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 1.33 2007/05/24 20:51:21 dillon 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     netisr_sync_port.mp_putport = netmsg_sync_putport;
 174 }
 175
 176 SYSINIT(netisr, SI_SUB_PROTO_BEGIN, SI_ORDER_FIRST, netisr_init, NULL);
 177
 178 /*
 179  * Finish initializing the message port for a netmsg service.  This also
 180  * registers the port for synchronous cleanup operations such as when an
 181  * ifnet is being destroyed.  There is no deregistration API yet.
 182  */
 183 void
 184 netmsg_service_port_init(lwkt_port_t port)
 185 {
 186     struct netmsg_port_registration *reg;
 187
 188     /*
 189      * Override the putport function.  Our custom function checks for
 190      * self-references and executes such commands synchronously.
 191      */
 192     if (netmsg_fwd_port_fn == NULL)
 193         netmsg_fwd_port_fn = port->mp_putport;
 194     KKASSERT(netmsg_fwd_port_fn == port->mp_putport);
 195     port->mp_putport = netmsg_put_port;
 196
 197     /*
 198      * Keep track of ports using the netmsg API so we can synchronize
 199      * certain operations (such as freeing an ifnet structure) across all
 200      * consumers.
 201      */
 202     reg = kmalloc(sizeof(*reg), M_TEMP, M_WAITOK|M_ZERO);
 203     reg->npr_port = port;
 204     TAILQ_INSERT_TAIL(&netreglist, reg, npr_entry);
 205 }
 206
 207 /*
 208  * This function synchronizes the caller with all netmsg services.  For
 209  * example, if an interface is being removed we must make sure that all
 210  * packets related to that interface complete processing before the structure
 211  * can actually be freed.  This sort of synchronization is an alternative to
 212  * ref-counting the netif, removing the ref counting overhead in favor of
 213  * placing additional overhead in the netif freeing sequence (where it is
 214  * inconsequential).
 215  */
 216 void
 217 netmsg_service_sync(void)
 218 {
 219     struct netmsg_port_registration *reg;
 220     struct netmsg smsg;
 221
 222     netmsg_init(&smsg, &curthread->td_msgport, 0, netmsg_sync_func);
 223
 224     TAILQ_FOREACH(reg, &netreglist, npr_entry) {
 225         lwkt_domsg(reg->npr_port, &smsg.nm_lmsg, 0);
 226     }
 227 }
 228
 229 /*
 230  * The netmsg function simply replies the message.  API semantics require
 231  * EASYNC to be returned if the netmsg function disposes of the message.
 232  */
 233 static void
 234 netmsg_sync_func(struct netmsg *msg)
 235 {
 236     lwkt_replymsg(&msg->nm_lmsg, 0);
 237 }
 238
 239 /*
 240  * Generic netmsg service loop.  Some protocols may roll their own but all
 241  * must do the basic command dispatch function call done here.
 242  */
 243 void
 244 netmsg_service_loop(void *arg)
 245 {
 246     struct netmsg *msg;
 247
 248     while ((msg = lwkt_waitport(&curthread->td_msgport, NULL))) {
 249         msg->nm_dispatch(msg);
 250     }
 251 }
 252
 253 /*
 254  * Call the netisr directly.
 255  * Queueing may be done in the msg port layer at its discretion.
 256  */
 257 void
 258 netisr_dispatch(int num, struct mbuf *m)
 259 {
 260     /* just queue it for now XXX JH */
 261     netisr_queue(num, m);
 262 }
 263
 264 /*
 265  * Same as netisr_dispatch(), but always queue.
 266  * This is either used in places where we are not confident that
 267  * direct dispatch is possible, or where queueing is required.
 268  */
 269 int
 270 netisr_queue(int num, struct mbuf *m)
 271 {
 272     struct netisr *ni;
 273     struct netmsg_packet *pmsg;
 274     lwkt_port_t port;
 275
 276     KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
 277         ("netisr_queue: bad isr %d", num));
 278
 279     ni = &netisrs[num];
 280     if (ni->ni_handler == NULL) {
 281         kprintf("netisr_queue: unregistered isr %d\n", num);
 282         return (EIO);
 283     }
 284
 285     if ((port = ni->ni_mport(&m)) == NULL)
 286         return (EIO);
 287
 288     pmsg = &m->m_hdr.mh_netmsg;
 289
 290     netmsg_init(&pmsg->nm_netmsg, &netisr_apanic_rport, 0, ni->ni_handler);
 291     pmsg->nm_packet = m;
 292     pmsg->nm_netmsg.nm_lmsg.u.ms_result = num;
 293     lwkt_sendmsg(port, &pmsg->nm_netmsg.nm_lmsg);
 294     return (0);
 295 }
 296
 297 void
 298 netisr_register(int num, lwkt_portfn_t mportfn, netisr_fn_t handler)
 299 {
 300     KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
 301         ("netisr_register: bad isr %d", num));
 302     netmsg_init(&netisrs[num].ni_netmsg, &netisr_adone_rport, 0, NULL);
 303     netisrs[num].ni_mport = mportfn;
 304     netisrs[num].ni_handler = handler;
 305 }
 306
 307 int
 308 netisr_unregister(int num)
 309 {
 310     KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
 311         ("unregister_netisr: bad isr number: %d\n", num));
 312
 313     /* XXX JH */
 314     return (0);
 315 }
 316
 317 /*
 318  * Return message port for default handler thread on CPU 0.
 319  */
 320 lwkt_port_t
 321 cpu0_portfn(struct mbuf **mptr)
 322 {
 323     return (&netisr_cpu[0].td_msgport);
 324 }
 325
 326 lwkt_port_t
 327 cpu_portfn(int cpu)
 328 {
 329     return (&netisr_cpu[cpu].td_msgport);
 330 }
 331
 332 /* ARGSUSED */
 333 lwkt_port_t
 334 cpu0_soport(struct socket *so __unused, struct sockaddr *nam __unused,
 335             int req __unused)
 336 {
 337     return (&netisr_cpu[0].td_msgport);
 338 }
 339
 340 lwkt_port_t
 341 sync_soport(struct socket *so __unused, struct sockaddr *nam __unused,
 342             int req __unused)
 343 {
 344     return (&netisr_sync_port);
 345 }
 346
 347 /*
 348  * schednetisr() is used to call the netisr handler from the appropriate
 349  * netisr thread for polling and other purposes.
 350  *
 351  * This function may be called from a hard interrupt or IPI and must be
 352  * MP SAFE and non-blocking.  We use a fixed per-cpu message instead of
 353  * trying to allocate one.  We must get ourselves onto the target cpu
 354  * to safely check the MSGF_DONE bit on the message but since the message
 355  * will be sent to that cpu anyway this does not add any extra work beyond
 356  * what lwkt_sendmsg() would have already had to do to schedule the target
 357  * thread.
 358  */
 359 static void
 360 schednetisr_remote(void *data)
 361 {
 362     int num = (int)data;
 363     struct netisr *ni = &netisrs[num];
 364     lwkt_port_t port = &netisr_cpu[0].td_msgport;
 365     struct netmsg *pmsg;
 366
 367     pmsg = &netisrs[num].ni_netmsg;
 368     crit_enter();
 369     if (pmsg->nm_lmsg.ms_flags & MSGF_DONE) {
 370         netmsg_init(pmsg, &netisr_adone_rport, 0, ni->ni_handler);
 371         pmsg->nm_lmsg.u.ms_result = num;
 372         lwkt_sendmsg(port, &pmsg->nm_lmsg);
 373     }
 374     crit_exit();
 375 }
 376
 377 void
 378 schednetisr(int num)
 379 {
 380     KASSERT((num > 0 && num <= (sizeof(netisrs)/sizeof(netisrs[0]))),
 381         ("schednetisr: bad isr %d", num));
 382 #ifdef SMP
 383     if (mycpu->gd_cpuid != 0)
 384         lwkt_send_ipiq(globaldata_find(0), schednetisr_remote, (void *)num);
 385     else
 386         schednetisr_remote((void *)num);
 387 #else
 388     schednetisr_remote((void *)num);
 389 #endif
 390 }
 391