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 1.49 2008/11/01 10:29:31 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 #define NETISR_GET_MPLOCK(ni) \
  60 do { \
  61     if (((ni)->ni_flags & NETISR_FLAG_MPSAFE) == 0) \
  62         get_mplock(); \
  63 } while (0)
  64
  65 #define NETISR_REL_MPLOCK(ni) \
  66 do { \
  67     if (((ni)->ni_flags & NETISR_FLAG_MPSAFE) == 0) \
  68         rel_mplock(); \
  69 } while (0)
  70
  71 static void netmsg_sync_func(struct netmsg *msg);
  72
  73 struct netmsg_port_registration {
  74     TAILQ_ENTRY(netmsg_port_registration) npr_entry;
  75     lwkt_port_t npr_port;
  76 };
  77
  78 static struct netisr netisrs[NETISR_MAX];
  79 static TAILQ_HEAD(,netmsg_port_registration) netreglist;
  80
  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;
  87
  88 static int (*netmsg_fwd_port_fn)(lwkt_port_t, lwkt_msg_t);
  89
  90 static int netisr_mpsafe_thread = NETMSG_SERVICE_ADAPTIVE;
  91 TUNABLE_INT("net.netisr.mpsafe_thread", &netisr_mpsafe_thread);
  92
  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)");
  97
  98 static __inline int
  99 NETISR_TO_MSGF(const struct netisr *ni)
 100 {
 101     int msg_flags = 0;
 102
 103     if (ni->ni_flags & NETISR_FLAG_MPSAFE)
 104         msg_flags |= MSGF_MPSAFE;
 105     return msg_flags;
 106 }
 107
 108 /*
 109  * netisr_afree_rport replymsg function, only used to handle async
 110  * messages which the sender has abandoned to their fate.
 111  */
 112 static void
 113 netisr_autofree_reply(lwkt_port_t port, lwkt_msg_t msg)
 114 {
 115     kfree(msg, M_LWKTMSG);
 116 }
 117
 118 /*
 119  * We need a custom putport function to handle the case where the
 120  * message target is the current thread's message port.  This case
 121  * can occur when the TCP or UDP stack does a direct callback to NFS and NFS
 122  * then turns around and executes a network operation synchronously.
 123  *
 124  * To prevent deadlocking, we must execute these self-referential messages
 125  * synchronously, effectively turning the message into a glorified direct
 126  * procedure call back into the protocol stack.  The operation must be
 127  * complete on return or we will deadlock, so panic if it isn't.
 128  */
 129 static int
 130 netmsg_put_port(lwkt_port_t port, lwkt_msg_t lmsg)
 131 {
 132     netmsg_t netmsg = (void *)lmsg;
 133
 134     if ((lmsg->ms_flags & MSGF_SYNC) && port == &curthread->td_msgport) {
 135         netmsg->nm_dispatch(netmsg);
 136         if ((lmsg->ms_flags & MSGF_DONE) == 0)
 137             panic("netmsg_put_port: self-referential deadlock on netport");
 138         return(EASYNC);
 139     } else {
 140         return(netmsg_fwd_port_fn(port, lmsg));
 141     }
 142 }
 143
 144 /*
 145  * UNIX DOMAIN sockets still have to run their uipc functions synchronously,
 146  * because they depend on the user proc context for a number of things
 147  * (like creds) which we have not yet incorporated into the message structure.
 148  *
 149  * However, we maintain or message/port abstraction.  Having a special
 150  * synchronous port which runs the commands synchronously gives us the
 151  * ability to serialize operations in one place later on when we start
 152  * removing the BGL.
 153  */
 154 static int
 155 netmsg_sync_putport(lwkt_port_t port, lwkt_msg_t lmsg)
 156 {
 157     netmsg_t netmsg = (void *)lmsg;
 158
 159     KKASSERT((lmsg->ms_flags & MSGF_DONE) == 0);
 160
 161     lmsg->ms_target_port = port;        /* required for abort */
 162     netmsg->nm_dispatch(netmsg);
 163     return(EASYNC);
 164 }
 165
 166 static void
 167 netisr_init(void)
 168 {
 169     int i;
 170
 171     TAILQ_INIT(&netreglist);
 172
 173     /*
 174      * Create default per-cpu threads for generic protocol handling.
 175      */
 176     for (i = 0; i < ncpus; ++i) {
 177         lwkt_create(netmsg_service_loop, &netisr_mpsafe_thread, NULL,
 178                     &netisr_cpu[i], TDF_NETWORK | TDF_MPSAFE, i,
 179                     "netisr_cpu %d", i);
 180         netmsg_service_port_init(&netisr_cpu[i].td_msgport);
 181     }
 182
 183     /*
 184      * The netisr_afree_rport is a special reply port which automatically
 185      * frees the replied message.  The netisr_adone_rport simply marks
 186      * the message as being done.  The netisr_apanic_rport panics if
 187      * the message is replied to.
 188      */
 189     lwkt_initport_replyonly(&netisr_afree_rport, netisr_autofree_reply);
 190     lwkt_initport_replyonly_null(&netisr_adone_rport);
 191     lwkt_initport_panic(&netisr_apanic_rport);
 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_putonly(&netisr_sync_port, netmsg_sync_putport);
 198 }
 199
 200 SYSINIT(netisr, SI_SUB_PRE_DRIVERS, SI_ORDER_FIRST, netisr_init, NULL);
 201
 202 /*
 203  * Finish initializing the message port for a netmsg service.  This also
 204  * registers the port for synchronous cleanup operations such as when an
 205  * ifnet is being destroyed.  There is no deregistration API yet.
 206  */
 207 void
 208 netmsg_service_port_init(lwkt_port_t port)
 209 {
 210     struct netmsg_port_registration *reg;
 211
 212     /*
 213      * Override the putport function.  Our custom function checks for
 214      * self-references and executes such commands synchronously.
 215      */
 216     if (netmsg_fwd_port_fn == NULL)
 217         netmsg_fwd_port_fn = port->mp_putport;
 218     KKASSERT(netmsg_fwd_port_fn == port->mp_putport);
 219     port->mp_putport = netmsg_put_port;
 220
 221     /*
 222      * Keep track of ports using the netmsg API so we can synchronize
 223      * certain operations (such as freeing an ifnet structure) across all
 224      * consumers.
 225      */
 226     reg = kmalloc(sizeof(*reg), M_TEMP, M_WAITOK|M_ZERO);
 227     reg->npr_port = port;
 228     TAILQ_INSERT_TAIL(&netreglist, reg, npr_entry);
 229 }
 230
 231 /*
 232  * This function synchronizes the caller with all netmsg services.  For
 233  * example, if an interface is being removed we must make sure that all
 234  * packets related to that interface complete processing before the structure
 235  * can actually be freed.  This sort of synchronization is an alternative to
 236  * ref-counting the netif, removing the ref counting overhead in favor of
 237  * placing additional overhead in the netif freeing sequence (where it is
 238  * inconsequential).
 239  */
 240 void
 241 netmsg_service_sync(void)
 242 {
 243     struct netmsg_port_registration *reg;
 244     struct netmsg smsg;
 245
 246     netmsg_init(&smsg, &curthread->td_msgport, MSGF_MPSAFE, netmsg_sync_func);
 247
 248     TAILQ_FOREACH(reg, &netreglist, npr_entry) {
 249         lwkt_domsg(reg->npr_port, &smsg.nm_lmsg, 0);
 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 void
 258 netmsg_sync_func(struct netmsg *msg)
 259 {
 260     lwkt_replymsg(&msg->nm_lmsg, 0);
 261 }
 262
 263 /*
 264  * Return current BGL lock state (1:locked, 0: unlocked)
 265  */
 266 int
 267 netmsg_service(struct netmsg *msg, int mpsafe_mode, int mplocked)
 268 {
 269     /*
 270      * Adjust the mplock dynamically.
 271      */
 272     switch (mpsafe_mode) {
 273     case NETMSG_SERVICE_ADAPTIVE: /* Adaptive BGL */
 274         if (msg->nm_lmsg.ms_flags & MSGF_MPSAFE) {
 275             if (mplocked) {
 276                 rel_mplock();
 277                 mplocked = 0;
 278             }
 279             msg->nm_dispatch(msg);
 280             /* Leave mpunlocked */
 281         } else {
 282             if (!mplocked) {
 283                 get_mplock();
 284                 /* mplocked = 1; not needed */
 285             }
 286             msg->nm_dispatch(msg);
 287             rel_mplock();
 288             mplocked = 0;
 289             /* Leave mpunlocked, next msg might be mpsafe */
 290         }
 291         break;
 292
 293     case NETMSG_SERVICE_MPSAFE: /* No BGL */
 294         if (mplocked) {
 295             rel_mplock();
 296             mplocked = 0;
 297         }
 298         msg->nm_dispatch(msg);
 299         /* Leave mpunlocked */
 300         break;
 301
 302     default: /* BGL */
 303         if (!mplocked) {
 304             get_mplock();
 305             mplocked = 1;
 306         }
 307         msg->nm_dispatch(msg);
 308         /* Leave mplocked */
 309         break;
 310     }
 311     return mplocked;
 312 }
 313
 314 /*
 315  * Generic netmsg service loop.  Some protocols may roll their own but all
 316  * must do the basic command dispatch function call done here.
 317  */
 318 void
 319 netmsg_service_loop(void *arg)
 320 {
 321     struct netmsg *msg;
 322     int mplocked, *mpsafe_mode = arg;
 323
 324     /*
 325      * Thread was started with TDF_MPSAFE
 326      */
 327     mplocked = 0;
 328
 329     /*
 330      * Loop on netmsgs
 331      */
 332     while ((msg = lwkt_waitport(&curthread->td_msgport, 0))) {
 333         mplocked = netmsg_service(msg, *mpsafe_mode, mplocked);
 334     }
 335 }
 336
 337 /*
 338  * Call the netisr directly.
 339  * Queueing may be done in the msg port layer at its discretion.
 340  */
 341 void
 342 netisr_dispatch(int num, struct mbuf *m)
 343 {
 344     /* just queue it for now XXX JH */
 345     netisr_queue(num, m);
 346 }
 347
 348 /*
 349  * Same as netisr_dispatch(), but always queue.
 350  * This is either used in places where we are not confident that
 351  * direct dispatch is possible, or where queueing is required.
 352  */
 353 int
 354 netisr_queue(int num, struct mbuf *m)
 355 {
 356     struct netisr *ni;
 357     struct netmsg_packet *pmsg;
 358     lwkt_port_t port;
 359
 360     KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
 361             ("%s: bad isr %d", __func__, num));
 362
 363     ni = &netisrs[num];
 364     if (ni->ni_handler == NULL) {
 365         kprintf("%s: unregistered isr %d\n", __func__, num);
 366         m_freem(m);
 367         return (EIO);
 368     }
 369
 370     if ((port = ni->ni_mport(&m)) == NULL)
 371         return (EIO);
 372
 373     pmsg = &m->m_hdr.mh_netmsg;
 374
 375     netmsg_init(&pmsg->nm_netmsg, &netisr_apanic_rport, NETISR_TO_MSGF(ni),
 376                 ni->ni_handler);
 377     pmsg->nm_packet = m;
 378     pmsg->nm_netmsg.nm_lmsg.u.ms_result = num;
 379     lwkt_sendmsg(port, &pmsg->nm_netmsg.nm_lmsg);
 380     return (0);
 381 }
 382
 383 void
 384 netisr_register(int num, pkt_portfn_t mportfn,
 385                 pktinfo_portfn_t mportfn_pktinfo, netisr_fn_t handler,
 386                 uint32_t flags)
 387 {
 388     struct netisr *ni;
 389
 390     KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
 391         ("netisr_register: bad isr %d", num));
 392     ni = &netisrs[num];
 393
 394     ni->ni_mport = mportfn;
 395     ni->ni_mport_pktinfo = mportfn_pktinfo;
 396     ni->ni_handler = handler;
 397     ni->ni_flags = flags;
 398     netmsg_init(&ni->ni_netmsg, &netisr_adone_rport, NETISR_TO_MSGF(ni), NULL);
 399 }
 400
 401 int
 402 netisr_unregister(int num)
 403 {
 404     KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
 405         ("unregister_netisr: bad isr number: %d\n", num));
 406
 407     /* XXX JH */
 408     return (0);
 409 }
 410
 411 /*
 412  * Return message port for default handler thread on CPU 0.
 413  */
 414 lwkt_port_t
 415 cpu0_portfn(struct mbuf **mptr)
 416 {
 417     struct mbuf *m = *mptr;
 418     int cpu = 0;
 419
 420     m->m_pkthdr.hash = cpu;
 421     m->m_flags |= M_HASH;
 422     return (&netisr_cpu[cpu].td_msgport);
 423 }
 424
 425 lwkt_port_t
 426 cpu_portfn(int cpu)
 427 {
 428     return (&netisr_cpu[cpu].td_msgport);
 429 }
 430
 431 /*
 432  * If the current thread is a network protocol thread (TDF_NETWORK),
 433  * then return the current thread's message port.
 434  * XXX Else, return the current CPU's netisr message port.
 435  */
 436 lwkt_port_t
 437 cur_netport(void)
 438 {
 439     if (curthread->td_flags & TDF_NETWORK)
 440         return &curthread->td_msgport;
 441     else
 442         return cpu_portfn(mycpuid);
 443 }
 444
 445 /* ARGSUSED */
 446 lwkt_port_t
 447 cpu0_soport(struct socket *so __unused, struct sockaddr *nam __unused,
 448             struct mbuf **dummy __unused, int req __unused)
 449 {
 450     return (&netisr_cpu[0].td_msgport);
 451 }
 452
 453 lwkt_port_t
 454 cpu0_ctlport(int cmd __unused, struct sockaddr *sa __unused,
 455              void *extra __unused)
 456 {
 457     return (&netisr_cpu[0].td_msgport);
 458 }
 459
 460 lwkt_port_t
 461 sync_soport(struct socket *so __unused, struct sockaddr *nam __unused,
 462             struct mbuf **dummy __unused, int req __unused)
 463 {
 464     return (&netisr_sync_port);
 465 }
 466
 467 /*
 468  * schednetisr() is used to call the netisr handler from the appropriate
 469  * netisr thread for polling and other purposes.
 470  *
 471  * This function may be called from a hard interrupt or IPI and must be
 472  * MP SAFE and non-blocking.  We use a fixed per-cpu message instead of
 473  * trying to allocate one.  We must get ourselves onto the target cpu
 474  * to safely check the MSGF_DONE bit on the message but since the message
 475  * will be sent to that cpu anyway this does not add any extra work beyond
 476  * what lwkt_sendmsg() would have already had to do to schedule the target
 477  * thread.
 478  */
 479 static void
 480 schednetisr_remote(void *data)
 481 {
 482     int num = (int)(intptr_t)data;
 483     struct netisr *ni = &netisrs[num];
 484     lwkt_port_t port = &netisr_cpu[0].td_msgport;
 485     struct netmsg *pmsg;
 486
 487     pmsg = &netisrs[num].ni_netmsg;
 488     crit_enter();
 489     if (pmsg->nm_lmsg.ms_flags & MSGF_DONE) {
 490         netmsg_init(pmsg, &netisr_adone_rport, NETISR_TO_MSGF(ni),
 491                     ni->ni_handler);
 492         pmsg->nm_lmsg.u.ms_result = num;
 493         lwkt_sendmsg(port, &pmsg->nm_lmsg);
 494     }
 495     crit_exit();
 496 }
 497
 498 void
 499 schednetisr(int num)
 500 {
 501     KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
 502         ("schednetisr: bad isr %d", num));
 503 #ifdef SMP
 504     if (mycpu->gd_cpuid != 0)
 505         lwkt_send_ipiq(globaldata_find(0), schednetisr_remote, (void *)num);
 506     else
 507         schednetisr_remote((void *)(intptr_t)num);
 508 #else
 509     schednetisr_remote((void *)(intptr_t)num);
 510 #endif
 511 }
 512
 513 lwkt_port_t
 514 netisr_find_port(int num, struct mbuf **m0)
 515 {
 516     struct netisr *ni;
 517     lwkt_port_t port;
 518     struct mbuf *m = *m0;
 519
 520     *m0 = NULL;
 521
 522     KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
 523             ("%s: bad isr %d", __func__, num));
 524
 525     ni = &netisrs[num];
 526     if (ni->ni_mport == NULL) {
 527         kprintf("%s: unregistered isr %d\n", __func__, num);
 528         m_freem(m);
 529         return NULL;
 530     }
 531
 532     if ((port = ni->ni_mport(&m)) == NULL)
 533         return NULL;
 534
 535     *m0 = m;
 536     return port;
 537 }
 538
 539 void
 540 netisr_run(int num, struct mbuf *m)
 541 {
 542     struct netisr *ni;
 543     struct netmsg_packet *pmsg;
 544
 545     KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
 546             ("%s: bad isr %d", __func__, num));
 547
 548     ni = &netisrs[num];
 549     if (ni->ni_handler == NULL) {
 550         kprintf("%s: unregistered isr %d\n", __func__, num);
 551         m_freem(m);
 552         return;
 553     }
 554
 555     pmsg = &m->m_hdr.mh_netmsg;
 556
 557     netmsg_init(&pmsg->nm_netmsg, &netisr_apanic_rport, 0, ni->ni_handler);
 558     pmsg->nm_packet = m;
 559     pmsg->nm_netmsg.nm_lmsg.u.ms_result = num;
 560
 561     NETISR_GET_MPLOCK(ni);
 562     ni->ni_handler(&pmsg->nm_netmsg);
 563     NETISR_REL_MPLOCK(ni);
 564 }
 565
 566 lwkt_port_t
 567 pktinfo_portfn_cpu0(const struct pktinfo *dummy __unused,
 568                     struct mbuf *m)
 569 {
 570     m->m_pkthdr.hash = 0;
 571     return &netisr_cpu[0].td_msgport;
 572 }
 573
 574 lwkt_port_t
 575 pktinfo_portfn_notsupp(const struct pktinfo *dummy __unused,
 576                        struct mbuf *m __unused)
 577 {
 578     return NULL;
 579 }
 580
 581 lwkt_port_t
 582 netisr_find_pktinfo_port(const struct pktinfo *pi, struct mbuf *m)
 583 {
 584     struct netisr *ni;
 585     int num = pi->pi_netisr;
 586
 587     KASSERT(m->m_flags & M_HASH, ("packet does not contain hash\n"));
 588     KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
 589             ("%s: bad isr %d", __func__, num));
 590
 591     ni = &netisrs[num];
 592     if (ni->ni_mport_pktinfo == NULL) {
 593         kprintf("%s: unregistered isr %d\n", __func__, num);
 594         return NULL;
 595     }
 596     return ni->ni_mport_pktinfo(pi, m);
 597 }