xref: /dflybsd-src/sys/net/netisr.c (revision 681e352e6cf7f30481ae49f284ab7dab4168ba5a)
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/netisr.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 cpu0_cpufn(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 static struct thread netisr_cpu[MAXCPU];
95 static struct thread netisr_ded[NETISR_MAX];
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;
101 
102 static int (*netmsg_fwd_port_fn)(lwkt_port_t, lwkt_msg_t);
103 
104 SYSCTL_NODE(_net, OID_AUTO, netisr, CTLFLAG_RW, 0, "netisr");
105 
106 /*
107  * netisr_afree_rport replymsg function, only used to handle async
108  * messages which the sender has abandoned to their fate.
109  */
110 static void
111 netisr_autofree_reply(lwkt_port_t port, lwkt_msg_t msg)
112 {
113 	kfree(msg, M_LWKTMSG);
114 }
115 
116 static void
117 netisr_autofree_free_so_reply(lwkt_port_t port, lwkt_msg_t msg)
118 {
119 	sofree(((netmsg_t)msg)->base.nm_so);
120 	kfree(msg, M_LWKTMSG);
121 }
122 
123 /*
124  * We need a custom putport function to handle the case where the
125  * message target is the current thread's message port.  This case
126  * can occur when the TCP or UDP stack does a direct callback to NFS and NFS
127  * then turns around and executes a network operation synchronously.
128  *
129  * To prevent deadlocking, we must execute these self-referential messages
130  * synchronously, effectively turning the message into a glorified direct
131  * procedure call back into the protocol stack.  The operation must be
132  * complete on return or we will deadlock, so panic if it isn't.
133  *
134  * However, the target function is under no obligation to immediately
135  * reply the message.  It may forward it elsewhere.
136  */
137 static int
138 netmsg_put_port(lwkt_port_t port, lwkt_msg_t lmsg)
139 {
140 	netmsg_base_t nmsg = (void *)lmsg;
141 
142 	if ((lmsg->ms_flags & MSGF_SYNC) && port == &curthread->td_msgport) {
143 		nmsg->nm_dispatch((netmsg_t)nmsg);
144 		return(EASYNC);
145 	} else {
146 		return(netmsg_fwd_port_fn(port, lmsg));
147 	}
148 }
149 
150 /*
151  * UNIX DOMAIN sockets still have to run their uipc functions synchronously,
152  * because they depend on the user proc context for a number of things
153  * (like creds) which we have not yet incorporated into the message structure.
154  *
155  * However, we maintain or message/port abstraction.  Having a special
156  * synchronous port which runs the commands synchronously gives us the
157  * ability to serialize operations in one place later on when we start
158  * removing the BGL.
159  */
160 static int
161 netmsg_sync_putport(lwkt_port_t port, lwkt_msg_t lmsg)
162 {
163 	netmsg_base_t nmsg = (void *)lmsg;
164 
165 	KKASSERT((lmsg->ms_flags & MSGF_DONE) == 0);
166 
167 	lmsg->ms_target_port = port;	/* required for abort */
168 	nmsg->nm_dispatch((netmsg_t)nmsg);
169 	return(EASYNC);
170 }
171 
172 static void
173 netisr_init(void)
174 {
175 	int i;
176 
177 	TAILQ_INIT(&netreglist);
178 	TAILQ_INIT(&netrulist);
179 
180 	/*
181 	 * Create default per-cpu threads for generic protocol handling.
182 	 */
183 	for (i = 0; i < ncpus; ++i) {
184 		lwkt_create(netmsg_service_loop, NULL, NULL,
185 			    &netisr_cpu[i], TDF_NOSTART|TDF_FORCE_SPINPORT,
186 			    i, "netisr_cpu %d", i);
187 		netmsg_service_port_init(&netisr_cpu[i].td_msgport);
188 		lwkt_schedule(&netisr_cpu[i]);
189 	}
190 
191 	/*
192 	 * The netisr_afree_rport is a special reply port which automatically
193 	 * frees the replied message.  The netisr_adone_rport simply marks
194 	 * the message as being done.  The netisr_apanic_rport panics if
195 	 * the message is replied to.
196 	 */
197 	lwkt_initport_replyonly(&netisr_afree_rport, netisr_autofree_reply);
198 	lwkt_initport_replyonly(&netisr_afree_free_so_rport,
199 				netisr_autofree_free_so_reply);
200 	lwkt_initport_replyonly_null(&netisr_adone_rport);
201 	lwkt_initport_panic(&netisr_apanic_rport);
202 
203 	/*
204 	 * The netisr_syncport is a special port which executes the message
205 	 * synchronously and waits for it if EASYNC is returned.
206 	 */
207 	lwkt_initport_putonly(&netisr_sync_port, netmsg_sync_putport);
208 }
209 
210 SYSINIT(netisr, SI_SUB_PRE_DRIVERS, SI_ORDER_FIRST, netisr_init, NULL);
211 
212 /*
213  * Finish initializing the message port for a netmsg service.  This also
214  * registers the port for synchronous cleanup operations such as when an
215  * ifnet is being destroyed.  There is no deregistration API yet.
216  */
217 void
218 netmsg_service_port_init(lwkt_port_t port)
219 {
220 	struct netmsg_port_registration *reg;
221 
222 	/*
223 	 * Override the putport function.  Our custom function checks for
224 	 * self-references and executes such commands synchronously.
225 	 */
226 	if (netmsg_fwd_port_fn == NULL)
227 		netmsg_fwd_port_fn = port->mp_putport;
228 	KKASSERT(netmsg_fwd_port_fn == port->mp_putport);
229 	port->mp_putport = netmsg_put_port;
230 
231 	/*
232 	 * Keep track of ports using the netmsg API so we can synchronize
233 	 * certain operations (such as freeing an ifnet structure) across all
234 	 * consumers.
235 	 */
236 	reg = kmalloc(sizeof(*reg), M_TEMP, M_WAITOK|M_ZERO);
237 	reg->npr_port = port;
238 	TAILQ_INSERT_TAIL(&netreglist, reg, npr_entry);
239 }
240 
241 /*
242  * This function synchronizes the caller with all netmsg services.  For
243  * example, if an interface is being removed we must make sure that all
244  * packets related to that interface complete processing before the structure
245  * can actually be freed.  This sort of synchronization is an alternative to
246  * ref-counting the netif, removing the ref counting overhead in favor of
247  * placing additional overhead in the netif freeing sequence (where it is
248  * inconsequential).
249  */
250 void
251 netmsg_service_sync(void)
252 {
253 	struct netmsg_port_registration *reg;
254 	struct netmsg_base smsg;
255 
256 	netmsg_init(&smsg, NULL, &curthread->td_msgport, 0, netmsg_sync_handler);
257 
258 	TAILQ_FOREACH(reg, &netreglist, npr_entry) {
259 		lwkt_domsg(reg->npr_port, &smsg.lmsg, 0);
260 	}
261 }
262 
263 /*
264  * The netmsg function simply replies the message.  API semantics require
265  * EASYNC to be returned if the netmsg function disposes of the message.
266  */
267 void
268 netmsg_sync_handler(netmsg_t msg)
269 {
270 	lwkt_replymsg(&msg->lmsg, 0);
271 }
272 
273 /*
274  * Generic netmsg service loop.  Some protocols may roll their own but all
275  * must do the basic command dispatch function call done here.
276  */
277 static void
278 netmsg_service_loop(void *arg)
279 {
280 	struct netmsg_rollup *ru;
281 	netmsg_base_t msg;
282 	thread_t td = curthread;
283 	int limit;
284 
285 	td->td_type = TD_TYPE_NETISR;
286 
287 	while ((msg = lwkt_waitport(&td->td_msgport, 0))) {
288 		/*
289 		 * Run up to 512 pending netmsgs.
290 		 */
291 		limit = 512;
292 		do {
293 			KASSERT(msg->nm_dispatch != NULL,
294 				("netmsg_service isr %d badmsg",
295 				msg->lmsg.u.ms_result));
296 			if (msg->nm_so &&
297 			    msg->nm_so->so_port != &td->td_msgport) {
298 				/*
299 				 * Sockets undergoing connect or disconnect
300 				 * ops can change ports on us.  Chase the
301 				 * port.
302 				 */
303 				kprintf("netmsg_service_loop: Warning, "
304 					"port changed so=%p\n", msg->nm_so);
305 				lwkt_forwardmsg(msg->nm_so->so_port,
306 						&msg->lmsg);
307 			} else {
308 				/*
309 				 * We are on the correct port, dispatch it.
310 				 */
311 				msg->nm_dispatch((netmsg_t)msg);
312 			}
313 			if (--limit == 0)
314 				break;
315 		} while ((msg = lwkt_getport(&td->td_msgport)) != NULL);
316 
317 		/*
318 		 * Run all registered rollup functions for this cpu
319 		 * (e.g. tcp_willblock()).
320 		 */
321 		TAILQ_FOREACH(ru, &netrulist, ru_entry)
322 			ru->ru_func();
323 	}
324 }
325 
326 /*
327  * Forward a packet to a netisr service function.
328  *
329  * If the packet has not been assigned to a protocol thread we call
330  * the port characterization function to assign it.  The caller must
331  * clear M_HASH (or not have set it in the first place) if the caller
332  * wishes the packet to be recharacterized.
333  */
334 int
335 netisr_queue(int num, struct mbuf *m)
336 {
337 	struct netisr *ni;
338 	struct netmsg_packet *pmsg;
339 	lwkt_port_t port;
340 
341 	KASSERT((num > 0 && num <= NELEM(netisrs)),
342 		("Bad isr %d", num));
343 
344 	ni = &netisrs[num];
345 	if (ni->ni_handler == NULL) {
346 		kprintf("Unregistered isr %d\n", num);
347 		m_freem(m);
348 		return (EIO);
349 	}
350 
351 	/*
352 	 * Figure out which protocol thread to send to.  This does not
353 	 * have to be perfect but performance will be really good if it
354 	 * is correct.  Major protocol inputs such as ip_input() will
355 	 * re-characterize the packet as necessary.
356 	 */
357 	if ((m->m_flags & M_HASH) == 0) {
358 		ni->ni_cpufn(&m, 0);
359 		if (m == NULL) {
360 			m_freem(m);
361 			return (EIO);
362 		}
363 		if ((m->m_flags & M_HASH) == 0) {
364 			kprintf("netisr_queue(%d): packet hash failed\n", num);
365 			m_freem(m);
366 			return (EIO);
367 		}
368 	}
369 
370 	/*
371 	 * Get the protocol port based on the packet hash, initialize
372 	 * the netmsg, and send it off.
373 	 */
374 	port = netisr_portfn(m->m_pkthdr.hash);
375 	pmsg = &m->m_hdr.mh_netmsg;
376 	netmsg_init(&pmsg->base, NULL, &netisr_apanic_rport,
377 		    0, ni->ni_handler);
378 	pmsg->nm_packet = m;
379 	pmsg->base.lmsg.u.ms_result = num;
380 	lwkt_sendmsg(port, &pmsg->base.lmsg);
381 
382 	return (0);
383 }
384 
385 /*
386  * Run a netisr service function on the packet.
387  *
388  * The packet must have been correctly characterized!
389  */
390 int
391 netisr_handle(int num, struct mbuf *m)
392 {
393 	struct netisr *ni;
394 	struct netmsg_packet *pmsg;
395 	lwkt_port_t port;
396 
397 	/*
398 	 * Get the protocol port based on the packet hash
399 	 */
400 	KASSERT((m->m_flags & M_HASH), ("packet not characterized"));
401 	port = netisr_portfn(m->m_pkthdr.hash);
402 	KASSERT(&curthread->td_msgport == port, ("wrong msgport"));
403 
404 	KASSERT((num > 0 && num <= NELEM(netisrs)), ("bad isr %d", num));
405 	ni = &netisrs[num];
406 	if (ni->ni_handler == NULL) {
407 		kprintf("unregistered isr %d\n", num);
408 		m_freem(m);
409 		return EIO;
410 	}
411 
412 	/*
413 	 * Initialize the netmsg, and run the handler directly.
414 	 */
415 	pmsg = &m->m_hdr.mh_netmsg;
416 	netmsg_init(&pmsg->base, NULL, &netisr_apanic_rport,
417 		    0, ni->ni_handler);
418 	pmsg->nm_packet = m;
419 	pmsg->base.lmsg.u.ms_result = num;
420 	ni->ni_handler((netmsg_t)&pmsg->base);
421 
422 	return 0;
423 }
424 
425 /*
426  * Pre-characterization of a deeper portion of the packet for the
427  * requested isr.
428  *
429  * The base of the ISR type (e.g. IP) that we want to characterize is
430  * at (hoff) relative to the beginning of the mbuf.  This allows
431  * e.g. ether_input_chain() to not have to adjust the m_data/m_len.
432  */
433 void
434 netisr_characterize(int num, struct mbuf **mp, int hoff)
435 {
436 	struct netisr *ni;
437 	struct mbuf *m;
438 
439 	/*
440 	 * Validation
441 	 */
442 	m = *mp;
443 	KKASSERT(m != NULL);
444 
445 	if (num < 0 || num >= NETISR_MAX) {
446 		if (num == NETISR_MAX) {
447 			m->m_flags |= M_HASH;
448 			m->m_pkthdr.hash = 0;
449 			return;
450 		}
451 		panic("Bad isr %d", num);
452 	}
453 
454 	/*
455 	 * Valid netisr?
456 	 */
457 	ni = &netisrs[num];
458 	if (ni->ni_handler == NULL) {
459 		kprintf("Unregistered isr %d\n", num);
460 		m_freem(m);
461 		*mp = NULL;
462 	}
463 
464 	/*
465 	 * Characterize the packet
466 	 */
467 	if ((m->m_flags & M_HASH) == 0) {
468 		ni->ni_cpufn(mp, hoff);
469 		m = *mp;
470 		if (m && (m->m_flags & M_HASH) == 0)
471 			kprintf("netisr_queue(%d): packet hash failed\n", num);
472 	}
473 }
474 
475 void
476 netisr_init_dedicated(int num)
477 {
478 	KKASSERT(num > 0 && num < NETISR_MAX);
479 	KKASSERT(netisr_ded[num].td_pri == 0);
480 	lwkt_create(netmsg_service_loop, NULL, NULL,
481 		    &netisr_ded[num], TDF_NOSTART|TDF_FORCE_SPINPORT,
482 		    num % ncpus, "netisr_ded %d", num);
483 	netmsg_service_port_init(&netisr_ded[num].td_msgport);
484 	lwkt_schedule(&netisr_ded[num]);
485 }
486 
487 void
488 netisr_register(int num, netisr_fn_t handler, netisr_cpufn_t cpufn)
489 {
490 	struct netisr *ni;
491 
492 	KASSERT((num > 0 && num <= NELEM(netisrs)),
493 		("netisr_register: bad isr %d", num));
494 	KKASSERT(handler != NULL);
495 
496 	if (cpufn == NULL)
497 		cpufn = cpu0_cpufn;
498 
499 	ni = &netisrs[num];
500 
501 	ni->ni_handler = handler;
502 	ni->ni_hashck = netisr_nohashck;
503 	ni->ni_cpufn = cpufn;
504 	netmsg_init(&ni->ni_netmsg, NULL, &netisr_adone_rport, 0, NULL);
505 }
506 
507 void
508 netisr_register_hashcheck(int num, netisr_hashck_t hashck)
509 {
510 	struct netisr *ni;
511 
512 	KASSERT((num > 0 && num <= NELEM(netisrs)),
513 		("netisr_register: bad isr %d", num));
514 
515 	ni = &netisrs[num];
516 	ni->ni_hashck = hashck;
517 }
518 
519 void
520 netisr_register_rollup(netisr_ru_t ru_func, int prio)
521 {
522 	struct netmsg_rollup *new_ru, *ru;
523 
524 	new_ru = kmalloc(sizeof(*new_ru), M_TEMP, M_WAITOK|M_ZERO);
525 	new_ru->ru_func = ru_func;
526 	new_ru->ru_prio = prio;
527 
528 	/*
529 	 * Higher priority "rollup" appears first
530 	 */
531 	TAILQ_FOREACH(ru, &netrulist, ru_entry) {
532 		if (ru->ru_prio < new_ru->ru_prio) {
533 			TAILQ_INSERT_BEFORE(ru, new_ru, ru_entry);
534 			return;
535 		}
536 	}
537 	TAILQ_INSERT_TAIL(&netrulist, new_ru, ru_entry);
538 }
539 
540 /*
541  * Return the message port for the general protocol message servicing
542  * thread for a particular cpu.
543  *
544  * A standard cpu value returns the general lockless/asynchronous
545  * netisr thread for the cpu specified.
546  *
547  * A dedicated cpu value specifies a thread dedicated to a particular
548  * ISR.  Such threads can potentially stall or block for long periods
549  * of time (see arp_init() for an example).
550  */
551 lwkt_port_t
552 netisr_portfn(int cpu)
553 {
554 	if (__predict_false(cpu & NETISR_DEDICATED)) {
555 		cpu &= (NETISR_DEDICATED - 1);
556 		KKASSERT(cpu < NETISR_MAX && netisr_ded[cpu].td_pri != 0);
557 		return (&netisr_ded[cpu].td_msgport);
558 	} else {
559 		KKASSERT((uint32_t)cpu < ncpus);
560 		return (&netisr_cpu[cpu].td_msgport);
561 	}
562 }
563 
564 /*
565  * Return the current cpu's network protocol thread.
566  */
567 lwkt_port_t
568 cur_netport(void)
569 {
570 	return(netisr_portfn(mycpu->gd_cpuid));
571 }
572 
573 /*
574  * Return a default protocol control message processing thread port
575  */
576 lwkt_port_t
577 cpu0_ctlport(int cmd __unused, struct sockaddr *sa __unused,
578 	     void *extra __unused)
579 {
580 	return (&netisr_cpu[0].td_msgport);
581 }
582 
583 /*
584  * This is a default netisr packet characterization function which
585  * sets M_HASH.  If a netisr is registered with a NULL cpufn function
586  * this one is assigned.
587  *
588  * This function makes no attempt to validate the packet.
589  */
590 static void
591 cpu0_cpufn(struct mbuf **mp, int hoff __unused)
592 {
593 	struct mbuf *m = *mp;
594 
595 	m->m_flags |= M_HASH;
596 	m->m_pkthdr.hash = 0;
597 }
598 
599 /*
600  * schednetisr() is used to call the netisr handler from the appropriate
601  * netisr thread for polling and other purposes.
602  *
603  * This function may be called from a hard interrupt or IPI and must be
604  * MP SAFE and non-blocking.  We use a fixed per-cpu message instead of
605  * trying to allocate one.  We must get ourselves onto the target cpu
606  * to safely check the MSGF_DONE bit on the message but since the message
607  * will be sent to that cpu anyway this does not add any extra work beyond
608  * what lwkt_sendmsg() would have already had to do to schedule the target
609  * thread.
610  */
611 static void
612 schednetisr_remote(void *data)
613 {
614 	int num = (int)(intptr_t)data;
615 	struct netisr *ni = &netisrs[num];
616 	lwkt_port_t port = &netisr_cpu[0].td_msgport;
617 	netmsg_base_t pmsg;
618 
619 	pmsg = &netisrs[num].ni_netmsg;
620 	if (pmsg->lmsg.ms_flags & MSGF_DONE) {
621 		netmsg_init(pmsg, NULL, &netisr_adone_rport, 0, ni->ni_handler);
622 		pmsg->lmsg.u.ms_result = num;
623 		lwkt_sendmsg(port, &pmsg->lmsg);
624 	}
625 }
626 
627 void
628 schednetisr(int num)
629 {
630 	KASSERT((num > 0 && num <= NELEM(netisrs)),
631 		("schednetisr: bad isr %d", num));
632 	KKASSERT(netisrs[num].ni_handler != NULL);
633 	if (mycpu->gd_cpuid != 0) {
634 		lwkt_send_ipiq(globaldata_find(0),
635 			       schednetisr_remote, (void *)(intptr_t)num);
636 	} else {
637 		crit_enter();
638 		schednetisr_remote((void *)(intptr_t)num);
639 		crit_exit();
640 	}
641 }
642 
643 static void
644 netisr_barrier_dispatch(netmsg_t nmsg)
645 {
646 	struct netmsg_barrier *msg = (struct netmsg_barrier *)nmsg;
647 
648 	atomic_clear_cpumask(msg->br_cpumask, mycpu->gd_cpumask);
649 	if (*msg->br_cpumask == 0)
650 		wakeup(msg->br_cpumask);
651 
652 	for (;;) {
653 		uint32_t done = msg->br_done;
654 
655 		cpu_ccfence();
656 		if ((done & NETISR_BR_NOTDONE) == 0)
657 			break;
658 
659 		tsleep_interlock(&msg->br_done, 0);
660 		if (atomic_cmpset_int(&msg->br_done,
661 		    done, done | NETISR_BR_WAITDONE))
662 			tsleep(&msg->br_done, PINTERLOCKED, "nbrdsp", 0);
663 	}
664 
665 	lwkt_replymsg(&nmsg->lmsg, 0);
666 }
667 
668 struct netisr_barrier *
669 netisr_barrier_create(void)
670 {
671 	struct netisr_barrier *br;
672 
673 	br = kmalloc(sizeof(*br), M_LWKTMSG, M_WAITOK | M_ZERO);
674 	return br;
675 }
676 
677 void
678 netisr_barrier_set(struct netisr_barrier *br)
679 {
680 	volatile cpumask_t other_cpumask;
681 	int i, cur_cpuid;
682 
683 	KKASSERT(&curthread->td_msgport == netisr_portfn(0));
684 	KKASSERT(!br->br_isset);
685 
686 	other_cpumask = mycpu->gd_other_cpus & smp_active_mask;
687 	cur_cpuid = mycpuid;
688 
689 	for (i = 0; i < ncpus; ++i) {
690 		struct netmsg_barrier *msg;
691 
692 		if (i == cur_cpuid)
693 			continue;
694 
695 		msg = kmalloc(sizeof(struct netmsg_barrier),
696 			      M_LWKTMSG, M_WAITOK);
697 		netmsg_init(&msg->base, NULL, &netisr_afree_rport,
698 			    MSGF_PRIORITY, netisr_barrier_dispatch);
699 		msg->br_cpumask = &other_cpumask;
700 		msg->br_done = NETISR_BR_NOTDONE;
701 
702 		KKASSERT(br->br_msgs[i] == NULL);
703 		br->br_msgs[i] = msg;
704 	}
705 
706 	for (i = 0; i < ncpus; ++i) {
707 		if (i == cur_cpuid)
708 			continue;
709 		lwkt_sendmsg(netisr_portfn(i), &br->br_msgs[i]->base.lmsg);
710 	}
711 
712 	while (other_cpumask != 0) {
713 		tsleep_interlock(&other_cpumask, 0);
714 		if (other_cpumask != 0)
715 			tsleep(&other_cpumask, PINTERLOCKED, "nbrset", 0);
716 	}
717 	br->br_isset = 1;
718 }
719 
720 void
721 netisr_barrier_rem(struct netisr_barrier *br)
722 {
723 	int i, cur_cpuid;
724 
725 	KKASSERT(&curthread->td_msgport == netisr_portfn(0));
726 	KKASSERT(br->br_isset);
727 
728 	cur_cpuid = mycpuid;
729 	for (i = 0; i < ncpus; ++i) {
730 		struct netmsg_barrier *msg = br->br_msgs[i];
731 		uint32_t done;
732 
733 		msg = br->br_msgs[i];
734 		br->br_msgs[i] = NULL;
735 
736 		if (i == cur_cpuid)
737 			continue;
738 
739 		done = atomic_swap_int(&msg->br_done, 0);
740 		if (done & NETISR_BR_WAITDONE)
741 			wakeup(&msg->br_done);
742 	}
743 	br->br_isset = 0;
744 }
745 
746 static void
747 netisr_nohashck(struct mbuf *m, const struct pktinfo *pi __unused)
748 {
749 	m->m_flags &= ~M_HASH;
750 }
751 
752 void
753 netisr_hashcheck(int num, struct mbuf *m, const struct pktinfo *pi)
754 {
755 	struct netisr *ni;
756 
757 	if (num < 0 || num >= NETISR_MAX)
758 		panic("Bad isr %d", num);
759 
760 	/*
761 	 * Valid netisr?
762 	 */
763 	ni = &netisrs[num];
764 	if (ni->ni_handler == NULL)
765 		panic("Unregistered isr %d", num);
766 
767 	ni->ni_hashck(m, pi);
768 }
769