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 1.34 2007/05/31 11:00:25 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 <net/if.h> 51 #include <net/if_var.h> 52 #include <net/netisr.h> 53 #include <machine/cpufunc.h> 54 55 #include <sys/thread2.h> 56 #include <sys/msgport2.h> 57 #include <net/netmsg2.h> 58 59 static void netmsg_sync_func(struct netmsg *msg); 60 61 struct netmsg_port_registration { 62 TAILQ_ENTRY(netmsg_port_registration) npr_entry; 63 lwkt_port_t npr_port; 64 }; 65 66 static struct netisr netisrs[NETISR_MAX]; 67 static TAILQ_HEAD(,netmsg_port_registration) netreglist; 68 69 /* Per-CPU thread to handle any protocol. */ 70 struct thread netisr_cpu[MAXCPU]; 71 lwkt_port netisr_afree_rport; 72 lwkt_port netisr_adone_rport; 73 lwkt_port netisr_apanic_rport; 74 lwkt_port netisr_sync_port; 75 76 static int (*netmsg_fwd_port_fn)(lwkt_port_t, lwkt_msg_t); 77 78 /* 79 * netisr_afree_rport replymsg function, only used to handle async 80 * messages which the sender has abandoned to their fate. 81 */ 82 static void 83 netisr_autofree_reply(lwkt_port_t port, lwkt_msg_t msg) 84 { 85 kfree(msg, M_LWKTMSG); 86 } 87 88 /* 89 * We need a custom putport function to handle the case where the 90 * message target is the current thread's message port. This case 91 * can occur when the TCP or UDP stack does a direct callback to NFS and NFS 92 * then turns around and executes a network operation synchronously. 93 * 94 * To prevent deadlocking, we must execute these self-referential messages 95 * synchronously, effectively turning the message into a glorified direct 96 * procedure call back into the protocol stack. The operation must be 97 * complete on return or we will deadlock, so panic if it isn't. 98 */ 99 static int 100 netmsg_put_port(lwkt_port_t port, lwkt_msg_t lmsg) 101 { 102 netmsg_t netmsg = (void *)lmsg; 103 104 if ((lmsg->ms_flags & MSGF_SYNC) && port == &curthread->td_msgport) { 105 netmsg->nm_dispatch(netmsg); 106 if ((lmsg->ms_flags & MSGF_DONE) == 0) 107 panic("netmsg_put_port: self-referential deadlock on netport"); 108 return(EASYNC); 109 } else { 110 return(netmsg_fwd_port_fn(port, lmsg)); 111 } 112 } 113 114 /* 115 * UNIX DOMAIN sockets still have to run their uipc functions synchronously, 116 * because they depend on the user proc context for a number of things 117 * (like creds) which we have not yet incorporated into the message structure. 118 * 119 * However, we maintain or message/port abstraction. Having a special 120 * synchronous port which runs the commands synchronously gives us the 121 * ability to serialize operations in one place later on when we start 122 * removing the BGL. 123 * 124 * We clear MSGF_DONE prior to executing the message in order to close 125 * any potential replymsg races with the flags field. If a synchronous 126 * result code is returned we set MSGF_DONE again. MSGF_DONE's flag state 127 * must be correct or the caller will be confused. 128 */ 129 static int 130 netmsg_sync_putport(lwkt_port_t port, lwkt_msg_t lmsg) 131 { 132 netmsg_t netmsg = (void *)lmsg; 133 int error; 134 135 lmsg->ms_flags &= ~MSGF_DONE; 136 lmsg->ms_target_port = port; /* required for abort */ 137 netmsg->nm_dispatch(netmsg); 138 error = lwkt_waitmsg(lmsg, 0); 139 return(error); 140 } 141 142 static void 143 netisr_init(void) 144 { 145 int i; 146 147 TAILQ_INIT(&netreglist); 148 149 /* 150 * Create default per-cpu threads for generic protocol handling. 151 */ 152 for (i = 0; i < ncpus; ++i) { 153 lwkt_create(netmsg_service_loop, NULL, NULL, &netisr_cpu[i], 0, i, 154 "netisr_cpu %d", i); 155 netmsg_service_port_init(&netisr_cpu[i].td_msgport); 156 } 157 158 /* 159 * The netisr_afree_rport is a special reply port which automatically 160 * frees the replied message. The netisr_adone_rport simply marks 161 * the message as being done. The netisr_apanic_rport panics if 162 * the message is replied to. 163 */ 164 lwkt_initport_replyonly(&netisr_afree_rport, netisr_autofree_reply); 165 lwkt_initport_replyonly_null(&netisr_adone_rport); 166 lwkt_initport_panic(&netisr_apanic_rport); 167 168 /* 169 * The netisr_syncport is a special port which executes the message 170 * synchronously and waits for it if EASYNC is returned. 171 */ 172 lwkt_initport_putonly(&netisr_sync_port, netmsg_sync_putport); 173 } 174 175 SYSINIT(netisr, SI_SUB_PROTO_BEGIN, SI_ORDER_FIRST, netisr_init, NULL); 176 177 /* 178 * Finish initializing the message port for a netmsg service. This also 179 * registers the port for synchronous cleanup operations such as when an 180 * ifnet is being destroyed. There is no deregistration API yet. 181 */ 182 void 183 netmsg_service_port_init(lwkt_port_t port) 184 { 185 struct netmsg_port_registration *reg; 186 187 /* 188 * Override the putport function. Our custom function checks for 189 * self-references and executes such commands synchronously. 190 */ 191 if (netmsg_fwd_port_fn == NULL) 192 netmsg_fwd_port_fn = port->mp_putport; 193 KKASSERT(netmsg_fwd_port_fn == port->mp_putport); 194 port->mp_putport = netmsg_put_port; 195 196 /* 197 * Keep track of ports using the netmsg API so we can synchronize 198 * certain operations (such as freeing an ifnet structure) across all 199 * consumers. 200 */ 201 reg = kmalloc(sizeof(*reg), M_TEMP, M_WAITOK|M_ZERO); 202 reg->npr_port = port; 203 TAILQ_INSERT_TAIL(&netreglist, reg, npr_entry); 204 } 205 206 /* 207 * This function synchronizes the caller with all netmsg services. For 208 * example, if an interface is being removed we must make sure that all 209 * packets related to that interface complete processing before the structure 210 * can actually be freed. This sort of synchronization is an alternative to 211 * ref-counting the netif, removing the ref counting overhead in favor of 212 * placing additional overhead in the netif freeing sequence (where it is 213 * inconsequential). 214 */ 215 void 216 netmsg_service_sync(void) 217 { 218 struct netmsg_port_registration *reg; 219 struct netmsg smsg; 220 221 netmsg_init(&smsg, &curthread->td_msgport, 0, netmsg_sync_func); 222 223 TAILQ_FOREACH(reg, &netreglist, npr_entry) { 224 lwkt_domsg(reg->npr_port, &smsg.nm_lmsg, 0); 225 } 226 } 227 228 /* 229 * The netmsg function simply replies the message. API semantics require 230 * EASYNC to be returned if the netmsg function disposes of the message. 231 */ 232 static void 233 netmsg_sync_func(struct netmsg *msg) 234 { 235 lwkt_replymsg(&msg->nm_lmsg, 0); 236 } 237 238 /* 239 * Generic netmsg service loop. Some protocols may roll their own but all 240 * must do the basic command dispatch function call done here. 241 */ 242 void 243 netmsg_service_loop(void *arg) 244 { 245 struct netmsg *msg; 246 247 while ((msg = lwkt_waitport(&curthread->td_msgport, 0))) { 248 msg->nm_dispatch(msg); 249 } 250 } 251 252 /* 253 * Call the netisr directly. 254 * Queueing may be done in the msg port layer at its discretion. 255 */ 256 void 257 netisr_dispatch(int num, struct mbuf *m) 258 { 259 /* just queue it for now XXX JH */ 260 netisr_queue(num, m); 261 } 262 263 /* 264 * Same as netisr_dispatch(), but always queue. 265 * This is either used in places where we are not confident that 266 * direct dispatch is possible, or where queueing is required. 267 */ 268 int 269 netisr_queue(int num, struct mbuf *m) 270 { 271 struct netisr *ni; 272 struct netmsg_packet *pmsg; 273 lwkt_port_t port; 274 275 KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))), 276 ("netisr_queue: bad isr %d", num)); 277 278 ni = &netisrs[num]; 279 if (ni->ni_handler == NULL) { 280 kprintf("netisr_queue: unregistered isr %d\n", num); 281 return (EIO); 282 } 283 284 if ((port = ni->ni_mport(&m)) == NULL) 285 return (EIO); 286 287 pmsg = &m->m_hdr.mh_netmsg; 288 289 netmsg_init(&pmsg->nm_netmsg, &netisr_apanic_rport, 0, ni->ni_handler); 290 pmsg->nm_packet = m; 291 pmsg->nm_netmsg.nm_lmsg.u.ms_result = num; 292 lwkt_sendmsg(port, &pmsg->nm_netmsg.nm_lmsg); 293 return (0); 294 } 295 296 void 297 netisr_register(int num, lwkt_portfn_t mportfn, netisr_fn_t handler) 298 { 299 KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))), 300 ("netisr_register: bad isr %d", num)); 301 netmsg_init(&netisrs[num].ni_netmsg, &netisr_adone_rport, 0, NULL); 302 netisrs[num].ni_mport = mportfn; 303 netisrs[num].ni_handler = handler; 304 } 305 306 int 307 netisr_unregister(int num) 308 { 309 KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))), 310 ("unregister_netisr: bad isr number: %d\n", num)); 311 312 /* XXX JH */ 313 return (0); 314 } 315 316 /* 317 * Return message port for default handler thread on CPU 0. 318 */ 319 lwkt_port_t 320 cpu0_portfn(struct mbuf **mptr) 321 { 322 return (&netisr_cpu[0].td_msgport); 323 } 324 325 lwkt_port_t 326 cpu_portfn(int cpu) 327 { 328 return (&netisr_cpu[cpu].td_msgport); 329 } 330 331 /* ARGSUSED */ 332 lwkt_port_t 333 cpu0_soport(struct socket *so __unused, struct sockaddr *nam __unused, 334 int req __unused) 335 { 336 return (&netisr_cpu[0].td_msgport); 337 } 338 339 lwkt_port_t 340 sync_soport(struct socket *so __unused, struct sockaddr *nam __unused, 341 int req __unused) 342 { 343 return (&netisr_sync_port); 344 } 345 346 /* 347 * schednetisr() is used to call the netisr handler from the appropriate 348 * netisr thread for polling and other purposes. 349 * 350 * This function may be called from a hard interrupt or IPI and must be 351 * MP SAFE and non-blocking. We use a fixed per-cpu message instead of 352 * trying to allocate one. We must get ourselves onto the target cpu 353 * to safely check the MSGF_DONE bit on the message but since the message 354 * will be sent to that cpu anyway this does not add any extra work beyond 355 * what lwkt_sendmsg() would have already had to do to schedule the target 356 * thread. 357 */ 358 static void 359 schednetisr_remote(void *data) 360 { 361 int num = (int)data; 362 struct netisr *ni = &netisrs[num]; 363 lwkt_port_t port = &netisr_cpu[0].td_msgport; 364 struct netmsg *pmsg; 365 366 pmsg = &netisrs[num].ni_netmsg; 367 crit_enter(); 368 if (pmsg->nm_lmsg.ms_flags & MSGF_DONE) { 369 netmsg_init(pmsg, &netisr_adone_rport, 0, ni->ni_handler); 370 pmsg->nm_lmsg.u.ms_result = num; 371 lwkt_sendmsg(port, &pmsg->nm_lmsg); 372 } 373 crit_exit(); 374 } 375 376 void 377 schednetisr(int num) 378 { 379 KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))), 380 ("schednetisr: bad isr %d", num)); 381 #ifdef SMP 382 if (mycpu->gd_cpuid != 0) 383 lwkt_send_ipiq(globaldata_find(0), schednetisr_remote, (void *)num); 384 else 385 schednetisr_remote((void *)num); 386 #else 387 schednetisr_remote((void *)num); 388 #endif 389 } 390 391