xref: /dflybsd-src/sys/net/netisr.c (revision 617f6bd6553bb7523bb0198e7aaa5d9c36fb6e05)
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 #include <vm/vm_extern.h>
61 
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 *);
66 
67 struct netmsg_port_registration {
68 	TAILQ_ENTRY(netmsg_port_registration) npr_entry;
69 	lwkt_port_t	npr_port;
70 };
71 
72 struct netisr_rollup {
73 	TAILQ_ENTRY(netisr_rollup) ru_entry;
74 	netisr_ru_t	ru_func;
75 	int		ru_prio;
76 	void		*ru_key;
77 };
78 
79 struct netmsg_rollup {
80 	struct netmsg_base	base;
81 	netisr_ru_t		func;
82 	int			prio;
83 	void			*key;
84 };
85 
86 struct netmsg_barrier {
87 	struct netmsg_base	base;
88 	volatile cpumask_t	*br_cpumask;
89 	volatile uint32_t	br_done;
90 };
91 
92 #define NETISR_BR_NOTDONE	0x1
93 #define NETISR_BR_WAITDONE	0x80000000
94 
95 struct netisr_barrier {
96 	struct netmsg_barrier	*br_msgs[MAXCPU];
97 	int			br_isset;
98 };
99 
100 struct netisr_data {
101 	struct thread		thread;
102 #ifdef INVARIANTS
103 	void			*netlastfunc;
104 #endif
105 	TAILQ_HEAD(, netisr_rollup) netrulist;
106 };
107 
108 static struct netisr_data	*netisr_data[MAXCPU];
109 
110 static struct netisr netisrs[NETISR_MAX];
111 static TAILQ_HEAD(,netmsg_port_registration) netreglist;
112 
113 /* Per-CPU thread to handle any protocol.  */
114 struct thread *netisr_threads[MAXCPU];
115 
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;
121 
122 static int (*netmsg_fwd_port_fn)(lwkt_port_t, lwkt_msg_t);
123 
124 SYSCTL_NODE(_net, OID_AUTO, netisr, CTLFLAG_RW, 0, "netisr");
125 
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");
129 
130 int netisr_ncpus;
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");
134 
135 /*
136  * netisr_afree_rport replymsg function, only used to handle async
137  * messages which the sender has abandoned to their fate.
138  */
139 static void
140 netisr_autofree_reply(lwkt_port_t port, lwkt_msg_t msg)
141 {
142 	kfree(msg, M_LWKTMSG);
143 }
144 
145 static void
146 netisr_autofree_free_so_reply(lwkt_port_t port, lwkt_msg_t msg)
147 {
148 	sofree(((netmsg_t)msg)->base.nm_so);
149 	kfree(msg, M_LWKTMSG);
150 }
151 
152 /*
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.
157  *
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.
162  *
163  * However, the target function is under no obligation to immediately
164  * reply the message.  It may forward it elsewhere.
165  */
166 static int
167 netmsg_put_port(lwkt_port_t port, lwkt_msg_t lmsg)
168 {
169 	netmsg_base_t nmsg = (void *)lmsg;
170 
171 	if ((lmsg->ms_flags & MSGF_SYNC) && port == &curthread->td_msgport) {
172 		nmsg->nm_dispatch((netmsg_t)nmsg);
173 		return(EASYNC);
174 	} else {
175 		return(netmsg_fwd_port_fn(port, lmsg));
176 	}
177 }
178 
179 /*
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.
183  *
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
187  * removing the BGL.
188  */
189 static int
190 netmsg_sync_putport(lwkt_port_t port, lwkt_msg_t lmsg)
191 {
192 	netmsg_base_t nmsg = (void *)lmsg;
193 
194 	KKASSERT((lmsg->ms_flags & MSGF_DONE) == 0);
195 
196 	lmsg->ms_target_port = port;	/* required for abort */
197 	nmsg->nm_dispatch((netmsg_t)nmsg);
198 	return(EASYNC);
199 }
200 
201 static void
202 netisr_init(void)
203 {
204 	int i;
205 
206 	if (netisr_ncpus <= 0) {
207 		/* Default. */
208 		netisr_ncpus = ncpus2;
209 	} else if (netisr_ncpus > ncpus) {
210 		netisr_ncpus = ncpus;
211 	}
212 	if (netisr_ncpus > NETISR_CPUMAX)
213 		netisr_ncpus = NETISR_CPUMAX;
214 
215 	TAILQ_INIT(&netreglist);
216 
217 	/*
218 	 * Create default per-cpu threads for generic protocol handling.
219 	 */
220 	for (i = 0; i < ncpus; ++i) {
221 		struct netisr_data *nd;
222 
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);
227 		netisr_data[i] = nd;
228 
229 		lwkt_create(netmsg_service_loop, NULL, &netisr_threads[i],
230 		    &nd->thread, TDF_NOSTART|TDF_FORCE_SPINPORT|TDF_FIXEDCPU,
231 		    i, "netisr %d", i);
232 		netmsg_service_port_init(&netisr_threads[i]->td_msgport);
233 		lwkt_schedule(netisr_threads[i]);
234 	}
235 
236 	/*
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.
241 	 */
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);
247 
248 	/*
249 	 * The netisr_syncport is a special port which executes the message
250 	 * synchronously and waits for it if EASYNC is returned.
251 	 */
252 	lwkt_initport_putonly(&netisr_sync_port, netmsg_sync_putport);
253 }
254 SYSINIT(netisr, SI_SUB_PRE_DRIVERS, SI_ORDER_FIRST, netisr_init, NULL);
255 
256 /*
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.
260  */
261 static void
262 netmsg_service_port_init(lwkt_port_t port)
263 {
264 	struct netmsg_port_registration *reg;
265 
266 	/*
267 	 * Override the putport function.  Our custom function checks for
268 	 * self-references and executes such commands synchronously.
269 	 */
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;
274 
275 	/*
276 	 * Keep track of ports using the netmsg API so we can synchronize
277 	 * certain operations (such as freeing an ifnet structure) across all
278 	 * consumers.
279 	 */
280 	reg = kmalloc(sizeof(*reg), M_TEMP, M_WAITOK|M_ZERO);
281 	reg->npr_port = port;
282 	TAILQ_INSERT_TAIL(&netreglist, reg, npr_entry);
283 }
284 
285 /*
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
292  * inconsequential).
293  */
294 void
295 netmsg_service_sync(void)
296 {
297 	struct netmsg_port_registration *reg;
298 	struct netmsg_base smsg;
299 
300 	netmsg_init(&smsg, NULL, &curthread->td_msgport, 0, netmsg_sync_handler);
301 
302 	TAILQ_FOREACH(reg, &netreglist, npr_entry) {
303 		lwkt_domsg(reg->npr_port, &smsg.lmsg, 0);
304 	}
305 }
306 
307 /*
308  * The netmsg function simply replies the message.  API semantics require
309  * EASYNC to be returned if the netmsg function disposes of the message.
310  */
311 void
312 netmsg_sync_handler(netmsg_t msg)
313 {
314 	lwkt_replymsg(&msg->lmsg, 0);
315 }
316 
317 /*
318  * Generic netmsg service loop.  Some protocols may roll their own but all
319  * must do the basic command dispatch function call done here.
320  */
321 static void
322 netmsg_service_loop(void *arg)
323 {
324 	netmsg_base_t msg;
325 	thread_t td = curthread;
326 	int limit;
327 	struct netisr_data *nd = netisr_data[mycpuid];
328 
329 	td->td_type = TD_TYPE_NETISR;
330 
331 	while ((msg = lwkt_waitport(&td->td_msgport, 0))) {
332 		struct netisr_rollup *ru;
333 
334 		/*
335 		 * Run up to 512 pending netmsgs.
336 		 */
337 		limit = netisr_rollup_limit;
338 		do {
339 			KASSERT(msg->nm_dispatch != NULL,
340 				("netmsg_service isr %d badmsg",
341 				msg->lmsg.u.ms_result));
342 			/*
343 			 * Don't match so_port, if the msg explicitly
344 			 * asks us to ignore its so_port.
345 			 */
346 			if ((msg->lmsg.ms_flags & MSGF_IGNSOPORT) == 0 &&
347 			    msg->nm_so &&
348 			    msg->nm_so->so_port != &td->td_msgport) {
349 				/*
350 				 * Sockets undergoing connect or disconnect
351 				 * ops can change ports on us.  Chase the
352 				 * port.
353 				 */
354 #ifdef foo
355 				/*
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.
360 				 */
361 				kprintf("%s: Warning, port changed so=%p\n",
362 					__func__, msg->nm_so);
363 #endif
364 				lwkt_forwardmsg(msg->nm_so->so_port,
365 						&msg->lmsg);
366 			} else {
367 				/*
368 				 * We are on the correct port, dispatch it.
369 				 */
370 #ifdef INVARIANTS
371 				nd->netlastfunc = msg->nm_dispatch;
372 #endif
373 				msg->nm_dispatch((netmsg_t)msg);
374 			}
375 			if (--limit == 0)
376 				break;
377 		} while ((msg = lwkt_getport(&td->td_msgport)) != NULL);
378 
379 		/*
380 		 * Run all registered rollup functions for this cpu
381 		 * (e.g. tcp_willblock()).
382 		 */
383 		TAILQ_FOREACH(ru, &nd->netrulist, ru_entry)
384 			ru->ru_func();
385 	}
386 }
387 
388 /*
389  * Forward a packet to a netisr service function.
390  *
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.
395  */
396 int
397 netisr_queue(int num, struct mbuf *m)
398 {
399 	struct netisr *ni;
400 	struct netmsg_packet *pmsg;
401 	lwkt_port_t port;
402 
403 	KASSERT((num > 0 && num <= NELEM(netisrs)),
404 		("Bad isr %d", num));
405 
406 	ni = &netisrs[num];
407 	if (ni->ni_handler == NULL) {
408 		kprintf("%s: Unregistered isr %d\n", __func__, num);
409 		m_freem(m);
410 		return (EIO);
411 	}
412 
413 	/*
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.
418 	 */
419 	if ((m->m_flags & M_HASH) == 0) {
420 		ni->ni_hashfn(&m, 0);
421 		if (m == NULL)
422 			return (EIO);
423 		if ((m->m_flags & M_HASH) == 0) {
424 			kprintf("%s(%d): packet hash failed\n",
425 				__func__, num);
426 			m_freem(m);
427 			return (EIO);
428 		}
429 	}
430 
431 	/*
432 	 * Get the protocol port based on the packet hash, initialize
433 	 * the netmsg, and send it off.
434 	 */
435 	port = netisr_hashport(m->m_pkthdr.hash);
436 	pmsg = &m->m_hdr.mh_netmsg;
437 	netmsg_init(&pmsg->base, NULL, &netisr_apanic_rport,
438 		    0, ni->ni_handler);
439 	pmsg->nm_packet = m;
440 	pmsg->base.lmsg.u.ms_result = num;
441 	lwkt_sendmsg(port, &pmsg->base.lmsg);
442 
443 	return (0);
444 }
445 
446 /*
447  * Run a netisr service function on the packet.
448  *
449  * The packet must have been correctly characterized!
450  */
451 int
452 netisr_handle(int num, struct mbuf *m)
453 {
454 	struct netisr *ni;
455 	struct netmsg_packet *pmsg;
456 	lwkt_port_t port;
457 
458 	/*
459 	 * Get the protocol port based on the packet hash
460 	 */
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"));
464 
465 	KASSERT((num > 0 && num <= NELEM(netisrs)), ("bad isr %d", num));
466 	ni = &netisrs[num];
467 	if (ni->ni_handler == NULL) {
468 		kprintf("%s: unregistered isr %d\n", __func__, num);
469 		m_freem(m);
470 		return EIO;
471 	}
472 
473 	/*
474 	 * Initialize the netmsg, and run the handler directly.
475 	 */
476 	pmsg = &m->m_hdr.mh_netmsg;
477 	netmsg_init(&pmsg->base, NULL, &netisr_apanic_rport,
478 		    0, ni->ni_handler);
479 	pmsg->nm_packet = m;
480 	pmsg->base.lmsg.u.ms_result = num;
481 	ni->ni_handler((netmsg_t)&pmsg->base);
482 
483 	return 0;
484 }
485 
486 /*
487  * Pre-characterization of a deeper portion of the packet for the
488  * requested isr.
489  *
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.
493  */
494 void
495 netisr_characterize(int num, struct mbuf **mp, int hoff)
496 {
497 	struct netisr *ni;
498 	struct mbuf *m;
499 
500 	/*
501 	 * Validation
502 	 */
503 	m = *mp;
504 	KKASSERT(m != NULL);
505 
506 	if (num < 0 || num >= NETISR_MAX) {
507 		if (num == NETISR_MAX) {
508 			m_sethash(m, 0);
509 			return;
510 		}
511 		panic("Bad isr %d", num);
512 	}
513 
514 	/*
515 	 * Valid netisr?
516 	 */
517 	ni = &netisrs[num];
518 	if (ni->ni_handler == NULL) {
519 		kprintf("%s: Unregistered isr %d\n", __func__, num);
520 		m_freem(m);
521 		*mp = NULL;
522 	}
523 
524 	/*
525 	 * Characterize the packet
526 	 */
527 	if ((m->m_flags & M_HASH) == 0) {
528 		ni->ni_hashfn(mp, hoff);
529 		m = *mp;
530 		if (m && (m->m_flags & M_HASH) == 0) {
531 			kprintf("%s(%d): packet hash failed\n",
532 				__func__, num);
533 		}
534 	}
535 }
536 
537 void
538 netisr_register(int num, netisr_fn_t handler, netisr_hashfn_t hashfn)
539 {
540 	struct netisr *ni;
541 
542 	KASSERT((num > 0 && num <= NELEM(netisrs)),
543 		("netisr_register: bad isr %d", num));
544 	KKASSERT(handler != NULL);
545 
546 	if (hashfn == NULL)
547 		hashfn = netisr_hashfn0;
548 
549 	ni = &netisrs[num];
550 
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);
555 }
556 
557 void
558 netisr_register_hashcheck(int num, netisr_hashck_t hashck)
559 {
560 	struct netisr *ni;
561 
562 	KASSERT((num > 0 && num <= NELEM(netisrs)),
563 		("netisr_register: bad isr %d", num));
564 
565 	ni = &netisrs[num];
566 	ni->ni_hashck = hashck;
567 }
568 
569 static void
570 netisr_register_rollup_dispatch(netmsg_t nmsg)
571 {
572 	struct netmsg_rollup *nm = (struct netmsg_rollup *)nmsg;
573 	int cpuid = mycpuid;
574 	struct netisr_data *nd = netisr_data[cpuid];
575 	struct netisr_rollup *new_ru, *ru;
576 
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;
580 
581 	/*
582 	 * Higher priority "rollup" appears first
583 	 */
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);
587 			goto done;
588 		}
589 	}
590 	TAILQ_INSERT_TAIL(&nd->netrulist, new_ru, ru_entry);
591 done:
592 	if (cpuid == 0)
593 		nm->key = new_ru;
594 	KKASSERT(nm->key != NULL);
595 	new_ru->ru_key = nm->key;
596 
597 	netisr_forwardmsg_all(&nm->base, cpuid + 1);
598 }
599 
600 struct netisr_rollup *
601 netisr_register_rollup(netisr_ru_t func, int prio)
602 {
603 	struct netmsg_rollup nm;
604 
605 	netmsg_init(&nm.base, NULL, &curthread->td_msgport, MSGF_PRIORITY,
606 	    netisr_register_rollup_dispatch);
607 	nm.func = func;
608 	nm.prio = prio;
609 	nm.key = NULL;
610 	netisr_domsg_global(&nm.base);
611 
612 	KKASSERT(nm.key != NULL);
613 	return (nm.key);
614 }
615 
616 static void
617 netisr_unregister_rollup_dispatch(netmsg_t nmsg)
618 {
619 	struct netmsg_rollup *nm = (struct netmsg_rollup *)nmsg;
620 	int cpuid = mycpuid;
621 	struct netisr_data *nd = netisr_data[cpuid];
622 	struct netisr_rollup *ru;
623 
624 	TAILQ_FOREACH(ru, &nd->netrulist, ru_entry) {
625 		if (ru->ru_key == nm->key)
626 			break;
627 	}
628 	if (ru == NULL)
629 		panic("netisr: no rullup for %p", nm->key);
630 
631 	TAILQ_REMOVE(&nd->netrulist, ru, ru_entry);
632 	kfree(ru, M_TEMP);
633 
634 	netisr_forwardmsg_all(&nm->base, cpuid + 1);
635 }
636 
637 void
638 netisr_unregister_rollup(struct netisr_rollup *key)
639 {
640 	struct netmsg_rollup nm;
641 
642 	netmsg_init(&nm.base, NULL, &curthread->td_msgport, MSGF_PRIORITY,
643 	    netisr_unregister_rollup_dispatch);
644 	nm.key = key;
645 	netisr_domsg_global(&nm.base);
646 }
647 
648 /*
649  * Return a default protocol control message processing thread port
650  */
651 lwkt_port_t
652 cpu0_ctlport(int cmd __unused, struct sockaddr *sa __unused,
653     void *extra __unused, int *cpuid)
654 {
655 	*cpuid = 0;
656 	return netisr_cpuport(*cpuid);
657 }
658 
659 /*
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.
663  *
664  * This function makes no attempt to validate the packet.
665  */
666 static void
667 netisr_hashfn0(struct mbuf **mp, int hoff __unused)
668 {
669 
670 	m_sethash(*mp, 0);
671 }
672 
673 /*
674  * schednetisr() is used to call the netisr handler from the appropriate
675  * netisr thread for polling and other purposes.
676  *
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
683  * thread.
684  */
685 static void
686 schednetisr_remote(void *data)
687 {
688 	int num = (int)(intptr_t)data;
689 	struct netisr *ni = &netisrs[num];
690 	lwkt_port_t port = &netisr_threads[0]->td_msgport;
691 	netmsg_base_t pmsg;
692 
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);
698 	}
699 }
700 
701 void
702 schednetisr(int num)
703 {
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);
710 	} else {
711 		crit_enter();
712 		schednetisr_remote((void *)(intptr_t)num);
713 		crit_exit();
714 	}
715 }
716 
717 static void
718 netisr_barrier_dispatch(netmsg_t nmsg)
719 {
720 	struct netmsg_barrier *msg = (struct netmsg_barrier *)nmsg;
721 
722 	ATOMIC_CPUMASK_NANDBIT(*msg->br_cpumask, mycpu->gd_cpuid);
723 	if (CPUMASK_TESTZERO(*msg->br_cpumask))
724 		wakeup(msg->br_cpumask);
725 
726 	for (;;) {
727 		uint32_t done = msg->br_done;
728 
729 		cpu_ccfence();
730 		if ((done & NETISR_BR_NOTDONE) == 0)
731 			break;
732 
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);
737 	}
738 
739 	lwkt_replymsg(&nmsg->lmsg, 0);
740 }
741 
742 struct netisr_barrier *
743 netisr_barrier_create(void)
744 {
745 	struct netisr_barrier *br;
746 
747 	br = kmalloc(sizeof(*br), M_LWKTMSG, M_WAITOK | M_ZERO);
748 	return br;
749 }
750 
751 void
752 netisr_barrier_set(struct netisr_barrier *br)
753 {
754 	volatile cpumask_t other_cpumask;
755 	int i, cur_cpuid;
756 
757 	ASSERT_NETISR0;
758 	KKASSERT(!br->br_isset);
759 
760 	other_cpumask = mycpu->gd_other_cpus;
761 	CPUMASK_ANDMASK(other_cpumask, smp_active_mask);
762 	cur_cpuid = mycpuid;
763 
764 	for (i = 0; i < ncpus; ++i) {
765 		struct netmsg_barrier *msg;
766 
767 		if (i == cur_cpuid)
768 			continue;
769 
770 		msg = kmalloc(sizeof(struct netmsg_barrier),
771 			      M_LWKTMSG, M_WAITOK);
772 
773 		/*
774 		 * Don't use priority message here; mainly to keep
775 		 * it ordered w/ the previous data packets sent by
776 		 * the caller.
777 		 */
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;
782 
783 		KKASSERT(br->br_msgs[i] == NULL);
784 		br->br_msgs[i] = msg;
785 	}
786 
787 	for (i = 0; i < ncpus; ++i) {
788 		if (i == cur_cpuid)
789 			continue;
790 		lwkt_sendmsg(netisr_cpuport(i), &br->br_msgs[i]->base.lmsg);
791 	}
792 
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);
797 	}
798 	br->br_isset = 1;
799 }
800 
801 void
802 netisr_barrier_rem(struct netisr_barrier *br)
803 {
804 	int i, cur_cpuid;
805 
806 	ASSERT_NETISR0;
807 	KKASSERT(br->br_isset);
808 
809 	cur_cpuid = mycpuid;
810 	for (i = 0; i < ncpus; ++i) {
811 		struct netmsg_barrier *msg = br->br_msgs[i];
812 		uint32_t done;
813 
814 		msg = br->br_msgs[i];
815 		br->br_msgs[i] = NULL;
816 
817 		if (i == cur_cpuid)
818 			continue;
819 
820 		done = atomic_swap_int(&msg->br_done, 0);
821 		if (done & NETISR_BR_WAITDONE)
822 			wakeup(&msg->br_done);
823 	}
824 	br->br_isset = 0;
825 }
826 
827 static void
828 netisr_nohashck(struct mbuf *m, const struct pktinfo *pi __unused)
829 {
830 	m->m_flags &= ~M_HASH;
831 }
832 
833 void
834 netisr_hashcheck(int num, struct mbuf *m, const struct pktinfo *pi)
835 {
836 	struct netisr *ni;
837 
838 	if (num < 0 || num >= NETISR_MAX)
839 		panic("Bad isr %d", num);
840 
841 	/*
842 	 * Valid netisr?
843 	 */
844 	ni = &netisrs[num];
845 	if (ni->ni_handler == NULL)
846 		panic("Unregistered isr %d", num);
847 
848 	ni->ni_hashck(m, pi);
849 }
850