sys/kern/kern_event.c

   1 /*-
   2  * Copyright (c) 1999,2000,2001 Jonathan Lemon <jlemon@FreeBSD.org>
   3  * All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice, this list of conditions and the following disclaimer.
  10  * 2. Redistributions in binary form must reproduce the above copyright
  11  *    notice, this list of conditions and the following disclaimer in the
  12  *    documentation and/or other materials provided with the distribution.
  13  *
  14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  24  * SUCH DAMAGE.
  25  *
  26  * $FreeBSD: src/sys/kern/kern_event.c,v 1.2.2.10 2004/04/04 07:03:14 cperciva Exp $
  27  */
  28
  29 #include <sys/param.h>
  30 #include <sys/systm.h>
  31 #include <sys/kernel.h>
  32 #include <sys/proc.h>
  33 #include <sys/malloc.h>
  34 #include <sys/unistd.h>
  35 #include <sys/file.h>
  36 #include <sys/lock.h>
  37 #include <sys/fcntl.h>
  38 #include <sys/queue.h>
  39 #include <sys/event.h>
  40 #include <sys/eventvar.h>
  41 #include <sys/protosw.h>
  42 #include <sys/socket.h>
  43 #include <sys/socketvar.h>
  44 #include <sys/stat.h>
  45 #include <sys/sysctl.h>
  46 #include <sys/sysproto.h>
  47 #include <sys/thread.h>
  48 #include <sys/uio.h>
  49 #include <sys/signalvar.h>
  50 #include <sys/filio.h>
  51 #include <sys/ktr.h>
  52
  53 #include <sys/thread2.h>
  54 #include <sys/file2.h>
  55 #include <sys/mplock2.h>
  56
  57 #define EVENT_REGISTER  1
  58 #define EVENT_PROCESS   2
  59
  60 MALLOC_DEFINE(M_KQUEUE, "kqueue", "memory for kqueue system");
  61
  62 struct kevent_copyin_args {
  63         struct kevent_args      *ka;
  64         int                     pchanges;
  65 };
  66
  67 #define KNOTE_CACHE_MAX         8
  68
  69 struct knote_cache_list {
  70         struct klist            knote_cache;
  71         int                     knote_cache_cnt;
  72 } __cachealign;
  73
  74 static int      kqueue_scan(struct kqueue *kq, struct kevent *kevp, int count,
  75                     struct knote *marker);
  76 static int      kqueue_read(struct file *fp, struct uio *uio,
  77                     struct ucred *cred, int flags);
  78 static int      kqueue_write(struct file *fp, struct uio *uio,
  79                     struct ucred *cred, int flags);
  80 static int      kqueue_ioctl(struct file *fp, u_long com, caddr_t data,
  81                     struct ucred *cred, struct sysmsg *msg);
  82 static int      kqueue_kqfilter(struct file *fp, struct knote *kn);
  83 static int      kqueue_stat(struct file *fp, struct stat *st,
  84                     struct ucred *cred);
  85 static int      kqueue_close(struct file *fp);
  86 static void     kqueue_wakeup(struct kqueue *kq);
  87 static int      filter_attach(struct knote *kn);
  88 static int      filter_event(struct knote *kn, long hint);
  89
  90 /*
  91  * MPSAFE
  92  */
  93 static struct fileops kqueueops = {
  94         .fo_read = kqueue_read,
  95         .fo_write = kqueue_write,
  96         .fo_ioctl = kqueue_ioctl,
  97         .fo_kqfilter = kqueue_kqfilter,
  98         .fo_stat = kqueue_stat,
  99         .fo_close = kqueue_close,
 100         .fo_shutdown = nofo_shutdown
 101 };
 102
 103 static void     knote_attach(struct knote *kn);
 104 static void     knote_drop(struct knote *kn);
 105 static void     knote_detach_and_drop(struct knote *kn);
 106 static void     knote_enqueue(struct knote *kn);
 107 static void     knote_dequeue(struct knote *kn);
 108 static struct   knote *knote_alloc(void);
 109 static void     knote_free(struct knote *kn);
 110
 111 static void     precise_sleep_intr(systimer_t info, int in_ipi,
 112                                    struct intrframe *frame);
 113 static int      precise_sleep(void *ident, int flags, const char *wmesg,
 114                               int us);
 115
 116 static void     filt_kqdetach(struct knote *kn);
 117 static int      filt_kqueue(struct knote *kn, long hint);
 118 static int      filt_procattach(struct knote *kn);
 119 static void     filt_procdetach(struct knote *kn);
 120 static int      filt_proc(struct knote *kn, long hint);
 121 static int      filt_fileattach(struct knote *kn);
 122 static void     filt_timerexpire(void *knx);
 123 static int      filt_timerattach(struct knote *kn);
 124 static void     filt_timerdetach(struct knote *kn);
 125 static int      filt_timer(struct knote *kn, long hint);
 126 static int      filt_userattach(struct knote *kn);
 127 static void     filt_userdetach(struct knote *kn);
 128 static int      filt_user(struct knote *kn, long hint);
 129 static void     filt_usertouch(struct knote *kn, struct kevent *kev,
 130                                 u_long type);
 131 static int      filt_fsattach(struct knote *kn);
 132 static void     filt_fsdetach(struct knote *kn);
 133 static int      filt_fs(struct knote *kn, long hint);
 134
 135 static struct filterops file_filtops =
 136         { FILTEROP_ISFD | FILTEROP_MPSAFE, filt_fileattach, NULL, NULL };
 137 static struct filterops kqread_filtops =
 138         { FILTEROP_ISFD | FILTEROP_MPSAFE, NULL, filt_kqdetach, filt_kqueue };
 139 static struct filterops proc_filtops =
 140         { FILTEROP_MPSAFE, filt_procattach, filt_procdetach, filt_proc };
 141 static struct filterops timer_filtops =
 142         { FILTEROP_MPSAFE, filt_timerattach, filt_timerdetach, filt_timer };
 143 static struct filterops user_filtops =
 144         { FILTEROP_MPSAFE, filt_userattach, filt_userdetach, filt_user };
 145 static struct filterops fs_filtops =
 146         { FILTEROP_MPSAFE, filt_fsattach, filt_fsdetach, filt_fs };
 147
 148 static int              kq_ncallouts = 0;
 149 static int              kq_calloutmax = (4 * 1024);
 150 SYSCTL_INT(_kern, OID_AUTO, kq_calloutmax, CTLFLAG_RW,
 151     &kq_calloutmax, 0, "Maximum number of callouts allocated for kqueue");
 152 static int              kq_checkloop = 1000000;
 153 SYSCTL_INT(_kern, OID_AUTO, kq_checkloop, CTLFLAG_RW,
 154     &kq_checkloop, 0, "Maximum number of loops for kqueue scan");
 155 static int              kq_sleep_threshold = 20000;
 156 SYSCTL_INT(_kern, OID_AUTO, kq_sleep_threshold, CTLFLAG_RW,
 157     &kq_sleep_threshold, 0, "Minimum sleep duration without busy-looping");
 158
 159 #define KNOTE_ACTIVATE(kn) do {                                         \
 160         kn->kn_status |= KN_ACTIVE;                                     \
 161         if ((kn->kn_status & (KN_QUEUED | KN_DISABLED)) == 0)           \
 162                 knote_enqueue(kn);                                      \
 163 } while(0)
 164
 165 #define KN_HASHSIZE             64              /* XXX should be tunable */
 166 #define KN_HASH(val, mask)      (((val) ^ (val >> 8)) & (mask))
 167
 168 extern struct filterops aio_filtops;
 169 extern struct filterops sig_filtops;
 170
 171 /*
 172  * Table for for all system-defined filters.
 173  */
 174 static struct filterops *sysfilt_ops[] = {
 175         &file_filtops,                  /* EVFILT_READ */
 176         &file_filtops,                  /* EVFILT_WRITE */
 177         &aio_filtops,                   /* EVFILT_AIO */
 178         &file_filtops,                  /* EVFILT_VNODE */
 179         &proc_filtops,                  /* EVFILT_PROC */
 180         &sig_filtops,                   /* EVFILT_SIGNAL */
 181         &timer_filtops,                 /* EVFILT_TIMER */
 182         &file_filtops,                  /* EVFILT_EXCEPT */
 183         &user_filtops,                  /* EVFILT_USER */
 184         &fs_filtops,                    /* EVFILT_FS */
 185 };
 186
 187 static struct knote_cache_list  knote_cache_lists[MAXCPU];
 188
 189 /*
 190  * Acquire a knote, return non-zero on success, 0 on failure.
 191  *
 192  * If we cannot acquire the knote we sleep and return 0.  The knote
 193  * may be stale on return in this case and the caller must restart
 194  * whatever loop they are in.
 195  *
 196  * Related kq token must be held.
 197  */
 198 static __inline int
 199 knote_acquire(struct knote *kn)
 200 {
 201         if (kn->kn_status & KN_PROCESSING) {
 202                 kn->kn_status |= KN_WAITING | KN_REPROCESS;
 203                 tsleep(kn, 0, "kqepts", hz);
 204                 /* knote may be stale now */
 205                 return(0);
 206         }
 207         kn->kn_status |= KN_PROCESSING;
 208         return(1);
 209 }
 210
 211 /*
 212  * Release an acquired knote, clearing KN_PROCESSING and handling any
 213  * KN_REPROCESS events.
 214  *
 215  * Caller must be holding the related kq token
 216  *
 217  * Non-zero is returned if the knote is destroyed or detached.
 218  */
 219 static __inline int
 220 knote_release(struct knote *kn)
 221 {
 222         int ret;
 223
 224         while (kn->kn_status & KN_REPROCESS) {
 225                 kn->kn_status &= ~KN_REPROCESS;
 226                 if (kn->kn_status & KN_WAITING) {
 227                         kn->kn_status &= ~KN_WAITING;
 228                         wakeup(kn);
 229                 }
 230                 if (kn->kn_status & KN_DELETING) {
 231                         knote_detach_and_drop(kn);
 232                         return(1);
 233                         /* NOT REACHED */
 234                 }
 235                 if (filter_event(kn, 0))
 236                         KNOTE_ACTIVATE(kn);
 237         }
 238         if (kn->kn_status & KN_DETACHED)
 239                 ret = 1;
 240         else
 241                 ret = 0;
 242         kn->kn_status &= ~KN_PROCESSING;
 243         /* kn should not be accessed anymore */
 244         return ret;
 245 }
 246
 247 static int
 248 filt_fileattach(struct knote *kn)
 249 {
 250         return (fo_kqfilter(kn->kn_fp, kn));
 251 }
 252
 253 /*
 254  * MPSAFE
 255  */
 256 static int
 257 kqueue_kqfilter(struct file *fp, struct knote *kn)
 258 {
 259         struct kqueue *kq = (struct kqueue *)kn->kn_fp->f_data;
 260
 261         if (kn->kn_filter != EVFILT_READ)
 262                 return (EOPNOTSUPP);
 263
 264         kn->kn_fop = &kqread_filtops;
 265         knote_insert(&kq->kq_kqinfo.ki_note, kn);
 266         return (0);
 267 }
 268
 269 static void
 270 filt_kqdetach(struct knote *kn)
 271 {
 272         struct kqueue *kq = (struct kqueue *)kn->kn_fp->f_data;
 273
 274         knote_remove(&kq->kq_kqinfo.ki_note, kn);
 275 }
 276
 277 /*ARGSUSED*/
 278 static int
 279 filt_kqueue(struct knote *kn, long hint)
 280 {
 281         struct kqueue *kq = (struct kqueue *)kn->kn_fp->f_data;
 282
 283         kn->kn_data = kq->kq_count;
 284         return (kn->kn_data > 0);
 285 }
 286
 287 static int
 288 filt_procattach(struct knote *kn)
 289 {
 290         struct proc *p;
 291         int immediate;
 292
 293         immediate = 0;
 294         p = pfind(kn->kn_id);
 295         if (p == NULL && (kn->kn_sfflags & NOTE_EXIT)) {
 296                 p = zpfind(kn->kn_id);
 297                 immediate = 1;
 298         }
 299         if (p == NULL) {
 300                 return (ESRCH);
 301         }
 302         if (!PRISON_CHECK(curthread->td_ucred, p->p_ucred)) {
 303                 if (p)
 304                         PRELE(p);
 305                 return (EACCES);
 306         }
 307
 308         lwkt_gettoken(&p->p_token);
 309         kn->kn_ptr.p_proc = p;
 310         kn->kn_flags |= EV_CLEAR;               /* automatically set */
 311
 312         /*
 313          * internal flag indicating registration done by kernel
 314          */
 315         if (kn->kn_flags & EV_FLAG1) {
 316                 kn->kn_data = kn->kn_sdata;             /* ppid */
 317                 kn->kn_fflags = NOTE_CHILD;
 318                 kn->kn_flags &= ~EV_FLAG1;
 319         }
 320
 321         knote_insert(&p->p_klist, kn);
 322
 323         /*
 324          * Immediately activate any exit notes if the target process is a
 325          * zombie.  This is necessary to handle the case where the target
 326          * process, e.g. a child, dies before the kevent is negistered.
 327          */
 328         if (immediate && filt_proc(kn, NOTE_EXIT))
 329                 KNOTE_ACTIVATE(kn);
 330         lwkt_reltoken(&p->p_token);
 331         PRELE(p);
 332
 333         return (0);
 334 }
 335
 336 /*
 337  * The knote may be attached to a different process, which may exit,
 338  * leaving nothing for the knote to be attached to.  So when the process
 339  * exits, the knote is marked as DETACHED and also flagged as ONESHOT so
 340  * it will be deleted when read out.  However, as part of the knote deletion,
 341  * this routine is called, so a check is needed to avoid actually performing
 342  * a detach, because the original process does not exist any more.
 343  */
 344 static void
 345 filt_procdetach(struct knote *kn)
 346 {
 347         struct proc *p;
 348
 349         if (kn->kn_status & KN_DETACHED)
 350                 return;
 351         p = kn->kn_ptr.p_proc;
 352         knote_remove(&p->p_klist, kn);
 353 }
 354
 355 static int
 356 filt_proc(struct knote *kn, long hint)
 357 {
 358         u_int event;
 359
 360         /*
 361          * mask off extra data
 362          */
 363         event = (u_int)hint & NOTE_PCTRLMASK;
 364
 365         /*
 366          * if the user is interested in this event, record it.
 367          */
 368         if (kn->kn_sfflags & event)
 369                 kn->kn_fflags |= event;
 370
 371         /*
 372          * Process is gone, so flag the event as finished.  Detach the
 373          * knote from the process now because the process will be poof,
 374          * gone later on.
 375          */
 376         if (event == NOTE_EXIT) {
 377                 struct proc *p = kn->kn_ptr.p_proc;
 378                 if ((kn->kn_status & KN_DETACHED) == 0) {
 379                         PHOLD(p);
 380                         knote_remove(&p->p_klist, kn);
 381                         kn->kn_status |= KN_DETACHED;
 382                         kn->kn_data = p->p_xstat;
 383                         kn->kn_ptr.p_proc = NULL;
 384                         PRELE(p);
 385                 }
 386                 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT);
 387                 return (1);
 388         }
 389
 390         /*
 391          * process forked, and user wants to track the new process,
 392          * so attach a new knote to it, and immediately report an
 393          * event with the parent's pid.
 394          */
 395         if ((event == NOTE_FORK) && (kn->kn_sfflags & NOTE_TRACK)) {
 396                 struct kevent kev;
 397                 int error;
 398
 399                 /*
 400                  * register knote with new process.
 401                  */
 402                 kev.ident = hint & NOTE_PDATAMASK;      /* pid */
 403                 kev.filter = kn->kn_filter;
 404                 kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1;
 405                 kev.fflags = kn->kn_sfflags;
 406                 kev.data = kn->kn_id;                   /* parent */
 407                 kev.udata = kn->kn_kevent.udata;        /* preserve udata */
 408                 error = kqueue_register(kn->kn_kq, &kev);
 409                 if (error)
 410                         kn->kn_fflags |= NOTE_TRACKERR;
 411         }
 412
 413         return (kn->kn_fflags != 0);
 414 }
 415
 416 static void
 417 filt_timerreset(struct knote *kn)
 418 {
 419         struct callout *calloutp;
 420         struct timeval tv;
 421         int tticks;
 422
 423         tv.tv_sec = kn->kn_sdata / 1000;
 424         tv.tv_usec = (kn->kn_sdata % 1000) * 1000;
 425         tticks = tvtohz_high(&tv);
 426         calloutp = (struct callout *)kn->kn_hook;
 427         callout_reset(calloutp, tticks, filt_timerexpire, kn);
 428 }
 429
 430 /*
 431  * The callout interlocks with callout_terminate() but can still
 432  * race a deletion so if KN_DELETING is set we just don't touch
 433  * the knote.
 434  */
 435 static void
 436 filt_timerexpire(void *knx)
 437 {
 438         struct knote *kn = knx;
 439         struct kqueue *kq = kn->kn_kq;
 440
 441         lwkt_getpooltoken(kq);
 442
 443         /*
 444          * Open knote_acquire(), since we can't sleep in callout,
 445          * however, we do need to record this expiration.
 446          */
 447         kn->kn_data++;
 448         if (kn->kn_status & KN_PROCESSING) {
 449                 kn->kn_status |= KN_REPROCESS;
 450                 if ((kn->kn_status & KN_DELETING) == 0 &&
 451                     (kn->kn_flags & EV_ONESHOT) == 0)
 452                         filt_timerreset(kn);
 453                 lwkt_relpooltoken(kq);
 454                 return;
 455         }
 456         KASSERT((kn->kn_status & KN_DELETING) == 0,
 457             ("acquire a deleting knote %#x", kn->kn_status));
 458         kn->kn_status |= KN_PROCESSING;
 459
 460         KNOTE_ACTIVATE(kn);
 461         if ((kn->kn_flags & EV_ONESHOT) == 0)
 462                 filt_timerreset(kn);
 463
 464         knote_release(kn);
 465
 466         lwkt_relpooltoken(kq);
 467 }
 468
 469 /*
 470  * data contains amount of time to sleep, in milliseconds
 471  */
 472 static int
 473 filt_timerattach(struct knote *kn)
 474 {
 475         struct callout *calloutp;
 476         int prev_ncallouts;
 477
 478         prev_ncallouts = atomic_fetchadd_int(&kq_ncallouts, 1);
 479         if (prev_ncallouts >= kq_calloutmax) {
 480                 atomic_subtract_int(&kq_ncallouts, 1);
 481                 kn->kn_hook = NULL;
 482                 return (ENOMEM);
 483         }
 484
 485         kn->kn_flags |= EV_CLEAR;               /* automatically set */
 486         calloutp = kmalloc(sizeof(*calloutp), M_KQUEUE, M_WAITOK);
 487         callout_init_mp(calloutp);
 488         kn->kn_hook = (caddr_t)calloutp;
 489
 490         filt_timerreset(kn);
 491         return (0);
 492 }
 493
 494 /*
 495  * This function is called with the knote flagged locked but it is
 496  * still possible to race a callout event due to the callback blocking.
 497  * We must call callout_terminate() instead of callout_stop() to deal
 498  * with the race.
 499  */
 500 static void
 501 filt_timerdetach(struct knote *kn)
 502 {
 503         struct callout *calloutp;
 504
 505         calloutp = (struct callout *)kn->kn_hook;
 506         callout_terminate(calloutp);
 507         kfree(calloutp, M_KQUEUE);
 508         atomic_subtract_int(&kq_ncallouts, 1);
 509 }
 510
 511 static int
 512 filt_timer(struct knote *kn, long hint)
 513 {
 514
 515         return (kn->kn_data != 0);
 516 }
 517
 518 /*
 519  * EVFILT_USER
 520  */
 521 static int
 522 filt_userattach(struct knote *kn)
 523 {
 524         kn->kn_hook = NULL;
 525         if (kn->kn_fflags & NOTE_TRIGGER)
 526                 kn->kn_ptr.hookid = 1;
 527         else
 528                 kn->kn_ptr.hookid = 0;
 529         return 0;
 530 }
 531
 532 static void
 533 filt_userdetach(struct knote *kn)
 534 {
 535         /* nothing to do */
 536 }
 537
 538 static int
 539 filt_user(struct knote *kn, long hint)
 540 {
 541         return (kn->kn_ptr.hookid);
 542 }
 543
 544 static void
 545 filt_usertouch(struct knote *kn, struct kevent *kev, u_long type)
 546 {
 547         u_int ffctrl;
 548
 549         switch (type) {
 550         case EVENT_REGISTER:
 551                 if (kev->fflags & NOTE_TRIGGER)
 552                         kn->kn_ptr.hookid = 1;
 553
 554                 ffctrl = kev->fflags & NOTE_FFCTRLMASK;
 555                 kev->fflags &= NOTE_FFLAGSMASK;
 556                 switch (ffctrl) {
 557                 case NOTE_FFNOP:
 558                         break;
 559
 560                 case NOTE_FFAND:
 561                         kn->kn_sfflags &= kev->fflags;
 562                         break;
 563
 564                 case NOTE_FFOR:
 565                         kn->kn_sfflags |= kev->fflags;
 566                         break;
 567
 568                 case NOTE_FFCOPY:
 569                         kn->kn_sfflags = kev->fflags;
 570                         break;
 571
 572                 default:
 573                         /* XXX Return error? */
 574                         break;
 575                 }
 576                 kn->kn_sdata = kev->data;
 577
 578                 /*
 579                  * This is not the correct use of EV_CLEAR in an event
 580                  * modification, it should have been passed as a NOTE instead.
 581                  * But we need to maintain compatibility with Apple & FreeBSD.
 582                  *
 583                  * Note however that EV_CLEAR can still be used when doing
 584                  * the initial registration of the event and works as expected
 585                  * (clears the event on reception).
 586                  */
 587                 if (kev->flags & EV_CLEAR) {
 588                         kn->kn_ptr.hookid = 0;
 589                         kn->kn_data = 0;
 590                         kn->kn_fflags = 0;
 591                 }
 592                 break;
 593
 594         case EVENT_PROCESS:
 595                 *kev = kn->kn_kevent;
 596                 kev->fflags = kn->kn_sfflags;
 597                 kev->data = kn->kn_sdata;
 598                 if (kn->kn_flags & EV_CLEAR) {
 599                         kn->kn_ptr.hookid = 0;
 600                         /* kn_data, kn_fflags handled by parent */
 601                 }
 602                 break;
 603
 604         default:
 605                 panic("filt_usertouch() - invalid type (%ld)", type);
 606                 break;
 607         }
 608 }
 609
 610 /*
 611  * EVFILT_FS
 612  */
 613 struct klist fs_klist = SLIST_HEAD_INITIALIZER(&fs_klist);
 614
 615 static int
 616 filt_fsattach(struct knote *kn)
 617 {
 618         kn->kn_flags |= EV_CLEAR;
 619         knote_insert(&fs_klist, kn);
 620
 621         return (0);
 622 }
 623
 624 static void
 625 filt_fsdetach(struct knote *kn)
 626 {
 627         knote_remove(&fs_klist, kn);
 628 }
 629
 630 static int
 631 filt_fs(struct knote *kn, long hint)
 632 {
 633         kn->kn_fflags |= hint;
 634         return (kn->kn_fflags != 0);
 635 }
 636
 637 /*
 638  * Initialize a kqueue.
 639  *
 640  * NOTE: The lwp/proc code initializes a kqueue for select/poll ops.
 641  *
 642  * MPSAFE
 643  */
 644 void
 645 kqueue_init(struct kqueue *kq, struct filedesc *fdp)
 646 {
 647         TAILQ_INIT(&kq->kq_knpend);
 648         TAILQ_INIT(&kq->kq_knlist);
 649         kq->kq_count = 0;
 650         kq->kq_fdp = fdp;
 651         SLIST_INIT(&kq->kq_kqinfo.ki_note);
 652 }
 653
 654 /*
 655  * Terminate a kqueue.  Freeing the actual kq itself is left up to the
 656  * caller (it might be embedded in a lwp so we don't do it here).
 657  *
 658  * The kq's knlist must be completely eradicated so block on any
 659  * processing races.
 660  */
 661 void
 662 kqueue_terminate(struct kqueue *kq)
 663 {
 664         struct knote *kn;
 665
 666         lwkt_getpooltoken(kq);
 667         while ((kn = TAILQ_FIRST(&kq->kq_knlist)) != NULL) {
 668                 if (knote_acquire(kn))
 669                         knote_detach_and_drop(kn);
 670         }
 671         lwkt_relpooltoken(kq);
 672
 673         if (kq->kq_knhash) {
 674                 hashdestroy(kq->kq_knhash, M_KQUEUE, kq->kq_knhashmask);
 675                 kq->kq_knhash = NULL;
 676                 kq->kq_knhashmask = 0;
 677         }
 678 }
 679
 680 /*
 681  * MPSAFE
 682  */
 683 int
 684 sys_kqueue(struct kqueue_args *uap)
 685 {
 686         struct thread *td = curthread;
 687         struct kqueue *kq;
 688         struct file *fp;
 689         int fd, error;
 690
 691         error = falloc(td->td_lwp, &fp, &fd);
 692         if (error)
 693                 return (error);
 694         fp->f_flag = FREAD | FWRITE;
 695         fp->f_type = DTYPE_KQUEUE;
 696         fp->f_ops = &kqueueops;
 697
 698         kq = kmalloc(sizeof(struct kqueue), M_KQUEUE, M_WAITOK | M_ZERO);
 699         kqueue_init(kq, td->td_proc->p_fd);
 700         fp->f_data = kq;
 701
 702         fsetfd(kq->kq_fdp, fp, fd);
 703         uap->sysmsg_result = fd;
 704         fdrop(fp);
 705         return (error);
 706 }
 707
 708 /*
 709  * Copy 'count' items into the destination list pointed to by uap->eventlist.
 710  */
 711 static int
 712 kevent_copyout(void *arg, struct kevent *kevp, int count, int *res)
 713 {
 714         struct kevent_copyin_args *kap;
 715         int error;
 716
 717         kap = (struct kevent_copyin_args *)arg;
 718
 719         error = copyout(kevp, kap->ka->eventlist, count * sizeof(*kevp));
 720         if (error == 0) {
 721                 kap->ka->eventlist += count;
 722                 *res += count;
 723         } else {
 724                 *res = -1;
 725         }
 726
 727         return (error);
 728 }
 729
 730 /*
 731  * Copy at most 'max' items from the list pointed to by kap->changelist,
 732  * return number of items in 'events'.
 733  */
 734 static int
 735 kevent_copyin(void *arg, struct kevent *kevp, int max, int *events)
 736 {
 737         struct kevent_copyin_args *kap;
 738         int error, count;
 739
 740         kap = (struct kevent_copyin_args *)arg;
 741
 742         count = min(kap->ka->nchanges - kap->pchanges, max);
 743         error = copyin(kap->ka->changelist, kevp, count * sizeof *kevp);
 744         if (error == 0) {
 745                 kap->ka->changelist += count;
 746                 kap->pchanges += count;
 747                 *events = count;
 748         }
 749
 750         return (error);
 751 }
 752
 753 /*
 754  * MPSAFE
 755  */
 756 int
 757 kern_kevent(struct kqueue *kq, int nevents, int *res, void *uap,
 758             k_copyin_fn kevent_copyinfn, k_copyout_fn kevent_copyoutfn,
 759             struct timespec *tsp_in, int flags)
 760 {
 761         struct kevent *kevp;
 762         struct timespec *tsp, ats;
 763         int i, n, total, error, nerrors = 0;
 764         int lres;
 765         int limit = kq_checkloop;
 766         struct kevent kev[KQ_NEVENTS];
 767         struct knote marker;
 768         struct lwkt_token *tok;
 769
 770         if (tsp_in == NULL || tsp_in->tv_sec || tsp_in->tv_nsec)
 771                 atomic_set_int(&curthread->td_mpflags, TDF_MP_BATCH_DEMARC);
 772
 773         tsp = tsp_in;
 774         *res = 0;
 775
 776         for (;;) {
 777                 n = 0;
 778                 error = kevent_copyinfn(uap, kev, KQ_NEVENTS, &n);
 779                 if (error)
 780                         return error;
 781                 if (n == 0)
 782                         break;
 783                 for (i = 0; i < n; i++) {
 784                         kevp = &kev[i];
 785                         kevp->flags &= ~EV_SYSFLAGS;
 786                         error = kqueue_register(kq, kevp);
 787
 788                         /*
 789                          * If a registration returns an error we
 790                          * immediately post the error.  The kevent()
 791                          * call itself will fail with the error if
 792                          * no space is available for posting.
 793                          *
 794                          * Such errors normally bypass the timeout/blocking
 795                          * code.  However, if the copyoutfn function refuses
 796                          * to post the error (see sys_poll()), then we
 797                          * ignore it too.
 798                          */
 799                         if (error || (kevp->flags & EV_RECEIPT)) {
 800                                 kevp->flags = EV_ERROR;
 801                                 kevp->data = error;
 802                                 lres = *res;
 803                                 kevent_copyoutfn(uap, kevp, 1, res);
 804                                 if (*res < 0) {
 805                                         return error;
 806                                 } else if (lres != *res) {
 807                                         nevents--;
 808                                         nerrors++;
 809                                 }
 810                         }
 811                 }
 812         }
 813         if (nerrors)
 814                 return 0;
 815
 816         /*
 817          * Acquire/wait for events - setup timeout
 818          */
 819         if (tsp != NULL) {
 820                 if (tsp->tv_sec || tsp->tv_nsec) {
 821                         getnanouptime(&ats);
 822                         timespecadd(tsp, &ats);         /* tsp = target time */
 823                 }
 824         }
 825
 826         /*
 827          * Loop as required.
 828          *
 829          * Collect as many events as we can. Sleeping on successive
 830          * loops is disabled if copyoutfn has incremented (*res).
 831          *
 832          * The loop stops if an error occurs, all events have been
 833          * scanned (the marker has been reached), or fewer than the
 834          * maximum number of events is found.
 835          *
 836          * The copyoutfn function does not have to increment (*res) in
 837          * order for the loop to continue.
 838          *
 839          * NOTE: doselect() usually passes 0x7FFFFFFF for nevents.
 840          */
 841         total = 0;
 842         error = 0;
 843         marker.kn_filter = EVFILT_MARKER;
 844         marker.kn_status = KN_PROCESSING;
 845         tok = lwkt_token_pool_lookup(kq);
 846         lwkt_gettoken(tok);
 847         TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker, kn_tqe);
 848         lwkt_reltoken(tok);
 849         while ((n = nevents - total) > 0) {
 850                 if (n > KQ_NEVENTS)
 851                         n = KQ_NEVENTS;
 852
 853                 /*
 854                  * If no events are pending sleep until timeout (if any)
 855                  * or an event occurs.
 856                  *
 857                  * After the sleep completes the marker is moved to the
 858                  * end of the list, making any received events available
 859                  * to our scan.
 860                  */
 861                 if (kq->kq_count == 0 && *res == 0) {
 862                         int timeout, ustimeout = 0;
 863
 864                         if (tsp == NULL) {
 865                                 timeout = 0;
 866                         } else if (tsp->tv_sec == 0 && tsp->tv_nsec == 0) {
 867                                 error = EWOULDBLOCK;
 868                                 break;
 869                         } else {
 870                                 struct timespec atx = *tsp;
 871
 872                                 getnanouptime(&ats);
 873                                 timespecsub(&atx, &ats);
 874                                 if (atx.tv_sec < 0) {
 875                                         error = EWOULDBLOCK;
 876                                         break;
 877                                 } else {
 878                                         timeout = atx.tv_sec > 24 * 60 * 60 ?
 879                                             24 * 60 * 60 * hz :
 880                                             tstohz_high(&atx);
 881                                 }
 882                                 if (flags & KEVENT_TIMEOUT_PRECISE &&
 883                                     timeout != 0) {
 884                                         if (atx.tv_sec == 0 &&
 885                                             atx.tv_nsec < kq_sleep_threshold) {
 886                                                 DELAY(atx.tv_nsec / 1000);
 887                                                 error = EWOULDBLOCK;
 888                                                 break;
 889                                         } else if (atx.tv_sec < 2000) {
 890                                                 ustimeout = atx.tv_sec *
 891                                                     1000000 + atx.tv_nsec/1000;
 892                                         } else {
 893                                                 ustimeout = 2000000000;
 894                                         }
 895                                 }
 896                         }
 897
 898                         lwkt_gettoken(tok);
 899                         if (kq->kq_count == 0) {
 900                                 kq->kq_sleep_cnt++;
 901                                 if (__predict_false(kq->kq_sleep_cnt == 0)) {
 902                                         /*
 903                                          * Guard against possible wrapping.  And
 904                                          * set it to 2, so that kqueue_wakeup()
 905                                          * can wake everyone up.
 906                                          */
 907                                         kq->kq_sleep_cnt = 2;
 908                                 }
 909                                 if ((flags & KEVENT_TIMEOUT_PRECISE) &&
 910                                     timeout != 0) {
 911                                         error = precise_sleep(kq, PCATCH,
 912                                             "kqread", ustimeout);
 913                                 } else {
 914                                         error = tsleep(kq, PCATCH, "kqread",
 915                                             timeout);
 916                                 }
 917
 918                                 /* don't restart after signals... */
 919                                 if (error == ERESTART)
 920                                         error = EINTR;
 921                                 if (error) {
 922                                         lwkt_reltoken(tok);
 923                                         break;
 924                                 }
 925
 926                                 TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe);
 927                                 TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker,
 928                                     kn_tqe);
 929                         }
 930                         lwkt_reltoken(tok);
 931                 }
 932
 933                 /*
 934                  * Process all received events
 935                  * Account for all non-spurious events in our total
 936                  */
 937                 i = kqueue_scan(kq, kev, n, &marker);
 938                 if (i) {
 939                         lres = *res;
 940                         error = kevent_copyoutfn(uap, kev, i, res);
 941                         total += *res - lres;
 942                         if (error)
 943                                 break;
 944                 }
 945                 if (limit && --limit == 0)
 946                         panic("kqueue: checkloop failed i=%d", i);
 947
 948                 /*
 949                  * Normally when fewer events are returned than requested
 950                  * we can stop.  However, if only spurious events were
 951                  * collected the copyout will not bump (*res) and we have
 952                  * to continue.
 953                  */
 954                 if (i < n && *res)
 955                         break;
 956
 957                 /*
 958                  * Deal with an edge case where spurious events can cause
 959                  * a loop to occur without moving the marker.  This can
 960                  * prevent kqueue_scan() from picking up new events which
 961                  * race us.  We must be sure to move the marker for this
 962                  * case.
 963                  *
 964                  * NOTE: We do not want to move the marker if events
 965                  *       were scanned because normal kqueue operations
 966                  *       may reactivate events.  Moving the marker in
 967                  *       that case could result in duplicates for the
 968                  *       same event.
 969                  */
 970                 if (i == 0) {
 971                         lwkt_gettoken(tok);
 972                         TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe);
 973                         TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker, kn_tqe);
 974                         lwkt_reltoken(tok);
 975                 }
 976         }
 977         lwkt_gettoken(tok);
 978         TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe);
 979         lwkt_reltoken(tok);
 980
 981         /* Timeouts do not return EWOULDBLOCK. */
 982         if (error == EWOULDBLOCK)
 983                 error = 0;
 984         return error;
 985 }
 986
 987 /*
 988  * MPALMOSTSAFE
 989  */
 990 int
 991 sys_kevent(struct kevent_args *uap)
 992 {
 993         struct thread *td = curthread;
 994         struct proc *p = td->td_proc;
 995         struct timespec ts, *tsp;
 996         struct kqueue *kq;
 997         struct file *fp = NULL;
 998         struct kevent_copyin_args *kap, ka;
 999         int error;
1000
1001         if (uap->timeout) {
1002                 error = copyin(uap->timeout, &ts, sizeof(ts));
1003                 if (error)
1004                         return (error);
1005                 tsp = &ts;
1006         } else {
1007                 tsp = NULL;
1008         }
1009         fp = holdfp(p->p_fd, uap->fd, -1);
1010         if (fp == NULL)
1011                 return (EBADF);
1012         if (fp->f_type != DTYPE_KQUEUE) {
1013                 fdrop(fp);
1014                 return (EBADF);
1015         }
1016
1017         kq = (struct kqueue *)fp->f_data;
1018
1019         kap = &ka;
1020         kap->ka = uap;
1021         kap->pchanges = 0;
1022
1023         error = kern_kevent(kq, uap->nevents, &uap->sysmsg_result, kap,
1024                             kevent_copyin, kevent_copyout, tsp, 0);
1025
1026         fdrop(fp);
1027
1028         return (error);
1029 }
1030
1031 int
1032 kqueue_register(struct kqueue *kq, struct kevent *kev)
1033 {
1034         struct filedesc *fdp = kq->kq_fdp;
1035         struct klist *list = NULL;
1036         struct filterops *fops;
1037         struct file *fp = NULL;
1038         struct knote *kn = NULL;
1039         struct thread *td;
1040         int error = 0;
1041         struct knote_cache_list *cache_list;
1042
1043         if (kev->filter < 0) {
1044                 if (kev->filter + EVFILT_SYSCOUNT < 0)
1045                         return (EINVAL);
1046                 fops = sysfilt_ops[~kev->filter];       /* to 0-base index */
1047         } else {
1048                 /*
1049                  * XXX
1050                  * filter attach routine is responsible for insuring that
1051                  * the identifier can be attached to it.
1052                  */
1053                 return (EINVAL);
1054         }
1055
1056         if (fops->f_flags & FILTEROP_ISFD) {
1057                 /* validate descriptor */
1058                 fp = holdfp(fdp, kev->ident, -1);
1059                 if (fp == NULL)
1060                         return (EBADF);
1061         }
1062
1063         cache_list = &knote_cache_lists[mycpuid];
1064         if (SLIST_EMPTY(&cache_list->knote_cache)) {
1065                 struct knote *new_kn;
1066
1067                 new_kn = knote_alloc();
1068                 crit_enter();
1069                 SLIST_INSERT_HEAD(&cache_list->knote_cache, new_kn, kn_link);
1070                 cache_list->knote_cache_cnt++;
1071                 crit_exit();
1072         }
1073
1074         td = curthread;
1075         lwkt_getpooltoken(kq);
1076
1077         /*
1078          * Make sure that only one thread can register event on this kqueue,
1079          * so that we would not suffer any race, even if the registration
1080          * blocked, i.e. kq token was released, and the kqueue was shared
1081          * between threads (this should be rare though).
1082          */
1083         while (__predict_false(kq->kq_regtd != NULL && kq->kq_regtd != td)) {
1084                 kq->kq_state |= KQ_REGWAIT;
1085                 tsleep(&kq->kq_regtd, 0, "kqreg", 0);
1086         }
1087         if (__predict_false(kq->kq_regtd != NULL)) {
1088                 /* Recursive calling of kqueue_register() */
1089                 td = NULL;
1090         } else {
1091                 /* Owner of the kq_regtd, i.e. td != NULL */
1092                 kq->kq_regtd = td;
1093         }
1094
1095         if (fp != NULL) {
1096                 list = &fp->f_klist;
1097         } else if (kq->kq_knhashmask) {
1098                 list = &kq->kq_knhash[
1099                     KN_HASH((u_long)kev->ident, kq->kq_knhashmask)];
1100         }
1101         if (list != NULL) {
1102                 lwkt_getpooltoken(list);
1103 again:
1104                 SLIST_FOREACH(kn, list, kn_link) {
1105                         if (kn->kn_kq == kq &&
1106                             kn->kn_filter == kev->filter &&
1107                             kn->kn_id == kev->ident) {
1108                                 if (knote_acquire(kn) == 0)
1109                                         goto again;
1110                                 break;
1111                         }
1112                 }
1113                 lwkt_relpooltoken(list);
1114         }
1115
1116         /*
1117          * NOTE: At this point if kn is non-NULL we will have acquired
1118          *       it and set KN_PROCESSING.
1119          */
1120         if (kn == NULL && ((kev->flags & EV_ADD) == 0)) {
1121                 error = ENOENT;
1122                 goto done;
1123         }
1124
1125         /*
1126          * kn now contains the matching knote, or NULL if no match
1127          */
1128         if (kev->flags & EV_ADD) {
1129                 if (kn == NULL) {
1130                         crit_enter();
1131                         kn = SLIST_FIRST(&cache_list->knote_cache);
1132                         if (kn == NULL) {
1133                                 crit_exit();
1134                                 kn = knote_alloc();
1135                         } else {
1136                                 SLIST_REMOVE_HEAD(&cache_list->knote_cache,
1137                                     kn_link);
1138                                 cache_list->knote_cache_cnt--;
1139                                 crit_exit();
1140                         }
1141                         kn->kn_fp = fp;
1142                         kn->kn_kq = kq;
1143                         kn->kn_fop = fops;
1144
1145                         /*
1146                          * apply reference count to knote structure, and
1147                          * do not release it at the end of this routine.
1148                          */
1149                         fp = NULL;
1150
1151                         kn->kn_sfflags = kev->fflags;
1152                         kn->kn_sdata = kev->data;
1153                         kev->fflags = 0;
1154                         kev->data = 0;
1155                         kn->kn_kevent = *kev;
1156
1157                         /*
1158                          * KN_PROCESSING prevents the knote from getting
1159                          * ripped out from under us while we are trying
1160                          * to attach it, in case the attach blocks.
1161                          */
1162                         kn->kn_status = KN_PROCESSING;
1163                         knote_attach(kn);
1164                         if ((error = filter_attach(kn)) != 0) {
1165                                 kn->kn_status |= KN_DELETING | KN_REPROCESS;
1166                                 knote_drop(kn);
1167                                 goto done;
1168                         }
1169
1170                         /*
1171                          * Interlock against close races which either tried
1172                          * to remove our knote while we were blocked or missed
1173                          * it entirely prior to our attachment.  We do not
1174                          * want to end up with a knote on a closed descriptor.
1175                          */
1176                         if ((fops->f_flags & FILTEROP_ISFD) &&
1177                             checkfdclosed(fdp, kev->ident, kn->kn_fp)) {
1178                                 kn->kn_status |= KN_DELETING | KN_REPROCESS;
1179                         }
1180                 } else {
1181                         /*
1182                          * The user may change some filter values after the
1183                          * initial EV_ADD, but doing so will not reset any
1184                          * filter which have already been triggered.
1185                          */
1186                         KKASSERT(kn->kn_status & KN_PROCESSING);
1187                         if (fops == &user_filtops) {
1188                                 filt_usertouch(kn, kev, EVENT_REGISTER);
1189                         } else {
1190                                 kn->kn_sfflags = kev->fflags;
1191                                 kn->kn_sdata = kev->data;
1192                                 kn->kn_kevent.udata = kev->udata;
1193                         }
1194                 }
1195
1196                 /*
1197                  * Execute the filter event to immediately activate the
1198                  * knote if necessary.  If reprocessing events are pending
1199                  * due to blocking above we do not run the filter here
1200                  * but instead let knote_release() do it.  Otherwise we
1201                  * might run the filter on a deleted event.
1202                  */
1203                 if ((kn->kn_status & KN_REPROCESS) == 0) {
1204                         if (filter_event(kn, 0))
1205                                 KNOTE_ACTIVATE(kn);
1206                 }
1207         } else if (kev->flags & EV_DELETE) {
1208                 /*
1209                  * Delete the existing knote
1210                  */
1211                 knote_detach_and_drop(kn);
1212                 goto done;
1213         } else {
1214                 /*
1215                  * Modify an existing event.
1216                  *
1217                  * The user may change some filter values after the
1218                  * initial EV_ADD, but doing so will not reset any
1219                  * filter which have already been triggered.
1220                  */
1221                 KKASSERT(kn->kn_status & KN_PROCESSING);
1222                 if (fops == &user_filtops) {
1223                         filt_usertouch(kn, kev, EVENT_REGISTER);
1224                 } else {
1225                         kn->kn_sfflags = kev->fflags;
1226                         kn->kn_sdata = kev->data;
1227                         kn->kn_kevent.udata = kev->udata;
1228                 }
1229
1230                 /*
1231                  * Execute the filter event to immediately activate the
1232                  * knote if necessary.  If reprocessing events are pending
1233                  * due to blocking above we do not run the filter here
1234                  * but instead let knote_release() do it.  Otherwise we
1235                  * might run the filter on a deleted event.
1236                  */
1237                 if ((kn->kn_status & KN_REPROCESS) == 0) {
1238                         if (filter_event(kn, 0))
1239                                 KNOTE_ACTIVATE(kn);
1240                 }
1241         }
1242
1243         /*
1244          * Disablement does not deactivate a knote here.
1245          */
1246         if ((kev->flags & EV_DISABLE) &&
1247             ((kn->kn_status & KN_DISABLED) == 0)) {
1248                 kn->kn_status |= KN_DISABLED;
1249         }
1250
1251         /*
1252          * Re-enablement may have to immediately enqueue an active knote.
1253          */
1254         if ((kev->flags & EV_ENABLE) && (kn->kn_status & KN_DISABLED)) {
1255                 kn->kn_status &= ~KN_DISABLED;
1256                 if ((kn->kn_status & KN_ACTIVE) &&
1257                     ((kn->kn_status & KN_QUEUED) == 0)) {
1258                         knote_enqueue(kn);
1259                 }
1260         }
1261
1262         /*
1263          * Handle any required reprocessing
1264          */
1265         knote_release(kn);
1266         /* kn may be invalid now */
1267
1268 done:
1269         if (td != NULL) { /* Owner of the kq_regtd */
1270                 kq->kq_regtd = NULL;
1271                 if (__predict_false(kq->kq_state & KQ_REGWAIT)) {
1272                         kq->kq_state &= ~KQ_REGWAIT;
1273                         wakeup(&kq->kq_regtd);
1274                 }
1275         }
1276         lwkt_relpooltoken(kq);
1277         if (fp != NULL)
1278                 fdrop(fp);
1279         return (error);
1280 }
1281
1282 /*
1283  * Scan the kqueue, return the number of active events placed in kevp up
1284  * to count.
1285  *
1286  * Continuous mode events may get recycled, do not continue scanning past
1287  * marker unless no events have been collected.
1288  */
1289 static int
1290 kqueue_scan(struct kqueue *kq, struct kevent *kevp, int count,
1291             struct knote *marker)
1292 {
1293         struct knote *kn, local_marker;
1294         int total;
1295
1296         total = 0;
1297         local_marker.kn_filter = EVFILT_MARKER;
1298         local_marker.kn_status = KN_PROCESSING;
1299
1300         lwkt_getpooltoken(kq);
1301
1302         /*
1303          * Collect events.
1304          */
1305         TAILQ_INSERT_HEAD(&kq->kq_knpend, &local_marker, kn_tqe);
1306         while (count) {
1307                 kn = TAILQ_NEXT(&local_marker, kn_tqe);
1308                 if (kn->kn_filter == EVFILT_MARKER) {
1309                         /* Marker reached, we are done */
1310                         if (kn == marker)
1311                                 break;
1312
1313                         /* Move local marker past some other threads marker */
1314                         kn = TAILQ_NEXT(kn, kn_tqe);
1315                         TAILQ_REMOVE(&kq->kq_knpend, &local_marker, kn_tqe);
1316                         TAILQ_INSERT_BEFORE(kn, &local_marker, kn_tqe);
1317                         continue;
1318                 }
1319
1320                 /*
1321                  * We can't skip a knote undergoing processing, otherwise
1322                  * we risk not returning it when the user process expects
1323                  * it should be returned.  Sleep and retry.
1324                  */
1325                 if (knote_acquire(kn) == 0)
1326                         continue;
1327
1328                 /*
1329                  * Remove the event for processing.
1330                  *
1331                  * WARNING!  We must leave KN_QUEUED set to prevent the
1332                  *           event from being KNOTE_ACTIVATE()d while
1333                  *           the queue state is in limbo, in case we
1334                  *           block.
1335                  */
1336                 TAILQ_REMOVE(&kq->kq_knpend, kn, kn_tqe);
1337                 kq->kq_count--;
1338
1339                 /*
1340                  * We have to deal with an extremely important race against
1341                  * file descriptor close()s here.  The file descriptor can
1342                  * disappear MPSAFE, and there is a small window of
1343                  * opportunity between that and the call to knote_fdclose().
1344                  *
1345                  * If we hit that window here while doselect or dopoll is
1346                  * trying to delete a spurious event they will not be able
1347                  * to match up the event against a knote and will go haywire.
1348                  */
1349                 if ((kn->kn_fop->f_flags & FILTEROP_ISFD) &&
1350                     checkfdclosed(kq->kq_fdp, kn->kn_kevent.ident, kn->kn_fp)) {
1351                         kn->kn_status |= KN_DELETING | KN_REPROCESS;
1352                 }
1353
1354                 if (kn->kn_status & KN_DISABLED) {
1355                         /*
1356                          * If disabled we ensure the event is not queued
1357                          * but leave its active bit set.  On re-enablement
1358                          * the event may be immediately triggered.
1359                          */
1360                         kn->kn_status &= ~KN_QUEUED;
1361                 } else if ((kn->kn_flags & EV_ONESHOT) == 0 &&
1362                            (kn->kn_status & KN_DELETING) == 0 &&
1363                            filter_event(kn, 0) == 0) {
1364                         /*
1365                          * If not running in one-shot mode and the event
1366                          * is no longer present we ensure it is removed
1367                          * from the queue and ignore it.
1368                          */
1369                         kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE);
1370                 } else {
1371                         /*
1372                          * Post the event
1373                          */
1374                         if (kn->kn_fop == &user_filtops)
1375                                 filt_usertouch(kn, kevp, EVENT_PROCESS);
1376                         else
1377                                 *kevp = kn->kn_kevent;
1378                         ++kevp;
1379                         ++total;
1380                         --count;
1381
1382                         if (kn->kn_flags & EV_ONESHOT) {
1383                                 kn->kn_status &= ~KN_QUEUED;
1384                                 kn->kn_status |= KN_DELETING | KN_REPROCESS;
1385                         } else {
1386                                 if (kn->kn_flags & (EV_CLEAR | EV_DISPATCH)) {
1387                                         if (kn->kn_flags & EV_CLEAR) {
1388                                                 kn->kn_data = 0;
1389                                                 kn->kn_fflags = 0;
1390                                         }
1391                                         if (kn->kn_flags & EV_DISPATCH) {
1392                                                 kn->kn_status |= KN_DISABLED;
1393                                         }
1394                                         kn->kn_status &= ~(KN_QUEUED |
1395                                                            KN_ACTIVE);
1396                                 } else {
1397                                         TAILQ_INSERT_TAIL(&kq->kq_knpend, kn, kn_tqe);
1398                                         kq->kq_count++;
1399                                 }
1400                         }
1401                 }
1402
1403                 /*
1404                  * Handle any post-processing states
1405                  */
1406                 knote_release(kn);
1407         }
1408         TAILQ_REMOVE(&kq->kq_knpend, &local_marker, kn_tqe);
1409
1410         lwkt_relpooltoken(kq);
1411         return (total);
1412 }
1413
1414 /*
1415  * XXX
1416  * This could be expanded to call kqueue_scan, if desired.
1417  *
1418  * MPSAFE
1419  */
1420 static int
1421 kqueue_read(struct file *fp, struct uio *uio, struct ucred *cred, int flags)
1422 {
1423         return (ENXIO);
1424 }
1425
1426 /*
1427  * MPSAFE
1428  */
1429 static int
1430 kqueue_write(struct file *fp, struct uio *uio, struct ucred *cred, int flags)
1431 {
1432         return (ENXIO);
1433 }
1434
1435 /*
1436  * MPALMOSTSAFE
1437  */
1438 static int
1439 kqueue_ioctl(struct file *fp, u_long com, caddr_t data,
1440              struct ucred *cred, struct sysmsg *msg)
1441 {
1442         struct kqueue *kq;
1443         int error;
1444
1445         kq = (struct kqueue *)fp->f_data;
1446         lwkt_getpooltoken(kq);
1447         switch(com) {
1448         case FIOASYNC:
1449                 if (*(int *)data)
1450                         kq->kq_state |= KQ_ASYNC;
1451                 else
1452                         kq->kq_state &= ~KQ_ASYNC;
1453                 error = 0;
1454                 break;
1455         case FIOSETOWN:
1456                 error = fsetown(*(int *)data, &kq->kq_sigio);
1457                 break;
1458         default:
1459                 error = ENOTTY;
1460                 break;
1461         }
1462         lwkt_relpooltoken(kq);
1463         return (error);
1464 }
1465
1466 /*
1467  * MPSAFE
1468  */
1469 static int
1470 kqueue_stat(struct file *fp, struct stat *st, struct ucred *cred)
1471 {
1472         struct kqueue *kq = (struct kqueue *)fp->f_data;
1473
1474         bzero((void *)st, sizeof(*st));
1475         st->st_size = kq->kq_count;
1476         st->st_blksize = sizeof(struct kevent);
1477         st->st_mode = S_IFIFO;
1478         return (0);
1479 }
1480
1481 /*
1482  * MPSAFE
1483  */
1484 static int
1485 kqueue_close(struct file *fp)
1486 {
1487         struct kqueue *kq = (struct kqueue *)fp->f_data;
1488
1489         kqueue_terminate(kq);
1490
1491         fp->f_data = NULL;
1492         funsetown(&kq->kq_sigio);
1493
1494         kfree(kq, M_KQUEUE);
1495         return (0);
1496 }
1497
1498 static void
1499 kqueue_wakeup(struct kqueue *kq)
1500 {
1501         if (kq->kq_sleep_cnt) {
1502                 u_int sleep_cnt = kq->kq_sleep_cnt;
1503
1504                 kq->kq_sleep_cnt = 0;
1505                 if (sleep_cnt == 1)
1506                         wakeup_one(kq);
1507                 else
1508                         wakeup(kq);
1509         }
1510         KNOTE(&kq->kq_kqinfo.ki_note, 0);
1511 }
1512
1513 /*
1514  * Calls filterops f_attach function, acquiring mplock if filter is not
1515  * marked as FILTEROP_MPSAFE.
1516  *
1517  * Caller must be holding the related kq token
1518  */
1519 static int
1520 filter_attach(struct knote *kn)
1521 {
1522         int ret;
1523
1524         if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) {
1525                 ret = kn->kn_fop->f_attach(kn);
1526         } else {
1527                 get_mplock();
1528                 ret = kn->kn_fop->f_attach(kn);
1529                 rel_mplock();
1530         }
1531         return (ret);
1532 }
1533
1534 /*
1535  * Detach the knote and drop it, destroying the knote.
1536  *
1537  * Calls filterops f_detach function, acquiring mplock if filter is not
1538  * marked as FILTEROP_MPSAFE.
1539  *
1540  * Caller must be holding the related kq token
1541  */
1542 static void
1543 knote_detach_and_drop(struct knote *kn)
1544 {
1545         kn->kn_status |= KN_DELETING | KN_REPROCESS;
1546         if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) {
1547                 kn->kn_fop->f_detach(kn);
1548         } else {
1549                 get_mplock();
1550                 kn->kn_fop->f_detach(kn);
1551                 rel_mplock();
1552         }
1553         knote_drop(kn);
1554 }
1555
1556 /*
1557  * Calls filterops f_event function, acquiring mplock if filter is not
1558  * marked as FILTEROP_MPSAFE.
1559  *
1560  * If the knote is in the middle of being created or deleted we cannot
1561  * safely call the filter op.
1562  *
1563  * Caller must be holding the related kq token
1564  */
1565 static int
1566 filter_event(struct knote *kn, long hint)
1567 {
1568         int ret;
1569
1570         if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) {
1571                 ret = kn->kn_fop->f_event(kn, hint);
1572         } else {
1573                 get_mplock();
1574                 ret = kn->kn_fop->f_event(kn, hint);
1575                 rel_mplock();
1576         }
1577         return (ret);
1578 }
1579
1580 /*
1581  * Walk down a list of knotes, activating them if their event has triggered.
1582  *
1583  * If we encounter any knotes which are undergoing processing we just mark
1584  * them for reprocessing and do not try to [re]activate the knote.  However,
1585  * if a hint is being passed we have to wait and that makes things a bit
1586  * sticky.
1587  */
1588 void
1589 knote(struct klist *list, long hint)
1590 {
1591         struct kqueue *kq;
1592         struct knote *kn;
1593         struct knote *kntmp;
1594
1595         lwkt_getpooltoken(list);
1596 restart:
1597         SLIST_FOREACH(kn, list, kn_next) {
1598                 kq = kn->kn_kq;
1599                 lwkt_getpooltoken(kq);
1600
1601                 /* temporary verification hack */
1602                 SLIST_FOREACH(kntmp, list, kn_next) {
1603                         if (kn == kntmp)
1604                                 break;
1605                 }
1606                 if (kn != kntmp || kn->kn_kq != kq) {
1607                         lwkt_relpooltoken(kq);
1608                         goto restart;
1609                 }
1610
1611                 if (kn->kn_status & KN_PROCESSING) {
1612                         /*
1613                          * Someone else is processing the knote, ask the
1614                          * other thread to reprocess it and don't mess
1615                          * with it otherwise.
1616                          */
1617                         if (hint == 0) {
1618                                 kn->kn_status |= KN_REPROCESS;
1619                                 lwkt_relpooltoken(kq);
1620                                 continue;
1621                         }
1622
1623                         /*
1624                          * If the hint is non-zero we have to wait or risk
1625                          * losing the state the caller is trying to update.
1626                          *
1627                          * XXX This is a real problem, certain process
1628                          *     and signal filters will bump kn_data for
1629                          *     already-processed notes more than once if
1630                          *     we restart the list scan.  FIXME.
1631                          */
1632                         kn->kn_status |= KN_WAITING | KN_REPROCESS;
1633                         tsleep(kn, 0, "knotec", hz);
1634                         lwkt_relpooltoken(kq);
1635                         goto restart;
1636                 }
1637
1638                 /*
1639                  * Become the reprocessing master ourselves.
1640                  *
1641                  * If hint is non-zero running the event is mandatory
1642                  * when not deleting so do it whether reprocessing is
1643                  * set or not.
1644                  */
1645                 kn->kn_status |= KN_PROCESSING;
1646                 if ((kn->kn_status & KN_DELETING) == 0) {
1647                         if (filter_event(kn, hint))
1648                                 KNOTE_ACTIVATE(kn);
1649                 }
1650                 if (knote_release(kn)) {
1651                         lwkt_relpooltoken(kq);
1652                         goto restart;
1653                 }
1654                 lwkt_relpooltoken(kq);
1655         }
1656         lwkt_relpooltoken(list);
1657 }
1658
1659 /*
1660  * Insert knote at head of klist.
1661  *
1662  * This function may only be called via a filter function and thus
1663  * kq_token should already be held and marked for processing.
1664  */
1665 void
1666 knote_insert(struct klist *klist, struct knote *kn)
1667 {
1668         lwkt_getpooltoken(klist);
1669         KKASSERT(kn->kn_status & KN_PROCESSING);
1670         SLIST_INSERT_HEAD(klist, kn, kn_next);
1671         lwkt_relpooltoken(klist);
1672 }
1673
1674 /*
1675  * Remove knote from a klist
1676  *
1677  * This function may only be called via a filter function and thus
1678  * kq_token should already be held and marked for processing.
1679  */
1680 void
1681 knote_remove(struct klist *klist, struct knote *kn)
1682 {
1683         lwkt_getpooltoken(klist);
1684         KKASSERT(kn->kn_status & KN_PROCESSING);
1685         SLIST_REMOVE(klist, kn, knote, kn_next);
1686         lwkt_relpooltoken(klist);
1687 }
1688
1689 void
1690 knote_assume_knotes(struct kqinfo *src, struct kqinfo *dst,
1691                     struct filterops *ops, void *hook)
1692 {
1693         struct kqueue *kq;
1694         struct knote *kn;
1695
1696         lwkt_getpooltoken(&src->ki_note);
1697         lwkt_getpooltoken(&dst->ki_note);
1698         while ((kn = SLIST_FIRST(&src->ki_note)) != NULL) {
1699                 kq = kn->kn_kq;
1700                 lwkt_getpooltoken(kq);
1701                 if (SLIST_FIRST(&src->ki_note) != kn || kn->kn_kq != kq) {
1702                         lwkt_relpooltoken(kq);
1703                         continue;
1704                 }
1705                 if (knote_acquire(kn)) {
1706                         knote_remove(&src->ki_note, kn);
1707                         kn->kn_fop = ops;
1708                         kn->kn_hook = hook;
1709                         knote_insert(&dst->ki_note, kn);
1710                         knote_release(kn);
1711                         /* kn may be invalid now */
1712                 }
1713                 lwkt_relpooltoken(kq);
1714         }
1715         lwkt_relpooltoken(&dst->ki_note);
1716         lwkt_relpooltoken(&src->ki_note);
1717 }
1718
1719 /*
1720  * Remove all knotes referencing a specified fd
1721  */
1722 void
1723 knote_fdclose(struct file *fp, struct filedesc *fdp, int fd)
1724 {
1725         struct kqueue *kq;
1726         struct knote *kn;
1727         struct knote *kntmp;
1728
1729         lwkt_getpooltoken(&fp->f_klist);
1730 restart:
1731         SLIST_FOREACH(kn, &fp->f_klist, kn_link) {
1732                 if (kn->kn_kq->kq_fdp == fdp && kn->kn_id == fd) {
1733                         kq = kn->kn_kq;
1734                         lwkt_getpooltoken(kq);
1735
1736                         /* temporary verification hack */
1737                         SLIST_FOREACH(kntmp, &fp->f_klist, kn_link) {
1738                                 if (kn == kntmp)
1739                                         break;
1740                         }
1741                         if (kn != kntmp || kn->kn_kq->kq_fdp != fdp ||
1742                             kn->kn_id != fd || kn->kn_kq != kq) {
1743                                 lwkt_relpooltoken(kq);
1744                                 goto restart;
1745                         }
1746                         if (knote_acquire(kn))
1747                                 knote_detach_and_drop(kn);
1748                         lwkt_relpooltoken(kq);
1749                         goto restart;
1750                 }
1751         }
1752         lwkt_relpooltoken(&fp->f_klist);
1753 }
1754
1755 /*
1756  * Low level attach function.
1757  *
1758  * The knote should already be marked for processing.
1759  * Caller must hold the related kq token.
1760  */
1761 static void
1762 knote_attach(struct knote *kn)
1763 {
1764         struct klist *list;
1765         struct kqueue *kq = kn->kn_kq;
1766
1767         if (kn->kn_fop->f_flags & FILTEROP_ISFD) {
1768                 KKASSERT(kn->kn_fp);
1769                 list = &kn->kn_fp->f_klist;
1770         } else {
1771                 if (kq->kq_knhashmask == 0)
1772                         kq->kq_knhash = hashinit(KN_HASHSIZE, M_KQUEUE,
1773                                                  &kq->kq_knhashmask);
1774                 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
1775         }
1776         lwkt_getpooltoken(list);
1777         SLIST_INSERT_HEAD(list, kn, kn_link);
1778         lwkt_relpooltoken(list);
1779         TAILQ_INSERT_HEAD(&kq->kq_knlist, kn, kn_kqlink);
1780 }
1781
1782 /*
1783  * Low level drop function.
1784  *
1785  * The knote should already be marked for processing.
1786  * Caller must hold the related kq token.
1787  */
1788 static void
1789 knote_drop(struct knote *kn)
1790 {
1791         struct kqueue *kq;
1792         struct klist *list;
1793
1794         kq = kn->kn_kq;
1795
1796         if (kn->kn_fop->f_flags & FILTEROP_ISFD)
1797                 list = &kn->kn_fp->f_klist;
1798         else
1799                 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)];
1800
1801         lwkt_getpooltoken(list);
1802         SLIST_REMOVE(list, kn, knote, kn_link);
1803         lwkt_relpooltoken(list);
1804         TAILQ_REMOVE(&kq->kq_knlist, kn, kn_kqlink);
1805         if (kn->kn_status & KN_QUEUED)
1806                 knote_dequeue(kn);
1807         if (kn->kn_fop->f_flags & FILTEROP_ISFD) {
1808                 fdrop(kn->kn_fp);
1809                 kn->kn_fp = NULL;
1810         }
1811         knote_free(kn);
1812 }
1813
1814 /*
1815  * Low level enqueue function.
1816  *
1817  * The knote should already be marked for processing.
1818  * Caller must be holding the kq token
1819  */
1820 static void
1821 knote_enqueue(struct knote *kn)
1822 {
1823         struct kqueue *kq = kn->kn_kq;
1824
1825         KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued"));
1826         TAILQ_INSERT_TAIL(&kq->kq_knpend, kn, kn_tqe);
1827         kn->kn_status |= KN_QUEUED;
1828         ++kq->kq_count;
1829
1830         /*
1831          * Send SIGIO on request (typically set up as a mailbox signal)
1832          */
1833         if (kq->kq_sigio && (kq->kq_state & KQ_ASYNC) && kq->kq_count == 1)
1834                 pgsigio(kq->kq_sigio, SIGIO, 0);
1835
1836         kqueue_wakeup(kq);
1837 }
1838
1839 /*
1840  * Low level dequeue function.
1841  *
1842  * The knote should already be marked for processing.
1843  * Caller must be holding the kq token
1844  */
1845 static void
1846 knote_dequeue(struct knote *kn)
1847 {
1848         struct kqueue *kq = kn->kn_kq;
1849
1850         KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued"));
1851         TAILQ_REMOVE(&kq->kq_knpend, kn, kn_tqe);
1852         kn->kn_status &= ~KN_QUEUED;
1853         kq->kq_count--;
1854 }
1855
1856 static struct knote *
1857 knote_alloc(void)
1858 {
1859         return kmalloc(sizeof(struct knote), M_KQUEUE, M_WAITOK);
1860 }
1861
1862 static void
1863 knote_free(struct knote *kn)
1864 {
1865         struct knote_cache_list *cache_list;
1866
1867         cache_list = &knote_cache_lists[mycpuid];
1868         if (cache_list->knote_cache_cnt < KNOTE_CACHE_MAX) {
1869                 crit_enter();
1870                 SLIST_INSERT_HEAD(&cache_list->knote_cache, kn, kn_link);
1871                 cache_list->knote_cache_cnt++;
1872                 crit_exit();
1873                 return;
1874         }
1875         kfree(kn, M_KQUEUE);
1876 }
1877
1878 struct sleepinfo {
1879         void *ident;
1880         int timedout;
1881 };
1882
1883 static void
1884 precise_sleep_intr(systimer_t info, int in_ipi, struct intrframe *frame)
1885 {
1886         struct sleepinfo *si;
1887
1888         si = info->data;
1889         si->timedout = 1;
1890         wakeup(si->ident);
1891 }
1892
1893 static int
1894 precise_sleep(void *ident, int flags, const char *wmesg, int us)
1895 {
1896         struct systimer info;
1897         struct sleepinfo si = {
1898                 .ident = ident,
1899                 .timedout = 0,
1900         };
1901         int r;
1902
1903         tsleep_interlock(ident, flags);
1904         systimer_init_oneshot(&info, precise_sleep_intr, &si,
1905             us == 0 ? 1 : us);
1906         r = tsleep(ident, flags | PINTERLOCKED, wmesg, 0);
1907         systimer_del(&info);
1908         if (si.timedout)
1909                 r = EWOULDBLOCK;
1910
1911         return r;
1912 }