xref: /dflybsd-src/sys/net/netisr.c (revision d2d1103f52e6fb116ee65a9940477c5449933f28)
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  * $DragonFly: src/sys/net/netisr.c,v 1.49 2008/11/01 10:29:31 sephe Exp $
39  */
40 
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/kernel.h>
44 #include <sys/malloc.h>
45 #include <sys/msgport.h>
46 #include <sys/proc.h>
47 #include <sys/interrupt.h>
48 #include <sys/socket.h>
49 #include <sys/sysctl.h>
50 #include <sys/socketvar.h>
51 #include <net/if.h>
52 #include <net/if_var.h>
53 #include <net/netisr.h>
54 #include <machine/cpufunc.h>
55 
56 #include <sys/thread2.h>
57 #include <sys/msgport2.h>
58 #include <net/netmsg2.h>
59 #include <sys/mplock2.h>
60 
61 static void netmsg_sync_func(netmsg_t msg);
62 static void netmsg_service_loop(void *arg);
63 static void cpu0_cpufn(struct mbuf **mp, int hoff);
64 
65 struct netmsg_port_registration {
66 	TAILQ_ENTRY(netmsg_port_registration) npr_entry;
67 	lwkt_port_t	npr_port;
68 };
69 
70 struct netmsg_rollup {
71 	TAILQ_ENTRY(netmsg_rollup) ru_entry;
72 	netisr_ru_t	ru_func;
73 };
74 
75 static struct netisr netisrs[NETISR_MAX];
76 static TAILQ_HEAD(,netmsg_port_registration) netreglist;
77 static TAILQ_HEAD(,netmsg_rollup) netrulist;
78 
79 /* Per-CPU thread to handle any protocol.  */
80 static struct thread netisr_cpu[MAXCPU];
81 lwkt_port netisr_afree_rport;
82 lwkt_port netisr_adone_rport;
83 lwkt_port netisr_apanic_rport;
84 lwkt_port netisr_sync_port;
85 
86 static int (*netmsg_fwd_port_fn)(lwkt_port_t, lwkt_msg_t);
87 
88 SYSCTL_NODE(_net, OID_AUTO, netisr, CTLFLAG_RW, 0, "netisr");
89 
90 /*
91  * netisr_afree_rport replymsg function, only used to handle async
92  * messages which the sender has abandoned to their fate.
93  */
94 static void
95 netisr_autofree_reply(lwkt_port_t port, lwkt_msg_t msg)
96 {
97 	kfree(msg, M_LWKTMSG);
98 }
99 
100 /*
101  * We need a custom putport function to handle the case where the
102  * message target is the current thread's message port.  This case
103  * can occur when the TCP or UDP stack does a direct callback to NFS and NFS
104  * then turns around and executes a network operation synchronously.
105  *
106  * To prevent deadlocking, we must execute these self-referential messages
107  * synchronously, effectively turning the message into a glorified direct
108  * procedure call back into the protocol stack.  The operation must be
109  * complete on return or we will deadlock, so panic if it isn't.
110  *
111  * However, the target function is under no obligation to immediately
112  * reply the message.  It may forward it elsewhere.
113  */
114 static int
115 netmsg_put_port(lwkt_port_t port, lwkt_msg_t lmsg)
116 {
117 	netmsg_base_t nmsg = (void *)lmsg;
118 
119 	if ((lmsg->ms_flags & MSGF_SYNC) && port == &curthread->td_msgport) {
120 		nmsg->nm_dispatch((netmsg_t)nmsg);
121 		return(EASYNC);
122 	} else {
123 		return(netmsg_fwd_port_fn(port, lmsg));
124 	}
125 }
126 
127 /*
128  * UNIX DOMAIN sockets still have to run their uipc functions synchronously,
129  * because they depend on the user proc context for a number of things
130  * (like creds) which we have not yet incorporated into the message structure.
131  *
132  * However, we maintain or message/port abstraction.  Having a special
133  * synchronous port which runs the commands synchronously gives us the
134  * ability to serialize operations in one place later on when we start
135  * removing the BGL.
136  */
137 static int
138 netmsg_sync_putport(lwkt_port_t port, lwkt_msg_t lmsg)
139 {
140 	netmsg_base_t nmsg = (void *)lmsg;
141 
142 	KKASSERT((lmsg->ms_flags & MSGF_DONE) == 0);
143 
144 	lmsg->ms_target_port = port;	/* required for abort */
145 	nmsg->nm_dispatch((netmsg_t)nmsg);
146 	return(EASYNC);
147 }
148 
149 static void
150 netisr_init(void)
151 {
152 	int i;
153 
154 	TAILQ_INIT(&netreglist);
155 	TAILQ_INIT(&netrulist);
156 
157 	/*
158 	 * Create default per-cpu threads for generic protocol handling.
159 	 */
160 	for (i = 0; i < ncpus; ++i) {
161 		lwkt_create(netmsg_service_loop, NULL, NULL,
162 			    &netisr_cpu[i], TDF_STOPREQ, i,
163 			    "netisr_cpu %d", i);
164 		netmsg_service_port_init(&netisr_cpu[i].td_msgport);
165 		lwkt_schedule(&netisr_cpu[i]);
166 	}
167 
168 	/*
169 	 * The netisr_afree_rport is a special reply port which automatically
170 	 * frees the replied message.  The netisr_adone_rport simply marks
171 	 * the message as being done.  The netisr_apanic_rport panics if
172 	 * the message is replied to.
173 	 */
174 	lwkt_initport_replyonly(&netisr_afree_rport, netisr_autofree_reply);
175 	lwkt_initport_replyonly_null(&netisr_adone_rport);
176 	lwkt_initport_panic(&netisr_apanic_rport);
177 
178 	/*
179 	 * The netisr_syncport is a special port which executes the message
180 	 * synchronously and waits for it if EASYNC is returned.
181 	 */
182 	lwkt_initport_putonly(&netisr_sync_port, netmsg_sync_putport);
183 }
184 
185 SYSINIT(netisr, SI_SUB_PRE_DRIVERS, SI_ORDER_FIRST, netisr_init, NULL);
186 
187 /*
188  * Finish initializing the message port for a netmsg service.  This also
189  * registers the port for synchronous cleanup operations such as when an
190  * ifnet is being destroyed.  There is no deregistration API yet.
191  */
192 void
193 netmsg_service_port_init(lwkt_port_t port)
194 {
195 	struct netmsg_port_registration *reg;
196 
197 	/*
198 	 * Override the putport function.  Our custom function checks for
199 	 * self-references and executes such commands synchronously.
200 	 */
201 	if (netmsg_fwd_port_fn == NULL)
202 		netmsg_fwd_port_fn = port->mp_putport;
203 	KKASSERT(netmsg_fwd_port_fn == port->mp_putport);
204 	port->mp_putport = netmsg_put_port;
205 
206 	/*
207 	 * Keep track of ports using the netmsg API so we can synchronize
208 	 * certain operations (such as freeing an ifnet structure) across all
209 	 * consumers.
210 	 */
211 	reg = kmalloc(sizeof(*reg), M_TEMP, M_WAITOK|M_ZERO);
212 	reg->npr_port = port;
213 	TAILQ_INSERT_TAIL(&netreglist, reg, npr_entry);
214 }
215 
216 /*
217  * This function synchronizes the caller with all netmsg services.  For
218  * example, if an interface is being removed we must make sure that all
219  * packets related to that interface complete processing before the structure
220  * can actually be freed.  This sort of synchronization is an alternative to
221  * ref-counting the netif, removing the ref counting overhead in favor of
222  * placing additional overhead in the netif freeing sequence (where it is
223  * inconsequential).
224  */
225 void
226 netmsg_service_sync(void)
227 {
228 	struct netmsg_port_registration *reg;
229 	struct netmsg_base smsg;
230 
231 	netmsg_init(&smsg, NULL, &curthread->td_msgport, 0, netmsg_sync_func);
232 
233 	TAILQ_FOREACH(reg, &netreglist, npr_entry) {
234 		lwkt_domsg(reg->npr_port, &smsg.lmsg, 0);
235 	}
236 }
237 
238 /*
239  * The netmsg function simply replies the message.  API semantics require
240  * EASYNC to be returned if the netmsg function disposes of the message.
241  */
242 static void
243 netmsg_sync_func(netmsg_t msg)
244 {
245 	lwkt_replymsg(&msg->lmsg, 0);
246 }
247 
248 /*
249  * Generic netmsg service loop.  Some protocols may roll their own but all
250  * must do the basic command dispatch function call done here.
251  */
252 static void
253 netmsg_service_loop(void *arg)
254 {
255 	struct netmsg_rollup *ru;
256 	netmsg_base_t msg;
257 	thread_t td = curthread;;
258 	int limit;
259 
260 	while ((msg = lwkt_waitport(&td->td_msgport, 0))) {
261 		/*
262 		 * Run up to 512 pending netmsgs.
263 		 */
264 		limit = 512;
265 		do {
266 			KASSERT(msg->nm_dispatch != NULL,
267 				("netmsg_service isr %d badmsg\n",
268 				msg->lmsg.u.ms_result));
269 			if (msg->nm_so &&
270 			    msg->nm_so->so_port != &td->td_msgport) {
271 				/*
272 				 * Sockets undergoing connect or disconnect
273 				 * ops can change ports on us.  Chase the
274 				 * port.
275 				 */
276 				kprintf("netmsg_service_loop: Warning, "
277 					"port changed so=%p\n", msg->nm_so);
278 				lwkt_forwardmsg(msg->nm_so->so_port,
279 						&msg->lmsg);
280 			} else {
281 				/*
282 				 * We are on the correct port, dispatch it.
283 				 */
284 				msg->nm_dispatch((netmsg_t)msg);
285 			}
286 			if (--limit == 0)
287 				break;
288 		} while ((msg = lwkt_getport(&td->td_msgport)) != NULL);
289 
290 		/*
291 		 * Run all registered rollup functions for this cpu
292 		 * (e.g. tcp_willblock()).
293 		 */
294 		TAILQ_FOREACH(ru, &netrulist, ru_entry)
295 			ru->ru_func();
296 	}
297 }
298 
299 /*
300  * Forward a packet to a netisr service function.
301  *
302  * If the packet has not been assigned to a protocol thread we call
303  * the port characterization function to assign it.  The caller must
304  * clear M_HASH (or not have set it in the first place) if the caller
305  * wishes the packet to be recharacterized.
306  */
307 int
308 netisr_queue(int num, struct mbuf *m)
309 {
310 	struct netisr *ni;
311 	struct netmsg_packet *pmsg;
312 	lwkt_port_t port;
313 
314 	KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
315 		("Bad isr %d", num));
316 
317 	ni = &netisrs[num];
318 	if (ni->ni_handler == NULL) {
319 		kprintf("Unregistered isr %d\n", num);
320 		m_freem(m);
321 		return (EIO);
322 	}
323 
324 	/*
325 	 * Figure out which protocol thread to send to.  This does not
326 	 * have to be perfect but performance will be really good if it
327 	 * is correct.  Major protocol inputs such as ip_input() will
328 	 * re-characterize the packet as necessary.
329 	 */
330 	if ((m->m_flags & M_HASH) == 0) {
331 		ni->ni_cpufn(&m, 0);
332 		if (m == NULL) {
333 			m_freem(m);
334 			return (EIO);
335 		}
336 		if ((m->m_flags & M_HASH) == 0) {
337 			kprintf("netisr_queue(%d): packet hash failed\n", num);
338 			m_freem(m);
339 			return (EIO);
340 		}
341 	}
342 
343 	/*
344 	 * Get the protocol port based on the packet hash, initialize
345 	 * the netmsg, and send it off.
346 	 */
347 	port = cpu_portfn(m->m_pkthdr.hash);
348 	pmsg = &m->m_hdr.mh_netmsg;
349 	netmsg_init(&pmsg->base, NULL, &netisr_apanic_rport,
350 		    0, ni->ni_handler);
351 	pmsg->nm_packet = m;
352 	pmsg->base.lmsg.u.ms_result = num;
353 	lwkt_sendmsg(port, &pmsg->base.lmsg);
354 
355 	return (0);
356 }
357 
358 /*
359  * Pre-characterization of a deeper portion of the packet for the
360  * requested isr.
361  *
362  * The base of the ISR type (e.g. IP) that we want to characterize is
363  * at (hoff) relative to the beginning of the mbuf.  This allows
364  * e.g. ether_input_chain() to not have to adjust the m_data/m_len.
365  */
366 void
367 netisr_characterize(int num, struct mbuf **mp, int hoff)
368 {
369 	struct netisr *ni;
370 	struct mbuf *m;
371 
372 	/*
373 	 * Validation
374 	 */
375 	m = *mp;
376 	KKASSERT(m != NULL);
377 
378 	if (num < 0 || num >= NETISR_MAX) {
379 		if (num == NETISR_MAX) {
380 			m->m_flags |= M_HASH;
381 			m->m_pkthdr.hash = 0;
382 			return;
383 		}
384 		panic("Bad isr %d", num);
385 	}
386 
387 	/*
388 	 * Valid netisr?
389 	 */
390 	ni = &netisrs[num];
391 	if (ni->ni_handler == NULL) {
392 		kprintf("Unregistered isr %d\n", num);
393 		m_freem(m);
394 		*mp = NULL;
395 	}
396 
397 	/*
398 	 * Characterize the packet
399 	 */
400 	if ((m->m_flags & M_HASH) == 0) {
401 		ni->ni_cpufn(mp, hoff);
402 		m = *mp;
403 		if (m && (m->m_flags & M_HASH) == 0)
404 			kprintf("netisr_queue(%d): packet hash failed\n", num);
405 	}
406 }
407 
408 void
409 netisr_register(int num, netisr_fn_t handler, netisr_cpufn_t cpufn)
410 {
411 	struct netisr *ni;
412 
413 	KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
414 		("netisr_register: bad isr %d", num));
415 	KKASSERT(handler != NULL);
416 
417 	if (cpufn == NULL)
418 		cpufn = cpu0_cpufn;
419 
420 	ni = &netisrs[num];
421 
422 	ni->ni_handler = handler;
423 	ni->ni_cpufn = cpufn;
424 	netmsg_init(&ni->ni_netmsg, NULL, &netisr_adone_rport, 0, NULL);
425 }
426 
427 void
428 netisr_register_rollup(netisr_ru_t ru_func)
429 {
430 	struct netmsg_rollup *ru;
431 
432 	ru = kmalloc(sizeof(*ru), M_TEMP, M_WAITOK|M_ZERO);
433 	ru->ru_func = ru_func;
434 	TAILQ_INSERT_TAIL(&netrulist, ru, ru_entry);
435 }
436 
437 /*
438  * Return the message port for the general protocol message servicing
439  * thread for a particular cpu.
440  */
441 lwkt_port_t
442 cpu_portfn(int cpu)
443 {
444 	KKASSERT(cpu >= 0 && cpu < ncpus);
445 	return (&netisr_cpu[cpu].td_msgport);
446 }
447 
448 /*
449  * Return the current cpu's network protocol thread.
450  */
451 lwkt_port_t
452 cur_netport(void)
453 {
454 	return(cpu_portfn(mycpu->gd_cpuid));
455 }
456 
457 /*
458  * Return a default protocol control message processing thread port
459  */
460 lwkt_port_t
461 cpu0_ctlport(int cmd __unused, struct sockaddr *sa __unused,
462 	     void *extra __unused)
463 {
464 	return (&netisr_cpu[0].td_msgport);
465 }
466 
467 /*
468  * This is a default netisr packet characterization function which
469  * sets M_HASH.  If a netisr is registered with a NULL cpufn function
470  * this one is assigned.
471  *
472  * This function makes no attempt to validate the packet.
473  */
474 static void
475 cpu0_cpufn(struct mbuf **mp, int hoff __unused)
476 {
477 	struct mbuf *m = *mp;
478 
479 	m->m_flags |= M_HASH;
480 	m->m_pkthdr.hash = 0;
481 }
482 
483 /*
484  * schednetisr() is used to call the netisr handler from the appropriate
485  * netisr thread for polling and other purposes.
486  *
487  * This function may be called from a hard interrupt or IPI and must be
488  * MP SAFE and non-blocking.  We use a fixed per-cpu message instead of
489  * trying to allocate one.  We must get ourselves onto the target cpu
490  * to safely check the MSGF_DONE bit on the message but since the message
491  * will be sent to that cpu anyway this does not add any extra work beyond
492  * what lwkt_sendmsg() would have already had to do to schedule the target
493  * thread.
494  */
495 static void
496 schednetisr_remote(void *data)
497 {
498 	int num = (int)(intptr_t)data;
499 	struct netisr *ni = &netisrs[num];
500 	lwkt_port_t port = &netisr_cpu[0].td_msgport;
501 	netmsg_base_t pmsg;
502 
503 	pmsg = &netisrs[num].ni_netmsg;
504 	if (pmsg->lmsg.ms_flags & MSGF_DONE) {
505 		netmsg_init(pmsg, NULL, &netisr_adone_rport, 0, ni->ni_handler);
506 		pmsg->lmsg.u.ms_result = num;
507 		lwkt_sendmsg(port, &pmsg->lmsg);
508 	}
509 }
510 
511 void
512 schednetisr(int num)
513 {
514 	KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
515 		("schednetisr: bad isr %d", num));
516 	KKASSERT(netisrs[num].ni_handler != NULL);
517 #ifdef SMP
518 	if (mycpu->gd_cpuid != 0) {
519 		lwkt_send_ipiq(globaldata_find(0),
520 			       schednetisr_remote, (void *)(intptr_t)num);
521 	} else {
522 		crit_enter();
523 		schednetisr_remote((void *)(intptr_t)num);
524 		crit_exit();
525 	}
526 #else
527 	crit_enter();
528 	schednetisr_remote((void *)(intptr_t)num);
529 	crit_exit();
530 #endif
531 }
532