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.
7 * This code is derived from software contributed to The DragonFly Project
8 * by Jonathan Lemon, Jeffrey M. Hsu, and Matthew Dillon.
10 * Jonathan Lemon gave Jeffrey Hsu permission to combine his copyright
11 * into this one around July 8 2004.
13 * Redistribution and use in source and binary forms, with or without
14 * modification, are permitted provided that the following conditions
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.
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
38 * $DragonFly: src/sys/net/netisr.c,v 1.49 2008/11/01 10:29:31 sephe Exp $
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>
47 #include <sys/interrupt.h>
48 #include <sys/socket.h>
49 #include <sys/sysctl.h>
50 #include <sys/socketvar.h>
52 #include <net/if_var.h>
53 #include <net/netisr.h>
54 #include <machine/cpufunc.h>
56 #include <sys/thread2.h>
57 #include <sys/msgport2.h>
58 #include <net/netmsg2.h>
59 #include <sys/mplock2.h>
61 static void netmsg_sync_func(struct netmsg *msg);
62 static void netmsg_service_loop(void *arg);
63 static void cpu0_cpufn(struct mbuf **mp, int hoff);
65 struct netmsg_port_registration {
66 TAILQ_ENTRY(netmsg_port_registration) npr_entry;
70 struct netmsg_rollup {
71 TAILQ_ENTRY(netmsg_rollup) ru_entry;
75 static struct netisr netisrs[NETISR_MAX];
76 static TAILQ_HEAD(,netmsg_port_registration) netreglist;
77 static TAILQ_HEAD(,netmsg_rollup) netrulist;
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;
86 static int (*netmsg_fwd_port_fn)(lwkt_port_t, lwkt_msg_t);
88 SYSCTL_NODE(_net, OID_AUTO, netisr, CTLFLAG_RW, 0, "netisr");
91 * netisr_afree_rport replymsg function, only used to handle async
92 * messages which the sender has abandoned to their fate.
95 netisr_autofree_reply(lwkt_port_t port, lwkt_msg_t msg)
97 kfree(msg, M_LWKTMSG);
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.
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.
112 netmsg_put_port(lwkt_port_t port, lwkt_msg_t lmsg)
114 netmsg_t netmsg = (void *)lmsg;
116 if ((lmsg->ms_flags & MSGF_SYNC) && port == &curthread->td_msgport) {
117 netmsg->nm_dispatch(netmsg);
118 if ((lmsg->ms_flags & MSGF_DONE) == 0) {
119 panic("netmsg_put_port: self-referential "
120 "deadlock on netport");
124 return(netmsg_fwd_port_fn(port, lmsg));
129 * UNIX DOMAIN sockets still have to run their uipc functions synchronously,
130 * because they depend on the user proc context for a number of things
131 * (like creds) which we have not yet incorporated into the message structure.
133 * However, we maintain or message/port abstraction. Having a special
134 * synchronous port which runs the commands synchronously gives us the
135 * ability to serialize operations in one place later on when we start
139 netmsg_sync_putport(lwkt_port_t port, lwkt_msg_t lmsg)
141 netmsg_t netmsg = (void *)lmsg;
143 KKASSERT((lmsg->ms_flags & MSGF_DONE) == 0);
145 lmsg->ms_target_port = port; /* required for abort */
146 netmsg->nm_dispatch(netmsg);
155 TAILQ_INIT(&netreglist);
156 TAILQ_INIT(&netrulist);
159 * Create default per-cpu threads for generic protocol handling.
161 for (i = 0; i < ncpus; ++i) {
162 lwkt_create(netmsg_service_loop, NULL, NULL,
163 &netisr_cpu[i], TDF_STOPREQ, i,
165 netmsg_service_port_init(&netisr_cpu[i].td_msgport);
166 lwkt_schedule(&netisr_cpu[i]);
170 * The netisr_afree_rport is a special reply port which automatically
171 * frees the replied message. The netisr_adone_rport simply marks
172 * the message as being done. The netisr_apanic_rport panics if
173 * the message is replied to.
175 lwkt_initport_replyonly(&netisr_afree_rport, netisr_autofree_reply);
176 lwkt_initport_replyonly_null(&netisr_adone_rport);
177 lwkt_initport_panic(&netisr_apanic_rport);
180 * The netisr_syncport is a special port which executes the message
181 * synchronously and waits for it if EASYNC is returned.
183 lwkt_initport_putonly(&netisr_sync_port, netmsg_sync_putport);
186 SYSINIT(netisr, SI_SUB_PRE_DRIVERS, SI_ORDER_FIRST, netisr_init, NULL);
189 * Finish initializing the message port for a netmsg service. This also
190 * registers the port for synchronous cleanup operations such as when an
191 * ifnet is being destroyed. There is no deregistration API yet.
194 netmsg_service_port_init(lwkt_port_t port)
196 struct netmsg_port_registration *reg;
199 * Override the putport function. Our custom function checks for
200 * self-references and executes such commands synchronously.
202 if (netmsg_fwd_port_fn == NULL)
203 netmsg_fwd_port_fn = port->mp_putport;
204 KKASSERT(netmsg_fwd_port_fn == port->mp_putport);
205 port->mp_putport = netmsg_put_port;
208 * Keep track of ports using the netmsg API so we can synchronize
209 * certain operations (such as freeing an ifnet structure) across all
212 reg = kmalloc(sizeof(*reg), M_TEMP, M_WAITOK|M_ZERO);
213 reg->npr_port = port;
214 TAILQ_INSERT_TAIL(&netreglist, reg, npr_entry);
218 * This function synchronizes the caller with all netmsg services. For
219 * example, if an interface is being removed we must make sure that all
220 * packets related to that interface complete processing before the structure
221 * can actually be freed. This sort of synchronization is an alternative to
222 * ref-counting the netif, removing the ref counting overhead in favor of
223 * placing additional overhead in the netif freeing sequence (where it is
227 netmsg_service_sync(void)
229 struct netmsg_port_registration *reg;
232 netmsg_init(&smsg, NULL, &curthread->td_msgport, 0, netmsg_sync_func);
234 TAILQ_FOREACH(reg, &netreglist, npr_entry) {
235 lwkt_domsg(reg->npr_port, &smsg.nm_lmsg, 0);
240 * The netmsg function simply replies the message. API semantics require
241 * EASYNC to be returned if the netmsg function disposes of the message.
244 netmsg_sync_func(struct netmsg *msg)
246 lwkt_replymsg(&msg->nm_lmsg, 0);
250 * Generic netmsg service loop. Some protocols may roll their own but all
251 * must do the basic command dispatch function call done here.
254 netmsg_service_loop(void *arg)
256 struct netmsg_rollup *ru;
258 thread_t td = curthread;;
261 while ((msg = lwkt_waitport(&td->td_msgport, 0))) {
263 * Run up to 512 pending netmsgs.
267 KASSERT(msg->nm_dispatch != NULL,
268 ("netmsg_service isr %d badmsg\n",
269 msg->nm_lmsg.u.ms_result));
270 msg->nm_dispatch(msg);
273 } while ((msg = lwkt_getport(&td->td_msgport)) != NULL);
276 * Run all registered rollup functions for this cpu
277 * (e.g. tcp_willblock()).
279 TAILQ_FOREACH(ru, &netrulist, ru_entry)
285 * Forward a packet to a netisr service function.
287 * If the packet has not been assigned to a protocol thread we call
288 * the port characterization function to assign it. The caller must
289 * clear M_HASH (or not have set it in the first place) if the caller
290 * wishes the packet to be recharacterized.
293 netisr_queue(int num, struct mbuf *m)
296 struct netmsg_packet *pmsg;
299 KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
300 ("Bad isr %d", num));
303 if (ni->ni_handler == NULL) {
304 kprintf("Unregistered isr %d\n", num);
310 * Figure out which protocol thread to send to. This does not
311 * have to be perfect but performance will be really good if it
312 * is correct. Major protocol inputs such as ip_input() will
313 * re-characterize the packet as necessary.
315 if ((m->m_flags & M_HASH) == 0) {
321 if ((m->m_flags & M_HASH) == 0) {
322 kprintf("netisr_queue(%d): packet hash failed\n", num);
329 * Get the protocol port based on the packet hash, initialize
330 * the netmsg, and send it off.
332 port = cpu_portfn(m->m_pkthdr.hash);
333 pmsg = &m->m_hdr.mh_netmsg;
334 netmsg_init(&pmsg->nm_netmsg, NULL, &netisr_apanic_rport,
337 pmsg->nm_netmsg.nm_lmsg.u.ms_result = num;
338 lwkt_sendmsg(port, &pmsg->nm_netmsg.nm_lmsg);
344 * Pre-characterization of a deeper portion of the packet for the
347 * The base of the ISR type (e.g. IP) that we want to characterize is
348 * at (hoff) relative to the beginning of the mbuf. This allows
349 * e.g. ether_input_chain() to not have to adjust the m_data/m_len.
352 netisr_characterize(int num, struct mbuf **mp, int hoff)
363 if (num < 0 || num >= NETISR_MAX) {
364 if (num == NETISR_MAX) {
365 m->m_flags |= M_HASH;
366 m->m_pkthdr.hash = 0;
369 panic("Bad isr %d", num);
376 if (ni->ni_handler == NULL) {
377 kprintf("Unregistered isr %d\n", num);
383 * Characterize the packet
385 if ((m->m_flags & M_HASH) == 0) {
386 ni->ni_cpufn(mp, hoff);
388 if (m && (m->m_flags & M_HASH) == 0)
389 kprintf("netisr_queue(%d): packet hash failed\n", num);
394 netisr_register(int num, netisr_fn_t handler, netisr_cpufn_t cpufn)
398 KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
399 ("netisr_register: bad isr %d", num));
400 KKASSERT(handler != NULL);
407 ni->ni_handler = handler;
408 ni->ni_cpufn = cpufn;
409 netmsg_init(&ni->ni_netmsg, NULL, &netisr_adone_rport, 0, NULL);
413 netisr_register_rollup(netisr_ru_t ru_func)
415 struct netmsg_rollup *ru;
417 ru = kmalloc(sizeof(*ru), M_TEMP, M_WAITOK|M_ZERO);
418 ru->ru_func = ru_func;
419 TAILQ_INSERT_TAIL(&netrulist, ru, ru_entry);
423 * Return the message port for the general protocol message servicing
424 * thread for a particular cpu.
429 KKASSERT(cpu >= 0 && cpu < ncpus);
430 return (&netisr_cpu[cpu].td_msgport);
434 * Return the current cpu's network protocol thread.
439 return(cpu_portfn(mycpu->gd_cpuid));
443 * Return a default protocol mbuf processing thread port
446 cpu0_soport(struct socket *so __unused, struct sockaddr *nam __unused,
447 struct mbuf **dummy __unused)
449 return (&netisr_cpu[0].td_msgport);
453 * Return a default protocol control message processing thread port
456 cpu0_ctlport(int cmd __unused, struct sockaddr *sa __unused,
457 void *extra __unused)
459 return (&netisr_cpu[0].td_msgport);
463 * This is a dummy port that causes a message to be executed synchronously
464 * instead of being queued to a port.
467 sync_soport(struct socket *so __unused, struct sockaddr *nam __unused,
468 struct mbuf **dummy __unused)
470 return (&netisr_sync_port);
474 * This is a default netisr packet characterization function which
475 * sets M_HASH. If a netisr is registered with a NULL cpufn function
476 * this one is assigned.
478 * This function makes no attempt to validate the packet.
481 cpu0_cpufn(struct mbuf **mp, int hoff __unused)
483 struct mbuf *m = *mp;
485 m->m_flags |= M_HASH;
486 m->m_pkthdr.hash = 0;
490 * schednetisr() is used to call the netisr handler from the appropriate
491 * netisr thread for polling and other purposes.
493 * This function may be called from a hard interrupt or IPI and must be
494 * MP SAFE and non-blocking. We use a fixed per-cpu message instead of
495 * trying to allocate one. We must get ourselves onto the target cpu
496 * to safely check the MSGF_DONE bit on the message but since the message
497 * will be sent to that cpu anyway this does not add any extra work beyond
498 * what lwkt_sendmsg() would have already had to do to schedule the target
502 schednetisr_remote(void *data)
504 int num = (int)(intptr_t)data;
505 struct netisr *ni = &netisrs[num];
506 lwkt_port_t port = &netisr_cpu[0].td_msgport;
509 pmsg = &netisrs[num].ni_netmsg;
510 if (pmsg->nm_lmsg.ms_flags & MSGF_DONE) {
511 netmsg_init(pmsg, NULL, &netisr_adone_rport, 0, ni->ni_handler);
512 pmsg->nm_lmsg.u.ms_result = num;
513 lwkt_sendmsg(port, &pmsg->nm_lmsg);
520 KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
521 ("schednetisr: bad isr %d", num));
522 KKASSERT(netisrs[num].ni_handler != NULL);
524 if (mycpu->gd_cpuid != 0) {
525 lwkt_send_ipiq(globaldata_find(0),
526 schednetisr_remote, (void *)(intptr_t)num);
529 schednetisr_remote((void *)(intptr_t)num);
534 schednetisr_remote((void *)(intptr_t)num);