netisr: Add priority for netisr "rollup" functions
[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
c3c96e44
MD
60static void netmsg_service_loop(void *arg);
61static void cpu0_cpufn(struct mbuf **mp, int hoff);
e6f77b88 62static void netisr_nohashck(struct mbuf *, const struct pktinfo *);
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;
239bdb58 72 int ru_prio;
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
ef0fdad1 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
79f504ca 255 netmsg_init(&smsg, NULL, &curthread->td_msgport, 0, netmsg_sync_handler);
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 */
79f504ca
SZ
266void
267netmsg_sync_handler(netmsg_t msg)
5c703385 268{
002c1265 269 lwkt_replymsg(&msg->lmsg, 0);
bf82f9b7 270}
8bde602d 271
95f8b5ce 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;
3598cc14 281 thread_t td = curthread;
c3c96e44
MD
282 int limit;
283
f256b6c0
SZ
284 td->td_type = TD_TYPE_NETISR;
285
c3c96e44
MD
286 while ((msg = lwkt_waitport(&td->td_msgport, 0))) {
287 /*
288 * Run up to 512 pending netmsgs.
289 */
290 limit = 512;
291 do {
292 KASSERT(msg->nm_dispatch != NULL,
ed20d0e3 293 ("netmsg_service isr %d badmsg",
002c1265 294 msg->lmsg.u.ms_result));
0ce0603e
MD
295 if (msg->nm_so &&
296 msg->nm_so->so_port != &td->td_msgport) {
297 /*
298 * Sockets undergoing connect or disconnect
299 * ops can change ports on us. Chase the
300 * port.
301 */
302 kprintf("netmsg_service_loop: Warning, "
303 "port changed so=%p\n", msg->nm_so);
304 lwkt_forwardmsg(msg->nm_so->so_port,
002c1265 305 &msg->lmsg);
0ce0603e
MD
306 } else {
307 /*
308 * We are on the correct port, dispatch it.
309 */
002c1265 310 msg->nm_dispatch((netmsg_t)msg);
0ce0603e 311 }
c3c96e44
MD
312 if (--limit == 0)
313 break;
314 } while ((msg = lwkt_getport(&td->td_msgport)) != NULL);
315
316 /*
317 * Run all registered rollup functions for this cpu
318 * (e.g. tcp_willblock()).
319 */
320 TAILQ_FOREACH(ru, &netrulist, ru_entry)
321 ru->ru_func();
322 }
95f8b5ce
SZ
323}
324
bf82f9b7 325/*
c3c96e44
MD
326 * Forward a packet to a netisr service function.
327 *
328 * If the packet has not been assigned to a protocol thread we call
329 * the port characterization function to assign it. The caller must
330 * clear M_HASH (or not have set it in the first place) if the caller
331 * wishes the packet to be recharacterized.
bf82f9b7 332 */
c3c96e44
MD
333int
334netisr_queue(int num, struct mbuf *m)
bf82f9b7 335{
c3c96e44
MD
336 struct netisr *ni;
337 struct netmsg_packet *pmsg;
338 lwkt_port_t port;
339
c157ff7a 340 KASSERT((num > 0 && num <= NELEM(netisrs)),
c3c96e44
MD
341 ("Bad isr %d", num));
342
343 ni = &netisrs[num];
344 if (ni->ni_handler == NULL) {
345 kprintf("Unregistered isr %d\n", num);
346 m_freem(m);
347 return (EIO);
348 }
349
350 /*
351 * Figure out which protocol thread to send to. This does not
352 * have to be perfect but performance will be really good if it
353 * is correct. Major protocol inputs such as ip_input() will
354 * re-characterize the packet as necessary.
355 */
356 if ((m->m_flags & M_HASH) == 0) {
357 ni->ni_cpufn(&m, 0);
358 if (m == NULL) {
359 m_freem(m);
360 return (EIO);
361 }
362 if ((m->m_flags & M_HASH) == 0) {
363 kprintf("netisr_queue(%d): packet hash failed\n", num);
364 m_freem(m);
365 return (EIO);
366 }
367 }
368
369 /*
370 * Get the protocol port based on the packet hash, initialize
371 * the netmsg, and send it off.
372 */
3abced87 373 port = netisr_portfn(m->m_pkthdr.hash);
c3c96e44 374 pmsg = &m->m_hdr.mh_netmsg;
002c1265 375 netmsg_init(&pmsg->base, NULL, &netisr_apanic_rport,
c3c96e44
MD
376 0, ni->ni_handler);
377 pmsg->nm_packet = m;
002c1265
MD
378 pmsg->base.lmsg.u.ms_result = num;
379 lwkt_sendmsg(port, &pmsg->base.lmsg);
c3c96e44
MD
380
381 return (0);
8bde602d
JH
382}
383
ebe4c2ae
SZ
384/*
385 * Run a netisr service function on the packet.
386 *
387 * The packet must have been correctly characterized!
388 */
389int
390netisr_handle(int num, struct mbuf *m)
391{
392 struct netisr *ni;
393 struct netmsg_packet *pmsg;
394 lwkt_port_t port;
395
396 /*
397 * Get the protocol port based on the packet hash
398 */
ed20d0e3 399 KASSERT((m->m_flags & M_HASH), ("packet not characterized"));
3abced87 400 port = netisr_portfn(m->m_pkthdr.hash);
ed20d0e3 401 KASSERT(&curthread->td_msgport == port, ("wrong msgport"));
ebe4c2ae
SZ
402
403 KASSERT((num > 0 && num <= NELEM(netisrs)), ("bad isr %d", num));
404 ni = &netisrs[num];
405 if (ni->ni_handler == NULL) {
406 kprintf("unregistered isr %d\n", num);
407 m_freem(m);
408 return EIO;
409 }
410
411 /*
412 * Initialize the netmsg, and run the handler directly.
413 */
414 pmsg = &m->m_hdr.mh_netmsg;
415 netmsg_init(&pmsg->base, NULL, &netisr_apanic_rport,
416 0, ni->ni_handler);
417 pmsg->nm_packet = m;
418 pmsg->base.lmsg.u.ms_result = num;
419 ni->ni_handler((netmsg_t)&pmsg->base);
420
421 return 0;
422}
423
8bde602d 424/*
c3c96e44
MD
425 * Pre-characterization of a deeper portion of the packet for the
426 * requested isr.
427 *
428 * The base of the ISR type (e.g. IP) that we want to characterize is
429 * at (hoff) relative to the beginning of the mbuf. This allows
430 * e.g. ether_input_chain() to not have to adjust the m_data/m_len.
8bde602d 431 */
c3c96e44
MD
432void
433netisr_characterize(int num, struct mbuf **mp, int hoff)
ef0fdad1 434{
c3c96e44
MD
435 struct netisr *ni;
436 struct mbuf *m;
437
438 /*
439 * Validation
440 */
c3c96e44
MD
441 m = *mp;
442 KKASSERT(m != NULL);
443
e6318d16
MD
444 if (num < 0 || num >= NETISR_MAX) {
445 if (num == NETISR_MAX) {
446 m->m_flags |= M_HASH;
447 m->m_pkthdr.hash = 0;
448 return;
449 }
450 panic("Bad isr %d", num);
451 }
452
c3c96e44
MD
453 /*
454 * Valid netisr?
455 */
456 ni = &netisrs[num];
457 if (ni->ni_handler == NULL) {
458 kprintf("Unregistered isr %d\n", num);
459 m_freem(m);
460 *mp = NULL;
461 }
462
463 /*
464 * Characterize the packet
465 */
466 if ((m->m_flags & M_HASH) == 0) {
467 ni->ni_cpufn(mp, hoff);
468 m = *mp;
469 if (m && (m->m_flags & M_HASH) == 0)
470 kprintf("netisr_queue(%d): packet hash failed\n", num);
471 }
8bde602d
JH
472}
473
bf82f9b7 474void
c3c96e44 475netisr_register(int num, netisr_fn_t handler, netisr_cpufn_t cpufn)
8bde602d 476{
c3c96e44 477 struct netisr *ni;
ef0fdad1 478
c157ff7a 479 KASSERT((num > 0 && num <= NELEM(netisrs)),
c3c96e44
MD
480 ("netisr_register: bad isr %d", num));
481 KKASSERT(handler != NULL);
482
483 if (cpufn == NULL)
484 cpufn = cpu0_cpufn;
8bde602d 485
c3c96e44
MD
486 ni = &netisrs[num];
487
488 ni->ni_handler = handler;
e6f77b88 489 ni->ni_hashck = netisr_nohashck;
c3c96e44
MD
490 ni->ni_cpufn = cpufn;
491 netmsg_init(&ni->ni_netmsg, NULL, &netisr_adone_rport, 0, NULL);
ef0fdad1 492}
bf82f9b7 493
e6f77b88
SZ
494void
495netisr_register_hashcheck(int num, netisr_hashck_t hashck)
496{
497 struct netisr *ni;
498
499 KASSERT((num > 0 && num <= NELEM(netisrs)),
500 ("netisr_register: bad isr %d", num));
501
502 ni = &netisrs[num];
503 ni->ni_hashck = hashck;
504}
505
c3c96e44 506void
239bdb58 507netisr_register_rollup(netisr_ru_t ru_func, int prio)
bf82f9b7 508{
239bdb58
SZ
509 struct netmsg_rollup *new_ru, *ru;
510
511 new_ru = kmalloc(sizeof(*new_ru), M_TEMP, M_WAITOK|M_ZERO);
512 new_ru->ru_func = ru_func;
513 new_ru->ru_prio = prio;
97a43e72 514
239bdb58
SZ
515 /*
516 * Higher priority "rollup" appears first
517 */
518 TAILQ_FOREACH(ru, &netrulist, ru_entry) {
519 if (ru->ru_prio < new_ru->ru_prio) {
520 TAILQ_INSERT_BEFORE(ru, new_ru, ru_entry);
521 return;
522 }
523 }
524 TAILQ_INSERT_TAIL(&netrulist, new_ru, ru_entry);
bf82f9b7
MD
525}
526
c3c96e44
MD
527/*
528 * Return the message port for the general protocol message servicing
529 * thread for a particular cpu.
530 */
ecdefdda 531lwkt_port_t
3abced87 532netisr_portfn(int cpu)
ecdefdda 533{
c3c96e44
MD
534 KKASSERT(cpu >= 0 && cpu < ncpus);
535 return (&netisr_cpu[cpu].td_msgport);
ecdefdda
MD
536}
537
c244d613 538/*
c3c96e44 539 * Return the current cpu's network protocol thread.
c244d613
SZ
540 */
541lwkt_port_t
542cur_netport(void)
543{
3abced87 544 return(netisr_portfn(mycpu->gd_cpuid));
c244d613
SZ
545}
546
c3c96e44
MD
547/*
548 * Return a default protocol control message processing thread port
549 */
e3873585
SZ
550lwkt_port_t
551cpu0_ctlport(int cmd __unused, struct sockaddr *sa __unused,
552 void *extra __unused)
553{
c3c96e44 554 return (&netisr_cpu[0].td_msgport);
e3873585
SZ
555}
556
c3c96e44
MD
557/*
558 * This is a default netisr packet characterization function which
559 * sets M_HASH. If a netisr is registered with a NULL cpufn function
560 * this one is assigned.
561 *
562 * This function makes no attempt to validate the packet.
563 */
564static void
565cpu0_cpufn(struct mbuf **mp, int hoff __unused)
566{
567 struct mbuf *m = *mp;
568
569 m->m_flags |= M_HASH;
570 m->m_pkthdr.hash = 0;
3efe7008
MD
571}
572
bf82f9b7 573/*
a29576fc 574 * schednetisr() is used to call the netisr handler from the appropriate
9eeaa8a9 575 * netisr thread for polling and other purposes.
a29576fc
MD
576 *
577 * This function may be called from a hard interrupt or IPI and must be
578 * MP SAFE and non-blocking. We use a fixed per-cpu message instead of
579 * trying to allocate one. We must get ourselves onto the target cpu
580 * to safely check the MSGF_DONE bit on the message but since the message
581 * will be sent to that cpu anyway this does not add any extra work beyond
582 * what lwkt_sendmsg() would have already had to do to schedule the target
583 * thread.
bf82f9b7 584 */
a29576fc
MD
585static void
586schednetisr_remote(void *data)
bf82f9b7 587{
c3c96e44
MD
588 int num = (int)(intptr_t)data;
589 struct netisr *ni = &netisrs[num];
590 lwkt_port_t port = &netisr_cpu[0].td_msgport;
002c1265 591 netmsg_base_t pmsg;
c3c96e44
MD
592
593 pmsg = &netisrs[num].ni_netmsg;
002c1265 594 if (pmsg->lmsg.ms_flags & MSGF_DONE) {
c3c96e44 595 netmsg_init(pmsg, NULL, &netisr_adone_rport, 0, ni->ni_handler);
002c1265
MD
596 pmsg->lmsg.u.ms_result = num;
597 lwkt_sendmsg(port, &pmsg->lmsg);
c3c96e44 598 }
a29576fc
MD
599}
600
601void
602schednetisr(int num)
603{
c157ff7a 604 KASSERT((num > 0 && num <= NELEM(netisrs)),
c3c96e44
MD
605 ("schednetisr: bad isr %d", num));
606 KKASSERT(netisrs[num].ni_handler != NULL);
c3c96e44
MD
607 if (mycpu->gd_cpuid != 0) {
608 lwkt_send_ipiq(globaldata_find(0),
609 schednetisr_remote, (void *)(intptr_t)num);
610 } else {
611 crit_enter();
612 schednetisr_remote((void *)(intptr_t)num);
613 crit_exit();
614 }
bf82f9b7 615}
a91f9815 616
a91f9815
SZ
617static void
618netisr_barrier_dispatch(netmsg_t nmsg)
619{
620 struct netmsg_barrier *msg = (struct netmsg_barrier *)nmsg;
621
622 atomic_clear_cpumask(msg->br_cpumask, mycpu->gd_cpumask);
623 if (*msg->br_cpumask == 0)
624 wakeup(msg->br_cpumask);
625
d0c7a72a
SZ
626 for (;;) {
627 uint32_t done = msg->br_done;
628
ca3321f8 629 cpu_ccfence();
d0c7a72a
SZ
630 if ((done & NETISR_BR_NOTDONE) == 0)
631 break;
632
ca3321f8
SZ
633 tsleep_interlock(&msg->br_done, 0);
634 if (atomic_cmpset_int(&msg->br_done,
d0c7a72a 635 done, done | NETISR_BR_WAITDONE))
ca3321f8
SZ
636 tsleep(&msg->br_done, PINTERLOCKED, "nbrdsp", 0);
637 }
a91f9815
SZ
638
639 lwkt_replymsg(&nmsg->lmsg, 0);
640}
641
a91f9815
SZ
642struct netisr_barrier *
643netisr_barrier_create(void)
644{
645 struct netisr_barrier *br;
646
647 br = kmalloc(sizeof(*br), M_LWKTMSG, M_WAITOK | M_ZERO);
648 return br;
649}
650
651void
652netisr_barrier_set(struct netisr_barrier *br)
653{
0503d1d0 654 volatile cpumask_t other_cpumask;
a91f9815
SZ
655 int i, cur_cpuid;
656
3abced87 657 KKASSERT(&curthread->td_msgport == netisr_portfn(0));
a91f9815
SZ
658 KKASSERT(!br->br_isset);
659
660 other_cpumask = mycpu->gd_other_cpus & smp_active_mask;
661 cur_cpuid = mycpuid;
662
663 for (i = 0; i < ncpus; ++i) {
664 struct netmsg_barrier *msg;
665
666 if (i == cur_cpuid)
667 continue;
668
669 msg = kmalloc(sizeof(struct netmsg_barrier),
670 M_LWKTMSG, M_WAITOK);
671 netmsg_init(&msg->base, NULL, &netisr_afree_rport,
672 MSGF_PRIORITY, netisr_barrier_dispatch);
673 msg->br_cpumask = &other_cpumask;
ca3321f8 674 msg->br_done = NETISR_BR_NOTDONE;
a91f9815
SZ
675
676 KKASSERT(br->br_msgs[i] == NULL);
677 br->br_msgs[i] = msg;
678 }
679
680 for (i = 0; i < ncpus; ++i) {
681 if (i == cur_cpuid)
682 continue;
3abced87 683 lwkt_sendmsg(netisr_portfn(i), &br->br_msgs[i]->base.lmsg);
a91f9815
SZ
684 }
685
686 while (other_cpumask != 0) {
687 tsleep_interlock(&other_cpumask, 0);
688 if (other_cpumask != 0)
689 tsleep(&other_cpumask, PINTERLOCKED, "nbrset", 0);
690 }
a91f9815
SZ
691 br->br_isset = 1;
692}
693
694void
695netisr_barrier_rem(struct netisr_barrier *br)
696{
a91f9815
SZ
697 int i, cur_cpuid;
698
3abced87 699 KKASSERT(&curthread->td_msgport == netisr_portfn(0));
a91f9815
SZ
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 716 }
a91f9815
SZ
717 br->br_isset = 0;
718}
e6f77b88
SZ
719
720static void
721netisr_nohashck(struct mbuf *m, const struct pktinfo *pi __unused)
722{
723 m->m_flags &= ~M_HASH;
724}
725
726void
727netisr_hashcheck(int num, struct mbuf *m, const struct pktinfo *pi)
728{
729 struct netisr *ni;
730
731 if (num < 0 || num >= NETISR_MAX)
732 panic("Bad isr %d", num);
733
734 /*
735 * Valid netisr?
736 */
737 ni = &netisrs[num];
738 if (ni->ni_handler == NULL)
ed20d0e3 739 panic("Unregistered isr %d", num);
e6f77b88
SZ
740
741 ni->ni_hashck(m, pi);
742}