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 "opt_ktrace.h" 32 33 #include <sys/param.h> 34 #include <sys/systm.h> 35 #include <sys/kernel.h> 36 #include <sys/lock.h> 37 #include <sys/mutex.h> 38 #include <sys/proc.h> 39 #include <sys/malloc.h> 40 #include <sys/unistd.h> 41 #include <sys/file.h> 42 #include <sys/filedesc.h> 43 #include <sys/filio.h> 44 #include <sys/fcntl.h> 45 #include <sys/kthread.h> 46 #include <sys/selinfo.h> 47 #include <sys/queue.h> 48 #include <sys/event.h> 49 #include <sys/eventvar.h> 50 #include <sys/poll.h> 51 #include <sys/protosw.h> 52 #include <sys/sigio.h> 53 #include <sys/signalvar.h> 54 #include <sys/socket.h> 55 #include <sys/socketvar.h> 56 #include <sys/stat.h> 57 #include <sys/sysctl.h> 58 #include <sys/sysproto.h> 59 #include <sys/syscallsubr.h> 60 #include <sys/taskqueue.h> 61 #include <sys/uio.h> 62 #ifdef KTRACE 63 #include <sys/ktrace.h> 64 #endif 65 66 #include <vm/uma.h> 67 68 static MALLOC_DEFINE(M_KQUEUE, "kqueue", "memory for kqueue system"); 69 70 /* 71 * This lock is used if multiple kq locks are required. This possibly 72 * should be made into a per proc lock. 73 */ 74 static struct mtx kq_global; 75 MTX_SYSINIT(kq_global, &kq_global, "kqueue order", MTX_DEF); 76 #define KQ_GLOBAL_LOCK(lck, haslck) do { \ 77 if (!haslck) \ 78 mtx_lock(lck); \ 79 haslck = 1; \ 80 } while (0) 81 #define KQ_GLOBAL_UNLOCK(lck, haslck) do { \ 82 if (haslck) \ 83 mtx_unlock(lck); \ 84 haslck = 0; \ 85 } while (0) 86 87 TASKQUEUE_DEFINE_THREAD(kqueue); 88 89 static int kevent_copyout(void *arg, struct kevent *kevp, int count); 90 static int kevent_copyin(void *arg, struct kevent *kevp, int count); 91 static int kqueue_register(struct kqueue *kq, struct kevent *kev, 92 struct thread *td, int waitok); 93 static int kqueue_aquire(struct file *fp, struct kqueue **kqp); 94 static void kqueue_release(struct kqueue *kq, int locked); 95 static int kqueue_expand(struct kqueue *kq, struct filterops *fops, 96 uintptr_t ident, int waitok); 97 static void kqueue_task(void *arg, int pending); 98 static int kqueue_scan(struct kqueue *kq, int maxevents, 99 struct kevent_copyops *k_ops, 100 const struct timespec *timeout, 101 struct kevent *keva, struct thread *td); 102 static void kqueue_wakeup(struct kqueue *kq); 103 static struct filterops *kqueue_fo_find(int filt); 104 static void kqueue_fo_release(int filt); 105 106 static fo_rdwr_t kqueue_read; 107 static fo_rdwr_t kqueue_write; 108 static fo_ioctl_t kqueue_ioctl; 109 static fo_poll_t kqueue_poll; 110 static fo_kqfilter_t kqueue_kqfilter; 111 static fo_stat_t kqueue_stat; 112 static fo_close_t kqueue_close; 113 114 static struct fileops kqueueops = { 115 .fo_read = kqueue_read, 116 .fo_write = kqueue_write, 117 .fo_ioctl = kqueue_ioctl, 118 .fo_poll = kqueue_poll, 119 .fo_kqfilter = kqueue_kqfilter, 120 .fo_stat = kqueue_stat, 121 .fo_close = kqueue_close, 122 }; 123 124 static int knote_attach(struct knote *kn, struct kqueue *kq); 125 static void knote_drop(struct knote *kn, struct thread *td); 126 static void knote_enqueue(struct knote *kn); 127 static void knote_dequeue(struct knote *kn); 128 static void knote_init(void); 129 static struct knote *knote_alloc(int waitok); 130 static void knote_free(struct knote *kn); 131 132 static void filt_kqdetach(struct knote *kn); 133 static int filt_kqueue(struct knote *kn, long hint); 134 static int filt_procattach(struct knote *kn); 135 static void filt_procdetach(struct knote *kn); 136 static int filt_proc(struct knote *kn, long hint); 137 static int filt_fileattach(struct knote *kn); 138 static void filt_timerexpire(void *knx); 139 static int filt_timerattach(struct knote *kn); 140 static void filt_timerdetach(struct knote *kn); 141 static int filt_timer(struct knote *kn, long hint); 142 143 static struct filterops file_filtops = 144 { 1, filt_fileattach, NULL, NULL }; 145 static struct filterops kqread_filtops = 146 { 1, NULL, filt_kqdetach, filt_kqueue }; 147 /* XXX - move to kern_proc.c? */ 148 static struct filterops proc_filtops = 149 { 0, filt_procattach, filt_procdetach, filt_proc }; 150 static struct filterops timer_filtops = 151 { 0, filt_timerattach, filt_timerdetach, filt_timer }; 152 153 static uma_zone_t knote_zone; 154 static int kq_ncallouts = 0; 155 static int kq_calloutmax = (4 * 1024); 156 SYSCTL_INT(_kern, OID_AUTO, kq_calloutmax, CTLFLAG_RW, 157 &kq_calloutmax, 0, "Maximum number of callouts allocated for kqueue"); 158 159 /* XXX - ensure not KN_INFLUX?? */ 160 #define KNOTE_ACTIVATE(kn, islock) do { \ 161 if ((islock)) \ 162 mtx_assert(&(kn)->kn_kq->kq_lock, MA_OWNED); \ 163 else \ 164 KQ_LOCK((kn)->kn_kq); \ 165 (kn)->kn_status |= KN_ACTIVE; \ 166 if (((kn)->kn_status & (KN_QUEUED | KN_DISABLED)) == 0) \ 167 knote_enqueue((kn)); \ 168 if (!(islock)) \ 169 KQ_UNLOCK((kn)->kn_kq); \ 170 } while(0) 171 #define KQ_LOCK(kq) do { \ 172 mtx_lock(&(kq)->kq_lock); \ 173 } while (0) 174 #define KQ_FLUX_WAKEUP(kq) do { \ 175 if (((kq)->kq_state & KQ_FLUXWAIT) == KQ_FLUXWAIT) { \ 176 (kq)->kq_state &= ~KQ_FLUXWAIT; \ 177 wakeup((kq)); \ 178 } \ 179 } while (0) 180 #define KQ_UNLOCK_FLUX(kq) do { \ 181 KQ_FLUX_WAKEUP(kq); \ 182 mtx_unlock(&(kq)->kq_lock); \ 183 } while (0) 184 #define KQ_UNLOCK(kq) do { \ 185 mtx_unlock(&(kq)->kq_lock); \ 186 } while (0) 187 #define KQ_OWNED(kq) do { \ 188 mtx_assert(&(kq)->kq_lock, MA_OWNED); \ 189 } while (0) 190 #define KQ_NOTOWNED(kq) do { \ 191 mtx_assert(&(kq)->kq_lock, MA_NOTOWNED); \ 192 } while (0) 193 #define KN_LIST_LOCK(kn) do { \ 194 if (kn->kn_knlist != NULL) \ 195 kn->kn_knlist->kl_lock(kn->kn_knlist->kl_lockarg); \ 196 } while (0) 197 #define KN_LIST_UNLOCK(kn) do { \ 198 if (kn->kn_knlist != NULL) \ 199 kn->kn_knlist->kl_unlock(kn->kn_knlist->kl_lockarg); \ 200 } while (0) 201 #define KNL_ASSERT_LOCK(knl, islocked) do { \ 202 if (islocked) \ 203 KNL_ASSERT_LOCKED(knl); \ 204 else \ 205 KNL_ASSERT_UNLOCKED(knl); \ 206 } while (0) 207 #ifdef INVARIANTS 208 #define KNL_ASSERT_LOCKED(knl) do { \ 209 if (!knl->kl_locked((knl)->kl_lockarg)) \ 210 panic("knlist not locked, but should be"); \ 211 } while (0) 212 #define KNL_ASSERT_UNLOCKED(knl) do { \ 213 if (knl->kl_locked((knl)->kl_lockarg)) \ 214 panic("knlist locked, but should not be"); \ 215 } while (0) 216 #else /* !INVARIANTS */ 217 #define KNL_ASSERT_LOCKED(knl) do {} while(0) 218 #define KNL_ASSERT_UNLOCKED(knl) do {} while (0) 219 #endif /* INVARIANTS */ 220 221 #define KN_HASHSIZE 64 /* XXX should be tunable */ 222 #define KN_HASH(val, mask) (((val) ^ (val >> 8)) & (mask)) 223 224 static int 225 filt_nullattach(struct knote *kn) 226 { 227 228 return (ENXIO); 229 }; 230 231 struct filterops null_filtops = 232 { 0, filt_nullattach, NULL, NULL }; 233 234 /* XXX - make SYSINIT to add these, and move into respective modules. */ 235 extern struct filterops sig_filtops; 236 extern struct filterops fs_filtops; 237 238 /* 239 * Table for for all system-defined filters. 240 */ 241 static struct mtx filterops_lock; 242 MTX_SYSINIT(kqueue_filterops, &filterops_lock, "protect sysfilt_ops", 243 MTX_DEF); 244 static struct { 245 struct filterops *for_fop; 246 int for_refcnt; 247 } sysfilt_ops[EVFILT_SYSCOUNT] = { 248 { &file_filtops }, /* EVFILT_READ */ 249 { &file_filtops }, /* EVFILT_WRITE */ 250 { &null_filtops }, /* EVFILT_AIO */ 251 { &file_filtops }, /* EVFILT_VNODE */ 252 { &proc_filtops }, /* EVFILT_PROC */ 253 { &sig_filtops }, /* EVFILT_SIGNAL */ 254 { &timer_filtops }, /* EVFILT_TIMER */ 255 { &file_filtops }, /* EVFILT_NETDEV */ 256 { &fs_filtops }, /* EVFILT_FS */ 257 { &null_filtops }, /* EVFILT_LIO */ 258 }; 259 260 /* 261 * Simple redirection for all cdevsw style objects to call their fo_kqfilter 262 * method. 263 */ 264 static int 265 filt_fileattach(struct knote *kn) 266 { 267 268 return (fo_kqfilter(kn->kn_fp, kn)); 269 } 270 271 /*ARGSUSED*/ 272 static int 273 kqueue_kqfilter(struct file *fp, struct knote *kn) 274 { 275 struct kqueue *kq = kn->kn_fp->f_data; 276 277 if (kn->kn_filter != EVFILT_READ) 278 return (EINVAL); 279 280 kn->kn_status |= KN_KQUEUE; 281 kn->kn_fop = &kqread_filtops; 282 knlist_add(&kq->kq_sel.si_note, kn, 0); 283 284 return (0); 285 } 286 287 static void 288 filt_kqdetach(struct knote *kn) 289 { 290 struct kqueue *kq = kn->kn_fp->f_data; 291 292 knlist_remove(&kq->kq_sel.si_note, kn, 0); 293 } 294 295 /*ARGSUSED*/ 296 static int 297 filt_kqueue(struct knote *kn, long hint) 298 { 299 struct kqueue *kq = kn->kn_fp->f_data; 300 301 kn->kn_data = kq->kq_count; 302 return (kn->kn_data > 0); 303 } 304 305 /* XXX - move to kern_proc.c? */ 306 static int 307 filt_procattach(struct knote *kn) 308 { 309 struct proc *p; 310 int immediate; 311 int error; 312 313 immediate = 0; 314 p = pfind(kn->kn_id); 315 if (p == NULL && (kn->kn_sfflags & NOTE_EXIT)) { 316 p = zpfind(kn->kn_id); 317 immediate = 1; 318 } else if (p != NULL && (p->p_flag & P_WEXIT)) { 319 immediate = 1; 320 } 321 322 if (p == NULL) 323 return (ESRCH); 324 if ((error = p_cansee(curthread, p))) 325 return (error); 326 327 kn->kn_ptr.p_proc = p; 328 kn->kn_flags |= EV_CLEAR; /* automatically set */ 329 330 /* 331 * internal flag indicating registration done by kernel 332 */ 333 if (kn->kn_flags & EV_FLAG1) { 334 kn->kn_data = kn->kn_sdata; /* ppid */ 335 kn->kn_fflags = NOTE_CHILD; 336 kn->kn_flags &= ~EV_FLAG1; 337 } 338 339 if (immediate == 0) 340 knlist_add(&p->p_klist, kn, 1); 341 342 /* 343 * Immediately activate any exit notes if the target process is a 344 * zombie. This is necessary to handle the case where the target 345 * process, e.g. a child, dies before the kevent is registered. 346 */ 347 if (immediate && filt_proc(kn, NOTE_EXIT)) 348 KNOTE_ACTIVATE(kn, 0); 349 350 PROC_UNLOCK(p); 351 352 return (0); 353 } 354 355 /* 356 * The knote may be attached to a different process, which may exit, 357 * leaving nothing for the knote to be attached to. So when the process 358 * exits, the knote is marked as DETACHED and also flagged as ONESHOT so 359 * it will be deleted when read out. However, as part of the knote deletion, 360 * this routine is called, so a check is needed to avoid actually performing 361 * a detach, because the original process does not exist any more. 362 */ 363 /* XXX - move to kern_proc.c? */ 364 static void 365 filt_procdetach(struct knote *kn) 366 { 367 struct proc *p; 368 369 p = kn->kn_ptr.p_proc; 370 knlist_remove(&p->p_klist, kn, 0); 371 kn->kn_ptr.p_proc = NULL; 372 } 373 374 /* XXX - move to kern_proc.c? */ 375 static int 376 filt_proc(struct knote *kn, long hint) 377 { 378 struct proc *p = kn->kn_ptr.p_proc; 379 u_int event; 380 381 /* 382 * mask off extra data 383 */ 384 event = (u_int)hint & NOTE_PCTRLMASK; 385 386 /* 387 * if the user is interested in this event, record it. 388 */ 389 if (kn->kn_sfflags & event) 390 kn->kn_fflags |= event; 391 392 /* 393 * process is gone, so flag the event as finished. 394 */ 395 if (event == NOTE_EXIT) { 396 if (!(kn->kn_status & KN_DETACHED)) 397 knlist_remove_inevent(&p->p_klist, kn); 398 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 399 kn->kn_ptr.p_proc = NULL; 400 return (1); 401 } 402 403 /* 404 * process forked, and user wants to track the new process, 405 * so attach a new knote to it, and immediately report an 406 * event with the parent's pid. 407 */ 408 if ((event == NOTE_FORK) && (kn->kn_sfflags & NOTE_TRACK)) { 409 struct kevent kev; 410 int error; 411 412 /* 413 * register knote with new process. 414 */ 415 kev.ident = hint & NOTE_PDATAMASK; /* pid */ 416 kev.filter = kn->kn_filter; 417 kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1; 418 kev.fflags = kn->kn_sfflags; 419 kev.data = kn->kn_id; /* parent */ 420 kev.udata = kn->kn_kevent.udata; /* preserve udata */ 421 error = kqueue_register(kn->kn_kq, &kev, NULL, 0); 422 if (error) 423 kn->kn_fflags |= NOTE_TRACKERR; 424 } 425 426 return (kn->kn_fflags != 0); 427 } 428 429 static int 430 timertoticks(intptr_t data) 431 { 432 struct timeval tv; 433 int tticks; 434 435 tv.tv_sec = data / 1000; 436 tv.tv_usec = (data % 1000) * 1000; 437 tticks = tvtohz(&tv); 438 439 return tticks; 440 } 441 442 /* XXX - move to kern_timeout.c? */ 443 static void 444 filt_timerexpire(void *knx) 445 { 446 struct knote *kn = knx; 447 struct callout *calloutp; 448 449 kn->kn_data++; 450 KNOTE_ACTIVATE(kn, 0); /* XXX - handle locking */ 451 452 if ((kn->kn_flags & EV_ONESHOT) != EV_ONESHOT) { 453 calloutp = (struct callout *)kn->kn_hook; 454 callout_reset(calloutp, timertoticks(kn->kn_sdata), 455 filt_timerexpire, kn); 456 } 457 } 458 459 /* 460 * data contains amount of time to sleep, in milliseconds 461 */ 462 /* XXX - move to kern_timeout.c? */ 463 static int 464 filt_timerattach(struct knote *kn) 465 { 466 struct callout *calloutp; 467 468 atomic_add_int(&kq_ncallouts, 1); 469 470 if (kq_ncallouts >= kq_calloutmax) { 471 atomic_add_int(&kq_ncallouts, -1); 472 return (ENOMEM); 473 } 474 475 kn->kn_flags |= EV_CLEAR; /* automatically set */ 476 kn->kn_status &= ~KN_DETACHED; /* knlist_add usually sets it */ 477 MALLOC(calloutp, struct callout *, sizeof(*calloutp), 478 M_KQUEUE, M_WAITOK); 479 callout_init(calloutp, CALLOUT_MPSAFE); 480 kn->kn_hook = calloutp; 481 callout_reset(calloutp, timertoticks(kn->kn_sdata), filt_timerexpire, 482 kn); 483 484 return (0); 485 } 486 487 /* XXX - move to kern_timeout.c? */ 488 static void 489 filt_timerdetach(struct knote *kn) 490 { 491 struct callout *calloutp; 492 493 calloutp = (struct callout *)kn->kn_hook; 494 callout_drain(calloutp); 495 FREE(calloutp, M_KQUEUE); 496 atomic_add_int(&kq_ncallouts, -1); 497 kn->kn_status |= KN_DETACHED; /* knlist_remove usually clears it */ 498 } 499 500 /* XXX - move to kern_timeout.c? */ 501 static int 502 filt_timer(struct knote *kn, long hint) 503 { 504 505 return (kn->kn_data != 0); 506 } 507 508 /* 509 * MPSAFE 510 */ 511 int 512 kqueue(struct thread *td, struct kqueue_args *uap) 513 { 514 struct filedesc *fdp; 515 struct kqueue *kq; 516 struct file *fp; 517 int fd, error; 518 519 fdp = td->td_proc->p_fd; 520 error = falloc(td, &fp, &fd); 521 if (error) 522 goto done2; 523 524 /* An extra reference on `nfp' has been held for us by falloc(). */ 525 kq = malloc(sizeof *kq, M_KQUEUE, M_WAITOK | M_ZERO); 526 mtx_init(&kq->kq_lock, "kqueue", NULL, MTX_DEF|MTX_DUPOK); 527 TAILQ_INIT(&kq->kq_head); 528 kq->kq_fdp = fdp; 529 knlist_init(&kq->kq_sel.si_note, &kq->kq_lock, NULL, NULL, NULL); 530 TASK_INIT(&kq->kq_task, 0, kqueue_task, kq); 531 532 FILEDESC_LOCK_FAST(fdp); 533 SLIST_INSERT_HEAD(&fdp->fd_kqlist, kq, kq_list); 534 FILEDESC_UNLOCK_FAST(fdp); 535 536 FILE_LOCK(fp); 537 fp->f_flag = FREAD | FWRITE; 538 fp->f_type = DTYPE_KQUEUE; 539 fp->f_ops = &kqueueops; 540 fp->f_data = kq; 541 FILE_UNLOCK(fp); 542 fdrop(fp, td); 543 544 td->td_retval[0] = fd; 545 done2: 546 return (error); 547 } 548 549 #ifndef _SYS_SYSPROTO_H_ 550 struct kevent_args { 551 int fd; 552 const struct kevent *changelist; 553 int nchanges; 554 struct kevent *eventlist; 555 int nevents; 556 const struct timespec *timeout; 557 }; 558 #endif 559 /* 560 * MPSAFE 561 */ 562 int 563 kevent(struct thread *td, struct kevent_args *uap) 564 { 565 struct timespec ts, *tsp; 566 struct kevent_copyops k_ops = { uap, 567 kevent_copyout, 568 kevent_copyin}; 569 int error; 570 #ifdef KTRACE 571 struct uio ktruio; 572 struct iovec ktriov; 573 struct uio *ktruioin = NULL; 574 struct uio *ktruioout = NULL; 575 #endif 576 577 if (uap->timeout != NULL) { 578 error = copyin(uap->timeout, &ts, sizeof(ts)); 579 if (error) 580 return (error); 581 tsp = &ts; 582 } else 583 tsp = NULL; 584 585 #ifdef KTRACE 586 if (KTRPOINT(td, KTR_GENIO)) { 587 ktriov.iov_base = uap->changelist; 588 ktriov.iov_len = uap->nchanges * sizeof(struct kevent); 589 ktruio = (struct uio){ .uio_iov = &ktriov, .uio_iovcnt = 1, 590 .uio_segflg = UIO_USERSPACE, .uio_rw = UIO_READ, 591 .uio_td = td }; 592 ktruioin = cloneuio(&ktruio); 593 ktriov.iov_base = uap->eventlist; 594 ktriov.iov_len = uap->nevents * sizeof(struct kevent); 595 ktruioout = cloneuio(&ktruio); 596 } 597 #endif 598 599 error = kern_kevent(td, uap->fd, uap->nchanges, uap->nevents, 600 &k_ops, tsp); 601 602 #ifdef KTRACE 603 if (ktruioin != NULL) { 604 ktruioin->uio_resid = uap->nchanges * sizeof(struct kevent); 605 ktrgenio(uap->fd, UIO_WRITE, ktruioin, 0); 606 ktruioout->uio_resid = td->td_retval[0] * sizeof(struct kevent); 607 ktrgenio(uap->fd, UIO_READ, ktruioout, error); 608 } 609 #endif 610 611 return (error); 612 } 613 614 /* 615 * Copy 'count' items into the destination list pointed to by uap->eventlist. 616 */ 617 static int 618 kevent_copyout(void *arg, struct kevent *kevp, int count) 619 { 620 struct kevent_args *uap; 621 int error; 622 623 KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count)); 624 uap = (struct kevent_args *)arg; 625 626 error = copyout(kevp, uap->eventlist, count * sizeof *kevp); 627 if (error == 0) 628 uap->eventlist += count; 629 return (error); 630 } 631 632 /* 633 * Copy 'count' items from the list pointed to by uap->changelist. 634 */ 635 static int 636 kevent_copyin(void *arg, struct kevent *kevp, int count) 637 { 638 struct kevent_args *uap; 639 int error; 640 641 KASSERT(count <= KQ_NEVENTS, ("count (%d) > KQ_NEVENTS", count)); 642 uap = (struct kevent_args *)arg; 643 644 error = copyin(uap->changelist, kevp, count * sizeof *kevp); 645 if (error == 0) 646 uap->changelist += count; 647 return (error); 648 } 649 650 int 651 kern_kevent(struct thread *td, int fd, int nchanges, int nevents, 652 struct kevent_copyops *k_ops, const struct timespec *timeout) 653 { 654 struct kevent keva[KQ_NEVENTS]; 655 struct kevent *kevp, *changes; 656 struct kqueue *kq; 657 struct file *fp; 658 int i, n, nerrors, error; 659 660 if ((error = fget(td, fd, &fp)) != 0) 661 return (error); 662 if ((error = kqueue_aquire(fp, &kq)) != 0) 663 goto done_norel; 664 665 nerrors = 0; 666 667 while (nchanges > 0) { 668 n = nchanges > KQ_NEVENTS ? KQ_NEVENTS : nchanges; 669 error = k_ops->k_copyin(k_ops->arg, keva, n); 670 if (error) 671 goto done; 672 changes = keva; 673 for (i = 0; i < n; i++) { 674 kevp = &changes[i]; 675 if (!kevp->filter) 676 continue; 677 kevp->flags &= ~EV_SYSFLAGS; 678 error = kqueue_register(kq, kevp, td, 1); 679 if (error) { 680 if (nevents != 0) { 681 kevp->flags = EV_ERROR; 682 kevp->data = error; 683 (void) k_ops->k_copyout(k_ops->arg, 684 kevp, 1); 685 nevents--; 686 nerrors++; 687 } else { 688 goto done; 689 } 690 } 691 } 692 nchanges -= n; 693 } 694 if (nerrors) { 695 td->td_retval[0] = nerrors; 696 error = 0; 697 goto done; 698 } 699 700 error = kqueue_scan(kq, nevents, k_ops, timeout, keva, td); 701 done: 702 kqueue_release(kq, 0); 703 done_norel: 704 if (fp != NULL) 705 fdrop(fp, td); 706 return (error); 707 } 708 709 int 710 kqueue_add_filteropts(int filt, struct filterops *filtops) 711 { 712 int error; 713 714 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) { 715 printf( 716 "trying to add a filterop that is out of range: %d is beyond %d\n", 717 ~filt, EVFILT_SYSCOUNT); 718 return EINVAL; 719 } 720 mtx_lock(&filterops_lock); 721 if (sysfilt_ops[~filt].for_fop != &null_filtops && 722 sysfilt_ops[~filt].for_fop != NULL) 723 error = EEXIST; 724 else { 725 sysfilt_ops[~filt].for_fop = filtops; 726 sysfilt_ops[~filt].for_refcnt = 0; 727 } 728 mtx_unlock(&filterops_lock); 729 730 return (0); 731 } 732 733 int 734 kqueue_del_filteropts(int filt) 735 { 736 int error; 737 738 error = 0; 739 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) 740 return EINVAL; 741 742 mtx_lock(&filterops_lock); 743 if (sysfilt_ops[~filt].for_fop == &null_filtops || 744 sysfilt_ops[~filt].for_fop == NULL) 745 error = EINVAL; 746 else if (sysfilt_ops[~filt].for_refcnt != 0) 747 error = EBUSY; 748 else { 749 sysfilt_ops[~filt].for_fop = &null_filtops; 750 sysfilt_ops[~filt].for_refcnt = 0; 751 } 752 mtx_unlock(&filterops_lock); 753 754 return error; 755 } 756 757 static struct filterops * 758 kqueue_fo_find(int filt) 759 { 760 761 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) 762 return NULL; 763 764 mtx_lock(&filterops_lock); 765 sysfilt_ops[~filt].for_refcnt++; 766 if (sysfilt_ops[~filt].for_fop == NULL) 767 sysfilt_ops[~filt].for_fop = &null_filtops; 768 mtx_unlock(&filterops_lock); 769 770 return sysfilt_ops[~filt].for_fop; 771 } 772 773 static void 774 kqueue_fo_release(int filt) 775 { 776 777 if (filt > 0 || filt + EVFILT_SYSCOUNT < 0) 778 return; 779 780 mtx_lock(&filterops_lock); 781 KASSERT(sysfilt_ops[~filt].for_refcnt > 0, 782 ("filter object refcount not valid on release")); 783 sysfilt_ops[~filt].for_refcnt--; 784 mtx_unlock(&filterops_lock); 785 } 786 787 /* 788 * A ref to kq (obtained via kqueue_aquire) must be held. waitok will 789 * influence if memory allocation should wait. Make sure it is 0 if you 790 * hold any mutexes. 791 */ 792 static int 793 kqueue_register(struct kqueue *kq, struct kevent *kev, struct thread *td, int waitok) 794 { 795 struct filterops *fops; 796 struct file *fp; 797 struct knote *kn, *tkn; 798 int error, filt, event; 799 int haskqglobal; 800 801 fp = NULL; 802 kn = NULL; 803 error = 0; 804 haskqglobal = 0; 805 806 filt = kev->filter; 807 fops = kqueue_fo_find(filt); 808 if (fops == NULL) 809 return EINVAL; 810 811 tkn = knote_alloc(waitok); /* prevent waiting with locks */ 812 813 findkn: 814 if (fops->f_isfd) { 815 KASSERT(td != NULL, ("td is NULL")); 816 error = fget(td, kev->ident, &fp); 817 if (error) 818 goto done; 819 820 if ((kev->flags & EV_ADD) == EV_ADD && kqueue_expand(kq, fops, 821 kev->ident, 0) != 0) { 822 /* try again */ 823 fdrop(fp, td); 824 fp = NULL; 825 error = kqueue_expand(kq, fops, kev->ident, waitok); 826 if (error) 827 goto done; 828 goto findkn; 829 } 830 831 if (fp->f_type == DTYPE_KQUEUE) { 832 /* 833 * if we add some inteligence about what we are doing, 834 * we should be able to support events on ourselves. 835 * We need to know when we are doing this to prevent 836 * getting both the knlist lock and the kq lock since 837 * they are the same thing. 838 */ 839 if (fp->f_data == kq) { 840 error = EINVAL; 841 goto done; 842 } 843 844 KQ_GLOBAL_LOCK(&kq_global, haskqglobal); 845 } 846 847 KQ_LOCK(kq); 848 if (kev->ident < kq->kq_knlistsize) { 849 SLIST_FOREACH(kn, &kq->kq_knlist[kev->ident], kn_link) 850 if (kev->filter == kn->kn_filter) 851 break; 852 } 853 } else { 854 if ((kev->flags & EV_ADD) == EV_ADD) 855 kqueue_expand(kq, fops, kev->ident, waitok); 856 857 KQ_LOCK(kq); 858 if (kq->kq_knhashmask != 0) { 859 struct klist *list; 860 861 list = &kq->kq_knhash[ 862 KN_HASH((u_long)kev->ident, kq->kq_knhashmask)]; 863 SLIST_FOREACH(kn, list, kn_link) 864 if (kev->ident == kn->kn_id && 865 kev->filter == kn->kn_filter) 866 break; 867 } 868 } 869 870 /* knote is in the process of changing, wait for it to stablize. */ 871 if (kn != NULL && (kn->kn_status & KN_INFLUX) == KN_INFLUX) { 872 if (fp != NULL) { 873 fdrop(fp, td); 874 fp = NULL; 875 } 876 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal); 877 kq->kq_state |= KQ_FLUXWAIT; 878 msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqflxwt", 0); 879 goto findkn; 880 } 881 882 if (kn == NULL && ((kev->flags & EV_ADD) == 0)) { 883 KQ_UNLOCK(kq); 884 error = ENOENT; 885 goto done; 886 } 887 888 /* 889 * kn now contains the matching knote, or NULL if no match 890 */ 891 if (kev->flags & EV_ADD) { 892 if (kn == NULL) { 893 kn = tkn; 894 tkn = NULL; 895 if (kn == NULL) { 896 KQ_UNLOCK(kq); 897 error = ENOMEM; 898 goto done; 899 } 900 kn->kn_fp = fp; 901 kn->kn_kq = kq; 902 kn->kn_fop = fops; 903 /* 904 * apply reference counts to knote structure, and 905 * do not release it at the end of this routine. 906 */ 907 fops = NULL; 908 fp = NULL; 909 910 kn->kn_sfflags = kev->fflags; 911 kn->kn_sdata = kev->data; 912 kev->fflags = 0; 913 kev->data = 0; 914 kn->kn_kevent = *kev; 915 kn->kn_kevent.flags &= ~(EV_ADD | EV_DELETE | 916 EV_ENABLE | EV_DISABLE); 917 kn->kn_status = KN_INFLUX|KN_DETACHED; 918 919 error = knote_attach(kn, kq); 920 KQ_UNLOCK(kq); 921 if (error != 0) { 922 tkn = kn; 923 goto done; 924 } 925 926 if ((error = kn->kn_fop->f_attach(kn)) != 0) { 927 knote_drop(kn, td); 928 goto done; 929 } 930 KN_LIST_LOCK(kn); 931 } else { 932 /* 933 * The user may change some filter values after the 934 * initial EV_ADD, but doing so will not reset any 935 * filter which has already been triggered. 936 */ 937 kn->kn_status |= KN_INFLUX; 938 KQ_UNLOCK(kq); 939 KN_LIST_LOCK(kn); 940 kn->kn_sfflags = kev->fflags; 941 kn->kn_sdata = kev->data; 942 kn->kn_kevent.udata = kev->udata; 943 } 944 945 /* 946 * We can get here with kn->kn_knlist == NULL. 947 * This can happen when the initial attach event decides that 948 * the event is "completed" already. i.e. filt_procattach 949 * is called on a zombie process. It will call filt_proc 950 * which will remove it from the list, and NULL kn_knlist. 951 */ 952 event = kn->kn_fop->f_event(kn, 0); 953 KQ_LOCK(kq); 954 if (event) 955 KNOTE_ACTIVATE(kn, 1); 956 kn->kn_status &= ~KN_INFLUX; 957 KN_LIST_UNLOCK(kn); 958 } else if (kev->flags & EV_DELETE) { 959 kn->kn_status |= KN_INFLUX; 960 KQ_UNLOCK(kq); 961 if (!(kn->kn_status & KN_DETACHED)) 962 kn->kn_fop->f_detach(kn); 963 knote_drop(kn, td); 964 goto done; 965 } 966 967 if ((kev->flags & EV_DISABLE) && 968 ((kn->kn_status & KN_DISABLED) == 0)) { 969 kn->kn_status |= KN_DISABLED; 970 } 971 972 if ((kev->flags & EV_ENABLE) && (kn->kn_status & KN_DISABLED)) { 973 kn->kn_status &= ~KN_DISABLED; 974 if ((kn->kn_status & KN_ACTIVE) && 975 ((kn->kn_status & KN_QUEUED) == 0)) 976 knote_enqueue(kn); 977 } 978 KQ_UNLOCK_FLUX(kq); 979 980 done: 981 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal); 982 if (fp != NULL) 983 fdrop(fp, td); 984 if (tkn != NULL) 985 knote_free(tkn); 986 if (fops != NULL) 987 kqueue_fo_release(filt); 988 return (error); 989 } 990 991 static int 992 kqueue_aquire(struct file *fp, struct kqueue **kqp) 993 { 994 int error; 995 struct kqueue *kq; 996 997 error = 0; 998 999 FILE_LOCK(fp); 1000 do { 1001 kq = fp->f_data; 1002 if (fp->f_type != DTYPE_KQUEUE || kq == NULL) { 1003 error = EBADF; 1004 break; 1005 } 1006 *kqp = kq; 1007 KQ_LOCK(kq); 1008 if ((kq->kq_state & KQ_CLOSING) == KQ_CLOSING) { 1009 KQ_UNLOCK(kq); 1010 error = EBADF; 1011 break; 1012 } 1013 kq->kq_refcnt++; 1014 KQ_UNLOCK(kq); 1015 } while (0); 1016 FILE_UNLOCK(fp); 1017 1018 return error; 1019 } 1020 1021 static void 1022 kqueue_release(struct kqueue *kq, int locked) 1023 { 1024 if (locked) 1025 KQ_OWNED(kq); 1026 else 1027 KQ_LOCK(kq); 1028 kq->kq_refcnt--; 1029 if (kq->kq_refcnt == 1) 1030 wakeup(&kq->kq_refcnt); 1031 if (!locked) 1032 KQ_UNLOCK(kq); 1033 } 1034 1035 static void 1036 kqueue_schedtask(struct kqueue *kq) 1037 { 1038 1039 KQ_OWNED(kq); 1040 KASSERT(((kq->kq_state & KQ_TASKDRAIN) != KQ_TASKDRAIN), 1041 ("scheduling kqueue task while draining")); 1042 1043 if ((kq->kq_state & KQ_TASKSCHED) != KQ_TASKSCHED) { 1044 taskqueue_enqueue(taskqueue_kqueue, &kq->kq_task); 1045 kq->kq_state |= KQ_TASKSCHED; 1046 } 1047 } 1048 1049 /* 1050 * Expand the kq to make sure we have storage for fops/ident pair. 1051 * 1052 * Return 0 on success (or no work necessary), return errno on failure. 1053 * 1054 * Not calling hashinit w/ waitok (proper malloc flag) should be safe. 1055 * If kqueue_register is called from a non-fd context, there usually/should 1056 * be no locks held. 1057 */ 1058 static int 1059 kqueue_expand(struct kqueue *kq, struct filterops *fops, uintptr_t ident, 1060 int waitok) 1061 { 1062 struct klist *list, *tmp_knhash; 1063 u_long tmp_knhashmask; 1064 int size; 1065 int fd; 1066 int mflag = waitok ? M_WAITOK : M_NOWAIT; 1067 1068 KQ_NOTOWNED(kq); 1069 1070 if (fops->f_isfd) { 1071 fd = ident; 1072 if (kq->kq_knlistsize <= fd) { 1073 size = kq->kq_knlistsize; 1074 while (size <= fd) 1075 size += KQEXTENT; 1076 MALLOC(list, struct klist *, 1077 size * sizeof list, M_KQUEUE, mflag); 1078 if (list == NULL) 1079 return ENOMEM; 1080 KQ_LOCK(kq); 1081 if (kq->kq_knlistsize > fd) { 1082 FREE(list, M_KQUEUE); 1083 list = NULL; 1084 } else { 1085 if (kq->kq_knlist != NULL) { 1086 bcopy(kq->kq_knlist, list, 1087 kq->kq_knlistsize * sizeof list); 1088 FREE(kq->kq_knlist, M_KQUEUE); 1089 kq->kq_knlist = NULL; 1090 } 1091 bzero((caddr_t)list + 1092 kq->kq_knlistsize * sizeof list, 1093 (size - kq->kq_knlistsize) * sizeof list); 1094 kq->kq_knlistsize = size; 1095 kq->kq_knlist = list; 1096 } 1097 KQ_UNLOCK(kq); 1098 } 1099 } else { 1100 if (kq->kq_knhashmask == 0) { 1101 tmp_knhash = hashinit(KN_HASHSIZE, M_KQUEUE, 1102 &tmp_knhashmask); 1103 if (tmp_knhash == NULL) 1104 return ENOMEM; 1105 KQ_LOCK(kq); 1106 if (kq->kq_knhashmask == 0) { 1107 kq->kq_knhash = tmp_knhash; 1108 kq->kq_knhashmask = tmp_knhashmask; 1109 } else { 1110 free(tmp_knhash, M_KQUEUE); 1111 } 1112 KQ_UNLOCK(kq); 1113 } 1114 } 1115 1116 KQ_NOTOWNED(kq); 1117 return 0; 1118 } 1119 1120 static void 1121 kqueue_task(void *arg, int pending) 1122 { 1123 struct kqueue *kq; 1124 int haskqglobal; 1125 1126 haskqglobal = 0; 1127 kq = arg; 1128 1129 KQ_GLOBAL_LOCK(&kq_global, haskqglobal); 1130 KQ_LOCK(kq); 1131 1132 KNOTE_LOCKED(&kq->kq_sel.si_note, 0); 1133 1134 kq->kq_state &= ~KQ_TASKSCHED; 1135 if ((kq->kq_state & KQ_TASKDRAIN) == KQ_TASKDRAIN) { 1136 wakeup(&kq->kq_state); 1137 } 1138 KQ_UNLOCK(kq); 1139 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal); 1140 } 1141 1142 /* 1143 * Scan, update kn_data (if not ONESHOT), and copyout triggered events. 1144 * We treat KN_MARKER knotes as if they are INFLUX. 1145 */ 1146 static int 1147 kqueue_scan(struct kqueue *kq, int maxevents, struct kevent_copyops *k_ops, 1148 const struct timespec *tsp, struct kevent *keva, struct thread *td) 1149 { 1150 struct kevent *kevp; 1151 struct timeval atv, rtv, ttv; 1152 struct knote *kn, *marker; 1153 int count, timeout, nkev, error; 1154 int haskqglobal; 1155 1156 count = maxevents; 1157 nkev = 0; 1158 error = 0; 1159 haskqglobal = 0; 1160 1161 if (maxevents == 0) 1162 goto done_nl; 1163 1164 if (tsp != NULL) { 1165 TIMESPEC_TO_TIMEVAL(&atv, tsp); 1166 if (itimerfix(&atv)) { 1167 error = EINVAL; 1168 goto done_nl; 1169 } 1170 if (tsp->tv_sec == 0 && tsp->tv_nsec == 0) 1171 timeout = -1; 1172 else 1173 timeout = atv.tv_sec > 24 * 60 * 60 ? 1174 24 * 60 * 60 * hz : tvtohz(&atv); 1175 getmicrouptime(&rtv); 1176 timevaladd(&atv, &rtv); 1177 } else { 1178 atv.tv_sec = 0; 1179 atv.tv_usec = 0; 1180 timeout = 0; 1181 } 1182 marker = knote_alloc(1); 1183 if (marker == NULL) { 1184 error = ENOMEM; 1185 goto done_nl; 1186 } 1187 marker->kn_status = KN_MARKER; 1188 KQ_LOCK(kq); 1189 goto start; 1190 1191 retry: 1192 if (atv.tv_sec || atv.tv_usec) { 1193 getmicrouptime(&rtv); 1194 if (timevalcmp(&rtv, &atv, >=)) 1195 goto done; 1196 ttv = atv; 1197 timevalsub(&ttv, &rtv); 1198 timeout = ttv.tv_sec > 24 * 60 * 60 ? 1199 24 * 60 * 60 * hz : tvtohz(&ttv); 1200 } 1201 1202 start: 1203 kevp = keva; 1204 if (kq->kq_count == 0) { 1205 if (timeout < 0) { 1206 error = EWOULDBLOCK; 1207 } else { 1208 kq->kq_state |= KQ_SLEEP; 1209 error = msleep(kq, &kq->kq_lock, PSOCK | PCATCH, 1210 "kqread", timeout); 1211 } 1212 if (error == 0) 1213 goto retry; 1214 /* don't restart after signals... */ 1215 if (error == ERESTART) 1216 error = EINTR; 1217 else if (error == EWOULDBLOCK) 1218 error = 0; 1219 goto done; 1220 } 1221 1222 TAILQ_INSERT_TAIL(&kq->kq_head, marker, kn_tqe); 1223 while (count) { 1224 KQ_OWNED(kq); 1225 kn = TAILQ_FIRST(&kq->kq_head); 1226 1227 if ((kn->kn_status == KN_MARKER && kn != marker) || 1228 (kn->kn_status & KN_INFLUX) == KN_INFLUX) { 1229 kq->kq_state |= KQ_FLUXWAIT; 1230 error = msleep(kq, &kq->kq_lock, PSOCK, 1231 "kqflxwt", 0); 1232 continue; 1233 } 1234 1235 TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe); 1236 if ((kn->kn_status & KN_DISABLED) == KN_DISABLED) { 1237 kn->kn_status &= ~KN_QUEUED; 1238 kq->kq_count--; 1239 continue; 1240 } 1241 if (kn == marker) { 1242 KQ_FLUX_WAKEUP(kq); 1243 if (count == maxevents) 1244 goto retry; 1245 goto done; 1246 } 1247 KASSERT((kn->kn_status & KN_INFLUX) == 0, 1248 ("KN_INFLUX set when not suppose to be")); 1249 1250 if ((kn->kn_flags & EV_ONESHOT) == EV_ONESHOT) { 1251 kn->kn_status &= ~KN_QUEUED; 1252 kn->kn_status |= KN_INFLUX; 1253 kq->kq_count--; 1254 KQ_UNLOCK(kq); 1255 /* 1256 * We don't need to lock the list since we've marked 1257 * it _INFLUX. 1258 */ 1259 *kevp = kn->kn_kevent; 1260 if (!(kn->kn_status & KN_DETACHED)) 1261 kn->kn_fop->f_detach(kn); 1262 knote_drop(kn, td); 1263 KQ_LOCK(kq); 1264 kn = NULL; 1265 } else { 1266 kn->kn_status |= KN_INFLUX; 1267 KQ_UNLOCK(kq); 1268 if ((kn->kn_status & KN_KQUEUE) == KN_KQUEUE) 1269 KQ_GLOBAL_LOCK(&kq_global, haskqglobal); 1270 KN_LIST_LOCK(kn); 1271 if (kn->kn_fop->f_event(kn, 0) == 0) { 1272 KQ_LOCK(kq); 1273 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal); 1274 kn->kn_status &= 1275 ~(KN_QUEUED | KN_ACTIVE | KN_INFLUX); 1276 kq->kq_count--; 1277 KN_LIST_UNLOCK(kn); 1278 continue; 1279 } 1280 *kevp = kn->kn_kevent; 1281 KQ_LOCK(kq); 1282 KQ_GLOBAL_UNLOCK(&kq_global, haskqglobal); 1283 if (kn->kn_flags & EV_CLEAR) { 1284 kn->kn_data = 0; 1285 kn->kn_fflags = 0; 1286 kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE); 1287 kq->kq_count--; 1288 } else 1289 TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe); 1290 1291 kn->kn_status &= ~(KN_INFLUX); 1292 KN_LIST_UNLOCK(kn); 1293 } 1294 1295 /* we are returning a copy to the user */ 1296 kevp++; 1297 nkev++; 1298 count--; 1299 1300 if (nkev == KQ_NEVENTS) { 1301 KQ_UNLOCK_FLUX(kq); 1302 error = k_ops->k_copyout(k_ops->arg, keva, nkev); 1303 nkev = 0; 1304 kevp = keva; 1305 KQ_LOCK(kq); 1306 if (error) 1307 break; 1308 } 1309 } 1310 TAILQ_REMOVE(&kq->kq_head, marker, kn_tqe); 1311 done: 1312 KQ_OWNED(kq); 1313 KQ_UNLOCK_FLUX(kq); 1314 knote_free(marker); 1315 done_nl: 1316 KQ_NOTOWNED(kq); 1317 if (nkev != 0) 1318 error = k_ops->k_copyout(k_ops->arg, keva, nkev); 1319 td->td_retval[0] = maxevents - count; 1320 return (error); 1321 } 1322 1323 /* 1324 * XXX 1325 * This could be expanded to call kqueue_scan, if desired. 1326 */ 1327 /*ARGSUSED*/ 1328 static int 1329 kqueue_read(struct file *fp, struct uio *uio, struct ucred *active_cred, 1330 int flags, struct thread *td) 1331 { 1332 return (ENXIO); 1333 } 1334 1335 /*ARGSUSED*/ 1336 static int 1337 kqueue_write(struct file *fp, struct uio *uio, struct ucred *active_cred, 1338 int flags, struct thread *td) 1339 { 1340 return (ENXIO); 1341 } 1342 1343 /*ARGSUSED*/ 1344 static int 1345 kqueue_ioctl(struct file *fp, u_long cmd, void *data, 1346 struct ucred *active_cred, struct thread *td) 1347 { 1348 /* 1349 * Enabling sigio causes two major problems: 1350 * 1) infinite recursion: 1351 * Synopsys: kevent is being used to track signals and have FIOASYNC 1352 * set. On receipt of a signal this will cause a kqueue to recurse 1353 * into itself over and over. Sending the sigio causes the kqueue 1354 * to become ready, which in turn posts sigio again, forever. 1355 * Solution: this can be solved by setting a flag in the kqueue that 1356 * we have a SIGIO in progress. 1357 * 2) locking problems: 1358 * Synopsys: Kqueue is a leaf subsystem, but adding signalling puts 1359 * us above the proc and pgrp locks. 1360 * Solution: Post a signal using an async mechanism, being sure to 1361 * record a generation count in the delivery so that we do not deliver 1362 * a signal to the wrong process. 1363 * 1364 * Note, these two mechanisms are somewhat mutually exclusive! 1365 */ 1366 #if 0 1367 struct kqueue *kq; 1368 1369 kq = fp->f_data; 1370 switch (cmd) { 1371 case FIOASYNC: 1372 if (*(int *)data) { 1373 kq->kq_state |= KQ_ASYNC; 1374 } else { 1375 kq->kq_state &= ~KQ_ASYNC; 1376 } 1377 return (0); 1378 1379 case FIOSETOWN: 1380 return (fsetown(*(int *)data, &kq->kq_sigio)); 1381 1382 case FIOGETOWN: 1383 *(int *)data = fgetown(&kq->kq_sigio); 1384 return (0); 1385 } 1386 #endif 1387 1388 return (ENOTTY); 1389 } 1390 1391 /*ARGSUSED*/ 1392 static int 1393 kqueue_poll(struct file *fp, int events, struct ucred *active_cred, 1394 struct thread *td) 1395 { 1396 struct kqueue *kq; 1397 int revents = 0; 1398 int error; 1399 1400 if ((error = kqueue_aquire(fp, &kq))) 1401 return POLLERR; 1402 1403 KQ_LOCK(kq); 1404 if (events & (POLLIN | POLLRDNORM)) { 1405 if (kq->kq_count) { 1406 revents |= events & (POLLIN | POLLRDNORM); 1407 } else { 1408 selrecord(td, &kq->kq_sel); 1409 kq->kq_state |= KQ_SEL; 1410 } 1411 } 1412 kqueue_release(kq, 1); 1413 KQ_UNLOCK(kq); 1414 return (revents); 1415 } 1416 1417 /*ARGSUSED*/ 1418 static int 1419 kqueue_stat(struct file *fp, struct stat *st, struct ucred *active_cred, 1420 struct thread *td) 1421 { 1422 1423 bzero((void *)st, sizeof *st); 1424 /* 1425 * We no longer return kq_count because the unlocked value is useless. 1426 * If you spent all this time getting the count, why not spend your 1427 * syscall better by calling kevent? 1428 * 1429 * XXX - This is needed for libc_r. 1430 */ 1431 st->st_mode = S_IFIFO; 1432 return (0); 1433 } 1434 1435 /*ARGSUSED*/ 1436 static int 1437 kqueue_close(struct file *fp, struct thread *td) 1438 { 1439 struct kqueue *kq = fp->f_data; 1440 struct filedesc *fdp; 1441 struct knote *kn; 1442 int i; 1443 int error; 1444 1445 if ((error = kqueue_aquire(fp, &kq))) 1446 return error; 1447 1448 KQ_LOCK(kq); 1449 1450 KASSERT((kq->kq_state & KQ_CLOSING) != KQ_CLOSING, 1451 ("kqueue already closing")); 1452 kq->kq_state |= KQ_CLOSING; 1453 if (kq->kq_refcnt > 1) 1454 msleep(&kq->kq_refcnt, &kq->kq_lock, PSOCK, "kqclose", 0); 1455 1456 KASSERT(kq->kq_refcnt == 1, ("other refs are out there!")); 1457 fdp = kq->kq_fdp; 1458 1459 KASSERT(knlist_empty(&kq->kq_sel.si_note), 1460 ("kqueue's knlist not empty")); 1461 1462 for (i = 0; i < kq->kq_knlistsize; i++) { 1463 while ((kn = SLIST_FIRST(&kq->kq_knlist[i])) != NULL) { 1464 KASSERT((kn->kn_status & KN_INFLUX) == 0, 1465 ("KN_INFLUX set when not suppose to be")); 1466 kn->kn_status |= KN_INFLUX; 1467 KQ_UNLOCK(kq); 1468 if (!(kn->kn_status & KN_DETACHED)) 1469 kn->kn_fop->f_detach(kn); 1470 knote_drop(kn, td); 1471 KQ_LOCK(kq); 1472 } 1473 } 1474 if (kq->kq_knhashmask != 0) { 1475 for (i = 0; i <= kq->kq_knhashmask; i++) { 1476 while ((kn = SLIST_FIRST(&kq->kq_knhash[i])) != NULL) { 1477 KASSERT((kn->kn_status & KN_INFLUX) == 0, 1478 ("KN_INFLUX set when not suppose to be")); 1479 kn->kn_status |= KN_INFLUX; 1480 KQ_UNLOCK(kq); 1481 if (!(kn->kn_status & KN_DETACHED)) 1482 kn->kn_fop->f_detach(kn); 1483 knote_drop(kn, td); 1484 KQ_LOCK(kq); 1485 } 1486 } 1487 } 1488 1489 if ((kq->kq_state & KQ_TASKSCHED) == KQ_TASKSCHED) { 1490 kq->kq_state |= KQ_TASKDRAIN; 1491 msleep(&kq->kq_state, &kq->kq_lock, PSOCK, "kqtqdr", 0); 1492 } 1493 1494 if ((kq->kq_state & KQ_SEL) == KQ_SEL) { 1495 kq->kq_state &= ~KQ_SEL; 1496 selwakeuppri(&kq->kq_sel, PSOCK); 1497 } 1498 1499 KQ_UNLOCK(kq); 1500 1501 FILEDESC_LOCK_FAST(fdp); 1502 SLIST_REMOVE(&fdp->fd_kqlist, kq, kqueue, kq_list); 1503 FILEDESC_UNLOCK_FAST(fdp); 1504 1505 knlist_destroy(&kq->kq_sel.si_note); 1506 mtx_destroy(&kq->kq_lock); 1507 kq->kq_fdp = NULL; 1508 1509 if (kq->kq_knhash != NULL) 1510 free(kq->kq_knhash, M_KQUEUE); 1511 if (kq->kq_knlist != NULL) 1512 free(kq->kq_knlist, M_KQUEUE); 1513 1514 funsetown(&kq->kq_sigio); 1515 free(kq, M_KQUEUE); 1516 fp->f_data = NULL; 1517 1518 return (0); 1519 } 1520 1521 static void 1522 kqueue_wakeup(struct kqueue *kq) 1523 { 1524 KQ_OWNED(kq); 1525 1526 if ((kq->kq_state & KQ_SLEEP) == KQ_SLEEP) { 1527 kq->kq_state &= ~KQ_SLEEP; 1528 wakeup(kq); 1529 } 1530 if ((kq->kq_state & KQ_SEL) == KQ_SEL) { 1531 kq->kq_state &= ~KQ_SEL; 1532 selwakeuppri(&kq->kq_sel, PSOCK); 1533 } 1534 if (!knlist_empty(&kq->kq_sel.si_note)) 1535 kqueue_schedtask(kq); 1536 if ((kq->kq_state & KQ_ASYNC) == KQ_ASYNC) { 1537 pgsigio(&kq->kq_sigio, SIGIO, 0); 1538 } 1539 } 1540 1541 /* 1542 * Walk down a list of knotes, activating them if their event has triggered. 1543 * 1544 * There is a possibility to optimize in the case of one kq watching another. 1545 * Instead of scheduling a task to wake it up, you could pass enough state 1546 * down the chain to make up the parent kqueue. Make this code functional 1547 * first. 1548 */ 1549 void 1550 knote(struct knlist *list, long hint, int islocked) 1551 { 1552 struct kqueue *kq; 1553 struct knote *kn; 1554 1555 if (list == NULL) 1556 return; 1557 1558 KNL_ASSERT_LOCK(list, islocked); 1559 1560 if (!islocked) 1561 list->kl_lock(list->kl_lockarg); 1562 1563 /* 1564 * If we unlock the list lock (and set KN_INFLUX), we can eliminate 1565 * the kqueue scheduling, but this will introduce four 1566 * lock/unlock's for each knote to test. If we do, continue to use 1567 * SLIST_FOREACH, SLIST_FOREACH_SAFE is not safe in our case, it is 1568 * only safe if you want to remove the current item, which we are 1569 * not doing. 1570 */ 1571 SLIST_FOREACH(kn, &list->kl_list, kn_selnext) { 1572 kq = kn->kn_kq; 1573 if ((kn->kn_status & KN_INFLUX) != KN_INFLUX) { 1574 KQ_LOCK(kq); 1575 if ((kn->kn_status & KN_INFLUX) != KN_INFLUX) { 1576 kn->kn_status |= KN_HASKQLOCK; 1577 if (kn->kn_fop->f_event(kn, hint)) 1578 KNOTE_ACTIVATE(kn, 1); 1579 kn->kn_status &= ~KN_HASKQLOCK; 1580 } 1581 KQ_UNLOCK(kq); 1582 } 1583 kq = NULL; 1584 } 1585 if (!islocked) 1586 list->kl_unlock(list->kl_lockarg); 1587 } 1588 1589 /* 1590 * add a knote to a knlist 1591 */ 1592 void 1593 knlist_add(struct knlist *knl, struct knote *kn, int islocked) 1594 { 1595 KNL_ASSERT_LOCK(knl, islocked); 1596 KQ_NOTOWNED(kn->kn_kq); 1597 KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) == 1598 (KN_INFLUX|KN_DETACHED), ("knote not KN_INFLUX and KN_DETACHED")); 1599 if (!islocked) 1600 knl->kl_lock(knl->kl_lockarg); 1601 SLIST_INSERT_HEAD(&knl->kl_list, kn, kn_selnext); 1602 if (!islocked) 1603 knl->kl_unlock(knl->kl_lockarg); 1604 KQ_LOCK(kn->kn_kq); 1605 kn->kn_knlist = knl; 1606 kn->kn_status &= ~KN_DETACHED; 1607 KQ_UNLOCK(kn->kn_kq); 1608 } 1609 1610 static void 1611 knlist_remove_kq(struct knlist *knl, struct knote *kn, int knlislocked, int kqislocked) 1612 { 1613 KASSERT(!(!!kqislocked && !knlislocked), ("kq locked w/o knl locked")); 1614 KNL_ASSERT_LOCK(knl, knlislocked); 1615 mtx_assert(&kn->kn_kq->kq_lock, kqislocked ? MA_OWNED : MA_NOTOWNED); 1616 if (!kqislocked) 1617 KASSERT((kn->kn_status & (KN_INFLUX|KN_DETACHED)) == KN_INFLUX, 1618 ("knlist_remove called w/o knote being KN_INFLUX or already removed")); 1619 if (!knlislocked) 1620 knl->kl_lock(knl->kl_lockarg); 1621 SLIST_REMOVE(&knl->kl_list, kn, knote, kn_selnext); 1622 kn->kn_knlist = NULL; 1623 if (!knlislocked) 1624 knl->kl_unlock(knl->kl_lockarg); 1625 if (!kqislocked) 1626 KQ_LOCK(kn->kn_kq); 1627 kn->kn_status |= KN_DETACHED; 1628 if (!kqislocked) 1629 KQ_UNLOCK(kn->kn_kq); 1630 } 1631 1632 /* 1633 * remove all knotes from a specified klist 1634 */ 1635 void 1636 knlist_remove(struct knlist *knl, struct knote *kn, int islocked) 1637 { 1638 1639 knlist_remove_kq(knl, kn, islocked, 0); 1640 } 1641 1642 /* 1643 * remove knote from a specified klist while in f_event handler. 1644 */ 1645 void 1646 knlist_remove_inevent(struct knlist *knl, struct knote *kn) 1647 { 1648 1649 knlist_remove_kq(knl, kn, 1, 1650 (kn->kn_status & KN_HASKQLOCK) == KN_HASKQLOCK); 1651 } 1652 1653 int 1654 knlist_empty(struct knlist *knl) 1655 { 1656 KNL_ASSERT_LOCKED(knl); 1657 return SLIST_EMPTY(&knl->kl_list); 1658 } 1659 1660 static struct mtx knlist_lock; 1661 MTX_SYSINIT(knlist_lock, &knlist_lock, "knlist lock for lockless objects", 1662 MTX_DEF); 1663 static void knlist_mtx_lock(void *arg); 1664 static void knlist_mtx_unlock(void *arg); 1665 static int knlist_mtx_locked(void *arg); 1666 1667 static void 1668 knlist_mtx_lock(void *arg) 1669 { 1670 mtx_lock((struct mtx *)arg); 1671 } 1672 1673 static void 1674 knlist_mtx_unlock(void *arg) 1675 { 1676 mtx_unlock((struct mtx *)arg); 1677 } 1678 1679 static int 1680 knlist_mtx_locked(void *arg) 1681 { 1682 return (mtx_owned((struct mtx *)arg)); 1683 } 1684 1685 void 1686 knlist_init(struct knlist *knl, void *lock, void (*kl_lock)(void *), 1687 void (*kl_unlock)(void *), int (*kl_locked)(void *)) 1688 { 1689 1690 if (lock == NULL) 1691 knl->kl_lockarg = &knlist_lock; 1692 else 1693 knl->kl_lockarg = lock; 1694 1695 if (kl_lock == NULL) 1696 knl->kl_lock = knlist_mtx_lock; 1697 else 1698 knl->kl_lock = kl_lock; 1699 if (kl_unlock == NULL) 1700 knl->kl_unlock = knlist_mtx_unlock; 1701 else 1702 knl->kl_unlock = kl_unlock; 1703 if (kl_locked == NULL) 1704 knl->kl_locked = knlist_mtx_locked; 1705 else 1706 knl->kl_locked = kl_locked; 1707 1708 SLIST_INIT(&knl->kl_list); 1709 } 1710 1711 void 1712 knlist_destroy(struct knlist *knl) 1713 { 1714 1715 #ifdef INVARIANTS 1716 /* 1717 * if we run across this error, we need to find the offending 1718 * driver and have it call knlist_clear. 1719 */ 1720 if (!SLIST_EMPTY(&knl->kl_list)) 1721 printf("WARNING: destroying knlist w/ knotes on it!\n"); 1722 #endif 1723 1724 knl->kl_lockarg = knl->kl_lock = knl->kl_unlock = NULL; 1725 SLIST_INIT(&knl->kl_list); 1726 } 1727 1728 /* 1729 * Even if we are locked, we may need to drop the lock to allow any influx 1730 * knotes time to "settle". 1731 */ 1732 void 1733 knlist_cleardel(struct knlist *knl, struct thread *td, int islocked, int killkn) 1734 { 1735 struct knote *kn, *kn2; 1736 struct kqueue *kq; 1737 1738 if (islocked) 1739 KNL_ASSERT_LOCKED(knl); 1740 else { 1741 KNL_ASSERT_UNLOCKED(knl); 1742 again: /* need to reaquire lock since we have dropped it */ 1743 knl->kl_lock(knl->kl_lockarg); 1744 } 1745 1746 SLIST_FOREACH_SAFE(kn, &knl->kl_list, kn_selnext, kn2) { 1747 kq = kn->kn_kq; 1748 KQ_LOCK(kq); 1749 if ((kn->kn_status & KN_INFLUX)) { 1750 KQ_UNLOCK(kq); 1751 continue; 1752 } 1753 knlist_remove_kq(knl, kn, 1, 1); 1754 if (killkn) { 1755 kn->kn_status |= KN_INFLUX | KN_DETACHED; 1756 KQ_UNLOCK(kq); 1757 knote_drop(kn, td); 1758 } else { 1759 /* Make sure cleared knotes disappear soon */ 1760 kn->kn_flags |= (EV_EOF | EV_ONESHOT); 1761 KQ_UNLOCK(kq); 1762 } 1763 kq = NULL; 1764 } 1765 1766 if (!SLIST_EMPTY(&knl->kl_list)) { 1767 /* there are still KN_INFLUX remaining */ 1768 kn = SLIST_FIRST(&knl->kl_list); 1769 kq = kn->kn_kq; 1770 KQ_LOCK(kq); 1771 KASSERT(kn->kn_status & KN_INFLUX, 1772 ("knote removed w/o list lock")); 1773 knl->kl_unlock(knl->kl_lockarg); 1774 kq->kq_state |= KQ_FLUXWAIT; 1775 msleep(kq, &kq->kq_lock, PSOCK | PDROP, "kqkclr", 0); 1776 kq = NULL; 1777 goto again; 1778 } 1779 1780 if (islocked) 1781 KNL_ASSERT_LOCKED(knl); 1782 else { 1783 knl->kl_unlock(knl->kl_lockarg); 1784 KNL_ASSERT_UNLOCKED(knl); 1785 } 1786 } 1787 1788 /* 1789 * remove all knotes referencing a specified fd 1790 * must be called with FILEDESC lock. This prevents a race where a new fd 1791 * comes along and occupies the entry and we attach a knote to the fd. 1792 */ 1793 void 1794 knote_fdclose(struct thread *td, int fd) 1795 { 1796 struct filedesc *fdp = td->td_proc->p_fd; 1797 struct kqueue *kq; 1798 struct knote *kn; 1799 int influx; 1800 1801 FILEDESC_LOCK_ASSERT(fdp, MA_OWNED); 1802 1803 /* 1804 * We shouldn't have to worry about new kevents appearing on fd 1805 * since filedesc is locked. 1806 */ 1807 SLIST_FOREACH(kq, &fdp->fd_kqlist, kq_list) { 1808 KQ_LOCK(kq); 1809 1810 again: 1811 influx = 0; 1812 while (kq->kq_knlistsize > fd && 1813 (kn = SLIST_FIRST(&kq->kq_knlist[fd])) != NULL) { 1814 if (kn->kn_status & KN_INFLUX) { 1815 /* someone else might be waiting on our knote */ 1816 if (influx) 1817 wakeup(kq); 1818 kq->kq_state |= KQ_FLUXWAIT; 1819 msleep(kq, &kq->kq_lock, PSOCK, "kqflxwt", 0); 1820 goto again; 1821 } 1822 kn->kn_status |= KN_INFLUX; 1823 KQ_UNLOCK(kq); 1824 if (!(kn->kn_status & KN_DETACHED)) 1825 kn->kn_fop->f_detach(kn); 1826 knote_drop(kn, td); 1827 influx = 1; 1828 KQ_LOCK(kq); 1829 } 1830 KQ_UNLOCK_FLUX(kq); 1831 } 1832 } 1833 1834 static int 1835 knote_attach(struct knote *kn, struct kqueue *kq) 1836 { 1837 struct klist *list; 1838 1839 KASSERT(kn->kn_status & KN_INFLUX, ("knote not marked INFLUX")); 1840 KQ_OWNED(kq); 1841 1842 if (kn->kn_fop->f_isfd) { 1843 if (kn->kn_id >= kq->kq_knlistsize) 1844 return ENOMEM; 1845 list = &kq->kq_knlist[kn->kn_id]; 1846 } else { 1847 if (kq->kq_knhash == NULL) 1848 return ENOMEM; 1849 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)]; 1850 } 1851 1852 SLIST_INSERT_HEAD(list, kn, kn_link); 1853 1854 return 0; 1855 } 1856 1857 /* 1858 * knote must already have been detached using the f_detach method. 1859 * no lock need to be held, it is assumed that the KN_INFLUX flag is set 1860 * to prevent other removal. 1861 */ 1862 static void 1863 knote_drop(struct knote *kn, struct thread *td) 1864 { 1865 struct kqueue *kq; 1866 struct klist *list; 1867 1868 kq = kn->kn_kq; 1869 1870 KQ_NOTOWNED(kq); 1871 KASSERT((kn->kn_status & KN_INFLUX) == KN_INFLUX, 1872 ("knote_drop called without KN_INFLUX set in kn_status")); 1873 1874 KQ_LOCK(kq); 1875 if (kn->kn_fop->f_isfd) 1876 list = &kq->kq_knlist[kn->kn_id]; 1877 else 1878 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)]; 1879 1880 if (!SLIST_EMPTY(list)) 1881 SLIST_REMOVE(list, kn, knote, kn_link); 1882 if (kn->kn_status & KN_QUEUED) 1883 knote_dequeue(kn); 1884 KQ_UNLOCK_FLUX(kq); 1885 1886 if (kn->kn_fop->f_isfd) { 1887 fdrop(kn->kn_fp, td); 1888 kn->kn_fp = NULL; 1889 } 1890 kqueue_fo_release(kn->kn_kevent.filter); 1891 kn->kn_fop = NULL; 1892 knote_free(kn); 1893 } 1894 1895 static void 1896 knote_enqueue(struct knote *kn) 1897 { 1898 struct kqueue *kq = kn->kn_kq; 1899 1900 KQ_OWNED(kn->kn_kq); 1901 KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued")); 1902 1903 TAILQ_INSERT_TAIL(&kq->kq_head, kn, kn_tqe); 1904 kn->kn_status |= KN_QUEUED; 1905 kq->kq_count++; 1906 kqueue_wakeup(kq); 1907 } 1908 1909 static void 1910 knote_dequeue(struct knote *kn) 1911 { 1912 struct kqueue *kq = kn->kn_kq; 1913 1914 KQ_OWNED(kn->kn_kq); 1915 KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued")); 1916 1917 TAILQ_REMOVE(&kq->kq_head, kn, kn_tqe); 1918 kn->kn_status &= ~KN_QUEUED; 1919 kq->kq_count--; 1920 } 1921 1922 static void 1923 knote_init(void) 1924 { 1925 1926 knote_zone = uma_zcreate("KNOTE", sizeof(struct knote), NULL, NULL, 1927 NULL, NULL, UMA_ALIGN_PTR, 0); 1928 } 1929 SYSINIT(knote, SI_SUB_PSEUDO, SI_ORDER_ANY, knote_init, NULL) 1930 1931 static struct knote * 1932 knote_alloc(int waitok) 1933 { 1934 return ((struct knote *)uma_zalloc(knote_zone, 1935 (waitok ? M_WAITOK : M_NOWAIT)|M_ZERO)); 1936 } 1937 1938 static void 1939 knote_free(struct knote *kn) 1940 { 1941 if (kn != NULL) 1942 uma_zfree(knote_zone, kn); 1943 } 1944 1945 /* 1946 * Register the kev w/ the kq specified by fd. 1947 */ 1948 int 1949 kqfd_register(int fd, struct kevent *kev, struct thread *td, int waitok) 1950 { 1951 struct kqueue *kq; 1952 struct file *fp; 1953 int error; 1954 1955 if ((error = fget(td, fd, &fp)) != 0) 1956 return (error); 1957 if ((error = kqueue_aquire(fp, &kq)) != 0) 1958 goto noaquire; 1959 1960 error = kqueue_register(kq, kev, td, waitok); 1961 1962 kqueue_release(kq, 0); 1963 1964 noaquire: 1965 fdrop(fp, td); 1966 1967 return error; 1968 } 1969