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 #include <sys/spinlock.h> 53 54 #include <sys/thread2.h> 55 #include <sys/file2.h> 56 #include <sys/mplock2.h> 57 #include <sys/spinlock2.h> 58 59 #define EVENT_REGISTER 1 60 #define EVENT_PROCESS 2 61 62 MALLOC_DEFINE(M_KQUEUE, "kqueue", "memory for kqueue system"); 63 64 struct kevent_copyin_args { 65 struct kevent_args *ka; 66 int pchanges; 67 }; 68 69 #define KNOTE_CACHE_MAX 8 70 71 struct knote_cache_list { 72 struct klist knote_cache; 73 int knote_cache_cnt; 74 } __cachealign; 75 76 static int kqueue_scan(struct kqueue *kq, struct kevent *kevp, int count, 77 struct knote *marker, int closedcounter); 78 static int kqueue_read(struct file *fp, struct uio *uio, 79 struct ucred *cred, int flags); 80 static int kqueue_write(struct file *fp, struct uio *uio, 81 struct ucred *cred, int flags); 82 static int kqueue_ioctl(struct file *fp, u_long com, caddr_t data, 83 struct ucred *cred, struct sysmsg *msg); 84 static int kqueue_kqfilter(struct file *fp, struct knote *kn); 85 static int kqueue_stat(struct file *fp, struct stat *st, 86 struct ucred *cred); 87 static int kqueue_close(struct file *fp); 88 static void kqueue_wakeup(struct kqueue *kq); 89 static int filter_attach(struct knote *kn); 90 static int filter_event(struct knote *kn, long hint); 91 92 /* 93 * MPSAFE 94 */ 95 static struct fileops kqueueops = { 96 .fo_read = kqueue_read, 97 .fo_write = kqueue_write, 98 .fo_ioctl = kqueue_ioctl, 99 .fo_kqfilter = kqueue_kqfilter, 100 .fo_stat = kqueue_stat, 101 .fo_close = kqueue_close, 102 .fo_shutdown = nofo_shutdown 103 }; 104 105 static void knote_attach(struct knote *kn); 106 static void knote_drop(struct knote *kn); 107 static void knote_detach_and_drop(struct knote *kn); 108 static void knote_enqueue(struct knote *kn); 109 static void knote_dequeue(struct knote *kn); 110 static struct knote *knote_alloc(void); 111 static void knote_free(struct knote *kn); 112 113 static void precise_sleep_intr(systimer_t info, int in_ipi, 114 struct intrframe *frame); 115 static int precise_sleep(void *ident, int flags, const char *wmesg, 116 int us); 117 118 static void filt_kqdetach(struct knote *kn); 119 static int filt_kqueue(struct knote *kn, long hint); 120 static int filt_procattach(struct knote *kn); 121 static void filt_procdetach(struct knote *kn); 122 static int filt_proc(struct knote *kn, long hint); 123 static int filt_fileattach(struct knote *kn); 124 static void filt_timerexpire(void *knx); 125 static int filt_timerattach(struct knote *kn); 126 static void filt_timerdetach(struct knote *kn); 127 static int filt_timer(struct knote *kn, long hint); 128 static int filt_userattach(struct knote *kn); 129 static void filt_userdetach(struct knote *kn); 130 static int filt_user(struct knote *kn, long hint); 131 static void filt_usertouch(struct knote *kn, struct kevent *kev, 132 u_long type); 133 static int filt_fsattach(struct knote *kn); 134 static void filt_fsdetach(struct knote *kn); 135 static int filt_fs(struct knote *kn, long hint); 136 137 static struct filterops file_filtops = 138 { FILTEROP_ISFD | FILTEROP_MPSAFE, filt_fileattach, NULL, NULL }; 139 static struct filterops kqread_filtops = 140 { FILTEROP_ISFD | FILTEROP_MPSAFE, NULL, filt_kqdetach, filt_kqueue }; 141 static struct filterops proc_filtops = 142 { FILTEROP_MPSAFE, filt_procattach, filt_procdetach, filt_proc }; 143 static struct filterops timer_filtops = 144 { FILTEROP_MPSAFE, filt_timerattach, filt_timerdetach, filt_timer }; 145 static struct filterops user_filtops = 146 { FILTEROP_MPSAFE, filt_userattach, filt_userdetach, filt_user }; 147 static struct filterops fs_filtops = 148 { FILTEROP_MPSAFE, filt_fsattach, filt_fsdetach, filt_fs }; 149 150 static int kq_ncallouts = 0; 151 static int kq_calloutmax = (4 * 1024); 152 SYSCTL_INT(_kern, OID_AUTO, kq_calloutmax, CTLFLAG_RW, 153 &kq_calloutmax, 0, "Maximum number of callouts allocated for kqueue"); 154 static int kq_checkloop = 1000000; 155 SYSCTL_INT(_kern, OID_AUTO, kq_checkloop, CTLFLAG_RW, 156 &kq_checkloop, 0, "Maximum number of loops for kqueue scan"); 157 static int kq_sleep_threshold = 20000; 158 SYSCTL_INT(_kern, OID_AUTO, kq_sleep_threshold, CTLFLAG_RW, 159 &kq_sleep_threshold, 0, "Minimum sleep duration without busy-looping"); 160 161 #define KNOTE_ACTIVATE(kn) do { \ 162 kn->kn_status |= KN_ACTIVE; \ 163 if ((kn->kn_status & (KN_QUEUED | KN_DISABLED)) == 0) \ 164 knote_enqueue(kn); \ 165 } while(0) 166 167 #define KN_HASHSIZE 64 /* XXX should be tunable */ 168 #define KN_HASH(val, mask) (((val) ^ (val >> 8)) & (mask)) 169 170 extern struct filterops aio_filtops; 171 extern struct filterops sig_filtops; 172 173 /* 174 * Table for for all system-defined filters. 175 */ 176 static struct filterops *sysfilt_ops[] = { 177 &file_filtops, /* EVFILT_READ */ 178 &file_filtops, /* EVFILT_WRITE */ 179 &aio_filtops, /* EVFILT_AIO */ 180 &file_filtops, /* EVFILT_VNODE */ 181 &proc_filtops, /* EVFILT_PROC */ 182 &sig_filtops, /* EVFILT_SIGNAL */ 183 &timer_filtops, /* EVFILT_TIMER */ 184 &file_filtops, /* EVFILT_EXCEPT */ 185 &user_filtops, /* EVFILT_USER */ 186 &fs_filtops, /* EVFILT_FS */ 187 }; 188 189 static struct knote_cache_list knote_cache_lists[MAXCPU]; 190 191 /* 192 * Acquire a knote, return non-zero on success, 0 on failure. 193 * 194 * If we cannot acquire the knote we sleep and return 0. The knote 195 * may be stale on return in this case and the caller must restart 196 * whatever loop they are in. 197 * 198 * Related kq token must be held. 199 */ 200 static __inline int 201 knote_acquire(struct knote *kn) 202 { 203 if (kn->kn_status & KN_PROCESSING) { 204 kn->kn_status |= KN_WAITING | KN_REPROCESS; 205 tsleep(kn, 0, "kqepts", hz); 206 /* knote may be stale now */ 207 return(0); 208 } 209 kn->kn_status |= KN_PROCESSING; 210 return(1); 211 } 212 213 /* 214 * Release an acquired knote, clearing KN_PROCESSING and handling any 215 * KN_REPROCESS events. 216 * 217 * Caller must be holding the related kq token 218 * 219 * Non-zero is returned if the knote is destroyed or detached. 220 */ 221 static __inline int 222 knote_release(struct knote *kn) 223 { 224 int ret; 225 226 while (kn->kn_status & KN_REPROCESS) { 227 kn->kn_status &= ~KN_REPROCESS; 228 if (kn->kn_status & KN_WAITING) { 229 kn->kn_status &= ~KN_WAITING; 230 wakeup(kn); 231 } 232 if (kn->kn_status & KN_DELETING) { 233 knote_detach_and_drop(kn); 234 return(1); 235 /* NOT REACHED */ 236 } 237 if (filter_event(kn, 0)) 238 KNOTE_ACTIVATE(kn); 239 } 240 if (kn->kn_status & KN_DETACHED) 241 ret = 1; 242 else 243 ret = 0; 244 kn->kn_status &= ~KN_PROCESSING; 245 /* kn should not be accessed anymore */ 246 return ret; 247 } 248 249 static int 250 filt_fileattach(struct knote *kn) 251 { 252 return (fo_kqfilter(kn->kn_fp, kn)); 253 } 254 255 /* 256 * MPSAFE 257 */ 258 static int 259 kqueue_kqfilter(struct file *fp, struct knote *kn) 260 { 261 struct kqueue *kq = (struct kqueue *)kn->kn_fp->f_data; 262 263 if (kn->kn_filter != EVFILT_READ) 264 return (EOPNOTSUPP); 265 266 kn->kn_fop = &kqread_filtops; 267 knote_insert(&kq->kq_kqinfo.ki_note, kn); 268 return (0); 269 } 270 271 static void 272 filt_kqdetach(struct knote *kn) 273 { 274 struct kqueue *kq = (struct kqueue *)kn->kn_fp->f_data; 275 276 knote_remove(&kq->kq_kqinfo.ki_note, kn); 277 } 278 279 /*ARGSUSED*/ 280 static int 281 filt_kqueue(struct knote *kn, long hint) 282 { 283 struct kqueue *kq = (struct kqueue *)kn->kn_fp->f_data; 284 285 kn->kn_data = kq->kq_count; 286 return (kn->kn_data > 0); 287 } 288 289 static int 290 filt_procattach(struct knote *kn) 291 { 292 struct proc *p; 293 int immediate; 294 295 immediate = 0; 296 p = pfind(kn->kn_id); 297 if (p == NULL && (kn->kn_sfflags & NOTE_EXIT)) { 298 p = zpfind(kn->kn_id); 299 immediate = 1; 300 } 301 if (p == NULL) { 302 return (ESRCH); 303 } 304 if (!PRISON_CHECK(curthread->td_ucred, p->p_ucred)) { 305 if (p) 306 PRELE(p); 307 return (EACCES); 308 } 309 310 lwkt_gettoken(&p->p_token); 311 kn->kn_ptr.p_proc = p; 312 kn->kn_flags |= EV_CLEAR; /* automatically set */ 313 314 /* 315 * internal flag indicating registration done by kernel 316 */ 317 if (kn->kn_flags & EV_FLAG1) { 318 kn->kn_data = kn->kn_sdata; /* ppid */ 319 kn->kn_fflags = NOTE_CHILD; 320 kn->kn_flags &= ~EV_FLAG1; 321 } 322 323 knote_insert(&p->p_klist, kn); 324 325 /* 326 * Immediately activate any exit notes if the target process is a 327 * zombie. This is necessary to handle the case where the target 328 * process, e.g. a child, dies before the kevent is negistered. 329 */ 330 if (immediate && filt_proc(kn, NOTE_EXIT)) 331 KNOTE_ACTIVATE(kn); 332 lwkt_reltoken(&p->p_token); 333 PRELE(p); 334 335 return (0); 336 } 337 338 /* 339 * The knote may be attached to a different process, which may exit, 340 * leaving nothing for the knote to be attached to. So when the process 341 * exits, the knote is marked as DETACHED and also flagged as ONESHOT so 342 * it will be deleted when read out. However, as part of the knote deletion, 343 * this routine is called, so a check is needed to avoid actually performing 344 * a detach, because the original process does not exist any more. 345 */ 346 static void 347 filt_procdetach(struct knote *kn) 348 { 349 struct proc *p; 350 351 if (kn->kn_status & KN_DETACHED) 352 return; 353 p = kn->kn_ptr.p_proc; 354 knote_remove(&p->p_klist, kn); 355 } 356 357 static int 358 filt_proc(struct knote *kn, long hint) 359 { 360 u_int event; 361 362 /* 363 * mask off extra data 364 */ 365 event = (u_int)hint & NOTE_PCTRLMASK; 366 367 /* 368 * if the user is interested in this event, record it. 369 */ 370 if (kn->kn_sfflags & event) 371 kn->kn_fflags |= event; 372 373 /* 374 * Process is gone, so flag the event as finished. Detach the 375 * knote from the process now because the process will be poof, 376 * gone later on. 377 */ 378 if (event == NOTE_EXIT) { 379 struct proc *p = kn->kn_ptr.p_proc; 380 if ((kn->kn_status & KN_DETACHED) == 0) { 381 PHOLD(p); 382 knote_remove(&p->p_klist, kn); 383 kn->kn_status |= KN_DETACHED; 384 kn->kn_data = p->p_xstat; 385 kn->kn_ptr.p_proc = NULL; 386 PRELE(p); 387 } 388 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT); 389 return (1); 390 } 391 392 /* 393 * process forked, and user wants to track the new process, 394 * so attach a new knote to it, and immediately report an 395 * event with the parent's pid. 396 */ 397 if ((event == NOTE_FORK) && (kn->kn_sfflags & NOTE_TRACK)) { 398 struct kevent kev; 399 int error; 400 int n; 401 402 /* 403 * register knote with new process. 404 */ 405 kev.ident = hint & NOTE_PDATAMASK; /* pid */ 406 kev.filter = kn->kn_filter; 407 kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1; 408 kev.fflags = kn->kn_sfflags; 409 kev.data = kn->kn_id; /* parent */ 410 kev.udata = kn->kn_kevent.udata; /* preserve udata */ 411 n = 1; 412 error = kqueue_register(kn->kn_kq, &kev, &n); 413 if (error) 414 kn->kn_fflags |= NOTE_TRACKERR; 415 } 416 417 return (kn->kn_fflags != 0); 418 } 419 420 static void 421 filt_timerreset(struct knote *kn) 422 { 423 struct callout *calloutp; 424 struct timeval tv; 425 int tticks; 426 427 tv.tv_sec = kn->kn_sdata / 1000; 428 tv.tv_usec = (kn->kn_sdata % 1000) * 1000; 429 tticks = tvtohz_high(&tv); 430 calloutp = (struct callout *)kn->kn_hook; 431 callout_reset(calloutp, tticks, filt_timerexpire, kn); 432 } 433 434 /* 435 * The callout interlocks with callout_terminate() but can still 436 * race a deletion so if KN_DELETING is set we just don't touch 437 * the knote. 438 */ 439 static void 440 filt_timerexpire(void *knx) 441 { 442 struct knote *kn = knx; 443 struct kqueue *kq = kn->kn_kq; 444 445 lwkt_getpooltoken(kq); 446 447 /* 448 * Open knote_acquire(), since we can't sleep in callout, 449 * however, we do need to record this expiration. 450 */ 451 kn->kn_data++; 452 if (kn->kn_status & KN_PROCESSING) { 453 kn->kn_status |= KN_REPROCESS; 454 if ((kn->kn_status & KN_DELETING) == 0 && 455 (kn->kn_flags & EV_ONESHOT) == 0) 456 filt_timerreset(kn); 457 lwkt_relpooltoken(kq); 458 return; 459 } 460 KASSERT((kn->kn_status & KN_DELETING) == 0, 461 ("acquire a deleting knote %#x", kn->kn_status)); 462 kn->kn_status |= KN_PROCESSING; 463 464 KNOTE_ACTIVATE(kn); 465 if ((kn->kn_flags & EV_ONESHOT) == 0) 466 filt_timerreset(kn); 467 468 knote_release(kn); 469 470 lwkt_relpooltoken(kq); 471 } 472 473 /* 474 * data contains amount of time to sleep, in milliseconds 475 */ 476 static int 477 filt_timerattach(struct knote *kn) 478 { 479 struct callout *calloutp; 480 int prev_ncallouts; 481 482 prev_ncallouts = atomic_fetchadd_int(&kq_ncallouts, 1); 483 if (prev_ncallouts >= kq_calloutmax) { 484 atomic_subtract_int(&kq_ncallouts, 1); 485 kn->kn_hook = NULL; 486 return (ENOMEM); 487 } 488 489 kn->kn_flags |= EV_CLEAR; /* automatically set */ 490 calloutp = kmalloc(sizeof(*calloutp), M_KQUEUE, M_WAITOK); 491 callout_init_mp(calloutp); 492 kn->kn_hook = (caddr_t)calloutp; 493 494 filt_timerreset(kn); 495 return (0); 496 } 497 498 /* 499 * This function is called with the knote flagged locked but it is 500 * still possible to race a callout event due to the callback blocking. 501 * We must call callout_terminate() instead of callout_stop() to deal 502 * with the race. 503 */ 504 static void 505 filt_timerdetach(struct knote *kn) 506 { 507 struct callout *calloutp; 508 509 calloutp = (struct callout *)kn->kn_hook; 510 callout_terminate(calloutp); 511 kn->kn_hook = NULL; 512 kfree(calloutp, M_KQUEUE); 513 atomic_subtract_int(&kq_ncallouts, 1); 514 } 515 516 static int 517 filt_timer(struct knote *kn, long hint) 518 { 519 return (kn->kn_data != 0); 520 } 521 522 /* 523 * EVFILT_USER 524 */ 525 static int 526 filt_userattach(struct knote *kn) 527 { 528 u_int ffctrl; 529 530 kn->kn_hook = NULL; 531 if (kn->kn_sfflags & NOTE_TRIGGER) 532 kn->kn_ptr.hookid = 1; 533 else 534 kn->kn_ptr.hookid = 0; 535 536 ffctrl = kn->kn_sfflags & NOTE_FFCTRLMASK; 537 kn->kn_sfflags &= NOTE_FFLAGSMASK; 538 switch (ffctrl) { 539 case NOTE_FFNOP: 540 break; 541 542 case NOTE_FFAND: 543 kn->kn_fflags &= kn->kn_sfflags; 544 break; 545 546 case NOTE_FFOR: 547 kn->kn_fflags |= kn->kn_sfflags; 548 break; 549 550 case NOTE_FFCOPY: 551 kn->kn_fflags = kn->kn_sfflags; 552 break; 553 554 default: 555 /* XXX Return error? */ 556 break; 557 } 558 /* We just happen to copy this value as well. Undocumented. */ 559 kn->kn_data = kn->kn_sdata; 560 561 return 0; 562 } 563 564 static void 565 filt_userdetach(struct knote *kn) 566 { 567 /* nothing to do */ 568 } 569 570 static int 571 filt_user(struct knote *kn, long hint) 572 { 573 return (kn->kn_ptr.hookid); 574 } 575 576 static void 577 filt_usertouch(struct knote *kn, struct kevent *kev, u_long type) 578 { 579 u_int ffctrl; 580 581 switch (type) { 582 case EVENT_REGISTER: 583 if (kev->fflags & NOTE_TRIGGER) 584 kn->kn_ptr.hookid = 1; 585 586 ffctrl = kev->fflags & NOTE_FFCTRLMASK; 587 kev->fflags &= NOTE_FFLAGSMASK; 588 switch (ffctrl) { 589 case NOTE_FFNOP: 590 break; 591 592 case NOTE_FFAND: 593 kn->kn_fflags &= kev->fflags; 594 break; 595 596 case NOTE_FFOR: 597 kn->kn_fflags |= kev->fflags; 598 break; 599 600 case NOTE_FFCOPY: 601 kn->kn_fflags = kev->fflags; 602 break; 603 604 default: 605 /* XXX Return error? */ 606 break; 607 } 608 /* We just happen to copy this value as well. Undocumented. */ 609 kn->kn_data = kev->data; 610 611 /* 612 * This is not the correct use of EV_CLEAR in an event 613 * modification, it should have been passed as a NOTE instead. 614 * But we need to maintain compatibility with Apple & FreeBSD. 615 * 616 * Note however that EV_CLEAR can still be used when doing 617 * the initial registration of the event and works as expected 618 * (clears the event on reception). 619 */ 620 if (kev->flags & EV_CLEAR) { 621 kn->kn_ptr.hookid = 0; 622 /* 623 * Clearing kn->kn_data is fine, since it gets set 624 * every time anyway. We just shouldn't clear 625 * kn->kn_fflags here, since that would limit the 626 * possible uses of this API. NOTE_FFAND or 627 * NOTE_FFCOPY should be used for explicitly clearing 628 * kn->kn_fflags. 629 */ 630 kn->kn_data = 0; 631 } 632 break; 633 634 case EVENT_PROCESS: 635 *kev = kn->kn_kevent; 636 kev->fflags = kn->kn_fflags; 637 kev->data = kn->kn_data; 638 if (kn->kn_flags & EV_CLEAR) { 639 kn->kn_ptr.hookid = 0; 640 /* kn_data, kn_fflags handled by parent */ 641 } 642 break; 643 644 default: 645 panic("filt_usertouch() - invalid type (%ld)", type); 646 break; 647 } 648 } 649 650 /* 651 * EVFILT_FS 652 */ 653 struct klist fs_klist = SLIST_HEAD_INITIALIZER(&fs_klist); 654 655 static int 656 filt_fsattach(struct knote *kn) 657 { 658 kn->kn_flags |= EV_CLEAR; 659 knote_insert(&fs_klist, kn); 660 661 return (0); 662 } 663 664 static void 665 filt_fsdetach(struct knote *kn) 666 { 667 knote_remove(&fs_klist, kn); 668 } 669 670 static int 671 filt_fs(struct knote *kn, long hint) 672 { 673 kn->kn_fflags |= hint; 674 return (kn->kn_fflags != 0); 675 } 676 677 /* 678 * Initialize a kqueue. 679 * 680 * NOTE: The lwp/proc code initializes a kqueue for select/poll ops. 681 * 682 * MPSAFE 683 */ 684 void 685 kqueue_init(struct kqueue *kq, struct filedesc *fdp) 686 { 687 TAILQ_INIT(&kq->kq_knpend); 688 TAILQ_INIT(&kq->kq_knlist); 689 kq->kq_count = 0; 690 kq->kq_fdp = fdp; 691 SLIST_INIT(&kq->kq_kqinfo.ki_note); 692 } 693 694 /* 695 * Terminate a kqueue. Freeing the actual kq itself is left up to the 696 * caller (it might be embedded in a lwp so we don't do it here). 697 * 698 * The kq's knlist must be completely eradicated so block on any 699 * processing races. 700 */ 701 void 702 kqueue_terminate(struct kqueue *kq) 703 { 704 struct knote *kn; 705 706 lwkt_getpooltoken(kq); 707 while ((kn = TAILQ_FIRST(&kq->kq_knlist)) != NULL) { 708 if (knote_acquire(kn)) 709 knote_detach_and_drop(kn); 710 } 711 lwkt_relpooltoken(kq); 712 713 if (kq->kq_knhash) { 714 hashdestroy(kq->kq_knhash, M_KQUEUE, kq->kq_knhashmask); 715 kq->kq_knhash = NULL; 716 kq->kq_knhashmask = 0; 717 } 718 } 719 720 /* 721 * MPSAFE 722 */ 723 int 724 sys_kqueue(struct kqueue_args *uap) 725 { 726 struct thread *td = curthread; 727 struct kqueue *kq; 728 struct file *fp; 729 int fd, error; 730 731 error = falloc(td->td_lwp, &fp, &fd); 732 if (error) 733 return (error); 734 fp->f_flag = FREAD | FWRITE; 735 fp->f_type = DTYPE_KQUEUE; 736 fp->f_ops = &kqueueops; 737 738 kq = kmalloc(sizeof(struct kqueue), M_KQUEUE, M_WAITOK | M_ZERO); 739 kqueue_init(kq, td->td_proc->p_fd); 740 fp->f_data = kq; 741 742 fsetfd(kq->kq_fdp, fp, fd); 743 uap->sysmsg_result = fd; 744 fdrop(fp); 745 return (error); 746 } 747 748 /* 749 * Copy 'count' items into the destination list pointed to by uap->eventlist. 750 */ 751 static int 752 kevent_copyout(void *arg, struct kevent *kevp, int count, int *res) 753 { 754 struct kevent_copyin_args *kap; 755 int error; 756 757 kap = (struct kevent_copyin_args *)arg; 758 759 error = copyout(kevp, kap->ka->eventlist, count * sizeof(*kevp)); 760 if (error == 0) { 761 kap->ka->eventlist += count; 762 *res += count; 763 } else { 764 *res = -1; 765 } 766 767 return (error); 768 } 769 770 /* 771 * Copy at most 'max' items from the list pointed to by kap->changelist, 772 * return number of items in 'events'. 773 */ 774 static int 775 kevent_copyin(void *arg, struct kevent *kevp, int max, int *events) 776 { 777 struct kevent_copyin_args *kap; 778 int error, count; 779 780 kap = (struct kevent_copyin_args *)arg; 781 782 count = min(kap->ka->nchanges - kap->pchanges, max); 783 error = copyin(kap->ka->changelist, kevp, count * sizeof *kevp); 784 if (error == 0) { 785 kap->ka->changelist += count; 786 kap->pchanges += count; 787 *events = count; 788 } 789 790 return (error); 791 } 792 793 /* 794 * MPSAFE 795 */ 796 int 797 kern_kevent(struct kqueue *kq, int nevents, int *res, void *uap, 798 k_copyin_fn kevent_copyinfn, k_copyout_fn kevent_copyoutfn, 799 struct timespec *tsp_in, int flags) 800 { 801 struct kevent *kevp; 802 struct timespec *tsp, ats; 803 int i, n, total, error, nerrors = 0; 804 int gobbled; 805 int lres; 806 int limit = kq_checkloop; 807 int closedcounter; 808 struct kevent kev[KQ_NEVENTS]; 809 struct knote marker; 810 struct lwkt_token *tok; 811 812 if (tsp_in == NULL || tsp_in->tv_sec || tsp_in->tv_nsec) 813 atomic_set_int(&curthread->td_mpflags, TDF_MP_BATCH_DEMARC); 814 815 tsp = tsp_in; 816 *res = 0; 817 818 closedcounter = kq->kq_fdp->fd_closedcounter; 819 820 for (;;) { 821 n = 0; 822 error = kevent_copyinfn(uap, kev, KQ_NEVENTS, &n); 823 if (error) 824 return error; 825 if (n == 0) 826 break; 827 for (i = 0; i < n; ++i) 828 kev[i].flags &= ~EV_SYSFLAGS; 829 for (i = 0; i < n; ++i) { 830 gobbled = n - i; 831 error = kqueue_register(kq, &kev[i], &gobbled); 832 i += gobbled - 1; 833 kevp = &kev[i]; 834 835 /* 836 * If a registration returns an error we 837 * immediately post the error. The kevent() 838 * call itself will fail with the error if 839 * no space is available for posting. 840 * 841 * Such errors normally bypass the timeout/blocking 842 * code. However, if the copyoutfn function refuses 843 * to post the error (see sys_poll()), then we 844 * ignore it too. 845 */ 846 if (error || (kevp->flags & EV_RECEIPT)) { 847 kevp->flags = EV_ERROR; 848 kevp->data = error; 849 lres = *res; 850 kevent_copyoutfn(uap, kevp, 1, res); 851 if (*res < 0) { 852 return error; 853 } else if (lres != *res) { 854 nevents--; 855 nerrors++; 856 } 857 } 858 } 859 } 860 if (nerrors) 861 return 0; 862 863 /* 864 * Acquire/wait for events - setup timeout 865 */ 866 if (tsp != NULL) { 867 if (tsp->tv_sec || tsp->tv_nsec) { 868 getnanouptime(&ats); 869 timespecadd(tsp, &ats); /* tsp = target time */ 870 } 871 } 872 873 /* 874 * Loop as required. 875 * 876 * Collect as many events as we can. Sleeping on successive 877 * loops is disabled if copyoutfn has incremented (*res). 878 * 879 * The loop stops if an error occurs, all events have been 880 * scanned (the marker has been reached), or fewer than the 881 * maximum number of events is found. 882 * 883 * The copyoutfn function does not have to increment (*res) in 884 * order for the loop to continue. 885 * 886 * NOTE: doselect() usually passes 0x7FFFFFFF for nevents. 887 */ 888 total = 0; 889 error = 0; 890 marker.kn_filter = EVFILT_MARKER; 891 marker.kn_status = KN_PROCESSING; 892 tok = lwkt_token_pool_lookup(kq); 893 lwkt_gettoken(tok); 894 TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker, kn_tqe); 895 lwkt_reltoken(tok); 896 while ((n = nevents - total) > 0) { 897 if (n > KQ_NEVENTS) 898 n = KQ_NEVENTS; 899 900 /* 901 * If no events are pending sleep until timeout (if any) 902 * or an event occurs. 903 * 904 * After the sleep completes the marker is moved to the 905 * end of the list, making any received events available 906 * to our scan. 907 */ 908 if (kq->kq_count == 0 && *res == 0) { 909 int timeout, ustimeout = 0; 910 911 if (tsp == NULL) { 912 timeout = 0; 913 } else if (tsp->tv_sec == 0 && tsp->tv_nsec == 0) { 914 error = EWOULDBLOCK; 915 break; 916 } else { 917 struct timespec atx = *tsp; 918 919 getnanouptime(&ats); 920 timespecsub(&atx, &ats); 921 if (atx.tv_sec < 0) { 922 error = EWOULDBLOCK; 923 break; 924 } else { 925 timeout = atx.tv_sec > 24 * 60 * 60 ? 926 24 * 60 * 60 * hz : 927 tstohz_high(&atx); 928 } 929 if (flags & KEVENT_TIMEOUT_PRECISE && 930 timeout != 0) { 931 if (atx.tv_sec == 0 && 932 atx.tv_nsec < kq_sleep_threshold) { 933 DELAY(atx.tv_nsec / 1000); 934 error = EWOULDBLOCK; 935 break; 936 } else if (atx.tv_sec < 2000) { 937 ustimeout = atx.tv_sec * 938 1000000 + atx.tv_nsec/1000; 939 } else { 940 ustimeout = 2000000000; 941 } 942 } 943 } 944 945 lwkt_gettoken(tok); 946 if (kq->kq_count == 0) { 947 kq->kq_sleep_cnt++; 948 if (__predict_false(kq->kq_sleep_cnt == 0)) { 949 /* 950 * Guard against possible wrapping. And 951 * set it to 2, so that kqueue_wakeup() 952 * can wake everyone up. 953 */ 954 kq->kq_sleep_cnt = 2; 955 } 956 if ((flags & KEVENT_TIMEOUT_PRECISE) && 957 timeout != 0) { 958 error = precise_sleep(kq, PCATCH, 959 "kqread", ustimeout); 960 } else { 961 error = tsleep(kq, PCATCH, "kqread", 962 timeout); 963 } 964 965 /* don't restart after signals... */ 966 if (error == ERESTART) 967 error = EINTR; 968 if (error) { 969 lwkt_reltoken(tok); 970 break; 971 } 972 973 TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe); 974 TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker, 975 kn_tqe); 976 } 977 lwkt_reltoken(tok); 978 } 979 980 /* 981 * Process all received events 982 * Account for all non-spurious events in our total 983 */ 984 i = kqueue_scan(kq, kev, n, &marker, closedcounter); 985 if (i) { 986 lres = *res; 987 error = kevent_copyoutfn(uap, kev, i, res); 988 total += *res - lres; 989 if (error) 990 break; 991 } 992 if (limit && --limit == 0) 993 panic("kqueue: checkloop failed i=%d", i); 994 995 /* 996 * Normally when fewer events are returned than requested 997 * we can stop. However, if only spurious events were 998 * collected the copyout will not bump (*res) and we have 999 * to continue. 1000 */ 1001 if (i < n && *res) 1002 break; 1003 1004 /* 1005 * Deal with an edge case where spurious events can cause 1006 * a loop to occur without moving the marker. This can 1007 * prevent kqueue_scan() from picking up new events which 1008 * race us. We must be sure to move the marker for this 1009 * case. 1010 * 1011 * NOTE: We do not want to move the marker if events 1012 * were scanned because normal kqueue operations 1013 * may reactivate events. Moving the marker in 1014 * that case could result in duplicates for the 1015 * same event. 1016 */ 1017 if (i == 0) { 1018 lwkt_gettoken(tok); 1019 TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe); 1020 TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker, kn_tqe); 1021 lwkt_reltoken(tok); 1022 } 1023 } 1024 lwkt_gettoken(tok); 1025 TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe); 1026 lwkt_reltoken(tok); 1027 1028 /* Timeouts do not return EWOULDBLOCK. */ 1029 if (error == EWOULDBLOCK) 1030 error = 0; 1031 return error; 1032 } 1033 1034 /* 1035 * MPALMOSTSAFE 1036 */ 1037 int 1038 sys_kevent(struct kevent_args *uap) 1039 { 1040 struct thread *td = curthread; 1041 struct timespec ts, *tsp; 1042 struct kqueue *kq; 1043 struct file *fp = NULL; 1044 struct kevent_copyin_args *kap, ka; 1045 int error; 1046 1047 if (uap->timeout) { 1048 error = copyin(uap->timeout, &ts, sizeof(ts)); 1049 if (error) 1050 return (error); 1051 tsp = &ts; 1052 } else { 1053 tsp = NULL; 1054 } 1055 fp = holdfp(td, uap->fd, -1); 1056 if (fp == NULL) 1057 return (EBADF); 1058 if (fp->f_type != DTYPE_KQUEUE) { 1059 fdrop(fp); 1060 return (EBADF); 1061 } 1062 1063 kq = (struct kqueue *)fp->f_data; 1064 1065 kap = &ka; 1066 kap->ka = uap; 1067 kap->pchanges = 0; 1068 1069 error = kern_kevent(kq, uap->nevents, &uap->sysmsg_result, kap, 1070 kevent_copyin, kevent_copyout, tsp, 0); 1071 1072 dropfp(td, uap->fd, fp); 1073 1074 return (error); 1075 } 1076 1077 /* 1078 * Efficiently load multiple file pointers. This significantly reduces 1079 * threaded overhead. When doing simple polling we can depend on the 1080 * per-thread (fd,fp) cache. With more descriptors, we batch. 1081 */ 1082 static 1083 void 1084 floadkevfps(thread_t td, struct filedesc *fdp, struct kevent *kev, 1085 struct file **fp, int climit) 1086 { 1087 struct filterops *fops; 1088 int tdcache; 1089 1090 if (climit <= 2 && td->td_proc && td->td_proc->p_fd == fdp) { 1091 tdcache = 1; 1092 } else { 1093 tdcache = 0; 1094 spin_lock_shared(&fdp->fd_spin); 1095 } 1096 1097 while (climit) { 1098 *fp = NULL; 1099 if (kev->filter < 0 && 1100 kev->filter + EVFILT_SYSCOUNT >= 0) { 1101 fops = sysfilt_ops[~kev->filter]; 1102 if (fops->f_flags & FILTEROP_ISFD) { 1103 if (tdcache) { 1104 *fp = holdfp(td, kev->ident, -1); 1105 } else { 1106 *fp = holdfp_fdp_locked(fdp, 1107 kev->ident, -1); 1108 } 1109 } 1110 } 1111 --climit; 1112 ++fp; 1113 ++kev; 1114 } 1115 if (tdcache == 0) 1116 spin_unlock_shared(&fdp->fd_spin); 1117 } 1118 1119 /* 1120 * Register up to *countp kev's. Always registers at least 1. 1121 * 1122 * The number registered is returned in *countp. 1123 * 1124 * If an error occurs or a kev is flagged EV_RECEIPT, it is 1125 * processed and included in *countp, and processing then 1126 * stops. 1127 */ 1128 int 1129 kqueue_register(struct kqueue *kq, struct kevent *kev, int *countp) 1130 { 1131 struct filedesc *fdp = kq->kq_fdp; 1132 struct klist *list = NULL; 1133 struct filterops *fops; 1134 struct file *fp[KQ_NEVENTS]; 1135 struct knote *kn = NULL; 1136 struct thread *td; 1137 int error; 1138 int count; 1139 int climit; 1140 int closedcounter; 1141 struct knote_cache_list *cache_list; 1142 1143 td = curthread; 1144 climit = *countp; 1145 if (climit > KQ_NEVENTS) 1146 climit = KQ_NEVENTS; 1147 closedcounter = fdp->fd_closedcounter; 1148 floadkevfps(td, fdp, kev, fp, climit); 1149 1150 lwkt_getpooltoken(kq); 1151 count = 0; 1152 1153 /* 1154 * To avoid races, only one thread can register events on this 1155 * kqueue at a time. 1156 */ 1157 while (__predict_false(kq->kq_regtd != NULL && kq->kq_regtd != td)) { 1158 kq->kq_state |= KQ_REGWAIT; 1159 tsleep(&kq->kq_regtd, 0, "kqreg", 0); 1160 } 1161 if (__predict_false(kq->kq_regtd != NULL)) { 1162 /* Recursive calling of kqueue_register() */ 1163 td = NULL; 1164 } else { 1165 /* Owner of the kq_regtd, i.e. td != NULL */ 1166 kq->kq_regtd = td; 1167 } 1168 1169 loop: 1170 if (kev->filter < 0) { 1171 if (kev->filter + EVFILT_SYSCOUNT < 0) { 1172 error = EINVAL; 1173 ++count; 1174 goto done; 1175 } 1176 fops = sysfilt_ops[~kev->filter]; /* to 0-base index */ 1177 } else { 1178 /* 1179 * XXX 1180 * filter attach routine is responsible for insuring that 1181 * the identifier can be attached to it. 1182 */ 1183 error = EINVAL; 1184 ++count; 1185 goto done; 1186 } 1187 1188 if (fops->f_flags & FILTEROP_ISFD) { 1189 /* validate descriptor */ 1190 if (fp[count] == NULL) { 1191 error = EBADF; 1192 ++count; 1193 goto done; 1194 } 1195 } 1196 1197 cache_list = &knote_cache_lists[mycpuid]; 1198 if (SLIST_EMPTY(&cache_list->knote_cache)) { 1199 struct knote *new_kn; 1200 1201 new_kn = knote_alloc(); 1202 crit_enter(); 1203 SLIST_INSERT_HEAD(&cache_list->knote_cache, new_kn, kn_link); 1204 cache_list->knote_cache_cnt++; 1205 crit_exit(); 1206 } 1207 1208 if (fp[count] != NULL) { 1209 list = &fp[count]->f_klist; 1210 } else if (kq->kq_knhashmask) { 1211 list = &kq->kq_knhash[ 1212 KN_HASH((u_long)kev->ident, kq->kq_knhashmask)]; 1213 } 1214 if (list != NULL) { 1215 lwkt_getpooltoken(list); 1216 again: 1217 SLIST_FOREACH(kn, list, kn_link) { 1218 if (kn->kn_kq == kq && 1219 kn->kn_filter == kev->filter && 1220 kn->kn_id == kev->ident) { 1221 if (knote_acquire(kn) == 0) 1222 goto again; 1223 break; 1224 } 1225 } 1226 lwkt_relpooltoken(list); 1227 } 1228 1229 /* 1230 * NOTE: At this point if kn is non-NULL we will have acquired 1231 * it and set KN_PROCESSING. 1232 */ 1233 if (kn == NULL && ((kev->flags & EV_ADD) == 0)) { 1234 error = ENOENT; 1235 ++count; 1236 goto done; 1237 } 1238 1239 /* 1240 * kn now contains the matching knote, or NULL if no match 1241 */ 1242 if (kev->flags & EV_ADD) { 1243 if (kn == NULL) { 1244 crit_enter(); 1245 kn = SLIST_FIRST(&cache_list->knote_cache); 1246 if (kn == NULL) { 1247 crit_exit(); 1248 kn = knote_alloc(); 1249 } else { 1250 SLIST_REMOVE_HEAD(&cache_list->knote_cache, 1251 kn_link); 1252 cache_list->knote_cache_cnt--; 1253 crit_exit(); 1254 } 1255 kn->kn_fp = fp[count]; 1256 kn->kn_kq = kq; 1257 kn->kn_fop = fops; 1258 1259 /* 1260 * apply reference count to knote structure, and 1261 * do not release it at the end of this routine. 1262 */ 1263 fp[count] = NULL; /* safety */ 1264 1265 kn->kn_sfflags = kev->fflags; 1266 kn->kn_sdata = kev->data; 1267 kev->fflags = 0; 1268 kev->data = 0; 1269 kn->kn_kevent = *kev; 1270 1271 /* 1272 * KN_PROCESSING prevents the knote from getting 1273 * ripped out from under us while we are trying 1274 * to attach it, in case the attach blocks. 1275 */ 1276 kn->kn_status = KN_PROCESSING; 1277 knote_attach(kn); 1278 if ((error = filter_attach(kn)) != 0) { 1279 kn->kn_status |= KN_DELETING | KN_REPROCESS; 1280 knote_drop(kn); 1281 ++count; 1282 goto done; 1283 } 1284 1285 /* 1286 * Interlock against close races which either tried 1287 * to remove our knote while we were blocked or missed 1288 * it entirely prior to our attachment. We do not 1289 * want to end up with a knote on a closed descriptor. 1290 */ 1291 if ((fops->f_flags & FILTEROP_ISFD) && 1292 checkfdclosed(curthread, fdp, kev->ident, kn->kn_fp, 1293 closedcounter)) { 1294 kn->kn_status |= KN_DELETING | KN_REPROCESS; 1295 } 1296 } else { 1297 /* 1298 * The user may change some filter values after the 1299 * initial EV_ADD, but doing so will not reset any 1300 * filter which have already been triggered. 1301 */ 1302 KKASSERT(kn->kn_status & KN_PROCESSING); 1303 if (fops == &user_filtops) { 1304 filt_usertouch(kn, kev, EVENT_REGISTER); 1305 } else { 1306 kn->kn_sfflags = kev->fflags; 1307 kn->kn_sdata = kev->data; 1308 kn->kn_kevent.udata = kev->udata; 1309 } 1310 } 1311 1312 /* 1313 * Execute the filter event to immediately activate the 1314 * knote if necessary. If reprocessing events are pending 1315 * due to blocking above we do not run the filter here 1316 * but instead let knote_release() do it. Otherwise we 1317 * might run the filter on a deleted event. 1318 */ 1319 if ((kn->kn_status & KN_REPROCESS) == 0) { 1320 if (filter_event(kn, 0)) 1321 KNOTE_ACTIVATE(kn); 1322 } 1323 } else if (kev->flags & EV_DELETE) { 1324 /* 1325 * Delete the existing knote 1326 */ 1327 knote_detach_and_drop(kn); 1328 error = 0; 1329 ++count; 1330 goto done; 1331 } else { 1332 /* 1333 * Modify an existing event. 1334 * 1335 * The user may change some filter values after the 1336 * initial EV_ADD, but doing so will not reset any 1337 * filter which have already been triggered. 1338 */ 1339 KKASSERT(kn->kn_status & KN_PROCESSING); 1340 if (fops == &user_filtops) { 1341 filt_usertouch(kn, kev, EVENT_REGISTER); 1342 } else { 1343 kn->kn_sfflags = kev->fflags; 1344 kn->kn_sdata = kev->data; 1345 kn->kn_kevent.udata = kev->udata; 1346 } 1347 1348 /* 1349 * Execute the filter event to immediately activate the 1350 * knote if necessary. If reprocessing events are pending 1351 * due to blocking above we do not run the filter here 1352 * but instead let knote_release() do it. Otherwise we 1353 * might run the filter on a deleted event. 1354 */ 1355 if ((kn->kn_status & KN_REPROCESS) == 0) { 1356 if (filter_event(kn, 0)) 1357 KNOTE_ACTIVATE(kn); 1358 } 1359 } 1360 1361 /* 1362 * Disablement does not deactivate a knote here. 1363 */ 1364 if ((kev->flags & EV_DISABLE) && 1365 ((kn->kn_status & KN_DISABLED) == 0)) { 1366 kn->kn_status |= KN_DISABLED; 1367 } 1368 1369 /* 1370 * Re-enablement may have to immediately enqueue an active knote. 1371 */ 1372 if ((kev->flags & EV_ENABLE) && (kn->kn_status & KN_DISABLED)) { 1373 kn->kn_status &= ~KN_DISABLED; 1374 if ((kn->kn_status & KN_ACTIVE) && 1375 ((kn->kn_status & KN_QUEUED) == 0)) { 1376 knote_enqueue(kn); 1377 } 1378 } 1379 1380 /* 1381 * Handle any required reprocessing 1382 */ 1383 knote_release(kn); 1384 /* kn may be invalid now */ 1385 1386 /* 1387 * Loop control. We stop on errors (above), and also stop after 1388 * processing EV_RECEIPT, so the caller can process it. 1389 */ 1390 ++count; 1391 if (kev->flags & EV_RECEIPT) { 1392 error = 0; 1393 goto done; 1394 } 1395 ++kev; 1396 if (count < climit) { 1397 if (fp[count-1]) /* drop unprocessed fp */ 1398 fdrop(fp[count-1]); 1399 goto loop; 1400 } 1401 1402 /* 1403 * Cleanup 1404 */ 1405 done: 1406 if (td != NULL) { /* Owner of the kq_regtd */ 1407 kq->kq_regtd = NULL; 1408 if (__predict_false(kq->kq_state & KQ_REGWAIT)) { 1409 kq->kq_state &= ~KQ_REGWAIT; 1410 wakeup(&kq->kq_regtd); 1411 } 1412 } 1413 lwkt_relpooltoken(kq); 1414 1415 /* 1416 * Drop unprocessed file pointers 1417 */ 1418 *countp = count; 1419 if (count && fp[count-1]) 1420 fdrop(fp[count-1]); 1421 while (count < climit) { 1422 if (fp[count]) 1423 fdrop(fp[count]); 1424 ++count; 1425 } 1426 return (error); 1427 } 1428 1429 /* 1430 * Scan the kqueue, return the number of active events placed in kevp up 1431 * to count. 1432 * 1433 * Continuous mode events may get recycled, do not continue scanning past 1434 * marker unless no events have been collected. 1435 */ 1436 static int 1437 kqueue_scan(struct kqueue *kq, struct kevent *kevp, int count, 1438 struct knote *marker, int closedcounter) 1439 { 1440 struct knote *kn, local_marker; 1441 thread_t td = curthread; 1442 int total; 1443 1444 total = 0; 1445 local_marker.kn_filter = EVFILT_MARKER; 1446 local_marker.kn_status = KN_PROCESSING; 1447 1448 lwkt_getpooltoken(kq); 1449 1450 /* 1451 * Collect events. 1452 */ 1453 TAILQ_INSERT_HEAD(&kq->kq_knpend, &local_marker, kn_tqe); 1454 while (count) { 1455 kn = TAILQ_NEXT(&local_marker, kn_tqe); 1456 if (kn->kn_filter == EVFILT_MARKER) { 1457 /* Marker reached, we are done */ 1458 if (kn == marker) 1459 break; 1460 1461 /* Move local marker past some other threads marker */ 1462 kn = TAILQ_NEXT(kn, kn_tqe); 1463 TAILQ_REMOVE(&kq->kq_knpend, &local_marker, kn_tqe); 1464 TAILQ_INSERT_BEFORE(kn, &local_marker, kn_tqe); 1465 continue; 1466 } 1467 1468 /* 1469 * We can't skip a knote undergoing processing, otherwise 1470 * we risk not returning it when the user process expects 1471 * it should be returned. Sleep and retry. 1472 */ 1473 if (knote_acquire(kn) == 0) 1474 continue; 1475 1476 /* 1477 * Remove the event for processing. 1478 * 1479 * WARNING! We must leave KN_QUEUED set to prevent the 1480 * event from being KNOTE_ACTIVATE()d while 1481 * the queue state is in limbo, in case we 1482 * block. 1483 */ 1484 TAILQ_REMOVE(&kq->kq_knpend, kn, kn_tqe); 1485 kq->kq_count--; 1486 1487 /* 1488 * We have to deal with an extremely important race against 1489 * file descriptor close()s here. The file descriptor can 1490 * disappear MPSAFE, and there is a small window of 1491 * opportunity between that and the call to knote_fdclose(). 1492 * 1493 * If we hit that window here while doselect or dopoll is 1494 * trying to delete a spurious event they will not be able 1495 * to match up the event against a knote and will go haywire. 1496 */ 1497 if ((kn->kn_fop->f_flags & FILTEROP_ISFD) && 1498 checkfdclosed(td, kq->kq_fdp, kn->kn_kevent.ident, 1499 kn->kn_fp, closedcounter)) { 1500 kn->kn_status |= KN_DELETING | KN_REPROCESS; 1501 } 1502 1503 if (kn->kn_status & KN_DISABLED) { 1504 /* 1505 * If disabled we ensure the event is not queued 1506 * but leave its active bit set. On re-enablement 1507 * the event may be immediately triggered. 1508 */ 1509 kn->kn_status &= ~KN_QUEUED; 1510 } else if ((kn->kn_flags & EV_ONESHOT) == 0 && 1511 (kn->kn_status & KN_DELETING) == 0 && 1512 filter_event(kn, 0) == 0) { 1513 /* 1514 * If not running in one-shot mode and the event 1515 * is no longer present we ensure it is removed 1516 * from the queue and ignore it. 1517 */ 1518 kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE); 1519 } else { 1520 /* 1521 * Post the event 1522 */ 1523 if (kn->kn_fop == &user_filtops) 1524 filt_usertouch(kn, kevp, EVENT_PROCESS); 1525 else 1526 *kevp = kn->kn_kevent; 1527 ++kevp; 1528 ++total; 1529 --count; 1530 1531 if (kn->kn_flags & EV_ONESHOT) { 1532 kn->kn_status &= ~KN_QUEUED; 1533 kn->kn_status |= KN_DELETING | KN_REPROCESS; 1534 } else { 1535 if (kn->kn_flags & (EV_CLEAR | EV_DISPATCH)) { 1536 if (kn->kn_flags & EV_CLEAR) { 1537 kn->kn_data = 0; 1538 kn->kn_fflags = 0; 1539 } 1540 if (kn->kn_flags & EV_DISPATCH) { 1541 kn->kn_status |= KN_DISABLED; 1542 } 1543 kn->kn_status &= ~(KN_QUEUED | 1544 KN_ACTIVE); 1545 } else { 1546 TAILQ_INSERT_TAIL(&kq->kq_knpend, kn, kn_tqe); 1547 kq->kq_count++; 1548 } 1549 } 1550 } 1551 1552 /* 1553 * Handle any post-processing states 1554 */ 1555 knote_release(kn); 1556 } 1557 TAILQ_REMOVE(&kq->kq_knpend, &local_marker, kn_tqe); 1558 1559 lwkt_relpooltoken(kq); 1560 return (total); 1561 } 1562 1563 /* 1564 * XXX 1565 * This could be expanded to call kqueue_scan, if desired. 1566 * 1567 * MPSAFE 1568 */ 1569 static int 1570 kqueue_read(struct file *fp, struct uio *uio, struct ucred *cred, int flags) 1571 { 1572 return (ENXIO); 1573 } 1574 1575 /* 1576 * MPSAFE 1577 */ 1578 static int 1579 kqueue_write(struct file *fp, struct uio *uio, struct ucred *cred, int flags) 1580 { 1581 return (ENXIO); 1582 } 1583 1584 /* 1585 * MPALMOSTSAFE 1586 */ 1587 static int 1588 kqueue_ioctl(struct file *fp, u_long com, caddr_t data, 1589 struct ucred *cred, struct sysmsg *msg) 1590 { 1591 struct kqueue *kq; 1592 int error; 1593 1594 kq = (struct kqueue *)fp->f_data; 1595 lwkt_getpooltoken(kq); 1596 switch(com) { 1597 case FIOASYNC: 1598 if (*(int *)data) 1599 kq->kq_state |= KQ_ASYNC; 1600 else 1601 kq->kq_state &= ~KQ_ASYNC; 1602 error = 0; 1603 break; 1604 case FIOSETOWN: 1605 error = fsetown(*(int *)data, &kq->kq_sigio); 1606 break; 1607 default: 1608 error = ENOTTY; 1609 break; 1610 } 1611 lwkt_relpooltoken(kq); 1612 return (error); 1613 } 1614 1615 /* 1616 * MPSAFE 1617 */ 1618 static int 1619 kqueue_stat(struct file *fp, struct stat *st, struct ucred *cred) 1620 { 1621 struct kqueue *kq = (struct kqueue *)fp->f_data; 1622 1623 bzero((void *)st, sizeof(*st)); 1624 st->st_size = kq->kq_count; 1625 st->st_blksize = sizeof(struct kevent); 1626 st->st_mode = S_IFIFO; 1627 return (0); 1628 } 1629 1630 /* 1631 * MPSAFE 1632 */ 1633 static int 1634 kqueue_close(struct file *fp) 1635 { 1636 struct kqueue *kq = (struct kqueue *)fp->f_data; 1637 1638 kqueue_terminate(kq); 1639 1640 fp->f_data = NULL; 1641 funsetown(&kq->kq_sigio); 1642 1643 kfree(kq, M_KQUEUE); 1644 return (0); 1645 } 1646 1647 static void 1648 kqueue_wakeup(struct kqueue *kq) 1649 { 1650 if (kq->kq_sleep_cnt) { 1651 u_int sleep_cnt = kq->kq_sleep_cnt; 1652 1653 kq->kq_sleep_cnt = 0; 1654 if (sleep_cnt == 1) 1655 wakeup_one(kq); 1656 else 1657 wakeup(kq); 1658 } 1659 KNOTE(&kq->kq_kqinfo.ki_note, 0); 1660 } 1661 1662 /* 1663 * Calls filterops f_attach function, acquiring mplock if filter is not 1664 * marked as FILTEROP_MPSAFE. 1665 * 1666 * Caller must be holding the related kq token 1667 */ 1668 static int 1669 filter_attach(struct knote *kn) 1670 { 1671 int ret; 1672 1673 if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) { 1674 ret = kn->kn_fop->f_attach(kn); 1675 } else { 1676 get_mplock(); 1677 ret = kn->kn_fop->f_attach(kn); 1678 rel_mplock(); 1679 } 1680 return (ret); 1681 } 1682 1683 /* 1684 * Detach the knote and drop it, destroying the knote. 1685 * 1686 * Calls filterops f_detach function, acquiring mplock if filter is not 1687 * marked as FILTEROP_MPSAFE. 1688 * 1689 * Caller must be holding the related kq token 1690 */ 1691 static void 1692 knote_detach_and_drop(struct knote *kn) 1693 { 1694 kn->kn_status |= KN_DELETING | KN_REPROCESS; 1695 if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) { 1696 kn->kn_fop->f_detach(kn); 1697 } else { 1698 get_mplock(); 1699 kn->kn_fop->f_detach(kn); 1700 rel_mplock(); 1701 } 1702 knote_drop(kn); 1703 } 1704 1705 /* 1706 * Calls filterops f_event function, acquiring mplock if filter is not 1707 * marked as FILTEROP_MPSAFE. 1708 * 1709 * If the knote is in the middle of being created or deleted we cannot 1710 * safely call the filter op. 1711 * 1712 * Caller must be holding the related kq token 1713 */ 1714 static int 1715 filter_event(struct knote *kn, long hint) 1716 { 1717 int ret; 1718 1719 if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) { 1720 ret = kn->kn_fop->f_event(kn, hint); 1721 } else { 1722 get_mplock(); 1723 ret = kn->kn_fop->f_event(kn, hint); 1724 rel_mplock(); 1725 } 1726 return (ret); 1727 } 1728 1729 /* 1730 * Walk down a list of knotes, activating them if their event has triggered. 1731 * 1732 * If we encounter any knotes which are undergoing processing we just mark 1733 * them for reprocessing and do not try to [re]activate the knote. However, 1734 * if a hint is being passed we have to wait and that makes things a bit 1735 * sticky. 1736 */ 1737 void 1738 knote(struct klist *list, long hint) 1739 { 1740 struct kqueue *kq; 1741 struct knote *kn; 1742 struct knote *kntmp; 1743 1744 lwkt_getpooltoken(list); 1745 restart: 1746 SLIST_FOREACH(kn, list, kn_next) { 1747 kq = kn->kn_kq; 1748 lwkt_getpooltoken(kq); 1749 1750 /* temporary verification hack */ 1751 SLIST_FOREACH(kntmp, list, kn_next) { 1752 if (kn == kntmp) 1753 break; 1754 } 1755 if (kn != kntmp || kn->kn_kq != kq) { 1756 lwkt_relpooltoken(kq); 1757 goto restart; 1758 } 1759 1760 if (kn->kn_status & KN_PROCESSING) { 1761 /* 1762 * Someone else is processing the knote, ask the 1763 * other thread to reprocess it and don't mess 1764 * with it otherwise. 1765 */ 1766 if (hint == 0) { 1767 kn->kn_status |= KN_REPROCESS; 1768 lwkt_relpooltoken(kq); 1769 continue; 1770 } 1771 1772 /* 1773 * If the hint is non-zero we have to wait or risk 1774 * losing the state the caller is trying to update. 1775 * 1776 * XXX This is a real problem, certain process 1777 * and signal filters will bump kn_data for 1778 * already-processed notes more than once if 1779 * we restart the list scan. FIXME. 1780 */ 1781 kn->kn_status |= KN_WAITING | KN_REPROCESS; 1782 tsleep(kn, 0, "knotec", hz); 1783 lwkt_relpooltoken(kq); 1784 goto restart; 1785 } 1786 1787 /* 1788 * Become the reprocessing master ourselves. 1789 * 1790 * If hint is non-zero running the event is mandatory 1791 * when not deleting so do it whether reprocessing is 1792 * set or not. 1793 */ 1794 kn->kn_status |= KN_PROCESSING; 1795 if ((kn->kn_status & KN_DELETING) == 0) { 1796 if (filter_event(kn, hint)) 1797 KNOTE_ACTIVATE(kn); 1798 } 1799 if (knote_release(kn)) { 1800 lwkt_relpooltoken(kq); 1801 goto restart; 1802 } 1803 lwkt_relpooltoken(kq); 1804 } 1805 lwkt_relpooltoken(list); 1806 } 1807 1808 /* 1809 * Insert knote at head of klist. 1810 * 1811 * This function may only be called via a filter function and thus 1812 * kq_token should already be held and marked for processing. 1813 */ 1814 void 1815 knote_insert(struct klist *klist, struct knote *kn) 1816 { 1817 lwkt_getpooltoken(klist); 1818 KKASSERT(kn->kn_status & KN_PROCESSING); 1819 SLIST_INSERT_HEAD(klist, kn, kn_next); 1820 lwkt_relpooltoken(klist); 1821 } 1822 1823 /* 1824 * Remove knote from a klist 1825 * 1826 * This function may only be called via a filter function and thus 1827 * kq_token should already be held and marked for processing. 1828 */ 1829 void 1830 knote_remove(struct klist *klist, struct knote *kn) 1831 { 1832 lwkt_getpooltoken(klist); 1833 KKASSERT(kn->kn_status & KN_PROCESSING); 1834 SLIST_REMOVE(klist, kn, knote, kn_next); 1835 lwkt_relpooltoken(klist); 1836 } 1837 1838 void 1839 knote_assume_knotes(struct kqinfo *src, struct kqinfo *dst, 1840 struct filterops *ops, void *hook) 1841 { 1842 struct kqueue *kq; 1843 struct knote *kn; 1844 1845 lwkt_getpooltoken(&src->ki_note); 1846 lwkt_getpooltoken(&dst->ki_note); 1847 while ((kn = SLIST_FIRST(&src->ki_note)) != NULL) { 1848 kq = kn->kn_kq; 1849 lwkt_getpooltoken(kq); 1850 if (SLIST_FIRST(&src->ki_note) != kn || kn->kn_kq != kq) { 1851 lwkt_relpooltoken(kq); 1852 continue; 1853 } 1854 if (knote_acquire(kn)) { 1855 knote_remove(&src->ki_note, kn); 1856 kn->kn_fop = ops; 1857 kn->kn_hook = hook; 1858 knote_insert(&dst->ki_note, kn); 1859 knote_release(kn); 1860 /* kn may be invalid now */ 1861 } 1862 lwkt_relpooltoken(kq); 1863 } 1864 lwkt_relpooltoken(&dst->ki_note); 1865 lwkt_relpooltoken(&src->ki_note); 1866 } 1867 1868 /* 1869 * Remove all knotes referencing a specified fd 1870 */ 1871 void 1872 knote_fdclose(struct file *fp, struct filedesc *fdp, int fd) 1873 { 1874 struct kqueue *kq; 1875 struct knote *kn; 1876 struct knote *kntmp; 1877 1878 lwkt_getpooltoken(&fp->f_klist); 1879 restart: 1880 SLIST_FOREACH(kn, &fp->f_klist, kn_link) { 1881 if (kn->kn_kq->kq_fdp == fdp && kn->kn_id == fd) { 1882 kq = kn->kn_kq; 1883 lwkt_getpooltoken(kq); 1884 1885 /* temporary verification hack */ 1886 SLIST_FOREACH(kntmp, &fp->f_klist, kn_link) { 1887 if (kn == kntmp) 1888 break; 1889 } 1890 if (kn != kntmp || kn->kn_kq->kq_fdp != fdp || 1891 kn->kn_id != fd || kn->kn_kq != kq) { 1892 lwkt_relpooltoken(kq); 1893 goto restart; 1894 } 1895 if (knote_acquire(kn)) 1896 knote_detach_and_drop(kn); 1897 lwkt_relpooltoken(kq); 1898 goto restart; 1899 } 1900 } 1901 lwkt_relpooltoken(&fp->f_klist); 1902 } 1903 1904 /* 1905 * Low level attach function. 1906 * 1907 * The knote should already be marked for processing. 1908 * Caller must hold the related kq token. 1909 */ 1910 static void 1911 knote_attach(struct knote *kn) 1912 { 1913 struct klist *list; 1914 struct kqueue *kq = kn->kn_kq; 1915 1916 if (kn->kn_fop->f_flags & FILTEROP_ISFD) { 1917 KKASSERT(kn->kn_fp); 1918 list = &kn->kn_fp->f_klist; 1919 } else { 1920 if (kq->kq_knhashmask == 0) 1921 kq->kq_knhash = hashinit(KN_HASHSIZE, M_KQUEUE, 1922 &kq->kq_knhashmask); 1923 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)]; 1924 } 1925 lwkt_getpooltoken(list); 1926 SLIST_INSERT_HEAD(list, kn, kn_link); 1927 lwkt_relpooltoken(list); 1928 TAILQ_INSERT_HEAD(&kq->kq_knlist, kn, kn_kqlink); 1929 } 1930 1931 /* 1932 * Low level drop function. 1933 * 1934 * The knote should already be marked for processing. 1935 * Caller must hold the related kq token. 1936 */ 1937 static void 1938 knote_drop(struct knote *kn) 1939 { 1940 struct kqueue *kq; 1941 struct klist *list; 1942 1943 kq = kn->kn_kq; 1944 1945 if (kn->kn_fop->f_flags & FILTEROP_ISFD) 1946 list = &kn->kn_fp->f_klist; 1947 else 1948 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)]; 1949 1950 lwkt_getpooltoken(list); 1951 SLIST_REMOVE(list, kn, knote, kn_link); 1952 lwkt_relpooltoken(list); 1953 TAILQ_REMOVE(&kq->kq_knlist, kn, kn_kqlink); 1954 if (kn->kn_status & KN_QUEUED) 1955 knote_dequeue(kn); 1956 if (kn->kn_fop->f_flags & FILTEROP_ISFD) { 1957 fdrop(kn->kn_fp); 1958 kn->kn_fp = NULL; 1959 } 1960 knote_free(kn); 1961 } 1962 1963 /* 1964 * Low level enqueue function. 1965 * 1966 * The knote should already be marked for processing. 1967 * Caller must be holding the kq token 1968 */ 1969 static void 1970 knote_enqueue(struct knote *kn) 1971 { 1972 struct kqueue *kq = kn->kn_kq; 1973 1974 KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued")); 1975 TAILQ_INSERT_TAIL(&kq->kq_knpend, kn, kn_tqe); 1976 kn->kn_status |= KN_QUEUED; 1977 ++kq->kq_count; 1978 1979 /* 1980 * Send SIGIO on request (typically set up as a mailbox signal) 1981 */ 1982 if (kq->kq_sigio && (kq->kq_state & KQ_ASYNC) && kq->kq_count == 1) 1983 pgsigio(kq->kq_sigio, SIGIO, 0); 1984 1985 kqueue_wakeup(kq); 1986 } 1987 1988 /* 1989 * Low level dequeue function. 1990 * 1991 * The knote should already be marked for processing. 1992 * Caller must be holding the kq token 1993 */ 1994 static void 1995 knote_dequeue(struct knote *kn) 1996 { 1997 struct kqueue *kq = kn->kn_kq; 1998 1999 KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued")); 2000 TAILQ_REMOVE(&kq->kq_knpend, kn, kn_tqe); 2001 kn->kn_status &= ~KN_QUEUED; 2002 kq->kq_count--; 2003 } 2004 2005 static struct knote * 2006 knote_alloc(void) 2007 { 2008 return kmalloc(sizeof(struct knote), M_KQUEUE, M_WAITOK); 2009 } 2010 2011 static void 2012 knote_free(struct knote *kn) 2013 { 2014 struct knote_cache_list *cache_list; 2015 2016 cache_list = &knote_cache_lists[mycpuid]; 2017 if (cache_list->knote_cache_cnt < KNOTE_CACHE_MAX) { 2018 crit_enter(); 2019 SLIST_INSERT_HEAD(&cache_list->knote_cache, kn, kn_link); 2020 cache_list->knote_cache_cnt++; 2021 crit_exit(); 2022 return; 2023 } 2024 kfree(kn, M_KQUEUE); 2025 } 2026 2027 struct sleepinfo { 2028 void *ident; 2029 int timedout; 2030 }; 2031 2032 static void 2033 precise_sleep_intr(systimer_t info, int in_ipi, struct intrframe *frame) 2034 { 2035 struct sleepinfo *si; 2036 2037 si = info->data; 2038 si->timedout = 1; 2039 wakeup(si->ident); 2040 } 2041 2042 static int 2043 precise_sleep(void *ident, int flags, const char *wmesg, int us) 2044 { 2045 struct systimer info; 2046 struct sleepinfo si = { 2047 .ident = ident, 2048 .timedout = 0, 2049 }; 2050 int r; 2051 2052 tsleep_interlock(ident, flags); 2053 systimer_init_oneshot(&info, precise_sleep_intr, &si, 2054 us == 0 ? 1 : us); 2055 r = tsleep(ident, flags | PINTERLOCKED, wmesg, 0); 2056 systimer_del(&info); 2057 if (si.timedout) 2058 r = EWOULDBLOCK; 2059 2060 return r; 2061 } 2062