1 /*- 2 * Copyright (c) 1999,2000,2001 Jonathan Lemon <jlemon@FreeBSD.org> 3 * Copyright 2004 John-Mark Gurney <jmg@FreeBSD.org> 4 * All rights reserved. 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 AUTHOR 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 AUTHOR 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 28 #include <sys/cdefs.h> 29 __FBSDID("$FreeBSD$"); 30 31 #include <sys/param.h> 32 #include <sys/systm.h> 33 #include <sys/kernel.h> 34 #include <sys/lock.h> 35 #include <sys/mutex.h> 36 #include <sys/proc.h> 37 #include <sys/malloc.h> 38 #include <sys/unistd.h> 39 #include <sys/file.h> 40 #include <sys/filedesc.h> 41 #include <sys/filio.h> 42 #include <sys/fcntl.h> 43 #include <sys/kthread.h> 44 #include <sys/selinfo.h> 45 #include <sys/queue.h> 46 #include <sys/event.h> 47 #include <sys/eventvar.h> 48 #include <sys/poll.h> 49 #include <sys/protosw.h> 50 #include <sys/sigio.h> 51 #include <sys/signalvar.h> 52 #include <sys/socket.h> 53 #include <sys/socketvar.h> 54 #include <sys/stat.h> 55 #include <sys/sysctl.h> 56 #include <sys/sysproto.h> 57 #include <sys/syscallsubr.h> 58 #include <sys/taskqueue.h> 59 #include <sys/uio.h> 60 61 #include <vm/uma.h> 62 63 static MALLOC_DEFINE(M_KQUEUE, "kqueue", "memory for kqueue system"); 64 65 /* 66 * This lock is used if multiple kq locks are required. This possibly 67 * should be made into a per proc lock. 68 */ 69 static struct mtx kq_global; 70 MTX_SYSINIT(kq_global, &kq_global, "kqueue order", MTX_DEF); 71 #define KQ_GLOBAL_LOCK(lck, haslck) do { \ 72 if (!haslck) \ 73 mtx_lock(lck); \ 74 haslck = 1; \ 75 } while (0) 76 #define KQ_GLOBAL_UNLOCK(lck, haslck) do { \ 77 if (haslck) \ 78 mtx_unlock(lck); \ 79 haslck = 0; \ 80 } while (0) 81 82 TASKQUEUE_DEFINE_THREAD(kqueue); 83 84 static int kevent_copyout(struct kevent **eventlist, enum uio_seg eventseg, 85 struct kevent *kevp, int count); 86 static int kqueue_aquire(struct file *fp, struct kqueue **kqp); 87 static void kqueue_release(struct kqueue *kq, int locked); 88 static int kqueue_expand(struct kqueue *kq, struct filterops *fops, 89 uintptr_t ident, int waitok); 90 static void kqueue_task(void *arg, int pending); 91 static int kqueue_scan(struct kqueue *kq, int maxevents, 92 struct kevent *eventlist, enum uio_seg eventseg, 93 const struct timespec *timeout, struct kevent *keva, 94 struct thread *td); 95 static void kqueue_wakeup(struct kqueue *kq); 96 static struct filterops *kqueue_fo_find(int filt); 97 static void kqueue_fo_release(int filt); 98 99 static fo_rdwr_t kqueue_read; 100 static fo_rdwr_t kqueue_write; 101 static fo_ioctl_t kqueue_ioctl; 102 static fo_poll_t kqueue_poll; 103 static fo_kqfilter_t kqueue_kqfilter; 104 static fo_stat_t kqueue_stat; 105 static fo_close_t kqueue_close; 106 107 static struct fileops kqueueops = { 108 .fo_read = kqueue_read, 109 .fo_write = kqueue_write, 110 .fo_ioctl = kqueue_ioctl, 111 .fo_poll = kqueue_poll, 112 .fo_kqfilter = kqueue_kqfilter, 113 .fo_stat = kqueue_stat, 114 .fo_close = kqueue_close, 115 }; 116 117 static int knote_attach(struct knote *kn, struct kqueue *kq); 118 static void knote_drop(struct knote *kn, struct thread *td); 119 static void knote_enqueue(struct knote *kn); 120 static void knote_dequeue(struct knote *kn); 121 static void knote_init(void); 122 static struct knote *knote_alloc(int waitok); 123 static void knote_free(struct knote *kn); 124 125 static void filt_kqdetach(struct knote *kn); 126 static int filt_kqueue(struct knote *kn, long hint); 127 static int filt_procattach(struct knote *kn); 128 static void filt_procdetach(struct knote *kn); 129 static int filt_proc(struct knote *kn, long hint); 130 static int filt_fileattach(struct knote *kn); 131 static void filt_timerexpire(void *knx); 132 static int filt_timerattach(struct knote *kn); 133 static void filt_timerdetach(struct knote *kn); 134 static int filt_timer(struct knote *kn, long hint); 135 136 static struct filterops file_filtops = 137 { 1, filt_fileattach, NULL, NULL }; 138 static struct filterops kqread_filtops = 139 { 1, NULL, filt_kqdetach, filt_kqueue }; 140 /* XXX - move to kern_proc.c? */ 141 static struct filterops proc_filtops = 142 { 0, filt_procattach, filt_procdetach, filt_proc }; 143 static struct filterops timer_filtops = 144 { 0, filt_timerattach, filt_timerdetach, filt_timer }; 145 146 static uma_zone_t knote_zone; 147 static int kq_ncallouts = 0; 148 static int kq_calloutmax = (4 * 1024); 149 SYSCTL_INT(_kern, OID_AUTO, kq_calloutmax, CTLFLAG_RW, 150 &kq_calloutmax, 0, "Maximum number of callouts allocated for kqueue"); 151 152 /* XXX - ensure not KN_INFLUX?? */ 153 #define KNOTE_ACTIVATE(kn, islock) do { \ 154 if ((islock)) \ 155 mtx_assert(&(kn)->kn_kq->kq_lock, MA_OWNED); \ 156 else \ 157 KQ_LOCK((kn)->kn_kq); \ 158 (kn)->kn_status |= KN_ACTIVE; \ 159 if (((kn)->kn_status & (KN_QUEUED | KN_DISABLED)) == 0) \ 160 knote_enqueue((kn)); \ 161 if (!(islock)) \ 162 KQ_UNLOCK((kn)->kn_kq); \ 163 } while(0) 164 #define KQ_LOCK(kq) do { \ 165 mtx_lock(&(kq)->kq_lock); \ 166 } while (0) 167 #define KQ_FLUX_WAKEUP(kq) do { \ 168 if (((kq)->kq_state & KQ_FLUXWAIT) == KQ_FLUXWAIT) { \ 169 (kq)->kq_state &= ~KQ_FLUXWAIT; \ 170 wakeup((kq)); \ 171 } \ 172 } while (0) 173 #define KQ_UNLOCK_FLUX(kq) do { \ 174 KQ_FLUX_WAKEUP(kq); \ 175 mtx_unlock(&(kq)->kq_lock); \ 176 } while (0) 177 #define KQ_UNLOCK(kq) do { \ 178 mtx_unlock(&(kq)->kq_lock); \ 179 } while (0) 180 #define KQ_OWNED(kq) do { \ 181 mtx_assert(&(kq)->kq_lock, MA_OWNED); \ 182 } while (0) 183 #define KQ_NOTOWNED(kq) do { \ 184 mtx_assert(&(kq)->kq_lock, MA_NOTOWNED); \ 185 } while (0) 186 #define KN_LIST_LOCK(kn) do { \ 187 if (kn->kn_knlist != NULL) \ 188 mtx_lock(kn->kn_knlist->kl_lock); \ 189 } while (0) 190 #define KN_LIST_UNLOCK(kn) do { \ 191 if (kn->kn_knlist != NULL) \ 192 mtx_unlock(kn->kn_knlist->kl_lock); \ 193 } while (0) 194 195 #define KN_HASHSIZE 64 /* XXX should be tunable */ 196 #define KN_HASH(val, mask) (((val) ^ (val >> 8)) & (mask)) 197 198 static int 199 filt_nullattach(struct knote *kn) 200 { 201 202 return (ENXIO); 203 }; 204 205 struct filterops null_filtops = 206 { 0, filt_nullattach, NULL, NULL }; 207 208 /* XXX - make SYSINIT to add these, and move into respective modules. */ 209 extern struct filterops sig_filtops; 210 extern struct filterops fs_filtops; 211 212 /* 213 * Table for for all system-defined filters. 214 */ 215 static struct mtx filterops_lock; 216 MTX_SYSINIT(kqueue_filterops, &filterops_lock, "protect sysfilt_ops", 217 MTX_DEF); 218 static struct { 219 struct filterops *for_fop; 220 int for_refcnt; 221 } sysfilt_ops[EVFILT_SYSCOUNT] = { 222 { &file_filtops }, /* EVFILT_READ */ 223 { &file_filtops }, /* EVFILT_WRITE */ 224 { &null_filtops }, /* EVFILT_AIO */ 225 { &file_filtops }, /* EVFILT_VNODE */ 226 { &proc_filtops }, /* EVFILT_PROC */ 227 { &sig_filtops }, /* EVFILT_SIGNAL */ 228 { &timer_filtops }, /* EVFILT_TIMER */ 229 { &file_filtops }, /* EVFILT_NETDEV */ 230 { &fs_filtops }, /* EVFILT_FS */ 231 }; 232 233 /* 234 * Simple redirection for all cdevsw style objects to call their fo_kqfilter 235 * method. 236 */ 237 static int 238 filt_fileattach(struct knote *kn) 239 { 240 241 return (fo_kqfilter(kn->kn_fp, kn)); 242 } 243 244 /*ARGSUSED*/ 245 static int 246 kqueue_kqfilter(struct file *fp, struct knote *kn) 247 { 248 struct kqueue *kq = kn->kn_fp->f_data; 249 250 if (kn->kn_filter != EVFILT_READ) 251 return (EINVAL); 252 253 kn->kn_status |= KN_KQUEUE; 254 kn->kn_fop = &kqread_filtops; 255 knlist_add(&kq->kq_sel.si_note, kn, 0); 256 257 return (0); 258 } 259 260 static void 261 filt_kqdetach(struct knote *kn) 262 { 263 struct kqueue *kq = kn->kn_fp->f_data; 264 265 knlist_remove(&kq->kq_sel.si_note, kn, 0); 266 } 267 268 /*ARGSUSED*/ 269 static int 270 filt_kqueue(struct knote *kn, long hint) 271 { 272 struct kqueue *kq = kn->kn_fp->f_data; 273 274 kn->kn_data = kq->kq_count; 275 return (kn->kn_data > 0); 276 } 277 278 /* XXX - move to kern_proc.c? */ 279 static int 280 filt_procattach(struct knote *kn) 281 { 282 struct proc *p; 283 int immediate; 284 int error; 285 286 immediate = 0; 287 p = pfind(kn->kn_id); 288 if (p == NULL && (kn->kn_sfflags & NOTE_EXIT)) { 289 p = zpfind(kn->kn_id); 290 immediate = 1; 291 } else if (p != NULL && (p->p_flag & P_WEXIT)) { 292 immediate = 1; 293 } 294 295 if (p == NULL) 296 return (ESRCH); 297 if ((error = p_cansee(curthread, p))) 298 return (error); 299 300 kn->kn_ptr.p_proc = p; 301 kn->kn_flags |= EV_CLEAR; /* automatically set */ 302 303 /* 304 * internal flag indicating registration done by kernel 305 */ 306 if (kn->kn_flags & EV_FLAG1) { 307 kn->kn_data = kn->kn_sdata; /* ppid */ 308 kn->kn_fflags = NOTE_CHILD; 309 kn->kn_flags &= ~EV_FLAG1; 310 } 311 312 if (immediate == 0) 313 knlist_add(&p->p_klist, kn, 1); 314 315 /* 316 * Immediately activate any exit notes if the target process is a 317 * zombie. This is necessary to handle the case where the target 318 * process, e.g. a child, dies before the kevent is registered. 319 */ 320 if (immediate && filt_proc(kn, NOTE_EXIT)) 321 KNOTE_ACTIVATE(kn, 0); 322 323 PROC_UNLOCK(p); 324 325 return (0); 326 } 327 328 /* 329 * The knote may be attached to a different process, which may exit, 330 * leaving nothing for the knote to be attached to. So when the process 331 * exits, the knote is marked as DETACHED and also flagged as ONESHOT so 332 * it will be deleted when read out. However, as part of the knote deletion, 333 * this routine is called, so a check is needed to avoid actually performing 334 * a detach, because the original process does not exist any more. 335 */ 336 /* XXX - move to kern_proc.c? */ 337 static void 338 filt_procdetach(struct knote *kn) 339 { 340 struct proc *p; 341 342 p = kn->kn_ptr.p_proc; 343 knlist_remove(&p->p_klist, kn, 0); 344 kn->kn_ptr.p_proc = NULL; 345 } 346 347 /* XXX - move to kern_proc.c? */ 348 static int 349 filt_proc(struct knote *kn, long hint) 350 { 351 struct proc *p = kn->kn_ptr.p_proc; 352 u_int event; 353 354 /* 355 * mask off extra data 356 */ 357 event = (u_int)hint & NOTE_PCTRLMASK; 358 359 /* 360 * if the user is interested in this event, record it. 361 */ 362 if (kn->kn_sfflags & event) 363 kn->kn_fflags |= event; 364 365 /* 366 * process is gone, so flag the event as finished. 367 */ 368 if (event == NOTE_EXIT) { 369 if (!(kn->kn_status & KN_DETACHED)) 370 knlist_remove_inevent(&p->p_klist, kn); 371 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 372 kn->kn_ptr.p_proc = NULL; 373 return (1); 374 } 375 376 /* 377 * process forked, and user wants to track the new process, 378 * so attach a new knote to it, and immediately report an 379 * event with the parent's pid. 380 */ 381 if ((event == NOTE_FORK) && (kn->kn_sfflags & NOTE_TRACK)) { 382 struct kevent kev; 383 int error; 384 385 /* 386 * register knote with new process. 387 */ 388 kev.ident = hint & NOTE_PDATAMASK; /* pid */ 389 kev.filter = kn->kn_filter; 390 kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1; 391 kev.fflags = kn->kn_sfflags; 392 kev.data = kn->kn_id; /* parent */ 393 kev.udata = kn->kn_kevent.udata; /* preserve udata */ 394 error = kqueue_register(kn->kn_kq, &kev, NULL, 0); 395 if (error) 396 kn->kn_fflags |= NOTE_TRACKERR; 397 } 398 399 return (kn->kn_fflags != 0); 400 } 401 402 static int 403 timertoticks(intptr_t data) 404 { 405 struct timeval tv; 406 int tticks; 407 408 tv.tv_sec = data / 1000; 409 tv.tv_usec = (data % 1000) * 1000; 410 tticks = tvtohz(&tv); 411 412 return tticks; 413 } 414 415 /* XXX - move to kern_timeout.c? */ 416 static void 417 filt_timerexpire(void *knx) 418 { 419 struct knote *kn = knx; 420 struct callout *calloutp; 421 422 kn->kn_data++; 423 KNOTE_ACTIVATE(kn, 0); /* XXX - handle locking */ 424 425 if ((kn->kn_flags & EV_ONESHOT) != EV_ONESHOT) { 426 calloutp = (struct callout *)kn->kn_hook; 427 callout_reset(calloutp, timertoticks(kn->kn_sdata), 428 filt_timerexpire, kn); 429 } 430 } 431 432 /* 433 * data contains amount of time to sleep, in milliseconds 434 */ 435 /* XXX - move to kern_timeout.c? */ 436 static int 437 filt_timerattach(struct knote *kn) 438 { 439 struct callout *calloutp; 440 441 atomic_add_int(&kq_ncallouts, 1); 442 443 if (kq_ncallouts >= kq_calloutmax) { 444 atomic_add_int(&kq_ncallouts, -1); 445 return (ENOMEM); 446 } 447 448 kn->kn_flags |= EV_CLEAR; /* automatically set */ 449 kn->kn_status &= ~KN_DETACHED; /* knlist_add usually sets it */ 450 MALLOC(calloutp, struct callout *, sizeof(*calloutp), 451 M_KQUEUE, M_WAITOK); 452 callout_init(calloutp, CALLOUT_MPSAFE); 453 kn->kn_hook = calloutp; 454 callout_reset(calloutp, timertoticks(kn->kn_sdata), filt_timerexpire, 455 kn); 456 457 return (0); 458 } 459 460 /* XXX - move to kern_timeout.c? */ 461 static void 462 filt_timerdetach(struct knote *kn) 463 { 464 struct callout *calloutp; 465 466 calloutp = (struct callout *)kn->kn_hook; 467 callout_drain(calloutp); 468 FREE(calloutp, M_KQUEUE); 469 atomic_add_int(&kq_ncallouts, -1); 470 kn->kn_status |= KN_DETACHED; /* knlist_remove usually clears it */ 471 } 472 473 /* XXX - move to kern_timeout.c? */ 474 static int 475 filt_timer(struct knote *kn, long hint) 476 { 477 478 return (kn->kn_data != 0); 479 } 480 481 /* 482 * MPSAFE 483 */ 484 int 485 kqueue(struct thread *td, struct kqueue_args *uap) 486 { 487 struct filedesc *fdp; 488 struct kqueue *kq; 489 struct file *fp; 490 int fd, error; 491 492 fdp = td->td_proc->p_fd; 493 error = falloc(td, &fp, &fd); 494 if (error) 495 goto done2; 496 497 /* An extra reference on `nfp' has been held for us by falloc(). */ 498 kq = malloc(sizeof *kq, M_KQUEUE, M_WAITOK | M_ZERO); 499 mtx_init(&kq->kq_lock, "kqueue", NULL, MTX_DEF|MTX_DUPOK); 500 TAILQ_INIT(&kq->kq_head); 501 kq->kq_fdp = fdp; 502 knlist_init(&kq->kq_sel.si_note, &kq->kq_lock); 503 TASK_INIT(&kq->kq_task, 0, kqueue_task, kq); 504 505 FILEDESC_LOCK_FAST(fdp); 506 SLIST_INSERT_HEAD(&fdp->fd_kqlist, kq, kq_list); 507 FILEDESC_UNLOCK_FAST(fdp); 508 509 FILE_LOCK(fp); 510 fp->f_flag = FREAD | FWRITE; 511 fp->f_type = DTYPE_KQUEUE; 512 fp->f_ops = &kqueueops; 513 fp->f_data = kq; 514 FILE_UNLOCK(fp); 515 fdrop(fp, td); 516 517 td->td_retval[0] = fd; 518 done2: 519 return (error); 520 } 521 522 #ifndef _SYS_SYSPROTO_H_ 523 struct kevent_args { 524 int fd; 525 const struct kevent *changelist; 526 int nchanges; 527 struct kevent *eventlist; 528 int nevents; 529 const struct timespec *timeout; 530 }; 531 #endif 532 /* 533 * MPSAFE 534 */ 535 int 536 kevent(struct thread *td, struct kevent_args *uap) 537 { 538 struct timespec ts, *tsp; 539 int error; 540 541 if (uap->timeout != NULL) { 542 error = copyin(uap->timeout, &ts, sizeof(ts)); 543 if (error) 544 return (error); 545 tsp = &ts; 546 } else 547 tsp = NULL; 548 549 return (kern_kevent(td, uap->fd, uap->changelist, uap->nchanges, 550 UIO_USERSPACE, uap->eventlist, uap->nevents, UIO_USERSPACE, tsp)); 551 } 552 553 /* 554 * Copy 'count' items into the destination list pointd to by *eventlist. The 555 * eventlist and nevents values are updated to point after the copied out 556 * item(s) upon return. 557 */ 558 static int 559 kevent_copyout(struct kevent **eventlist, enum uio_seg eventseg, 560 struct kevent *kevp, int count) 561 { 562 int error; 563 564 if (eventseg == UIO_USERSPACE) 565 error = copyout(kevp, *eventlist, 566 sizeof(struct kevent) * count); 567 else { 568 bcopy(kevp, *eventlist, sizeof(struct kevent) * count); 569 error = 0; 570 } 571 *eventlist += count; 572 return (error); 573 } 574 575 int 576 kern_kevent(struct thread *td, int fd, struct kevent *changelist, int nchanges, 577 enum uio_seg changeseg, struct kevent *eventlist, int nevents, 578 enum uio_seg eventseg, const struct timespec *timeout) 579 { 580 struct kevent keva[KQ_NEVENTS]; 581 struct kevent *kevp, *changes; 582 struct kqueue *kq; 583 struct file *fp; 584 int i, n, nerrors, error; 585 586 if ((error = fget(td, fd, &fp)) != 0) 587 return (error); 588 if ((error = kqueue_aquire(fp, &kq)) != 0) 589 goto done_norel; 590 591 nerrors = 0; 592 593 while (nchanges > 0) { 594 if (changeseg == UIO_USERSPACE) { 595 n = nchanges > KQ_NEVENTS ? KQ_NEVENTS : nchanges; 596 error = copyin(changelist, keva, n * sizeof *keva); 597 if (error) 598 goto done; 599 changes = keva; 600 } else { 601 changes = changelist; 602 n = nchanges; 603 } 604 for (i = 0; i < n; i++) { 605 kevp = &changes[i]; 606 kevp->flags &= ~EV_SYSFLAGS; 607 error = kqueue_register(kq, kevp, td, 1); 608 if (error) { 609 if (nevents != 0) { 610 kevp->flags = EV_ERROR; 611 kevp->data = error; 612 (void) kevent_copyout(&eventlist, 613 eventseg, kevp, 1); 614 nevents--; 615 nerrors++; 616 } else { 617 goto done; 618 } 619 } 620 } 621 nchanges -= n; 622 changelist += n; 623 } 624 if (nerrors) { 625 td->td_retval[0] = nerrors; 626 error = 0; 627 goto done; 628 } 629 630 error = kqueue_scan(kq, nevents, eventlist, eventseg, timeout, 631 keva, td); 632 done: 633 kqueue_release(kq, 0); 634 done_norel: 635 if (fp != NULL) 636 fdrop(fp, td); 637 return (error); 638 } 639 640 int 641 kqueue_add_filteropts(int filt, struct filterops *filtops) 642 { 643 int error; 644 645 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) { 646 printf( 647 "trying to add a filterop that is out of range: %d is beyond %d\n", 648 ~filt, EVFILT_SYSCOUNT); 649 return EINVAL; 650 } 651 mtx_lock(&filterops_lock); 652 if (sysfilt_ops[~filt].for_fop != &null_filtops && 653 sysfilt_ops[~filt].for_fop != NULL) 654 error = EEXIST; 655 else { 656 sysfilt_ops[~filt].for_fop = filtops; 657 sysfilt_ops[~filt].for_refcnt = 0; 658 } 659 mtx_unlock(&filterops_lock); 660 661 return (0); 662 } 663 664 int 665 kqueue_del_filteropts(int filt) 666 { 667 int error; 668 669 error = 0; 670 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) 671 return EINVAL; 672 673 mtx_lock(&filterops_lock); 674 if (sysfilt_ops[~filt].for_fop == &null_filtops || 675 sysfilt_ops[~filt].for_fop == NULL) 676 error = EINVAL; 677 else if (sysfilt_ops[~filt].for_refcnt != 0) 678 error = EBUSY; 679 else { 680 sysfilt_ops[~filt].for_fop = &null_filtops; 681 sysfilt_ops[~filt].for_refcnt = 0; 682 } 683 mtx_unlock(&filterops_lock); 684 685 return error; 686 } 687 688 static struct filterops * 689 kqueue_fo_find(int filt) 690 { 691 692 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) 693 return NULL; 694 695 mtx_lock(&filterops_lock); 696 sysfilt_ops[~filt].for_refcnt++; 697 if (sysfilt_ops[~filt].for_fop == NULL) 698 sysfilt_ops[~filt].for_fop = &null_filtops; 699 mtx_unlock(&filterops_lock); 700 701 return sysfilt_ops[~filt].for_fop; 702 } 703 704 static void 705 kqueue_fo_release(int filt) 706 { 707 708 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) 709 return; 710 711 mtx_lock(&filterops_lock); 712 KASSERT(sysfilt_ops[~filt].for_refcnt > 0, 713 ("filter object refcount not valid on release")); 714 sysfilt_ops[~filt].for_refcnt--; 715 mtx_unlock(&filterops_lock); 716 } 717 718 /* 719 * A ref to kq (obtained via kqueue_aquire) should be held. waitok will 720 * influence if memory allocation should wait. Make sure it is 0 if you 721 * hold any mutexes. 722 */ 723 int 724 kqueue_register(struct kqueue *kq, struct kevent *kev, struct thread *td, int waitok) 725 { 726 struct filedesc *fdp; 727 struct filterops *fops; 728 struct file *fp; 729 struct knote *kn, *tkn; 730 int error, filt, event; 731 int haskqglobal; 732 int fd; 733 734 fdp = NULL; 735 fp = NULL; 736 kn = NULL; 737 error = 0; 738 haskqglobal = 0; 739 740 filt = kev->filter; 741 fops = kqueue_fo_find(filt); 742 if (fops == NULL) 743 return EINVAL; 744 745 tkn = knote_alloc(waitok); /* prevent waiting with locks */ 746 747 findkn: 748 if (fops->f_isfd) { 749 KASSERT(td != NULL, ("td is NULL")); 750 fdp = td->td_proc->p_fd; 751 FILEDESC_LOCK(fdp); 752 /* validate descriptor */ 753 fd = kev->ident; 754 if (fd < 0 || fd >= fdp->fd_nfiles || 755 (fp = fdp->fd_ofiles[fd]) == NULL) { 756 FILEDESC_UNLOCK(fdp); 757 error = EBADF; 758 goto done; 759 } 760 fhold(fp); 761 762 if ((kev->flags & EV_ADD) == EV_ADD && kqueue_expand(kq, fops, 763 kev->ident, 0) != 0) { 764 /* unlock and try again */ 765 FILEDESC_UNLOCK(fdp); 766 fdrop(fp, td); 767 fp = NULL; 768 error = kqueue_expand(kq, fops, kev->ident, waitok); 769 if (error) 770 goto done; 771 goto findkn; 772 } 773 774 if (fp->f_type == DTYPE_KQUEUE) { 775 /* 776 * if we add some inteligence about what we are doing, 777 * we should be able to support events on ourselves. 778 * We need to know when we are doing this to prevent 779 * getting both the knlist lock and the kq lock since 780 * they are the same thing. 781 */ 782 if (fp->f_data == kq) { 783 FILEDESC_UNLOCK(fdp); 784 error = EINVAL; 785 goto done_noglobal; 786 } 787 788 KQ_GLOBAL_LOCK(&kq_global, haskqglobal); 789 } 790 791 FILEDESC_UNLOCK(fdp); 792 KQ_LOCK(kq); 793 if (kev->ident < kq->kq_knlistsize) { 794 SLIST_FOREACH(kn, &kq->kq_knlist[kev->ident], kn_link) 795 if (kev->filter == kn->kn_filter) 796 break; 797 } 798 } else { 799 if ((kev->flags & EV_ADD) == EV_ADD) 800 kqueue_expand(kq, fops, kev->ident, waitok); 801 802 KQ_LOCK(kq); 803 if (kq->kq_knhashmask != 0) { 804 struct klist *list; 805 806 list = &kq->kq_knhash[ 807 KN_HASH((u_long)kev->ident, kq->kq_knhashmask)]; 808 SLIST_FOREACH(kn, list, kn_link) 809 if (kev->ident == kn->kn_id && 810 kev->filter == kn->kn_filter) 811 break; 812 } 813 } 814 815 /* knote is in the process of changing, wait for it to stablize. */ 816 if (kn != NULL && (kn->kn_status & KN_INFLUX) == KN_INFLUX) { 817 if (fp != NULL) { 818 fdrop(fp, td); 819 fp = NULL; 820 } 821 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal); 822 kq->kq_state |= KQ_FLUXWAIT; 823 msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqflxwt", 0); 824 goto findkn; 825 } 826 827 if (kn == NULL && ((kev->flags & EV_ADD) == 0)) { 828 KQ_UNLOCK(kq); 829 error = ENOENT; 830 goto done; 831 } 832 833 /* 834 * kn now contains the matching knote, or NULL if no match 835 */ 836 if (kev->flags & EV_ADD) { 837 if (kn == NULL) { 838 kn = tkn; 839 tkn = NULL; 840 if (kn == NULL) { 841 error = ENOMEM; 842 goto done; 843 } 844 kn->kn_fp = fp; 845 kn->kn_kq = kq; 846 kn->kn_fop = fops; 847 /* 848 * apply reference counts to knote structure, and 849 * do not release it at the end of this routine. 850 */ 851 fops = NULL; 852 fp = NULL; 853 854 kn->kn_sfflags = kev->fflags; 855 kn->kn_sdata = kev->data; 856 kev->fflags = 0; 857 kev->data = 0; 858 kn->kn_kevent = *kev; 859 kn->kn_status = KN_INFLUX|KN_DETACHED; 860 861 error = knote_attach(kn, kq); 862 KQ_UNLOCK(kq); 863 if (error != 0) { 864 tkn = kn; 865 goto done; 866 } 867 868 if ((error = kn->kn_fop->f_attach(kn)) != 0) { 869 knote_drop(kn, td); 870 goto done; 871 } 872 KN_LIST_LOCK(kn); 873 } else { 874 /* 875 * The user may change some filter values after the 876 * initial EV_ADD, but doing so will not reset any 877 * filter which has already been triggered. 878 */ 879 kn->kn_status |= KN_INFLUX; 880 KQ_UNLOCK(kq); 881 KN_LIST_LOCK(kn); 882 kn->kn_sfflags = kev->fflags; 883 kn->kn_sdata = kev->data; 884 kn->kn_kevent.udata = kev->udata; 885 } 886 887 /* 888 * We can get here with kn->kn_knlist == NULL. 889 * This can happen when the initial attach event decides that 890 * the event is "completed" already. i.e. filt_procattach 891 * is called on a zombie process. It will call filt_proc 892 * which will remove it from the list, and NULL kn_knlist. 893 */ 894 event = kn->kn_fop->f_event(kn, 0); 895 KN_LIST_UNLOCK(kn); 896 KQ_LOCK(kq); 897 if (event) 898 KNOTE_ACTIVATE(kn, 1); 899 kn->kn_status &= ~KN_INFLUX; 900 } else if (kev->flags & EV_DELETE) { 901 kn->kn_status |= KN_INFLUX; 902 KQ_UNLOCK(kq); 903 if (!(kn->kn_status & KN_DETACHED)) 904 kn->kn_fop->f_detach(kn); 905 knote_drop(kn, td); 906 goto done; 907 } 908 909 if ((kev->flags & EV_DISABLE) && 910 ((kn->kn_status & KN_DISABLED) == 0)) { 911 kn->kn_status |= KN_DISABLED; 912 } 913 914 if ((kev->flags & EV_ENABLE) && (kn->kn_status & KN_DISABLED)) { 915 kn->kn_status &= ~KN_DISABLED; 916 if ((kn->kn_status & KN_ACTIVE) && 917 ((kn->kn_status & KN_QUEUED) == 0)) 918 knote_enqueue(kn); 919 } 920 KQ_UNLOCK_FLUX(kq); 921 922 done: 923 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal); 924 done_noglobal: 925 if (fp != NULL) 926 fdrop(fp, td); 927 if (tkn != NULL) 928 knote_free(tkn); 929 if (fops != NULL) 930 kqueue_fo_release(filt); 931 return (error); 932 } 933 934 static int 935 kqueue_aquire(struct file *fp, struct kqueue **kqp) 936 { 937 int error; 938 struct kqueue *kq; 939 940 error = 0; 941 942 FILE_LOCK(fp); 943 do { 944 kq = fp->f_data; 945 if (fp->f_type != DTYPE_KQUEUE || kq == NULL) { 946 error = EBADF; 947 break; 948 } 949 *kqp = kq; 950 KQ_LOCK(kq); 951 if ((kq->kq_state & KQ_CLOSING) == KQ_CLOSING) { 952 KQ_UNLOCK(kq); 953 error = EBADF; 954 break; 955 } 956 kq->kq_refcnt++; 957 KQ_UNLOCK(kq); 958 } while (0); 959 FILE_UNLOCK(fp); 960 961 return error; 962 } 963 964 static void 965 kqueue_release(struct kqueue *kq, int locked) 966 { 967 if (locked) 968 KQ_OWNED(kq); 969 else 970 KQ_LOCK(kq); 971 kq->kq_refcnt--; 972 if (kq->kq_refcnt == 1) 973 wakeup(&kq->kq_refcnt); 974 if (!locked) 975 KQ_UNLOCK(kq); 976 } 977 978 static void 979 kqueue_schedtask(struct kqueue *kq) 980 { 981 982 KQ_OWNED(kq); 983 KASSERT(((kq->kq_state & KQ_TASKDRAIN) != KQ_TASKDRAIN), 984 ("scheduling kqueue task while draining")); 985 986 if ((kq->kq_state & KQ_TASKSCHED) != KQ_TASKSCHED) { 987 taskqueue_enqueue(taskqueue_kqueue, &kq->kq_task); 988 kq->kq_state |= KQ_TASKSCHED; 989 } 990 } 991 992 /* 993 * Expand the kq to make sure we have storage for fops/ident pair. 994 * 995 * Return 0 on success (or no work necessary), return errno on failure. 996 * 997 * Not calling hashinit w/ waitok (proper malloc flag) should be safe. 998 * If kqueue_register is called from a non-fd context, there usually/should 999 * be no locks held. 1000 */ 1001 static int 1002 kqueue_expand(struct kqueue *kq, struct filterops *fops, uintptr_t ident, 1003 int waitok) 1004 { 1005 struct klist *list, *tmp_knhash; 1006 u_long tmp_knhashmask; 1007 int size; 1008 int fd; 1009 int mflag = waitok ? M_WAITOK : M_NOWAIT; 1010 1011 KQ_NOTOWNED(kq); 1012 1013 if (fops->f_isfd) { 1014 fd = ident; 1015 if (kq->kq_knlistsize <= fd) { 1016 size = kq->kq_knlistsize; 1017 while (size <= fd) 1018 size += KQEXTENT; 1019 MALLOC(list, struct klist *, 1020 size * sizeof list, M_KQUEUE, mflag); 1021 if (list == NULL) 1022 return ENOMEM; 1023 KQ_LOCK(kq); 1024 if (kq->kq_knlistsize > fd) { 1025 FREE(list, M_KQUEUE); 1026 list = NULL; 1027 } else { 1028 if (kq->kq_knlist != NULL) { 1029 bcopy(kq->kq_knlist, list, 1030 kq->kq_knlistsize * sizeof list); 1031 FREE(kq->kq_knlist, M_KQUEUE); 1032 kq->kq_knlist = NULL; 1033 } 1034 bzero((caddr_t)list + 1035 kq->kq_knlistsize * sizeof list, 1036 (size - kq->kq_knlistsize) * sizeof list); 1037 kq->kq_knlistsize = size; 1038 kq->kq_knlist = list; 1039 } 1040 KQ_UNLOCK(kq); 1041 } 1042 } else { 1043 if (kq->kq_knhashmask == 0) { 1044 tmp_knhash = hashinit(KN_HASHSIZE, M_KQUEUE, 1045 &tmp_knhashmask); 1046 if (tmp_knhash == NULL) 1047 return ENOMEM; 1048 KQ_LOCK(kq); 1049 if (kq->kq_knhashmask == 0) { 1050 kq->kq_knhash = tmp_knhash; 1051 kq->kq_knhashmask = tmp_knhashmask; 1052 } else { 1053 free(tmp_knhash, M_KQUEUE); 1054 } 1055 KQ_UNLOCK(kq); 1056 } 1057 } 1058 1059 KQ_NOTOWNED(kq); 1060 return 0; 1061 } 1062 1063 static void 1064 kqueue_task(void *arg, int pending) 1065 { 1066 struct kqueue *kq; 1067 int haskqglobal; 1068 1069 haskqglobal = 0; 1070 kq = arg; 1071 1072 KQ_GLOBAL_LOCK(&kq_global, haskqglobal); 1073 KQ_LOCK(kq); 1074 1075 KNOTE_LOCKED(&kq->kq_sel.si_note, 0); 1076 1077 kq->kq_state &= ~KQ_TASKSCHED; 1078 if ((kq->kq_state & KQ_TASKDRAIN) == KQ_TASKDRAIN) { 1079 wakeup(&kq->kq_state); 1080 } 1081 KQ_UNLOCK(kq); 1082 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal); 1083 } 1084 1085 /* 1086 * Scan, update kn_data (if not ONESHOT), and copyout triggered events. 1087 * We treat KN_MARKER knotes as if they are INFLUX. 1088 */ 1089 static int 1090 kqueue_scan(struct kqueue *kq, int maxevents, struct kevent *eventlist, 1091 enum uio_seg eventseg, const struct timespec *tsp, struct kevent *keva, 1092 struct thread *td) 1093 { 1094 struct kevent *kevp; 1095 struct timeval atv, rtv, ttv; 1096 struct knote *kn, *marker; 1097 int count, timeout, nkev, error; 1098 int haskqglobal; 1099 1100 count = maxevents; 1101 nkev = 0; 1102 error = 0; 1103 haskqglobal = 0; 1104 1105 if (maxevents == 0) 1106 goto done_nl; 1107 1108 if (tsp != NULL) { 1109 TIMESPEC_TO_TIMEVAL(&atv, tsp); 1110 if (itimerfix(&atv)) { 1111 error = EINVAL; 1112 goto done_nl; 1113 } 1114 if (tsp->tv_sec == 0 && tsp->tv_nsec == 0) 1115 timeout = -1; 1116 else 1117 timeout = atv.tv_sec > 24 * 60 * 60 ? 1118 24 * 60 * 60 * hz : tvtohz(&atv); 1119 getmicrouptime(&rtv); 1120 timevaladd(&atv, &rtv); 1121 } else { 1122 atv.tv_sec = 0; 1123 atv.tv_usec = 0; 1124 timeout = 0; 1125 } 1126 marker = knote_alloc(1); 1127 if (marker == NULL) { 1128 error = ENOMEM; 1129 goto done_nl; 1130 } 1131 marker->kn_status = KN_MARKER; 1132 KQ_LOCK(kq); 1133 goto start; 1134 1135 retry: 1136 if (atv.tv_sec || atv.tv_usec) { 1137 getmicrouptime(&rtv); 1138 if (timevalcmp(&rtv, &atv, >=)) 1139 goto done; 1140 ttv = atv; 1141 timevalsub(&ttv, &rtv); 1142 timeout = ttv.tv_sec > 24 * 60 * 60 ? 1143 24 * 60 * 60 * hz : tvtohz(&ttv); 1144 } 1145 1146 start: 1147 kevp = keva; 1148 if (kq->kq_count == 0) { 1149 if (timeout < 0) { 1150 error = EWOULDBLOCK; 1151 } else { 1152 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal); 1153 kq->kq_state |= KQ_SLEEP; 1154 error = msleep(kq, &kq->kq_lock, PSOCK | PCATCH, 1155 "kqread", timeout); 1156 } 1157 if (error == 0) 1158 goto retry; 1159 /* don't restart after signals... */ 1160 if (error == ERESTART) 1161 error = EINTR; 1162 else if (error == EWOULDBLOCK) 1163 error = 0; 1164 goto done; 1165 } 1166 1167 TAILQ_INSERT_TAIL(&kq->kq_head, marker, kn_tqe); 1168 while (count) { 1169 KQ_OWNED(kq); 1170 kn = TAILQ_FIRST(&kq->kq_head); 1171 1172 if ((kn->kn_status == KN_MARKER && kn != marker) || 1173 (kn->kn_status & KN_INFLUX) == KN_INFLUX) { 1174 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal); 1175 kq->kq_state |= KQ_FLUXWAIT; 1176 error = msleep(kq, &kq->kq_lock, PSOCK, 1177 "kqflxwt", 0); 1178 continue; 1179 } 1180 1181 TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe); 1182 if ((kn->kn_status & KN_DISABLED) == KN_DISABLED) { 1183 kn->kn_status &= ~KN_QUEUED; 1184 kq->kq_count--; 1185 continue; 1186 } 1187 if (kn == marker) { 1188 KQ_FLUX_WAKEUP(kq); 1189 if (count == maxevents) 1190 goto retry; 1191 goto done; 1192 } 1193 KASSERT((kn->kn_status & KN_INFLUX) == 0, 1194 ("KN_INFLUX set when not suppose to be")); 1195 1196 if ((kn->kn_flags & EV_ONESHOT) == EV_ONESHOT) { 1197 kn->kn_status &= ~KN_QUEUED; 1198 kn->kn_status |= KN_INFLUX; 1199 kq->kq_count--; 1200 KQ_UNLOCK(kq); 1201 /* 1202 * We don't need to lock the list since we've marked 1203 * it _INFLUX. 1204 */ 1205 *kevp = kn->kn_kevent; 1206 if (!(kn->kn_status & KN_DETACHED)) 1207 kn->kn_fop->f_detach(kn); 1208 knote_drop(kn, td); 1209 KQ_LOCK(kq); 1210 kn = NULL; 1211 } else { 1212 kn->kn_status |= KN_INFLUX; 1213 KQ_UNLOCK(kq); 1214 if ((kn->kn_status & KN_KQUEUE) == KN_KQUEUE) 1215 KQ_GLOBAL_LOCK(&kq_global, haskqglobal); 1216 KN_LIST_LOCK(kn); 1217 if (kn->kn_fop->f_event(kn, 0) == 0) { 1218 KN_LIST_UNLOCK(kn); 1219 KQ_LOCK(kq); 1220 kn->kn_status &= 1221 ~(KN_QUEUED | KN_ACTIVE | KN_INFLUX); 1222 kq->kq_count--; 1223 continue; 1224 } 1225 *kevp = kn->kn_kevent; 1226 KQ_LOCK(kq); 1227 if (kn->kn_flags & EV_CLEAR) { 1228 kn->kn_data = 0; 1229 kn->kn_fflags = 0; 1230 kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE); 1231 kq->kq_count--; 1232 } else 1233 TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe); 1234 KN_LIST_UNLOCK(kn); 1235 kn->kn_status &= ~(KN_INFLUX); 1236 } 1237 1238 /* we are returning a copy to the user */ 1239 kevp++; 1240 nkev++; 1241 count--; 1242 1243 if (nkev == KQ_NEVENTS) { 1244 KQ_UNLOCK_FLUX(kq); 1245 error = kevent_copyout(&eventlist, eventseg, keva, 1246 nkev); 1247 nkev = 0; 1248 kevp = keva; 1249 KQ_LOCK(kq); 1250 if (error) 1251 break; 1252 } 1253 } 1254 TAILQ_REMOVE(&kq->kq_head, marker, kn_tqe); 1255 done: 1256 KQ_OWNED(kq); 1257 KQ_UNLOCK_FLUX(kq); 1258 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal); 1259 knote_free(marker); 1260 done_nl: 1261 KQ_NOTOWNED(kq); 1262 if (nkev != 0) 1263 error = kevent_copyout(&eventlist, eventseg, keva, nkev); 1264 td->td_retval[0] = maxevents - count; 1265 return (error); 1266 } 1267 1268 /* 1269 * XXX 1270 * This could be expanded to call kqueue_scan, if desired. 1271 */ 1272 /*ARGSUSED*/ 1273 static int 1274 kqueue_read(struct file *fp, struct uio *uio, struct ucred *active_cred, 1275 int flags, struct thread *td) 1276 { 1277 return (ENXIO); 1278 } 1279 1280 /*ARGSUSED*/ 1281 static int 1282 kqueue_write(struct file *fp, struct uio *uio, struct ucred *active_cred, 1283 int flags, struct thread *td) 1284 { 1285 return (ENXIO); 1286 } 1287 1288 /*ARGSUSED*/ 1289 static int 1290 kqueue_ioctl(struct file *fp, u_long cmd, void *data, 1291 struct ucred *active_cred, struct thread *td) 1292 { 1293 /* 1294 * Enabling sigio causes two major problems: 1295 * 1) infinite recursion: 1296 * Synopsys: kevent is being used to track signals and have FIOASYNC 1297 * set. On receipt of a signal this will cause a kqueue to recurse 1298 * into itself over and over. Sending the sigio causes the kqueue 1299 * to become ready, which in turn posts sigio again, forever. 1300 * Solution: this can be solved by setting a flag in the kqueue that 1301 * we have a SIGIO in progress. 1302 * 2) locking problems: 1303 * Synopsys: Kqueue is a leaf subsystem, but adding signalling puts 1304 * us above the proc and pgrp locks. 1305 * Solution: Post a signal using an async mechanism, being sure to 1306 * record a generation count in the delivery so that we do not deliver 1307 * a signal to the wrong process. 1308 * 1309 * Note, these two mechanisms are somewhat mutually exclusive! 1310 */ 1311 #if 0 1312 struct kqueue *kq; 1313 1314 kq = fp->f_data; 1315 switch (cmd) { 1316 case FIOASYNC: 1317 if (*(int *)data) { 1318 kq->kq_state |= KQ_ASYNC; 1319 } else { 1320 kq->kq_state &= ~KQ_ASYNC; 1321 } 1322 return (0); 1323 1324 case FIOSETOWN: 1325 return (fsetown(*(int *)data, &kq->kq_sigio)); 1326 1327 case FIOGETOWN: 1328 *(int *)data = fgetown(&kq->kq_sigio); 1329 return (0); 1330 } 1331 #endif 1332 1333 return (ENOTTY); 1334 } 1335 1336 /*ARGSUSED*/ 1337 static int 1338 kqueue_poll(struct file *fp, int events, struct ucred *active_cred, 1339 struct thread *td) 1340 { 1341 struct kqueue *kq; 1342 int revents = 0; 1343 int error; 1344 1345 if ((error = kqueue_aquire(fp, &kq))) 1346 return POLLERR; 1347 1348 KQ_LOCK(kq); 1349 if (events & (POLLIN | POLLRDNORM)) { 1350 if (kq->kq_count) { 1351 revents |= events & (POLLIN | POLLRDNORM); 1352 } else { 1353 selrecord(td, &kq->kq_sel); 1354 kq->kq_state |= KQ_SEL; 1355 } 1356 } 1357 kqueue_release(kq, 1); 1358 KQ_UNLOCK(kq); 1359 return (revents); 1360 } 1361 1362 /*ARGSUSED*/ 1363 static int 1364 kqueue_stat(struct file *fp, struct stat *st, struct ucred *active_cred, 1365 struct thread *td) 1366 { 1367 1368 return (ENXIO); 1369 } 1370 1371 /*ARGSUSED*/ 1372 static int 1373 kqueue_close(struct file *fp, struct thread *td) 1374 { 1375 struct kqueue *kq = fp->f_data; 1376 struct filedesc *fdp; 1377 struct knote *kn; 1378 int i; 1379 int error; 1380 1381 if ((error = kqueue_aquire(fp, &kq))) 1382 return error; 1383 1384 KQ_LOCK(kq); 1385 1386 KASSERT((kq->kq_state & KQ_CLOSING) != KQ_CLOSING, 1387 ("kqueue already closing")); 1388 kq->kq_state |= KQ_CLOSING; 1389 if (kq->kq_refcnt > 1) 1390 msleep(&kq->kq_refcnt, &kq->kq_lock, PSOCK, "kqclose", 0); 1391 1392 KASSERT(kq->kq_refcnt == 1, ("other refs are out there!")); 1393 fdp = kq->kq_fdp; 1394 1395 KASSERT(knlist_empty(&kq->kq_sel.si_note), 1396 ("kqueue's knlist not empty")); 1397 1398 for (i = 0; i < kq->kq_knlistsize; i++) { 1399 while ((kn = SLIST_FIRST(&kq->kq_knlist[i])) != NULL) { 1400 KASSERT((kn->kn_status & KN_INFLUX) == 0, 1401 ("KN_INFLUX set when not suppose to be")); 1402 kn->kn_status |= KN_INFLUX; 1403 KQ_UNLOCK(kq); 1404 if (!(kn->kn_status & KN_DETACHED)) 1405 kn->kn_fop->f_detach(kn); 1406 knote_drop(kn, td); 1407 KQ_LOCK(kq); 1408 } 1409 } 1410 if (kq->kq_knhashmask != 0) { 1411 for (i = 0; i <= kq->kq_knhashmask; i++) { 1412 while ((kn = SLIST_FIRST(&kq->kq_knhash[i])) != NULL) { 1413 KASSERT((kn->kn_status & KN_INFLUX) == 0, 1414 ("KN_INFLUX set when not suppose to be")); 1415 kn->kn_status |= KN_INFLUX; 1416 KQ_UNLOCK(kq); 1417 if (!(kn->kn_status & KN_DETACHED)) 1418 kn->kn_fop->f_detach(kn); 1419 knote_drop(kn, td); 1420 KQ_LOCK(kq); 1421 } 1422 } 1423 } 1424 1425 if ((kq->kq_state & KQ_TASKSCHED) == KQ_TASKSCHED) { 1426 kq->kq_state |= KQ_TASKDRAIN; 1427 msleep(&kq->kq_state, &kq->kq_lock, PSOCK, "kqtqdr", 0); 1428 } 1429 1430 if ((kq->kq_state & KQ_SEL) == KQ_SEL) { 1431 kq->kq_state &= ~KQ_SEL; 1432 selwakeuppri(&kq->kq_sel, PSOCK); 1433 } 1434 1435 KQ_UNLOCK(kq); 1436 1437 FILEDESC_LOCK_FAST(fdp); 1438 SLIST_REMOVE(&fdp->fd_kqlist, kq, kqueue, kq_list); 1439 FILEDESC_UNLOCK_FAST(fdp); 1440 1441 knlist_destroy(&kq->kq_sel.si_note); 1442 mtx_destroy(&kq->kq_lock); 1443 kq->kq_fdp = NULL; 1444 1445 if (kq->kq_knhash != NULL) 1446 free(kq->kq_knhash, M_KQUEUE); 1447 if (kq->kq_knlist != NULL) 1448 free(kq->kq_knlist, M_KQUEUE); 1449 1450 funsetown(&kq->kq_sigio); 1451 free(kq, M_KQUEUE); 1452 fp->f_data = NULL; 1453 1454 return (0); 1455 } 1456 1457 static void 1458 kqueue_wakeup(struct kqueue *kq) 1459 { 1460 KQ_OWNED(kq); 1461 1462 if ((kq->kq_state & KQ_SLEEP) == KQ_SLEEP) { 1463 kq->kq_state &= ~KQ_SLEEP; 1464 wakeup(kq); 1465 } 1466 if ((kq->kq_state & KQ_SEL) == KQ_SEL) { 1467 kq->kq_state &= ~KQ_SEL; 1468 selwakeuppri(&kq->kq_sel, PSOCK); 1469 } 1470 if (!knlist_empty(&kq->kq_sel.si_note)) 1471 kqueue_schedtask(kq); 1472 if ((kq->kq_state & KQ_ASYNC) == KQ_ASYNC) { 1473 pgsigio(&kq->kq_sigio, SIGIO, 0); 1474 } 1475 } 1476 1477 /* 1478 * Walk down a list of knotes, activating them if their event has triggered. 1479 * 1480 * There is a possibility to optimize in the case of one kq watching another. 1481 * Instead of scheduling a task to wake it up, you could pass enough state 1482 * down the chain to make up the parent kqueue. Make this code functional 1483 * first. 1484 */ 1485 void 1486 knote(struct knlist *list, long hint, int islocked) 1487 { 1488 struct kqueue *kq; 1489 struct knote *kn; 1490 1491 if (list == NULL) 1492 return; 1493 1494 mtx_assert(list->kl_lock, islocked ? MA_OWNED : MA_NOTOWNED); 1495 if (!islocked) 1496 mtx_lock(list->kl_lock); 1497 /* 1498 * If we unlock the list lock (and set KN_INFLUX), we can eliminate 1499 * the kqueue scheduling, but this will introduce four 1500 * lock/unlock's for each knote to test. If we do, continue to use 1501 * SLIST_FOREACH, SLIST_FOREACH_SAFE is not safe in our case, it is 1502 * only safe if you want to remove the current item, which we are 1503 * not doing. 1504 */ 1505 SLIST_FOREACH(kn, &list->kl_list, kn_selnext) { 1506 kq = kn->kn_kq; 1507 if ((kn->kn_status & KN_INFLUX) != KN_INFLUX) { 1508 KQ_LOCK(kq); 1509 if ((kn->kn_status & KN_INFLUX) != KN_INFLUX) { 1510 kn->kn_status |= KN_HASKQLOCK; 1511 if (kn->kn_fop->f_event(kn, hint)) 1512 KNOTE_ACTIVATE(kn, 1); 1513 kn->kn_status &= ~KN_HASKQLOCK; 1514 } 1515 KQ_UNLOCK(kq); 1516 } 1517 kq = NULL; 1518 } 1519 if (!islocked) 1520 mtx_unlock(list->kl_lock); 1521 } 1522 1523 /* 1524 * add a knote to a knlist 1525 */ 1526 void 1527 knlist_add(struct knlist *knl, struct knote *kn, int islocked) 1528 { 1529 mtx_assert(knl->kl_lock, islocked ? MA_OWNED : MA_NOTOWNED); 1530 KQ_NOTOWNED(kn->kn_kq); 1531 KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) == 1532 (KN_INFLUX|KN_DETACHED), ("knote not KN_INFLUX and KN_DETACHED")); 1533 if (!islocked) 1534 mtx_lock(knl->kl_lock); 1535 SLIST_INSERT_HEAD(&knl->kl_list, kn, kn_selnext); 1536 if (!islocked) 1537 mtx_unlock(knl->kl_lock); 1538 KQ_LOCK(kn->kn_kq); 1539 kn->kn_knlist = knl; 1540 kn->kn_status &= ~KN_DETACHED; 1541 KQ_UNLOCK(kn->kn_kq); 1542 } 1543 1544 static void 1545 knlist_remove_kq(struct knlist *knl, struct knote *kn, int knlislocked, int kqislocked) 1546 { 1547 KASSERT(!(!!kqislocked && !knlislocked), ("kq locked w/o knl locked")); 1548 mtx_assert(knl->kl_lock, knlislocked ? MA_OWNED : MA_NOTOWNED); 1549 mtx_assert(&kn->kn_kq->kq_lock, kqislocked ? MA_OWNED : MA_NOTOWNED); 1550 if (!kqislocked) 1551 KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) == KN_INFLUX, 1552 ("knlist_remove called w/o knote being KN_INFLUX or already removed")); 1553 if (!knlislocked) 1554 mtx_lock(knl->kl_lock); 1555 SLIST_REMOVE(&knl->kl_list, kn, knote, kn_selnext); 1556 kn->kn_knlist = NULL; 1557 if (!knlislocked) 1558 mtx_unlock(knl->kl_lock); 1559 if (!kqislocked) 1560 KQ_LOCK(kn->kn_kq); 1561 kn->kn_status |= KN_DETACHED; 1562 if (!kqislocked) 1563 KQ_UNLOCK(kn->kn_kq); 1564 } 1565 1566 /* 1567 * remove all knotes from a specified klist 1568 */ 1569 void 1570 knlist_remove(struct knlist *knl, struct knote *kn, int islocked) 1571 { 1572 1573 knlist_remove_kq(knl, kn, islocked, 0); 1574 } 1575 1576 /* 1577 * remove knote from a specified klist while in f_event handler. 1578 */ 1579 void 1580 knlist_remove_inevent(struct knlist *knl, struct knote *kn) 1581 { 1582 1583 knlist_remove_kq(knl, kn, 1, 1584 (kn->kn_status & KN_HASKQLOCK) == KN_HASKQLOCK); 1585 } 1586 1587 int 1588 knlist_empty(struct knlist *knl) 1589 { 1590 1591 mtx_assert(knl->kl_lock, MA_OWNED); 1592 return SLIST_EMPTY(&knl->kl_list); 1593 } 1594 1595 static struct mtx knlist_lock; 1596 MTX_SYSINIT(knlist_lock, &knlist_lock, "knlist lock for lockless objects", 1597 MTX_DEF); 1598 1599 void 1600 knlist_init(struct knlist *knl, struct mtx *mtx) 1601 { 1602 1603 if (mtx == NULL) 1604 knl->kl_lock = &knlist_lock; 1605 else 1606 knl->kl_lock = mtx; 1607 1608 SLIST_INIT(&knl->kl_list); 1609 } 1610 1611 void 1612 knlist_destroy(struct knlist *knl) 1613 { 1614 1615 #ifdef INVARIANTS 1616 /* 1617 * if we run across this error, we need to find the offending 1618 * driver and have it call knlist_clear. 1619 */ 1620 if (!SLIST_EMPTY(&knl->kl_list)) 1621 printf("WARNING: destroying knlist w/ knotes on it!\n"); 1622 #endif 1623 1624 knl->kl_lock = NULL; 1625 SLIST_INIT(&knl->kl_list); 1626 } 1627 1628 /* 1629 * Even if we are locked, we may need to drop the lock to allow any influx 1630 * knotes time to "settle". 1631 */ 1632 void 1633 knlist_cleardel(struct knlist *knl, struct thread *td, int islocked, int killkn) 1634 { 1635 struct knote *kn; 1636 struct kqueue *kq; 1637 1638 if (islocked) 1639 mtx_assert(knl->kl_lock, MA_OWNED); 1640 else { 1641 mtx_assert(knl->kl_lock, MA_NOTOWNED); 1642 again: /* need to reaquire lock since we have dropped it */ 1643 mtx_lock(knl->kl_lock); 1644 } 1645 1646 SLIST_FOREACH(kn, &knl->kl_list, kn_selnext) { 1647 kq = kn->kn_kq; 1648 KQ_LOCK(kq); 1649 if ((kn->kn_status & KN_INFLUX)) { 1650 KQ_UNLOCK(kq); 1651 continue; 1652 } 1653 knlist_remove_kq(knl, kn, 1, 1); 1654 if (killkn) { 1655 kn->kn_status |= KN_INFLUX | KN_DETACHED; 1656 KQ_UNLOCK(kq); 1657 knote_drop(kn, td); 1658 } else { 1659 /* Make sure cleared knotes disappear soon */ 1660 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 1661 KQ_UNLOCK(kq); 1662 } 1663 kq = NULL; 1664 } 1665 1666 if (!SLIST_EMPTY(&knl->kl_list)) { 1667 /* there are still KN_INFLUX remaining */ 1668 kn = SLIST_FIRST(&knl->kl_list); 1669 kq = kn->kn_kq; 1670 KQ_LOCK(kq); 1671 KASSERT(kn->kn_status & KN_INFLUX, 1672 ("knote removed w/o list lock")); 1673 mtx_unlock(knl->kl_lock); 1674 kq->kq_state |= KQ_FLUXWAIT; 1675 msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqkclr", 0); 1676 kq = NULL; 1677 goto again; 1678 } 1679 1680 if (islocked) 1681 mtx_assert(knl->kl_lock, MA_OWNED); 1682 else { 1683 mtx_unlock(knl->kl_lock); 1684 mtx_assert(knl->kl_lock, MA_NOTOWNED); 1685 } 1686 } 1687 1688 /* 1689 * remove all knotes referencing a specified fd 1690 * must be called with FILEDESC lock. This prevents a race where a new fd 1691 * comes along and occupies the entry and we attach a knote to the fd. 1692 */ 1693 void 1694 knote_fdclose(struct thread *td, int fd) 1695 { 1696 struct filedesc *fdp = td->td_proc->p_fd; 1697 struct kqueue *kq; 1698 struct knote *kn; 1699 int influx; 1700 1701 FILEDESC_LOCK_ASSERT(fdp, MA_OWNED); 1702 1703 /* 1704 * We shouldn't have to worry about new kevents appearing on fd 1705 * since filedesc is locked. 1706 */ 1707 SLIST_FOREACH(kq, &fdp->fd_kqlist, kq_list) { 1708 KQ_LOCK(kq); 1709 1710 again: 1711 influx = 0; 1712 while (kq->kq_knlistsize > fd && 1713 (kn = SLIST_FIRST(&kq->kq_knlist[fd])) != NULL) { 1714 if (kn->kn_status & KN_INFLUX) { 1715 /* someone else might be waiting on our knote */ 1716 if (influx) 1717 wakeup(kq); 1718 kq->kq_state |= KQ_FLUXWAIT; 1719 msleep(kq, &kq->kq_lock, PSOCK, "kqflxwt", 0); 1720 goto again; 1721 } 1722 kn->kn_status |= KN_INFLUX; 1723 KQ_UNLOCK(kq); 1724 if (!(kn->kn_status & KN_DETACHED)) 1725 kn->kn_fop->f_detach(kn); 1726 knote_drop(kn, td); 1727 influx = 1; 1728 KQ_LOCK(kq); 1729 } 1730 KQ_UNLOCK_FLUX(kq); 1731 } 1732 } 1733 1734 static int 1735 knote_attach(struct knote *kn, struct kqueue *kq) 1736 { 1737 struct klist *list; 1738 1739 KASSERT(kn->kn_status & KN_INFLUX, ("knote not marked INFLUX")); 1740 KQ_OWNED(kq); 1741 1742 if (kn->kn_fop->f_isfd) { 1743 if (kn->kn_id >= kq->kq_knlistsize) 1744 return ENOMEM; 1745 list = &kq->kq_knlist[kn->kn_id]; 1746 } else { 1747 if (kq->kq_knhash == NULL) 1748 return ENOMEM; 1749 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)]; 1750 } 1751 1752 SLIST_INSERT_HEAD(list, kn, kn_link); 1753 1754 return 0; 1755 } 1756 1757 /* 1758 * knote must already have been detatched using the f_detach method. 1759 * no lock need to be held, it is assumed that the KN_INFLUX flag is set 1760 * to prevent other removal. 1761 */ 1762 static void 1763 knote_drop(struct knote *kn, struct thread *td) 1764 { 1765 struct kqueue *kq; 1766 struct klist *list; 1767 1768 kq = kn->kn_kq; 1769 1770 KQ_NOTOWNED(kq); 1771 KASSERT((kn->kn_status & KN_INFLUX) == KN_INFLUX, 1772 ("knote_drop called without KN_INFLUX set in kn_status")); 1773 1774 KQ_LOCK(kq); 1775 if (kn->kn_fop->f_isfd) 1776 list = &kq->kq_knlist[kn->kn_id]; 1777 else 1778 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)]; 1779 1780 SLIST_REMOVE(list, kn, knote, kn_link); 1781 if (kn->kn_status & KN_QUEUED) 1782 knote_dequeue(kn); 1783 KQ_UNLOCK_FLUX(kq); 1784 1785 if (kn->kn_fop->f_isfd) { 1786 fdrop(kn->kn_fp, td); 1787 kn->kn_fp = NULL; 1788 } 1789 kqueue_fo_release(kn->kn_kevent.filter); 1790 kn->kn_fop = NULL; 1791 knote_free(kn); 1792 } 1793 1794 static void 1795 knote_enqueue(struct knote *kn) 1796 { 1797 struct kqueue *kq = kn->kn_kq; 1798 1799 KQ_OWNED(kn->kn_kq); 1800 KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued")); 1801 1802 TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe); 1803 kn->kn_status |= KN_QUEUED; 1804 kq->kq_count++; 1805 kqueue_wakeup(kq); 1806 } 1807 1808 static void 1809 knote_dequeue(struct knote *kn) 1810 { 1811 struct kqueue *kq = kn->kn_kq; 1812 1813 KQ_OWNED(kn->kn_kq); 1814 KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued")); 1815 1816 TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe); 1817 kn->kn_status &= ~KN_QUEUED; 1818 kq->kq_count--; 1819 } 1820 1821 static void 1822 knote_init(void) 1823 { 1824 1825 knote_zone = uma_zcreate("KNOTE", sizeof(struct knote), NULL, NULL, 1826 NULL, NULL, UMA_ALIGN_PTR, 0); 1827 } 1828 SYSINIT(knote, SI_SUB_PSEUDO, SI_ORDER_ANY, knote_init, NULL) 1829 1830 static struct knote * 1831 knote_alloc(int waitok) 1832 { 1833 return ((struct knote *)uma_zalloc(knote_zone, 1834 (waitok ? M_WAITOK : M_NOWAIT)|M_ZERO)); 1835 } 1836 1837 static void 1838 knote_free(struct knote *kn) 1839 { 1840 if (kn != NULL) 1841 uma_zfree(knote_zone, kn); 1842 } 1843