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(void *arg, struct kevent *kevp, int count); 85 static int kevent_copyin(void *arg, 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_copyops *k_ops, 93 const struct timespec *timeout, 94 struct kevent *keva, 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 kn->kn_knlist->kl_lock(kn->kn_knlist->kl_lockarg); \ 189 } while (0) 190 #define KN_LIST_UNLOCK(kn) do { \ 191 if (kn->kn_knlist != NULL) \ 192 kn->kn_knlist->kl_unlock(kn->kn_knlist->kl_lockarg); \ 193 } while (0) 194 #define KNL_ASSERT_LOCK(knl, islocked) do { \ 195 if (islocked) \ 196 KNL_ASSERT_LOCKED(knl); \ 197 else \ 198 KNL_ASSERT_UNLOCKED(knl); \ 199 } while (0) 200 #ifdef INVARIANTS 201 #define KNL_ASSERT_LOCKED(knl) do { \ 202 if (!knl->kl_locked((knl)->kl_lockarg)) \ 203 panic("knlist not locked, but should be"); \ 204 } while (0) 205 #define KNL_ASSERT_UNLOCKED(knl) do { \ 206 if (knl->kl_locked((knl)->kl_lockarg)) \ 207 panic("knlist locked, but should not be"); \ 208 } while (0) 209 #else /* !INVARIANTS */ 210 #define KNL_ASSERT_LOCKED(knl) do {} while(0) 211 #define KNL_ASSERT_UNLOCKED(knl) do {} while (0) 212 #endif /* INVARIANTS */ 213 214 #define KN_HASHSIZE 64 /* XXX should be tunable */ 215 #define KN_HASH(val, mask) (((val) ^ (val >> 8)) & (mask)) 216 217 static int 218 filt_nullattach(struct knote *kn) 219 { 220 221 return (ENXIO); 222 }; 223 224 struct filterops null_filtops = 225 { 0, filt_nullattach, NULL, NULL }; 226 227 /* XXX - make SYSINIT to add these, and move into respective modules. */ 228 extern struct filterops sig_filtops; 229 extern struct filterops fs_filtops; 230 231 /* 232 * Table for for all system-defined filters. 233 */ 234 static struct mtx filterops_lock; 235 MTX_SYSINIT(kqueue_filterops, &filterops_lock, "protect sysfilt_ops", 236 MTX_DEF); 237 static struct { 238 struct filterops *for_fop; 239 int for_refcnt; 240 } sysfilt_ops[EVFILT_SYSCOUNT] = { 241 { &file_filtops }, /* EVFILT_READ */ 242 { &file_filtops }, /* EVFILT_WRITE */ 243 { &null_filtops }, /* EVFILT_AIO */ 244 { &file_filtops }, /* EVFILT_VNODE */ 245 { &proc_filtops }, /* EVFILT_PROC */ 246 { &sig_filtops }, /* EVFILT_SIGNAL */ 247 { &timer_filtops }, /* EVFILT_TIMER */ 248 { &file_filtops }, /* EVFILT_NETDEV */ 249 { &fs_filtops }, /* EVFILT_FS */ 250 }; 251 252 /* 253 * Simple redirection for all cdevsw style objects to call their fo_kqfilter 254 * method. 255 */ 256 static int 257 filt_fileattach(struct knote *kn) 258 { 259 260 return (fo_kqfilter(kn->kn_fp, kn)); 261 } 262 263 /*ARGSUSED*/ 264 static int 265 kqueue_kqfilter(struct file *fp, struct knote *kn) 266 { 267 struct kqueue *kq = kn->kn_fp->f_data; 268 269 if (kn->kn_filter != EVFILT_READ) 270 return (EINVAL); 271 272 kn->kn_status |= KN_KQUEUE; 273 kn->kn_fop = &kqread_filtops; 274 knlist_add(&kq->kq_sel.si_note, kn, 0); 275 276 return (0); 277 } 278 279 static void 280 filt_kqdetach(struct knote *kn) 281 { 282 struct kqueue *kq = kn->kn_fp->f_data; 283 284 knlist_remove(&kq->kq_sel.si_note, kn, 0); 285 } 286 287 /*ARGSUSED*/ 288 static int 289 filt_kqueue(struct knote *kn, long hint) 290 { 291 struct kqueue *kq = kn->kn_fp->f_data; 292 293 kn->kn_data = kq->kq_count; 294 return (kn->kn_data > 0); 295 } 296 297 /* XXX - move to kern_proc.c? */ 298 static int 299 filt_procattach(struct knote *kn) 300 { 301 struct proc *p; 302 int immediate; 303 int error; 304 305 immediate = 0; 306 p = pfind(kn->kn_id); 307 if (p == NULL && (kn->kn_sfflags & NOTE_EXIT)) { 308 p = zpfind(kn->kn_id); 309 immediate = 1; 310 } else if (p != NULL && (p->p_flag & P_WEXIT)) { 311 immediate = 1; 312 } 313 314 if (p == NULL) 315 return (ESRCH); 316 if ((error = p_cansee(curthread, p))) 317 return (error); 318 319 kn->kn_ptr.p_proc = p; 320 kn->kn_flags |= EV_CLEAR; /* automatically set */ 321 322 /* 323 * internal flag indicating registration done by kernel 324 */ 325 if (kn->kn_flags & EV_FLAG1) { 326 kn->kn_data = kn->kn_sdata; /* ppid */ 327 kn->kn_fflags = NOTE_CHILD; 328 kn->kn_flags &= ~EV_FLAG1; 329 } 330 331 if (immediate == 0) 332 knlist_add(&p->p_klist, kn, 1); 333 334 /* 335 * Immediately activate any exit notes if the target process is a 336 * zombie. This is necessary to handle the case where the target 337 * process, e.g. a child, dies before the kevent is registered. 338 */ 339 if (immediate && filt_proc(kn, NOTE_EXIT)) 340 KNOTE_ACTIVATE(kn, 0); 341 342 PROC_UNLOCK(p); 343 344 return (0); 345 } 346 347 /* 348 * The knote may be attached to a different process, which may exit, 349 * leaving nothing for the knote to be attached to. So when the process 350 * exits, the knote is marked as DETACHED and also flagged as ONESHOT so 351 * it will be deleted when read out. However, as part of the knote deletion, 352 * this routine is called, so a check is needed to avoid actually performing 353 * a detach, because the original process does not exist any more. 354 */ 355 /* XXX - move to kern_proc.c? */ 356 static void 357 filt_procdetach(struct knote *kn) 358 { 359 struct proc *p; 360 361 p = kn->kn_ptr.p_proc; 362 knlist_remove(&p->p_klist, kn, 0); 363 kn->kn_ptr.p_proc = NULL; 364 } 365 366 /* XXX - move to kern_proc.c? */ 367 static int 368 filt_proc(struct knote *kn, long hint) 369 { 370 struct proc *p = kn->kn_ptr.p_proc; 371 u_int event; 372 373 /* 374 * mask off extra data 375 */ 376 event = (u_int)hint & NOTE_PCTRLMASK; 377 378 /* 379 * if the user is interested in this event, record it. 380 */ 381 if (kn->kn_sfflags & event) 382 kn->kn_fflags |= event; 383 384 /* 385 * process is gone, so flag the event as finished. 386 */ 387 if (event == NOTE_EXIT) { 388 if (!(kn->kn_status & KN_DETACHED)) 389 knlist_remove_inevent(&p->p_klist, kn); 390 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 391 kn->kn_ptr.p_proc = NULL; 392 return (1); 393 } 394 395 /* 396 * process forked, and user wants to track the new process, 397 * so attach a new knote to it, and immediately report an 398 * event with the parent's pid. 399 */ 400 if ((event == NOTE_FORK) && (kn->kn_sfflags & NOTE_TRACK)) { 401 struct kevent kev; 402 int error; 403 404 /* 405 * register knote with new process. 406 */ 407 kev.ident = hint & NOTE_PDATAMASK; /* pid */ 408 kev.filter = kn->kn_filter; 409 kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1; 410 kev.fflags = kn->kn_sfflags; 411 kev.data = kn->kn_id; /* parent */ 412 kev.udata = kn->kn_kevent.udata; /* preserve udata */ 413 error = kqueue_register(kn->kn_kq, &kev, NULL, 0); 414 if (error) 415 kn->kn_fflags |= NOTE_TRACKERR; 416 } 417 418 return (kn->kn_fflags != 0); 419 } 420 421 static int 422 timertoticks(intptr_t data) 423 { 424 struct timeval tv; 425 int tticks; 426 427 tv.tv_sec = data / 1000; 428 tv.tv_usec = (data % 1000) * 1000; 429 tticks = tvtohz(&tv); 430 431 return tticks; 432 } 433 434 /* XXX - move to kern_timeout.c? */ 435 static void 436 filt_timerexpire(void *knx) 437 { 438 struct knote *kn = knx; 439 struct callout *calloutp; 440 441 kn->kn_data++; 442 KNOTE_ACTIVATE(kn, 0); /* XXX - handle locking */ 443 444 if ((kn->kn_flags & EV_ONESHOT) != EV_ONESHOT) { 445 calloutp = (struct callout *)kn->kn_hook; 446 callout_reset(calloutp, timertoticks(kn->kn_sdata), 447 filt_timerexpire, kn); 448 } 449 } 450 451 /* 452 * data contains amount of time to sleep, in milliseconds 453 */ 454 /* XXX - move to kern_timeout.c? */ 455 static int 456 filt_timerattach(struct knote *kn) 457 { 458 struct callout *calloutp; 459 460 atomic_add_int(&kq_ncallouts, 1); 461 462 if (kq_ncallouts >= kq_calloutmax) { 463 atomic_add_int(&kq_ncallouts, -1); 464 return (ENOMEM); 465 } 466 467 kn->kn_flags |= EV_CLEAR; /* automatically set */ 468 kn->kn_status &= ~KN_DETACHED; /* knlist_add usually sets it */ 469 MALLOC(calloutp, struct callout *, sizeof(*calloutp), 470 M_KQUEUE, M_WAITOK); 471 callout_init(calloutp, CALLOUT_MPSAFE); 472 kn->kn_hook = calloutp; 473 callout_reset(calloutp, timertoticks(kn->kn_sdata), filt_timerexpire, 474 kn); 475 476 return (0); 477 } 478 479 /* XXX - move to kern_timeout.c? */ 480 static void 481 filt_timerdetach(struct knote *kn) 482 { 483 struct callout *calloutp; 484 485 calloutp = (struct callout *)kn->kn_hook; 486 callout_drain(calloutp); 487 FREE(calloutp, M_KQUEUE); 488 atomic_add_int(&kq_ncallouts, -1); 489 kn->kn_status |= KN_DETACHED; /* knlist_remove usually clears it */ 490 } 491 492 /* XXX - move to kern_timeout.c? */ 493 static int 494 filt_timer(struct knote *kn, long hint) 495 { 496 497 return (kn->kn_data != 0); 498 } 499 500 /* 501 * MPSAFE 502 */ 503 int 504 kqueue(struct thread *td, struct kqueue_args *uap) 505 { 506 struct filedesc *fdp; 507 struct kqueue *kq; 508 struct file *fp; 509 int fd, error; 510 511 fdp = td->td_proc->p_fd; 512 error = falloc(td, &fp, &fd); 513 if (error) 514 goto done2; 515 516 /* An extra reference on `nfp' has been held for us by falloc(). */ 517 kq = malloc(sizeof *kq, M_KQUEUE, M_WAITOK | M_ZERO); 518 mtx_init(&kq->kq_lock, "kqueue", NULL, MTX_DEF|MTX_DUPOK); 519 TAILQ_INIT(&kq->kq_head); 520 kq->kq_fdp = fdp; 521 knlist_init(&kq->kq_sel.si_note, &kq->kq_lock, NULL, NULL, NULL); 522 TASK_INIT(&kq->kq_task, 0, kqueue_task, kq); 523 524 FILEDESC_LOCK_FAST(fdp); 525 SLIST_INSERT_HEAD(&fdp->fd_kqlist, kq, kq_list); 526 FILEDESC_UNLOCK_FAST(fdp); 527 528 FILE_LOCK(fp); 529 fp->f_flag = FREAD | FWRITE; 530 fp->f_type = DTYPE_KQUEUE; 531 fp->f_ops = &kqueueops; 532 fp->f_data = kq; 533 FILE_UNLOCK(fp); 534 fdrop(fp, td); 535 536 td->td_retval[0] = fd; 537 done2: 538 return (error); 539 } 540 541 #ifndef _SYS_SYSPROTO_H_ 542 struct kevent_args { 543 int fd; 544 const struct kevent *changelist; 545 int nchanges; 546 struct kevent *eventlist; 547 int nevents; 548 const struct timespec *timeout; 549 }; 550 #endif 551 /* 552 * MPSAFE 553 */ 554 int 555 kevent(struct thread *td, struct kevent_args *uap) 556 { 557 struct timespec ts, *tsp; 558 struct kevent_copyops k_ops = { uap, 559 kevent_copyout, 560 kevent_copyin}; 561 int error; 562 563 if (uap->timeout != NULL) { 564 error = copyin(uap->timeout, &ts, sizeof(ts)); 565 if (error) 566 return (error); 567 tsp = &ts; 568 } else 569 tsp = NULL; 570 571 return (kern_kevent(td, uap->fd, uap->nchanges, uap->nevents, 572 &k_ops, tsp)); 573 } 574 575 /* 576 * Copy 'count' items into the destination list pointed to by uap->eventlist. 577 */ 578 static int 579 kevent_copyout(void *arg, struct kevent *kevp, int count) 580 { 581 struct kevent_args *uap; 582 int error; 583 584 KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count)); 585 uap = (struct kevent_args *)arg; 586 587 error = copyout(kevp, uap->eventlist, count * sizeof *kevp); 588 if (error == 0) 589 uap->eventlist += count; 590 return (error); 591 } 592 593 /* 594 * Copy 'count' items from the list pointed to by uap->changelist. 595 */ 596 static int 597 kevent_copyin(void *arg, struct kevent *kevp, int count) 598 { 599 struct kevent_args *uap; 600 int error; 601 602 KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count)); 603 uap = (struct kevent_args *)arg; 604 605 error = copyin(uap->changelist, kevp, count * sizeof *kevp); 606 if (error == 0) 607 uap->changelist += count; 608 return (error); 609 } 610 611 int 612 kern_kevent(struct thread *td, int fd, int nchanges, int nevents, 613 struct kevent_copyops *k_ops, const struct timespec *timeout) 614 { 615 struct kevent keva[KQ_NEVENTS]; 616 struct kevent *kevp, *changes; 617 struct kqueue *kq; 618 struct file *fp; 619 int i, n, nerrors, error; 620 621 if ((error = fget(td, fd, &fp)) != 0) 622 return (error); 623 if ((error = kqueue_aquire(fp, &kq)) != 0) 624 goto done_norel; 625 626 nerrors = 0; 627 628 while (nchanges > 0) { 629 n = nchanges > KQ_NEVENTS ? KQ_NEVENTS : nchanges; 630 error = k_ops->k_copyin(k_ops->arg, keva, n); 631 if (error) 632 goto done; 633 changes = keva; 634 for (i = 0; i < n; i++) { 635 kevp = &changes[i]; 636 kevp->flags &= ~EV_SYSFLAGS; 637 error = kqueue_register(kq, kevp, td, 1); 638 if (error) { 639 if (nevents != 0) { 640 kevp->flags = EV_ERROR; 641 kevp->data = error; 642 (void) k_ops->k_copyout(k_ops->arg, 643 kevp, 1); 644 nevents--; 645 nerrors++; 646 } else { 647 goto done; 648 } 649 } 650 } 651 nchanges -= n; 652 } 653 if (nerrors) { 654 td->td_retval[0] = nerrors; 655 error = 0; 656 goto done; 657 } 658 659 error = kqueue_scan(kq, nevents, k_ops, timeout, keva, td); 660 done: 661 kqueue_release(kq, 0); 662 done_norel: 663 if (fp != NULL) 664 fdrop(fp, td); 665 return (error); 666 } 667 668 int 669 kqueue_add_filteropts(int filt, struct filterops *filtops) 670 { 671 int error; 672 673 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) { 674 printf( 675 "trying to add a filterop that is out of range: %d is beyond %d\n", 676 ~filt, EVFILT_SYSCOUNT); 677 return EINVAL; 678 } 679 mtx_lock(&filterops_lock); 680 if (sysfilt_ops[~filt].for_fop != &null_filtops && 681 sysfilt_ops[~filt].for_fop != NULL) 682 error = EEXIST; 683 else { 684 sysfilt_ops[~filt].for_fop = filtops; 685 sysfilt_ops[~filt].for_refcnt = 0; 686 } 687 mtx_unlock(&filterops_lock); 688 689 return (0); 690 } 691 692 int 693 kqueue_del_filteropts(int filt) 694 { 695 int error; 696 697 error = 0; 698 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) 699 return EINVAL; 700 701 mtx_lock(&filterops_lock); 702 if (sysfilt_ops[~filt].for_fop == &null_filtops || 703 sysfilt_ops[~filt].for_fop == NULL) 704 error = EINVAL; 705 else if (sysfilt_ops[~filt].for_refcnt != 0) 706 error = EBUSY; 707 else { 708 sysfilt_ops[~filt].for_fop = &null_filtops; 709 sysfilt_ops[~filt].for_refcnt = 0; 710 } 711 mtx_unlock(&filterops_lock); 712 713 return error; 714 } 715 716 static struct filterops * 717 kqueue_fo_find(int filt) 718 { 719 720 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) 721 return NULL; 722 723 mtx_lock(&filterops_lock); 724 sysfilt_ops[~filt].for_refcnt++; 725 if (sysfilt_ops[~filt].for_fop == NULL) 726 sysfilt_ops[~filt].for_fop = &null_filtops; 727 mtx_unlock(&filterops_lock); 728 729 return sysfilt_ops[~filt].for_fop; 730 } 731 732 static void 733 kqueue_fo_release(int filt) 734 { 735 736 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) 737 return; 738 739 mtx_lock(&filterops_lock); 740 KASSERT(sysfilt_ops[~filt].for_refcnt > 0, 741 ("filter object refcount not valid on release")); 742 sysfilt_ops[~filt].for_refcnt--; 743 mtx_unlock(&filterops_lock); 744 } 745 746 /* 747 * A ref to kq (obtained via kqueue_aquire) should be held. waitok will 748 * influence if memory allocation should wait. Make sure it is 0 if you 749 * hold any mutexes. 750 */ 751 int 752 kqueue_register(struct kqueue *kq, struct kevent *kev, struct thread *td, int waitok) 753 { 754 struct filedesc *fdp; 755 struct filterops *fops; 756 struct file *fp; 757 struct knote *kn, *tkn; 758 int error, filt, event; 759 int haskqglobal; 760 int fd; 761 762 fdp = NULL; 763 fp = NULL; 764 kn = NULL; 765 error = 0; 766 haskqglobal = 0; 767 768 filt = kev->filter; 769 fops = kqueue_fo_find(filt); 770 if (fops == NULL) 771 return EINVAL; 772 773 tkn = knote_alloc(waitok); /* prevent waiting with locks */ 774 775 findkn: 776 if (fops->f_isfd) { 777 KASSERT(td != NULL, ("td is NULL")); 778 fdp = td->td_proc->p_fd; 779 FILEDESC_LOCK(fdp); 780 /* validate descriptor */ 781 fd = kev->ident; 782 if (fd < 0 || fd >= fdp->fd_nfiles || 783 (fp = fdp->fd_ofiles[fd]) == NULL) { 784 FILEDESC_UNLOCK(fdp); 785 error = EBADF; 786 goto done; 787 } 788 fhold(fp); 789 790 if ((kev->flags & EV_ADD) == EV_ADD && kqueue_expand(kq, fops, 791 kev->ident, 0) != 0) { 792 /* unlock and try again */ 793 FILEDESC_UNLOCK(fdp); 794 fdrop(fp, td); 795 fp = NULL; 796 error = kqueue_expand(kq, fops, kev->ident, waitok); 797 if (error) 798 goto done; 799 goto findkn; 800 } 801 802 if (fp->f_type == DTYPE_KQUEUE) { 803 /* 804 * if we add some inteligence about what we are doing, 805 * we should be able to support events on ourselves. 806 * We need to know when we are doing this to prevent 807 * getting both the knlist lock and the kq lock since 808 * they are the same thing. 809 */ 810 if (fp->f_data == kq) { 811 FILEDESC_UNLOCK(fdp); 812 error = EINVAL; 813 goto done_noglobal; 814 } 815 816 KQ_GLOBAL_LOCK(&kq_global, haskqglobal); 817 } 818 819 FILEDESC_UNLOCK(fdp); 820 KQ_LOCK(kq); 821 if (kev->ident < kq->kq_knlistsize) { 822 SLIST_FOREACH(kn, &kq->kq_knlist[kev->ident], kn_link) 823 if (kev->filter == kn->kn_filter) 824 break; 825 } 826 } else { 827 if ((kev->flags & EV_ADD) == EV_ADD) 828 kqueue_expand(kq, fops, kev->ident, waitok); 829 830 KQ_LOCK(kq); 831 if (kq->kq_knhashmask != 0) { 832 struct klist *list; 833 834 list = &kq->kq_knhash[ 835 KN_HASH((u_long)kev->ident, kq->kq_knhashmask)]; 836 SLIST_FOREACH(kn, list, kn_link) 837 if (kev->ident == kn->kn_id && 838 kev->filter == kn->kn_filter) 839 break; 840 } 841 } 842 843 /* knote is in the process of changing, wait for it to stablize. */ 844 if (kn != NULL && (kn->kn_status & KN_INFLUX) == KN_INFLUX) { 845 if (fp != NULL) { 846 fdrop(fp, td); 847 fp = NULL; 848 } 849 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal); 850 kq->kq_state |= KQ_FLUXWAIT; 851 msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqflxwt", 0); 852 goto findkn; 853 } 854 855 if (kn == NULL && ((kev->flags & EV_ADD) == 0)) { 856 KQ_UNLOCK(kq); 857 error = ENOENT; 858 goto done; 859 } 860 861 /* 862 * kn now contains the matching knote, or NULL if no match 863 */ 864 if (kev->flags & EV_ADD) { 865 if (kn == NULL) { 866 kn = tkn; 867 tkn = NULL; 868 if (kn == NULL) { 869 error = ENOMEM; 870 goto done; 871 } 872 kn->kn_fp = fp; 873 kn->kn_kq = kq; 874 kn->kn_fop = fops; 875 /* 876 * apply reference counts to knote structure, and 877 * do not release it at the end of this routine. 878 */ 879 fops = NULL; 880 fp = NULL; 881 882 kn->kn_sfflags = kev->fflags; 883 kn->kn_sdata = kev->data; 884 kev->fflags = 0; 885 kev->data = 0; 886 kn->kn_kevent = *kev; 887 kn->kn_status = KN_INFLUX|KN_DETACHED; 888 889 error = knote_attach(kn, kq); 890 KQ_UNLOCK(kq); 891 if (error != 0) { 892 tkn = kn; 893 goto done; 894 } 895 896 if ((error = kn->kn_fop->f_attach(kn)) != 0) { 897 knote_drop(kn, td); 898 goto done; 899 } 900 KN_LIST_LOCK(kn); 901 } else { 902 /* 903 * The user may change some filter values after the 904 * initial EV_ADD, but doing so will not reset any 905 * filter which has already been triggered. 906 */ 907 kn->kn_status |= KN_INFLUX; 908 KQ_UNLOCK(kq); 909 KN_LIST_LOCK(kn); 910 kn->kn_sfflags = kev->fflags; 911 kn->kn_sdata = kev->data; 912 kn->kn_kevent.udata = kev->udata; 913 } 914 915 /* 916 * We can get here with kn->kn_knlist == NULL. 917 * This can happen when the initial attach event decides that 918 * the event is "completed" already. i.e. filt_procattach 919 * is called on a zombie process. It will call filt_proc 920 * which will remove it from the list, and NULL kn_knlist. 921 */ 922 event = kn->kn_fop->f_event(kn, 0); 923 KN_LIST_UNLOCK(kn); 924 KQ_LOCK(kq); 925 if (event) 926 KNOTE_ACTIVATE(kn, 1); 927 kn->kn_status &= ~KN_INFLUX; 928 } else if (kev->flags & EV_DELETE) { 929 kn->kn_status |= KN_INFLUX; 930 KQ_UNLOCK(kq); 931 if (!(kn->kn_status & KN_DETACHED)) 932 kn->kn_fop->f_detach(kn); 933 knote_drop(kn, td); 934 goto done; 935 } 936 937 if ((kev->flags & EV_DISABLE) && 938 ((kn->kn_status & KN_DISABLED) == 0)) { 939 kn->kn_status |= KN_DISABLED; 940 } 941 942 if ((kev->flags & EV_ENABLE) && (kn->kn_status & KN_DISABLED)) { 943 kn->kn_status &= ~KN_DISABLED; 944 if ((kn->kn_status & KN_ACTIVE) && 945 ((kn->kn_status & KN_QUEUED) == 0)) 946 knote_enqueue(kn); 947 } 948 KQ_UNLOCK_FLUX(kq); 949 950 done: 951 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal); 952 done_noglobal: 953 if (fp != NULL) 954 fdrop(fp, td); 955 if (tkn != NULL) 956 knote_free(tkn); 957 if (fops != NULL) 958 kqueue_fo_release(filt); 959 return (error); 960 } 961 962 static int 963 kqueue_aquire(struct file *fp, struct kqueue **kqp) 964 { 965 int error; 966 struct kqueue *kq; 967 968 error = 0; 969 970 FILE_LOCK(fp); 971 do { 972 kq = fp->f_data; 973 if (fp->f_type != DTYPE_KQUEUE || kq == NULL) { 974 error = EBADF; 975 break; 976 } 977 *kqp = kq; 978 KQ_LOCK(kq); 979 if ((kq->kq_state & KQ_CLOSING) == KQ_CLOSING) { 980 KQ_UNLOCK(kq); 981 error = EBADF; 982 break; 983 } 984 kq->kq_refcnt++; 985 KQ_UNLOCK(kq); 986 } while (0); 987 FILE_UNLOCK(fp); 988 989 return error; 990 } 991 992 static void 993 kqueue_release(struct kqueue *kq, int locked) 994 { 995 if (locked) 996 KQ_OWNED(kq); 997 else 998 KQ_LOCK(kq); 999 kq->kq_refcnt--; 1000 if (kq->kq_refcnt == 1) 1001 wakeup(&kq->kq_refcnt); 1002 if (!locked) 1003 KQ_UNLOCK(kq); 1004 } 1005 1006 static void 1007 kqueue_schedtask(struct kqueue *kq) 1008 { 1009 1010 KQ_OWNED(kq); 1011 KASSERT(((kq->kq_state & KQ_TASKDRAIN) != KQ_TASKDRAIN), 1012 ("scheduling kqueue task while draining")); 1013 1014 if ((kq->kq_state & KQ_TASKSCHED) != KQ_TASKSCHED) { 1015 taskqueue_enqueue(taskqueue_kqueue, &kq->kq_task); 1016 kq->kq_state |= KQ_TASKSCHED; 1017 } 1018 } 1019 1020 /* 1021 * Expand the kq to make sure we have storage for fops/ident pair. 1022 * 1023 * Return 0 on success (or no work necessary), return errno on failure. 1024 * 1025 * Not calling hashinit w/ waitok (proper malloc flag) should be safe. 1026 * If kqueue_register is called from a non-fd context, there usually/should 1027 * be no locks held. 1028 */ 1029 static int 1030 kqueue_expand(struct kqueue *kq, struct filterops *fops, uintptr_t ident, 1031 int waitok) 1032 { 1033 struct klist *list, *tmp_knhash; 1034 u_long tmp_knhashmask; 1035 int size; 1036 int fd; 1037 int mflag = waitok ? M_WAITOK : M_NOWAIT; 1038 1039 KQ_NOTOWNED(kq); 1040 1041 if (fops->f_isfd) { 1042 fd = ident; 1043 if (kq->kq_knlistsize <= fd) { 1044 size = kq->kq_knlistsize; 1045 while (size <= fd) 1046 size += KQEXTENT; 1047 MALLOC(list, struct klist *, 1048 size * sizeof list, M_KQUEUE, mflag); 1049 if (list == NULL) 1050 return ENOMEM; 1051 KQ_LOCK(kq); 1052 if (kq->kq_knlistsize > fd) { 1053 FREE(list, M_KQUEUE); 1054 list = NULL; 1055 } else { 1056 if (kq->kq_knlist != NULL) { 1057 bcopy(kq->kq_knlist, list, 1058 kq->kq_knlistsize * sizeof list); 1059 FREE(kq->kq_knlist, M_KQUEUE); 1060 kq->kq_knlist = NULL; 1061 } 1062 bzero((caddr_t)list + 1063 kq->kq_knlistsize * sizeof list, 1064 (size - kq->kq_knlistsize) * sizeof list); 1065 kq->kq_knlistsize = size; 1066 kq->kq_knlist = list; 1067 } 1068 KQ_UNLOCK(kq); 1069 } 1070 } else { 1071 if (kq->kq_knhashmask == 0) { 1072 tmp_knhash = hashinit(KN_HASHSIZE, M_KQUEUE, 1073 &tmp_knhashmask); 1074 if (tmp_knhash == NULL) 1075 return ENOMEM; 1076 KQ_LOCK(kq); 1077 if (kq->kq_knhashmask == 0) { 1078 kq->kq_knhash = tmp_knhash; 1079 kq->kq_knhashmask = tmp_knhashmask; 1080 } else { 1081 free(tmp_knhash, M_KQUEUE); 1082 } 1083 KQ_UNLOCK(kq); 1084 } 1085 } 1086 1087 KQ_NOTOWNED(kq); 1088 return 0; 1089 } 1090 1091 static void 1092 kqueue_task(void *arg, int pending) 1093 { 1094 struct kqueue *kq; 1095 int haskqglobal; 1096 1097 haskqglobal = 0; 1098 kq = arg; 1099 1100 KQ_GLOBAL_LOCK(&kq_global, haskqglobal); 1101 KQ_LOCK(kq); 1102 1103 KNOTE_LOCKED(&kq->kq_sel.si_note, 0); 1104 1105 kq->kq_state &= ~KQ_TASKSCHED; 1106 if ((kq->kq_state & KQ_TASKDRAIN) == KQ_TASKDRAIN) { 1107 wakeup(&kq->kq_state); 1108 } 1109 KQ_UNLOCK(kq); 1110 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal); 1111 } 1112 1113 /* 1114 * Scan, update kn_data (if not ONESHOT), and copyout triggered events. 1115 * We treat KN_MARKER knotes as if they are INFLUX. 1116 */ 1117 static int 1118 kqueue_scan(struct kqueue *kq, int maxevents, struct kevent_copyops *k_ops, 1119 const struct timespec *tsp, struct kevent *keva, struct thread *td) 1120 { 1121 struct kevent *kevp; 1122 struct timeval atv, rtv, ttv; 1123 struct knote *kn, *marker; 1124 int count, timeout, nkev, error; 1125 int haskqglobal; 1126 1127 count = maxevents; 1128 nkev = 0; 1129 error = 0; 1130 haskqglobal = 0; 1131 1132 if (maxevents == 0) 1133 goto done_nl; 1134 1135 if (tsp != NULL) { 1136 TIMESPEC_TO_TIMEVAL(&atv, tsp); 1137 if (itimerfix(&atv)) { 1138 error = EINVAL; 1139 goto done_nl; 1140 } 1141 if (tsp->tv_sec == 0 && tsp->tv_nsec == 0) 1142 timeout = -1; 1143 else 1144 timeout = atv.tv_sec > 24 * 60 * 60 ? 1145 24 * 60 * 60 * hz : tvtohz(&atv); 1146 getmicrouptime(&rtv); 1147 timevaladd(&atv, &rtv); 1148 } else { 1149 atv.tv_sec = 0; 1150 atv.tv_usec = 0; 1151 timeout = 0; 1152 } 1153 marker = knote_alloc(1); 1154 if (marker == NULL) { 1155 error = ENOMEM; 1156 goto done_nl; 1157 } 1158 marker->kn_status = KN_MARKER; 1159 KQ_LOCK(kq); 1160 goto start; 1161 1162 retry: 1163 if (atv.tv_sec || atv.tv_usec) { 1164 getmicrouptime(&rtv); 1165 if (timevalcmp(&rtv, &atv, >=)) 1166 goto done; 1167 ttv = atv; 1168 timevalsub(&ttv, &rtv); 1169 timeout = ttv.tv_sec > 24 * 60 * 60 ? 1170 24 * 60 * 60 * hz : tvtohz(&ttv); 1171 } 1172 1173 start: 1174 kevp = keva; 1175 if (kq->kq_count == 0) { 1176 if (timeout < 0) { 1177 error = EWOULDBLOCK; 1178 } else { 1179 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal); 1180 kq->kq_state |= KQ_SLEEP; 1181 error = msleep(kq, &kq->kq_lock, PSOCK | PCATCH, 1182 "kqread", timeout); 1183 } 1184 if (error == 0) 1185 goto retry; 1186 /* don't restart after signals... */ 1187 if (error == ERESTART) 1188 error = EINTR; 1189 else if (error == EWOULDBLOCK) 1190 error = 0; 1191 goto done; 1192 } 1193 1194 TAILQ_INSERT_TAIL(&kq->kq_head, marker, kn_tqe); 1195 while (count) { 1196 KQ_OWNED(kq); 1197 kn = TAILQ_FIRST(&kq->kq_head); 1198 1199 if ((kn->kn_status == KN_MARKER && kn != marker) || 1200 (kn->kn_status & KN_INFLUX) == KN_INFLUX) { 1201 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal); 1202 kq->kq_state |= KQ_FLUXWAIT; 1203 error = msleep(kq, &kq->kq_lock, PSOCK, 1204 "kqflxwt", 0); 1205 continue; 1206 } 1207 1208 TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe); 1209 if ((kn->kn_status & KN_DISABLED) == KN_DISABLED) { 1210 kn->kn_status &= ~KN_QUEUED; 1211 kq->kq_count--; 1212 continue; 1213 } 1214 if (kn == marker) { 1215 KQ_FLUX_WAKEUP(kq); 1216 if (count == maxevents) 1217 goto retry; 1218 goto done; 1219 } 1220 KASSERT((kn->kn_status & KN_INFLUX) == 0, 1221 ("KN_INFLUX set when not suppose to be")); 1222 1223 if ((kn->kn_flags & EV_ONESHOT) == EV_ONESHOT) { 1224 kn->kn_status &= ~KN_QUEUED; 1225 kn->kn_status |= KN_INFLUX; 1226 kq->kq_count--; 1227 KQ_UNLOCK(kq); 1228 /* 1229 * We don't need to lock the list since we've marked 1230 * it _INFLUX. 1231 */ 1232 *kevp = kn->kn_kevent; 1233 if (!(kn->kn_status & KN_DETACHED)) 1234 kn->kn_fop->f_detach(kn); 1235 knote_drop(kn, td); 1236 KQ_LOCK(kq); 1237 kn = NULL; 1238 } else { 1239 kn->kn_status |= KN_INFLUX; 1240 KQ_UNLOCK(kq); 1241 if ((kn->kn_status & KN_KQUEUE) == KN_KQUEUE) 1242 KQ_GLOBAL_LOCK(&kq_global, haskqglobal); 1243 KN_LIST_LOCK(kn); 1244 if (kn->kn_fop->f_event(kn, 0) == 0) { 1245 KQ_LOCK(kq); 1246 kn->kn_status &= 1247 ~(KN_QUEUED | KN_ACTIVE | KN_INFLUX); 1248 kq->kq_count--; 1249 KN_LIST_UNLOCK(kn); 1250 continue; 1251 } 1252 *kevp = kn->kn_kevent; 1253 KQ_LOCK(kq); 1254 if (kn->kn_flags & EV_CLEAR) { 1255 kn->kn_data = 0; 1256 kn->kn_fflags = 0; 1257 kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE); 1258 kq->kq_count--; 1259 } else 1260 TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe); 1261 1262 kn->kn_status &= ~(KN_INFLUX); 1263 KN_LIST_UNLOCK(kn); 1264 1265 } 1266 1267 /* we are returning a copy to the user */ 1268 kevp++; 1269 nkev++; 1270 count--; 1271 1272 if (nkev == KQ_NEVENTS) { 1273 KQ_UNLOCK_FLUX(kq); 1274 error = k_ops->k_copyout(k_ops->arg, keva, nkev); 1275 nkev = 0; 1276 kevp = keva; 1277 KQ_LOCK(kq); 1278 if (error) 1279 break; 1280 } 1281 } 1282 TAILQ_REMOVE(&kq->kq_head, marker, kn_tqe); 1283 done: 1284 KQ_OWNED(kq); 1285 KQ_UNLOCK_FLUX(kq); 1286 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal); 1287 knote_free(marker); 1288 done_nl: 1289 KQ_NOTOWNED(kq); 1290 if (nkev != 0) 1291 error = k_ops->k_copyout(k_ops->arg, keva, nkev); 1292 td->td_retval[0] = maxevents - count; 1293 return (error); 1294 } 1295 1296 /* 1297 * XXX 1298 * This could be expanded to call kqueue_scan, if desired. 1299 */ 1300 /*ARGSUSED*/ 1301 static int 1302 kqueue_read(struct file *fp, struct uio *uio, struct ucred *active_cred, 1303 int flags, struct thread *td) 1304 { 1305 return (ENXIO); 1306 } 1307 1308 /*ARGSUSED*/ 1309 static int 1310 kqueue_write(struct file *fp, struct uio *uio, struct ucred *active_cred, 1311 int flags, struct thread *td) 1312 { 1313 return (ENXIO); 1314 } 1315 1316 /*ARGSUSED*/ 1317 static int 1318 kqueue_ioctl(struct file *fp, u_long cmd, void *data, 1319 struct ucred *active_cred, struct thread *td) 1320 { 1321 /* 1322 * Enabling sigio causes two major problems: 1323 * 1) infinite recursion: 1324 * Synopsys: kevent is being used to track signals and have FIOASYNC 1325 * set. On receipt of a signal this will cause a kqueue to recurse 1326 * into itself over and over. Sending the sigio causes the kqueue 1327 * to become ready, which in turn posts sigio again, forever. 1328 * Solution: this can be solved by setting a flag in the kqueue that 1329 * we have a SIGIO in progress. 1330 * 2) locking problems: 1331 * Synopsys: Kqueue is a leaf subsystem, but adding signalling puts 1332 * us above the proc and pgrp locks. 1333 * Solution: Post a signal using an async mechanism, being sure to 1334 * record a generation count in the delivery so that we do not deliver 1335 * a signal to the wrong process. 1336 * 1337 * Note, these two mechanisms are somewhat mutually exclusive! 1338 */ 1339 #if 0 1340 struct kqueue *kq; 1341 1342 kq = fp->f_data; 1343 switch (cmd) { 1344 case FIOASYNC: 1345 if (*(int *)data) { 1346 kq->kq_state |= KQ_ASYNC; 1347 } else { 1348 kq->kq_state &= ~KQ_ASYNC; 1349 } 1350 return (0); 1351 1352 case FIOSETOWN: 1353 return (fsetown(*(int *)data, &kq->kq_sigio)); 1354 1355 case FIOGETOWN: 1356 *(int *)data = fgetown(&kq->kq_sigio); 1357 return (0); 1358 } 1359 #endif 1360 1361 return (ENOTTY); 1362 } 1363 1364 /*ARGSUSED*/ 1365 static int 1366 kqueue_poll(struct file *fp, int events, struct ucred *active_cred, 1367 struct thread *td) 1368 { 1369 struct kqueue *kq; 1370 int revents = 0; 1371 int error; 1372 1373 if ((error = kqueue_aquire(fp, &kq))) 1374 return POLLERR; 1375 1376 KQ_LOCK(kq); 1377 if (events & (POLLIN | POLLRDNORM)) { 1378 if (kq->kq_count) { 1379 revents |= events & (POLLIN | POLLRDNORM); 1380 } else { 1381 selrecord(td, &kq->kq_sel); 1382 kq->kq_state |= KQ_SEL; 1383 } 1384 } 1385 kqueue_release(kq, 1); 1386 KQ_UNLOCK(kq); 1387 return (revents); 1388 } 1389 1390 /*ARGSUSED*/ 1391 static int 1392 kqueue_stat(struct file *fp, struct stat *st, struct ucred *active_cred, 1393 struct thread *td) 1394 { 1395 1396 bzero((void *)st, sizeof *st); 1397 /* 1398 * We no longer return kq_count because the unlocked value is useless. 1399 * If you spent all this time getting the count, why not spend your 1400 * syscall better by calling kevent? 1401 * 1402 * XXX - This is needed for libc_r. 1403 */ 1404 st->st_mode = S_IFIFO; 1405 return (0); 1406 } 1407 1408 /*ARGSUSED*/ 1409 static int 1410 kqueue_close(struct file *fp, struct thread *td) 1411 { 1412 struct kqueue *kq = fp->f_data; 1413 struct filedesc *fdp; 1414 struct knote *kn; 1415 int i; 1416 int error; 1417 1418 if ((error = kqueue_aquire(fp, &kq))) 1419 return error; 1420 1421 KQ_LOCK(kq); 1422 1423 KASSERT((kq->kq_state & KQ_CLOSING) != KQ_CLOSING, 1424 ("kqueue already closing")); 1425 kq->kq_state |= KQ_CLOSING; 1426 if (kq->kq_refcnt > 1) 1427 msleep(&kq->kq_refcnt, &kq->kq_lock, PSOCK, "kqclose", 0); 1428 1429 KASSERT(kq->kq_refcnt == 1, ("other refs are out there!")); 1430 fdp = kq->kq_fdp; 1431 1432 KASSERT(knlist_empty(&kq->kq_sel.si_note), 1433 ("kqueue's knlist not empty")); 1434 1435 for (i = 0; i < kq->kq_knlistsize; i++) { 1436 while ((kn = SLIST_FIRST(&kq->kq_knlist[i])) != NULL) { 1437 KASSERT((kn->kn_status & KN_INFLUX) == 0, 1438 ("KN_INFLUX set when not suppose to be")); 1439 kn->kn_status |= KN_INFLUX; 1440 KQ_UNLOCK(kq); 1441 if (!(kn->kn_status & KN_DETACHED)) 1442 kn->kn_fop->f_detach(kn); 1443 knote_drop(kn, td); 1444 KQ_LOCK(kq); 1445 } 1446 } 1447 if (kq->kq_knhashmask != 0) { 1448 for (i = 0; i <= kq->kq_knhashmask; i++) { 1449 while ((kn = SLIST_FIRST(&kq->kq_knhash[i])) != NULL) { 1450 KASSERT((kn->kn_status & KN_INFLUX) == 0, 1451 ("KN_INFLUX set when not suppose to be")); 1452 kn->kn_status |= KN_INFLUX; 1453 KQ_UNLOCK(kq); 1454 if (!(kn->kn_status & KN_DETACHED)) 1455 kn->kn_fop->f_detach(kn); 1456 knote_drop(kn, td); 1457 KQ_LOCK(kq); 1458 } 1459 } 1460 } 1461 1462 if ((kq->kq_state & KQ_TASKSCHED) == KQ_TASKSCHED) { 1463 kq->kq_state |= KQ_TASKDRAIN; 1464 msleep(&kq->kq_state, &kq->kq_lock, PSOCK, "kqtqdr", 0); 1465 } 1466 1467 if ((kq->kq_state & KQ_SEL) == KQ_SEL) { 1468 kq->kq_state &= ~KQ_SEL; 1469 selwakeuppri(&kq->kq_sel, PSOCK); 1470 } 1471 1472 KQ_UNLOCK(kq); 1473 1474 FILEDESC_LOCK_FAST(fdp); 1475 SLIST_REMOVE(&fdp->fd_kqlist, kq, kqueue, kq_list); 1476 FILEDESC_UNLOCK_FAST(fdp); 1477 1478 knlist_destroy(&kq->kq_sel.si_note); 1479 mtx_destroy(&kq->kq_lock); 1480 kq->kq_fdp = NULL; 1481 1482 if (kq->kq_knhash != NULL) 1483 free(kq->kq_knhash, M_KQUEUE); 1484 if (kq->kq_knlist != NULL) 1485 free(kq->kq_knlist, M_KQUEUE); 1486 1487 funsetown(&kq->kq_sigio); 1488 free(kq, M_KQUEUE); 1489 fp->f_data = NULL; 1490 1491 return (0); 1492 } 1493 1494 static void 1495 kqueue_wakeup(struct kqueue *kq) 1496 { 1497 KQ_OWNED(kq); 1498 1499 if ((kq->kq_state & KQ_SLEEP) == KQ_SLEEP) { 1500 kq->kq_state &= ~KQ_SLEEP; 1501 wakeup(kq); 1502 } 1503 if ((kq->kq_state & KQ_SEL) == KQ_SEL) { 1504 kq->kq_state &= ~KQ_SEL; 1505 selwakeuppri(&kq->kq_sel, PSOCK); 1506 } 1507 if (!knlist_empty(&kq->kq_sel.si_note)) 1508 kqueue_schedtask(kq); 1509 if ((kq->kq_state & KQ_ASYNC) == KQ_ASYNC) { 1510 pgsigio(&kq->kq_sigio, SIGIO, 0); 1511 } 1512 } 1513 1514 /* 1515 * Walk down a list of knotes, activating them if their event has triggered. 1516 * 1517 * There is a possibility to optimize in the case of one kq watching another. 1518 * Instead of scheduling a task to wake it up, you could pass enough state 1519 * down the chain to make up the parent kqueue. Make this code functional 1520 * first. 1521 */ 1522 void 1523 knote(struct knlist *list, long hint, int islocked) 1524 { 1525 struct kqueue *kq; 1526 struct knote *kn; 1527 1528 if (list == NULL) 1529 return; 1530 1531 KNL_ASSERT_LOCK(list, islocked); 1532 1533 if (!islocked) 1534 list->kl_lock(list->kl_lockarg); 1535 1536 /* 1537 * If we unlock the list lock (and set KN_INFLUX), we can eliminate 1538 * the kqueue scheduling, but this will introduce four 1539 * lock/unlock's for each knote to test. If we do, continue to use 1540 * SLIST_FOREACH, SLIST_FOREACH_SAFE is not safe in our case, it is 1541 * only safe if you want to remove the current item, which we are 1542 * not doing. 1543 */ 1544 SLIST_FOREACH(kn, &list->kl_list, kn_selnext) { 1545 kq = kn->kn_kq; 1546 if ((kn->kn_status & KN_INFLUX) != KN_INFLUX) { 1547 KQ_LOCK(kq); 1548 if ((kn->kn_status & KN_INFLUX) != KN_INFLUX) { 1549 kn->kn_status |= KN_HASKQLOCK; 1550 if (kn->kn_fop->f_event(kn, hint)) 1551 KNOTE_ACTIVATE(kn, 1); 1552 kn->kn_status &= ~KN_HASKQLOCK; 1553 } 1554 KQ_UNLOCK(kq); 1555 } 1556 kq = NULL; 1557 } 1558 if (!islocked) 1559 list->kl_unlock(list->kl_lockarg); 1560 } 1561 1562 /* 1563 * add a knote to a knlist 1564 */ 1565 void 1566 knlist_add(struct knlist *knl, struct knote *kn, int islocked) 1567 { 1568 KNL_ASSERT_LOCK(knl, islocked); 1569 KQ_NOTOWNED(kn->kn_kq); 1570 KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) == 1571 (KN_INFLUX|KN_DETACHED), ("knote not KN_INFLUX and KN_DETACHED")); 1572 if (!islocked) 1573 knl->kl_lock(knl->kl_lockarg); 1574 SLIST_INSERT_HEAD(&knl->kl_list, kn, kn_selnext); 1575 if (!islocked) 1576 knl->kl_unlock(knl->kl_lockarg); 1577 KQ_LOCK(kn->kn_kq); 1578 kn->kn_knlist = knl; 1579 kn->kn_status &= ~KN_DETACHED; 1580 KQ_UNLOCK(kn->kn_kq); 1581 } 1582 1583 static void 1584 knlist_remove_kq(struct knlist *knl, struct knote *kn, int knlislocked, int kqislocked) 1585 { 1586 KASSERT(!(!!kqislocked && !knlislocked), ("kq locked w/o knl locked")); 1587 KNL_ASSERT_LOCK(knl, knlislocked); 1588 mtx_assert(&kn->kn_kq->kq_lock, kqislocked ? MA_OWNED : MA_NOTOWNED); 1589 if (!kqislocked) 1590 KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) == KN_INFLUX, 1591 ("knlist_remove called w/o knote being KN_INFLUX or already removed")); 1592 if (!knlislocked) 1593 knl->kl_lock(knl->kl_lockarg); 1594 SLIST_REMOVE(&knl->kl_list, kn, knote, kn_selnext); 1595 kn->kn_knlist = NULL; 1596 if (!knlislocked) 1597 knl->kl_unlock(knl->kl_lockarg); 1598 if (!kqislocked) 1599 KQ_LOCK(kn->kn_kq); 1600 kn->kn_status |= KN_DETACHED; 1601 if (!kqislocked) 1602 KQ_UNLOCK(kn->kn_kq); 1603 } 1604 1605 /* 1606 * remove all knotes from a specified klist 1607 */ 1608 void 1609 knlist_remove(struct knlist *knl, struct knote *kn, int islocked) 1610 { 1611 1612 knlist_remove_kq(knl, kn, islocked, 0); 1613 } 1614 1615 /* 1616 * remove knote from a specified klist while in f_event handler. 1617 */ 1618 void 1619 knlist_remove_inevent(struct knlist *knl, struct knote *kn) 1620 { 1621 1622 knlist_remove_kq(knl, kn, 1, 1623 (kn->kn_status & KN_HASKQLOCK) == KN_HASKQLOCK); 1624 } 1625 1626 int 1627 knlist_empty(struct knlist *knl) 1628 { 1629 KNL_ASSERT_LOCKED(knl); 1630 return SLIST_EMPTY(&knl->kl_list); 1631 } 1632 1633 static struct mtx knlist_lock; 1634 MTX_SYSINIT(knlist_lock, &knlist_lock, "knlist lock for lockless objects", 1635 MTX_DEF); 1636 static void knlist_mtx_lock(void *arg); 1637 static void knlist_mtx_unlock(void *arg); 1638 static int knlist_mtx_locked(void *arg); 1639 1640 static void 1641 knlist_mtx_lock(void *arg) 1642 { 1643 mtx_lock((struct mtx *)arg); 1644 } 1645 1646 static void 1647 knlist_mtx_unlock(void *arg) 1648 { 1649 mtx_unlock((struct mtx *)arg); 1650 } 1651 1652 static int 1653 knlist_mtx_locked(void *arg) 1654 { 1655 return (mtx_owned((struct mtx *)arg)); 1656 } 1657 1658 void 1659 knlist_init(struct knlist *knl, void *lock, void (*kl_lock)(void *), 1660 void (*kl_unlock)(void *), int (*kl_locked)(void *)) 1661 { 1662 1663 if (lock == NULL) 1664 knl->kl_lockarg = &knlist_lock; 1665 else 1666 knl->kl_lockarg = lock; 1667 1668 if (kl_lock == NULL) 1669 knl->kl_lock = knlist_mtx_lock; 1670 else 1671 knl->kl_lock = kl_lock; 1672 if (kl_lock == NULL) 1673 knl->kl_unlock = knlist_mtx_unlock; 1674 else 1675 knl->kl_unlock = kl_unlock; 1676 if (kl_locked == NULL) 1677 knl->kl_locked = knlist_mtx_locked; 1678 else 1679 knl->kl_locked = kl_locked; 1680 1681 SLIST_INIT(&knl->kl_list); 1682 } 1683 1684 void 1685 knlist_destroy(struct knlist *knl) 1686 { 1687 1688 #ifdef INVARIANTS 1689 /* 1690 * if we run across this error, we need to find the offending 1691 * driver and have it call knlist_clear. 1692 */ 1693 if (!SLIST_EMPTY(&knl->kl_list)) 1694 printf("WARNING: destroying knlist w/ knotes on it!\n"); 1695 #endif 1696 1697 knl->kl_lockarg = knl->kl_lock = knl->kl_unlock = NULL; 1698 SLIST_INIT(&knl->kl_list); 1699 } 1700 1701 /* 1702 * Even if we are locked, we may need to drop the lock to allow any influx 1703 * knotes time to "settle". 1704 */ 1705 void 1706 knlist_cleardel(struct knlist *knl, struct thread *td, int islocked, int killkn) 1707 { 1708 struct knote *kn; 1709 struct kqueue *kq; 1710 1711 if (islocked) 1712 KNL_ASSERT_LOCKED(knl); 1713 else { 1714 KNL_ASSERT_UNLOCKED(knl); 1715 again: /* need to reaquire lock since we have dropped it */ 1716 knl->kl_lock(knl->kl_lockarg); 1717 } 1718 1719 SLIST_FOREACH(kn, &knl->kl_list, kn_selnext) { 1720 kq = kn->kn_kq; 1721 KQ_LOCK(kq); 1722 if ((kn->kn_status & KN_INFLUX)) { 1723 KQ_UNLOCK(kq); 1724 continue; 1725 } 1726 knlist_remove_kq(knl, kn, 1, 1); 1727 if (killkn) { 1728 kn->kn_status |= KN_INFLUX | KN_DETACHED; 1729 KQ_UNLOCK(kq); 1730 knote_drop(kn, td); 1731 } else { 1732 /* Make sure cleared knotes disappear soon */ 1733 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 1734 KQ_UNLOCK(kq); 1735 } 1736 kq = NULL; 1737 } 1738 1739 if (!SLIST_EMPTY(&knl->kl_list)) { 1740 /* there are still KN_INFLUX remaining */ 1741 kn = SLIST_FIRST(&knl->kl_list); 1742 kq = kn->kn_kq; 1743 KQ_LOCK(kq); 1744 KASSERT(kn->kn_status & KN_INFLUX, 1745 ("knote removed w/o list lock")); 1746 knl->kl_unlock(knl->kl_lockarg); 1747 kq->kq_state |= KQ_FLUXWAIT; 1748 msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqkclr", 0); 1749 kq = NULL; 1750 goto again; 1751 } 1752 1753 if (islocked) 1754 KNL_ASSERT_LOCKED(knl); 1755 else { 1756 knl->kl_unlock(knl->kl_lockarg); 1757 KNL_ASSERT_UNLOCKED(knl); 1758 } 1759 } 1760 1761 /* 1762 * remove all knotes referencing a specified fd 1763 * must be called with FILEDESC lock. This prevents a race where a new fd 1764 * comes along and occupies the entry and we attach a knote to the fd. 1765 */ 1766 void 1767 knote_fdclose(struct thread *td, int fd) 1768 { 1769 struct filedesc *fdp = td->td_proc->p_fd; 1770 struct kqueue *kq; 1771 struct knote *kn; 1772 int influx; 1773 1774 FILEDESC_LOCK_ASSERT(fdp, MA_OWNED); 1775 1776 /* 1777 * We shouldn't have to worry about new kevents appearing on fd 1778 * since filedesc is locked. 1779 */ 1780 SLIST_FOREACH(kq, &fdp->fd_kqlist, kq_list) { 1781 KQ_LOCK(kq); 1782 1783 again: 1784 influx = 0; 1785 while (kq->kq_knlistsize > fd && 1786 (kn = SLIST_FIRST(&kq->kq_knlist[fd])) != NULL) { 1787 if (kn->kn_status & KN_INFLUX) { 1788 /* someone else might be waiting on our knote */ 1789 if (influx) 1790 wakeup(kq); 1791 kq->kq_state |= KQ_FLUXWAIT; 1792 msleep(kq, &kq->kq_lock, PSOCK, "kqflxwt", 0); 1793 goto again; 1794 } 1795 kn->kn_status |= KN_INFLUX; 1796 KQ_UNLOCK(kq); 1797 if (!(kn->kn_status & KN_DETACHED)) 1798 kn->kn_fop->f_detach(kn); 1799 knote_drop(kn, td); 1800 influx = 1; 1801 KQ_LOCK(kq); 1802 } 1803 KQ_UNLOCK_FLUX(kq); 1804 } 1805 } 1806 1807 static int 1808 knote_attach(struct knote *kn, struct kqueue *kq) 1809 { 1810 struct klist *list; 1811 1812 KASSERT(kn->kn_status & KN_INFLUX, ("knote not marked INFLUX")); 1813 KQ_OWNED(kq); 1814 1815 if (kn->kn_fop->f_isfd) { 1816 if (kn->kn_id >= kq->kq_knlistsize) 1817 return ENOMEM; 1818 list = &kq->kq_knlist[kn->kn_id]; 1819 } else { 1820 if (kq->kq_knhash == NULL) 1821 return ENOMEM; 1822 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)]; 1823 } 1824 1825 SLIST_INSERT_HEAD(list, kn, kn_link); 1826 1827 return 0; 1828 } 1829 1830 /* 1831 * knote must already have been detatched using the f_detach method. 1832 * no lock need to be held, it is assumed that the KN_INFLUX flag is set 1833 * to prevent other removal. 1834 */ 1835 static void 1836 knote_drop(struct knote *kn, struct thread *td) 1837 { 1838 struct kqueue *kq; 1839 struct klist *list; 1840 1841 kq = kn->kn_kq; 1842 1843 KQ_NOTOWNED(kq); 1844 KASSERT((kn->kn_status & KN_INFLUX) == KN_INFLUX, 1845 ("knote_drop called without KN_INFLUX set in kn_status")); 1846 1847 KQ_LOCK(kq); 1848 if (kn->kn_fop->f_isfd) 1849 list = &kq->kq_knlist[kn->kn_id]; 1850 else 1851 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)]; 1852 1853 SLIST_REMOVE(list, kn, knote, kn_link); 1854 if (kn->kn_status & KN_QUEUED) 1855 knote_dequeue(kn); 1856 KQ_UNLOCK_FLUX(kq); 1857 1858 if (kn->kn_fop->f_isfd) { 1859 fdrop(kn->kn_fp, td); 1860 kn->kn_fp = NULL; 1861 } 1862 kqueue_fo_release(kn->kn_kevent.filter); 1863 kn->kn_fop = NULL; 1864 knote_free(kn); 1865 } 1866 1867 static void 1868 knote_enqueue(struct knote *kn) 1869 { 1870 struct kqueue *kq = kn->kn_kq; 1871 1872 KQ_OWNED(kn->kn_kq); 1873 KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued")); 1874 1875 TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe); 1876 kn->kn_status |= KN_QUEUED; 1877 kq->kq_count++; 1878 kqueue_wakeup(kq); 1879 } 1880 1881 static void 1882 knote_dequeue(struct knote *kn) 1883 { 1884 struct kqueue *kq = kn->kn_kq; 1885 1886 KQ_OWNED(kn->kn_kq); 1887 KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued")); 1888 1889 TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe); 1890 kn->kn_status &= ~KN_QUEUED; 1891 kq->kq_count--; 1892 } 1893 1894 static void 1895 knote_init(void) 1896 { 1897 1898 knote_zone = uma_zcreate("KNOTE", sizeof(struct knote), NULL, NULL, 1899 NULL, NULL, UMA_ALIGN_PTR, 0); 1900 } 1901 SYSINIT(knote, SI_SUB_PSEUDO, SI_ORDER_ANY, knote_init, NULL) 1902 1903 static struct knote * 1904 knote_alloc(int waitok) 1905 { 1906 return ((struct knote *)uma_zalloc(knote_zone, 1907 (waitok ? M_WAITOK : M_NOWAIT)|M_ZERO)); 1908 } 1909 1910 static void 1911 knote_free(struct knote *kn) 1912 { 1913 if (kn != NULL) 1914 uma_zfree(knote_zone, kn); 1915 } 1916