1 /*- 2 * Copyright (c) 1999,2000,2001 Jonathan Lemon <jlemon@FreeBSD.org> 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice, this list of conditions and the following disclaimer. 10 * 2. Redistributions in binary form must reproduce the above copyright 11 * notice, this list of conditions and the following disclaimer in the 12 * documentation and/or other materials provided with the distribution. 13 * 14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 24 * SUCH DAMAGE. 25 * 26 * $FreeBSD: src/sys/kern/kern_event.c,v 1.2.2.10 2004/04/04 07:03:14 cperciva Exp $ 27 */ 28 29 #include <sys/param.h> 30 #include <sys/systm.h> 31 #include <sys/kernel.h> 32 #include <sys/proc.h> 33 #include <sys/malloc.h> 34 #include <sys/unistd.h> 35 #include <sys/file.h> 36 #include <sys/lock.h> 37 #include <sys/fcntl.h> 38 #include <sys/queue.h> 39 #include <sys/event.h> 40 #include <sys/eventvar.h> 41 #include <sys/protosw.h> 42 #include <sys/socket.h> 43 #include <sys/socketvar.h> 44 #include <sys/stat.h> 45 #include <sys/sysctl.h> 46 #include <sys/sysproto.h> 47 #include <sys/thread.h> 48 #include <sys/uio.h> 49 #include <sys/signalvar.h> 50 #include <sys/filio.h> 51 #include <sys/ktr.h> 52 53 #include <sys/thread2.h> 54 #include <sys/file2.h> 55 #include <sys/mplock2.h> 56 57 #define EVENT_REGISTER 1 58 #define EVENT_PROCESS 2 59 60 MALLOC_DEFINE(M_KQUEUE, "kqueue", "memory for kqueue system"); 61 62 struct kevent_copyin_args { 63 struct kevent_args *ka; 64 int pchanges; 65 }; 66 67 #define KNOTE_CACHE_MAX 8 68 69 struct knote_cache_list { 70 struct klist knote_cache; 71 int knote_cache_cnt; 72 } __cachealign; 73 74 static int kqueue_scan(struct kqueue *kq, struct kevent *kevp, int count, 75 struct knote *marker); 76 static int kqueue_read(struct file *fp, struct uio *uio, 77 struct ucred *cred, int flags); 78 static int kqueue_write(struct file *fp, struct uio *uio, 79 struct ucred *cred, int flags); 80 static int kqueue_ioctl(struct file *fp, u_long com, caddr_t data, 81 struct ucred *cred, struct sysmsg *msg); 82 static int kqueue_kqfilter(struct file *fp, struct knote *kn); 83 static int kqueue_stat(struct file *fp, struct stat *st, 84 struct ucred *cred); 85 static int kqueue_close(struct file *fp); 86 static void kqueue_wakeup(struct kqueue *kq); 87 static int filter_attach(struct knote *kn); 88 static int filter_event(struct knote *kn, long hint); 89 90 /* 91 * MPSAFE 92 */ 93 static struct fileops kqueueops = { 94 .fo_read = kqueue_read, 95 .fo_write = kqueue_write, 96 .fo_ioctl = kqueue_ioctl, 97 .fo_kqfilter = kqueue_kqfilter, 98 .fo_stat = kqueue_stat, 99 .fo_close = kqueue_close, 100 .fo_shutdown = nofo_shutdown 101 }; 102 103 static void knote_attach(struct knote *kn); 104 static void knote_drop(struct knote *kn); 105 static void knote_detach_and_drop(struct knote *kn); 106 static void knote_enqueue(struct knote *kn); 107 static void knote_dequeue(struct knote *kn); 108 static struct knote *knote_alloc(void); 109 static void knote_free(struct knote *kn); 110 111 static void filt_kqdetach(struct knote *kn); 112 static int filt_kqueue(struct knote *kn, long hint); 113 static int filt_procattach(struct knote *kn); 114 static void filt_procdetach(struct knote *kn); 115 static int filt_proc(struct knote *kn, long hint); 116 static int filt_fileattach(struct knote *kn); 117 static void filt_timerexpire(void *knx); 118 static int filt_timerattach(struct knote *kn); 119 static void filt_timerdetach(struct knote *kn); 120 static int filt_timer(struct knote *kn, long hint); 121 static int filt_userattach(struct knote *kn); 122 static void filt_userdetach(struct knote *kn); 123 static int filt_user(struct knote *kn, long hint); 124 static void filt_usertouch(struct knote *kn, struct kevent *kev, 125 u_long type); 126 127 static struct filterops file_filtops = 128 { FILTEROP_ISFD | FILTEROP_MPSAFE, filt_fileattach, NULL, NULL }; 129 static struct filterops kqread_filtops = 130 { FILTEROP_ISFD | FILTEROP_MPSAFE, NULL, filt_kqdetach, filt_kqueue }; 131 static struct filterops proc_filtops = 132 { 0, filt_procattach, filt_procdetach, filt_proc }; 133 static struct filterops timer_filtops = 134 { FILTEROP_MPSAFE, filt_timerattach, filt_timerdetach, filt_timer }; 135 static struct filterops user_filtops = 136 { FILTEROP_MPSAFE, filt_userattach, filt_userdetach, filt_user }; 137 138 static int kq_ncallouts = 0; 139 static int kq_calloutmax = (4 * 1024); 140 SYSCTL_INT(_kern, OID_AUTO, kq_calloutmax, CTLFLAG_RW, 141 &kq_calloutmax, 0, "Maximum number of callouts allocated for kqueue"); 142 static int kq_checkloop = 1000000; 143 SYSCTL_INT(_kern, OID_AUTO, kq_checkloop, CTLFLAG_RW, 144 &kq_checkloop, 0, "Maximum number of loops for kqueue scan"); 145 146 #define KNOTE_ACTIVATE(kn) do { \ 147 kn->kn_status |= KN_ACTIVE; \ 148 if ((kn->kn_status & (KN_QUEUED | KN_DISABLED)) == 0) \ 149 knote_enqueue(kn); \ 150 } while(0) 151 152 #define KN_HASHSIZE 64 /* XXX should be tunable */ 153 #define KN_HASH(val, mask) (((val) ^ (val >> 8)) & (mask)) 154 155 extern struct filterops aio_filtops; 156 extern struct filterops sig_filtops; 157 158 /* 159 * Table for for all system-defined filters. 160 */ 161 static struct filterops *sysfilt_ops[] = { 162 &file_filtops, /* EVFILT_READ */ 163 &file_filtops, /* EVFILT_WRITE */ 164 &aio_filtops, /* EVFILT_AIO */ 165 &file_filtops, /* EVFILT_VNODE */ 166 &proc_filtops, /* EVFILT_PROC */ 167 &sig_filtops, /* EVFILT_SIGNAL */ 168 &timer_filtops, /* EVFILT_TIMER */ 169 &file_filtops, /* EVFILT_EXCEPT */ 170 &user_filtops, /* EVFILT_USER */ 171 }; 172 173 static struct knote_cache_list knote_cache_lists[MAXCPU]; 174 175 /* 176 * Acquire a knote, return non-zero on success, 0 on failure. 177 * 178 * If we cannot acquire the knote we sleep and return 0. The knote 179 * may be stale on return in this case and the caller must restart 180 * whatever loop they are in. 181 * 182 * Related kq token must be held. 183 */ 184 static __inline int 185 knote_acquire(struct knote *kn) 186 { 187 if (kn->kn_status & KN_PROCESSING) { 188 kn->kn_status |= KN_WAITING | KN_REPROCESS; 189 tsleep(kn, 0, "kqepts", hz); 190 /* knote may be stale now */ 191 return(0); 192 } 193 kn->kn_status |= KN_PROCESSING; 194 return(1); 195 } 196 197 /* 198 * Release an acquired knote, clearing KN_PROCESSING and handling any 199 * KN_REPROCESS events. 200 * 201 * Caller must be holding the related kq token 202 * 203 * Non-zero is returned if the knote is destroyed or detached. 204 */ 205 static __inline void 206 knote_release(struct knote *kn) 207 { 208 while (kn->kn_status & KN_REPROCESS) { 209 kn->kn_status &= ~KN_REPROCESS; 210 if (kn->kn_status & KN_WAITING) { 211 kn->kn_status &= ~KN_WAITING; 212 wakeup(kn); 213 } 214 if (kn->kn_status & KN_DELETING) { 215 knote_detach_and_drop(kn); 216 return; 217 /* NOT REACHED */ 218 } 219 if (filter_event(kn, 0)) 220 KNOTE_ACTIVATE(kn); 221 } 222 kn->kn_status &= ~KN_PROCESSING; 223 /* kn should not be accessed anymore */ 224 } 225 226 static int 227 filt_fileattach(struct knote *kn) 228 { 229 return (fo_kqfilter(kn->kn_fp, kn)); 230 } 231 232 /* 233 * MPSAFE 234 */ 235 static int 236 kqueue_kqfilter(struct file *fp, struct knote *kn) 237 { 238 struct kqueue *kq = (struct kqueue *)kn->kn_fp->f_data; 239 240 if (kn->kn_filter != EVFILT_READ) 241 return (EOPNOTSUPP); 242 243 kn->kn_fop = &kqread_filtops; 244 knote_insert(&kq->kq_kqinfo.ki_note, kn); 245 return (0); 246 } 247 248 static void 249 filt_kqdetach(struct knote *kn) 250 { 251 struct kqueue *kq = (struct kqueue *)kn->kn_fp->f_data; 252 253 knote_remove(&kq->kq_kqinfo.ki_note, kn); 254 } 255 256 /*ARGSUSED*/ 257 static int 258 filt_kqueue(struct knote *kn, long hint) 259 { 260 struct kqueue *kq = (struct kqueue *)kn->kn_fp->f_data; 261 262 kn->kn_data = kq->kq_count; 263 return (kn->kn_data > 0); 264 } 265 266 static int 267 filt_procattach(struct knote *kn) 268 { 269 struct proc *p; 270 int immediate; 271 272 immediate = 0; 273 p = pfind(kn->kn_id); 274 if (p == NULL && (kn->kn_sfflags & NOTE_EXIT)) { 275 p = zpfind(kn->kn_id); 276 immediate = 1; 277 } 278 if (p == NULL) { 279 return (ESRCH); 280 } 281 if (!PRISON_CHECK(curthread->td_ucred, p->p_ucred)) { 282 if (p) 283 PRELE(p); 284 return (EACCES); 285 } 286 287 lwkt_gettoken(&p->p_token); 288 kn->kn_ptr.p_proc = p; 289 kn->kn_flags |= EV_CLEAR; /* automatically set */ 290 291 /* 292 * internal flag indicating registration done by kernel 293 */ 294 if (kn->kn_flags & EV_FLAG1) { 295 kn->kn_data = kn->kn_sdata; /* ppid */ 296 kn->kn_fflags = NOTE_CHILD; 297 kn->kn_flags &= ~EV_FLAG1; 298 } 299 300 knote_insert(&p->p_klist, kn); 301 302 /* 303 * Immediately activate any exit notes if the target process is a 304 * zombie. This is necessary to handle the case where the target 305 * process, e.g. a child, dies before the kevent is negistered. 306 */ 307 if (immediate && filt_proc(kn, NOTE_EXIT)) 308 KNOTE_ACTIVATE(kn); 309 lwkt_reltoken(&p->p_token); 310 PRELE(p); 311 312 return (0); 313 } 314 315 /* 316 * The knote may be attached to a different process, which may exit, 317 * leaving nothing for the knote to be attached to. So when the process 318 * exits, the knote is marked as DETACHED and also flagged as ONESHOT so 319 * it will be deleted when read out. However, as part of the knote deletion, 320 * this routine is called, so a check is needed to avoid actually performing 321 * a detach, because the original process does not exist any more. 322 */ 323 static void 324 filt_procdetach(struct knote *kn) 325 { 326 struct proc *p; 327 328 if (kn->kn_status & KN_DETACHED) 329 return; 330 p = kn->kn_ptr.p_proc; 331 knote_remove(&p->p_klist, kn); 332 } 333 334 static int 335 filt_proc(struct knote *kn, long hint) 336 { 337 u_int event; 338 339 /* 340 * mask off extra data 341 */ 342 event = (u_int)hint & NOTE_PCTRLMASK; 343 344 /* 345 * if the user is interested in this event, record it. 346 */ 347 if (kn->kn_sfflags & event) 348 kn->kn_fflags |= event; 349 350 /* 351 * Process is gone, so flag the event as finished. Detach the 352 * knote from the process now because the process will be poof, 353 * gone later on. 354 */ 355 if (event == NOTE_EXIT) { 356 struct proc *p = kn->kn_ptr.p_proc; 357 if ((kn->kn_status & KN_DETACHED) == 0) { 358 PHOLD(p); 359 knote_remove(&p->p_klist, kn); 360 kn->kn_status |= KN_DETACHED; 361 kn->kn_data = p->p_xstat; 362 kn->kn_ptr.p_proc = NULL; 363 PRELE(p); 364 } 365 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT); 366 return (1); 367 } 368 369 /* 370 * process forked, and user wants to track the new process, 371 * so attach a new knote to it, and immediately report an 372 * event with the parent's pid. 373 */ 374 if ((event == NOTE_FORK) && (kn->kn_sfflags & NOTE_TRACK)) { 375 struct kevent kev; 376 int error; 377 378 /* 379 * register knote with new process. 380 */ 381 kev.ident = hint & NOTE_PDATAMASK; /* pid */ 382 kev.filter = kn->kn_filter; 383 kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1; 384 kev.fflags = kn->kn_sfflags; 385 kev.data = kn->kn_id; /* parent */ 386 kev.udata = kn->kn_kevent.udata; /* preserve udata */ 387 error = kqueue_register(kn->kn_kq, &kev); 388 if (error) 389 kn->kn_fflags |= NOTE_TRACKERR; 390 } 391 392 return (kn->kn_fflags != 0); 393 } 394 395 static void 396 filt_timerreset(struct knote *kn) 397 { 398 struct callout *calloutp; 399 struct timeval tv; 400 int tticks; 401 402 tv.tv_sec = kn->kn_sdata / 1000; 403 tv.tv_usec = (kn->kn_sdata % 1000) * 1000; 404 tticks = tvtohz_high(&tv); 405 calloutp = (struct callout *)kn->kn_hook; 406 callout_reset(calloutp, tticks, filt_timerexpire, kn); 407 } 408 409 /* 410 * The callout interlocks with callout_terminate() but can still 411 * race a deletion so if KN_DELETING is set we just don't touch 412 * the knote. 413 */ 414 static void 415 filt_timerexpire(void *knx) 416 { 417 struct knote *kn = knx; 418 struct kqueue *kq = kn->kn_kq; 419 420 lwkt_getpooltoken(kq); 421 422 /* 423 * Open knote_acquire(), since we can't sleep in callout, 424 * however, we do need to record this expiration. 425 */ 426 kn->kn_data++; 427 if (kn->kn_status & KN_PROCESSING) { 428 kn->kn_status |= KN_REPROCESS; 429 if ((kn->kn_status & KN_DELETING) == 0 && 430 (kn->kn_flags & EV_ONESHOT) == 0) 431 filt_timerreset(kn); 432 lwkt_relpooltoken(kq); 433 return; 434 } 435 KASSERT((kn->kn_status & KN_DELETING) == 0, 436 ("acquire a deleting knote %#x", kn->kn_status)); 437 kn->kn_status |= KN_PROCESSING; 438 439 KNOTE_ACTIVATE(kn); 440 if ((kn->kn_flags & EV_ONESHOT) == 0) 441 filt_timerreset(kn); 442 443 knote_release(kn); 444 445 lwkt_relpooltoken(kq); 446 } 447 448 /* 449 * data contains amount of time to sleep, in milliseconds 450 */ 451 static int 452 filt_timerattach(struct knote *kn) 453 { 454 struct callout *calloutp; 455 int prev_ncallouts; 456 457 prev_ncallouts = atomic_fetchadd_int(&kq_ncallouts, 1); 458 if (prev_ncallouts >= kq_calloutmax) { 459 atomic_subtract_int(&kq_ncallouts, 1); 460 kn->kn_hook = NULL; 461 return (ENOMEM); 462 } 463 464 kn->kn_flags |= EV_CLEAR; /* automatically set */ 465 calloutp = kmalloc(sizeof(*calloutp), M_KQUEUE, M_WAITOK); 466 callout_init_mp(calloutp); 467 kn->kn_hook = (caddr_t)calloutp; 468 469 filt_timerreset(kn); 470 return (0); 471 } 472 473 /* 474 * This function is called with the knote flagged locked but it is 475 * still possible to race a callout event due to the callback blocking. 476 * We must call callout_terminate() instead of callout_stop() to deal 477 * with the race. 478 */ 479 static void 480 filt_timerdetach(struct knote *kn) 481 { 482 struct callout *calloutp; 483 484 calloutp = (struct callout *)kn->kn_hook; 485 callout_terminate(calloutp); 486 kfree(calloutp, M_KQUEUE); 487 atomic_subtract_int(&kq_ncallouts, 1); 488 } 489 490 static int 491 filt_timer(struct knote *kn, long hint) 492 { 493 494 return (kn->kn_data != 0); 495 } 496 497 /* 498 * EVFILT_USER 499 */ 500 static int 501 filt_userattach(struct knote *kn) 502 { 503 kn->kn_hook = NULL; 504 if (kn->kn_fflags & NOTE_TRIGGER) 505 kn->kn_ptr.hookid = 1; 506 else 507 kn->kn_ptr.hookid = 0; 508 return 0; 509 } 510 511 static void 512 filt_userdetach(struct knote *kn) 513 { 514 /* nothing to do */ 515 } 516 517 static int 518 filt_user(struct knote *kn, long hint) 519 { 520 return (kn->kn_ptr.hookid); 521 } 522 523 static void 524 filt_usertouch(struct knote *kn, struct kevent *kev, u_long type) 525 { 526 u_int ffctrl; 527 528 switch (type) { 529 case EVENT_REGISTER: 530 if (kev->fflags & NOTE_TRIGGER) 531 kn->kn_ptr.hookid = 1; 532 533 ffctrl = kev->fflags & NOTE_FFCTRLMASK; 534 kev->fflags &= NOTE_FFLAGSMASK; 535 switch (ffctrl) { 536 case NOTE_FFNOP: 537 break; 538 539 case NOTE_FFAND: 540 kn->kn_sfflags &= kev->fflags; 541 break; 542 543 case NOTE_FFOR: 544 kn->kn_sfflags |= kev->fflags; 545 break; 546 547 case NOTE_FFCOPY: 548 kn->kn_sfflags = kev->fflags; 549 break; 550 551 default: 552 /* XXX Return error? */ 553 break; 554 } 555 kn->kn_sdata = kev->data; 556 557 /* 558 * This is not the correct use of EV_CLEAR in an event 559 * modification, it should have been passed as a NOTE instead. 560 * But we need to maintain compatibility with Apple & FreeBSD. 561 * 562 * Note however that EV_CLEAR can still be used when doing 563 * the initial registration of the event and works as expected 564 * (clears the event on reception). 565 */ 566 if (kev->flags & EV_CLEAR) { 567 kn->kn_ptr.hookid = 0; 568 kn->kn_data = 0; 569 kn->kn_fflags = 0; 570 } 571 break; 572 573 case EVENT_PROCESS: 574 *kev = kn->kn_kevent; 575 kev->fflags = kn->kn_sfflags; 576 kev->data = kn->kn_sdata; 577 if (kn->kn_flags & EV_CLEAR) { 578 kn->kn_ptr.hookid = 0; 579 /* kn_data, kn_fflags handled by parent */ 580 } 581 break; 582 583 default: 584 panic("filt_usertouch() - invalid type (%ld)", type); 585 break; 586 } 587 } 588 589 /* 590 * Initialize a kqueue. 591 * 592 * NOTE: The lwp/proc code initializes a kqueue for select/poll ops. 593 * 594 * MPSAFE 595 */ 596 void 597 kqueue_init(struct kqueue *kq, struct filedesc *fdp) 598 { 599 TAILQ_INIT(&kq->kq_knpend); 600 TAILQ_INIT(&kq->kq_knlist); 601 kq->kq_count = 0; 602 kq->kq_fdp = fdp; 603 SLIST_INIT(&kq->kq_kqinfo.ki_note); 604 } 605 606 /* 607 * Terminate a kqueue. Freeing the actual kq itself is left up to the 608 * caller (it might be embedded in a lwp so we don't do it here). 609 * 610 * The kq's knlist must be completely eradicated so block on any 611 * processing races. 612 */ 613 void 614 kqueue_terminate(struct kqueue *kq) 615 { 616 struct lwkt_token *tok; 617 struct knote *kn; 618 619 tok = lwkt_token_pool_lookup(kq); 620 lwkt_gettoken(tok); 621 while ((kn = TAILQ_FIRST(&kq->kq_knlist)) != NULL) { 622 if (knote_acquire(kn)) 623 knote_detach_and_drop(kn); 624 } 625 lwkt_reltoken(tok); 626 627 if (kq->kq_knhash) { 628 hashdestroy(kq->kq_knhash, M_KQUEUE, kq->kq_knhashmask); 629 kq->kq_knhash = NULL; 630 kq->kq_knhashmask = 0; 631 } 632 } 633 634 /* 635 * MPSAFE 636 */ 637 int 638 sys_kqueue(struct kqueue_args *uap) 639 { 640 struct thread *td = curthread; 641 struct kqueue *kq; 642 struct file *fp; 643 int fd, error; 644 645 error = falloc(td->td_lwp, &fp, &fd); 646 if (error) 647 return (error); 648 fp->f_flag = FREAD | FWRITE; 649 fp->f_type = DTYPE_KQUEUE; 650 fp->f_ops = &kqueueops; 651 652 kq = kmalloc(sizeof(struct kqueue), M_KQUEUE, M_WAITOK | M_ZERO); 653 kqueue_init(kq, td->td_proc->p_fd); 654 fp->f_data = kq; 655 656 fsetfd(kq->kq_fdp, fp, fd); 657 uap->sysmsg_result = fd; 658 fdrop(fp); 659 return (error); 660 } 661 662 /* 663 * Copy 'count' items into the destination list pointed to by uap->eventlist. 664 */ 665 static int 666 kevent_copyout(void *arg, struct kevent *kevp, int count, int *res) 667 { 668 struct kevent_copyin_args *kap; 669 int error; 670 671 kap = (struct kevent_copyin_args *)arg; 672 673 error = copyout(kevp, kap->ka->eventlist, count * sizeof(*kevp)); 674 if (error == 0) { 675 kap->ka->eventlist += count; 676 *res += count; 677 } else { 678 *res = -1; 679 } 680 681 return (error); 682 } 683 684 /* 685 * Copy at most 'max' items from the list pointed to by kap->changelist, 686 * return number of items in 'events'. 687 */ 688 static int 689 kevent_copyin(void *arg, struct kevent *kevp, int max, int *events) 690 { 691 struct kevent_copyin_args *kap; 692 int error, count; 693 694 kap = (struct kevent_copyin_args *)arg; 695 696 count = min(kap->ka->nchanges - kap->pchanges, max); 697 error = copyin(kap->ka->changelist, kevp, count * sizeof *kevp); 698 if (error == 0) { 699 kap->ka->changelist += count; 700 kap->pchanges += count; 701 *events = count; 702 } 703 704 return (error); 705 } 706 707 /* 708 * MPSAFE 709 */ 710 int 711 kern_kevent(struct kqueue *kq, int nevents, int *res, void *uap, 712 k_copyin_fn kevent_copyinfn, k_copyout_fn kevent_copyoutfn, 713 struct timespec *tsp_in) 714 { 715 struct kevent *kevp; 716 struct timespec *tsp, ats; 717 int i, n, total, error, nerrors = 0; 718 int lres; 719 int limit = kq_checkloop; 720 struct kevent kev[KQ_NEVENTS]; 721 struct knote marker; 722 struct lwkt_token *tok; 723 724 if (tsp_in == NULL || tsp_in->tv_sec || tsp_in->tv_nsec) 725 atomic_set_int(&curthread->td_mpflags, TDF_MP_BATCH_DEMARC); 726 727 tsp = tsp_in; 728 *res = 0; 729 730 for (;;) { 731 n = 0; 732 error = kevent_copyinfn(uap, kev, KQ_NEVENTS, &n); 733 if (error) 734 return error; 735 if (n == 0) 736 break; 737 for (i = 0; i < n; i++) { 738 kevp = &kev[i]; 739 kevp->flags &= ~EV_SYSFLAGS; 740 error = kqueue_register(kq, kevp); 741 742 /* 743 * If a registration returns an error we 744 * immediately post the error. The kevent() 745 * call itself will fail with the error if 746 * no space is available for posting. 747 * 748 * Such errors normally bypass the timeout/blocking 749 * code. However, if the copyoutfn function refuses 750 * to post the error (see sys_poll()), then we 751 * ignore it too. 752 */ 753 if (error || (kevp->flags & EV_RECEIPT)) { 754 kevp->flags = EV_ERROR; 755 kevp->data = error; 756 lres = *res; 757 kevent_copyoutfn(uap, kevp, 1, res); 758 if (*res < 0) { 759 return error; 760 } else if (lres != *res) { 761 nevents--; 762 nerrors++; 763 } 764 } 765 } 766 } 767 if (nerrors) 768 return 0; 769 770 /* 771 * Acquire/wait for events - setup timeout 772 */ 773 if (tsp != NULL) { 774 if (tsp->tv_sec || tsp->tv_nsec) { 775 getnanouptime(&ats); 776 timespecadd(tsp, &ats); /* tsp = target time */ 777 } 778 } 779 780 /* 781 * Loop as required. 782 * 783 * Collect as many events as we can. Sleeping on successive 784 * loops is disabled if copyoutfn has incremented (*res). 785 * 786 * The loop stops if an error occurs, all events have been 787 * scanned (the marker has been reached), or fewer than the 788 * maximum number of events is found. 789 * 790 * The copyoutfn function does not have to increment (*res) in 791 * order for the loop to continue. 792 * 793 * NOTE: doselect() usually passes 0x7FFFFFFF for nevents. 794 */ 795 total = 0; 796 error = 0; 797 marker.kn_filter = EVFILT_MARKER; 798 marker.kn_status = KN_PROCESSING; 799 tok = lwkt_token_pool_lookup(kq); 800 lwkt_gettoken(tok); 801 TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker, kn_tqe); 802 lwkt_reltoken(tok); 803 while ((n = nevents - total) > 0) { 804 if (n > KQ_NEVENTS) 805 n = KQ_NEVENTS; 806 807 /* 808 * If no events are pending sleep until timeout (if any) 809 * or an event occurs. 810 * 811 * After the sleep completes the marker is moved to the 812 * end of the list, making any received events available 813 * to our scan. 814 */ 815 if (kq->kq_count == 0 && *res == 0) { 816 int timeout; 817 818 if (tsp == NULL) { 819 timeout = 0; 820 } else if (tsp->tv_sec == 0 && tsp->tv_nsec == 0) { 821 error = EWOULDBLOCK; 822 break; 823 } else { 824 struct timespec atx = *tsp; 825 826 getnanouptime(&ats); 827 timespecsub(&atx, &ats); 828 if (atx.tv_sec < 0) { 829 error = EWOULDBLOCK; 830 break; 831 } else { 832 timeout = atx.tv_sec > 24 * 60 * 60 ? 833 24 * 60 * 60 * hz : 834 tstohz_high(&atx); 835 } 836 } 837 838 lwkt_gettoken(tok); 839 if (kq->kq_count == 0) { 840 kq->kq_sleep_cnt++; 841 error = tsleep(kq, PCATCH, "kqread", timeout); 842 843 /* don't restart after signals... */ 844 if (error == ERESTART) 845 error = EINTR; 846 if (error) { 847 lwkt_reltoken(tok); 848 break; 849 } 850 851 TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe); 852 TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker, 853 kn_tqe); 854 } 855 lwkt_reltoken(tok); 856 } 857 858 /* 859 * Process all received events 860 * Account for all non-spurious events in our total 861 */ 862 i = kqueue_scan(kq, kev, n, &marker); 863 if (i) { 864 lres = *res; 865 error = kevent_copyoutfn(uap, kev, i, res); 866 total += *res - lres; 867 if (error) 868 break; 869 } 870 if (limit && --limit == 0) 871 panic("kqueue: checkloop failed i=%d", i); 872 873 /* 874 * Normally when fewer events are returned than requested 875 * we can stop. However, if only spurious events were 876 * collected the copyout will not bump (*res) and we have 877 * to continue. 878 */ 879 if (i < n && *res) 880 break; 881 882 /* 883 * Deal with an edge case where spurious events can cause 884 * a loop to occur without moving the marker. This can 885 * prevent kqueue_scan() from picking up new events which 886 * race us. We must be sure to move the marker for this 887 * case. 888 * 889 * NOTE: We do not want to move the marker if events 890 * were scanned because normal kqueue operations 891 * may reactivate events. Moving the marker in 892 * that case could result in duplicates for the 893 * same event. 894 */ 895 if (i == 0) { 896 lwkt_gettoken(tok); 897 TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe); 898 TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker, kn_tqe); 899 lwkt_reltoken(tok); 900 } 901 } 902 lwkt_gettoken(tok); 903 TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe); 904 lwkt_reltoken(tok); 905 906 /* Timeouts do not return EWOULDBLOCK. */ 907 if (error == EWOULDBLOCK) 908 error = 0; 909 return error; 910 } 911 912 /* 913 * MPALMOSTSAFE 914 */ 915 int 916 sys_kevent(struct kevent_args *uap) 917 { 918 struct thread *td = curthread; 919 struct proc *p = td->td_proc; 920 struct timespec ts, *tsp; 921 struct kqueue *kq; 922 struct file *fp = NULL; 923 struct kevent_copyin_args *kap, ka; 924 int error; 925 926 if (uap->timeout) { 927 error = copyin(uap->timeout, &ts, sizeof(ts)); 928 if (error) 929 return (error); 930 tsp = &ts; 931 } else { 932 tsp = NULL; 933 } 934 fp = holdfp(p->p_fd, uap->fd, -1); 935 if (fp == NULL) 936 return (EBADF); 937 if (fp->f_type != DTYPE_KQUEUE) { 938 fdrop(fp); 939 return (EBADF); 940 } 941 942 kq = (struct kqueue *)fp->f_data; 943 944 kap = &ka; 945 kap->ka = uap; 946 kap->pchanges = 0; 947 948 error = kern_kevent(kq, uap->nevents, &uap->sysmsg_result, kap, 949 kevent_copyin, kevent_copyout, tsp); 950 951 fdrop(fp); 952 953 return (error); 954 } 955 956 int 957 kqueue_register(struct kqueue *kq, struct kevent *kev) 958 { 959 struct filedesc *fdp = kq->kq_fdp; 960 struct klist *list = NULL; 961 struct filterops *fops; 962 struct file *fp = NULL; 963 struct knote *kn = NULL; 964 struct thread *td; 965 int error = 0; 966 struct knote_cache_list *cache_list; 967 968 if (kev->filter < 0) { 969 if (kev->filter + EVFILT_SYSCOUNT < 0) 970 return (EINVAL); 971 fops = sysfilt_ops[~kev->filter]; /* to 0-base index */ 972 } else { 973 /* 974 * XXX 975 * filter attach routine is responsible for insuring that 976 * the identifier can be attached to it. 977 */ 978 return (EINVAL); 979 } 980 981 if (fops->f_flags & FILTEROP_ISFD) { 982 /* validate descriptor */ 983 fp = holdfp(fdp, kev->ident, -1); 984 if (fp == NULL) 985 return (EBADF); 986 } 987 988 cache_list = &knote_cache_lists[mycpuid]; 989 if (SLIST_EMPTY(&cache_list->knote_cache)) { 990 struct knote *new_kn; 991 992 new_kn = knote_alloc(); 993 SLIST_INSERT_HEAD(&cache_list->knote_cache, new_kn, kn_link); 994 cache_list->knote_cache_cnt++; 995 } 996 997 td = curthread; 998 lwkt_getpooltoken(kq); 999 1000 /* 1001 * Make sure that only one thread can register event on this kqueue, 1002 * so that we would not suffer any race, even if the registration 1003 * blocked, i.e. kq token was released, and the kqueue was shared 1004 * between threads (this should be rare though). 1005 */ 1006 while (__predict_false(kq->kq_regtd != NULL && kq->kq_regtd != td)) { 1007 kq->kq_state |= KQ_REGWAIT; 1008 tsleep(&kq->kq_regtd, 0, "kqreg", 0); 1009 } 1010 if (__predict_false(kq->kq_regtd != NULL)) { 1011 /* Recursive calling of kqueue_register() */ 1012 td = NULL; 1013 } else { 1014 /* Owner of the kq_regtd, i.e. td != NULL */ 1015 kq->kq_regtd = td; 1016 } 1017 1018 if (fp != NULL) { 1019 list = &fp->f_klist; 1020 } else if (kq->kq_knhashmask) { 1021 list = &kq->kq_knhash[ 1022 KN_HASH((u_long)kev->ident, kq->kq_knhashmask)]; 1023 } 1024 if (list != NULL) { 1025 lwkt_getpooltoken(list); 1026 again: 1027 SLIST_FOREACH(kn, list, kn_link) { 1028 if (kn->kn_kq == kq && 1029 kn->kn_filter == kev->filter && 1030 kn->kn_id == kev->ident) { 1031 if (knote_acquire(kn) == 0) 1032 goto again; 1033 break; 1034 } 1035 } 1036 lwkt_relpooltoken(list); 1037 } 1038 1039 /* 1040 * NOTE: At this point if kn is non-NULL we will have acquired 1041 * it and set KN_PROCESSING. 1042 */ 1043 if (kn == NULL && ((kev->flags & EV_ADD) == 0)) { 1044 error = ENOENT; 1045 goto done; 1046 } 1047 1048 /* 1049 * kn now contains the matching knote, or NULL if no match 1050 */ 1051 if (kev->flags & EV_ADD) { 1052 if (kn == NULL) { 1053 kn = SLIST_FIRST(&cache_list->knote_cache); 1054 if (kn == NULL) { 1055 kn = knote_alloc(); 1056 } else { 1057 SLIST_REMOVE_HEAD(&cache_list->knote_cache, 1058 kn_link); 1059 cache_list->knote_cache_cnt--; 1060 } 1061 kn->kn_fp = fp; 1062 kn->kn_kq = kq; 1063 kn->kn_fop = fops; 1064 1065 /* 1066 * apply reference count to knote structure, and 1067 * do not release it at the end of this routine. 1068 */ 1069 fp = NULL; 1070 1071 kn->kn_sfflags = kev->fflags; 1072 kn->kn_sdata = kev->data; 1073 kev->fflags = 0; 1074 kev->data = 0; 1075 kn->kn_kevent = *kev; 1076 1077 /* 1078 * KN_PROCESSING prevents the knote from getting 1079 * ripped out from under us while we are trying 1080 * to attach it, in case the attach blocks. 1081 */ 1082 kn->kn_status = KN_PROCESSING; 1083 knote_attach(kn); 1084 if ((error = filter_attach(kn)) != 0) { 1085 kn->kn_status |= KN_DELETING | KN_REPROCESS; 1086 knote_drop(kn); 1087 goto done; 1088 } 1089 1090 /* 1091 * Interlock against close races which either tried 1092 * to remove our knote while we were blocked or missed 1093 * it entirely prior to our attachment. We do not 1094 * want to end up with a knote on a closed descriptor. 1095 */ 1096 if ((fops->f_flags & FILTEROP_ISFD) && 1097 checkfdclosed(fdp, kev->ident, kn->kn_fp)) { 1098 kn->kn_status |= KN_DELETING | KN_REPROCESS; 1099 } 1100 } else { 1101 /* 1102 * The user may change some filter values after the 1103 * initial EV_ADD, but doing so will not reset any 1104 * filter which have already been triggered. 1105 */ 1106 KKASSERT(kn->kn_status & KN_PROCESSING); 1107 if (fops == &user_filtops) { 1108 filt_usertouch(kn, kev, EVENT_REGISTER); 1109 } else { 1110 kn->kn_sfflags = kev->fflags; 1111 kn->kn_sdata = kev->data; 1112 kn->kn_kevent.udata = kev->udata; 1113 } 1114 } 1115 1116 /* 1117 * Execute the filter event to immediately activate the 1118 * knote if necessary. If reprocessing events are pending 1119 * due to blocking above we do not run the filter here 1120 * but instead let knote_release() do it. Otherwise we 1121 * might run the filter on a deleted event. 1122 */ 1123 if ((kn->kn_status & KN_REPROCESS) == 0) { 1124 if (filter_event(kn, 0)) 1125 KNOTE_ACTIVATE(kn); 1126 } 1127 } else if (kev->flags & EV_DELETE) { 1128 /* 1129 * Delete the existing knote 1130 */ 1131 knote_detach_and_drop(kn); 1132 goto done; 1133 } else { 1134 /* 1135 * Modify an existing event. 1136 * 1137 * The user may change some filter values after the 1138 * initial EV_ADD, but doing so will not reset any 1139 * filter which have already been triggered. 1140 */ 1141 KKASSERT(kn->kn_status & KN_PROCESSING); 1142 if (fops == &user_filtops) { 1143 filt_usertouch(kn, kev, EVENT_REGISTER); 1144 } else { 1145 kn->kn_sfflags = kev->fflags; 1146 kn->kn_sdata = kev->data; 1147 kn->kn_kevent.udata = kev->udata; 1148 } 1149 1150 /* 1151 * Execute the filter event to immediately activate the 1152 * knote if necessary. If reprocessing events are pending 1153 * due to blocking above we do not run the filter here 1154 * but instead let knote_release() do it. Otherwise we 1155 * might run the filter on a deleted event. 1156 */ 1157 if ((kn->kn_status & KN_REPROCESS) == 0) { 1158 if (filter_event(kn, 0)) 1159 KNOTE_ACTIVATE(kn); 1160 } 1161 } 1162 1163 /* 1164 * Disablement does not deactivate a knote here. 1165 */ 1166 if ((kev->flags & EV_DISABLE) && 1167 ((kn->kn_status & KN_DISABLED) == 0)) { 1168 kn->kn_status |= KN_DISABLED; 1169 } 1170 1171 /* 1172 * Re-enablement may have to immediately enqueue an active knote. 1173 */ 1174 if ((kev->flags & EV_ENABLE) && (kn->kn_status & KN_DISABLED)) { 1175 kn->kn_status &= ~KN_DISABLED; 1176 if ((kn->kn_status & KN_ACTIVE) && 1177 ((kn->kn_status & KN_QUEUED) == 0)) { 1178 knote_enqueue(kn); 1179 } 1180 } 1181 1182 /* 1183 * Handle any required reprocessing 1184 */ 1185 knote_release(kn); 1186 /* kn may be invalid now */ 1187 1188 done: 1189 if (td != NULL) { /* Owner of the kq_regtd */ 1190 kq->kq_regtd = NULL; 1191 if (__predict_false(kq->kq_state & KQ_REGWAIT)) { 1192 kq->kq_state &= ~KQ_REGWAIT; 1193 wakeup(&kq->kq_regtd); 1194 } 1195 } 1196 lwkt_relpooltoken(kq); 1197 if (fp != NULL) 1198 fdrop(fp); 1199 return (error); 1200 } 1201 1202 /* 1203 * Scan the kqueue, return the number of active events placed in kevp up 1204 * to count. 1205 * 1206 * Continuous mode events may get recycled, do not continue scanning past 1207 * marker unless no events have been collected. 1208 */ 1209 static int 1210 kqueue_scan(struct kqueue *kq, struct kevent *kevp, int count, 1211 struct knote *marker) 1212 { 1213 struct knote *kn, local_marker; 1214 int total; 1215 1216 total = 0; 1217 local_marker.kn_filter = EVFILT_MARKER; 1218 local_marker.kn_status = KN_PROCESSING; 1219 1220 lwkt_getpooltoken(kq); 1221 1222 /* 1223 * Collect events. 1224 */ 1225 TAILQ_INSERT_HEAD(&kq->kq_knpend, &local_marker, kn_tqe); 1226 while (count) { 1227 kn = TAILQ_NEXT(&local_marker, kn_tqe); 1228 if (kn->kn_filter == EVFILT_MARKER) { 1229 /* Marker reached, we are done */ 1230 if (kn == marker) 1231 break; 1232 1233 /* Move local marker past some other threads marker */ 1234 kn = TAILQ_NEXT(kn, kn_tqe); 1235 TAILQ_REMOVE(&kq->kq_knpend, &local_marker, kn_tqe); 1236 TAILQ_INSERT_BEFORE(kn, &local_marker, kn_tqe); 1237 continue; 1238 } 1239 1240 /* 1241 * We can't skip a knote undergoing processing, otherwise 1242 * we risk not returning it when the user process expects 1243 * it should be returned. Sleep and retry. 1244 */ 1245 if (knote_acquire(kn) == 0) 1246 continue; 1247 1248 /* 1249 * Remove the event for processing. 1250 * 1251 * WARNING! We must leave KN_QUEUED set to prevent the 1252 * event from being KNOTE_ACTIVATE()d while 1253 * the queue state is in limbo, in case we 1254 * block. 1255 */ 1256 TAILQ_REMOVE(&kq->kq_knpend, kn, kn_tqe); 1257 kq->kq_count--; 1258 1259 /* 1260 * We have to deal with an extremely important race against 1261 * file descriptor close()s here. The file descriptor can 1262 * disappear MPSAFE, and there is a small window of 1263 * opportunity between that and the call to knote_fdclose(). 1264 * 1265 * If we hit that window here while doselect or dopoll is 1266 * trying to delete a spurious event they will not be able 1267 * to match up the event against a knote and will go haywire. 1268 */ 1269 if ((kn->kn_fop->f_flags & FILTEROP_ISFD) && 1270 checkfdclosed(kq->kq_fdp, kn->kn_kevent.ident, kn->kn_fp)) { 1271 kn->kn_status |= KN_DELETING | KN_REPROCESS; 1272 } 1273 1274 if (kn->kn_status & KN_DISABLED) { 1275 /* 1276 * If disabled we ensure the event is not queued 1277 * but leave its active bit set. On re-enablement 1278 * the event may be immediately triggered. 1279 */ 1280 kn->kn_status &= ~KN_QUEUED; 1281 } else if ((kn->kn_flags & EV_ONESHOT) == 0 && 1282 (kn->kn_status & KN_DELETING) == 0 && 1283 filter_event(kn, 0) == 0) { 1284 /* 1285 * If not running in one-shot mode and the event 1286 * is no longer present we ensure it is removed 1287 * from the queue and ignore it. 1288 */ 1289 kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE); 1290 } else { 1291 /* 1292 * Post the event 1293 */ 1294 if (kn->kn_fop == &user_filtops) 1295 filt_usertouch(kn, kevp, EVENT_PROCESS); 1296 else 1297 *kevp = kn->kn_kevent; 1298 ++kevp; 1299 ++total; 1300 --count; 1301 1302 if (kn->kn_flags & EV_ONESHOT) { 1303 kn->kn_status &= ~KN_QUEUED; 1304 kn->kn_status |= KN_DELETING | KN_REPROCESS; 1305 } else { 1306 if (kn->kn_flags & (EV_CLEAR | EV_DISPATCH)) { 1307 if (kn->kn_flags & EV_CLEAR) { 1308 kn->kn_data = 0; 1309 kn->kn_fflags = 0; 1310 } 1311 if (kn->kn_flags & EV_DISPATCH) { 1312 kn->kn_status |= KN_DISABLED; 1313 } 1314 kn->kn_status &= ~(KN_QUEUED | 1315 KN_ACTIVE); 1316 } else { 1317 TAILQ_INSERT_TAIL(&kq->kq_knpend, kn, kn_tqe); 1318 kq->kq_count++; 1319 } 1320 } 1321 } 1322 1323 /* 1324 * Handle any post-processing states 1325 */ 1326 knote_release(kn); 1327 } 1328 TAILQ_REMOVE(&kq->kq_knpend, &local_marker, kn_tqe); 1329 1330 lwkt_relpooltoken(kq); 1331 return (total); 1332 } 1333 1334 /* 1335 * XXX 1336 * This could be expanded to call kqueue_scan, if desired. 1337 * 1338 * MPSAFE 1339 */ 1340 static int 1341 kqueue_read(struct file *fp, struct uio *uio, struct ucred *cred, int flags) 1342 { 1343 return (ENXIO); 1344 } 1345 1346 /* 1347 * MPSAFE 1348 */ 1349 static int 1350 kqueue_write(struct file *fp, struct uio *uio, struct ucred *cred, int flags) 1351 { 1352 return (ENXIO); 1353 } 1354 1355 /* 1356 * MPALMOSTSAFE 1357 */ 1358 static int 1359 kqueue_ioctl(struct file *fp, u_long com, caddr_t data, 1360 struct ucred *cred, struct sysmsg *msg) 1361 { 1362 struct lwkt_token *tok; 1363 struct kqueue *kq; 1364 int error; 1365 1366 kq = (struct kqueue *)fp->f_data; 1367 tok = lwkt_token_pool_lookup(kq); 1368 lwkt_gettoken(tok); 1369 1370 switch(com) { 1371 case FIOASYNC: 1372 if (*(int *)data) 1373 kq->kq_state |= KQ_ASYNC; 1374 else 1375 kq->kq_state &= ~KQ_ASYNC; 1376 error = 0; 1377 break; 1378 case FIOSETOWN: 1379 error = fsetown(*(int *)data, &kq->kq_sigio); 1380 break; 1381 default: 1382 error = ENOTTY; 1383 break; 1384 } 1385 lwkt_reltoken(tok); 1386 return (error); 1387 } 1388 1389 /* 1390 * MPSAFE 1391 */ 1392 static int 1393 kqueue_stat(struct file *fp, struct stat *st, struct ucred *cred) 1394 { 1395 struct kqueue *kq = (struct kqueue *)fp->f_data; 1396 1397 bzero((void *)st, sizeof(*st)); 1398 st->st_size = kq->kq_count; 1399 st->st_blksize = sizeof(struct kevent); 1400 st->st_mode = S_IFIFO; 1401 return (0); 1402 } 1403 1404 /* 1405 * MPSAFE 1406 */ 1407 static int 1408 kqueue_close(struct file *fp) 1409 { 1410 struct kqueue *kq = (struct kqueue *)fp->f_data; 1411 1412 kqueue_terminate(kq); 1413 1414 fp->f_data = NULL; 1415 funsetown(&kq->kq_sigio); 1416 1417 kfree(kq, M_KQUEUE); 1418 return (0); 1419 } 1420 1421 static void 1422 kqueue_wakeup(struct kqueue *kq) 1423 { 1424 if (kq->kq_sleep_cnt) { 1425 if (kq->kq_sleep_cnt == 1) 1426 wakeup_one(kq); 1427 else 1428 wakeup(kq); 1429 kq->kq_sleep_cnt = 0; 1430 } 1431 KNOTE(&kq->kq_kqinfo.ki_note, 0); 1432 } 1433 1434 /* 1435 * Calls filterops f_attach function, acquiring mplock if filter is not 1436 * marked as FILTEROP_MPSAFE. 1437 * 1438 * Caller must be holding the related kq token 1439 */ 1440 static int 1441 filter_attach(struct knote *kn) 1442 { 1443 int ret; 1444 1445 if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) { 1446 ret = kn->kn_fop->f_attach(kn); 1447 } else { 1448 get_mplock(); 1449 ret = kn->kn_fop->f_attach(kn); 1450 rel_mplock(); 1451 } 1452 return (ret); 1453 } 1454 1455 /* 1456 * Detach the knote and drop it, destroying the knote. 1457 * 1458 * Calls filterops f_detach function, acquiring mplock if filter is not 1459 * marked as FILTEROP_MPSAFE. 1460 * 1461 * Caller must be holding the related kq token 1462 */ 1463 static void 1464 knote_detach_and_drop(struct knote *kn) 1465 { 1466 kn->kn_status |= KN_DELETING | KN_REPROCESS; 1467 if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) { 1468 kn->kn_fop->f_detach(kn); 1469 } else { 1470 get_mplock(); 1471 kn->kn_fop->f_detach(kn); 1472 rel_mplock(); 1473 } 1474 knote_drop(kn); 1475 } 1476 1477 /* 1478 * Calls filterops f_event function, acquiring mplock if filter is not 1479 * marked as FILTEROP_MPSAFE. 1480 * 1481 * If the knote is in the middle of being created or deleted we cannot 1482 * safely call the filter op. 1483 * 1484 * Caller must be holding the related kq token 1485 */ 1486 static int 1487 filter_event(struct knote *kn, long hint) 1488 { 1489 int ret; 1490 1491 if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) { 1492 ret = kn->kn_fop->f_event(kn, hint); 1493 } else { 1494 get_mplock(); 1495 ret = kn->kn_fop->f_event(kn, hint); 1496 rel_mplock(); 1497 } 1498 return (ret); 1499 } 1500 1501 /* 1502 * Walk down a list of knotes, activating them if their event has triggered. 1503 * 1504 * If we encounter any knotes which are undergoing processing we just mark 1505 * them for reprocessing and do not try to [re]activate the knote. However, 1506 * if a hint is being passed we have to wait and that makes things a bit 1507 * sticky. 1508 */ 1509 void 1510 knote(struct klist *list, long hint) 1511 { 1512 struct knote *kn, marker; 1513 1514 marker.kn_filter = EVFILT_MARKER; 1515 marker.kn_status = KN_PROCESSING; 1516 1517 lwkt_getpooltoken(list); 1518 if (SLIST_EMPTY(list)) { 1519 lwkt_relpooltoken(list); 1520 return; 1521 } 1522 1523 SLIST_INSERT_HEAD(list, &marker, kn_next); 1524 while ((kn = SLIST_NEXT(&marker, kn_next)) != NULL) { 1525 struct kqueue *kq; 1526 int last_knote = 0; 1527 1528 if (kn->kn_filter == EVFILT_MARKER) { 1529 /* Skip marker */ 1530 SLIST_REMOVE(list, &marker, knote, kn_next); 1531 if (SLIST_NEXT(kn, kn_next) == NULL) 1532 goto done; 1533 SLIST_INSERT_AFTER(kn, &marker, kn_next); 1534 continue; 1535 } 1536 1537 kq = kn->kn_kq; 1538 lwkt_getpooltoken(kq); 1539 1540 if (kn != SLIST_NEXT(&marker, kn_next) || kn->kn_kq != kq) { 1541 /* 1542 * Don't move the marker; check the knote after 1543 * the marker again. 1544 */ 1545 lwkt_relpooltoken(kq); 1546 continue; 1547 } 1548 1549 if (kn->kn_status & KN_PROCESSING) { 1550 /* 1551 * Someone else is processing the knote, ask the 1552 * other thread to reprocess it and don't mess 1553 * with it otherwise. 1554 */ 1555 if (hint == 0) { 1556 /* 1557 * Move the marker w/ the kq token, so that 1558 * this knote will not be ripped behind our 1559 * back. 1560 */ 1561 SLIST_REMOVE(list, &marker, knote, kn_next); 1562 if (SLIST_NEXT(kn, kn_next) != NULL) 1563 SLIST_INSERT_AFTER(kn, &marker, kn_next); 1564 else 1565 last_knote = 1; 1566 kn->kn_status |= KN_REPROCESS; 1567 lwkt_relpooltoken(kq); 1568 1569 if (last_knote) 1570 goto done; 1571 continue; 1572 } 1573 1574 /* 1575 * If the hint is non-zero we have to wait or risk 1576 * losing the state the caller is trying to update. 1577 */ 1578 kn->kn_status |= KN_WAITING | KN_REPROCESS; 1579 tsleep(kn, 0, "knotec", hz); 1580 1581 /* 1582 * Don't move the marker; check this knote again, 1583 * hopefully it is still after the marker. Or it 1584 * was deleted and we would check the next knote. 1585 */ 1586 lwkt_relpooltoken(kq); 1587 continue; 1588 } 1589 1590 /* 1591 * Become the reprocessing master ourselves. 1592 */ 1593 KASSERT((kn->kn_status & KN_DELETING) == 0, 1594 ("acquire a deleting knote %#x", kn->kn_status)); 1595 kn->kn_status |= KN_PROCESSING; 1596 1597 /* Move the marker */ 1598 SLIST_REMOVE(list, &marker, knote, kn_next); 1599 if (SLIST_NEXT(kn, kn_next) != NULL) 1600 SLIST_INSERT_AFTER(kn, &marker, kn_next); 1601 else 1602 last_knote = 1; 1603 1604 /* 1605 * If hint is non-zero running the event is mandatory 1606 * so do it whether reprocessing is set or not. 1607 */ 1608 if (filter_event(kn, hint)) 1609 KNOTE_ACTIVATE(kn); 1610 1611 knote_release(kn); 1612 lwkt_relpooltoken(kq); 1613 1614 if (last_knote) 1615 goto done; 1616 } 1617 SLIST_REMOVE(list, &marker, knote, kn_next); 1618 done: 1619 lwkt_relpooltoken(list); 1620 } 1621 1622 /* 1623 * Insert knote at head of klist. 1624 * 1625 * This function may only be called via a filter function and thus 1626 * kq_token should already be held and marked for processing. 1627 */ 1628 void 1629 knote_insert(struct klist *klist, struct knote *kn) 1630 { 1631 lwkt_getpooltoken(klist); 1632 KKASSERT(kn->kn_status & KN_PROCESSING); 1633 SLIST_INSERT_HEAD(klist, kn, kn_next); 1634 lwkt_relpooltoken(klist); 1635 } 1636 1637 /* 1638 * Remove knote from a klist 1639 * 1640 * This function may only be called via a filter function and thus 1641 * kq_token should already be held and marked for processing. 1642 */ 1643 void 1644 knote_remove(struct klist *klist, struct knote *kn) 1645 { 1646 lwkt_getpooltoken(klist); 1647 KKASSERT(kn->kn_status & KN_PROCESSING); 1648 SLIST_REMOVE(klist, kn, knote, kn_next); 1649 lwkt_relpooltoken(klist); 1650 } 1651 1652 void 1653 knote_assume_knotes(struct kqinfo *src, struct kqinfo *dst, 1654 struct filterops *ops, void *hook) 1655 { 1656 struct knote *kn, marker; 1657 int has_note; 1658 1659 marker.kn_filter = EVFILT_MARKER; 1660 marker.kn_status = KN_PROCESSING; 1661 1662 lwkt_getpooltoken(&src->ki_note); 1663 if (SLIST_EMPTY(&src->ki_note)) { 1664 lwkt_relpooltoken(&src->ki_note); 1665 return; 1666 } 1667 lwkt_getpooltoken(&dst->ki_note); 1668 1669 restart: 1670 has_note = 0; 1671 SLIST_INSERT_HEAD(&src->ki_note, &marker, kn_next); 1672 while ((kn = SLIST_NEXT(&marker, kn_next)) != NULL) { 1673 struct kqueue *kq; 1674 1675 if (kn->kn_filter == EVFILT_MARKER) { 1676 /* Skip marker */ 1677 SLIST_REMOVE(&src->ki_note, &marker, knote, kn_next); 1678 SLIST_INSERT_AFTER(kn, &marker, kn_next); 1679 continue; 1680 } 1681 1682 kq = kn->kn_kq; 1683 lwkt_getpooltoken(kq); 1684 1685 if (kn != SLIST_NEXT(&marker, kn_next) || kn->kn_kq != kq) { 1686 /* 1687 * Don't move the marker; check the knote after 1688 * the marker again. 1689 */ 1690 lwkt_relpooltoken(kq); 1691 continue; 1692 } 1693 1694 /* Move marker */ 1695 SLIST_REMOVE(&src->ki_note, &marker, knote, kn_next); 1696 SLIST_INSERT_AFTER(kn, &marker, kn_next); 1697 1698 has_note = 1; 1699 if (knote_acquire(kn)) { 1700 knote_remove(&src->ki_note, kn); 1701 kn->kn_fop = ops; 1702 kn->kn_hook = hook; 1703 knote_insert(&dst->ki_note, kn); 1704 knote_release(kn); 1705 /* kn may be invalid now */ 1706 } 1707 lwkt_relpooltoken(kq); 1708 } 1709 SLIST_REMOVE(&src->ki_note, &marker, knote, kn_next); 1710 if (has_note) { 1711 /* Keep draining, until nothing left */ 1712 goto restart; 1713 } 1714 1715 lwkt_relpooltoken(&dst->ki_note); 1716 lwkt_relpooltoken(&src->ki_note); 1717 } 1718 1719 /* 1720 * Remove all knotes referencing a specified fd 1721 */ 1722 void 1723 knote_fdclose(struct file *fp, struct filedesc *fdp, int fd) 1724 { 1725 struct kqueue *kq; 1726 struct knote *kn; 1727 struct knote *kntmp; 1728 1729 lwkt_getpooltoken(&fp->f_klist); 1730 restart: 1731 SLIST_FOREACH(kn, &fp->f_klist, kn_link) { 1732 if (kn->kn_kq->kq_fdp == fdp && kn->kn_id == fd) { 1733 kq = kn->kn_kq; 1734 lwkt_getpooltoken(kq); 1735 1736 /* temporary verification hack */ 1737 SLIST_FOREACH(kntmp, &fp->f_klist, kn_link) { 1738 if (kn == kntmp) 1739 break; 1740 } 1741 if (kn != kntmp || kn->kn_kq->kq_fdp != fdp || 1742 kn->kn_id != fd || kn->kn_kq != kq) { 1743 lwkt_relpooltoken(kq); 1744 goto restart; 1745 } 1746 if (knote_acquire(kn)) 1747 knote_detach_and_drop(kn); 1748 lwkt_relpooltoken(kq); 1749 goto restart; 1750 } 1751 } 1752 lwkt_relpooltoken(&fp->f_klist); 1753 } 1754 1755 /* 1756 * Low level attach function. 1757 * 1758 * The knote should already be marked for processing. 1759 * Caller must hold the related kq token. 1760 */ 1761 static void 1762 knote_attach(struct knote *kn) 1763 { 1764 struct klist *list; 1765 struct kqueue *kq = kn->kn_kq; 1766 1767 if (kn->kn_fop->f_flags & FILTEROP_ISFD) { 1768 KKASSERT(kn->kn_fp); 1769 list = &kn->kn_fp->f_klist; 1770 } else { 1771 if (kq->kq_knhashmask == 0) 1772 kq->kq_knhash = hashinit(KN_HASHSIZE, M_KQUEUE, 1773 &kq->kq_knhashmask); 1774 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)]; 1775 } 1776 lwkt_getpooltoken(list); 1777 SLIST_INSERT_HEAD(list, kn, kn_link); 1778 lwkt_relpooltoken(list); 1779 TAILQ_INSERT_HEAD(&kq->kq_knlist, kn, kn_kqlink); 1780 } 1781 1782 /* 1783 * Low level drop function. 1784 * 1785 * The knote should already be marked for processing. 1786 * Caller must hold the related kq token. 1787 */ 1788 static void 1789 knote_drop(struct knote *kn) 1790 { 1791 struct kqueue *kq; 1792 struct klist *list; 1793 1794 kq = kn->kn_kq; 1795 1796 if (kn->kn_fop->f_flags & FILTEROP_ISFD) 1797 list = &kn->kn_fp->f_klist; 1798 else 1799 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)]; 1800 1801 lwkt_getpooltoken(list); 1802 SLIST_REMOVE(list, kn, knote, kn_link); 1803 lwkt_relpooltoken(list); 1804 TAILQ_REMOVE(&kq->kq_knlist, kn, kn_kqlink); 1805 if (kn->kn_status & KN_QUEUED) 1806 knote_dequeue(kn); 1807 if (kn->kn_fop->f_flags & FILTEROP_ISFD) { 1808 fdrop(kn->kn_fp); 1809 kn->kn_fp = NULL; 1810 } 1811 knote_free(kn); 1812 } 1813 1814 /* 1815 * Low level enqueue function. 1816 * 1817 * The knote should already be marked for processing. 1818 * Caller must be holding the kq token 1819 */ 1820 static void 1821 knote_enqueue(struct knote *kn) 1822 { 1823 struct kqueue *kq = kn->kn_kq; 1824 1825 KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued")); 1826 TAILQ_INSERT_TAIL(&kq->kq_knpend, kn, kn_tqe); 1827 kn->kn_status |= KN_QUEUED; 1828 ++kq->kq_count; 1829 1830 /* 1831 * Send SIGIO on request (typically set up as a mailbox signal) 1832 */ 1833 if (kq->kq_sigio && (kq->kq_state & KQ_ASYNC) && kq->kq_count == 1) 1834 pgsigio(kq->kq_sigio, SIGIO, 0); 1835 1836 kqueue_wakeup(kq); 1837 } 1838 1839 /* 1840 * Low level dequeue function. 1841 * 1842 * The knote should already be marked for processing. 1843 * Caller must be holding the kq token 1844 */ 1845 static void 1846 knote_dequeue(struct knote *kn) 1847 { 1848 struct kqueue *kq = kn->kn_kq; 1849 1850 KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued")); 1851 TAILQ_REMOVE(&kq->kq_knpend, kn, kn_tqe); 1852 kn->kn_status &= ~KN_QUEUED; 1853 kq->kq_count--; 1854 } 1855 1856 static struct knote * 1857 knote_alloc(void) 1858 { 1859 return kmalloc(sizeof(struct knote), M_KQUEUE, M_WAITOK); 1860 } 1861 1862 static void 1863 knote_free(struct knote *kn) 1864 { 1865 struct knote_cache_list *cache_list; 1866 1867 cache_list = &knote_cache_lists[mycpuid]; 1868 if (cache_list->knote_cache_cnt < KNOTE_CACHE_MAX) { 1869 SLIST_INSERT_HEAD(&cache_list->knote_cache, kn, kn_link); 1870 cache_list->knote_cache_cnt++; 1871 return; 1872 } 1873 kfree(kn, M_KQUEUE); 1874 } 1875