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