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 1.46 2008/09/23 11:28:49 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>
51 #include <net/if_var.h>
52 #include <net/netisr.h>
53 #include <machine/cpufunc.h>
55 #include <sys/thread2.h>
56 #include <sys/msgport2.h>
57 #include <net/netmsg2.h>
59 #define NETISR_GET_MPLOCK(ni) \
61 if (((ni)->ni_flags & NETISR_FLAG_MPSAFE) == 0) \
65 #define NETISR_REL_MPLOCK(ni) \
67 if (((ni)->ni_flags & NETISR_FLAG_MPSAFE) == 0) \
71 static void netmsg_sync_func(struct netmsg *msg);
73 struct netmsg_port_registration {
74 TAILQ_ENTRY(netmsg_port_registration) npr_entry;
78 static struct netisr netisrs[NETISR_MAX];
79 static TAILQ_HEAD(,netmsg_port_registration) netreglist;
81 /* Per-CPU thread to handle any protocol. */
82 struct thread netisr_cpu[MAXCPU];
83 lwkt_port netisr_afree_rport;
84 lwkt_port netisr_adone_rport;
85 lwkt_port netisr_apanic_rport;
86 lwkt_port netisr_sync_port;
88 static int (*netmsg_fwd_port_fn)(lwkt_port_t, lwkt_msg_t);
90 static int netisr_mpsafe_thread = 0;
91 TUNABLE_INT("net.netisr.mpsafe_thread", &netisr_mpsafe_thread);
93 SYSCTL_NODE(_net, OID_AUTO, netisr, CTLFLAG_RW, 0, "netisr");
94 SYSCTL_INT(_net_netisr, OID_AUTO, mpsafe_thread, CTLFLAG_RW,
95 &netisr_mpsafe_thread, 0,
96 "0:BGL, 1:Adaptive BGL, 2:No BGL(experimental)");
99 * netisr_afree_rport replymsg function, only used to handle async
100 * messages which the sender has abandoned to their fate.
103 netisr_autofree_reply(lwkt_port_t port, lwkt_msg_t msg)
105 kfree(msg, M_LWKTMSG);
109 * We need a custom putport function to handle the case where the
110 * message target is the current thread's message port. This case
111 * can occur when the TCP or UDP stack does a direct callback to NFS and NFS
112 * then turns around and executes a network operation synchronously.
114 * To prevent deadlocking, we must execute these self-referential messages
115 * synchronously, effectively turning the message into a glorified direct
116 * procedure call back into the protocol stack. The operation must be
117 * complete on return or we will deadlock, so panic if it isn't.
120 netmsg_put_port(lwkt_port_t port, lwkt_msg_t lmsg)
122 netmsg_t netmsg = (void *)lmsg;
124 if ((lmsg->ms_flags & MSGF_SYNC) && port == &curthread->td_msgport) {
125 netmsg->nm_dispatch(netmsg);
126 if ((lmsg->ms_flags & MSGF_DONE) == 0)
127 panic("netmsg_put_port: self-referential deadlock on netport");
130 return(netmsg_fwd_port_fn(port, lmsg));
135 * UNIX DOMAIN sockets still have to run their uipc functions synchronously,
136 * because they depend on the user proc context for a number of things
137 * (like creds) which we have not yet incorporated into the message structure.
139 * However, we maintain or message/port abstraction. Having a special
140 * synchronous port which runs the commands synchronously gives us the
141 * ability to serialize operations in one place later on when we start
145 netmsg_sync_putport(lwkt_port_t port, lwkt_msg_t lmsg)
147 netmsg_t netmsg = (void *)lmsg;
149 KKASSERT((lmsg->ms_flags & MSGF_DONE) == 0);
151 lmsg->ms_target_port = port; /* required for abort */
152 netmsg->nm_dispatch(netmsg);
161 TAILQ_INIT(&netreglist);
164 * Create default per-cpu threads for generic protocol handling.
166 for (i = 0; i < ncpus; ++i) {
167 lwkt_create(netmsg_service_loop, &netisr_mpsafe_thread, NULL,
168 &netisr_cpu[i], TDF_NETWORK | TDF_MPSAFE, i,
170 netmsg_service_port_init(&netisr_cpu[i].td_msgport);
174 * The netisr_afree_rport is a special reply port which automatically
175 * frees the replied message. The netisr_adone_rport simply marks
176 * the message as being done. The netisr_apanic_rport panics if
177 * the message is replied to.
179 lwkt_initport_replyonly(&netisr_afree_rport, netisr_autofree_reply);
180 lwkt_initport_replyonly_null(&netisr_adone_rport);
181 lwkt_initport_panic(&netisr_apanic_rport);
184 * The netisr_syncport is a special port which executes the message
185 * synchronously and waits for it if EASYNC is returned.
187 lwkt_initport_putonly(&netisr_sync_port, netmsg_sync_putport);
190 SYSINIT(netisr, SI_SUB_PRE_DRIVERS, SI_ORDER_FIRST, netisr_init, NULL);
193 * Finish initializing the message port for a netmsg service. This also
194 * registers the port for synchronous cleanup operations such as when an
195 * ifnet is being destroyed. There is no deregistration API yet.
198 netmsg_service_port_init(lwkt_port_t port)
200 struct netmsg_port_registration *reg;
203 * Override the putport function. Our custom function checks for
204 * self-references and executes such commands synchronously.
206 if (netmsg_fwd_port_fn == NULL)
207 netmsg_fwd_port_fn = port->mp_putport;
208 KKASSERT(netmsg_fwd_port_fn == port->mp_putport);
209 port->mp_putport = netmsg_put_port;
212 * Keep track of ports using the netmsg API so we can synchronize
213 * certain operations (such as freeing an ifnet structure) across all
216 reg = kmalloc(sizeof(*reg), M_TEMP, M_WAITOK|M_ZERO);
217 reg->npr_port = port;
218 TAILQ_INSERT_TAIL(&netreglist, reg, npr_entry);
222 * This function synchronizes the caller with all netmsg services. For
223 * example, if an interface is being removed we must make sure that all
224 * packets related to that interface complete processing before the structure
225 * can actually be freed. This sort of synchronization is an alternative to
226 * ref-counting the netif, removing the ref counting overhead in favor of
227 * placing additional overhead in the netif freeing sequence (where it is
231 netmsg_service_sync(void)
233 struct netmsg_port_registration *reg;
236 netmsg_init(&smsg, &curthread->td_msgport, 0, netmsg_sync_func);
238 TAILQ_FOREACH(reg, &netreglist, npr_entry) {
239 lwkt_domsg(reg->npr_port, &smsg.nm_lmsg, 0);
244 * The netmsg function simply replies the message. API semantics require
245 * EASYNC to be returned if the netmsg function disposes of the message.
248 netmsg_sync_func(struct netmsg *msg)
250 lwkt_replymsg(&msg->nm_lmsg, 0);
254 * Return current BGL lock state (1:locked, 0: unlocked)
257 netmsg_service(struct netmsg *msg, int mpsafe_mode, int mplocked)
260 * Adjust the mplock dynamically.
262 switch (mpsafe_mode) {
263 case NETMSG_SERVICE_ADAPTIVE: /* Adaptive BGL */
264 if (msg->nm_lmsg.ms_flags & MSGF_MPSAFE) {
269 msg->nm_dispatch(msg);
270 /* Leave mpunlocked */
274 /* mplocked = 1; not needed */
276 msg->nm_dispatch(msg);
279 /* Leave mpunlocked, next msg might be mpsafe */
283 case NETMSG_SERVICE_MPSAFE: /* No BGL */
288 msg->nm_dispatch(msg);
289 /* Leave mpunlocked */
297 msg->nm_dispatch(msg);
305 * Generic netmsg service loop. Some protocols may roll their own but all
306 * must do the basic command dispatch function call done here.
309 netmsg_service_loop(void *arg)
312 int mplocked, *mpsafe_mode = arg;
315 * Thread was started with TDF_MPSAFE
322 while ((msg = lwkt_waitport(&curthread->td_msgport, 0))) {
323 mplocked = netmsg_service(msg, *mpsafe_mode, mplocked);
328 * Call the netisr directly.
329 * Queueing may be done in the msg port layer at its discretion.
332 netisr_dispatch(int num, struct mbuf *m)
334 /* just queue it for now XXX JH */
335 netisr_queue(num, m);
339 * Same as netisr_dispatch(), but always queue.
340 * This is either used in places where we are not confident that
341 * direct dispatch is possible, or where queueing is required.
344 netisr_queue(int num, struct mbuf *m)
347 struct netmsg_packet *pmsg;
350 KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
351 ("%s: bad isr %d", __func__, num));
354 if (ni->ni_handler == NULL) {
355 kprintf("%s: unregistered isr %d\n", __func__, num);
360 if ((port = ni->ni_mport(&m)) == NULL)
363 pmsg = &m->m_hdr.mh_netmsg;
365 netmsg_init(&pmsg->nm_netmsg, &netisr_apanic_rport,
366 (ni->ni_flags & NETISR_FLAG_MPSAFE) ? MSGF_MPSAFE : 0,
369 pmsg->nm_netmsg.nm_lmsg.u.ms_result = num;
370 lwkt_sendmsg(port, &pmsg->nm_netmsg.nm_lmsg);
375 netisr_register(int num, lwkt_portfn_t mportfn, netisr_fn_t handler,
378 KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
379 ("netisr_register: bad isr %d", num));
380 netmsg_init(&netisrs[num].ni_netmsg, &netisr_adone_rport,
381 (flags & NETISR_FLAG_MPSAFE) ? MSGF_MPSAFE : 0, NULL);
382 netisrs[num].ni_mport = mportfn;
383 netisrs[num].ni_handler = handler;
384 netisrs[num].ni_flags = flags;
388 netisr_unregister(int num)
390 KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
391 ("unregister_netisr: bad isr number: %d\n", num));
398 * Return message port for default handler thread on CPU 0.
401 cpu0_portfn(struct mbuf **mptr)
403 return (&netisr_cpu[0].td_msgport);
409 return (&netisr_cpu[cpu].td_msgport);
414 cpu0_soport(struct socket *so __unused, struct sockaddr *nam __unused,
415 struct mbuf **dummy __unused, int req __unused)
417 return (&netisr_cpu[0].td_msgport);
421 sync_soport(struct socket *so __unused, struct sockaddr *nam __unused,
422 struct mbuf **dummy __unused, int req __unused)
424 return (&netisr_sync_port);
428 * schednetisr() is used to call the netisr handler from the appropriate
429 * netisr thread for polling and other purposes.
431 * This function may be called from a hard interrupt or IPI and must be
432 * MP SAFE and non-blocking. We use a fixed per-cpu message instead of
433 * trying to allocate one. We must get ourselves onto the target cpu
434 * to safely check the MSGF_DONE bit on the message but since the message
435 * will be sent to that cpu anyway this does not add any extra work beyond
436 * what lwkt_sendmsg() would have already had to do to schedule the target
440 schednetisr_remote(void *data)
443 struct netisr *ni = &netisrs[num];
444 lwkt_port_t port = &netisr_cpu[0].td_msgport;
447 pmsg = &netisrs[num].ni_netmsg;
449 if (pmsg->nm_lmsg.ms_flags & MSGF_DONE) {
450 netmsg_init(pmsg, &netisr_adone_rport, 0, ni->ni_handler);
451 pmsg->nm_lmsg.u.ms_result = num;
452 lwkt_sendmsg(port, &pmsg->nm_lmsg);
460 KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
461 ("schednetisr: bad isr %d", num));
463 if (mycpu->gd_cpuid != 0)
464 lwkt_send_ipiq(globaldata_find(0), schednetisr_remote, (void *)num);
466 schednetisr_remote((void *)num);
468 schednetisr_remote((void *)num);
473 netisr_find_port(int num, struct mbuf **m0)
477 struct mbuf *m = *m0;
481 KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
482 ("%s: bad isr %d", __func__, num));
485 if (ni->ni_mport == NULL) {
486 kprintf("%s: unregistered isr %d\n", __func__, num);
491 if ((port = ni->ni_mport(&m)) == NULL)
499 netisr_run(int num, struct mbuf *m)
502 struct netmsg_packet *pmsg;
504 KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
505 ("%s: bad isr %d", __func__, num));
508 if (ni->ni_handler == NULL) {
509 kprintf("%s: unregistered isr %d\n", __func__, num);
514 pmsg = &m->m_hdr.mh_netmsg;
516 netmsg_init(&pmsg->nm_netmsg, &netisr_apanic_rport, 0, ni->ni_handler);
518 pmsg->nm_netmsg.nm_lmsg.u.ms_result = num;
520 NETISR_GET_MPLOCK(ni);
521 ni->ni_handler(&pmsg->nm_netmsg);
522 NETISR_REL_MPLOCK(ni);