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 1.30 2007/03/04 18:51:59 swildner 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
  58 static int netmsg_sync_func(struct netmsg *msg);
  59
  60 struct netmsg_port_registration {
  61     TAILQ_ENTRY(netmsg_port_registration) npr_entry;
  62     lwkt_port_t npr_port;
  63 };
  64
  65 static struct netisr netisrs[NETISR_MAX];
  66 static TAILQ_HEAD(,netmsg_port_registration) netreglist;
  67
  68 /* Per-CPU thread to handle any protocol.  */
  69 struct thread netisr_cpu[MAXCPU];
  70 lwkt_port netisr_afree_rport;
  71 lwkt_port netisr_adone_rport;
  72 lwkt_port netisr_apanic_rport;
  73 lwkt_port netisr_sync_port;
  74
  75 /*
  76  * netisr_afree_rport replymsg function, only used to handle async
  77  * messages which the sender has abandoned to their fate.
  78  */
  79 static void
  80 netisr_autofree_reply(lwkt_port_t port, lwkt_msg_t msg)
  81 {
  82     kfree(msg, M_LWKTMSG);
  83 }
  84
  85 static void
  86 netisr_autopanic_reply(lwkt_port_t port, lwkt_msg_t msg)
  87 {
  88     panic("unreplyable msg %p was replied!", msg);
  89 }
  90
  91 /*
  92  * We must construct a custom putport function (which runs in the context
  93  * of the message originator)
  94  *
  95  * Our custom putport must check for self-referential messages, which can
  96  * occur when the so_upcall routine is called (e.g. nfs).  Self referential
  97  * messages are executed synchronously.  However, we must panic if the message
  98  * is not marked DONE on completion because the self-referential case cannot
  99  * block without deadlocking.
 100  *
 101  * note: ms_target_port does not need to be set when returning a synchronous
 102  * error code.
 103  */
 104 static int
 105 netmsg_put_port(lwkt_port_t port, lwkt_msg_t lmsg)
 106 {
 107     int error;
 108
 109     if ((lmsg->ms_flags & MSGF_ASYNC) == 0 && port->mp_td == curthread) {
 110         error = lmsg->ms_cmd.cm_func(lmsg);
 111         if (error == EASYNC && (lmsg->ms_flags & MSGF_DONE) == 0)
 112             panic("netmsg_put_port: self-referential deadlock on netport");
 113         return(error);
 114     } else {
 115         return(lwkt_default_putport(port, lmsg));
 116     }
 117 }
 118
 119 /*
 120  * UNIX DOMAIN sockets still have to run their uipc functions synchronously,
 121  * because they depend on the user proc context for a number of things
 122  * (like creds) which we have not yet incorporated into the message structure.
 123  *
 124  * However, we maintain or message/port abstraction.  Having a special
 125  * synchronous port which runs the commands synchronously gives us the
 126  * ability to serialize operations in one place later on when we start
 127  * removing the BGL.
 128  *
 129  * We clear MSGF_DONE prior to executing the message in order to close
 130  * any potential replymsg races with the flags field.  If a synchronous
 131  * result code is returned we set MSGF_DONE again.  MSGF_DONE's flag state
 132  * must be correct or the caller will be confused.
 133  */
 134 static int
 135 netmsg_sync_putport(lwkt_port_t port, lwkt_msg_t lmsg)
 136 {
 137     int error;
 138
 139     lmsg->ms_flags &= ~MSGF_DONE;
 140     lmsg->ms_target_port = port;        /* required for abort */
 141     error = lmsg->ms_cmd.cm_func(lmsg);
 142     if (error == EASYNC)
 143         error = lwkt_waitmsg(lmsg);
 144     else
 145         lmsg->ms_flags |= MSGF_DONE;
 146     return(error);
 147 }
 148
 149 static void
 150 netmsg_sync_abortport(lwkt_port_t port, lwkt_msg_t lmsg)
 151 {
 152     lmsg->ms_abort_port = lmsg->ms_reply_port;
 153     lmsg->ms_flags |= MSGF_ABORTED;
 154     lmsg->ms_abort.cm_func(lmsg);
 155 }
 156
 157 static void
 158 netisr_init(void)
 159 {
 160     int i;
 161
 162     TAILQ_INIT(&netreglist);
 163
 164     /*
 165      * Create default per-cpu threads for generic protocol handling.
 166      */
 167     for (i = 0; i < ncpus; ++i) {
 168         lwkt_create(netmsg_service_loop, NULL, NULL, &netisr_cpu[i], 0, i,
 169                     "netisr_cpu %d", i);
 170         netmsg_service_port_init(&netisr_cpu[i].td_msgport);
 171     }
 172
 173     /*
 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.
 178      */
 179     lwkt_initport(&netisr_afree_rport, NULL);
 180     netisr_afree_rport.mp_replyport = netisr_autofree_reply;
 181     lwkt_initport_null_rport(&netisr_adone_rport, NULL);
 182     lwkt_initport(&netisr_apanic_rport, NULL);
 183     netisr_apanic_rport.mp_replyport = netisr_autopanic_reply;
 184
 185     /*
 186      * The netisr_syncport is a special port which executes the message
 187      * synchronously and waits for it if EASYNC is returned.
 188      */
 189     lwkt_initport(&netisr_sync_port, NULL);
 190     netisr_sync_port.mp_putport = netmsg_sync_putport;
 191     netisr_sync_port.mp_abortport = netmsg_sync_abortport;
 192 }
 193
 194 SYSINIT(netisr, SI_SUB_PROTO_BEGIN, SI_ORDER_FIRST, netisr_init, NULL);
 195
 196 /*
 197  * Finish initializing the message port for a netmsg service.  This also
 198  * registers the port for synchronous cleanup operations such as when an
 199  * ifnet is being destroyed.  There is no deregistration API yet.
 200  */
 201 void
 202 netmsg_service_port_init(lwkt_port_t port)
 203 {
 204     struct netmsg_port_registration *reg;
 205
 206     /*
 207      * Override the putport function.  Our custom function checks for
 208      * self-references and executes such commands synchronously.
 209      */
 210     port->mp_putport = netmsg_put_port;
 211
 212     /*
 213      * Keep track of ports using the netmsg API so we can synchronize
 214      * certain operations (such as freeing an ifnet structure) across all
 215      * consumers.
 216      */
 217     reg = kmalloc(sizeof(*reg), M_TEMP, M_WAITOK|M_ZERO);
 218     reg->npr_port = port;
 219     TAILQ_INSERT_TAIL(&netreglist, reg, npr_entry);
 220 }
 221
 222 /*
 223  * This function synchronizes the caller with all netmsg services.  For
 224  * example, if an interface is being removed we must make sure that all
 225  * packets related to that interface complete processing before the structure
 226  * can actually be freed.  This sort of synchronization is an alternative to
 227  * ref-counting the netif, removing the ref counting overhead in favor of
 228  * placing additional overhead in the netif freeing sequence (where it is
 229  * inconsequential).
 230  */
 231 void
 232 netmsg_service_sync(void)
 233 {
 234     struct netmsg_port_registration *reg;
 235     struct netmsg smsg;
 236
 237     lwkt_initmsg(&smsg.nm_lmsg, &curthread->td_msgport, 0,
 238                 lwkt_cmd_func((void *)netmsg_sync_func), lwkt_cmd_op_none);
 239
 240     TAILQ_FOREACH(reg, &netreglist, npr_entry) {
 241         lwkt_domsg(reg->npr_port, &smsg.nm_lmsg);
 242     }
 243 }
 244
 245 /*
 246  * The netmsg function simply replies the message.  API semantics require
 247  * EASYNC to be returned if the netmsg function disposes of the message.
 248  */
 249 static int
 250 netmsg_sync_func(struct netmsg *msg)
 251 {
 252     lwkt_replymsg(&msg->nm_lmsg, 0);
 253     return(EASYNC);
 254 }
 255
 256 /*
 257  * Generic netmsg service loop.  Some protocols may roll their own but all
 258  * must do the basic command dispatch function call done here.
 259  */
 260 void
 261 netmsg_service_loop(void *arg)
 262 {
 263     struct netmsg *msg;
 264
 265     while ((msg = lwkt_waitport(&curthread->td_msgport, NULL))) {
 266         msg->nm_lmsg.ms_cmd.cm_func(&msg->nm_lmsg);
 267     }
 268 }
 269
 270 /*
 271  * Call the netisr directly.
 272  * Queueing may be done in the msg port layer at its discretion.
 273  */
 274 void
 275 netisr_dispatch(int num, struct mbuf *m)
 276 {
 277     /* just queue it for now XXX JH */
 278     netisr_queue(num, m);
 279 }
 280
 281 /*
 282  * Same as netisr_dispatch(), but always queue.
 283  * This is either used in places where we are not confident that
 284  * direct dispatch is possible, or where queueing is required.
 285  */
 286 int
 287 netisr_queue(int num, struct mbuf *m)
 288 {
 289     struct netisr *ni;
 290     struct netmsg_packet *pmsg;
 291     lwkt_port_t port;
 292
 293     KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
 294         ("netisr_queue: bad isr %d", num));
 295
 296     ni = &netisrs[num];
 297     if (ni->ni_handler == NULL) {
 298         kprintf("netisr_queue: unregistered isr %d\n", num);
 299         return (EIO);
 300     }
 301
 302     if ((port = ni->ni_mport(&m)) == NULL)
 303         return (EIO);
 304
 305     pmsg = &m->m_hdr.mh_netmsg;
 306
 307     lwkt_initmsg(&pmsg->nm_lmsg, &netisr_apanic_rport, 0,
 308                 lwkt_cmd_func((void *)ni->ni_handler), lwkt_cmd_op_none);
 309     pmsg->nm_packet = m;
 310     pmsg->nm_lmsg.u.ms_result = num;
 311     lwkt_sendmsg(port, &pmsg->nm_lmsg);
 312     return (0);
 313 }
 314
 315 void
 316 netisr_register(int num, lwkt_portfn_t mportfn, netisr_fn_t handler)
 317 {
 318     KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
 319         ("netisr_register: bad isr %d", num));
 320     lwkt_initmsg(&netisrs[num].ni_netmsg.nm_lmsg, &netisr_adone_rport, 0,
 321             lwkt_cmd_op_none, lwkt_cmd_op_none);
 322     netisrs[num].ni_mport = mportfn;
 323     netisrs[num].ni_handler = handler;
 324 }
 325
 326 int
 327 netisr_unregister(int num)
 328 {
 329     KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
 330         ("unregister_netisr: bad isr number: %d\n", num));
 331
 332     /* XXX JH */
 333     return (0);
 334 }
 335
 336 /*
 337  * Return message port for default handler thread on CPU 0.
 338  */
 339 lwkt_port_t
 340 cpu0_portfn(struct mbuf **mptr)
 341 {
 342     return (&netisr_cpu[0].td_msgport);
 343 }
 344
 345 lwkt_port_t
 346 cpu_portfn(int cpu)
 347 {
 348     return (&netisr_cpu[cpu].td_msgport);
 349 }
 350
 351 /* ARGSUSED */
 352 lwkt_port_t
 353 cpu0_soport(struct socket *so __unused, struct sockaddr *nam __unused,
 354             int req __unused)
 355 {
 356     return (&netisr_cpu[0].td_msgport);
 357 }
 358
 359 lwkt_port_t
 360 sync_soport(struct socket *so __unused, struct sockaddr *nam __unused,
 361             int req __unused)
 362 {
 363     return (&netisr_sync_port);
 364 }
 365
 366 /*
 367  * schednetisr() is used to call the netisr handler from the appropriate
 368  * netisr thread for polling and other purposes.
 369  *
 370  * This function may be called from a hard interrupt or IPI and must be
 371  * MP SAFE and non-blocking.  We use a fixed per-cpu message instead of
 372  * trying to allocate one.  We must get ourselves onto the target cpu
 373  * to safely check the MSGF_DONE bit on the message but since the message
 374  * will be sent to that cpu anyway this does not add any extra work beyond
 375  * what lwkt_sendmsg() would have already had to do to schedule the target
 376  * thread.
 377  */
 378 static void
 379 schednetisr_remote(void *data)
 380 {
 381     int num = (int)data;
 382     struct netisr *ni = &netisrs[num];
 383     lwkt_port_t port = &netisr_cpu[0].td_msgport;
 384     struct netmsg *pmsg;
 385
 386     pmsg = &netisrs[num].ni_netmsg;
 387     crit_enter();
 388     if (pmsg->nm_lmsg.ms_flags & MSGF_DONE) {
 389         lwkt_initmsg(&pmsg->nm_lmsg, &netisr_adone_rport, 0,
 390                     lwkt_cmd_func((void *)ni->ni_handler), lwkt_cmd_op_none);
 391         pmsg->nm_lmsg.u.ms_result = num;
 392         lwkt_sendmsg(port, &pmsg->nm_lmsg);
 393     }
 394     crit_exit();
 395 }
 396
 397 void
 398 schednetisr(int num)
 399 {
 400     KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
 401         ("schednetisr: bad isr %d", num));
 402 #ifdef SMP
 403     if (mycpu->gd_cpuid != 0)
 404         lwkt_send_ipiq(globaldata_find(0), schednetisr_remote, (void *)num);
 405     else
 406         schednetisr_remote((void *)num);
 407 #else
 408     schednetisr_remote((void *)num);
 409 #endif
 410 }
 411