sys/net/netisr.c

   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 1.43 2008/06/23 11:57:19 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 static int netisr_mpsafe_thread = 0;
  79 TUNABLE_INT("netisr.mpsafe_thread", &netisr_mpsafe_thread);
  80
  81 /*
  82  * netisr_afree_rport replymsg function, only used to handle async
  83  * messages which the sender has abandoned to their fate.
  84  */
  85 static void
  86 netisr_autofree_reply(lwkt_port_t port, lwkt_msg_t msg)
  87 {
  88     kfree(msg, M_LWKTMSG);
  89 }
  90
  91 /*
  92  * We need a custom putport function to handle the case where the
  93  * message target is the current thread's message port.  This case
  94  * can occur when the TCP or UDP stack does a direct callback to NFS and NFS
  95  * then turns around and executes a network operation synchronously.
  96  *
  97  * To prevent deadlocking, we must execute these self-referential messages
  98  * synchronously, effectively turning the message into a glorified direct
  99  * procedure call back into the protocol stack.  The operation must be
 100  * complete on return or we will deadlock, so panic if it isn't.
 101  */
 102 static int
 103 netmsg_put_port(lwkt_port_t port, lwkt_msg_t lmsg)
 104 {
 105     netmsg_t netmsg = (void *)lmsg;
 106
 107     if ((lmsg->ms_flags & MSGF_SYNC) && port == &curthread->td_msgport) {
 108         netmsg->nm_dispatch(netmsg);
 109         if ((lmsg->ms_flags & MSGF_DONE) == 0)
 110             panic("netmsg_put_port: self-referential deadlock on netport");
 111         return(EASYNC);
 112     } else {
 113         return(netmsg_fwd_port_fn(port, lmsg));
 114     }
 115 }
 116
 117 /*
 118  * UNIX DOMAIN sockets still have to run their uipc functions synchronously,
 119  * because they depend on the user proc context for a number of things
 120  * (like creds) which we have not yet incorporated into the message structure.
 121  *
 122  * However, we maintain or message/port abstraction.  Having a special
 123  * synchronous port which runs the commands synchronously gives us the
 124  * ability to serialize operations in one place later on when we start
 125  * removing the BGL.
 126  */
 127 static int
 128 netmsg_sync_putport(lwkt_port_t port, lwkt_msg_t lmsg)
 129 {
 130     netmsg_t netmsg = (void *)lmsg;
 131
 132     KKASSERT((lmsg->ms_flags & MSGF_DONE) == 0);
 133
 134     lmsg->ms_target_port = port;        /* required for abort */
 135     netmsg->nm_dispatch(netmsg);
 136     return(EASYNC);
 137 }
 138
 139 static void
 140 netisr_init(void)
 141 {
 142     int i;
 143
 144     TAILQ_INIT(&netreglist);
 145
 146     /*
 147      * Create default per-cpu threads for generic protocol handling.
 148      */
 149     for (i = 0; i < ncpus; ++i) {
 150         lwkt_create(netisr_mpsafe_thread ?
 151                     netmsg_service_loop_mpsafe : netmsg_service_loop,
 152                     NULL, NULL, &netisr_cpu[i],
 153                     0, i, "netisr_cpu %d", i);
 154         netmsg_service_port_init(&netisr_cpu[i].td_msgport);
 155     }
 156
 157     /*
 158      * The netisr_afree_rport is a special reply port which automatically
 159      * frees the replied message.  The netisr_adone_rport simply marks
 160      * the message as being done.  The netisr_apanic_rport panics if
 161      * the message is replied to.
 162      */
 163     lwkt_initport_replyonly(&netisr_afree_rport, netisr_autofree_reply);
 164     lwkt_initport_replyonly_null(&netisr_adone_rport);
 165     lwkt_initport_panic(&netisr_apanic_rport);
 166
 167     /*
 168      * The netisr_syncport is a special port which executes the message
 169      * synchronously and waits for it if EASYNC is returned.
 170      */
 171     lwkt_initport_putonly(&netisr_sync_port, netmsg_sync_putport);
 172 }
 173
 174 SYSINIT(netisr, SI_SUB_PRE_DRIVERS, SI_ORDER_FIRST, netisr_init, NULL);
 175
 176 /*
 177  * Finish initializing the message port for a netmsg service.  This also
 178  * registers the port for synchronous cleanup operations such as when an
 179  * ifnet is being destroyed.  There is no deregistration API yet.
 180  */
 181 void
 182 netmsg_service_port_init(lwkt_port_t port)
 183 {
 184     struct netmsg_port_registration *reg;
 185
 186     /*
 187      * Override the putport function.  Our custom function checks for
 188      * self-references and executes such commands synchronously.
 189      */
 190     if (netmsg_fwd_port_fn == NULL)
 191         netmsg_fwd_port_fn = port->mp_putport;
 192     KKASSERT(netmsg_fwd_port_fn == port->mp_putport);
 193     port->mp_putport = netmsg_put_port;
 194
 195     /*
 196      * Keep track of ports using the netmsg API so we can synchronize
 197      * certain operations (such as freeing an ifnet structure) across all
 198      * consumers.
 199      */
 200     reg = kmalloc(sizeof(*reg), M_TEMP, M_WAITOK|M_ZERO);
 201     reg->npr_port = port;
 202     TAILQ_INSERT_TAIL(&netreglist, reg, npr_entry);
 203 }
 204
 205 /*
 206  * This function synchronizes the caller with all netmsg services.  For
 207  * example, if an interface is being removed we must make sure that all
 208  * packets related to that interface complete processing before the structure
 209  * can actually be freed.  This sort of synchronization is an alternative to
 210  * ref-counting the netif, removing the ref counting overhead in favor of
 211  * placing additional overhead in the netif freeing sequence (where it is
 212  * inconsequential).
 213  */
 214 void
 215 netmsg_service_sync(void)
 216 {
 217     struct netmsg_port_registration *reg;
 218     struct netmsg smsg;
 219
 220     netmsg_init(&smsg, &curthread->td_msgport, 0, netmsg_sync_func);
 221
 222     TAILQ_FOREACH(reg, &netreglist, npr_entry) {
 223         lwkt_domsg(reg->npr_port, &smsg.nm_lmsg, 0);
 224     }
 225 }
 226
 227 /*
 228  * The netmsg function simply replies the message.  API semantics require
 229  * EASYNC to be returned if the netmsg function disposes of the message.
 230  */
 231 static void
 232 netmsg_sync_func(struct netmsg *msg)
 233 {
 234     lwkt_replymsg(&msg->nm_lmsg, 0);
 235 }
 236
 237 /*
 238  * Generic netmsg service loop.  Some protocols may roll their own but all
 239  * must do the basic command dispatch function call done here.
 240  */
 241 void
 242 netmsg_service_loop(void *arg)
 243 {
 244     struct netmsg *msg;
 245
 246     while ((msg = lwkt_waitport(&curthread->td_msgport, 0))) {
 247         msg->nm_dispatch(msg);
 248     }
 249 }
 250
 251 /*
 252  * MPSAFE version of netmsg_service_loop()
 253  */
 254 void
 255 netmsg_service_loop_mpsafe(void *arg)
 256 {
 257     rel_mplock();
 258     netmsg_service_loop(arg);
 259 }
 260
 261 /*
 262  * Call the netisr directly.
 263  * Queueing may be done in the msg port layer at its discretion.
 264  */
 265 void
 266 netisr_dispatch(int num, struct mbuf *m)
 267 {
 268     /* just queue it for now XXX JH */
 269     netisr_queue(num, m);
 270 }
 271
 272 /*
 273  * Same as netisr_dispatch(), but always queue.
 274  * This is either used in places where we are not confident that
 275  * direct dispatch is possible, or where queueing is required.
 276  */
 277 int
 278 netisr_queue(int num, struct mbuf *m)
 279 {
 280     struct netisr *ni;
 281     struct netmsg_packet *pmsg;
 282     lwkt_port_t port;
 283
 284     KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
 285             ("%s: bad isr %d", __func__, num));
 286
 287     ni = &netisrs[num];
 288     if (ni->ni_handler == NULL) {
 289         kprintf("%s: unregistered isr %d\n", __func__, num);
 290         m_freem(m);
 291         return (EIO);
 292     }
 293
 294     if ((port = ni->ni_mport(&m)) == NULL)
 295         return (EIO);
 296
 297     pmsg = &m->m_hdr.mh_netmsg;
 298
 299     netmsg_init(&pmsg->nm_netmsg, &netisr_apanic_rport, 0, ni->ni_handler);
 300     pmsg->nm_packet = m;
 301     pmsg->nm_netmsg.nm_lmsg.u.ms_result = num;
 302     lwkt_sendmsg(port, &pmsg->nm_netmsg.nm_lmsg);
 303     return (0);
 304 }
 305
 306 void
 307 netisr_register(int num, lwkt_portfn_t mportfn, netisr_fn_t handler)
 308 {
 309     KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
 310         ("netisr_register: bad isr %d", num));
 311     netmsg_init(&netisrs[num].ni_netmsg, &netisr_adone_rport, 0, NULL);
 312     netisrs[num].ni_mport = mportfn;
 313     netisrs[num].ni_handler = handler;
 314 }
 315
 316 int
 317 netisr_unregister(int num)
 318 {
 319     KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
 320         ("unregister_netisr: bad isr number: %d\n", num));
 321
 322     /* XXX JH */
 323     return (0);
 324 }
 325
 326 /*
 327  * Return message port for default handler thread on CPU 0.
 328  */
 329 lwkt_port_t
 330 cpu0_portfn(struct mbuf **mptr)
 331 {
 332     return (&netisr_cpu[0].td_msgport);
 333 }
 334
 335 lwkt_port_t
 336 cpu_portfn(int cpu)
 337 {
 338     return (&netisr_cpu[cpu].td_msgport);
 339 }
 340
 341 /* ARGSUSED */
 342 lwkt_port_t
 343 cpu0_soport(struct socket *so __unused, struct sockaddr *nam __unused,
 344             struct mbuf **dummy __unused, int req __unused)
 345 {
 346     return (&netisr_cpu[0].td_msgport);
 347 }
 348
 349 lwkt_port_t
 350 sync_soport(struct socket *so __unused, struct sockaddr *nam __unused,
 351             struct mbuf **dummy __unused, int req __unused)
 352 {
 353     return (&netisr_sync_port);
 354 }
 355
 356 /*
 357  * schednetisr() is used to call the netisr handler from the appropriate
 358  * netisr thread for polling and other purposes.
 359  *
 360  * This function may be called from a hard interrupt or IPI and must be
 361  * MP SAFE and non-blocking.  We use a fixed per-cpu message instead of
 362  * trying to allocate one.  We must get ourselves onto the target cpu
 363  * to safely check the MSGF_DONE bit on the message but since the message
 364  * will be sent to that cpu anyway this does not add any extra work beyond
 365  * what lwkt_sendmsg() would have already had to do to schedule the target
 366  * thread.
 367  */
 368 static void
 369 schednetisr_remote(void *data)
 370 {
 371     int num = (int)data;
 372     struct netisr *ni = &netisrs[num];
 373     lwkt_port_t port = &netisr_cpu[0].td_msgport;
 374     struct netmsg *pmsg;
 375
 376     pmsg = &netisrs[num].ni_netmsg;
 377     crit_enter();
 378     if (pmsg->nm_lmsg.ms_flags & MSGF_DONE) {
 379         netmsg_init(pmsg, &netisr_adone_rport, 0, ni->ni_handler);
 380         pmsg->nm_lmsg.u.ms_result = num;
 381         lwkt_sendmsg(port, &pmsg->nm_lmsg);
 382     }
 383     crit_exit();
 384 }
 385
 386 void
 387 schednetisr(int num)
 388 {
 389     KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
 390         ("schednetisr: bad isr %d", num));
 391 #ifdef SMP
 392     if (mycpu->gd_cpuid != 0)
 393         lwkt_send_ipiq(globaldata_find(0), schednetisr_remote, (void *)num);
 394     else
 395         schednetisr_remote((void *)num);
 396 #else
 397     schednetisr_remote((void *)num);
 398 #endif
 399 }
 400
 401 lwkt_port_t
 402 netisr_mport(int num, struct mbuf **m0)
 403 {
 404     struct netisr *ni;
 405     struct netmsg_packet *pmsg;
 406     lwkt_port_t port;
 407     struct mbuf *m = *m0;
 408
 409     *m0 = NULL;
 410
 411     KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
 412             ("%s: bad isr %d", __func__, num));
 413
 414     ni = &netisrs[num];
 415     if (ni->ni_handler == NULL) {
 416         kprintf("%s: unregistered isr %d\n", __func__, num);
 417         m_freem(m);
 418         return NULL;
 419     }
 420
 421     if ((port = ni->ni_mport(&m)) == NULL)
 422         return NULL;
 423
 424     pmsg = &m->m_hdr.mh_netmsg;
 425
 426     netmsg_init(&pmsg->nm_netmsg, &netisr_apanic_rport, 0, ni->ni_handler);
 427     pmsg->nm_packet = m;
 428     pmsg->nm_netmsg.nm_lmsg.u.ms_result = num;
 429
 430     *m0 = m;
 431     return port;
 432 }
 433
 434 lwkt_port_t
 435 netisr_find_port(int num, struct mbuf **m0)
 436 {
 437     struct netisr *ni;
 438     lwkt_port_t port;
 439     struct mbuf *m = *m0;
 440
 441     *m0 = NULL;
 442
 443     KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
 444             ("%s: bad isr %d", __func__, num));
 445
 446     ni = &netisrs[num];
 447     if (ni->ni_mport == NULL) {
 448         kprintf("%s: unregistered isr %d\n", __func__, num);
 449         m_freem(m);
 450         return NULL;
 451     }
 452
 453     if ((port = ni->ni_mport(&m)) == NULL)
 454         return NULL;
 455
 456     *m0 = m;
 457     return port;
 458 }
 459
 460 void
 461 netisr_run(int num, struct mbuf *m)
 462 {
 463     struct netisr *ni;
 464     struct netmsg_packet *pmsg;
 465
 466     KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
 467             ("%s: bad isr %d", __func__, num));
 468
 469     ni = &netisrs[num];
 470     if (ni->ni_handler == NULL) {
 471         kprintf("%s: unregistered isr %d\n", __func__, num);
 472         m_freem(m);
 473         return;
 474     }
 475
 476     pmsg = &m->m_hdr.mh_netmsg;
 477
 478     netmsg_init(&pmsg->nm_netmsg, &netisr_apanic_rport, 0, ni->ni_handler);
 479     pmsg->nm_packet = m;
 480     pmsg->nm_netmsg.nm_lmsg.u.ms_result = num;
 481
 482     ni->ni_handler(&pmsg->nm_netmsg);
 483 }