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 1.23 2005/01/19 17:30:52 dillon Exp $ 39 */ 40 41 /* 42 * Copyright (c) 2003, 2004 Jeffrey M. Hsu. All rights reserved. 43 * 44 * License terms: all terms for the DragonFly license above plus the following: 45 * 46 * 4. All advertising materials mentioning features or use of this software 47 * must display the following acknowledgement: 48 * 49 * This product includes software developed by Jeffrey M. Hsu 50 * for the DragonFly Project. 51 * 52 * This requirement may be waived with permission from Jeffrey Hsu. 53 * This requirement will sunset and may be removed on July 8 2005, 54 * after which the standard DragonFly license (as shown above) will 55 * apply. 56 */ 57 58 #include <sys/param.h> 59 #include <sys/systm.h> 60 #include <sys/kernel.h> 61 #include <sys/malloc.h> 62 #include <sys/msgport.h> 63 #include <sys/proc.h> 64 #include <sys/interrupt.h> 65 #include <sys/socket.h> 66 #include <sys/sysctl.h> 67 #include <net/if.h> 68 #include <net/if_var.h> 69 #include <net/netisr.h> 70 #include <machine/cpufunc.h> 71 #include <machine/ipl.h> 72 73 #include <sys/thread2.h> 74 #include <sys/msgport2.h> 75 76 static int netmsg_sync_func(struct netmsg *msg); 77 78 struct netmsg_port_registration { 79 TAILQ_ENTRY(netmsg_port_registration) npr_entry; 80 lwkt_port_t npr_port; 81 }; 82 83 static struct netisr netisrs[NETISR_MAX]; 84 static TAILQ_HEAD(,netmsg_port_registration) netreglist; 85 86 /* Per-CPU thread to handle any protocol. */ 87 struct thread netisr_cpu[MAXCPU]; 88 lwkt_port netisr_afree_rport; 89 lwkt_port netisr_adone_rport; 90 lwkt_port netisr_sync_port; 91 92 /* 93 * netisr_afree_rport replymsg function, only used to handle async 94 * messages which the sender has abandoned to their fate. 95 */ 96 static void 97 netisr_autofree_reply(lwkt_port_t port, lwkt_msg_t msg) 98 { 99 free(msg, M_LWKTMSG); 100 } 101 102 /* 103 * We must construct a custom putport function (which runs in the context 104 * of the message originator) 105 * 106 * Our custom putport must check for self-referential messages, which can 107 * occur when the so_upcall routine is called (e.g. nfs). Self referential 108 * messages are executed synchronously. However, we must panic if the message 109 * is not marked DONE on completion because the self-referential case cannot 110 * block without deadlocking. 111 * 112 * note: ms_target_port does not need to be set when returning a synchronous 113 * error code. 114 */ 115 static int 116 netmsg_put_port(lwkt_port_t port, lwkt_msg_t lmsg) 117 { 118 int error; 119 120 if ((lmsg->ms_flags & MSGF_ASYNC) == 0 && port->mp_td == curthread) { 121 error = lmsg->ms_cmd.cm_func(lmsg); 122 if (error == EASYNC && (lmsg->ms_flags & MSGF_DONE) == 0) 123 panic("netmsg_put_port: self-referential deadlock on netport"); 124 return(error); 125 } else { 126 return(lwkt_default_putport(port, lmsg)); 127 } 128 } 129 130 /* 131 * UNIX DOMAIN sockets still have to run their uipc functions synchronously, 132 * because they depend on the user proc context for a number of things 133 * (like creds) which we have not yet incorporated into the message structure. 134 * 135 * However, we maintain or message/port abstraction. Having a special 136 * synchronous port which runs the commands synchronously gives us the 137 * ability to serialize operations in one place later on when we start 138 * removing the BGL. 139 * 140 * We clear MSGF_DONE prior to executing the message in order to close 141 * any potential replymsg races with the flags field. If a synchronous 142 * result code is returned we set MSGF_DONE again. MSGF_DONE's flag state 143 * must be correct or the caller will be confused. 144 */ 145 static int 146 netmsg_sync_putport(lwkt_port_t port, lwkt_msg_t lmsg) 147 { 148 int error; 149 150 lmsg->ms_flags &= ~MSGF_DONE; 151 lmsg->ms_target_port = port; /* required for abort */ 152 error = lmsg->ms_cmd.cm_func(lmsg); 153 if (error == EASYNC) 154 error = lwkt_waitmsg(lmsg); 155 else 156 lmsg->ms_flags |= MSGF_DONE; 157 return(error); 158 } 159 160 static void 161 netmsg_sync_abortport(lwkt_port_t port, lwkt_msg_t lmsg) 162 { 163 lmsg->ms_abort_port = lmsg->ms_reply_port; 164 lmsg->ms_flags |= MSGF_ABORTED; 165 lmsg->ms_abort.cm_func(lmsg); 166 } 167 168 static void 169 netisr_init(void) 170 { 171 int i; 172 173 TAILQ_INIT(&netreglist); 174 175 /* 176 * Create default per-cpu threads for generic protocol handling. 177 */ 178 for (i = 0; i < ncpus; ++i) { 179 lwkt_create(netmsg_service_loop, NULL, NULL, &netisr_cpu[i], 0, i, 180 "netisr_cpu %d", i); 181 netmsg_service_port_init(&netisr_cpu[i].td_msgport); 182 } 183 184 /* 185 * The netisr_afree_rport is a special reply port which automatically 186 * frees the replied message. The netisr_adone_rport() simply marks 187 * the message as being done. 188 */ 189 lwkt_initport(&netisr_afree_rport, NULL); 190 netisr_afree_rport.mp_replyport = netisr_autofree_reply; 191 lwkt_initport_null_rport(&netisr_adone_rport, NULL); 192 193 /* 194 * The netisr_syncport is a special port which executes the message 195 * synchronously and waits for it if EASYNC is returned. 196 */ 197 lwkt_initport(&netisr_sync_port, NULL); 198 netisr_sync_port.mp_putport = netmsg_sync_putport; 199 netisr_sync_port.mp_abortport = netmsg_sync_abortport; 200 } 201 202 SYSINIT(netisr, SI_SUB_PROTO_BEGIN, SI_ORDER_FIRST, netisr_init, NULL); 203 204 /* 205 * Finish initializing the message port for a netmsg service. This also 206 * registers the port for synchronous cleanup operations such as when an 207 * ifnet is being destroyed. There is no deregistration API yet. 208 */ 209 void 210 netmsg_service_port_init(lwkt_port_t port) 211 { 212 struct netmsg_port_registration *reg; 213 214 /* 215 * Override the putport function. Our custom function checks for 216 * self-references and executes such commands synchronously. 217 */ 218 port->mp_putport = netmsg_put_port; 219 220 /* 221 * Keep track of ports using the netmsg API so we can synchronize 222 * certain operations (such as freeing an ifnet structure) across all 223 * consumers. 224 */ 225 reg = malloc(sizeof(*reg), M_TEMP, M_WAITOK|M_ZERO); 226 reg->npr_port = port; 227 TAILQ_INSERT_TAIL(&netreglist, reg, npr_entry); 228 } 229 230 /* 231 * This function synchronizes the caller with all netmsg services. For 232 * example, if an interface is being removed we must make sure that all 233 * packets related to that interface complete processing before the structure 234 * can actually be freed. This sort of synchronization is an alternative to 235 * ref-counting the netif, removing the ref counting overhead in favor of 236 * placing additional overhead in the netif freeing sequence (where it is 237 * inconsequential). 238 */ 239 void 240 netmsg_service_sync(void) 241 { 242 struct netmsg_port_registration *reg; 243 struct netmsg smsg; 244 245 lwkt_initmsg(&smsg.nm_lmsg, &curthread->td_msgport, 0, 246 lwkt_cmd_func((void *)netmsg_sync_func), lwkt_cmd_op_none); 247 248 TAILQ_FOREACH(reg, &netreglist, npr_entry) { 249 lwkt_domsg(reg->npr_port, &smsg.nm_lmsg); 250 } 251 } 252 253 /* 254 * The netmsg function simply replies the message. API semantics require 255 * EASYNC to be returned if the netmsg function disposes of the message. 256 */ 257 static 258 int 259 netmsg_sync_func(struct netmsg *msg) 260 { 261 lwkt_replymsg(&msg->nm_lmsg, 0); 262 return(EASYNC); 263 } 264 265 /* 266 * Generic netmsg service loop. Some protocols may roll their own but all 267 * must do the basic command dispatch function call done here. 268 */ 269 void 270 netmsg_service_loop(void *arg) 271 { 272 struct netmsg *msg; 273 274 while ((msg = lwkt_waitport(&curthread->td_msgport, NULL))) { 275 msg->nm_lmsg.ms_cmd.cm_func(&msg->nm_lmsg); 276 } 277 } 278 279 /* 280 * Call the netisr directly. 281 * Queueing may be done in the msg port layer at its discretion. 282 */ 283 void 284 netisr_dispatch(int num, struct mbuf *m) 285 { 286 /* just queue it for now XXX JH */ 287 netisr_queue(num, m); 288 } 289 290 /* 291 * Same as netisr_dispatch(), but always queue. 292 * This is either used in places where we are not confident that 293 * direct dispatch is possible, or where queueing is required. 294 */ 295 int 296 netisr_queue(int num, struct mbuf *m) 297 { 298 struct netisr *ni; 299 struct netmsg_packet *pmsg; 300 lwkt_port_t port; 301 302 KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))), 303 ("netisr_queue: bad isr %d", num)); 304 305 ni = &netisrs[num]; 306 if (ni->ni_handler == NULL) { 307 printf("netisr_queue: unregistered isr %d\n", num); 308 return (EIO); 309 } 310 311 if ((port = ni->ni_mport(&m)) == NULL) 312 return (EIO); 313 314 /* use better message allocation system with limits later XXX JH */ 315 pmsg = malloc(sizeof(struct netmsg_packet), M_LWKTMSG, M_WAITOK); 316 317 lwkt_initmsg(&pmsg->nm_lmsg, &netisr_afree_rport, 0, 318 lwkt_cmd_func((void *)ni->ni_handler), lwkt_cmd_op_none); 319 pmsg->nm_packet = m; 320 pmsg->nm_lmsg.u.ms_result = num; 321 lwkt_sendmsg(port, &pmsg->nm_lmsg); 322 return (0); 323 } 324 325 void 326 netisr_register(int num, lwkt_portfn_t mportfn, netisr_fn_t handler) 327 { 328 KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))), 329 ("netisr_register: bad isr %d", num)); 330 lwkt_initmsg(&netisrs[num].ni_netmsg.nm_lmsg, &netisr_adone_rport, 0, 331 lwkt_cmd_op_none, lwkt_cmd_op_none); 332 netisrs[num].ni_mport = mportfn; 333 netisrs[num].ni_handler = handler; 334 } 335 336 int 337 netisr_unregister(int num) 338 { 339 KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))), 340 ("unregister_netisr: bad isr number: %d\n", num)); 341 342 /* XXX JH */ 343 return (0); 344 } 345 346 /* 347 * Return message port for default handler thread on CPU 0. 348 */ 349 lwkt_port_t 350 cpu0_portfn(struct mbuf **mptr) 351 { 352 return (&netisr_cpu[0].td_msgport); 353 } 354 355 /* ARGSUSED */ 356 lwkt_port_t 357 cpu0_soport(struct socket *so __unused, struct sockaddr *nam __unused, 358 int req __unused) 359 { 360 return (&netisr_cpu[0].td_msgport); 361 } 362 363 lwkt_port_t 364 sync_soport(struct socket *so __unused, struct sockaddr *nam __unused, 365 int req __unused) 366 { 367 return (&netisr_sync_port); 368 } 369 370 /* 371 * schednetisr() is used to call the netisr handler from the appropriate 372 * netisr thread for polling and other purposes. 373 * 374 * This function may be called from a hard interrupt or IPI and must be 375 * MP SAFE and non-blocking. We use a fixed per-cpu message instead of 376 * trying to allocate one. We must get ourselves onto the target cpu 377 * to safely check the MSGF_DONE bit on the message but since the message 378 * will be sent to that cpu anyway this does not add any extra work beyond 379 * what lwkt_sendmsg() would have already had to do to schedule the target 380 * thread. 381 */ 382 static void 383 schednetisr_remote(void *data) 384 { 385 int num = (int)data; 386 struct netisr *ni = &netisrs[num]; 387 lwkt_port_t port = &netisr_cpu[0].td_msgport; 388 struct netmsg *pmsg; 389 390 pmsg = &netisrs[num].ni_netmsg; 391 crit_enter(); 392 if (pmsg->nm_lmsg.ms_flags & MSGF_DONE) { 393 lwkt_initmsg(&pmsg->nm_lmsg, &netisr_adone_rport, 0, 394 lwkt_cmd_func((void *)ni->ni_handler), lwkt_cmd_op_none); 395 pmsg->nm_lmsg.u.ms_result = num; 396 lwkt_sendmsg(port, &pmsg->nm_lmsg); 397 } 398 crit_exit(); 399 } 400 401 void 402 schednetisr(int num) 403 { 404 KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))), 405 ("schednetisr: bad isr %d", num)); 406 #ifdef SMP 407 if (mycpu->gd_cpuid != 0) 408 lwkt_send_ipiq(globaldata_find(0), schednetisr_remote, (void *)num); 409 else 410 schednetisr_remote((void *)num); 411 #else 412 schednetisr_remote((void *)num); 413 #endif 414 } 415 416