
302 			 * asks us to ignore its so_port.
303 			 */
304 			if ((msg->lmsg.ms_flags & MSGF_IGNSOPORT) == 0 &&
305 			    msg->nm_so &&
306 			    msg->nm_so->so_port != &td->td_msgport) {
307 				/*
308 				 * Sockets undergoing connect or disconnect
309 				 * ops can change ports on us.  Chase the
310 				 * port.
311 				 */
312 #ifdef foo
313 				/*
314 				 * This could be quite common for protocols
315 				 * which support asynchronous pru_connect,
316 				 * e.g. TCP, so kprintf socket port chasing
317 				 * could be too verbose for the console.
318 				 */
319 				kprintf("%s: Warning, port changed so=%p\n",
320 					__func__, msg->nm_so);
321 #endif
322 				lwkt_forwardmsg(msg->nm_so->so_port,
323 						&msg->lmsg);
324 			} else {
325 				/*
326 				 * We are on the correct port, dispatch it.
327 				 */
328 				msg->nm_dispatch((netmsg_t)msg);
329 			}
330 			if (--limit == 0)
331 				break;
332 		} while ((msg = lwkt_getport(&td->td_msgport)) != NULL);
333 
334 		/*
335 		 * Run all registered rollup functions for this cpu
336 		 * (e.g. tcp_willblock()).
337 		 */
338 		TAILQ_FOREACH(ru, &netrulist, ru_entry)
339 			ru->ru_func();
340 	}
341 }
342 
343 /*
344  * Forward a packet to a netisr service function.
345  *
346  * If the packet has not been assigned to a protocol thread we call
347  * the port characterization function to assign it.  The caller must
348  * clear M_HASH (or not have set it in the first place) if the caller
349  * wishes the packet to be recharacterized.
350  */
351 int
352 netisr_queue(int num, struct mbuf *m)
353 {
354 	struct netisr *ni;
355 	struct netmsg_packet *pmsg;
356 	lwkt_port_t port;
357 
358 	KASSERT((num > 0 && num <= NELEM(netisrs)),
359 		("Bad isr %d", num));
360 
361 	ni = &netisrs[num];
362 	if (ni->ni_handler == NULL) {
363 		kprintf("%s: Unregistered isr %d\n", __func__, num);
364 		m_freem(m);
365 		return (EIO);
366 	}
367 
368 	/*
369 	 * Figure out which protocol thread to send to.  This does not
370 	 * have to be perfect but performance will be really good if it
371 	 * is correct.  Major protocol inputs such as ip_input() will
372 	 * re-characterize the packet as necessary.
373 	 */
374 	if ((m->m_flags & M_HASH) == 0) {
375 		ni->ni_hashfn(&m, 0);
376 		if (m == NULL)
377 			return (EIO);
378 		if ((m->m_flags & M_HASH) == 0) {
379 			kprintf("%s(%d): packet hash failed\n",
380 				__func__, num);
381 			m_freem(m);
382 			return (EIO);
383 		}
384 	}
385 
386 	/*
387 	 * Get the protocol port based on the packet hash, initialize
388 	 * the netmsg, and send it off.
389 	 */
390 	port = netisr_hashport(m->m_pkthdr.hash);
391 	pmsg = &m->m_hdr.mh_netmsg;
392 	netmsg_init(&pmsg->base, NULL, &netisr_apanic_rport,
393 		    0, ni->ni_handler);
394 	pmsg->nm_packet = m;
395 	pmsg->base.lmsg.u.ms_result = num;
396 	lwkt_sendmsg(port, &pmsg->base.lmsg);
397 
398 	return (0);
399 }
400 
401 /*
402  * Run a netisr service function on the packet.
403  *
404  * The packet must have been correctly characterized!
405  */
406 int
407 netisr_handle(int num, struct mbuf *m)
408 {
409 	struct netisr *ni;
410 	struct netmsg_packet *pmsg;
411 	lwkt_port_t port;
412 
413 	/*
414 	 * Get the protocol port based on the packet hash
415 	 */
416 	KASSERT((m->m_flags & M_HASH), ("packet not characterized"));
417 	port = netisr_hashport(m->m_pkthdr.hash);
418 	KASSERT(&curthread->td_msgport == port, ("wrong msgport"));
419 
420 	KASSERT((num > 0 && num <= NELEM(netisrs)), ("bad isr %d", num));
421 	ni = &netisrs[num];
422 	if (ni->ni_handler == NULL) {
423 		kprintf("%s: unregistered isr %d\n", __func__, num);
424 		m_freem(m);
425 		return EIO;
426 	}
427 
428 	/*
429 	 * Initialize the netmsg, and run the handler directly.
430 	 */
431 	pmsg = &m->m_hdr.mh_netmsg;
432 	netmsg_init(&pmsg->base, NULL, &netisr_apanic_rport,
433 		    0, ni->ni_handler);
434 	pmsg->nm_packet = m;
435 	pmsg->base.lmsg.u.ms_result = num;
436 	ni->ni_handler((netmsg_t)&pmsg->base);
437 
438 	return 0;
439 }
440 
441 /*
442  * Pre-characterization of a deeper portion of the packet for the
443  * requested isr.
444  *
445  * The base of the ISR type (e.g. IP) that we want to characterize is
446  * at (hoff) relative to the beginning of the mbuf.  This allows
447  * e.g. ether_characterize() to not have to adjust the m_data/m_len.
448  */
449 void
450 netisr_characterize(int num, struct mbuf **mp, int hoff)
451 {
452 	struct netisr *ni;
453 	struct mbuf *m;
454 
455 	/*
456 	 * Validation
457 	 */
458 	m = *mp;
459 	KKASSERT(m != NULL);
460 
461 	if (num < 0 || num >= NETISR_MAX) {
462 		if (num == NETISR_MAX) {
463 			m->m_flags |= M_HASH;
464 			m->m_pkthdr.hash = 0;
465 			return;
466 		}
467 		panic("Bad isr %d", num);
468 	}
469 
470 	/*
471 	 * Valid netisr?
472 	 */
473 	ni = &netisrs[num];
474 	if (ni->ni_handler == NULL) {
475 		kprintf("%s: Unregistered isr %d\n", __func__, num);
476 		m_freem(m);
477 		*mp = NULL;
478 	}
479 
480 	/*
481 	 * Characterize the packet
482 	 */
483 	if ((m->m_flags & M_HASH) == 0) {
484 		ni->ni_hashfn(mp, hoff);
485 		m = *mp;
486 		if (m && (m->m_flags & M_HASH) == 0) {
487 			kprintf("%s(%d): packet hash failed\n",
488 				__func__, num);
489 		}
490 	}
491 }
492 
493 void
494 netisr_register(int num, netisr_fn_t handler, netisr_hashfn_t hashfn)
495 {
496 	struct netisr *ni;
497 
498 	KASSERT((num > 0 && num <= NELEM(netisrs)),
499 		("netisr_register: bad isr %d", num));
500 	KKASSERT(handler != NULL);
501 
502 	if (hashfn == NULL)
503 		hashfn = netisr_hashfn0;
504 
505 	ni = &netisrs[num];
506 
507 	ni->ni_handler = handler;
508 	ni->ni_hashck = netisr_nohashck;
509 	ni->ni_hashfn = hashfn;
510 	netmsg_init(&ni->ni_netmsg, NULL, &netisr_adone_rport, 0, NULL);
511 }
512 
513 void
514 netisr_register_hashcheck(int num, netisr_hashck_t hashck)
515 {
516 	struct netisr *ni;
517 
518 	KASSERT((num > 0 && num <= NELEM(netisrs)),
519 		("netisr_register: bad isr %d", num));
520 
521 	ni = &netisrs[num];
522 	ni->ni_hashck = hashck;
523 }
524 
525 void
526 netisr_register_rollup(netisr_ru_t ru_func, int prio)
527 {
528 	struct netmsg_rollup *new_ru, *ru;
529 
530 	new_ru = kmalloc(sizeof(*new_ru), M_TEMP, M_WAITOK|M_ZERO);
531 	new_ru->ru_func = ru_func;
532 	new_ru->ru_prio = prio;
533 
534 	/*
535 	 * Higher priority "rollup" appears first
536 	 */
537 	TAILQ_FOREACH(ru, &netrulist, ru_entry) {
538 		if (ru->ru_prio < new_ru->ru_prio) {
539 			TAILQ_INSERT_BEFORE(ru, new_ru, ru_entry);
540 			return;
541 		}
542 	}
543 	TAILQ_INSERT_TAIL(&netrulist, new_ru, ru_entry);
544 }
545 
546 /*
547  * Return a default protocol control message processing thread port
548  */
549 lwkt_port_t
550 cpu0_ctlport(int cmd __unused, struct sockaddr *sa __unused,
551     void *extra __unused, int *cpuid)
552 {
553 	*cpuid = 0;
554 	return netisr_cpuport(*cpuid);
555 }
556 
557 /*
558  * This is a default netisr packet characterization function which
559  * sets M_HASH.  If a netisr is registered with a NULL hashfn function
560  * this one is assigned.
561  *
562  * This function makes no attempt to validate the packet.
563  */
564 static void
565 netisr_hashfn0(struct mbuf **mp, int hoff __unused)
566 {
567 	struct mbuf *m = *mp;
568 
569 	m->m_flags |= M_HASH;
570 	m->m_pkthdr.hash = 0;
571 }
572 
573 /*
574  * schednetisr() is used to call the netisr handler from the appropriate
575  * netisr thread for polling and other purposes.
576  *
577  * This function may be called from a hard interrupt or IPI and must be
578  * MP SAFE and non-blocking.  We use a fixed per-cpu message instead of
579  * trying to allocate one.  We must get ourselves onto the target cpu
580  * to safely check the MSGF_DONE bit on the message but since the message
581  * will be sent to that cpu anyway this does not add any extra work beyond
582  * what lwkt_sendmsg() would have already had to do to schedule the target
583  * thread.
584  */
585 static void
586 schednetisr_remote(void *data)
587 {
588 	int num = (int)(intptr_t)data;
589 	struct netisr *ni = &netisrs[num];
590 	lwkt_port_t port = &netisr_cpu[0].td_msgport;
591 	netmsg_base_t pmsg;
592 
593 	pmsg = &netisrs[num].ni_netmsg;
594 	if (pmsg->lmsg.ms_flags & MSGF_DONE) {
595 		netmsg_init(pmsg, NULL, &netisr_adone_rport, 0, ni->ni_handler);
596 		pmsg->lmsg.u.ms_result = num;
597 		lwkt_sendmsg(port, &pmsg->lmsg);
598 	}
599 }
600 
601 void
602 schednetisr(int num)
603 {
604 	KASSERT((num > 0 && num <= NELEM(netisrs)),
605 		("schednetisr: bad isr %d", num));
606 	KKASSERT(netisrs[num].ni_handler != NULL);
607 	if (mycpu->gd_cpuid != 0) {
608 		lwkt_send_ipiq(globaldata_find(0),
609 			       schednetisr_remote, (void *)(intptr_t)num);
610 	} else {
611 		crit_enter();
612 		schednetisr_remote((void *)(intptr_t)num);
613 		crit_exit();
614 	}
615 }
616 
617 static void
618 netisr_barrier_dispatch(netmsg_t nmsg)
619 {
620 	struct netmsg_barrier *msg = (struct netmsg_barrier *)nmsg;
621 
622 	ATOMIC_CPUMASK_NANDBIT(*msg->br_cpumask, mycpu->gd_cpuid);
623 	if (CPUMASK_TESTZERO(*msg->br_cpumask))
624 		wakeup(msg->br_cpumask);
625 
626 	for (;;) {
627 		uint32_t done = msg->br_done;
628 
629 		cpu_ccfence();
630 		if ((done & NETISR_BR_NOTDONE) == 0)
631 			break;
632 
633 		tsleep_interlock(&msg->br_done, 0);
634 		if (atomic_cmpset_int(&msg->br_done,
635 		    done, done | NETISR_BR_WAITDONE))
636 			tsleep(&msg->br_done, PINTERLOCKED, "nbrdsp", 0);
637 	}
638 
639 	lwkt_replymsg(&nmsg->lmsg, 0);
640 }
641 
642 struct netisr_barrier *
643 netisr_barrier_create(void)
644 {
645 	struct netisr_barrier *br;
646 
647 	br = kmalloc(sizeof(*br), M_LWKTMSG, M_WAITOK | M_ZERO);
648 	return br;
649 }
650 
651 void
652 netisr_barrier_set(struct netisr_barrier *br)
653 {
654 	volatile cpumask_t other_cpumask;
655 	int i, cur_cpuid;
656 
657 	ASSERT_IN_NETISR(0);
658 	KKASSERT(!br->br_isset);
659 
660 	other_cpumask = mycpu->gd_other_cpus;
661 	CPUMASK_ANDMASK(other_cpumask, smp_active_mask);
662 	cur_cpuid = mycpuid;
663 
664 	for (i = 0; i < ncpus; ++i) {
665 		struct netmsg_barrier *msg;
666 
667 		if (i == cur_cpuid)
668 			continue;
669 
670 		msg = kmalloc(sizeof(struct netmsg_barrier),
671 			      M_LWKTMSG, M_WAITOK);
672 
673 		/*
674 		 * Don't use priority message here; mainly to keep
675 		 * it ordered w/ the previous data packets sent by
676 		 * the caller.
677 		 */
678 		netmsg_init(&msg->base, NULL, &netisr_afree_rport, 0,
679 			    netisr_barrier_dispatch);
680 		msg->br_cpumask = &other_cpumask;
681 		msg->br_done = NETISR_BR_NOTDONE;
682 
683 		KKASSERT(br->br_msgs[i] == NULL);
684 		br->br_msgs[i] = msg;
685 	}
686 
687 	for (i = 0; i < ncpus; ++i) {
688 		if (i == cur_cpuid)
689 			continue;
690 		lwkt_sendmsg(netisr_cpuport(i), &br->br_msgs[i]->base.lmsg);
691 	}
692 
693 	while (CPUMASK_TESTNZERO(other_cpumask)) {
694 		tsleep_interlock(&other_cpumask, 0);
695 		if (CPUMASK_TESTNZERO(other_cpumask))
696 			tsleep(&other_cpumask, PINTERLOCKED, "nbrset", 0);
697 	}
698 	br->br_isset = 1;
699 }
700 
701 void
702 netisr_barrier_rem(struct netisr_barrier *br)
703 {
704 	int i, cur_cpuid;
705 
706 	ASSERT_IN_NETISR(0);
707 	KKASSERT(br->br_isset);
708 
709 	cur_cpuid = mycpuid;
710 	for (i = 0; i < ncpus; ++i) {
711 		struct netmsg_barrier *msg = br->br_msgs[i];
712 		uint32_t done;
713 
714 		msg = br->br_msgs[i];
715 		br->br_msgs[i] = NULL;
716 
717 		if (i == cur_cpuid)
718 			continue;
719 
720 		done = atomic_swap_int(&msg->br_done, 0);
721 		if (done & NETISR_BR_WAITDONE)
722 			wakeup(&msg->br_done);
723 	}
724 	br->br_isset = 0;
725 }
726 
727 static void
728 netisr_nohashck(struct mbuf *m, const struct pktinfo *pi __unused)
729 {
730 	m->m_flags &= ~M_HASH;
731 }
732 
733 void
734 netisr_hashcheck(int num, struct mbuf *m, const struct pktinfo *pi)
735 {
736 	struct netisr *ni;
737 
738 	if (num < 0 || num >= NETISR_MAX)
739 		panic("Bad isr %d", num);
740 
741 	/*
742 	 * Valid netisr?
743 	 */
744 	ni = &netisrs[num];
745 	if (ni->ni_handler == NULL)
746 		panic("Unregistered isr %d", num);
747 
748 	ni->ni_hashck(m, pi);
749 }
750