sys/kern/kern_poll.c

   1 /*-
   2  * Copyright (c) 2001-2002 Luigi Rizzo
   3  *
   4  * Supported by: the Xorp Project (www.xorp.org)
   5  *
   6  * Redistribution and use in source and binary forms, with or without
   7  * modification, are permitted provided that the following conditions
   8  * are met:
   9  * 1. Redistributions of source code must retain the above copyright
  10  *    notice, this list of conditions and the following disclaimer.
  11  * 2. Redistributions in binary form must reproduce the above copyright
  12  *    notice, this list of conditions and the following disclaimer in the
  13  *    documentation and/or other materials provided with the distribution.
  14  *
  15  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND
  16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE
  19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  25  * SUCH DAMAGE.
  26  *
  27  * $FreeBSD: src/sys/kern/kern_poll.c,v 1.2.2.4 2002/06/27 23:26:33 luigi Exp $
  28  * $DragonFly: src/sys/kern/kern_poll.c,v 1.33 2007/09/12 12:02:09 sephe Exp $
  29  */
  30
  31 #include "opt_polling.h"
  32
  33 #include <sys/param.h>
  34 #include <sys/systm.h>
  35 #include <sys/kernel.h>
  36 #include <sys/socket.h>                 /* needed by net/if.h           */
  37 #include <sys/sysctl.h>
  38
  39 #include <sys/thread2.h>
  40 #include <sys/msgport2.h>
  41
  42 #include <net/if.h>                     /* for IFF_* flags              */
  43 #include <net/netisr.h>                 /* for NETISR_POLL              */
  44
  45 /*
  46  * Polling support for [network] device drivers.
  47  *
  48  * Drivers which support this feature try to register with the
  49  * polling code.
  50  *
  51  * If registration is successful, the driver must disable interrupts,
  52  * and further I/O is performed through the handler, which is invoked
  53  * (at least once per clock tick) with 3 arguments: the "arg" passed at
  54  * register time (a struct ifnet pointer), a command, and a "count" limit.
  55  *
  56  * The command can be one of the following:
  57  *  POLL_ONLY: quick move of "count" packets from input/output queues.
  58  *  POLL_AND_CHECK_STATUS: as above, plus check status registers or do
  59  *      other more expensive operations. This command is issued periodically
  60  *      but less frequently than POLL_ONLY.
  61  *  POLL_DEREGISTER: deregister and return to interrupt mode.
  62  *  POLL_REGISTER: register and disable interrupts
  63  *
  64  * The first two commands are only issued if the interface is marked as
  65  * 'IFF_UP, IFF_RUNNING and IFF_POLLING', the last two only if IFF_RUNNING
  66  * is set.
  67  *
  68  * The count limit specifies how much work the handler can do during the
  69  * call -- typically this is the number of packets to be received, or
  70  * transmitted, etc. (drivers are free to interpret this number, as long
  71  * as the max time spent in the function grows roughly linearly with the
  72  * count).
  73  *
  74  * Deregistration can be requested by the driver itself (typically in the
  75  * *_stop() routine), or by the polling code, by invoking the handler.
  76  *
  77  * Polling can be globally enabled or disabled with the sysctl variable
  78  * kern.polling.enable (default is 0, disabled)
  79  *
  80  * A second variable controls the sharing of CPU between polling/kernel
  81  * network processing, and other activities (typically userlevel tasks):
  82  * kern.polling.user_frac (between 0 and 100, default 50) sets the share
  83  * of CPU allocated to user tasks. CPU is allocated proportionally to the
  84  * shares, by dynamically adjusting the "count" (poll_burst).
  85  *
  86  * Other parameters can should be left to their default values.
  87  * The following constraints hold
  88  *
  89  *      1 <= poll_each_burst <= poll_burst <= poll_burst_max
  90  *      MIN_POLL_BURST_MAX <= poll_burst_max <= MAX_POLL_BURST_MAX
  91  */
  92
  93 #define MIN_POLL_BURST_MAX      10
  94 #define MAX_POLL_BURST_MAX      1000
  95
  96 #ifndef DEVICE_POLLING_FREQ_MAX
  97 #define DEVICE_POLLING_FREQ_MAX         30000
  98 #endif
  99 #define DEVICE_POLLING_FREQ_DEFAULT     2000
 100
 101 #define POLL_LIST_LEN  128
 102 struct pollrec {
 103         struct ifnet    *ifp;
 104 };
 105
 106 #define POLLCTX_MAX     32
 107
 108 struct pollctx {
 109         struct sysctl_ctx_list  poll_sysctl_ctx;
 110         struct sysctl_oid       *poll_sysctl_tree;
 111
 112         uint32_t                poll_burst;
 113         uint32_t                poll_each_burst;
 114         uint32_t                poll_burst_max;
 115         uint32_t                user_frac;
 116         int                     reg_frac_count;
 117         uint32_t                reg_frac;
 118         uint32_t                short_ticks;
 119         uint32_t                lost_polls;
 120         uint32_t                pending_polls;
 121         int                     residual_burst;
 122         uint32_t                phase;
 123         uint32_t                suspect;
 124         uint32_t                stalled;
 125         struct timeval          poll_start_t;
 126         struct timeval          prev_t;
 127
 128         uint32_t                poll_handlers; /* next free entry in pr[]. */
 129         struct pollrec          pr[POLL_LIST_LEN];
 130
 131         int                     poll_cpuid;
 132         struct systimer         pollclock;
 133         int                     polling_enabled;
 134         int                     pollhz;
 135 };
 136
 137 static struct pollctx   *poll_context[POLLCTX_MAX];
 138
 139 SYSCTL_NODE(_kern, OID_AUTO, polling, CTLFLAG_RW, 0,
 140         "Device polling parameters");
 141
 142 static int      poll_defcpu = -1;
 143 SYSCTL_INT(_kern_polling, OID_AUTO, defcpu, CTLFLAG_RD,
 144         &poll_defcpu, 0, "default CPU# to run device polling");
 145
 146 static uint32_t poll_cpumask = 0x1;
 147 #ifdef notyet
 148 TUNABLE_INT("kern.polling.cpumask", &poll_cpumask);
 149 #endif
 150
 151 static int      polling_enabled = 0;    /* global polling enable */
 152 TUNABLE_INT("kern.polling.enable", &polling_enabled);
 153
 154 static int      pollhz = DEVICE_POLLING_FREQ_DEFAULT;
 155 TUNABLE_INT("kern.polling.pollhz", &pollhz);
 156
 157 /* The two netisr handlers */
 158 static void     netisr_poll(struct netmsg *);
 159 static void     netisr_pollmore(struct netmsg *);
 160
 161 /* Systimer handler */
 162 static void     pollclock(systimer_t, struct intrframe *);
 163
 164 /* Sysctl handlers */
 165 static int      sysctl_pollhz(SYSCTL_HANDLER_ARGS);
 166 static int      sysctl_polling(SYSCTL_HANDLER_ARGS);
 167 static void     poll_add_sysctl(struct sysctl_ctx_list *,
 168                                 struct sysctl_oid_list *, struct pollctx *);
 169
 170 void            init_device_poll(void);         /* init routine */
 171 void            init_device_poll_pcpu(int);     /* per-cpu init routine */
 172
 173 /*
 174  * register relevant netisr. Called from kern_clock.c:
 175  */
 176 void
 177 init_device_poll(void)
 178 {
 179         netisr_register(NETISR_POLL, cpu0_portfn, netisr_poll);
 180         netisr_register(NETISR_POLLMORE, cpu0_portfn, netisr_pollmore);
 181 }
 182
 183 void
 184 init_device_poll_pcpu(int cpuid)
 185 {
 186         struct pollctx *pctx;
 187         char cpuid_str[3];
 188
 189         if (((1 << cpuid) & poll_cpumask) == 0)
 190                 return;
 191
 192         pctx = kmalloc(sizeof(*pctx), M_DEVBUF, M_WAITOK | M_ZERO);
 193
 194         pctx->poll_burst = 5;
 195         pctx->poll_each_burst = 5;
 196         pctx->poll_burst_max = 150; /* good for 100Mbit net and HZ=1000 */
 197         pctx->user_frac = 50;
 198         pctx->reg_frac = 20;
 199         pctx->polling_enabled = polling_enabled;
 200         pctx->pollhz = pollhz;
 201         pctx->poll_cpuid = cpuid;
 202
 203         KASSERT(cpuid < POLLCTX_MAX, ("cpu id must < %d", cpuid));
 204         poll_context[cpuid] = pctx;
 205
 206         if (poll_defcpu < 0) {
 207                 poll_defcpu = cpuid;
 208
 209                 /*
 210                  * Initialize global sysctl nodes, for compat
 211                  */
 212                 poll_add_sysctl(NULL, SYSCTL_STATIC_CHILDREN(_kern_polling),
 213                                 pctx);
 214         }
 215
 216         /*
 217          * Initialize per-cpu sysctl nodes
 218          */
 219         ksnprintf(cpuid_str, sizeof(cpuid_str), "%d", pctx->poll_cpuid);
 220
 221         sysctl_ctx_init(&pctx->poll_sysctl_ctx);
 222         pctx->poll_sysctl_tree = SYSCTL_ADD_NODE(&pctx->poll_sysctl_ctx,
 223                                  SYSCTL_STATIC_CHILDREN(_kern_polling),
 224                                  OID_AUTO, cpuid_str, CTLFLAG_RD, 0, "");
 225         poll_add_sysctl(&pctx->poll_sysctl_ctx,
 226                         SYSCTL_CHILDREN(pctx->poll_sysctl_tree), pctx);
 227
 228         /*
 229          * Initialize systimer
 230          */
 231         systimer_init_periodic_nq(&pctx->pollclock, pollclock, pctx,
 232                                   pctx->polling_enabled ? pctx->pollhz : 1);
 233 }
 234
 235 /*
 236  * Set the polling frequency
 237  */
 238 static int
 239 sysctl_pollhz(SYSCTL_HANDLER_ARGS)
 240 {
 241         struct pollctx *pctx = arg1;
 242         int error, phz;
 243
 244         phz = pctx->pollhz;
 245         error = sysctl_handle_int(oidp, &phz, 0, req);
 246         if (error || req->newptr == NULL)
 247                 return error;
 248         if (phz <= 0)
 249                 return EINVAL;
 250         else if (phz > DEVICE_POLLING_FREQ_MAX)
 251                 phz = DEVICE_POLLING_FREQ_MAX;
 252
 253         crit_enter();
 254         pctx->pollhz = phz;
 255         if (pctx->polling_enabled)
 256                 systimer_adjust_periodic(&pctx->pollclock, phz);
 257         crit_exit();
 258         return 0;
 259 }
 260
 261 /*
 262  * Master enable.  If polling is disabled, cut the polling systimer
 263  * frequency to 1hz.
 264  */
 265 static int
 266 sysctl_polling(SYSCTL_HANDLER_ARGS)
 267 {
 268         struct pollctx *pctx = arg1;
 269         int error, enabled;
 270
 271         enabled = pctx->polling_enabled;
 272         error = sysctl_handle_int(oidp, &enabled, 0, req);
 273         if (error || req->newptr == NULL)
 274                 return error;
 275
 276         crit_enter();
 277         pctx->polling_enabled = enabled;
 278         if (pctx->polling_enabled)
 279                 systimer_adjust_periodic(&pctx->pollclock, pollhz);
 280         else
 281                 systimer_adjust_periodic(&pctx->pollclock, 1);
 282         crit_exit();
 283         return 0;
 284 }
 285
 286 /*
 287  * Hook from hardclock. Tries to schedule a netisr, but keeps track
 288  * of lost ticks due to the previous handler taking too long.
 289  * Normally, this should not happen, because polling handler should
 290  * run for a short time. However, in some cases (e.g. when there are
 291  * changes in link status etc.) the drivers take a very long time
 292  * (even in the order of milliseconds) to reset and reconfigure the
 293  * device, causing apparent lost polls.
 294  *
 295  * The first part of the code is just for debugging purposes, and tries
 296  * to count how often hardclock ticks are shorter than they should,
 297  * meaning either stray interrupts or delayed events.
 298  *
 299  * WARNING! called from fastint or IPI, the MP lock might not be held.
 300  */
 301 static void
 302 pollclock(systimer_t info, struct intrframe *frame __unused)
 303 {
 304         struct pollctx *pctx = info->data;
 305         struct timeval t;
 306         int delta;
 307
 308         if (pctx->poll_handlers == 0)
 309                 return;
 310
 311         microuptime(&t);
 312         delta = (t.tv_usec - pctx->prev_t.tv_usec) +
 313                 (t.tv_sec - pctx->prev_t.tv_sec)*1000000;
 314         if (delta * hz < 500000)
 315                 pctx->short_ticks++;
 316         else
 317                 pctx->prev_t = t;
 318
 319         if (pctx->pending_polls > 100) {
 320                 /*
 321                  * Too much, assume it has stalled (not always true
 322                  * see comment above).
 323                  */
 324                 pctx->stalled++;
 325                 pctx->pending_polls = 0;
 326                 pctx->phase = 0;
 327         }
 328
 329         if (pctx->phase <= 2) {
 330                 if (pctx->phase != 0)
 331                         pctx->suspect++;
 332                 pctx->phase = 1;
 333                 schednetisr(NETISR_POLL);
 334                 pctx->phase = 2;
 335         }
 336         if (pctx->pending_polls++ > 0)
 337                 pctx->lost_polls++;
 338 }
 339
 340 /*
 341  * netisr_pollmore is called after other netisr's, possibly scheduling
 342  * another NETISR_POLL call, or adapting the burst size for the next cycle.
 343  *
 344  * It is very bad to fetch large bursts of packets from a single card at once,
 345  * because the burst could take a long time to be completely processed, or
 346  * could saturate the intermediate queue (ipintrq or similar) leading to
 347  * losses or unfairness. To reduce the problem, and also to account better for
 348  * time spent in network-related processing, we split the burst in smaller
 349  * chunks of fixed size, giving control to the other netisr's between chunks.
 350  * This helps in improving the fairness, reducing livelock (because we
 351  * emulate more closely the "process to completion" that we have with
 352  * fastforwarding) and accounting for the work performed in low level
 353  * handling and forwarding.
 354  */
 355
 356 /* ARGSUSED */
 357 static void
 358 netisr_pollmore(struct netmsg *msg)
 359 {
 360         struct pollctx *pctx;
 361         struct timeval t;
 362         int kern_load, cpuid;
 363
 364         cpuid = mycpu->gd_cpuid;
 365         KKASSERT(cpuid < POLLCTX_MAX);
 366
 367         pctx = poll_context[cpuid];
 368         KKASSERT(pctx != NULL);
 369         KKASSERT(pctx->poll_cpuid == cpuid);
 370
 371         crit_enter();
 372         lwkt_replymsg(&msg->nm_lmsg, 0);
 373         pctx->phase = 5;
 374         if (pctx->residual_burst > 0) {
 375                 schednetisr(NETISR_POLL);
 376                 /* will run immediately on return, followed by netisrs */
 377                 goto out;
 378         }
 379         /* here we can account time spent in netisr's in this tick */
 380         microuptime(&t);
 381         kern_load = (t.tv_usec - pctx->poll_start_t.tv_usec) +
 382                 (t.tv_sec - pctx->poll_start_t.tv_sec)*1000000; /* us */
 383         kern_load = (kern_load * hz) / 10000;                   /* 0..100 */
 384         if (kern_load > (100 - pctx->user_frac)) { /* try decrease ticks */
 385                 if (pctx->poll_burst > 1)
 386                         pctx->poll_burst--;
 387         } else {
 388                 if (pctx->poll_burst < pctx->poll_burst_max)
 389                         pctx->poll_burst++;
 390         }
 391
 392         pctx->pending_polls--;
 393         if (pctx->pending_polls == 0) { /* we are done */
 394                 pctx->phase = 0;
 395         } else {
 396                 /*
 397                  * Last cycle was long and caused us to miss one or more
 398                  * hardclock ticks. Restart processing again, but slightly
 399                  * reduce the burst size to prevent that this happens again.
 400                  */
 401                 pctx->poll_burst -= (pctx->poll_burst / 8);
 402                 if (pctx->poll_burst < 1)
 403                         pctx->poll_burst = 1;
 404                 schednetisr(NETISR_POLL);
 405                 pctx->phase = 6;
 406         }
 407 out:
 408         crit_exit();
 409 }
 410
 411 /*
 412  * netisr_poll is scheduled by schednetisr when appropriate, typically once
 413  * per tick.
 414  *
 415  * Note that the message is replied immediately in order to allow a new
 416  * ISR to be scheduled in the handler.
 417  *
 418  * XXX each registration should indicate whether it needs a critical
 419  * section to operate.
 420  */
 421 /* ARGSUSED */
 422 static void
 423 netisr_poll(struct netmsg *msg)
 424 {
 425         struct pollctx *pctx;
 426         int i, cycles, cpuid;
 427         enum poll_cmd arg = POLL_ONLY;
 428
 429         cpuid = mycpu->gd_cpuid;
 430         KKASSERT(cpuid < POLLCTX_MAX);
 431
 432         pctx = poll_context[cpuid];
 433         KKASSERT(pctx != NULL);
 434         KKASSERT(pctx->poll_cpuid == cpuid);
 435
 436         lwkt_replymsg(&msg->nm_lmsg, 0);
 437         crit_enter();
 438         pctx->phase = 3;
 439         if (pctx->residual_burst == 0) { /* first call in this tick */
 440                 microuptime(&pctx->poll_start_t);
 441                 /*
 442                  * Check that paremeters are consistent with runtime
 443                  * variables. Some of these tests could be done at sysctl
 444                  * time, but the savings would be very limited because we
 445                  * still have to check against reg_frac_count and
 446                  * poll_each_burst. So, instead of writing separate sysctl
 447                  * handlers, we do all here.
 448                  */
 449
 450                 if (pctx->reg_frac > hz)
 451                         pctx->reg_frac = hz;
 452                 else if (pctx->reg_frac < 1)
 453                         pctx->reg_frac = 1;
 454                 if (pctx->reg_frac_count > pctx->reg_frac)
 455                         pctx->reg_frac_count = pctx->reg_frac - 1;
 456                 if (pctx->reg_frac_count-- == 0) {
 457                         arg = POLL_AND_CHECK_STATUS;
 458                         pctx->reg_frac_count = pctx->reg_frac - 1;
 459                 }
 460                 if (pctx->poll_burst_max < MIN_POLL_BURST_MAX)
 461                         pctx->poll_burst_max = MIN_POLL_BURST_MAX;
 462                 else if (pctx->poll_burst_max > MAX_POLL_BURST_MAX)
 463                         pctx->poll_burst_max = MAX_POLL_BURST_MAX;
 464
 465                 if (pctx->poll_each_burst < 1)
 466                         pctx->poll_each_burst = 1;
 467                 else if (pctx->poll_each_burst > pctx->poll_burst_max)
 468                         pctx->poll_each_burst = pctx->poll_burst_max;
 469
 470                 pctx->residual_burst = pctx->poll_burst;
 471         }
 472         cycles = (pctx->residual_burst < pctx->poll_each_burst) ?
 473                 pctx->residual_burst : pctx->poll_each_burst;
 474         pctx->residual_burst -= cycles;
 475
 476         if (pctx->polling_enabled) {
 477                 for (i = 0 ; i < pctx->poll_handlers ; i++) {
 478                         struct ifnet *ifp = pctx->pr[i].ifp;
 479
 480                         if ((ifp->if_flags & (IFF_UP|IFF_RUNNING|IFF_POLLING))
 481                             == (IFF_UP|IFF_RUNNING|IFF_POLLING)) {
 482                                 if (lwkt_serialize_try(ifp->if_serializer)) {
 483                                         ifp->if_poll(ifp, arg, cycles);
 484                                         lwkt_serialize_exit(ifp->if_serializer);
 485                                 }
 486                         }
 487                 }
 488         } else {        /* unregister */
 489                 for (i = 0 ; i < pctx->poll_handlers ; i++) {
 490                         struct ifnet *ifp = pctx->pr[i].ifp;
 491
 492                         if ((ifp->if_flags & IFF_POLLING) == 0)
 493                                 continue;
 494                         /*
 495                          * Only call the interface deregistration
 496                          * function if the interface is still
 497                          * running.
 498                          */
 499                         lwkt_serialize_enter(ifp->if_serializer);
 500                         ifp->if_flags &= ~IFF_POLLING;
 501                         if (ifp->if_flags & IFF_RUNNING)
 502                                 ifp->if_poll(ifp, POLL_DEREGISTER, 1);
 503                         lwkt_serialize_exit(ifp->if_serializer);
 504                 }
 505                 pctx->residual_burst = 0;
 506                 pctx->poll_handlers = 0;
 507         }
 508         schednetisr(NETISR_POLLMORE);
 509         pctx->phase = 4;
 510         crit_exit();
 511 }
 512
 513 /*
 514  * Try to register routine for polling. Returns 1 if successful
 515  * (and polling should be enabled), 0 otherwise.
 516  *
 517  * Called from mainline code only, not called from an interrupt.
 518  */
 519 int
 520 ether_poll_register(struct ifnet *ifp)
 521 {
 522         struct pollctx *pctx;
 523         int rc;
 524
 525         if (poll_defcpu < 0)
 526                 return 0;
 527         KKASSERT(poll_defcpu < POLLCTX_MAX);
 528
 529         pctx = poll_context[poll_defcpu];
 530         KKASSERT(pctx != NULL);
 531         KKASSERT(pctx->poll_cpuid == poll_defcpu);
 532
 533         if (pctx->polling_enabled == 0) /* polling disabled, cannot register */
 534                 return 0;
 535         if (ifp->if_flags & IFF_POLLING)        /* already polling      */
 536                 return 0;
 537         if (ifp->if_poll == NULL)               /* no polling support   */
 538                 return 0;
 539
 540         /*
 541          * Attempt to register.  Interlock with IFF_POLLING.
 542          */
 543         crit_enter();   /* XXX MP - not mp safe */
 544         lwkt_serialize_enter(ifp->if_serializer);
 545         ifp->if_flags |= IFF_POLLING;
 546         if (ifp->if_flags & IFF_RUNNING)
 547                 ifp->if_poll(ifp, POLL_REGISTER, 0);
 548         lwkt_serialize_exit(ifp->if_serializer);
 549         if ((ifp->if_flags & IFF_POLLING) == 0) {
 550                 crit_exit();
 551                 return 0;
 552         }
 553
 554         /*
 555          * Check if there is room.  If there isn't, deregister.
 556          */
 557         if (pctx->poll_handlers >= POLL_LIST_LEN) {
 558                 /*
 559                  * List full, cannot register more entries.
 560                  * This should never happen; if it does, it is probably a
 561                  * broken driver trying to register multiple times. Checking
 562                  * this at runtime is expensive, and won't solve the problem
 563                  * anyways, so just report a few times and then give up.
 564                  */
 565                 static int verbose = 10;        /* XXX */
 566                 if (verbose >0) {
 567                         kprintf("poll handlers list full, "
 568                                 "maybe a broken driver ?\n");
 569                         verbose--;
 570                 }
 571                 lwkt_serialize_enter(ifp->if_serializer);
 572                 ifp->if_flags &= ~IFF_POLLING;
 573                 if (ifp->if_flags & IFF_RUNNING)
 574                         ifp->if_poll(ifp, POLL_DEREGISTER, 0);
 575                 lwkt_serialize_exit(ifp->if_serializer);
 576                 rc = 0;
 577         } else {
 578                 pctx->pr[pctx->poll_handlers].ifp = ifp;
 579                 pctx->poll_handlers++;
 580                 rc = 1;
 581         }
 582         crit_exit();
 583         return (rc);
 584 }
 585
 586 /*
 587  * Remove interface from the polling list.  Occurs when polling is turned
 588  * off.  Called from mainline code only, not called from an interrupt.
 589  */
 590 int
 591 ether_poll_deregister(struct ifnet *ifp)
 592 {
 593         struct pollctx *pctx;
 594         int i;
 595
 596         KKASSERT(ifp != NULL);
 597
 598         if (poll_defcpu < 0)
 599                 return 0;
 600         KKASSERT(poll_defcpu < POLLCTX_MAX);
 601
 602         pctx = poll_context[poll_defcpu];
 603         KKASSERT(pctx != NULL);
 604         KKASSERT(pctx->poll_cpuid == poll_defcpu);
 605
 606         crit_enter();
 607         if ((ifp->if_flags & IFF_POLLING) == 0) {
 608                 crit_exit();
 609                 return 0;
 610         }
 611         for (i = 0 ; i < pctx->poll_handlers ; i++) {
 612                 if (pctx->pr[i].ifp == ifp) /* found it */
 613                         break;
 614         }
 615         ifp->if_flags &= ~IFF_POLLING; /* found or not... */
 616         if (i == pctx->poll_handlers) {
 617                 crit_exit();
 618                 kprintf("ether_poll_deregister: ifp not found!!!\n");
 619                 return 0;
 620         }
 621         pctx->poll_handlers--;
 622         if (i < pctx->poll_handlers) { /* Last entry replaces this one. */
 623                 pctx->pr[i].ifp = pctx->pr[pctx->poll_handlers].ifp;
 624         }
 625         crit_exit();
 626
 627         /*
 628          * Only call the deregistration function if the interface is still
 629          * in a run state.
 630          */
 631         if (ifp->if_flags & IFF_RUNNING) {
 632                 lwkt_serialize_enter(ifp->if_serializer);
 633                 ifp->if_poll(ifp, POLL_DEREGISTER, 1);
 634                 lwkt_serialize_exit(ifp->if_serializer);
 635         }
 636         return (1);
 637 }
 638
 639 static void
 640 poll_add_sysctl(struct sysctl_ctx_list *ctx, struct sysctl_oid_list *parent,
 641                 struct pollctx *pctx)
 642 {
 643         SYSCTL_ADD_PROC(ctx, parent, OID_AUTO, "enable",
 644                         CTLTYPE_INT | CTLFLAG_RW, pctx, 0, sysctl_polling,
 645                         "I", "Polling enabled");
 646
 647         SYSCTL_ADD_PROC(ctx, parent, OID_AUTO, "pollhz",
 648                         CTLTYPE_INT | CTLFLAG_RW, pctx, 0, sysctl_pollhz,
 649                         "I", "Device polling frequency");
 650
 651         SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "phase", CTLFLAG_RD,
 652                         &pctx->phase, 0, "Polling phase");
 653
 654         SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "suspect", CTLFLAG_RW,
 655                         &pctx->suspect, 0, "suspect event");
 656
 657         SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "stalled", CTLFLAG_RW,
 658                         &pctx->stalled, 0, "potential stalls");
 659
 660         SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "burst", CTLFLAG_RW,
 661                         &pctx->poll_burst, 0, "Current polling burst size");
 662
 663         SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "each_burst", CTLFLAG_RW,
 664                         &pctx->poll_each_burst, 0, "Max size of each burst");
 665
 666         SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "burst_max", CTLFLAG_RW,
 667                         &pctx->poll_burst_max, 0, "Max Polling burst size");
 668
 669         SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "user_frac", CTLFLAG_RW,
 670                         &pctx->user_frac, 0,
 671                         "Desired user fraction of cpu time");
 672
 673         SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "reg_frac", CTLFLAG_RW,
 674                         &pctx->reg_frac, 0,
 675                         "Every this many cycles poll register");
 676
 677         SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "short_ticks", CTLFLAG_RW,
 678                         &pctx->short_ticks, 0,
 679                         "Hardclock ticks shorter than they should be");
 680
 681         SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "lost_polls", CTLFLAG_RW,
 682                         &pctx->lost_polls, 0,
 683                         "How many times we would have lost a poll tick");
 684
 685         SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "pending_polls", CTLFLAG_RD,
 686                         &pctx->pending_polls, 0, "Do we need to poll again");
 687
 688         SYSCTL_ADD_INT(ctx, parent, OID_AUTO, "residual_burst", CTLFLAG_RW,
 689                        &pctx->residual_burst, 0,
 690                        "# of residual cycles in burst");
 691
 692         SYSCTL_ADD_UINT(ctx, parent, OID_AUTO, "handlers", CTLFLAG_RD,
 693                         &pctx->poll_handlers, 0,
 694                         "Number of registered poll handlers");
 695 }