kernel - Major signal path adjustments to fix races, tsleep race fixes, +more
[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);
e6f77b88 63static void netisr_nohashck(struct mbuf *, const struct pktinfo *);
5c703385
MD
64
65struct netmsg_port_registration {
c3c96e44
MD
66 TAILQ_ENTRY(netmsg_port_registration) npr_entry;
67 lwkt_port_t npr_port;
68};
69
70struct netmsg_rollup {
71 TAILQ_ENTRY(netmsg_rollup) ru_entry;
72 netisr_ru_t ru_func;
5c703385
MD
73};
74
a91f9815
SZ
75struct netmsg_barrier {
76 struct netmsg_base base;
0503d1d0 77 volatile cpumask_t *br_cpumask;
ca3321f8 78 volatile uint32_t br_done;
a91f9815
SZ
79};
80
d0c7a72a
SZ
81#define NETISR_BR_NOTDONE 0x1
82#define NETISR_BR_WAITDONE 0x80000000
ca3321f8 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,
4643740a 184 &netisr_cpu[i], TDF_NOSTART|TDF_FORCE_SPINPORT,
392cd266 185 i, "netisr_cpu %d", i);
c3c96e44
MD
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/*
ebe4c2ae
SZ
383 * Run a netisr service function on the packet.
384 *
385 * The packet must have been correctly characterized!
386 */
387int
388netisr_handle(int num, struct mbuf *m)
389{
390 struct netisr *ni;
391 struct netmsg_packet *pmsg;
392 lwkt_port_t port;
393
394 /*
395 * Get the protocol port based on the packet hash
396 */
397 KASSERT((m->m_flags & M_HASH), ("packet not characterized\n"));
398 port = cpu_portfn(m->m_pkthdr.hash);
399 KASSERT(&curthread->td_msgport == port, ("wrong msgport\n"));
400
401 KASSERT((num > 0 && num <= NELEM(netisrs)), ("bad isr %d", num));
402 ni = &netisrs[num];
403 if (ni->ni_handler == NULL) {
404 kprintf("unregistered isr %d\n", num);
405 m_freem(m);
406 return EIO;
407 }
408
409 /*
410 * Initialize the netmsg, and run the handler directly.
411 */
412 pmsg = &m->m_hdr.mh_netmsg;
413 netmsg_init(&pmsg->base, NULL, &netisr_apanic_rport,
414 0, ni->ni_handler);
415 pmsg->nm_packet = m;
416 pmsg->base.lmsg.u.ms_result = num;
417 ni->ni_handler((netmsg_t)&pmsg->base);
418
419 return 0;
420}
421
422/*
c3c96e44
MD
423 * Pre-characterization of a deeper portion of the packet for the
424 * requested isr.
425 *
426 * The base of the ISR type (e.g. IP) that we want to characterize is
427 * at (hoff) relative to the beginning of the mbuf. This allows
428 * e.g. ether_input_chain() to not have to adjust the m_data/m_len.
8bde602d 429 */
c3c96e44
MD
430void
431netisr_characterize(int num, struct mbuf **mp, int hoff)
ef0fdad1 432{
c3c96e44
MD
433 struct netisr *ni;
434 struct mbuf *m;
435
436 /*
437 * Validation
438 */
c3c96e44
MD
439 m = *mp;
440 KKASSERT(m != NULL);
441
e6318d16
MD
442 if (num < 0 || num >= NETISR_MAX) {
443 if (num == NETISR_MAX) {
444 m->m_flags |= M_HASH;
445 m->m_pkthdr.hash = 0;
446 return;
447 }
448 panic("Bad isr %d", num);
449 }
450
c3c96e44
MD
451 /*
452 * Valid netisr?
453 */
454 ni = &netisrs[num];
455 if (ni->ni_handler == NULL) {
456 kprintf("Unregistered isr %d\n", num);
457 m_freem(m);
458 *mp = NULL;
459 }
460
461 /*
462 * Characterize the packet
463 */
464 if ((m->m_flags & M_HASH) == 0) {
465 ni->ni_cpufn(mp, hoff);
466 m = *mp;
467 if (m && (m->m_flags & M_HASH) == 0)
468 kprintf("netisr_queue(%d): packet hash failed\n", num);
469 }
8bde602d
JH
470}
471
bf82f9b7 472void
c3c96e44 473netisr_register(int num, netisr_fn_t handler, netisr_cpufn_t cpufn)
8bde602d 474{
c3c96e44 475 struct netisr *ni;
ef0fdad1 476
c157ff7a 477 KASSERT((num > 0 && num <= NELEM(netisrs)),
c3c96e44
MD
478 ("netisr_register: bad isr %d", num));
479 KKASSERT(handler != NULL);
480
481 if (cpufn == NULL)
482 cpufn = cpu0_cpufn;
8bde602d 483
c3c96e44
MD
484 ni = &netisrs[num];
485
486 ni->ni_handler = handler;
e6f77b88 487 ni->ni_hashck = netisr_nohashck;
c3c96e44
MD
488 ni->ni_cpufn = cpufn;
489 netmsg_init(&ni->ni_netmsg, NULL, &netisr_adone_rport, 0, NULL);
ef0fdad1 490}
bf82f9b7 491
c3c96e44 492void
e6f77b88
SZ
493netisr_register_hashcheck(int num, netisr_hashck_t hashck)
494{
495 struct netisr *ni;
496
497 KASSERT((num > 0 && num <= NELEM(netisrs)),
498 ("netisr_register: bad isr %d", num));
499
500 ni = &netisrs[num];
501 ni->ni_hashck = hashck;
502}
503
504void
c3c96e44 505netisr_register_rollup(netisr_ru_t ru_func)
bf82f9b7 506{
c3c96e44 507 struct netmsg_rollup *ru;
97a43e72 508
c3c96e44
MD
509 ru = kmalloc(sizeof(*ru), M_TEMP, M_WAITOK|M_ZERO);
510 ru->ru_func = ru_func;
511 TAILQ_INSERT_TAIL(&netrulist, ru, ru_entry);
bf82f9b7
MD
512}
513
c3c96e44
MD
514/*
515 * Return the message port for the general protocol message servicing
516 * thread for a particular cpu.
517 */
ecdefdda
MD
518lwkt_port_t
519cpu_portfn(int cpu)
520{
c3c96e44
MD
521 KKASSERT(cpu >= 0 && cpu < ncpus);
522 return (&netisr_cpu[cpu].td_msgport);
ecdefdda
MD
523}
524
c244d613 525/*
c3c96e44 526 * Return the current cpu's network protocol thread.
c244d613
SZ
527 */
528lwkt_port_t
529cur_netport(void)
530{
c3c96e44 531 return(cpu_portfn(mycpu->gd_cpuid));
c244d613
SZ
532}
533
c3c96e44 534/*
c3c96e44
MD
535 * Return a default protocol control message processing thread port
536 */
3efe7008 537lwkt_port_t
e3873585
SZ
538cpu0_ctlport(int cmd __unused, struct sockaddr *sa __unused,
539 void *extra __unused)
540{
c3c96e44 541 return (&netisr_cpu[0].td_msgport);
e3873585
SZ
542}
543
c3c96e44 544/*
c3c96e44
MD
545 * This is a default netisr packet characterization function which
546 * sets M_HASH. If a netisr is registered with a NULL cpufn function
547 * this one is assigned.
548 *
549 * This function makes no attempt to validate the packet.
550 */
551static void
552cpu0_cpufn(struct mbuf **mp, int hoff __unused)
553{
554 struct mbuf *m = *mp;
555
556 m->m_flags |= M_HASH;
557 m->m_pkthdr.hash = 0;
3efe7008
MD
558}
559
bf82f9b7 560/*
a29576fc 561 * schednetisr() is used to call the netisr handler from the appropriate
9eeaa8a9 562 * netisr thread for polling and other purposes.
a29576fc
MD
563 *
564 * This function may be called from a hard interrupt or IPI and must be
565 * MP SAFE and non-blocking. We use a fixed per-cpu message instead of
566 * trying to allocate one. We must get ourselves onto the target cpu
567 * to safely check the MSGF_DONE bit on the message but since the message
568 * will be sent to that cpu anyway this does not add any extra work beyond
569 * what lwkt_sendmsg() would have already had to do to schedule the target
570 * thread.
bf82f9b7 571 */
a29576fc
MD
572static void
573schednetisr_remote(void *data)
bf82f9b7 574{
c3c96e44
MD
575 int num = (int)(intptr_t)data;
576 struct netisr *ni = &netisrs[num];
577 lwkt_port_t port = &netisr_cpu[0].td_msgport;
002c1265 578 netmsg_base_t pmsg;
c3c96e44
MD
579
580 pmsg = &netisrs[num].ni_netmsg;
002c1265 581 if (pmsg->lmsg.ms_flags & MSGF_DONE) {
c3c96e44 582 netmsg_init(pmsg, NULL, &netisr_adone_rport, 0, ni->ni_handler);
002c1265
MD
583 pmsg->lmsg.u.ms_result = num;
584 lwkt_sendmsg(port, &pmsg->lmsg);
c3c96e44 585 }
a29576fc
MD
586}
587
588void
589schednetisr(int num)
590{
c157ff7a 591 KASSERT((num > 0 && num <= NELEM(netisrs)),
c3c96e44
MD
592 ("schednetisr: bad isr %d", num));
593 KKASSERT(netisrs[num].ni_handler != NULL);
a29576fc 594#ifdef SMP
c3c96e44
MD
595 if (mycpu->gd_cpuid != 0) {
596 lwkt_send_ipiq(globaldata_find(0),
597 schednetisr_remote, (void *)(intptr_t)num);
598 } else {
599 crit_enter();
600 schednetisr_remote((void *)(intptr_t)num);
601 crit_exit();
602 }
a29576fc 603#else
c3c96e44
MD
604 crit_enter();
605 schednetisr_remote((void *)(intptr_t)num);
606 crit_exit();
a29576fc 607#endif
bf82f9b7 608}
a91f9815
SZ
609
610#ifdef SMP
611
612static void
613netisr_barrier_dispatch(netmsg_t nmsg)
614{
615 struct netmsg_barrier *msg = (struct netmsg_barrier *)nmsg;
616
617 atomic_clear_cpumask(msg->br_cpumask, mycpu->gd_cpumask);
618 if (*msg->br_cpumask == 0)
619 wakeup(msg->br_cpumask);
620
d0c7a72a
SZ
621 for (;;) {
622 uint32_t done = msg->br_done;
623
ca3321f8 624 cpu_ccfence();
d0c7a72a
SZ
625 if ((done & NETISR_BR_NOTDONE) == 0)
626 break;
627
ca3321f8
SZ
628 tsleep_interlock(&msg->br_done, 0);
629 if (atomic_cmpset_int(&msg->br_done,
d0c7a72a 630 done, done | NETISR_BR_WAITDONE))
ca3321f8
SZ
631 tsleep(&msg->br_done, PINTERLOCKED, "nbrdsp", 0);
632 }
a91f9815
SZ
633
634 lwkt_replymsg(&nmsg->lmsg, 0);
635}
636
637#endif
638
639struct netisr_barrier *
640netisr_barrier_create(void)
641{
642 struct netisr_barrier *br;
643
644 br = kmalloc(sizeof(*br), M_LWKTMSG, M_WAITOK | M_ZERO);
645 return br;
646}
647
648void
649netisr_barrier_set(struct netisr_barrier *br)
650{
651#ifdef SMP
0503d1d0 652 volatile cpumask_t other_cpumask;
a91f9815
SZ
653 int i, cur_cpuid;
654
655 KKASSERT(&curthread->td_msgport == cpu_portfn(0));
656 KKASSERT(!br->br_isset);
657
658 other_cpumask = mycpu->gd_other_cpus & smp_active_mask;
659 cur_cpuid = mycpuid;
660
661 for (i = 0; i < ncpus; ++i) {
662 struct netmsg_barrier *msg;
663
664 if (i == cur_cpuid)
665 continue;
666
667 msg = kmalloc(sizeof(struct netmsg_barrier),
668 M_LWKTMSG, M_WAITOK);
669 netmsg_init(&msg->base, NULL, &netisr_afree_rport,
670 MSGF_PRIORITY, netisr_barrier_dispatch);
671 msg->br_cpumask = &other_cpumask;
ca3321f8 672 msg->br_done = NETISR_BR_NOTDONE;
a91f9815
SZ
673
674 KKASSERT(br->br_msgs[i] == NULL);
675 br->br_msgs[i] = msg;
676 }
677
678 for (i = 0; i < ncpus; ++i) {
679 if (i == cur_cpuid)
680 continue;
681 lwkt_sendmsg(cpu_portfn(i), &br->br_msgs[i]->base.lmsg);
682 }
683
684 while (other_cpumask != 0) {
685 tsleep_interlock(&other_cpumask, 0);
686 if (other_cpumask != 0)
687 tsleep(&other_cpumask, PINTERLOCKED, "nbrset", 0);
688 }
689#endif
690 br->br_isset = 1;
691}
692
693void
694netisr_barrier_rem(struct netisr_barrier *br)
695{
696#ifdef SMP
697 int i, cur_cpuid;
698
699 KKASSERT(&curthread->td_msgport == cpu_portfn(0));
700 KKASSERT(br->br_isset);
701
702 cur_cpuid = mycpuid;
703 for (i = 0; i < ncpus; ++i) {
704 struct netmsg_barrier *msg = br->br_msgs[i];
d0c7a72a 705 uint32_t done;
a91f9815
SZ
706
707 msg = br->br_msgs[i];
708 br->br_msgs[i] = NULL;
709
710 if (i == cur_cpuid)
711 continue;
712
d0c7a72a
SZ
713 done = atomic_swap_int(&msg->br_done, 0);
714 if (done & NETISR_BR_WAITDONE)
715 wakeup(&msg->br_done);
a91f9815
SZ
716 }
717#endif
718 br->br_isset = 0;
719}
e6f77b88
SZ
720
721static void
722netisr_nohashck(struct mbuf *m, const struct pktinfo *pi __unused)
723{
724 m->m_flags &= ~M_HASH;
725}
726
727void
728netisr_hashcheck(int num, struct mbuf *m, const struct pktinfo *pi)
729{
730 struct netisr *ni;
731
732 if (num < 0 || num >= NETISR_MAX)
733 panic("Bad isr %d", num);
734
735 /*
736 * Valid netisr?
737 */
738 ni = &netisrs[num];
739 if (ni->ni_handler == NULL)
740 panic("Unregistered isr %d\n", num);
741
742 ni->ni_hashck(m, pi);
743}