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 knote *kn; 617 618 lwkt_getpooltoken(kq); 619 while ((kn = TAILQ_FIRST(&kq->kq_knlist)) != NULL) { 620 if (knote_acquire(kn)) 621 knote_detach_and_drop(kn); 622 } 623 lwkt_relpooltoken(kq); 624 625 if (kq->kq_knhash) { 626 hashdestroy(kq->kq_knhash, M_KQUEUE, kq->kq_knhashmask); 627 kq->kq_knhash = NULL; 628 kq->kq_knhashmask = 0; 629 } 630 } 631 632 /* 633 * MPSAFE 634 */ 635 int 636 sys_kqueue(struct kqueue_args *uap) 637 { 638 struct thread *td = curthread; 639 struct kqueue *kq; 640 struct file *fp; 641 int fd, error; 642 643 error = falloc(td->td_lwp, &fp, &fd); 644 if (error) 645 return (error); 646 fp->f_flag = FREAD | FWRITE; 647 fp->f_type = DTYPE_KQUEUE; 648 fp->f_ops = &kqueueops; 649 650 kq = kmalloc(sizeof(struct kqueue), M_KQUEUE, M_WAITOK | M_ZERO); 651 kqueue_init(kq, td->td_proc->p_fd); 652 fp->f_data = kq; 653 654 fsetfd(kq->kq_fdp, fp, fd); 655 uap->sysmsg_result = fd; 656 fdrop(fp); 657 return (error); 658 } 659 660 /* 661 * Copy 'count' items into the destination list pointed to by uap->eventlist. 662 */ 663 static int 664 kevent_copyout(void *arg, struct kevent *kevp, int count, int *res) 665 { 666 struct kevent_copyin_args *kap; 667 int error; 668 669 kap = (struct kevent_copyin_args *)arg; 670 671 error = copyout(kevp, kap->ka->eventlist, count * sizeof(*kevp)); 672 if (error == 0) { 673 kap->ka->eventlist += count; 674 *res += count; 675 } else { 676 *res = -1; 677 } 678 679 return (error); 680 } 681 682 /* 683 * Copy at most 'max' items from the list pointed to by kap->changelist, 684 * return number of items in 'events'. 685 */ 686 static int 687 kevent_copyin(void *arg, struct kevent *kevp, int max, int *events) 688 { 689 struct kevent_copyin_args *kap; 690 int error, count; 691 692 kap = (struct kevent_copyin_args *)arg; 693 694 count = min(kap->ka->nchanges - kap->pchanges, max); 695 error = copyin(kap->ka->changelist, kevp, count * sizeof *kevp); 696 if (error == 0) { 697 kap->ka->changelist += count; 698 kap->pchanges += count; 699 *events = count; 700 } 701 702 return (error); 703 } 704 705 /* 706 * MPSAFE 707 */ 708 int 709 kern_kevent(struct kqueue *kq, int nevents, int *res, void *uap, 710 k_copyin_fn kevent_copyinfn, k_copyout_fn kevent_copyoutfn, 711 struct timespec *tsp_in) 712 { 713 struct kevent *kevp; 714 struct timespec *tsp, ats; 715 int i, n, total, error, nerrors = 0; 716 int lres; 717 int limit = kq_checkloop; 718 struct kevent kev[KQ_NEVENTS]; 719 struct knote marker; 720 struct lwkt_token *tok; 721 722 if (tsp_in == NULL || tsp_in->tv_sec || tsp_in->tv_nsec) 723 atomic_set_int(&curthread->td_mpflags, TDF_MP_BATCH_DEMARC); 724 725 tsp = tsp_in; 726 *res = 0; 727 728 for (;;) { 729 n = 0; 730 error = kevent_copyinfn(uap, kev, KQ_NEVENTS, &n); 731 if (error) 732 return error; 733 if (n == 0) 734 break; 735 for (i = 0; i < n; i++) { 736 kevp = &kev[i]; 737 kevp->flags &= ~EV_SYSFLAGS; 738 error = kqueue_register(kq, kevp); 739 740 /* 741 * If a registration returns an error we 742 * immediately post the error. The kevent() 743 * call itself will fail with the error if 744 * no space is available for posting. 745 * 746 * Such errors normally bypass the timeout/blocking 747 * code. However, if the copyoutfn function refuses 748 * to post the error (see sys_poll()), then we 749 * ignore it too. 750 */ 751 if (error || (kevp->flags & EV_RECEIPT)) { 752 kevp->flags = EV_ERROR; 753 kevp->data = error; 754 lres = *res; 755 kevent_copyoutfn(uap, kevp, 1, res); 756 if (*res < 0) { 757 return error; 758 } else if (lres != *res) { 759 nevents--; 760 nerrors++; 761 } 762 } 763 } 764 } 765 if (nerrors) 766 return 0; 767 768 /* 769 * Acquire/wait for events - setup timeout 770 */ 771 if (tsp != NULL) { 772 if (tsp->tv_sec || tsp->tv_nsec) { 773 getnanouptime(&ats); 774 timespecadd(tsp, &ats); /* tsp = target time */ 775 } 776 } 777 778 /* 779 * Loop as required. 780 * 781 * Collect as many events as we can. Sleeping on successive 782 * loops is disabled if copyoutfn has incremented (*res). 783 * 784 * The loop stops if an error occurs, all events have been 785 * scanned (the marker has been reached), or fewer than the 786 * maximum number of events is found. 787 * 788 * The copyoutfn function does not have to increment (*res) in 789 * order for the loop to continue. 790 * 791 * NOTE: doselect() usually passes 0x7FFFFFFF for nevents. 792 */ 793 total = 0; 794 error = 0; 795 marker.kn_filter = EVFILT_MARKER; 796 marker.kn_status = KN_PROCESSING; 797 tok = lwkt_token_pool_lookup(kq); 798 lwkt_gettoken(tok); 799 TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker, kn_tqe); 800 lwkt_reltoken(tok); 801 while ((n = nevents - total) > 0) { 802 if (n > KQ_NEVENTS) 803 n = KQ_NEVENTS; 804 805 /* 806 * If no events are pending sleep until timeout (if any) 807 * or an event occurs. 808 * 809 * After the sleep completes the marker is moved to the 810 * end of the list, making any received events available 811 * to our scan. 812 */ 813 if (kq->kq_count == 0 && *res == 0) { 814 int timeout; 815 816 if (tsp == NULL) { 817 timeout = 0; 818 } else if (tsp->tv_sec == 0 && tsp->tv_nsec == 0) { 819 error = EWOULDBLOCK; 820 break; 821 } else { 822 struct timespec atx = *tsp; 823 824 getnanouptime(&ats); 825 timespecsub(&atx, &ats); 826 if (atx.tv_sec < 0) { 827 error = EWOULDBLOCK; 828 break; 829 } else { 830 timeout = atx.tv_sec > 24 * 60 * 60 ? 831 24 * 60 * 60 * hz : 832 tstohz_high(&atx); 833 } 834 } 835 836 lwkt_gettoken(tok); 837 if (kq->kq_count == 0) { 838 kq->kq_sleep_cnt++; 839 if (__predict_false(kq->kq_sleep_cnt == 0)) { 840 /* 841 * Guard against possible wrapping. And 842 * set it to 2, so that kqueue_wakeup() 843 * can wake everyone up. 844 */ 845 kq->kq_sleep_cnt = 2; 846 } 847 error = tsleep(kq, PCATCH, "kqread", timeout); 848 849 /* don't restart after signals... */ 850 if (error == ERESTART) 851 error = EINTR; 852 if (error) { 853 lwkt_reltoken(tok); 854 break; 855 } 856 857 TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe); 858 TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker, 859 kn_tqe); 860 } 861 lwkt_reltoken(tok); 862 } 863 864 /* 865 * Process all received events 866 * Account for all non-spurious events in our total 867 */ 868 i = kqueue_scan(kq, kev, n, &marker); 869 if (i) { 870 lres = *res; 871 error = kevent_copyoutfn(uap, kev, i, res); 872 total += *res - lres; 873 if (error) 874 break; 875 } 876 if (limit && --limit == 0) 877 panic("kqueue: checkloop failed i=%d", i); 878 879 /* 880 * Normally when fewer events are returned than requested 881 * we can stop. However, if only spurious events were 882 * collected the copyout will not bump (*res) and we have 883 * to continue. 884 */ 885 if (i < n && *res) 886 break; 887 888 /* 889 * Deal with an edge case where spurious events can cause 890 * a loop to occur without moving the marker. This can 891 * prevent kqueue_scan() from picking up new events which 892 * race us. We must be sure to move the marker for this 893 * case. 894 * 895 * NOTE: We do not want to move the marker if events 896 * were scanned because normal kqueue operations 897 * may reactivate events. Moving the marker in 898 * that case could result in duplicates for the 899 * same event. 900 */ 901 if (i == 0) { 902 lwkt_gettoken(tok); 903 TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe); 904 TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker, kn_tqe); 905 lwkt_reltoken(tok); 906 } 907 } 908 lwkt_gettoken(tok); 909 TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe); 910 lwkt_reltoken(tok); 911 912 /* Timeouts do not return EWOULDBLOCK. */ 913 if (error == EWOULDBLOCK) 914 error = 0; 915 return error; 916 } 917 918 /* 919 * MPALMOSTSAFE 920 */ 921 int 922 sys_kevent(struct kevent_args *uap) 923 { 924 struct thread *td = curthread; 925 struct proc *p = td->td_proc; 926 struct timespec ts, *tsp; 927 struct kqueue *kq; 928 struct file *fp = NULL; 929 struct kevent_copyin_args *kap, ka; 930 int error; 931 932 if (uap->timeout) { 933 error = copyin(uap->timeout, &ts, sizeof(ts)); 934 if (error) 935 return (error); 936 tsp = &ts; 937 } else { 938 tsp = NULL; 939 } 940 fp = holdfp(p->p_fd, uap->fd, -1); 941 if (fp == NULL) 942 return (EBADF); 943 if (fp->f_type != DTYPE_KQUEUE) { 944 fdrop(fp); 945 return (EBADF); 946 } 947 948 kq = (struct kqueue *)fp->f_data; 949 950 kap = &ka; 951 kap->ka = uap; 952 kap->pchanges = 0; 953 954 error = kern_kevent(kq, uap->nevents, &uap->sysmsg_result, kap, 955 kevent_copyin, kevent_copyout, tsp); 956 957 fdrop(fp); 958 959 return (error); 960 } 961 962 int 963 kqueue_register(struct kqueue *kq, struct kevent *kev) 964 { 965 struct filedesc *fdp = kq->kq_fdp; 966 struct klist *list = NULL; 967 struct filterops *fops; 968 struct file *fp = NULL; 969 struct knote *kn = NULL; 970 struct thread *td; 971 int error = 0; 972 struct knote_cache_list *cache_list; 973 974 if (kev->filter < 0) { 975 if (kev->filter + EVFILT_SYSCOUNT < 0) 976 return (EINVAL); 977 fops = sysfilt_ops[~kev->filter]; /* to 0-base index */ 978 } else { 979 /* 980 * XXX 981 * filter attach routine is responsible for insuring that 982 * the identifier can be attached to it. 983 */ 984 return (EINVAL); 985 } 986 987 if (fops->f_flags & FILTEROP_ISFD) { 988 /* validate descriptor */ 989 fp = holdfp(fdp, kev->ident, -1); 990 if (fp == NULL) 991 return (EBADF); 992 } 993 994 cache_list = &knote_cache_lists[mycpuid]; 995 if (SLIST_EMPTY(&cache_list->knote_cache)) { 996 struct knote *new_kn; 997 998 new_kn = knote_alloc(); 999 crit_enter(); 1000 SLIST_INSERT_HEAD(&cache_list->knote_cache, new_kn, kn_link); 1001 cache_list->knote_cache_cnt++; 1002 crit_exit(); 1003 } 1004 1005 td = curthread; 1006 lwkt_getpooltoken(kq); 1007 1008 /* 1009 * Make sure that only one thread can register event on this kqueue, 1010 * so that we would not suffer any race, even if the registration 1011 * blocked, i.e. kq token was released, and the kqueue was shared 1012 * between threads (this should be rare though). 1013 */ 1014 while (__predict_false(kq->kq_regtd != NULL && kq->kq_regtd != td)) { 1015 kq->kq_state |= KQ_REGWAIT; 1016 tsleep(&kq->kq_regtd, 0, "kqreg", 0); 1017 } 1018 if (__predict_false(kq->kq_regtd != NULL)) { 1019 /* Recursive calling of kqueue_register() */ 1020 td = NULL; 1021 } else { 1022 /* Owner of the kq_regtd, i.e. td != NULL */ 1023 kq->kq_regtd = td; 1024 } 1025 1026 if (fp != NULL) { 1027 list = &fp->f_klist; 1028 } else if (kq->kq_knhashmask) { 1029 list = &kq->kq_knhash[ 1030 KN_HASH((u_long)kev->ident, kq->kq_knhashmask)]; 1031 } 1032 if (list != NULL) { 1033 lwkt_getpooltoken(list); 1034 again: 1035 SLIST_FOREACH(kn, list, kn_link) { 1036 if (kn->kn_kq == kq && 1037 kn->kn_filter == kev->filter && 1038 kn->kn_id == kev->ident) { 1039 if (knote_acquire(kn) == 0) 1040 goto again; 1041 break; 1042 } 1043 } 1044 lwkt_relpooltoken(list); 1045 } 1046 1047 /* 1048 * NOTE: At this point if kn is non-NULL we will have acquired 1049 * it and set KN_PROCESSING. 1050 */ 1051 if (kn == NULL && ((kev->flags & EV_ADD) == 0)) { 1052 error = ENOENT; 1053 goto done; 1054 } 1055 1056 /* 1057 * kn now contains the matching knote, or NULL if no match 1058 */ 1059 if (kev->flags & EV_ADD) { 1060 if (kn == NULL) { 1061 crit_enter(); 1062 kn = SLIST_FIRST(&cache_list->knote_cache); 1063 if (kn == NULL) { 1064 crit_exit(); 1065 kn = knote_alloc(); 1066 } else { 1067 SLIST_REMOVE_HEAD(&cache_list->knote_cache, 1068 kn_link); 1069 cache_list->knote_cache_cnt--; 1070 crit_exit(); 1071 } 1072 kn->kn_fp = fp; 1073 kn->kn_kq = kq; 1074 kn->kn_fop = fops; 1075 1076 /* 1077 * apply reference count to knote structure, and 1078 * do not release it at the end of this routine. 1079 */ 1080 fp = NULL; 1081 1082 kn->kn_sfflags = kev->fflags; 1083 kn->kn_sdata = kev->data; 1084 kev->fflags = 0; 1085 kev->data = 0; 1086 kn->kn_kevent = *kev; 1087 1088 /* 1089 * KN_PROCESSING prevents the knote from getting 1090 * ripped out from under us while we are trying 1091 * to attach it, in case the attach blocks. 1092 */ 1093 kn->kn_status = KN_PROCESSING; 1094 knote_attach(kn); 1095 if ((error = filter_attach(kn)) != 0) { 1096 kn->kn_status |= KN_DELETING | KN_REPROCESS; 1097 knote_drop(kn); 1098 goto done; 1099 } 1100 1101 /* 1102 * Interlock against close races which either tried 1103 * to remove our knote while we were blocked or missed 1104 * it entirely prior to our attachment. We do not 1105 * want to end up with a knote on a closed descriptor. 1106 */ 1107 if ((fops->f_flags & FILTEROP_ISFD) && 1108 checkfdclosed(fdp, kev->ident, kn->kn_fp)) { 1109 kn->kn_status |= KN_DELETING | KN_REPROCESS; 1110 } 1111 } else { 1112 /* 1113 * The user may change some filter values after the 1114 * initial EV_ADD, but doing so will not reset any 1115 * filter which have already been triggered. 1116 */ 1117 KKASSERT(kn->kn_status & KN_PROCESSING); 1118 if (fops == &user_filtops) { 1119 filt_usertouch(kn, kev, EVENT_REGISTER); 1120 } else { 1121 kn->kn_sfflags = kev->fflags; 1122 kn->kn_sdata = kev->data; 1123 kn->kn_kevent.udata = kev->udata; 1124 } 1125 } 1126 1127 /* 1128 * Execute the filter event to immediately activate the 1129 * knote if necessary. If reprocessing events are pending 1130 * due to blocking above we do not run the filter here 1131 * but instead let knote_release() do it. Otherwise we 1132 * might run the filter on a deleted event. 1133 */ 1134 if ((kn->kn_status & KN_REPROCESS) == 0) { 1135 if (filter_event(kn, 0)) 1136 KNOTE_ACTIVATE(kn); 1137 } 1138 } else if (kev->flags & EV_DELETE) { 1139 /* 1140 * Delete the existing knote 1141 */ 1142 knote_detach_and_drop(kn); 1143 goto done; 1144 } else { 1145 /* 1146 * Modify an existing event. 1147 * 1148 * The user may change some filter values after the 1149 * initial EV_ADD, but doing so will not reset any 1150 * filter which have already been triggered. 1151 */ 1152 KKASSERT(kn->kn_status & KN_PROCESSING); 1153 if (fops == &user_filtops) { 1154 filt_usertouch(kn, kev, EVENT_REGISTER); 1155 } else { 1156 kn->kn_sfflags = kev->fflags; 1157 kn->kn_sdata = kev->data; 1158 kn->kn_kevent.udata = kev->udata; 1159 } 1160 1161 /* 1162 * Execute the filter event to immediately activate the 1163 * knote if necessary. If reprocessing events are pending 1164 * due to blocking above we do not run the filter here 1165 * but instead let knote_release() do it. Otherwise we 1166 * might run the filter on a deleted event. 1167 */ 1168 if ((kn->kn_status & KN_REPROCESS) == 0) { 1169 if (filter_event(kn, 0)) 1170 KNOTE_ACTIVATE(kn); 1171 } 1172 } 1173 1174 /* 1175 * Disablement does not deactivate a knote here. 1176 */ 1177 if ((kev->flags & EV_DISABLE) && 1178 ((kn->kn_status & KN_DISABLED) == 0)) { 1179 kn->kn_status |= KN_DISABLED; 1180 } 1181 1182 /* 1183 * Re-enablement may have to immediately enqueue an active knote. 1184 */ 1185 if ((kev->flags & EV_ENABLE) && (kn->kn_status & KN_DISABLED)) { 1186 kn->kn_status &= ~KN_DISABLED; 1187 if ((kn->kn_status & KN_ACTIVE) && 1188 ((kn->kn_status & KN_QUEUED) == 0)) { 1189 knote_enqueue(kn); 1190 } 1191 } 1192 1193 /* 1194 * Handle any required reprocessing 1195 */ 1196 knote_release(kn); 1197 /* kn may be invalid now */ 1198 1199 done: 1200 if (td != NULL) { /* Owner of the kq_regtd */ 1201 kq->kq_regtd = NULL; 1202 if (__predict_false(kq->kq_state & KQ_REGWAIT)) { 1203 kq->kq_state &= ~KQ_REGWAIT; 1204 wakeup(&kq->kq_regtd); 1205 } 1206 } 1207 lwkt_relpooltoken(kq); 1208 if (fp != NULL) 1209 fdrop(fp); 1210 return (error); 1211 } 1212 1213 /* 1214 * Scan the kqueue, return the number of active events placed in kevp up 1215 * to count. 1216 * 1217 * Continuous mode events may get recycled, do not continue scanning past 1218 * marker unless no events have been collected. 1219 */ 1220 static int 1221 kqueue_scan(struct kqueue *kq, struct kevent *kevp, int count, 1222 struct knote *marker) 1223 { 1224 struct knote *kn, local_marker; 1225 int total; 1226 1227 total = 0; 1228 local_marker.kn_filter = EVFILT_MARKER; 1229 local_marker.kn_status = KN_PROCESSING; 1230 1231 lwkt_getpooltoken(kq); 1232 1233 /* 1234 * Collect events. 1235 */ 1236 TAILQ_INSERT_HEAD(&kq->kq_knpend, &local_marker, kn_tqe); 1237 while (count) { 1238 kn = TAILQ_NEXT(&local_marker, kn_tqe); 1239 if (kn->kn_filter == EVFILT_MARKER) { 1240 /* Marker reached, we are done */ 1241 if (kn == marker) 1242 break; 1243 1244 /* Move local marker past some other threads marker */ 1245 kn = TAILQ_NEXT(kn, kn_tqe); 1246 TAILQ_REMOVE(&kq->kq_knpend, &local_marker, kn_tqe); 1247 TAILQ_INSERT_BEFORE(kn, &local_marker, kn_tqe); 1248 continue; 1249 } 1250 1251 /* 1252 * We can't skip a knote undergoing processing, otherwise 1253 * we risk not returning it when the user process expects 1254 * it should be returned. Sleep and retry. 1255 */ 1256 if (knote_acquire(kn) == 0) 1257 continue; 1258 1259 /* 1260 * Remove the event for processing. 1261 * 1262 * WARNING! We must leave KN_QUEUED set to prevent the 1263 * event from being KNOTE_ACTIVATE()d while 1264 * the queue state is in limbo, in case we 1265 * block. 1266 */ 1267 TAILQ_REMOVE(&kq->kq_knpend, kn, kn_tqe); 1268 kq->kq_count--; 1269 1270 /* 1271 * We have to deal with an extremely important race against 1272 * file descriptor close()s here. The file descriptor can 1273 * disappear MPSAFE, and there is a small window of 1274 * opportunity between that and the call to knote_fdclose(). 1275 * 1276 * If we hit that window here while doselect or dopoll is 1277 * trying to delete a spurious event they will not be able 1278 * to match up the event against a knote and will go haywire. 1279 */ 1280 if ((kn->kn_fop->f_flags & FILTEROP_ISFD) && 1281 checkfdclosed(kq->kq_fdp, kn->kn_kevent.ident, kn->kn_fp)) { 1282 kn->kn_status |= KN_DELETING | KN_REPROCESS; 1283 } 1284 1285 if (kn->kn_status & KN_DISABLED) { 1286 /* 1287 * If disabled we ensure the event is not queued 1288 * but leave its active bit set. On re-enablement 1289 * the event may be immediately triggered. 1290 */ 1291 kn->kn_status &= ~KN_QUEUED; 1292 } else if ((kn->kn_flags & EV_ONESHOT) == 0 && 1293 (kn->kn_status & KN_DELETING) == 0 && 1294 filter_event(kn, 0) == 0) { 1295 /* 1296 * If not running in one-shot mode and the event 1297 * is no longer present we ensure it is removed 1298 * from the queue and ignore it. 1299 */ 1300 kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE); 1301 } else { 1302 /* 1303 * Post the event 1304 */ 1305 if (kn->kn_fop == &user_filtops) 1306 filt_usertouch(kn, kevp, EVENT_PROCESS); 1307 else 1308 *kevp = kn->kn_kevent; 1309 ++kevp; 1310 ++total; 1311 --count; 1312 1313 if (kn->kn_flags & EV_ONESHOT) { 1314 kn->kn_status &= ~KN_QUEUED; 1315 kn->kn_status |= KN_DELETING | KN_REPROCESS; 1316 } else { 1317 if (kn->kn_flags & (EV_CLEAR | EV_DISPATCH)) { 1318 if (kn->kn_flags & EV_CLEAR) { 1319 kn->kn_data = 0; 1320 kn->kn_fflags = 0; 1321 } 1322 if (kn->kn_flags & EV_DISPATCH) { 1323 kn->kn_status |= KN_DISABLED; 1324 } 1325 kn->kn_status &= ~(KN_QUEUED | 1326 KN_ACTIVE); 1327 } else { 1328 TAILQ_INSERT_TAIL(&kq->kq_knpend, kn, kn_tqe); 1329 kq->kq_count++; 1330 } 1331 } 1332 } 1333 1334 /* 1335 * Handle any post-processing states 1336 */ 1337 knote_release(kn); 1338 } 1339 TAILQ_REMOVE(&kq->kq_knpend, &local_marker, kn_tqe); 1340 1341 lwkt_relpooltoken(kq); 1342 return (total); 1343 } 1344 1345 /* 1346 * XXX 1347 * This could be expanded to call kqueue_scan, if desired. 1348 * 1349 * MPSAFE 1350 */ 1351 static int 1352 kqueue_read(struct file *fp, struct uio *uio, struct ucred *cred, int flags) 1353 { 1354 return (ENXIO); 1355 } 1356 1357 /* 1358 * MPSAFE 1359 */ 1360 static int 1361 kqueue_write(struct file *fp, struct uio *uio, struct ucred *cred, int flags) 1362 { 1363 return (ENXIO); 1364 } 1365 1366 /* 1367 * MPALMOSTSAFE 1368 */ 1369 static int 1370 kqueue_ioctl(struct file *fp, u_long com, caddr_t data, 1371 struct ucred *cred, struct sysmsg *msg) 1372 { 1373 struct kqueue *kq; 1374 int error; 1375 1376 kq = (struct kqueue *)fp->f_data; 1377 lwkt_getpooltoken(kq); 1378 switch(com) { 1379 case FIOASYNC: 1380 if (*(int *)data) 1381 kq->kq_state |= KQ_ASYNC; 1382 else 1383 kq->kq_state &= ~KQ_ASYNC; 1384 error = 0; 1385 break; 1386 case FIOSETOWN: 1387 error = fsetown(*(int *)data, &kq->kq_sigio); 1388 break; 1389 default: 1390 error = ENOTTY; 1391 break; 1392 } 1393 lwkt_relpooltoken(kq); 1394 return (error); 1395 } 1396 1397 /* 1398 * MPSAFE 1399 */ 1400 static int 1401 kqueue_stat(struct file *fp, struct stat *st, struct ucred *cred) 1402 { 1403 struct kqueue *kq = (struct kqueue *)fp->f_data; 1404 1405 bzero((void *)st, sizeof(*st)); 1406 st->st_size = kq->kq_count; 1407 st->st_blksize = sizeof(struct kevent); 1408 st->st_mode = S_IFIFO; 1409 return (0); 1410 } 1411 1412 /* 1413 * MPSAFE 1414 */ 1415 static int 1416 kqueue_close(struct file *fp) 1417 { 1418 struct kqueue *kq = (struct kqueue *)fp->f_data; 1419 1420 kqueue_terminate(kq); 1421 1422 fp->f_data = NULL; 1423 funsetown(&kq->kq_sigio); 1424 1425 kfree(kq, M_KQUEUE); 1426 return (0); 1427 } 1428 1429 static void 1430 kqueue_wakeup(struct kqueue *kq) 1431 { 1432 if (kq->kq_sleep_cnt) { 1433 u_int sleep_cnt = kq->kq_sleep_cnt; 1434 1435 kq->kq_sleep_cnt = 0; 1436 if (sleep_cnt == 1) 1437 wakeup_one(kq); 1438 else 1439 wakeup(kq); 1440 } 1441 KNOTE(&kq->kq_kqinfo.ki_note, 0); 1442 } 1443 1444 /* 1445 * Calls filterops f_attach function, acquiring mplock if filter is not 1446 * marked as FILTEROP_MPSAFE. 1447 * 1448 * Caller must be holding the related kq token 1449 */ 1450 static int 1451 filter_attach(struct knote *kn) 1452 { 1453 int ret; 1454 1455 if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) { 1456 ret = kn->kn_fop->f_attach(kn); 1457 } else { 1458 get_mplock(); 1459 ret = kn->kn_fop->f_attach(kn); 1460 rel_mplock(); 1461 } 1462 return (ret); 1463 } 1464 1465 /* 1466 * Detach the knote and drop it, destroying the knote. 1467 * 1468 * Calls filterops f_detach function, acquiring mplock if filter is not 1469 * marked as FILTEROP_MPSAFE. 1470 * 1471 * Caller must be holding the related kq token 1472 */ 1473 static void 1474 knote_detach_and_drop(struct knote *kn) 1475 { 1476 kn->kn_status |= KN_DELETING | KN_REPROCESS; 1477 if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) { 1478 kn->kn_fop->f_detach(kn); 1479 } else { 1480 get_mplock(); 1481 kn->kn_fop->f_detach(kn); 1482 rel_mplock(); 1483 } 1484 knote_drop(kn); 1485 } 1486 1487 /* 1488 * Calls filterops f_event function, acquiring mplock if filter is not 1489 * marked as FILTEROP_MPSAFE. 1490 * 1491 * If the knote is in the middle of being created or deleted we cannot 1492 * safely call the filter op. 1493 * 1494 * Caller must be holding the related kq token 1495 */ 1496 static int 1497 filter_event(struct knote *kn, long hint) 1498 { 1499 int ret; 1500 1501 if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) { 1502 ret = kn->kn_fop->f_event(kn, hint); 1503 } else { 1504 get_mplock(); 1505 ret = kn->kn_fop->f_event(kn, hint); 1506 rel_mplock(); 1507 } 1508 return (ret); 1509 } 1510 1511 /* 1512 * Walk down a list of knotes, activating them if their event has triggered. 1513 * 1514 * If we encounter any knotes which are undergoing processing we just mark 1515 * them for reprocessing and do not try to [re]activate the knote. However, 1516 * if a hint is being passed we have to wait and that makes things a bit 1517 * sticky. 1518 */ 1519 void 1520 knote(struct klist *list, long hint) 1521 { 1522 struct knote *kn, marker; 1523 1524 marker.kn_filter = EVFILT_MARKER; 1525 marker.kn_status = KN_PROCESSING; 1526 1527 lwkt_getpooltoken(list); 1528 if (SLIST_EMPTY(list)) { 1529 lwkt_relpooltoken(list); 1530 return; 1531 } 1532 1533 SLIST_INSERT_HEAD(list, &marker, kn_next); 1534 while ((kn = SLIST_NEXT(&marker, kn_next)) != NULL) { 1535 struct kqueue *kq; 1536 int last_knote = 0; 1537 1538 if (kn->kn_filter == EVFILT_MARKER) { 1539 /* Skip marker */ 1540 SLIST_REMOVE(list, &marker, knote, kn_next); 1541 if (SLIST_NEXT(kn, kn_next) == NULL) 1542 goto done; 1543 SLIST_INSERT_AFTER(kn, &marker, kn_next); 1544 continue; 1545 } 1546 1547 kq = kn->kn_kq; 1548 lwkt_getpooltoken(kq); 1549 1550 if (kn != SLIST_NEXT(&marker, kn_next) || kn->kn_kq != kq) { 1551 /* 1552 * Don't move the marker; check the knote after 1553 * the marker again. 1554 */ 1555 lwkt_relpooltoken(kq); 1556 continue; 1557 } 1558 1559 if (kn->kn_status & KN_PROCESSING) { 1560 /* 1561 * Someone else is processing the knote, ask the 1562 * other thread to reprocess it and don't mess 1563 * with it otherwise. 1564 */ 1565 if (hint == 0) { 1566 /* 1567 * Move the marker w/ the kq token, so that 1568 * this knote will not be ripped behind our 1569 * back. 1570 */ 1571 SLIST_REMOVE(list, &marker, knote, kn_next); 1572 if (SLIST_NEXT(kn, kn_next) != NULL) 1573 SLIST_INSERT_AFTER(kn, &marker, kn_next); 1574 else 1575 last_knote = 1; 1576 kn->kn_status |= KN_REPROCESS; 1577 lwkt_relpooltoken(kq); 1578 1579 if (last_knote) 1580 goto done; 1581 continue; 1582 } 1583 1584 /* 1585 * If the hint is non-zero we have to wait or risk 1586 * losing the state the caller is trying to update. 1587 */ 1588 kn->kn_status |= KN_WAITING | KN_REPROCESS; 1589 tsleep(kn, 0, "knotec", hz); 1590 1591 /* 1592 * Don't move the marker; check this knote again, 1593 * hopefully it is still after the marker. Or it 1594 * was deleted and we would check the next knote. 1595 */ 1596 lwkt_relpooltoken(kq); 1597 continue; 1598 } 1599 1600 /* 1601 * Become the reprocessing master ourselves. 1602 */ 1603 KASSERT((kn->kn_status & KN_DELETING) == 0, 1604 ("acquire a deleting knote %#x", kn->kn_status)); 1605 kn->kn_status |= KN_PROCESSING; 1606 1607 /* Move the marker */ 1608 SLIST_REMOVE(list, &marker, knote, kn_next); 1609 if (SLIST_NEXT(kn, kn_next) != NULL) 1610 SLIST_INSERT_AFTER(kn, &marker, kn_next); 1611 else 1612 last_knote = 1; 1613 1614 /* 1615 * If hint is non-zero running the event is mandatory 1616 * so do it whether reprocessing is set or not. 1617 */ 1618 if (filter_event(kn, hint)) 1619 KNOTE_ACTIVATE(kn); 1620 1621 knote_release(kn); 1622 lwkt_relpooltoken(kq); 1623 1624 if (last_knote) 1625 goto done; 1626 } 1627 SLIST_REMOVE(list, &marker, knote, kn_next); 1628 done: 1629 lwkt_relpooltoken(list); 1630 } 1631 1632 /* 1633 * Insert knote at head of klist. 1634 * 1635 * This function may only be called via a filter function and thus 1636 * kq_token should already be held and marked for processing. 1637 */ 1638 void 1639 knote_insert(struct klist *klist, struct knote *kn) 1640 { 1641 lwkt_getpooltoken(klist); 1642 KKASSERT(kn->kn_status & KN_PROCESSING); 1643 SLIST_INSERT_HEAD(klist, kn, kn_next); 1644 lwkt_relpooltoken(klist); 1645 } 1646 1647 /* 1648 * Remove knote from a klist 1649 * 1650 * This function may only be called via a filter function and thus 1651 * kq_token should already be held and marked for processing. 1652 */ 1653 void 1654 knote_remove(struct klist *klist, struct knote *kn) 1655 { 1656 lwkt_getpooltoken(klist); 1657 KKASSERT(kn->kn_status & KN_PROCESSING); 1658 SLIST_REMOVE(klist, kn, knote, kn_next); 1659 lwkt_relpooltoken(klist); 1660 } 1661 1662 void 1663 knote_assume_knotes(struct kqinfo *src, struct kqinfo *dst, 1664 struct filterops *ops, void *hook) 1665 { 1666 struct knote *kn, marker; 1667 int has_note; 1668 1669 marker.kn_filter = EVFILT_MARKER; 1670 marker.kn_status = KN_PROCESSING; 1671 1672 lwkt_getpooltoken(&src->ki_note); 1673 if (SLIST_EMPTY(&src->ki_note)) { 1674 lwkt_relpooltoken(&src->ki_note); 1675 return; 1676 } 1677 lwkt_getpooltoken(&dst->ki_note); 1678 1679 restart: 1680 has_note = 0; 1681 SLIST_INSERT_HEAD(&src->ki_note, &marker, kn_next); 1682 while ((kn = SLIST_NEXT(&marker, kn_next)) != NULL) { 1683 struct kqueue *kq; 1684 1685 if (kn->kn_filter == EVFILT_MARKER) { 1686 /* Skip marker */ 1687 SLIST_REMOVE(&src->ki_note, &marker, knote, kn_next); 1688 SLIST_INSERT_AFTER(kn, &marker, kn_next); 1689 continue; 1690 } 1691 1692 kq = kn->kn_kq; 1693 lwkt_getpooltoken(kq); 1694 1695 if (kn != SLIST_NEXT(&marker, kn_next) || kn->kn_kq != kq) { 1696 /* 1697 * Don't move the marker; check the knote after 1698 * the marker again. 1699 */ 1700 lwkt_relpooltoken(kq); 1701 continue; 1702 } 1703 1704 /* Move marker */ 1705 SLIST_REMOVE(&src->ki_note, &marker, knote, kn_next); 1706 SLIST_INSERT_AFTER(kn, &marker, kn_next); 1707 1708 has_note = 1; 1709 if (knote_acquire(kn)) { 1710 knote_remove(&src->ki_note, kn); 1711 kn->kn_fop = ops; 1712 kn->kn_hook = hook; 1713 knote_insert(&dst->ki_note, kn); 1714 knote_release(kn); 1715 /* kn may be invalid now */ 1716 } 1717 lwkt_relpooltoken(kq); 1718 } 1719 SLIST_REMOVE(&src->ki_note, &marker, knote, kn_next); 1720 if (has_note) { 1721 /* Keep draining, until nothing left */ 1722 goto restart; 1723 } 1724 1725 lwkt_relpooltoken(&dst->ki_note); 1726 lwkt_relpooltoken(&src->ki_note); 1727 } 1728 1729 /* 1730 * Remove all knotes referencing a specified fd 1731 */ 1732 void 1733 knote_fdclose(struct file *fp, struct filedesc *fdp, int fd) 1734 { 1735 struct kqueue *kq; 1736 struct knote *kn; 1737 struct knote *kntmp; 1738 1739 lwkt_getpooltoken(&fp->f_klist); 1740 restart: 1741 SLIST_FOREACH(kn, &fp->f_klist, kn_link) { 1742 if (kn->kn_kq->kq_fdp == fdp && kn->kn_id == fd) { 1743 kq = kn->kn_kq; 1744 lwkt_getpooltoken(kq); 1745 1746 /* temporary verification hack */ 1747 SLIST_FOREACH(kntmp, &fp->f_klist, kn_link) { 1748 if (kn == kntmp) 1749 break; 1750 } 1751 if (kn != kntmp || kn->kn_kq->kq_fdp != fdp || 1752 kn->kn_id != fd || kn->kn_kq != kq) { 1753 lwkt_relpooltoken(kq); 1754 goto restart; 1755 } 1756 if (knote_acquire(kn)) 1757 knote_detach_and_drop(kn); 1758 lwkt_relpooltoken(kq); 1759 goto restart; 1760 } 1761 } 1762 lwkt_relpooltoken(&fp->f_klist); 1763 } 1764 1765 /* 1766 * Low level attach function. 1767 * 1768 * The knote should already be marked for processing. 1769 * Caller must hold the related kq token. 1770 */ 1771 static void 1772 knote_attach(struct knote *kn) 1773 { 1774 struct klist *list; 1775 struct kqueue *kq = kn->kn_kq; 1776 1777 if (kn->kn_fop->f_flags & FILTEROP_ISFD) { 1778 KKASSERT(kn->kn_fp); 1779 list = &kn->kn_fp->f_klist; 1780 } else { 1781 if (kq->kq_knhashmask == 0) 1782 kq->kq_knhash = hashinit(KN_HASHSIZE, M_KQUEUE, 1783 &kq->kq_knhashmask); 1784 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)]; 1785 } 1786 lwkt_getpooltoken(list); 1787 SLIST_INSERT_HEAD(list, kn, kn_link); 1788 lwkt_relpooltoken(list); 1789 TAILQ_INSERT_HEAD(&kq->kq_knlist, kn, kn_kqlink); 1790 } 1791 1792 /* 1793 * Low level drop function. 1794 * 1795 * The knote should already be marked for processing. 1796 * Caller must hold the related kq token. 1797 */ 1798 static void 1799 knote_drop(struct knote *kn) 1800 { 1801 struct kqueue *kq; 1802 struct klist *list; 1803 1804 kq = kn->kn_kq; 1805 1806 if (kn->kn_fop->f_flags & FILTEROP_ISFD) 1807 list = &kn->kn_fp->f_klist; 1808 else 1809 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)]; 1810 1811 lwkt_getpooltoken(list); 1812 SLIST_REMOVE(list, kn, knote, kn_link); 1813 lwkt_relpooltoken(list); 1814 TAILQ_REMOVE(&kq->kq_knlist, kn, kn_kqlink); 1815 if (kn->kn_status & KN_QUEUED) 1816 knote_dequeue(kn); 1817 if (kn->kn_fop->f_flags & FILTEROP_ISFD) { 1818 fdrop(kn->kn_fp); 1819 kn->kn_fp = NULL; 1820 } 1821 knote_free(kn); 1822 } 1823 1824 /* 1825 * Low level enqueue function. 1826 * 1827 * The knote should already be marked for processing. 1828 * Caller must be holding the kq token 1829 */ 1830 static void 1831 knote_enqueue(struct knote *kn) 1832 { 1833 struct kqueue *kq = kn->kn_kq; 1834 1835 KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued")); 1836 TAILQ_INSERT_TAIL(&kq->kq_knpend, kn, kn_tqe); 1837 kn->kn_status |= KN_QUEUED; 1838 ++kq->kq_count; 1839 1840 /* 1841 * Send SIGIO on request (typically set up as a mailbox signal) 1842 */ 1843 if (kq->kq_sigio && (kq->kq_state & KQ_ASYNC) && kq->kq_count == 1) 1844 pgsigio(kq->kq_sigio, SIGIO, 0); 1845 1846 kqueue_wakeup(kq); 1847 } 1848 1849 /* 1850 * Low level dequeue function. 1851 * 1852 * The knote should already be marked for processing. 1853 * Caller must be holding the kq token 1854 */ 1855 static void 1856 knote_dequeue(struct knote *kn) 1857 { 1858 struct kqueue *kq = kn->kn_kq; 1859 1860 KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued")); 1861 TAILQ_REMOVE(&kq->kq_knpend, kn, kn_tqe); 1862 kn->kn_status &= ~KN_QUEUED; 1863 kq->kq_count--; 1864 } 1865 1866 static struct knote * 1867 knote_alloc(void) 1868 { 1869 return kmalloc(sizeof(struct knote), M_KQUEUE, M_WAITOK); 1870 } 1871 1872 static void 1873 knote_free(struct knote *kn) 1874 { 1875 struct knote_cache_list *cache_list; 1876 1877 cache_list = &knote_cache_lists[mycpuid]; 1878 if (cache_list->knote_cache_cnt < KNOTE_CACHE_MAX) { 1879 crit_enter(); 1880 SLIST_INSERT_HEAD(&cache_list->knote_cache, kn, kn_link); 1881 cache_list->knote_cache_cnt++; 1882 crit_exit(); 1883 return; 1884 } 1885 kfree(kn, M_KQUEUE); 1886 } 1887