Add TDF_NETWORK lwkt flag, so various assertion could be performed to make sure
[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.
37 *
cbb1e737 38 * $DragonFly: src/sys/net/netisr.c,v 1.45 2008/09/20 04:31:02 sephe Exp $
ef0fdad1
MD
39 */
40
41#include <sys/param.h>
42#include <sys/systm.h>
bf82f9b7 43#include <sys/kernel.h>
9eeaa8a9 44#include <sys/malloc.h>
bf82f9b7 45#include <sys/msgport.h>
ef0fdad1
MD
46#include <sys/proc.h>
47#include <sys/interrupt.h>
8bde602d
JH
48#include <sys/socket.h>
49#include <sys/sysctl.h>
50#include <net/if.h>
51#include <net/if_var.h>
ef0fdad1
MD
52#include <net/netisr.h>
53#include <machine/cpufunc.h>
ef0fdad1 54
3227f1b8
MD
55#include <sys/thread2.h>
56#include <sys/msgport2.h>
4599cf19 57#include <net/netmsg2.h>
3227f1b8 58
4599cf19 59static void netmsg_sync_func(struct netmsg *msg);
5c703385
MD
60
61struct netmsg_port_registration {
62 TAILQ_ENTRY(netmsg_port_registration) npr_entry;
63 lwkt_port_t npr_port;
64};
65
bf82f9b7 66static struct netisr netisrs[NETISR_MAX];
5c703385 67static TAILQ_HEAD(,netmsg_port_registration) netreglist;
bf82f9b7
MD
68
69/* Per-CPU thread to handle any protocol. */
70struct thread netisr_cpu[MAXCPU];
3227f1b8 71lwkt_port netisr_afree_rport;
a29576fc 72lwkt_port netisr_adone_rport;
6aad077d 73lwkt_port netisr_apanic_rport;
3efe7008 74lwkt_port netisr_sync_port;
3227f1b8 75
fb0f29c4
MD
76static int (*netmsg_fwd_port_fn)(lwkt_port_t, lwkt_msg_t);
77
37eacd1a
SZ
78static int netisr_mpsafe_thread = 0;
79TUNABLE_INT("netisr.mpsafe_thread", &netisr_mpsafe_thread);
80
3227f1b8
MD
81/*
82 * netisr_afree_rport replymsg function, only used to handle async
83 * messages which the sender has abandoned to their fate.
84 */
85static void
86netisr_autofree_reply(lwkt_port_t port, lwkt_msg_t msg)
87{
efda3bd0 88 kfree(msg, M_LWKTMSG);
3227f1b8 89}
ef0fdad1 90
dc22b3aa 91/*
fb0f29c4
MD
92 * We need a custom putport function to handle the case where the
93 * message target is the current thread's message port. This case
94 * can occur when the TCP or UDP stack does a direct callback to NFS and NFS
95 * then turns around and executes a network operation synchronously.
3efe7008 96 *
fb0f29c4
MD
97 * To prevent deadlocking, we must execute these self-referential messages
98 * synchronously, effectively turning the message into a glorified direct
99 * procedure call back into the protocol stack. The operation must be
100 * complete on return or we will deadlock, so panic if it isn't.
dc22b3aa 101 */
5c703385 102static int
dc22b3aa
JH
103netmsg_put_port(lwkt_port_t port, lwkt_msg_t lmsg)
104{
4599cf19 105 netmsg_t netmsg = (void *)lmsg;
dc22b3aa 106
fb0f29c4 107 if ((lmsg->ms_flags & MSGF_SYNC) && port == &curthread->td_msgport) {
4599cf19
MD
108 netmsg->nm_dispatch(netmsg);
109 if ((lmsg->ms_flags & MSGF_DONE) == 0)
dc22b3aa 110 panic("netmsg_put_port: self-referential deadlock on netport");
4599cf19 111 return(EASYNC);
dc22b3aa 112 } else {
fb0f29c4 113 return(netmsg_fwd_port_fn(port, lmsg));
dc22b3aa
JH
114 }
115}
116
3efe7008
MD
117/*
118 * UNIX DOMAIN sockets still have to run their uipc functions synchronously,
119 * because they depend on the user proc context for a number of things
120 * (like creds) which we have not yet incorporated into the message structure.
121 *
122 * However, we maintain or message/port abstraction. Having a special
123 * synchronous port which runs the commands synchronously gives us the
124 * ability to serialize operations in one place later on when we start
125 * removing the BGL.
3efe7008
MD
126 */
127static int
128netmsg_sync_putport(lwkt_port_t port, lwkt_msg_t lmsg)
129{
4599cf19 130 netmsg_t netmsg = (void *)lmsg;
3efe7008 131
e0383bf3
MD
132 KKASSERT((lmsg->ms_flags & MSGF_DONE) == 0);
133
3efe7008 134 lmsg->ms_target_port = port; /* required for abort */
4599cf19 135 netmsg->nm_dispatch(netmsg);
e0383bf3 136 return(EASYNC);
3efe7008
MD
137}
138
ef0fdad1 139static void
bf82f9b7 140netisr_init(void)
ef0fdad1 141{
bf82f9b7
MD
142 int i;
143
5c703385
MD
144 TAILQ_INIT(&netreglist);
145
3efe7008
MD
146 /*
147 * Create default per-cpu threads for generic protocol handling.
148 */
3227f1b8 149 for (i = 0; i < ncpus; ++i) {
37eacd1a
SZ
150 lwkt_create(netisr_mpsafe_thread ?
151 netmsg_service_loop_mpsafe : netmsg_service_loop,
152 NULL, NULL, &netisr_cpu[i],
cbb1e737 153 TDF_NETWORK, i, "netisr_cpu %d", i);
5c703385 154 netmsg_service_port_init(&netisr_cpu[i].td_msgport);
3227f1b8 155 }
3efe7008
MD
156
157 /*
158 * The netisr_afree_rport is a special reply port which automatically
6aad077d
MD
159 * frees the replied message. The netisr_adone_rport simply marks
160 * the message as being done. The netisr_apanic_rport panics if
161 * the message is replied to.
3efe7008 162 */
fb0f29c4
MD
163 lwkt_initport_replyonly(&netisr_afree_rport, netisr_autofree_reply);
164 lwkt_initport_replyonly_null(&netisr_adone_rport);
165 lwkt_initport_panic(&netisr_apanic_rport);
3efe7008
MD
166
167 /*
168 * The netisr_syncport is a special port which executes the message
169 * synchronously and waits for it if EASYNC is returned.
170 */
fb0f29c4 171 lwkt_initport_putonly(&netisr_sync_port, netmsg_sync_putport);
ef0fdad1
MD
172}
173
b2632176 174SYSINIT(netisr, SI_SUB_PRE_DRIVERS, SI_ORDER_FIRST, netisr_init, NULL);
bf82f9b7 175
5c703385
MD
176/*
177 * Finish initializing the message port for a netmsg service. This also
178 * registers the port for synchronous cleanup operations such as when an
179 * ifnet is being destroyed. There is no deregistration API yet.
180 */
181void
182netmsg_service_port_init(lwkt_port_t port)
183{
184 struct netmsg_port_registration *reg;
185
186 /*
187 * Override the putport function. Our custom function checks for
188 * self-references and executes such commands synchronously.
189 */
fb0f29c4
MD
190 if (netmsg_fwd_port_fn == NULL)
191 netmsg_fwd_port_fn = port->mp_putport;
192 KKASSERT(netmsg_fwd_port_fn == port->mp_putport);
5c703385
MD
193 port->mp_putport = netmsg_put_port;
194
195 /*
196 * Keep track of ports using the netmsg API so we can synchronize
197 * certain operations (such as freeing an ifnet structure) across all
198 * consumers.
199 */
efda3bd0 200 reg = kmalloc(sizeof(*reg), M_TEMP, M_WAITOK|M_ZERO);
5c703385
MD
201 reg->npr_port = port;
202 TAILQ_INSERT_TAIL(&netreglist, reg, npr_entry);
203}
204
205/*
206 * This function synchronizes the caller with all netmsg services. For
207 * example, if an interface is being removed we must make sure that all
208 * packets related to that interface complete processing before the structure
209 * can actually be freed. This sort of synchronization is an alternative to
210 * ref-counting the netif, removing the ref counting overhead in favor of
211 * placing additional overhead in the netif freeing sequence (where it is
212 * inconsequential).
213 */
214void
215netmsg_service_sync(void)
216{
217 struct netmsg_port_registration *reg;
218 struct netmsg smsg;
219
4599cf19 220 netmsg_init(&smsg, &curthread->td_msgport, 0, netmsg_sync_func);
5c703385
MD
221
222 TAILQ_FOREACH(reg, &netreglist, npr_entry) {
a22c590e 223 lwkt_domsg(reg->npr_port, &smsg.nm_lmsg, 0);
5c703385
MD
224 }
225}
226
227/*
228 * The netmsg function simply replies the message. API semantics require
229 * EASYNC to be returned if the netmsg function disposes of the message.
230 */
4599cf19 231static void
5c703385
MD
232netmsg_sync_func(struct netmsg *msg)
233{
234 lwkt_replymsg(&msg->nm_lmsg, 0);
5c703385
MD
235}
236
237/*
238 * Generic netmsg service loop. Some protocols may roll their own but all
239 * must do the basic command dispatch function call done here.
240 */
8bde602d 241void
bf82f9b7 242netmsg_service_loop(void *arg)
8bde602d 243{
12967093
MD
244 struct netmsg *msg;
245
9740d7df 246 while ((msg = lwkt_waitport(&curthread->td_msgport, 0))) {
4599cf19 247 msg->nm_dispatch(msg);
a29576fc 248 }
bf82f9b7 249}
8bde602d 250
95f8b5ce
SZ
251/*
252 * MPSAFE version of netmsg_service_loop()
253 */
254void
255netmsg_service_loop_mpsafe(void *arg)
256{
257 rel_mplock();
5007c9d2 258 netmsg_service_loop(arg);
95f8b5ce
SZ
259}
260
bf82f9b7
MD
261/*
262 * Call the netisr directly.
263 * Queueing may be done in the msg port layer at its discretion.
264 */
265void
266netisr_dispatch(int num, struct mbuf *m)
267{
268 /* just queue it for now XXX JH */
269 netisr_queue(num, m);
8bde602d
JH
270}
271
272/*
273 * Same as netisr_dispatch(), but always queue.
274 * This is either used in places where we are not confident that
275 * direct dispatch is possible, or where queueing is required.
276 */
ef0fdad1 277int
8bde602d 278netisr_queue(int num, struct mbuf *m)
ef0fdad1 279{
5944299a 280 struct netisr *ni;
9eeaa8a9 281 struct netmsg_packet *pmsg;
bf82f9b7 282 lwkt_port_t port;
8bde602d
JH
283
284 KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
ff86c8e0 285 ("%s: bad isr %d", __func__, num));
8bde602d 286
5944299a
MD
287 ni = &netisrs[num];
288 if (ni->ni_handler == NULL) {
ff86c8e0 289 kprintf("%s: unregistered isr %d\n", __func__, num);
68b67450 290 m_freem(m);
9eeaa8a9 291 return (EIO);
5944299a
MD
292 }
293
b01ae44a 294 if ((port = ni->ni_mport(&m)) == NULL)
9eeaa8a9
JH
295 return (EIO);
296
6aad077d 297 pmsg = &m->m_hdr.mh_netmsg;
8bde602d 298
4599cf19 299 netmsg_init(&pmsg->nm_netmsg, &netisr_apanic_rport, 0, ni->ni_handler);
bf82f9b7 300 pmsg->nm_packet = m;
4599cf19
MD
301 pmsg->nm_netmsg.nm_lmsg.u.ms_result = num;
302 lwkt_sendmsg(port, &pmsg->nm_netmsg.nm_lmsg);
8bde602d
JH
303 return (0);
304}
305
bf82f9b7
MD
306void
307netisr_register(int num, lwkt_portfn_t mportfn, netisr_fn_t handler)
8bde602d
JH
308{
309 KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
1748bf82 310 ("netisr_register: bad isr %d", num));
4599cf19 311 netmsg_init(&netisrs[num].ni_netmsg, &netisr_adone_rport, 0, NULL);
bf82f9b7 312 netisrs[num].ni_mport = mportfn;
8bde602d 313 netisrs[num].ni_handler = handler;
ef0fdad1
MD
314}
315
316int
8bde602d 317netisr_unregister(int num)
ef0fdad1 318{
8bde602d
JH
319 KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
320 ("unregister_netisr: bad isr number: %d\n", num));
321
bf82f9b7 322 /* XXX JH */
ef0fdad1
MD
323 return (0);
324}
bf82f9b7
MD
325
326/*
327 * Return message port for default handler thread on CPU 0.
328 */
329lwkt_port_t
b01ae44a 330cpu0_portfn(struct mbuf **mptr)
bf82f9b7
MD
331{
332 return (&netisr_cpu[0].td_msgport);
333}
334
ecdefdda
MD
335lwkt_port_t
336cpu_portfn(int cpu)
337{
338 return (&netisr_cpu[cpu].td_msgport);
339}
340
9eeaa8a9
JH
341/* ARGSUSED */
342lwkt_port_t
1c770505 343cpu0_soport(struct socket *so __unused, struct sockaddr *nam __unused,
934c6849 344 struct mbuf **dummy __unused, int req __unused)
9eeaa8a9
JH
345{
346 return (&netisr_cpu[0].td_msgport);
347}
348
3efe7008
MD
349lwkt_port_t
350sync_soport(struct socket *so __unused, struct sockaddr *nam __unused,
934c6849 351 struct mbuf **dummy __unused, int req __unused)
3efe7008
MD
352{
353 return (&netisr_sync_port);
354}
355
bf82f9b7 356/*
a29576fc 357 * schednetisr() is used to call the netisr handler from the appropriate
9eeaa8a9 358 * netisr thread for polling and other purposes.
a29576fc
MD
359 *
360 * This function may be called from a hard interrupt or IPI and must be
361 * MP SAFE and non-blocking. We use a fixed per-cpu message instead of
362 * trying to allocate one. We must get ourselves onto the target cpu
363 * to safely check the MSGF_DONE bit on the message but since the message
364 * will be sent to that cpu anyway this does not add any extra work beyond
365 * what lwkt_sendmsg() would have already had to do to schedule the target
366 * thread.
bf82f9b7 367 */
a29576fc
MD
368static void
369schednetisr_remote(void *data)
bf82f9b7 370{
a29576fc 371 int num = (int)data;
bf82f9b7 372 struct netisr *ni = &netisrs[num];
bf82f9b7 373 lwkt_port_t port = &netisr_cpu[0].td_msgport;
a29576fc 374 struct netmsg *pmsg;
bf82f9b7 375
a29576fc
MD
376 pmsg = &netisrs[num].ni_netmsg;
377 crit_enter();
378 if (pmsg->nm_lmsg.ms_flags & MSGF_DONE) {
4599cf19 379 netmsg_init(pmsg, &netisr_adone_rport, 0, ni->ni_handler);
a29576fc 380 pmsg->nm_lmsg.u.ms_result = num;
b76bed62
MD
381 lwkt_sendmsg(port, &pmsg->nm_lmsg);
382 }
a29576fc
MD
383 crit_exit();
384}
385
386void
387schednetisr(int num)
388{
389 KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
390 ("schednetisr: bad isr %d", num));
391#ifdef SMP
392 if (mycpu->gd_cpuid != 0)
393 lwkt_send_ipiq(globaldata_find(0), schednetisr_remote, (void *)num);
394 else
395 schednetisr_remote((void *)num);
396#else
397 schednetisr_remote((void *)num);
398#endif
bf82f9b7 399}
a29576fc 400
29bc1092
SZ
401lwkt_port_t
402netisr_find_port(int num, struct mbuf **m0)
403{
404 struct netisr *ni;
405 lwkt_port_t port;
406 struct mbuf *m = *m0;
407
408 *m0 = NULL;
409
410 KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
411 ("%s: bad isr %d", __func__, num));
412
413 ni = &netisrs[num];
414 if (ni->ni_mport == NULL) {
415 kprintf("%s: unregistered isr %d\n", __func__, num);
416 m_freem(m);
417 return NULL;
418 }
419
420 if ((port = ni->ni_mport(&m)) == NULL)
421 return NULL;
422
423 *m0 = m;
424 return port;
425}
426
427void
428netisr_run(int num, struct mbuf *m)
429{
430 struct netisr *ni;
431 struct netmsg_packet *pmsg;
432
433 KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
434 ("%s: bad isr %d", __func__, num));
435
436 ni = &netisrs[num];
437 if (ni->ni_handler == NULL) {
438 kprintf("%s: unregistered isr %d\n", __func__, num);
439 m_freem(m);
440 return;
441 }
442
443 pmsg = &m->m_hdr.mh_netmsg;
444
445 netmsg_init(&pmsg->nm_netmsg, &netisr_apanic_rport, 0, ni->ni_handler);
446 pmsg->nm_packet = m;
447 pmsg->nm_netmsg.nm_lmsg.u.ms_result = num;
448
449 ni->ni_handler(&pmsg->nm_netmsg);
450}