kern_event.c revision 135240
1238438Sdteske/*- 2238438Sdteske * Copyright (c) 1999,2000,2001 Jonathan Lemon <jlemon@FreeBSD.org> 3249746Sdteske * Copyright 2004 John-Mark Gurney <jmg@FreeBSD.org> 4252980Sdteske * All rights reserved. 5238438Sdteske * 6238438Sdteske * Redistribution and use in source and binary forms, with or without 7238438Sdteske * modification, are permitted provided that the following conditions 8238438Sdteske * are met: 9238438Sdteske * 1. Redistributions of source code must retain the above copyright 10238438Sdteske * notice, this list of conditions and the following disclaimer. 11238438Sdteske * 2. Redistributions in binary form must reproduce the above copyright 12238438Sdteske * notice, this list of conditions and the following disclaimer in the 13238438Sdteske * documentation and/or other materials provided with the distribution. 14238438Sdteske * 15238438Sdteske * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 16252987Sdteske * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 17238438Sdteske * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 18238438Sdteske * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 19238438Sdteske * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 20252987Sdteske * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 21238438Sdteske * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 22238438Sdteske * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 23238438Sdteske * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 24238438Sdteske * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 25238438Sdteske * SUCH DAMAGE. 26238438Sdteske */ 27238438Sdteske 28238438Sdteske#include <sys/cdefs.h> 29238438Sdteske__FBSDID("$FreeBSD: head/sys/kern/kern_event.c 135240 2004-09-14 18:38:16Z jmg $"); 30238438Sdteske 31240684Sdteske#include <sys/param.h> 32240684Sdteske#include <sys/systm.h> 33244675Sdteske#include <sys/kernel.h> 34240684Sdteske#include <sys/lock.h> 35240684Sdteske#include <sys/mutex.h> 36238438Sdteske#include <sys/proc.h> 37240684Sdteske#include <sys/malloc.h> 38238438Sdteske#include <sys/unistd.h> 39238438Sdteske#include <sys/file.h> 40260678Sdteske#include <sys/filedesc.h> 41260678Sdteske#include <sys/filio.h> 42238438Sdteske#include <sys/fcntl.h> 43238438Sdteske#include <sys/kthread.h> 44238438Sdteske#include <sys/selinfo.h> 45238438Sdteske#include <sys/queue.h> 46238438Sdteske#include <sys/event.h> 47238438Sdteske#include <sys/eventvar.h> 48238438Sdteske#include <sys/poll.h> 49238438Sdteske#include <sys/protosw.h> 50238438Sdteske#include <sys/sigio.h> 51238438Sdteske#include <sys/signalvar.h> 52238438Sdteske#include <sys/socket.h> 53238438Sdteske#include <sys/socketvar.h> 54238438Sdteske#include <sys/stat.h> 55238438Sdteske#include <sys/sysctl.h> 56238438Sdteske#include <sys/sysproto.h> 57238438Sdteske#include <sys/taskqueue.h> 58238438Sdteske#include <sys/uio.h> 59250633Sdteske 60238438Sdteske#include <vm/uma.h> 61252178Sdteske 62238438SdteskeMALLOC_DEFINE(M_KQUEUE, "kqueue", "memory for kqueue system"); 63238438Sdteske/* 64238438Sdteske * This lock is used if multiple kq locks are required. This possibly 65238438Sdteske * should be made into a per proc lock. 66238438Sdteske */ 67238438Sdteskestatic struct mtx kq_global; 68238438SdteskeMTX_SYSINIT(kq_global, &kq_global, "kqueue order", MTX_DEF); 69238438Sdteske#define KQ_GLOBAL_LOCK(lck, haslck) do { \ 70238438Sdteske if (!haslck) \ 71238438Sdteske mtx_lock(lck); \ 72238438Sdteske haslck = 1; \ 73238438Sdteske} while (0) 74238438Sdteske#define KQ_GLOBAL_UNLOCK(lck, haslck) do { \ 75238438Sdteske if (haslck) \ 76238438Sdteske mtx_unlock(lck); \ 77238438Sdteske haslck = 0; \ 78251908Sdteske} while (0) 79238438Sdteske 80251908SdteskeTASKQUEUE_DEFINE_THREAD(kqueue); 81238438Sdteske 82238438Sdteskestatic int kqueue_aquire(struct file *fp, struct kqueue **kqp); 83238438Sdteskestatic void kqueue_release(struct kqueue *kq, int locked); 84238438Sdteskestatic int kqueue_expand(struct kqueue *kq, struct filterops *fops, 85238438Sdteske uintptr_t ident, int waitok); 86static void kqueue_task(void *arg, int pending); 87static int kqueue_scan(struct kqueue *kq, int maxevents, 88 struct kevent *ulistp, const struct timespec *timeout, 89 struct kevent *keva, struct thread *td); 90static void kqueue_wakeup(struct kqueue *kq); 91static struct filterops *kqueue_fo_find(int filt); 92static void kqueue_fo_release(int filt); 93 94static fo_rdwr_t kqueue_read; 95static fo_rdwr_t kqueue_write; 96static fo_ioctl_t kqueue_ioctl; 97static fo_poll_t kqueue_poll; 98static fo_kqfilter_t kqueue_kqfilter; 99static fo_stat_t kqueue_stat; 100static fo_close_t kqueue_close; 101 102static struct fileops kqueueops = { 103 .fo_read = kqueue_read, 104 .fo_write = kqueue_write, 105 .fo_ioctl = kqueue_ioctl, 106 .fo_poll = kqueue_poll, 107 .fo_kqfilter = kqueue_kqfilter, 108 .fo_stat = kqueue_stat, 109 .fo_close = kqueue_close, 110}; 111 112static int knote_attach(struct knote *kn, struct kqueue *kq); 113static void knote_drop(struct knote *kn, struct thread *td); 114static void knote_enqueue(struct knote *kn); 115static void knote_dequeue(struct knote *kn); 116static void knote_init(void); 117static struct knote *knote_alloc(int waitok); 118static void knote_free(struct knote *kn); 119 120static void filt_kqdetach(struct knote *kn); 121static int filt_kqueue(struct knote *kn, long hint); 122static int filt_procattach(struct knote *kn); 123static void filt_procdetach(struct knote *kn); 124static int filt_proc(struct knote *kn, long hint); 125static int filt_fileattach(struct knote *kn); 126static void filt_timerexpire(void *knx); 127static int filt_timerattach(struct knote *kn); 128static void filt_timerdetach(struct knote *kn); 129static int filt_timer(struct knote *kn, long hint); 130 131static struct filterops file_filtops = 132 { 1, filt_fileattach, NULL, NULL }; 133static struct filterops kqread_filtops = 134 { 1, NULL, filt_kqdetach, filt_kqueue }; 135/* XXX - move to kern_proc.c? */ 136static struct filterops proc_filtops = 137 { 0, filt_procattach, filt_procdetach, filt_proc }; 138static struct filterops timer_filtops = 139 { 0, filt_timerattach, filt_timerdetach, filt_timer }; 140 141static uma_zone_t knote_zone; 142static int kq_ncallouts = 0; 143static int kq_calloutmax = (4 * 1024); 144SYSCTL_INT(_kern, OID_AUTO, kq_calloutmax, CTLFLAG_RW, 145 &kq_calloutmax, 0, "Maximum number of callouts allocated for kqueue"); 146 147/* XXX - ensure not KN_INFLUX?? */ 148#define KNOTE_ACTIVATE(kn, islock) do { \ 149 if ((islock)) \ 150 mtx_assert(&(kn)->kn_kq->kq_lock, MA_OWNED); \ 151 else \ 152 KQ_LOCK((kn)->kn_kq); \ 153 (kn)->kn_status |= KN_ACTIVE; \ 154 if (((kn)->kn_status & (KN_QUEUED | KN_DISABLED)) == 0) \ 155 knote_enqueue((kn)); \ 156 if (!(islock)) \ 157 KQ_UNLOCK((kn)->kn_kq); \ 158} while(0) 159#define KQ_LOCK(kq) do { \ 160 mtx_lock(&(kq)->kq_lock); \ 161} while (0) 162#define KQ_FLUX_WAKEUP(kq) do { \ 163 if (((kq)->kq_state & KQ_FLUXWAIT) == KQ_FLUXWAIT) { \ 164 (kq)->kq_state &= ~KQ_FLUXWAIT; \ 165 wakeup((kq)); \ 166 } \ 167} while (0) 168#define KQ_UNLOCK_FLUX(kq) do { \ 169 KQ_FLUX_WAKEUP(kq); \ 170 mtx_unlock(&(kq)->kq_lock); \ 171} while (0) 172#define KQ_UNLOCK(kq) do { \ 173 mtx_unlock(&(kq)->kq_lock); \ 174} while (0) 175#define KQ_OWNED(kq) do { \ 176 mtx_assert(&(kq)->kq_lock, MA_OWNED); \ 177} while (0) 178#define KQ_NOTOWNED(kq) do { \ 179 mtx_assert(&(kq)->kq_lock, MA_NOTOWNED); \ 180} while (0) 181#define KN_LIST_LOCK(kn) do { \ 182 if (kn->kn_knlist != NULL) \ 183 mtx_lock(kn->kn_knlist->kl_lock); \ 184} while (0) 185#define KN_LIST_UNLOCK(kn) do { \ 186 if (kn->kn_knlist != NULL) \ 187 mtx_unlock(kn->kn_knlist->kl_lock); \ 188} while (0) 189 190#define KN_HASHSIZE 64 /* XXX should be tunable */ 191#define KN_HASH(val, mask) (((val) ^ (val >> 8)) & (mask)) 192 193static int 194filt_nullattach(struct knote *kn) 195{ 196 197 return (ENXIO); 198}; 199 200struct filterops null_filtops = 201 { 0, filt_nullattach, NULL, NULL }; 202 203/* XXX - make SYSINIT to add these, and move into respective modules. */ 204extern struct filterops sig_filtops; 205extern struct filterops fs_filtops; 206 207/* 208 * Table for for all system-defined filters. 209 */ 210static struct mtx filterops_lock; 211MTX_SYSINIT(kqueue_filterops, &filterops_lock, "protect sysfilt_ops", 212 MTX_DEF); 213static struct { 214 struct filterops *for_fop; 215 int for_refcnt; 216} sysfilt_ops[EVFILT_SYSCOUNT] = { 217 { &file_filtops }, /* EVFILT_READ */ 218 { &file_filtops }, /* EVFILT_WRITE */ 219 { &null_filtops }, /* EVFILT_AIO */ 220 { &file_filtops }, /* EVFILT_VNODE */ 221 { &proc_filtops }, /* EVFILT_PROC */ 222 { &sig_filtops }, /* EVFILT_SIGNAL */ 223 { &timer_filtops }, /* EVFILT_TIMER */ 224 { &file_filtops }, /* EVFILT_NETDEV */ 225 { &fs_filtops }, /* EVFILT_FS */ 226}; 227 228/* 229 * Simple redirection for all cdevsw style objects to call their fo_kqfilter 230 * method. 231 */ 232static int 233filt_fileattach(struct knote *kn) 234{ 235 236 return (fo_kqfilter(kn->kn_fp, kn)); 237} 238 239/*ARGSUSED*/ 240static int 241kqueue_kqfilter(struct file *fp, struct knote *kn) 242{ 243 struct kqueue *kq = kn->kn_fp->f_data; 244 245 if (kn->kn_filter != EVFILT_READ) 246 return (EINVAL); 247 248 kn->kn_status |= KN_KQUEUE; 249 kn->kn_fop = &kqread_filtops; 250 knlist_add(&kq->kq_sel.si_note, kn, 0); 251 252 return (0); 253} 254 255static void 256filt_kqdetach(struct knote *kn) 257{ 258 struct kqueue *kq = kn->kn_fp->f_data; 259 260 knlist_remove(&kq->kq_sel.si_note, kn, 0); 261} 262 263/*ARGSUSED*/ 264static int 265filt_kqueue(struct knote *kn, long hint) 266{ 267 struct kqueue *kq = kn->kn_fp->f_data; 268 269 kn->kn_data = kq->kq_count; 270 return (kn->kn_data > 0); 271} 272 273/* XXX - move to kern_proc.c? */ 274static int 275filt_procattach(struct knote *kn) 276{ 277 struct proc *p; 278 int immediate; 279 int error; 280 281 immediate = 0; 282 p = pfind(kn->kn_id); 283 if (p == NULL && (kn->kn_sfflags & NOTE_EXIT)) { 284 p = zpfind(kn->kn_id); 285 immediate = 1; 286 } else if (p != NULL && (p->p_flag & P_WEXIT)) { 287 immediate = 1; 288 } 289 290 if (p == NULL) 291 return (ESRCH); 292 if ((error = p_cansee(curthread, p))) 293 return (error); 294 295 kn->kn_ptr.p_proc = p; 296 kn->kn_flags |= EV_CLEAR; /* automatically set */ 297 298 /* 299 * internal flag indicating registration done by kernel 300 */ 301 if (kn->kn_flags & EV_FLAG1) { 302 kn->kn_data = kn->kn_sdata; /* ppid */ 303 kn->kn_fflags = NOTE_CHILD; 304 kn->kn_flags &= ~EV_FLAG1; 305 } 306 307 if (immediate == 0) 308 knlist_add(&p->p_klist, kn, 1); 309 310 /* 311 * Immediately activate any exit notes if the target process is a 312 * zombie. This is necessary to handle the case where the target 313 * process, e.g. a child, dies before the kevent is registered. 314 */ 315 if (immediate && filt_proc(kn, NOTE_EXIT)) 316 KNOTE_ACTIVATE(kn, 0); 317 318 PROC_UNLOCK(p); 319 320 return (0); 321} 322 323/* 324 * The knote may be attached to a different process, which may exit, 325 * leaving nothing for the knote to be attached to. So when the process 326 * exits, the knote is marked as DETACHED and also flagged as ONESHOT so 327 * it will be deleted when read out. However, as part of the knote deletion, 328 * this routine is called, so a check is needed to avoid actually performing 329 * a detach, because the original process does not exist any more. 330 */ 331/* XXX - move to kern_proc.c? */ 332static void 333filt_procdetach(struct knote *kn) 334{ 335 struct proc *p; 336 337 p = kn->kn_ptr.p_proc; 338 knlist_remove(&p->p_klist, kn, 0); 339 kn->kn_ptr.p_proc = NULL; 340} 341 342/* XXX - move to kern_proc.c? */ 343static int 344filt_proc(struct knote *kn, long hint) 345{ 346 struct proc *p = kn->kn_ptr.p_proc; 347 u_int event; 348 349 /* 350 * mask off extra data 351 */ 352 event = (u_int)hint & NOTE_PCTRLMASK; 353 354 /* 355 * if the user is interested in this event, record it. 356 */ 357 if (kn->kn_sfflags & event) 358 kn->kn_fflags |= event; 359 360 /* 361 * process is gone, so flag the event as finished. 362 */ 363 if (event == NOTE_EXIT) { 364 if (!(kn->kn_status & KN_DETACHED)) 365 knlist_remove_inevent(&p->p_klist, kn); 366 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 367 kn->kn_ptr.p_proc = NULL; 368 return (1); 369 } 370 371 /* 372 * process forked, and user wants to track the new process, 373 * so attach a new knote to it, and immediately report an 374 * event with the parent's pid. 375 */ 376 if ((event == NOTE_FORK) && (kn->kn_sfflags & NOTE_TRACK)) { 377 struct kevent kev; 378 int error; 379 380 /* 381 * register knote with new process. 382 */ 383 kev.ident = hint & NOTE_PDATAMASK; /* pid */ 384 kev.filter = kn->kn_filter; 385 kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1; 386 kev.fflags = kn->kn_sfflags; 387 kev.data = kn->kn_id; /* parent */ 388 kev.udata = kn->kn_kevent.udata; /* preserve udata */ 389 error = kqueue_register(kn->kn_kq, &kev, NULL, 0); 390 if (error) 391 kn->kn_fflags |= NOTE_TRACKERR; 392 } 393 394 return (kn->kn_fflags != 0); 395} 396 397static int 398timertoticks(intptr_t data) 399{ 400 struct timeval tv; 401 int tticks; 402 403 tv.tv_sec = data / 1000; 404 tv.tv_usec = (data % 1000) * 1000; 405 tticks = tvtohz(&tv); 406 407 return tticks; 408} 409 410/* XXX - move to kern_timeout.c? */ 411static void 412filt_timerexpire(void *knx) 413{ 414 struct knote *kn = knx; 415 struct callout *calloutp; 416 417 kn->kn_data++; 418 KNOTE_ACTIVATE(kn, 0); /* XXX - handle locking */ 419 420 if ((kn->kn_flags & EV_ONESHOT) != EV_ONESHOT) { 421 calloutp = (struct callout *)kn->kn_hook; 422 callout_reset(calloutp, timertoticks(kn->kn_sdata), 423 filt_timerexpire, kn); 424 } 425} 426 427/* 428 * data contains amount of time to sleep, in milliseconds 429 */ 430/* XXX - move to kern_timeout.c? */ 431static int 432filt_timerattach(struct knote *kn) 433{ 434 struct callout *calloutp; 435 436 atomic_add_int(&kq_ncallouts, 1); 437 438 if (kq_ncallouts >= kq_calloutmax) { 439 atomic_add_int(&kq_ncallouts, -1); 440 return (ENOMEM); 441 } 442 443 kn->kn_flags |= EV_CLEAR; /* automatically set */ 444 MALLOC(calloutp, struct callout *, sizeof(*calloutp), 445 M_KQUEUE, M_WAITOK); 446 callout_init(calloutp, 1); 447 kn->kn_hook = calloutp; 448 callout_reset(calloutp, timertoticks(kn->kn_sdata), filt_timerexpire, 449 kn); 450 451 return (0); 452} 453 454/* XXX - move to kern_timeout.c? */ 455static void 456filt_timerdetach(struct knote *kn) 457{ 458 struct callout *calloutp; 459 460 calloutp = (struct callout *)kn->kn_hook; 461 callout_drain(calloutp); 462 FREE(calloutp, M_KQUEUE); 463 atomic_add_int(&kq_ncallouts, -1); 464} 465 466/* XXX - move to kern_timeout.c? */ 467static int 468filt_timer(struct knote *kn, long hint) 469{ 470 471 return (kn->kn_data != 0); 472} 473 474/* 475 * MPSAFE 476 */ 477int 478kqueue(struct thread *td, struct kqueue_args *uap) 479{ 480 struct filedesc *fdp; 481 struct kqueue *kq; 482 struct file *fp; 483 int fd, error; 484 485 fdp = td->td_proc->p_fd; 486 error = falloc(td, &fp, &fd); 487 if (error) 488 goto done2; 489 490 /* An extra reference on `nfp' has been held for us by falloc(). */ 491 kq = malloc(sizeof *kq, M_KQUEUE, M_WAITOK | M_ZERO); 492 mtx_init(&kq->kq_lock, "kqueue", NULL, MTX_DEF|MTX_DUPOK); 493 TAILQ_INIT(&kq->kq_head); 494 kq->kq_fdp = fdp; 495 knlist_init(&kq->kq_sel.si_note, &kq->kq_lock); 496 TASK_INIT(&kq->kq_task, 0, kqueue_task, kq); 497 498 FILEDESC_LOCK(fdp); 499 SLIST_INSERT_HEAD(&fdp->fd_kqlist, kq, kq_list); 500 FILEDESC_UNLOCK(fdp); 501 502 FILE_LOCK(fp); 503 fp->f_flag = FREAD | FWRITE; 504 fp->f_type = DTYPE_KQUEUE; 505 fp->f_ops = &kqueueops; 506 fp->f_data = kq; 507 FILE_UNLOCK(fp); 508 fdrop(fp, td); 509 510 td->td_retval[0] = fd; 511done2: 512 return (error); 513} 514 515#ifndef _SYS_SYSPROTO_H_ 516struct kevent_args { 517 int fd; 518 const struct kevent *changelist; 519 int nchanges; 520 struct kevent *eventlist; 521 int nevents; 522 const struct timespec *timeout; 523}; 524#endif 525/* 526 * MPSAFE 527 */ 528int 529kevent(struct thread *td, struct kevent_args *uap) 530{ 531 struct kevent keva[KQ_NEVENTS]; 532 struct kevent *kevp; 533 struct kqueue *kq; 534 struct file *fp; 535 struct timespec ts; 536 int i, n, nerrors, error; 537 538 if ((error = fget(td, uap->fd, &fp)) != 0) 539 return (error); 540 if ((error = kqueue_aquire(fp, &kq)) != 0) 541 goto done_norel; 542 543 if (uap->timeout != NULL) { 544 error = copyin(uap->timeout, &ts, sizeof(ts)); 545 if (error) 546 goto done; 547 uap->timeout = &ts; 548 } 549 550 nerrors = 0; 551 552 while (uap->nchanges > 0) { 553 n = uap->nchanges > KQ_NEVENTS ? KQ_NEVENTS : uap->nchanges; 554 error = copyin(uap->changelist, keva, 555 n * sizeof *keva); 556 if (error) 557 goto done; 558 for (i = 0; i < n; i++) { 559 kevp = &keva[i]; 560 kevp->flags &= ~EV_SYSFLAGS; 561 error = kqueue_register(kq, kevp, td, 1); 562 if (error) { 563 if (uap->nevents != 0) { 564 kevp->flags = EV_ERROR; 565 kevp->data = error; 566 (void) copyout(kevp, 567 uap->eventlist, 568 sizeof(*kevp)); 569 uap->eventlist++; 570 uap->nevents--; 571 nerrors++; 572 } else { 573 goto done; 574 } 575 } 576 } 577 uap->nchanges -= n; 578 uap->changelist += n; 579 } 580 if (nerrors) { 581 td->td_retval[0] = nerrors; 582 error = 0; 583 goto done; 584 } 585 586 error = kqueue_scan(kq, uap->nevents, uap->eventlist, uap->timeout, 587 keva, td); 588done: 589 kqueue_release(kq, 0); 590done_norel: 591 if (fp != NULL) 592 fdrop(fp, td); 593 return (error); 594} 595 596int 597kqueue_add_filteropts(int filt, struct filterops *filtops) 598{ 599 int error; 600 601 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) { 602 printf( 603"trying to add a filterop that is out of range: %d is beyond %d\n", 604 ~filt, EVFILT_SYSCOUNT); 605 return EINVAL; 606 } 607 mtx_lock(&filterops_lock); 608 if (sysfilt_ops[~filt].for_fop != &null_filtops && 609 sysfilt_ops[~filt].for_fop != NULL) 610 error = EEXIST; 611 else { 612 sysfilt_ops[~filt].for_fop = filtops; 613 sysfilt_ops[~filt].for_refcnt = 0; 614 } 615 mtx_unlock(&filterops_lock); 616 617 return (0); 618} 619 620int 621kqueue_del_filteropts(int filt) 622{ 623 int error; 624 625 error = 0; 626 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) 627 return EINVAL; 628 629 mtx_lock(&filterops_lock); 630 if (sysfilt_ops[~filt].for_fop == &null_filtops || 631 sysfilt_ops[~filt].for_fop == NULL) 632 error = EINVAL; 633 else if (sysfilt_ops[~filt].for_refcnt != 0) 634 error = EBUSY; 635 else { 636 sysfilt_ops[~filt].for_fop = &null_filtops; 637 sysfilt_ops[~filt].for_refcnt = 0; 638 } 639 mtx_unlock(&filterops_lock); 640 641 return error; 642} 643 644static struct filterops * 645kqueue_fo_find(int filt) 646{ 647 648 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) 649 return NULL; 650 651 mtx_lock(&filterops_lock); 652 sysfilt_ops[~filt].for_refcnt++; 653 if (sysfilt_ops[~filt].for_fop == NULL) 654 sysfilt_ops[~filt].for_fop = &null_filtops; 655 mtx_unlock(&filterops_lock); 656 657 return sysfilt_ops[~filt].for_fop; 658} 659 660static void 661kqueue_fo_release(int filt) 662{ 663 664 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) 665 return; 666 667 mtx_lock(&filterops_lock); 668 KASSERT(sysfilt_ops[~filt].for_refcnt > 0, 669 ("filter object refcount not valid on release")); 670 sysfilt_ops[~filt].for_refcnt--; 671 mtx_unlock(&filterops_lock); 672} 673 674/* 675 * A ref to kq (obtained via kqueue_aquire) should be held. waitok will 676 * influence if memory allocation should wait. Make sure it is 0 if you 677 * hold any mutexes. 678 */ 679int 680kqueue_register(struct kqueue *kq, struct kevent *kev, struct thread *td, int waitok) 681{ 682 struct filedesc *fdp; 683 struct filterops *fops; 684 struct file *fp; 685 struct knote *kn, *tkn; 686 int error, filt, event; 687 int haskqglobal; 688 int fd; 689 690 fdp = NULL; 691 fp = NULL; 692 kn = NULL; 693 error = 0; 694 haskqglobal = 0; 695 696 filt = kev->filter; 697 fops = kqueue_fo_find(filt); 698 if (fops == NULL) 699 return EINVAL; 700 701 tkn = knote_alloc(waitok); /* prevent waiting with locks */ 702 703findkn: 704 if (fops->f_isfd) { 705 KASSERT(td != NULL, ("td is NULL")); 706 fdp = td->td_proc->p_fd; 707 FILEDESC_LOCK(fdp); 708 /* validate descriptor */ 709 fd = kev->ident; 710 if (fd < 0 || fd >= fdp->fd_nfiles || 711 (fp = fdp->fd_ofiles[fd]) == NULL) { 712 FILEDESC_UNLOCK(fdp); 713 error = EBADF; 714 goto done; 715 } 716 fhold(fp); 717 718 if ((kev->flags & EV_ADD) == EV_ADD && kqueue_expand(kq, fops, 719 kev->ident, 0) != 0) { 720 /* unlock and try again */ 721 FILEDESC_UNLOCK(fdp); 722 fdrop(fp, td); 723 fp = NULL; 724 error = kqueue_expand(kq, fops, kev->ident, waitok); 725 if (error) 726 goto done; 727 goto findkn; 728 } 729 730 if (fp->f_type == DTYPE_KQUEUE) { 731 /* 732 * if we add some inteligence about what we are doing, 733 * we should be able to support events on ourselves. 734 * We need to know when we are doing this to prevent 735 * getting both the knlist lock and the kq lock since 736 * they are the same thing. 737 */ 738 if (fp->f_data == kq) { 739 FILEDESC_UNLOCK(fdp); 740 error = EINVAL; 741 goto done_noglobal; 742 } 743 744 KQ_GLOBAL_LOCK(&kq_global, haskqglobal); 745 } 746 747 KQ_LOCK(kq); 748 if (kev->ident < kq->kq_knlistsize) { 749 SLIST_FOREACH(kn, &kq->kq_knlist[kev->ident], kn_link) 750 if (kev->filter == kn->kn_filter) 751 break; 752 } 753 FILEDESC_UNLOCK(fdp); 754 } else { 755 if ((kev->flags & EV_ADD) == EV_ADD) 756 kqueue_expand(kq, fops, kev->ident, waitok); 757 758 KQ_LOCK(kq); 759 if (kq->kq_knhashmask != 0) { 760 struct klist *list; 761 762 list = &kq->kq_knhash[ 763 KN_HASH((u_long)kev->ident, kq->kq_knhashmask)]; 764 SLIST_FOREACH(kn, list, kn_link) 765 if (kev->ident == kn->kn_id && 766 kev->filter == kn->kn_filter) 767 break; 768 } 769 } 770 771 /* knote is in the process of changing, wait for it to stablize. */ 772 if (kn != NULL && (kn->kn_status & KN_INFLUX) == KN_INFLUX) { 773 if (fp != NULL) { 774 fdrop(fp, td); 775 fp = NULL; 776 } 777 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal); 778 kq->kq_state |= KQ_FLUXWAIT; 779 msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqflxwt", 0); 780 goto findkn; 781 } 782 783 if (kn == NULL && ((kev->flags & EV_ADD) == 0)) { 784 KQ_UNLOCK(kq); 785 error = ENOENT; 786 goto done; 787 } 788 789 /* 790 * kn now contains the matching knote, or NULL if no match 791 */ 792 if (kev->flags & EV_ADD) { 793 if (kn == NULL) { 794 kn = tkn; 795 tkn = NULL; 796 if (kn == NULL) { 797 error = ENOMEM; 798 goto done; 799 } 800 kn->kn_fp = fp; 801 kn->kn_kq = kq; 802 kn->kn_fop = fops; 803 /* 804 * apply reference counts to knote structure, and 805 * do not release it at the end of this routine. 806 */ 807 fops = NULL; 808 fp = NULL; 809 810 kn->kn_sfflags = kev->fflags; 811 kn->kn_sdata = kev->data; 812 kev->fflags = 0; 813 kev->data = 0; 814 kn->kn_kevent = *kev; 815 kn->kn_status = KN_INFLUX|KN_DETACHED; 816 817 error = knote_attach(kn, kq); 818 KQ_UNLOCK(kq); 819 if (error != 0) { 820 tkn = kn; 821 goto done; 822 } 823 824 if ((error = kn->kn_fop->f_attach(kn)) != 0) { 825 knote_drop(kn, td); 826 goto done; 827 } 828 KN_LIST_LOCK(kn); 829 } else { 830 /* 831 * The user may change some filter values after the 832 * initial EV_ADD, but doing so will not reset any 833 * filter which has already been triggered. 834 */ 835 kn->kn_status |= KN_INFLUX; 836 KQ_UNLOCK(kq); 837 KN_LIST_LOCK(kn); 838 kn->kn_sfflags = kev->fflags; 839 kn->kn_sdata = kev->data; 840 kn->kn_kevent.udata = kev->udata; 841 } 842 843 /* 844 * We can get here with kn->kn_knlist == NULL. 845 * This can happen when the initial attach event decides that 846 * the event is "completed" already. i.e. filt_procattach 847 * is called on a zombie process. It will call filt_proc 848 * which will remove it from the list, and NULL kn_knlist. 849 */ 850 event = kn->kn_fop->f_event(kn, 0); 851 KN_LIST_UNLOCK(kn); 852 KQ_LOCK(kq); 853 if (event) 854 KNOTE_ACTIVATE(kn, 1); 855 kn->kn_status &= ~KN_INFLUX; 856 } else if (kev->flags & EV_DELETE) { 857 kn->kn_status |= KN_INFLUX; 858 KQ_UNLOCK(kq); 859 if (!(kn->kn_status & KN_DETACHED)) 860 kn->kn_fop->f_detach(kn); 861 knote_drop(kn, td); 862 goto done; 863 } 864 865 if ((kev->flags & EV_DISABLE) && 866 ((kn->kn_status & KN_DISABLED) == 0)) { 867 kn->kn_status |= KN_DISABLED; 868 } 869 870 if ((kev->flags & EV_ENABLE) && (kn->kn_status & KN_DISABLED)) { 871 kn->kn_status &= ~KN_DISABLED; 872 if ((kn->kn_status & KN_ACTIVE) && 873 ((kn->kn_status & KN_QUEUED) == 0)) 874 knote_enqueue(kn); 875 } 876 KQ_UNLOCK_FLUX(kq); 877 878done: 879 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal); 880done_noglobal: 881 if (fp != NULL) 882 fdrop(fp, td); 883 if (tkn != NULL) 884 knote_free(tkn); 885 if (fops != NULL) 886 kqueue_fo_release(filt); 887 return (error); 888} 889 890static int 891kqueue_aquire(struct file *fp, struct kqueue **kqp) 892{ 893 int error; 894 struct kqueue *kq; 895 896 error = 0; 897 898 FILE_LOCK(fp); 899 do { 900 kq = fp->f_data; 901 if (fp->f_type != DTYPE_KQUEUE || kq == NULL) { 902 error = EBADF; 903 break; 904 } 905 *kqp = kq; 906 KQ_LOCK(kq); 907 if ((kq->kq_state & KQ_CLOSING) == KQ_CLOSING) { 908 KQ_UNLOCK(kq); 909 error = EBADF; 910 break; 911 } 912 kq->kq_refcnt++; 913 KQ_UNLOCK(kq); 914 } while (0); 915 FILE_UNLOCK(fp); 916 917 return error; 918} 919 920static void 921kqueue_release(struct kqueue *kq, int locked) 922{ 923 if (locked) 924 KQ_OWNED(kq); 925 else 926 KQ_LOCK(kq); 927 kq->kq_refcnt--; 928 if (kq->kq_refcnt == 1) 929 wakeup(&kq->kq_refcnt); 930 if (!locked) 931 KQ_UNLOCK(kq); 932} 933 934static void 935kqueue_schedtask(struct kqueue *kq) 936{ 937 938 KQ_OWNED(kq); 939 KASSERT(((kq->kq_state & KQ_TASKDRAIN) != KQ_TASKDRAIN), 940 ("scheduling kqueue task while draining")); 941 942 if ((kq->kq_state & KQ_TASKSCHED) != KQ_TASKSCHED) { 943 taskqueue_enqueue(taskqueue_kqueue, &kq->kq_task); 944 kq->kq_state |= KQ_TASKSCHED; 945 } 946} 947 948/* 949 * Expand the kq to make sure we have storage for fops/ident pair. 950 * 951 * Return 0 on success (or no work necessary), return errno on failure. 952 * 953 * Not calling hashinit w/ waitok (proper malloc flag) should be safe. 954 * If kqueue_register is called from a non-fd context, there usually/should 955 * be no locks held. 956 */ 957static int 958kqueue_expand(struct kqueue *kq, struct filterops *fops, uintptr_t ident, 959 int waitok) 960{ 961 struct klist *list, *tmp_knhash; 962 u_long tmp_knhashmask; 963 int size; 964 int fd; 965 int mflag = waitok ? M_WAITOK : M_NOWAIT; 966 967 KQ_NOTOWNED(kq); 968 969 if (fops->f_isfd) { 970 fd = ident; 971 if (kq->kq_knlistsize <= fd) { 972 size = kq->kq_knlistsize; 973 while (size <= fd) 974 size += KQEXTENT; 975 MALLOC(list, struct klist *, 976 size * sizeof list, M_KQUEUE, mflag); 977 if (list == NULL) 978 return ENOMEM; 979 KQ_LOCK(kq); 980 if (kq->kq_knlistsize > fd) { 981 FREE(list, M_KQUEUE); 982 list = NULL; 983 } else { 984 if (kq->kq_knlist != NULL) { 985 bcopy(kq->kq_knlist, list, 986 kq->kq_knlistsize * sizeof list); 987 FREE(kq->kq_knlist, M_KQUEUE); 988 kq->kq_knlist = NULL; 989 } 990 bzero((caddr_t)list + 991 kq->kq_knlistsize * sizeof list, 992 (size - kq->kq_knlistsize) * sizeof list); 993 kq->kq_knlistsize = size; 994 kq->kq_knlist = list; 995 } 996 KQ_UNLOCK(kq); 997 } 998 } else { 999 if (kq->kq_knhashmask == 0) { 1000 tmp_knhash = hashinit(KN_HASHSIZE, M_KQUEUE, 1001 &tmp_knhashmask); 1002 if (tmp_knhash == NULL) 1003 return ENOMEM; 1004 KQ_LOCK(kq); 1005 if (kq->kq_knhashmask == 0) { 1006 kq->kq_knhash = tmp_knhash; 1007 kq->kq_knhashmask = tmp_knhashmask; 1008 } else { 1009 free(tmp_knhash, M_KQUEUE); 1010 } 1011 KQ_UNLOCK(kq); 1012 } 1013 } 1014 1015 KQ_NOTOWNED(kq); 1016 return 0; 1017} 1018 1019static void 1020kqueue_task(void *arg, int pending) 1021{ 1022 struct kqueue *kq; 1023 int haskqglobal; 1024 1025 haskqglobal = 0; 1026 kq = arg; 1027 1028 KQ_GLOBAL_LOCK(&kq_global, haskqglobal); 1029 KQ_LOCK(kq); 1030 1031 KNOTE_LOCKED(&kq->kq_sel.si_note, 0); 1032 1033 kq->kq_state &= ~KQ_TASKSCHED; 1034 if ((kq->kq_state & KQ_TASKDRAIN) == KQ_TASKDRAIN) { 1035 wakeup(&kq->kq_state); 1036 } 1037 KQ_UNLOCK(kq); 1038 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal); 1039} 1040 1041/* 1042 * Scan, update kn_data (if not ONESHOT), and copyout triggered events. 1043 * We treat KN_MARKER knotes as if they are INFLUX. 1044 */ 1045static int 1046kqueue_scan(struct kqueue *kq, int maxevents, struct kevent *ulistp, 1047 const struct timespec *tsp, struct kevent *keva, struct thread *td) 1048{ 1049 struct kevent *kevp; 1050 struct timeval atv, rtv, ttv; 1051 struct knote *kn, *marker; 1052 int count, timeout, nkev, error; 1053 int haskqglobal; 1054 1055 count = maxevents; 1056 nkev = 0; 1057 error = 0; 1058 haskqglobal = 0; 1059 1060 if (maxevents == 0) 1061 goto done_nl; 1062 1063 if (tsp != NULL) { 1064 TIMESPEC_TO_TIMEVAL(&atv, tsp); 1065 if (itimerfix(&atv)) { 1066 error = EINVAL; 1067 goto done_nl; 1068 } 1069 if (tsp->tv_sec == 0 && tsp->tv_nsec == 0) 1070 timeout = -1; 1071 else 1072 timeout = atv.tv_sec > 24 * 60 * 60 ? 1073 24 * 60 * 60 * hz : tvtohz(&atv); 1074 getmicrouptime(&rtv); 1075 timevaladd(&atv, &rtv); 1076 } else { 1077 atv.tv_sec = 0; 1078 atv.tv_usec = 0; 1079 timeout = 0; 1080 } 1081 marker = knote_alloc(1); 1082 if (marker == NULL) { 1083 error = ENOMEM; 1084 goto done_nl; 1085 } 1086 marker->kn_status = KN_MARKER; 1087 KQ_LOCK(kq); 1088 goto start; 1089 1090retry: 1091 if (atv.tv_sec || atv.tv_usec) { 1092 getmicrouptime(&rtv); 1093 if (timevalcmp(&rtv, &atv, >=)) 1094 goto done; 1095 ttv = atv; 1096 timevalsub(&ttv, &rtv); 1097 timeout = ttv.tv_sec > 24 * 60 * 60 ? 1098 24 * 60 * 60 * hz : tvtohz(&ttv); 1099 } 1100 1101start: 1102 kevp = keva; 1103 if (kq->kq_count == 0) { 1104 if (timeout < 0) { 1105 error = EWOULDBLOCK; 1106 } else { 1107 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal); 1108 kq->kq_state |= KQ_SLEEP; 1109 error = msleep(kq, &kq->kq_lock, PSOCK | PCATCH, 1110 "kqread", timeout); 1111 } 1112 if (error == 0) 1113 goto retry; 1114 /* don't restart after signals... */ 1115 if (error == ERESTART) 1116 error = EINTR; 1117 else if (error == EWOULDBLOCK) 1118 error = 0; 1119 goto done; 1120 } 1121 1122 TAILQ_INSERT_TAIL(&kq->kq_head, marker, kn_tqe); 1123 while (count) { 1124 KQ_OWNED(kq); 1125 kn = TAILQ_FIRST(&kq->kq_head); 1126 1127 if ((kn->kn_status == KN_MARKER && kn != marker) || 1128 (kn->kn_status & KN_INFLUX) == KN_INFLUX) { 1129 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal); 1130 kq->kq_state |= KQ_FLUXWAIT; 1131 error = msleep(kq, &kq->kq_lock, PSOCK, 1132 "kqflxwt", 0); 1133 continue; 1134 } 1135 1136 TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe); 1137 if ((kn->kn_status & KN_DISABLED) == KN_DISABLED) { 1138 kn->kn_status &= ~KN_QUEUED; 1139 kq->kq_count--; 1140 continue; 1141 } 1142 if (kn == marker) { 1143 KQ_FLUX_WAKEUP(kq); 1144 if (count == maxevents) 1145 goto retry; 1146 goto done; 1147 } 1148 KASSERT((kn->kn_status & KN_INFLUX) == 0, 1149 ("KN_INFLUX set when not suppose to be")); 1150 1151 if ((kn->kn_flags & EV_ONESHOT) == EV_ONESHOT) { 1152 kn->kn_status &= ~KN_QUEUED; 1153 kn->kn_status |= KN_INFLUX; 1154 kq->kq_count--; 1155 KQ_UNLOCK(kq); 1156 /* 1157 * We don't need to lock the list since we've marked 1158 * it _INFLUX. 1159 */ 1160 *kevp = kn->kn_kevent; 1161 if (!(kn->kn_status & KN_DETACHED)) 1162 kn->kn_fop->f_detach(kn); 1163 knote_drop(kn, td); 1164 KQ_LOCK(kq); 1165 kn = NULL; 1166 } else { 1167 kn->kn_status |= KN_INFLUX; 1168 KQ_UNLOCK(kq); 1169 if ((kn->kn_status & KN_KQUEUE) == KN_KQUEUE) 1170 KQ_GLOBAL_LOCK(&kq_global, haskqglobal); 1171 KN_LIST_LOCK(kn); 1172 if (kn->kn_fop->f_event(kn, 0) == 0) { 1173 KN_LIST_UNLOCK(kn); 1174 KQ_LOCK(kq); 1175 kn->kn_status &= 1176 ~(KN_QUEUED | KN_ACTIVE | KN_INFLUX); 1177 kq->kq_count--; 1178 continue; 1179 } 1180 *kevp = kn->kn_kevent; 1181 KQ_LOCK(kq); 1182 if (kn->kn_flags & EV_CLEAR) { 1183 kn->kn_data = 0; 1184 kn->kn_fflags = 0; 1185 kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE); 1186 kq->kq_count--; 1187 } else 1188 TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe); 1189 KN_LIST_UNLOCK(kn); 1190 kn->kn_status &= ~(KN_INFLUX); 1191 } 1192 1193 /* we are returning a copy to the user */ 1194 kevp++; 1195 nkev++; 1196 count--; 1197 1198 if (nkev == KQ_NEVENTS) { 1199 KQ_UNLOCK_FLUX(kq); 1200 error = copyout(keva, ulistp, sizeof *keva * nkev); 1201 ulistp += nkev; 1202 nkev = 0; 1203 kevp = keva; 1204 KQ_LOCK(kq); 1205 if (error) 1206 break; 1207 } 1208 } 1209 TAILQ_REMOVE(&kq->kq_head, marker, kn_tqe); 1210done: 1211 KQ_OWNED(kq); 1212 KQ_UNLOCK_FLUX(kq); 1213 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal); 1214 knote_free(marker); 1215done_nl: 1216 KQ_NOTOWNED(kq); 1217 if (nkev != 0) 1218 error = copyout(keva, ulistp, sizeof *keva * nkev); 1219 td->td_retval[0] = maxevents - count; 1220 return (error); 1221} 1222 1223/* 1224 * XXX 1225 * This could be expanded to call kqueue_scan, if desired. 1226 */ 1227/*ARGSUSED*/ 1228static int 1229kqueue_read(struct file *fp, struct uio *uio, struct ucred *active_cred, 1230 int flags, struct thread *td) 1231{ 1232 return (ENXIO); 1233} 1234 1235/*ARGSUSED*/ 1236static int 1237kqueue_write(struct file *fp, struct uio *uio, struct ucred *active_cred, 1238 int flags, struct thread *td) 1239{ 1240 return (ENXIO); 1241} 1242 1243/*ARGSUSED*/ 1244static int 1245kqueue_ioctl(struct file *fp, u_long cmd, void *data, 1246 struct ucred *active_cred, struct thread *td) 1247{ 1248 /* 1249 * Enabling sigio causes two major problems: 1250 * 1) infinite recursion: 1251 * Synopsys: kevent is being used to track signals and have FIOASYNC 1252 * set. On receipt of a signal this will cause a kqueue to recurse 1253 * into itself over and over. Sending the sigio causes the kqueue 1254 * to become ready, which in turn posts sigio again, forever. 1255 * Solution: this can be solved by setting a flag in the kqueue that 1256 * we have a SIGIO in progress. 1257 * 2) locking problems: 1258 * Synopsys: Kqueue is a leaf subsystem, but adding signalling puts 1259 * us above the proc and pgrp locks. 1260 * Solution: Post a signal using an async mechanism, being sure to 1261 * record a generation count in the delivery so that we do not deliver 1262 * a signal to the wrong process. 1263 * 1264 * Note, these two mechanisms are somewhat mutually exclusive! 1265 */ 1266#if 0 1267 struct kqueue *kq; 1268 1269 kq = fp->f_data; 1270 switch (cmd) { 1271 case FIOASYNC: 1272 if (*(int *)data) { 1273 kq->kq_state |= KQ_ASYNC; 1274 } else { 1275 kq->kq_state &= ~KQ_ASYNC; 1276 } 1277 return (0); 1278 1279 case FIOSETOWN: 1280 return (fsetown(*(int *)data, &kq->kq_sigio)); 1281 1282 case FIOGETOWN: 1283 *(int *)data = fgetown(&kq->kq_sigio); 1284 return (0); 1285 } 1286#endif 1287 1288 return (ENOTTY); 1289} 1290 1291/*ARGSUSED*/ 1292static int 1293kqueue_poll(struct file *fp, int events, struct ucred *active_cred, 1294 struct thread *td) 1295{ 1296 struct kqueue *kq; 1297 int revents = 0; 1298 int error; 1299 1300 if ((error = kqueue_aquire(fp, &kq))) 1301 return POLLERR; 1302 1303 KQ_LOCK(kq); 1304 if (events & (POLLIN | POLLRDNORM)) { 1305 if (kq->kq_count) { 1306 revents |= events & (POLLIN | POLLRDNORM); 1307 } else { 1308 selrecord(td, &kq->kq_sel); 1309 kq->kq_state |= KQ_SEL; 1310 } 1311 } 1312 kqueue_release(kq, 1); 1313 KQ_UNLOCK(kq); 1314 return (revents); 1315} 1316 1317/*ARGSUSED*/ 1318static int 1319kqueue_stat(struct file *fp, struct stat *st, struct ucred *active_cred, 1320 struct thread *td) 1321{ 1322 1323 return (ENXIO); 1324} 1325 1326/*ARGSUSED*/ 1327static int 1328kqueue_close(struct file *fp, struct thread *td) 1329{ 1330 struct kqueue *kq = fp->f_data; 1331 struct filedesc *fdp; 1332 struct knote *kn; 1333 int i; 1334 int error; 1335 1336 if ((error = kqueue_aquire(fp, &kq))) 1337 return error; 1338 1339 KQ_LOCK(kq); 1340 1341 KASSERT((kq->kq_state & KQ_CLOSING) != KQ_CLOSING, 1342 ("kqueue already closing")); 1343 kq->kq_state |= KQ_CLOSING; 1344 if (kq->kq_refcnt > 1) 1345 msleep(&kq->kq_refcnt, &kq->kq_lock, PSOCK, "kqclose", 0); 1346 1347 KASSERT(kq->kq_refcnt == 1, ("other refs are out there!")); 1348 fdp = kq->kq_fdp; 1349 1350 KASSERT(knlist_empty(&kq->kq_sel.si_note), 1351 ("kqueue's knlist not empty")); 1352 1353 for (i = 0; i < kq->kq_knlistsize; i++) { 1354 while ((kn = SLIST_FIRST(&kq->kq_knlist[i])) != NULL) { 1355 KASSERT((kn->kn_status & KN_INFLUX) == 0, 1356 ("KN_INFLUX set when not suppose to be")); 1357 kn->kn_status |= KN_INFLUX; 1358 KQ_UNLOCK(kq); 1359 if (!(kn->kn_status & KN_DETACHED)) 1360 kn->kn_fop->f_detach(kn); 1361 knote_drop(kn, td); 1362 KQ_LOCK(kq); 1363 } 1364 } 1365 if (kq->kq_knhashmask != 0) { 1366 for (i = 0; i <= kq->kq_knhashmask; i++) { 1367 while ((kn = SLIST_FIRST(&kq->kq_knhash[i])) != NULL) { 1368 KASSERT((kn->kn_status & KN_INFLUX) == 0, 1369 ("KN_INFLUX set when not suppose to be")); 1370 kn->kn_status |= KN_INFLUX; 1371 KQ_UNLOCK(kq); 1372 if (!(kn->kn_status & KN_DETACHED)) 1373 kn->kn_fop->f_detach(kn); 1374 knote_drop(kn, td); 1375 KQ_LOCK(kq); 1376 } 1377 } 1378 } 1379 1380 if ((kq->kq_state & KQ_TASKSCHED) == KQ_TASKSCHED) { 1381 kq->kq_state |= KQ_TASKDRAIN; 1382 msleep(&kq->kq_state, &kq->kq_lock, PSOCK, "kqtqdr", 0); 1383 } 1384 1385 if ((kq->kq_state & KQ_SEL) == KQ_SEL) { 1386 kq->kq_state &= ~KQ_SEL; 1387 selwakeuppri(&kq->kq_sel, PSOCK); 1388 } 1389 1390 KQ_UNLOCK(kq); 1391 1392 FILEDESC_LOCK(fdp); 1393 SLIST_REMOVE(&fdp->fd_kqlist, kq, kqueue, kq_list); 1394 FILEDESC_UNLOCK(fdp); 1395 1396 knlist_destroy(&kq->kq_sel.si_note); 1397 mtx_destroy(&kq->kq_lock); 1398 kq->kq_fdp = NULL; 1399 1400 if (kq->kq_knhash != NULL) 1401 free(kq->kq_knhash, M_KQUEUE); 1402 if (kq->kq_knlist != NULL) 1403 free(kq->kq_knlist, M_KQUEUE); 1404 1405 funsetown(&kq->kq_sigio); 1406 free(kq, M_KQUEUE); 1407 fp->f_data = NULL; 1408 1409 return (0); 1410} 1411 1412static void 1413kqueue_wakeup(struct kqueue *kq) 1414{ 1415 KQ_OWNED(kq); 1416 1417 if ((kq->kq_state & KQ_SLEEP) == KQ_SLEEP) { 1418 kq->kq_state &= ~KQ_SLEEP; 1419 wakeup(kq); 1420 } 1421 if ((kq->kq_state & KQ_SEL) == KQ_SEL) { 1422 kq->kq_state &= ~KQ_SEL; 1423 selwakeuppri(&kq->kq_sel, PSOCK); 1424 } 1425 if (!knlist_empty(&kq->kq_sel.si_note)) 1426 kqueue_schedtask(kq); 1427 if ((kq->kq_state & KQ_ASYNC) == KQ_ASYNC) { 1428 pgsigio(&kq->kq_sigio, SIGIO, 0); 1429 } 1430} 1431 1432/* 1433 * Walk down a list of knotes, activating them if their event has triggered. 1434 * 1435 * There is a possibility to optimize in the case of one kq watching another. 1436 * Instead of scheduling a task to wake it up, you could pass enough state 1437 * down the chain to make up the parent kqueue. Make this code functional 1438 * first. 1439 */ 1440void 1441knote(struct knlist *list, long hint, int islocked) 1442{ 1443 struct kqueue *kq; 1444 struct knote *kn; 1445 1446 if (list == NULL) 1447 return; 1448 1449 mtx_assert(list->kl_lock, islocked ? MA_OWNED : MA_NOTOWNED); 1450 if (!islocked) 1451 mtx_lock(list->kl_lock); 1452 /* 1453 * If we unlock the list lock (and set KN_INFLUX), we can eliminate 1454 * the kqueue scheduling, but this will introduce four 1455 * lock/unlock's for each knote to test. If we do, continue to use 1456 * SLIST_FOREACH, SLIST_FOREACH_SAFE is not safe in our case, it is 1457 * only safe if you want to remove the current item, which we are 1458 * not doing. 1459 */ 1460 SLIST_FOREACH(kn, &list->kl_list, kn_selnext) { 1461 kq = kn->kn_kq; 1462 if ((kn->kn_status & KN_INFLUX) != KN_INFLUX) { 1463 KQ_LOCK(kq); 1464 if ((kn->kn_status & KN_INFLUX) != KN_INFLUX) { 1465 kn->kn_status |= KN_HASKQLOCK; 1466 if (kn->kn_fop->f_event(kn, hint)) 1467 KNOTE_ACTIVATE(kn, 1); 1468 kn->kn_status &= ~KN_HASKQLOCK; 1469 } 1470 KQ_UNLOCK(kq); 1471 } 1472 kq = NULL; 1473 } 1474 if (!islocked) 1475 mtx_unlock(list->kl_lock); 1476} 1477 1478/* 1479 * add a knote to a knlist 1480 */ 1481void 1482knlist_add(struct knlist *knl, struct knote *kn, int islocked) 1483{ 1484 mtx_assert(knl->kl_lock, islocked ? MA_OWNED : MA_NOTOWNED); 1485 KQ_NOTOWNED(kn->kn_kq); 1486 KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) == 1487 (KN_INFLUX|KN_DETACHED), ("knote not KN_INFLUX and KN_DETACHED")); 1488 if (!islocked) 1489 mtx_lock(knl->kl_lock); 1490 SLIST_INSERT_HEAD(&knl->kl_list, kn, kn_selnext); 1491 if (!islocked) 1492 mtx_unlock(knl->kl_lock); 1493 KQ_LOCK(kn->kn_kq); 1494 kn->kn_knlist = knl; 1495 kn->kn_status &= ~KN_DETACHED; 1496 KQ_UNLOCK(kn->kn_kq); 1497} 1498 1499static void 1500knlist_remove_kq(struct knlist *knl, struct knote *kn, int knlislocked, int kqislocked) 1501{ 1502 KASSERT(!(!!kqislocked && !knlislocked), ("kq locked w/o knl locked")); 1503 mtx_assert(knl->kl_lock, knlislocked ? MA_OWNED : MA_NOTOWNED); 1504 mtx_assert(&kn->kn_kq->kq_lock, kqislocked ? MA_OWNED : MA_NOTOWNED); 1505 if (!kqislocked) 1506 KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) == KN_INFLUX, 1507 ("knlist_remove called w/o knote being KN_INFLUX or already removed")); 1508 if (!knlislocked) 1509 mtx_lock(knl->kl_lock); 1510 SLIST_REMOVE(&knl->kl_list, kn, knote, kn_selnext); 1511 kn->kn_knlist = NULL; 1512 if (!knlislocked) 1513 mtx_unlock(knl->kl_lock); 1514 if (!kqislocked) 1515 KQ_LOCK(kn->kn_kq); 1516 kn->kn_status |= KN_DETACHED; 1517 if (!kqislocked) 1518 KQ_UNLOCK(kn->kn_kq); 1519} 1520 1521/* 1522 * remove all knotes from a specified klist 1523 */ 1524void 1525knlist_remove(struct knlist *knl, struct knote *kn, int islocked) 1526{ 1527 1528 knlist_remove_kq(knl, kn, islocked, 0); 1529} 1530 1531/* 1532 * remove knote from a specified klist while in f_event handler. 1533 */ 1534void 1535knlist_remove_inevent(struct knlist *knl, struct knote *kn) 1536{ 1537 1538 knlist_remove_kq(knl, kn, 1, 1539 (kn->kn_status & KN_HASKQLOCK) == KN_HASKQLOCK); 1540} 1541 1542int 1543knlist_empty(struct knlist *knl) 1544{ 1545 1546 mtx_assert(knl->kl_lock, MA_OWNED); 1547 return SLIST_EMPTY(&knl->kl_list); 1548} 1549 1550static struct mtx knlist_lock; 1551MTX_SYSINIT(knlist_lock, &knlist_lock, "knlist lock for lockless objects", 1552 MTX_DEF); 1553 1554void 1555knlist_init(struct knlist *knl, struct mtx *mtx) 1556{ 1557 1558 if (mtx == NULL) 1559 knl->kl_lock = &knlist_lock; 1560 else 1561 knl->kl_lock = mtx; 1562 1563 SLIST_INIT(&knl->kl_list); 1564} 1565 1566void 1567knlist_destroy(struct knlist *knl) 1568{ 1569 1570#ifdef INVARIANTS 1571 /* 1572 * if we run across this error, we need to find the offending 1573 * driver and have it call knlist_clear. 1574 */ 1575 if (!SLIST_EMPTY(&knl->kl_list)) 1576 printf("WARNING: destroying knlist w/ knotes on it!\n"); 1577#endif 1578 1579 knl->kl_lock = NULL; 1580 SLIST_INIT(&knl->kl_list); 1581} 1582 1583/* 1584 * Even if we are locked, we may need to drop the lock to allow any influx 1585 * knotes time to "settle". 1586 */ 1587void 1588knlist_clear(struct knlist *knl, int islocked) 1589{ 1590 struct knote *kn; 1591 struct kqueue *kq; 1592 1593 if (islocked) 1594 mtx_assert(knl->kl_lock, MA_OWNED); 1595 else { 1596 mtx_assert(knl->kl_lock, MA_NOTOWNED); 1597again: /* need to reaquire lock since we have dropped it */ 1598 mtx_lock(knl->kl_lock); 1599 } 1600 1601 SLIST_FOREACH(kn, &knl->kl_list, kn_selnext) { 1602 kq = kn->kn_kq; 1603 KQ_LOCK(kq); 1604 if ((kn->kn_status & KN_INFLUX) && 1605 (kn->kn_status & KN_DETACHED) != KN_DETACHED) { 1606 KQ_UNLOCK(kq); 1607 continue; 1608 } 1609 /* Make sure cleared knotes disappear soon */ 1610 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 1611 knlist_remove_kq(knl, kn, 1, 1); 1612 KQ_UNLOCK(kq); 1613 kq = NULL; 1614 } 1615 1616 if (!SLIST_EMPTY(&knl->kl_list)) { 1617 /* there are still KN_INFLUX remaining */ 1618 kn = SLIST_FIRST(&knl->kl_list); 1619 kq = kn->kn_kq; 1620 KQ_LOCK(kq); 1621 KASSERT(kn->kn_status & KN_INFLUX, 1622 ("knote removed w/o list lock")); 1623 mtx_unlock(knl->kl_lock); 1624 kq->kq_state |= KQ_FLUXWAIT; 1625 msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqkclr", 0); 1626 kq = NULL; 1627 goto again; 1628 } 1629 1630 SLIST_INIT(&knl->kl_list); 1631 1632 if (islocked) 1633 mtx_assert(knl->kl_lock, MA_OWNED); 1634 else { 1635 mtx_unlock(knl->kl_lock); 1636 mtx_assert(knl->kl_lock, MA_NOTOWNED); 1637 } 1638} 1639 1640/* 1641 * remove all knotes referencing a specified fd 1642 * must be called with FILEDESC lock. This prevents a race where a new fd 1643 * comes along and occupies the entry and we attach a knote to the fd. 1644 */ 1645void 1646knote_fdclose(struct thread *td, int fd) 1647{ 1648 struct filedesc *fdp = td->td_proc->p_fd; 1649 struct kqueue *kq; 1650 struct knote *kn; 1651 int influx; 1652 1653 FILEDESC_LOCK_ASSERT(fdp, MA_OWNED); 1654 1655 /* 1656 * We shouldn't have to worry about new kevents appearing on fd 1657 * since filedesc is locked. 1658 */ 1659 SLIST_FOREACH(kq, &fdp->fd_kqlist, kq_list) { 1660 KQ_LOCK(kq); 1661 1662again: 1663 influx = 0; 1664 while (kq->kq_knlistsize > fd && 1665 (kn = SLIST_FIRST(&kq->kq_knlist[fd])) != NULL) { 1666 if (kn->kn_status & KN_INFLUX) { 1667 /* someone else might be waiting on our knote */ 1668 if (influx) 1669 wakeup(kq); 1670 kq->kq_state |= KQ_FLUXWAIT; 1671 msleep(kq, &kq->kq_lock, PSOCK, "kqflxwt", 0); 1672 goto again; 1673 } 1674 kn->kn_status |= KN_INFLUX; 1675 KQ_UNLOCK(kq); 1676 if (!(kn->kn_status & KN_DETACHED)) 1677 kn->kn_fop->f_detach(kn); 1678 knote_drop(kn, td); 1679 influx = 1; 1680 KQ_LOCK(kq); 1681 } 1682 KQ_UNLOCK_FLUX(kq); 1683 } 1684} 1685 1686static int 1687knote_attach(struct knote *kn, struct kqueue *kq) 1688{ 1689 struct klist *list; 1690 1691 KASSERT(kn->kn_status & KN_INFLUX, ("knote not marked INFLUX")); 1692 KQ_OWNED(kq); 1693 1694 if (kn->kn_fop->f_isfd) { 1695 if (kn->kn_id >= kq->kq_knlistsize) 1696 return ENOMEM; 1697 list = &kq->kq_knlist[kn->kn_id]; 1698 } else { 1699 if (kq->kq_knhash == NULL) 1700 return ENOMEM; 1701 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)]; 1702 } 1703 1704 SLIST_INSERT_HEAD(list, kn, kn_link); 1705 1706 return 0; 1707} 1708 1709/* 1710 * knote must already have been detatched using the f_detach method. 1711 * no lock need to be held, it is assumed that the KN_INFLUX flag is set 1712 * to prevent other removal. 1713 */ 1714static void 1715knote_drop(struct knote *kn, struct thread *td) 1716{ 1717 struct kqueue *kq; 1718 struct klist *list; 1719 1720 kq = kn->kn_kq; 1721 1722 KQ_NOTOWNED(kq); 1723 KASSERT((kn->kn_status & KN_INFLUX) == KN_INFLUX, 1724 ("knote_drop called without KN_INFLUX set in kn_status")); 1725 1726 KQ_LOCK(kq); 1727 if (kn->kn_fop->f_isfd) 1728 list = &kq->kq_knlist[kn->kn_id]; 1729 else 1730 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)]; 1731 1732 SLIST_REMOVE(list, kn, knote, kn_link); 1733 if (kn->kn_status & KN_QUEUED) 1734 knote_dequeue(kn); 1735 KQ_UNLOCK_FLUX(kq); 1736 1737 if (kn->kn_fop->f_isfd) { 1738 fdrop(kn->kn_fp, td); 1739 kn->kn_fp = NULL; 1740 } 1741 kqueue_fo_release(kn->kn_kevent.filter); 1742 kn->kn_fop = NULL; 1743 knote_free(kn); 1744} 1745 1746static void 1747knote_enqueue(struct knote *kn) 1748{ 1749 struct kqueue *kq = kn->kn_kq; 1750 1751 KQ_OWNED(kn->kn_kq); 1752 KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued")); 1753 1754 TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe); 1755 kn->kn_status |= KN_QUEUED; 1756 kq->kq_count++; 1757 kqueue_wakeup(kq); 1758} 1759 1760static void 1761knote_dequeue(struct knote *kn) 1762{ 1763 struct kqueue *kq = kn->kn_kq; 1764 1765 KQ_OWNED(kn->kn_kq); 1766 KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued")); 1767 1768 TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe); 1769 kn->kn_status &= ~KN_QUEUED; 1770 kq->kq_count--; 1771} 1772 1773static void 1774knote_init(void) 1775{ 1776 1777 knote_zone = uma_zcreate("KNOTE", sizeof(struct knote), NULL, NULL, 1778 NULL, NULL, UMA_ALIGN_PTR, 0); 1779} 1780SYSINIT(knote, SI_SUB_PSEUDO, SI_ORDER_ANY, knote_init, NULL) 1781 1782static struct knote * 1783knote_alloc(int waitok) 1784{ 1785 return ((struct knote *)uma_zalloc(knote_zone, 1786 (waitok ? M_WAITOK : M_NOWAIT)|M_ZERO)); 1787} 1788 1789static void 1790knote_free(struct knote *kn) 1791{ 1792 if (kn != NULL) 1793 uma_zfree(knote_zone, kn); 1794} 1795