netisr barrier: Avoid lockless wakeup/tsleep race
[dragonfly.git] / sys / net / netisr.c
CommitLineData
66d6c637
JH
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 *
d849e575
MD
10 * Jonathan Lemon gave Jeffrey Hsu permission to combine his copyright
11 * into this one around July 8 2004.
12 *
66d6c637
JH
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.
ef0fdad1
MD
37 */
38
39#include <sys/param.h>
40#include <sys/systm.h>
bf82f9b7 41#include <sys/kernel.h>
9eeaa8a9 42#include <sys/malloc.h>
bf82f9b7 43#include <sys/msgport.h>
ef0fdad1
MD
44#include <sys/proc.h>
45#include <sys/interrupt.h>
8bde602d
JH
46#include <sys/socket.h>
47#include <sys/sysctl.h>
48e7b118 48#include <sys/socketvar.h>
8bde602d
JH
49#include <net/if.h>
50#include <net/if_var.h>
ef0fdad1
MD
51#include <net/netisr.h>
52#include <machine/cpufunc.h>
a91f9815 53#include <machine/smp.h>
ef0fdad1 54
3227f1b8
MD
55#include <sys/thread2.h>
56#include <sys/msgport2.h>
4599cf19 57#include <net/netmsg2.h>
684a93c4 58#include <sys/mplock2.h>
3227f1b8 59
002c1265 60static void netmsg_sync_func(netmsg_t msg);
c3c96e44
MD
61static void netmsg_service_loop(void *arg);
62static void cpu0_cpufn(struct mbuf **mp, int hoff);
5c703385
MD
63
64struct netmsg_port_registration {
c3c96e44
MD
65 TAILQ_ENTRY(netmsg_port_registration) npr_entry;
66 lwkt_port_t npr_port;
67};
68
69struct netmsg_rollup {
70 TAILQ_ENTRY(netmsg_rollup) ru_entry;
71 netisr_ru_t ru_func;
5c703385
MD
72};
73
a91f9815
SZ
74struct netmsg_barrier {
75 struct netmsg_base base;
0503d1d0 76 volatile cpumask_t *br_cpumask;
ca3321f8 77 volatile uint32_t br_done;
a91f9815
SZ
78};
79
ca3321f8
SZ
80#define NETISR_BR_NOTDONE 0
81#define NETISR_BR_DONE 1
82#define NETISR_BR_WAITDONE 2
83
a91f9815
SZ
84struct netisr_barrier {
85 struct netmsg_barrier *br_msgs[MAXCPU];
86 int br_isset;
87};
88
bf82f9b7 89static struct netisr netisrs[NETISR_MAX];
5c703385 90static TAILQ_HEAD(,netmsg_port_registration) netreglist;
c3c96e44 91static TAILQ_HEAD(,netmsg_rollup) netrulist;
bf82f9b7
MD
92
93/* Per-CPU thread to handle any protocol. */
c3c96e44 94static struct thread netisr_cpu[MAXCPU];
3227f1b8 95lwkt_port netisr_afree_rport;
c3d495a1 96lwkt_port netisr_afree_free_so_rport;
a29576fc 97lwkt_port netisr_adone_rport;
6aad077d 98lwkt_port netisr_apanic_rport;
3efe7008 99lwkt_port netisr_sync_port;
3227f1b8 100
fb0f29c4
MD
101static int (*netmsg_fwd_port_fn)(lwkt_port_t, lwkt_msg_t);
102
92db3805 103SYSCTL_NODE(_net, OID_AUTO, netisr, CTLFLAG_RW, 0, "netisr");
ff4a1403 104
3227f1b8
MD
105/*
106 * netisr_afree_rport replymsg function, only used to handle async
107 * messages which the sender has abandoned to their fate.
108 */
109static void
110netisr_autofree_reply(lwkt_port_t port, lwkt_msg_t msg)
111{
c3c96e44 112 kfree(msg, M_LWKTMSG);
3227f1b8 113}
ef0fdad1 114
c3d495a1
MD
115static void
116netisr_autofree_free_so_reply(lwkt_port_t port, lwkt_msg_t msg)
117{
118 sofree(((netmsg_t)msg)->base.nm_so);
119 kfree(msg, M_LWKTMSG);
120}
121
dc22b3aa 122/*
fb0f29c4
MD
123 * We need a custom putport function to handle the case where the
124 * message target is the current thread's message port. This case
125 * can occur when the TCP or UDP stack does a direct callback to NFS and NFS
126 * then turns around and executes a network operation synchronously.
3efe7008 127 *
fb0f29c4
MD
128 * To prevent deadlocking, we must execute these self-referential messages
129 * synchronously, effectively turning the message into a glorified direct
130 * procedure call back into the protocol stack. The operation must be
131 * complete on return or we will deadlock, so panic if it isn't.
002c1265
MD
132 *
133 * However, the target function is under no obligation to immediately
134 * reply the message. It may forward it elsewhere.
dc22b3aa 135 */
5c703385 136static int
dc22b3aa
JH
137netmsg_put_port(lwkt_port_t port, lwkt_msg_t lmsg)
138{
002c1265 139 netmsg_base_t nmsg = (void *)lmsg;
c3c96e44
MD
140
141 if ((lmsg->ms_flags & MSGF_SYNC) && port == &curthread->td_msgport) {
002c1265 142 nmsg->nm_dispatch((netmsg_t)nmsg);
c3c96e44
MD
143 return(EASYNC);
144 } else {
145 return(netmsg_fwd_port_fn(port, lmsg));
146 }
dc22b3aa
JH
147}
148
3efe7008
MD
149/*
150 * UNIX DOMAIN sockets still have to run their uipc functions synchronously,
151 * because they depend on the user proc context for a number of things
152 * (like creds) which we have not yet incorporated into the message structure.
153 *
154 * However, we maintain or message/port abstraction. Having a special
155 * synchronous port which runs the commands synchronously gives us the
156 * ability to serialize operations in one place later on when we start
157 * removing the BGL.
3efe7008
MD
158 */
159static int
160netmsg_sync_putport(lwkt_port_t port, lwkt_msg_t lmsg)
161{
002c1265 162 netmsg_base_t nmsg = (void *)lmsg;
3efe7008 163
c3c96e44 164 KKASSERT((lmsg->ms_flags & MSGF_DONE) == 0);
e0383bf3 165
c3c96e44 166 lmsg->ms_target_port = port; /* required for abort */
002c1265 167 nmsg->nm_dispatch((netmsg_t)nmsg);
c3c96e44 168 return(EASYNC);
3efe7008
MD
169}
170
171static void
bf82f9b7 172netisr_init(void)
ef0fdad1 173{
c3c96e44
MD
174 int i;
175
176 TAILQ_INIT(&netreglist);
177 TAILQ_INIT(&netrulist);
178
179 /*
180 * Create default per-cpu threads for generic protocol handling.
181 */
182 for (i = 0; i < ncpus; ++i) {
183 lwkt_create(netmsg_service_loop, NULL, NULL,
184 &netisr_cpu[i], TDF_STOPREQ, i,
185 "netisr_cpu %d", i);
186 netmsg_service_port_init(&netisr_cpu[i].td_msgport);
187 lwkt_schedule(&netisr_cpu[i]);
188 }
189
190 /*
191 * The netisr_afree_rport is a special reply port which automatically
192 * frees the replied message. The netisr_adone_rport simply marks
193 * the message as being done. The netisr_apanic_rport panics if
194 * the message is replied to.
195 */
196 lwkt_initport_replyonly(&netisr_afree_rport, netisr_autofree_reply);
c3d495a1
MD
197 lwkt_initport_replyonly(&netisr_afree_free_so_rport,
198 netisr_autofree_free_so_reply);
c3c96e44
MD
199 lwkt_initport_replyonly_null(&netisr_adone_rport);
200 lwkt_initport_panic(&netisr_apanic_rport);
201
202 /*
203 * The netisr_syncport is a special port which executes the message
204 * synchronously and waits for it if EASYNC is returned.
205 */
206 lwkt_initport_putonly(&netisr_sync_port, netmsg_sync_putport);
ef0fdad1
MD
207}
208
b2632176 209SYSINIT(netisr, SI_SUB_PRE_DRIVERS, SI_ORDER_FIRST, netisr_init, NULL);
bf82f9b7 210
5c703385
MD
211/*
212 * Finish initializing the message port for a netmsg service. This also
213 * registers the port for synchronous cleanup operations such as when an
214 * ifnet is being destroyed. There is no deregistration API yet.
215 */
216void
217netmsg_service_port_init(lwkt_port_t port)
218{
c3c96e44
MD
219 struct netmsg_port_registration *reg;
220
221 /*
222 * Override the putport function. Our custom function checks for
223 * self-references and executes such commands synchronously.
224 */
225 if (netmsg_fwd_port_fn == NULL)
226 netmsg_fwd_port_fn = port->mp_putport;
227 KKASSERT(netmsg_fwd_port_fn == port->mp_putport);
228 port->mp_putport = netmsg_put_port;
229
230 /*
231 * Keep track of ports using the netmsg API so we can synchronize
232 * certain operations (such as freeing an ifnet structure) across all
233 * consumers.
234 */
235 reg = kmalloc(sizeof(*reg), M_TEMP, M_WAITOK|M_ZERO);
236 reg->npr_port = port;
237 TAILQ_INSERT_TAIL(&netreglist, reg, npr_entry);
5c703385
MD
238}
239
240/*
241 * This function synchronizes the caller with all netmsg services. For
242 * example, if an interface is being removed we must make sure that all
243 * packets related to that interface complete processing before the structure
244 * can actually be freed. This sort of synchronization is an alternative to
245 * ref-counting the netif, removing the ref counting overhead in favor of
246 * placing additional overhead in the netif freeing sequence (where it is
247 * inconsequential).
248 */
249void
250netmsg_service_sync(void)
251{
c3c96e44 252 struct netmsg_port_registration *reg;
002c1265 253 struct netmsg_base smsg;
5c703385 254
c3c96e44 255 netmsg_init(&smsg, NULL, &curthread->td_msgport, 0, netmsg_sync_func);
5c703385 256
c3c96e44 257 TAILQ_FOREACH(reg, &netreglist, npr_entry) {
002c1265 258 lwkt_domsg(reg->npr_port, &smsg.lmsg, 0);
c3c96e44 259 }
5c703385
MD
260}
261
262/*
263 * The netmsg function simply replies the message. API semantics require
264 * EASYNC to be returned if the netmsg function disposes of the message.
265 */
4599cf19 266static void
002c1265 267netmsg_sync_func(netmsg_t msg)
5c703385 268{
002c1265 269 lwkt_replymsg(&msg->lmsg, 0);
bf82f9b7 270}
8bde602d 271
bf82f9b7 272/*
92db3805
SZ
273 * Generic netmsg service loop. Some protocols may roll their own but all
274 * must do the basic command dispatch function call done here.
95f8b5ce 275 */
c3c96e44 276static void
92db3805 277netmsg_service_loop(void *arg)
95f8b5ce 278{
c3c96e44 279 struct netmsg_rollup *ru;
002c1265 280 netmsg_base_t msg;
c3c96e44
MD
281 thread_t td = curthread;;
282 int limit;
283
284 while ((msg = lwkt_waitport(&td->td_msgport, 0))) {
285 /*
286 * Run up to 512 pending netmsgs.
287 */
288 limit = 512;
289 do {
290 KASSERT(msg->nm_dispatch != NULL,
291 ("netmsg_service isr %d badmsg\n",
002c1265 292 msg->lmsg.u.ms_result));
0ce0603e
MD
293 if (msg->nm_so &&
294 msg->nm_so->so_port != &td->td_msgport) {
295 /*
296 * Sockets undergoing connect or disconnect
297 * ops can change ports on us. Chase the
298 * port.
299 */
300 kprintf("netmsg_service_loop: Warning, "
301 "port changed so=%p\n", msg->nm_so);
302 lwkt_forwardmsg(msg->nm_so->so_port,
002c1265 303 &msg->lmsg);
0ce0603e
MD
304 } else {
305 /*
306 * We are on the correct port, dispatch it.
307 */
002c1265 308 msg->nm_dispatch((netmsg_t)msg);
0ce0603e 309 }
c3c96e44
MD
310 if (--limit == 0)
311 break;
312 } while ((msg = lwkt_getport(&td->td_msgport)) != NULL);
313
314 /*
315 * Run all registered rollup functions for this cpu
316 * (e.g. tcp_willblock()).
317 */
318 TAILQ_FOREACH(ru, &netrulist, ru_entry)
319 ru->ru_func();
320 }
95f8b5ce
SZ
321}
322
323/*
c3c96e44
MD
324 * Forward a packet to a netisr service function.
325 *
326 * If the packet has not been assigned to a protocol thread we call
327 * the port characterization function to assign it. The caller must
328 * clear M_HASH (or not have set it in the first place) if the caller
329 * wishes the packet to be recharacterized.
bf82f9b7 330 */
c3c96e44
MD
331int
332netisr_queue(int num, struct mbuf *m)
bf82f9b7 333{
c3c96e44
MD
334 struct netisr *ni;
335 struct netmsg_packet *pmsg;
336 lwkt_port_t port;
337
c157ff7a 338 KASSERT((num > 0 && num <= NELEM(netisrs)),
c3c96e44
MD
339 ("Bad isr %d", num));
340
341 ni = &netisrs[num];
342 if (ni->ni_handler == NULL) {
343 kprintf("Unregistered isr %d\n", num);
344 m_freem(m);
345 return (EIO);
346 }
347
348 /*
349 * Figure out which protocol thread to send to. This does not
350 * have to be perfect but performance will be really good if it
351 * is correct. Major protocol inputs such as ip_input() will
352 * re-characterize the packet as necessary.
353 */
354 if ((m->m_flags & M_HASH) == 0) {
355 ni->ni_cpufn(&m, 0);
356 if (m == NULL) {
357 m_freem(m);
358 return (EIO);
359 }
360 if ((m->m_flags & M_HASH) == 0) {
361 kprintf("netisr_queue(%d): packet hash failed\n", num);
362 m_freem(m);
363 return (EIO);
364 }
365 }
366
367 /*
368 * Get the protocol port based on the packet hash, initialize
369 * the netmsg, and send it off.
370 */
371 port = cpu_portfn(m->m_pkthdr.hash);
372 pmsg = &m->m_hdr.mh_netmsg;
002c1265 373 netmsg_init(&pmsg->base, NULL, &netisr_apanic_rport,
c3c96e44
MD
374 0, ni->ni_handler);
375 pmsg->nm_packet = m;
002c1265
MD
376 pmsg->base.lmsg.u.ms_result = num;
377 lwkt_sendmsg(port, &pmsg->base.lmsg);
c3c96e44
MD
378
379 return (0);
8bde602d
JH
380}
381
382/*
c3c96e44
MD
383 * Pre-characterization of a deeper portion of the packet for the
384 * requested isr.
385 *
386 * The base of the ISR type (e.g. IP) that we want to characterize is
387 * at (hoff) relative to the beginning of the mbuf. This allows
388 * e.g. ether_input_chain() to not have to adjust the m_data/m_len.
8bde602d 389 */
c3c96e44
MD
390void
391netisr_characterize(int num, struct mbuf **mp, int hoff)
ef0fdad1 392{
c3c96e44
MD
393 struct netisr *ni;
394 struct mbuf *m;
395
396 /*
397 * Validation
398 */
c3c96e44
MD
399 m = *mp;
400 KKASSERT(m != NULL);
401
e6318d16
MD
402 if (num < 0 || num >= NETISR_MAX) {
403 if (num == NETISR_MAX) {
404 m->m_flags |= M_HASH;
405 m->m_pkthdr.hash = 0;
406 return;
407 }
408 panic("Bad isr %d", num);
409 }
410
c3c96e44
MD
411 /*
412 * Valid netisr?
413 */
414 ni = &netisrs[num];
415 if (ni->ni_handler == NULL) {
416 kprintf("Unregistered isr %d\n", num);
417 m_freem(m);
418 *mp = NULL;
419 }
420
421 /*
422 * Characterize the packet
423 */
424 if ((m->m_flags & M_HASH) == 0) {
425 ni->ni_cpufn(mp, hoff);
426 m = *mp;
427 if (m && (m->m_flags & M_HASH) == 0)
428 kprintf("netisr_queue(%d): packet hash failed\n", num);
429 }
8bde602d
JH
430}
431
bf82f9b7 432void
c3c96e44 433netisr_register(int num, netisr_fn_t handler, netisr_cpufn_t cpufn)
8bde602d 434{
c3c96e44 435 struct netisr *ni;
ef0fdad1 436
c157ff7a 437 KASSERT((num > 0 && num <= NELEM(netisrs)),
c3c96e44
MD
438 ("netisr_register: bad isr %d", num));
439 KKASSERT(handler != NULL);
440
441 if (cpufn == NULL)
442 cpufn = cpu0_cpufn;
8bde602d 443
c3c96e44
MD
444 ni = &netisrs[num];
445
446 ni->ni_handler = handler;
447 ni->ni_cpufn = cpufn;
448 netmsg_init(&ni->ni_netmsg, NULL, &netisr_adone_rport, 0, NULL);
ef0fdad1 449}
bf82f9b7 450
c3c96e44
MD
451void
452netisr_register_rollup(netisr_ru_t ru_func)
bf82f9b7 453{
c3c96e44 454 struct netmsg_rollup *ru;
97a43e72 455
c3c96e44
MD
456 ru = kmalloc(sizeof(*ru), M_TEMP, M_WAITOK|M_ZERO);
457 ru->ru_func = ru_func;
458 TAILQ_INSERT_TAIL(&netrulist, ru, ru_entry);
bf82f9b7
MD
459}
460
c3c96e44
MD
461/*
462 * Return the message port for the general protocol message servicing
463 * thread for a particular cpu.
464 */
ecdefdda
MD
465lwkt_port_t
466cpu_portfn(int cpu)
467{
c3c96e44
MD
468 KKASSERT(cpu >= 0 && cpu < ncpus);
469 return (&netisr_cpu[cpu].td_msgport);
ecdefdda
MD
470}
471
c244d613 472/*
c3c96e44 473 * Return the current cpu's network protocol thread.
c244d613
SZ
474 */
475lwkt_port_t
476cur_netport(void)
477{
c3c96e44 478 return(cpu_portfn(mycpu->gd_cpuid));
c244d613
SZ
479}
480
c3c96e44 481/*
c3c96e44
MD
482 * Return a default protocol control message processing thread port
483 */
3efe7008 484lwkt_port_t
e3873585
SZ
485cpu0_ctlport(int cmd __unused, struct sockaddr *sa __unused,
486 void *extra __unused)
487{
c3c96e44 488 return (&netisr_cpu[0].td_msgport);
e3873585
SZ
489}
490
c3c96e44 491/*
c3c96e44
MD
492 * This is a default netisr packet characterization function which
493 * sets M_HASH. If a netisr is registered with a NULL cpufn function
494 * this one is assigned.
495 *
496 * This function makes no attempt to validate the packet.
497 */
498static void
499cpu0_cpufn(struct mbuf **mp, int hoff __unused)
500{
501 struct mbuf *m = *mp;
502
503 m->m_flags |= M_HASH;
504 m->m_pkthdr.hash = 0;
3efe7008
MD
505}
506
bf82f9b7 507/*
a29576fc 508 * schednetisr() is used to call the netisr handler from the appropriate
9eeaa8a9 509 * netisr thread for polling and other purposes.
a29576fc
MD
510 *
511 * This function may be called from a hard interrupt or IPI and must be
512 * MP SAFE and non-blocking. We use a fixed per-cpu message instead of
513 * trying to allocate one. We must get ourselves onto the target cpu
514 * to safely check the MSGF_DONE bit on the message but since the message
515 * will be sent to that cpu anyway this does not add any extra work beyond
516 * what lwkt_sendmsg() would have already had to do to schedule the target
517 * thread.
bf82f9b7 518 */
a29576fc
MD
519static void
520schednetisr_remote(void *data)
bf82f9b7 521{
c3c96e44
MD
522 int num = (int)(intptr_t)data;
523 struct netisr *ni = &netisrs[num];
524 lwkt_port_t port = &netisr_cpu[0].td_msgport;
002c1265 525 netmsg_base_t pmsg;
c3c96e44
MD
526
527 pmsg = &netisrs[num].ni_netmsg;
002c1265 528 if (pmsg->lmsg.ms_flags & MSGF_DONE) {
c3c96e44 529 netmsg_init(pmsg, NULL, &netisr_adone_rport, 0, ni->ni_handler);
002c1265
MD
530 pmsg->lmsg.u.ms_result = num;
531 lwkt_sendmsg(port, &pmsg->lmsg);
c3c96e44 532 }
a29576fc
MD
533}
534
535void
536schednetisr(int num)
537{
c157ff7a 538 KASSERT((num > 0 && num <= NELEM(netisrs)),
c3c96e44
MD
539 ("schednetisr: bad isr %d", num));
540 KKASSERT(netisrs[num].ni_handler != NULL);
a29576fc 541#ifdef SMP
c3c96e44
MD
542 if (mycpu->gd_cpuid != 0) {
543 lwkt_send_ipiq(globaldata_find(0),
544 schednetisr_remote, (void *)(intptr_t)num);
545 } else {
546 crit_enter();
547 schednetisr_remote((void *)(intptr_t)num);
548 crit_exit();
549 }
a29576fc 550#else
c3c96e44
MD
551 crit_enter();
552 schednetisr_remote((void *)(intptr_t)num);
553 crit_exit();
a29576fc 554#endif
bf82f9b7 555}
a91f9815
SZ
556
557#ifdef SMP
558
559static void
560netisr_barrier_dispatch(netmsg_t nmsg)
561{
562 struct netmsg_barrier *msg = (struct netmsg_barrier *)nmsg;
563
564 atomic_clear_cpumask(msg->br_cpumask, mycpu->gd_cpumask);
565 if (*msg->br_cpumask == 0)
566 wakeup(msg->br_cpumask);
567
ca3321f8
SZ
568 while (msg->br_done == NETISR_BR_NOTDONE) {
569 cpu_ccfence();
570 tsleep_interlock(&msg->br_done, 0);
571 if (atomic_cmpset_int(&msg->br_done,
572 NETISR_BR_NOTDONE, NETISR_BR_WAITDONE))
573 tsleep(&msg->br_done, PINTERLOCKED, "nbrdsp", 0);
574 }
a91f9815
SZ
575
576 lwkt_replymsg(&nmsg->lmsg, 0);
577}
578
579#endif
580
581struct netisr_barrier *
582netisr_barrier_create(void)
583{
584 struct netisr_barrier *br;
585
586 br = kmalloc(sizeof(*br), M_LWKTMSG, M_WAITOK | M_ZERO);
587 return br;
588}
589
590void
591netisr_barrier_set(struct netisr_barrier *br)
592{
593#ifdef SMP
0503d1d0 594 volatile cpumask_t other_cpumask;
a91f9815
SZ
595 int i, cur_cpuid;
596
597 KKASSERT(&curthread->td_msgport == cpu_portfn(0));
598 KKASSERT(!br->br_isset);
599
600 other_cpumask = mycpu->gd_other_cpus & smp_active_mask;
601 cur_cpuid = mycpuid;
602
603 for (i = 0; i < ncpus; ++i) {
604 struct netmsg_barrier *msg;
605
606 if (i == cur_cpuid)
607 continue;
608
609 msg = kmalloc(sizeof(struct netmsg_barrier),
610 M_LWKTMSG, M_WAITOK);
611 netmsg_init(&msg->base, NULL, &netisr_afree_rport,
612 MSGF_PRIORITY, netisr_barrier_dispatch);
613 msg->br_cpumask = &other_cpumask;
ca3321f8 614 msg->br_done = NETISR_BR_NOTDONE;
a91f9815
SZ
615
616 KKASSERT(br->br_msgs[i] == NULL);
617 br->br_msgs[i] = msg;
618 }
619
620 for (i = 0; i < ncpus; ++i) {
621 if (i == cur_cpuid)
622 continue;
623 lwkt_sendmsg(cpu_portfn(i), &br->br_msgs[i]->base.lmsg);
624 }
625
626 while (other_cpumask != 0) {
627 tsleep_interlock(&other_cpumask, 0);
628 if (other_cpumask != 0)
629 tsleep(&other_cpumask, PINTERLOCKED, "nbrset", 0);
630 }
631#endif
632 br->br_isset = 1;
633}
634
635void
636netisr_barrier_rem(struct netisr_barrier *br)
637{
638#ifdef SMP
639 int i, cur_cpuid;
640
641 KKASSERT(&curthread->td_msgport == cpu_portfn(0));
642 KKASSERT(br->br_isset);
643
644 cur_cpuid = mycpuid;
645 for (i = 0; i < ncpus; ++i) {
646 struct netmsg_barrier *msg = br->br_msgs[i];
647
648 msg = br->br_msgs[i];
649 br->br_msgs[i] = NULL;
650
651 if (i == cur_cpuid)
652 continue;
653
ca3321f8
SZ
654 for (;;) {
655 if (atomic_cmpset_int(&msg->br_done,
656 NETISR_BR_WAITDONE, NETISR_BR_DONE)) {
657 wakeup(&msg->br_done);
658 break;
659 }
660 }
a91f9815
SZ
661 }
662#endif
663 br->br_isset = 0;
664}