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/sysmsg.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 static MALLOC_DEFINE(M_KQUEUE, "kqueue", "memory for kqueue system"); 63 64 struct kevent_copyin_args { 65 const struct kevent_args *ka; 66 struct kevent *eventlist; 67 const struct kevent *changelist; 68 int pchanges; 69 }; 70 71 #define KNOTE_CACHE_MAX 64 72 73 struct knote_cache_list { 74 struct klist knote_cache; 75 int knote_cache_cnt; 76 } __cachealign; 77 78 static int kqueue_scan(struct kqueue *kq, struct kevent *kevp, int count, 79 struct knote *marker, int closedcounter, int scan_flags); 80 static int kqueue_read(struct file *fp, struct uio *uio, 81 struct ucred *cred, int flags); 82 static int kqueue_write(struct file *fp, struct uio *uio, 83 struct ucred *cred, int flags); 84 static int kqueue_ioctl(struct file *fp, u_long com, caddr_t data, 85 struct ucred *cred, struct sysmsg *msg); 86 static int kqueue_kqfilter(struct file *fp, struct knote *kn); 87 static int kqueue_stat(struct file *fp, struct stat *st, 88 struct ucred *cred); 89 static int kqueue_close(struct file *fp); 90 static void kqueue_wakeup(struct kqueue *kq); 91 static int filter_attach(struct knote *kn); 92 static int filter_event(struct knote *kn, long hint); 93 94 /* 95 * MPSAFE 96 */ 97 static struct fileops kqueueops = { 98 .fo_read = kqueue_read, 99 .fo_write = kqueue_write, 100 .fo_ioctl = kqueue_ioctl, 101 .fo_kqfilter = kqueue_kqfilter, 102 .fo_stat = kqueue_stat, 103 .fo_close = kqueue_close, 104 .fo_shutdown = nofo_shutdown 105 }; 106 107 static void knote_attach(struct knote *kn); 108 static void knote_drop(struct knote *kn); 109 static void knote_detach_and_drop(struct knote *kn); 110 static void knote_enqueue(struct knote *kn); 111 static void knote_dequeue(struct knote *kn); 112 static struct knote *knote_alloc(void); 113 static void knote_free(struct knote *kn); 114 115 static void precise_sleep_intr(systimer_t info, int in_ipi, 116 struct intrframe *frame); 117 static int precise_sleep(void *ident, int flags, const char *wmesg, 118 int us); 119 120 static void filt_kqdetach(struct knote *kn); 121 static int filt_kqueue(struct knote *kn, long hint); 122 static int filt_procattach(struct knote *kn); 123 static void filt_procdetach(struct knote *kn); 124 static int filt_proc(struct knote *kn, long hint); 125 static int filt_fileattach(struct knote *kn); 126 static void filt_timerexpire(void *knx); 127 static int filt_timerattach(struct knote *kn); 128 static void filt_timerdetach(struct knote *kn); 129 static int filt_timer(struct knote *kn, long hint); 130 static int filt_userattach(struct knote *kn); 131 static void filt_userdetach(struct knote *kn); 132 static int filt_user(struct knote *kn, long hint); 133 static void filt_usertouch(struct knote *kn, struct kevent *kev, 134 u_long type); 135 static int filt_fsattach(struct knote *kn); 136 static void filt_fsdetach(struct knote *kn); 137 static int filt_fs(struct knote *kn, long hint); 138 139 static struct filterops file_filtops = 140 { FILTEROP_ISFD | FILTEROP_MPSAFE, filt_fileattach, NULL, NULL }; 141 static struct filterops kqread_filtops = 142 { FILTEROP_ISFD | FILTEROP_MPSAFE, NULL, filt_kqdetach, filt_kqueue }; 143 static struct filterops proc_filtops = 144 { FILTEROP_MPSAFE, filt_procattach, filt_procdetach, filt_proc }; 145 static struct filterops timer_filtops = 146 { FILTEROP_MPSAFE, filt_timerattach, filt_timerdetach, filt_timer }; 147 static struct filterops user_filtops = 148 { FILTEROP_MPSAFE, filt_userattach, filt_userdetach, filt_user }; 149 static struct filterops fs_filtops = 150 { FILTEROP_MPSAFE, filt_fsattach, filt_fsdetach, filt_fs }; 151 152 static int kq_ncallouts = 0; 153 static int kq_calloutmax = 65536; 154 SYSCTL_INT(_kern, OID_AUTO, kq_calloutmax, CTLFLAG_RW, 155 &kq_calloutmax, 0, "Maximum number of callouts allocated for kqueue"); 156 static int kq_checkloop = 1000000; 157 SYSCTL_INT(_kern, OID_AUTO, kq_checkloop, CTLFLAG_RW, 158 &kq_checkloop, 0, "Maximum number of loops for kqueue scan"); 159 static int kq_sleep_threshold = 20000; 160 SYSCTL_INT(_kern, OID_AUTO, kq_sleep_threshold, CTLFLAG_RW, 161 &kq_sleep_threshold, 0, "Minimum sleep duration without busy-looping"); 162 163 #define KNOTE_ACTIVATE(kn) do { \ 164 kn->kn_status |= KN_ACTIVE; \ 165 if ((kn->kn_status & (KN_QUEUED | KN_DISABLED)) == 0) \ 166 knote_enqueue(kn); \ 167 } while(0) 168 169 #define KN_HASHSIZE 64 /* XXX should be tunable */ 170 #define KN_HASH(val, mask) (((val) ^ (val >> 8)) & (mask)) 171 172 extern struct filterops aio_filtops; 173 extern struct filterops sig_filtops; 174 175 /* 176 * Table for for all system-defined filters. 177 */ 178 static struct filterops *sysfilt_ops[] = { 179 &file_filtops, /* EVFILT_READ */ 180 &file_filtops, /* EVFILT_WRITE */ 181 &aio_filtops, /* EVFILT_AIO */ 182 &file_filtops, /* EVFILT_VNODE */ 183 &proc_filtops, /* EVFILT_PROC */ 184 &sig_filtops, /* EVFILT_SIGNAL */ 185 &timer_filtops, /* EVFILT_TIMER */ 186 &file_filtops, /* EVFILT_EXCEPT */ 187 &user_filtops, /* EVFILT_USER */ 188 &fs_filtops, /* EVFILT_FS */ 189 }; 190 191 static struct knote_cache_list knote_cache_lists[MAXCPU]; 192 193 /* 194 * Acquire a knote, return non-zero on success, 0 on failure. 195 * 196 * If we cannot acquire the knote we sleep and return 0. The knote 197 * may be stale on return in this case and the caller must restart 198 * whatever loop they are in. 199 * 200 * Related kq token must be held. 201 */ 202 static __inline int 203 knote_acquire(struct knote *kn) 204 { 205 if (kn->kn_status & KN_PROCESSING) { 206 kn->kn_status |= KN_WAITING | KN_REPROCESS; 207 tsleep(kn, 0, "kqepts", hz); 208 /* knote may be stale now */ 209 return(0); 210 } 211 kn->kn_status |= KN_PROCESSING; 212 return(1); 213 } 214 215 /* 216 * Release an acquired knote, clearing KN_PROCESSING and handling any 217 * KN_REPROCESS events. 218 * 219 * Caller must be holding the related kq token 220 * 221 * Non-zero is returned if the knote is destroyed or detached. 222 */ 223 static __inline int 224 knote_release(struct knote *kn) 225 { 226 int ret; 227 228 while (kn->kn_status & KN_REPROCESS) { 229 kn->kn_status &= ~KN_REPROCESS; 230 if (kn->kn_status & KN_WAITING) { 231 kn->kn_status &= ~KN_WAITING; 232 wakeup(kn); 233 } 234 if (kn->kn_status & KN_DELETING) { 235 knote_detach_and_drop(kn); 236 return(1); 237 /* NOT REACHED */ 238 } 239 if (filter_event(kn, 0)) 240 KNOTE_ACTIVATE(kn); 241 } 242 if (kn->kn_status & KN_DETACHED) 243 ret = 1; 244 else 245 ret = 0; 246 kn->kn_status &= ~KN_PROCESSING; 247 /* kn should not be accessed anymore */ 248 return ret; 249 } 250 251 static int 252 filt_fileattach(struct knote *kn) 253 { 254 return (fo_kqfilter(kn->kn_fp, kn)); 255 } 256 257 /* 258 * MPSAFE 259 */ 260 static int 261 kqueue_kqfilter(struct file *fp, struct knote *kn) 262 { 263 struct kqueue *kq = (struct kqueue *)kn->kn_fp->f_data; 264 265 if (kn->kn_filter != EVFILT_READ) 266 return (EOPNOTSUPP); 267 268 kn->kn_fop = &kqread_filtops; 269 knote_insert(&kq->kq_kqinfo.ki_note, kn); 270 return (0); 271 } 272 273 static void 274 filt_kqdetach(struct knote *kn) 275 { 276 struct kqueue *kq = (struct kqueue *)kn->kn_fp->f_data; 277 278 knote_remove(&kq->kq_kqinfo.ki_note, kn); 279 } 280 281 /*ARGSUSED*/ 282 static int 283 filt_kqueue(struct knote *kn, long hint) 284 { 285 struct kqueue *kq = (struct kqueue *)kn->kn_fp->f_data; 286 287 kn->kn_data = kq->kq_count; 288 return (kn->kn_data > 0); 289 } 290 291 static int 292 filt_procattach(struct knote *kn) 293 { 294 struct proc *p; 295 int immediate; 296 297 immediate = 0; 298 p = pfind(kn->kn_id); 299 if (p == NULL && (kn->kn_sfflags & NOTE_EXIT)) { 300 p = zpfind(kn->kn_id); 301 immediate = 1; 302 } 303 if (p == NULL) { 304 return (ESRCH); 305 } 306 if (!PRISON_CHECK(curthread->td_ucred, p->p_ucred)) { 307 if (p) 308 PRELE(p); 309 return (EACCES); 310 } 311 312 lwkt_gettoken(&p->p_token); 313 kn->kn_ptr.p_proc = p; 314 kn->kn_flags |= EV_CLEAR; /* automatically set */ 315 316 /* 317 * internal flag indicating registration done by kernel 318 */ 319 if (kn->kn_flags & EV_FLAG1) { 320 kn->kn_data = kn->kn_sdata; /* ppid */ 321 kn->kn_fflags = NOTE_CHILD; 322 kn->kn_flags &= ~EV_FLAG1; 323 } 324 325 knote_insert(&p->p_klist, kn); 326 327 /* 328 * Immediately activate any exit notes if the target process is a 329 * zombie. This is necessary to handle the case where the target 330 * process, e.g. a child, dies before the kevent is negistered. 331 */ 332 if (immediate && filt_proc(kn, NOTE_EXIT)) 333 KNOTE_ACTIVATE(kn); 334 lwkt_reltoken(&p->p_token); 335 PRELE(p); 336 337 return (0); 338 } 339 340 /* 341 * The knote may be attached to a different process, which may exit, 342 * leaving nothing for the knote to be attached to. So when the process 343 * exits, the knote is marked as DETACHED and also flagged as ONESHOT so 344 * it will be deleted when read out. However, as part of the knote deletion, 345 * this routine is called, so a check is needed to avoid actually performing 346 * a detach, because the original process does not exist any more. 347 */ 348 static void 349 filt_procdetach(struct knote *kn) 350 { 351 struct proc *p; 352 353 if (kn->kn_status & KN_DETACHED) 354 return; 355 p = kn->kn_ptr.p_proc; 356 knote_remove(&p->p_klist, kn); 357 } 358 359 static int 360 filt_proc(struct knote *kn, long hint) 361 { 362 u_int event; 363 364 /* 365 * mask off extra data 366 */ 367 event = (u_int)hint & NOTE_PCTRLMASK; 368 369 /* 370 * if the user is interested in this event, record it. 371 */ 372 if (kn->kn_sfflags & event) 373 kn->kn_fflags |= event; 374 375 /* 376 * Process is gone, so flag the event as finished. Detach the 377 * knote from the process now because the process will be poof, 378 * gone later on. 379 */ 380 if (event == NOTE_EXIT) { 381 struct proc *p = kn->kn_ptr.p_proc; 382 if ((kn->kn_status & KN_DETACHED) == 0) { 383 PHOLD(p); 384 knote_remove(&p->p_klist, kn); 385 kn->kn_status |= KN_DETACHED; 386 kn->kn_data = p->p_xstat; 387 kn->kn_ptr.p_proc = NULL; 388 PRELE(p); 389 } 390 kn->kn_flags |= (EV_EOF | EV_NODATA | EV_ONESHOT); 391 return (1); 392 } 393 394 /* 395 * process forked, and user wants to track the new process, 396 * so attach a new knote to it, and immediately report an 397 * event with the parent's pid. 398 */ 399 if ((event == NOTE_FORK) && (kn->kn_sfflags & NOTE_TRACK)) { 400 struct kevent kev; 401 int error; 402 int n; 403 404 /* 405 * register knote with new process. 406 */ 407 kev.ident = hint & NOTE_PDATAMASK; /* pid */ 408 kev.filter = kn->kn_filter; 409 kev.flags = kn->kn_flags | EV_ADD | EV_ENABLE | EV_FLAG1; 410 kev.fflags = kn->kn_sfflags; 411 kev.data = kn->kn_id; /* parent */ 412 kev.udata = kn->kn_kevent.udata; /* preserve udata */ 413 n = 1; 414 error = kqueue_register(kn->kn_kq, &kev, &n); 415 if (error) 416 kn->kn_fflags |= NOTE_TRACKERR; 417 } 418 419 return (kn->kn_fflags != 0); 420 } 421 422 static void 423 filt_timerreset(struct knote *kn) 424 { 425 struct callout *calloutp; 426 struct timeval tv; 427 int tticks; 428 429 tv.tv_sec = kn->kn_sdata / 1000; 430 tv.tv_usec = (kn->kn_sdata % 1000) * 1000; 431 tticks = tvtohz_high(&tv); 432 calloutp = (struct callout *)kn->kn_hook; 433 callout_reset(calloutp, tticks, filt_timerexpire, kn); 434 } 435 436 /* 437 * The callout interlocks with callout_stop() but can still 438 * race a deletion so if KN_DELETING is set we just don't touch 439 * the knote. 440 */ 441 static void 442 filt_timerexpire(void *knx) 443 { 444 struct knote *kn = knx; 445 struct kqueue *kq = kn->kn_kq; 446 447 lwkt_getpooltoken(kq); 448 449 /* 450 * Open knote_acquire(), since we can't sleep in callout, 451 * however, we do need to record this expiration. 452 */ 453 kn->kn_data++; 454 if (kn->kn_status & KN_PROCESSING) { 455 kn->kn_status |= KN_REPROCESS; 456 if ((kn->kn_status & KN_DELETING) == 0 && 457 (kn->kn_flags & EV_ONESHOT) == 0) 458 filt_timerreset(kn); 459 lwkt_relpooltoken(kq); 460 return; 461 } 462 KASSERT((kn->kn_status & KN_DELETING) == 0, 463 ("acquire a deleting knote %#x", kn->kn_status)); 464 kn->kn_status |= KN_PROCESSING; 465 466 KNOTE_ACTIVATE(kn); 467 if ((kn->kn_flags & EV_ONESHOT) == 0) 468 filt_timerreset(kn); 469 470 knote_release(kn); 471 472 lwkt_relpooltoken(kq); 473 } 474 475 /* 476 * data contains amount of time to sleep, in milliseconds 477 */ 478 static int 479 filt_timerattach(struct knote *kn) 480 { 481 struct callout *calloutp; 482 int prev_ncallouts; 483 484 prev_ncallouts = atomic_fetchadd_int(&kq_ncallouts, 1); 485 if (prev_ncallouts >= kq_calloutmax) { 486 atomic_subtract_int(&kq_ncallouts, 1); 487 kn->kn_hook = NULL; 488 return (ENOMEM); 489 } 490 491 kn->kn_flags |= EV_CLEAR; /* automatically set */ 492 calloutp = kmalloc(sizeof(*calloutp), M_KQUEUE, M_WAITOK); 493 callout_init_mp(calloutp); 494 kn->kn_hook = (caddr_t)calloutp; 495 496 filt_timerreset(kn); 497 return (0); 498 } 499 500 /* 501 * This function is called with the knote flagged locked but it is 502 * still possible to race a callout event due to the callback blocking. 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 void 683 kqueue_init(struct kqueue *kq, struct filedesc *fdp) 684 { 685 bzero(kq, sizeof(*kq)); 686 TAILQ_INIT(&kq->kq_knpend); 687 TAILQ_INIT(&kq->kq_knlist); 688 kq->kq_fdp = fdp; 689 SLIST_INIT(&kq->kq_kqinfo.ki_note); 690 } 691 692 /* 693 * Terminate a kqueue. Freeing the actual kq itself is left up to the 694 * caller (it might be embedded in a lwp so we don't do it here). 695 * 696 * The kq's knlist must be completely eradicated so block on any 697 * processing races. 698 */ 699 void 700 kqueue_terminate(struct kqueue *kq) 701 { 702 struct knote *kn; 703 704 lwkt_getpooltoken(kq); 705 while ((kn = TAILQ_FIRST(&kq->kq_knlist)) != NULL) { 706 if (knote_acquire(kn)) 707 knote_detach_and_drop(kn); 708 } 709 lwkt_relpooltoken(kq); 710 711 if (kq->kq_knhash) { 712 hashdestroy(kq->kq_knhash, M_KQUEUE, kq->kq_knhashmask); 713 kq->kq_knhash = NULL; 714 kq->kq_knhashmask = 0; 715 } 716 } 717 718 /* 719 * MPSAFE 720 */ 721 int 722 sys_kqueue(struct sysmsg *sysmsg, const struct kqueue_args *uap) 723 { 724 struct thread *td = curthread; 725 struct kqueue *kq; 726 struct file *fp; 727 int fd, error; 728 729 error = falloc(td->td_lwp, &fp, &fd); 730 if (error) 731 return (error); 732 fp->f_flag = FREAD | FWRITE; 733 fp->f_type = DTYPE_KQUEUE; 734 fp->f_ops = &kqueueops; 735 736 kq = kmalloc(sizeof(struct kqueue), M_KQUEUE, M_WAITOK | M_ZERO); 737 kqueue_init(kq, td->td_proc->p_fd); 738 fp->f_data = kq; 739 740 fsetfd(kq->kq_fdp, fp, fd); 741 sysmsg->sysmsg_result = fd; 742 fdrop(fp); 743 return (error); 744 } 745 746 /* 747 * Copy 'count' items into the destination list pointed to by uap->eventlist. 748 */ 749 static int 750 kevent_copyout(void *arg, struct kevent *kevp, int count, int *res) 751 { 752 struct kevent_copyin_args *kap; 753 int error; 754 755 kap = (struct kevent_copyin_args *)arg; 756 757 error = copyout(kevp, kap->eventlist, count * sizeof(*kevp)); 758 if (error == 0) { 759 kap->eventlist += count; 760 *res += count; 761 } else { 762 *res = -1; 763 } 764 765 return (error); 766 } 767 768 /* 769 * Copy at most 'max' items from the list pointed to by kap->changelist, 770 * return number of items in 'events'. 771 */ 772 static int 773 kevent_copyin(void *arg, struct kevent *kevp, int max, int *events) 774 { 775 struct kevent_copyin_args *kap; 776 int error, count; 777 778 kap = (struct kevent_copyin_args *)arg; 779 780 count = min(kap->ka->nchanges - kap->pchanges, max); 781 error = copyin(kap->changelist, kevp, count * sizeof *kevp); 782 if (error == 0) { 783 kap->changelist += count; 784 kap->pchanges += count; 785 *events = count; 786 } 787 788 return (error); 789 } 790 791 /* 792 * MPSAFE 793 */ 794 int 795 kern_kevent(struct kqueue *kq, int nevents, int *res, void *uap, 796 k_copyin_fn kevent_copyinfn, k_copyout_fn kevent_copyoutfn, 797 struct timespec *tsp_in, int flags) 798 { 799 struct kevent *kevp; 800 struct timespec *tsp, ats; 801 int i, n, total, error, nerrors = 0; 802 int gobbled; 803 int lres; 804 int limit = kq_checkloop; 805 int closedcounter; 806 int scan_flags; 807 struct kevent kev[KQ_NEVENTS]; 808 struct knote marker; 809 struct lwkt_token *tok; 810 811 if (tsp_in == NULL || tsp_in->tv_sec || tsp_in->tv_nsec) 812 atomic_set_int(&curthread->td_mpflags, TDF_MP_BATCH_DEMARC); 813 814 tsp = tsp_in; 815 *res = 0; 816 817 closedcounter = kq->kq_fdp->fd_closedcounter; 818 819 for (;;) { 820 n = 0; 821 error = kevent_copyinfn(uap, kev, KQ_NEVENTS, &n); 822 if (error) 823 return error; 824 if (n == 0) 825 break; 826 for (i = 0; i < n; ++i) 827 kev[i].flags &= ~EV_SYSFLAGS; 828 for (i = 0; i < n; ++i) { 829 gobbled = n - i; 830 error = kqueue_register(kq, &kev[i], &gobbled); 831 i += gobbled - 1; 832 kevp = &kev[i]; 833 834 /* 835 * If a registration returns an error we 836 * immediately post the error. The kevent() 837 * call itself will fail with the error if 838 * no space is available for posting. 839 * 840 * Such errors normally bypass the timeout/blocking 841 * code. However, if the copyoutfn function refuses 842 * to post the error (see sys_poll()), then we 843 * ignore it too. 844 */ 845 if (error || (kevp->flags & EV_RECEIPT)) { 846 kevp->flags = EV_ERROR; 847 kevp->data = error; 848 lres = *res; 849 kevent_copyoutfn(uap, kevp, 1, res); 850 if (*res < 0) { 851 return error; 852 } else if (lres != *res) { 853 nevents--; 854 nerrors++; 855 } 856 } 857 } 858 } 859 if (nerrors) 860 return 0; 861 862 /* 863 * Acquire/wait for events - setup timeout 864 * 865 * If no timeout specified clean up the run path by clearing the 866 * PRECISE flag. 867 */ 868 if (tsp != NULL) { 869 if (tsp->tv_sec || tsp->tv_nsec) { 870 getnanouptime(&ats); 871 timespecadd(tsp, &ats, tsp); /* tsp = target time */ 872 } 873 } else { 874 flags &= ~KEVENT_TIMEOUT_PRECISE; 875 } 876 877 /* 878 * Loop as required. 879 * 880 * Collect as many events as we can. Sleeping on successive 881 * loops is disabled if copyoutfn has incremented (*res). 882 * 883 * The loop stops if an error occurs, all events have been 884 * scanned (the marker has been reached), or fewer than the 885 * maximum number of events is found. 886 * 887 * The copyoutfn function does not have to increment (*res) in 888 * order for the loop to continue. 889 * 890 * NOTE: doselect() usually passes 0x7FFFFFFF for nevents. 891 */ 892 total = 0; 893 error = 0; 894 marker.kn_filter = EVFILT_MARKER; 895 marker.kn_status = KN_PROCESSING; 896 897 tok = lwkt_token_pool_lookup(kq); 898 scan_flags = KEVENT_SCAN_INSERT_MARKER; 899 900 while ((n = nevents - total) > 0) { 901 if (n > KQ_NEVENTS) 902 n = KQ_NEVENTS; 903 904 /* 905 * Process all received events 906 * Account for all non-spurious events in our total 907 */ 908 i = kqueue_scan(kq, kev, n, &marker, closedcounter, scan_flags); 909 scan_flags = KEVENT_SCAN_KEEP_MARKER; 910 if (i) { 911 lres = *res; 912 error = kevent_copyoutfn(uap, kev, i, res); 913 total += *res - lres; 914 if (error) 915 break; 916 } 917 if (limit && --limit == 0) 918 panic("kqueue: checkloop failed i=%d", i); 919 920 /* 921 * Normally when fewer events are returned than requested 922 * we can stop. However, if only spurious events were 923 * collected the copyout will not bump (*res) and we have 924 * to continue. 925 */ 926 if (i < n && *res) 927 break; 928 929 /* 930 * If no events were recorded (no events happened or the events 931 * that did happen were all spurious), block until an event 932 * occurs or the timeout occurs and reload the marker. 933 * 934 * If we saturated n (i == n) loop up without sleeping to 935 * continue processing the list. 936 */ 937 if (i != n && kq->kq_count == 0 && *res == 0) { 938 int timeout; 939 int ustimeout; 940 941 if (tsp == NULL) { 942 timeout = 0; 943 ustimeout = 0; 944 } else if (tsp->tv_sec == 0 && tsp->tv_nsec == 0) { 945 error = EWOULDBLOCK; 946 break; 947 } else { 948 struct timespec atx = *tsp; 949 950 getnanouptime(&ats); 951 timespecsub(&atx, &ats, &atx); 952 if (atx.tv_sec < 0 || 953 (atx.tv_sec == 0 && atx.tv_nsec <= 0)) { 954 error = EWOULDBLOCK; 955 break; 956 } 957 if (flags & KEVENT_TIMEOUT_PRECISE) { 958 if (atx.tv_sec == 0 && 959 atx.tv_nsec < kq_sleep_threshold) { 960 ustimeout = kq_sleep_threshold / 961 1000; 962 } else if (atx.tv_sec < 60) { 963 ustimeout = 964 atx.tv_sec * 1000000 + 965 atx.tv_nsec / 1000; 966 } else { 967 ustimeout = 60 * 1000000; 968 } 969 if (ustimeout == 0) 970 ustimeout = 1; 971 timeout = 0; 972 } else if (atx.tv_sec > 60 * 60) { 973 timeout = 60 * 60 * hz; 974 ustimeout = 0; 975 } else { 976 timeout = tstohz_high(&atx); 977 ustimeout = 0; 978 } 979 } 980 981 lwkt_gettoken(tok); 982 if (kq->kq_count == 0) { 983 kq->kq_sleep_cnt++; 984 if (__predict_false(kq->kq_sleep_cnt == 0)) { 985 /* 986 * Guard against possible wrapping. And 987 * set it to 2, so that kqueue_wakeup() 988 * can wake everyone up. 989 */ 990 kq->kq_sleep_cnt = 2; 991 } 992 if (flags & KEVENT_TIMEOUT_PRECISE) { 993 error = precise_sleep(kq, PCATCH, 994 "kqread", ustimeout); 995 } else { 996 error = tsleep(kq, PCATCH, 997 "kqread", timeout); 998 } 999 1000 /* don't restart after signals... */ 1001 if (error == ERESTART) 1002 error = EINTR; 1003 if (error == EWOULDBLOCK) 1004 error = 0; 1005 if (error) { 1006 lwkt_reltoken(tok); 1007 break; 1008 } 1009 scan_flags = KEVENT_SCAN_RELOAD_MARKER; 1010 } 1011 lwkt_reltoken(tok); 1012 } 1013 1014 /* 1015 * Deal with an edge case where spurious events can cause 1016 * a loop to occur without moving the marker. This can 1017 * prevent kqueue_scan() from picking up new events which 1018 * race us. We must be sure to move the marker for this 1019 * case. 1020 * 1021 * NOTE: We do not want to move the marker if events 1022 * were scanned because normal kqueue operations 1023 * may reactivate events. Moving the marker in 1024 * that case could result in duplicates for the 1025 * same event. 1026 */ 1027 if (i == 0) 1028 scan_flags = KEVENT_SCAN_RELOAD_MARKER; 1029 } 1030 1031 /* 1032 * Remove the marker 1033 */ 1034 if (scan_flags != KEVENT_SCAN_INSERT_MARKER) { 1035 lwkt_gettoken(tok); 1036 TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe); 1037 lwkt_reltoken(tok); 1038 } 1039 1040 /* Timeouts do not return EWOULDBLOCK. */ 1041 if (error == EWOULDBLOCK) 1042 error = 0; 1043 return error; 1044 } 1045 1046 /* 1047 * MPALMOSTSAFE 1048 */ 1049 int 1050 sys_kevent(struct sysmsg *sysmsg, const struct kevent_args *uap) 1051 { 1052 struct thread *td = curthread; 1053 struct timespec ts, *tsp; 1054 struct kqueue *kq; 1055 struct file *fp = NULL; 1056 struct kevent_copyin_args *kap, ka; 1057 int error; 1058 1059 if (uap->timeout) { 1060 error = copyin(uap->timeout, &ts, sizeof(ts)); 1061 if (error) 1062 return (error); 1063 tsp = &ts; 1064 } else { 1065 tsp = NULL; 1066 } 1067 fp = holdfp(td, uap->fd, -1); 1068 if (fp == NULL) 1069 return (EBADF); 1070 if (fp->f_type != DTYPE_KQUEUE) { 1071 fdrop(fp); 1072 return (EBADF); 1073 } 1074 1075 kq = (struct kqueue *)fp->f_data; 1076 1077 kap = &ka; 1078 kap->ka = uap; 1079 kap->pchanges = 0; 1080 kap->eventlist = uap->eventlist; 1081 kap->changelist = uap->changelist; 1082 1083 error = kern_kevent(kq, uap->nevents, &sysmsg->sysmsg_result, kap, 1084 kevent_copyin, kevent_copyout, tsp, 0); 1085 1086 dropfp(td, uap->fd, fp); 1087 1088 return (error); 1089 } 1090 1091 /* 1092 * Efficiently load multiple file pointers. This significantly reduces 1093 * threaded overhead. When doing simple polling we can depend on the 1094 * per-thread (fd,fp) cache. With more descriptors, we batch. 1095 */ 1096 static 1097 void 1098 floadkevfps(thread_t td, struct filedesc *fdp, struct kevent *kev, 1099 struct file **fp, int climit) 1100 { 1101 struct filterops *fops; 1102 int tdcache; 1103 1104 if (climit <= 2 && td->td_proc && td->td_proc->p_fd == fdp) { 1105 tdcache = 1; 1106 } else { 1107 tdcache = 0; 1108 spin_lock_shared(&fdp->fd_spin); 1109 } 1110 1111 while (climit) { 1112 *fp = NULL; 1113 if (kev->filter < 0 && 1114 kev->filter + EVFILT_SYSCOUNT >= 0) { 1115 fops = sysfilt_ops[~kev->filter]; 1116 if (fops->f_flags & FILTEROP_ISFD) { 1117 if (tdcache) { 1118 *fp = holdfp(td, kev->ident, -1); 1119 } else { 1120 *fp = holdfp_fdp_locked(fdp, 1121 kev->ident, -1); 1122 } 1123 } 1124 } 1125 --climit; 1126 ++fp; 1127 ++kev; 1128 } 1129 if (tdcache == 0) 1130 spin_unlock_shared(&fdp->fd_spin); 1131 } 1132 1133 /* 1134 * Register up to *countp kev's. Always registers at least 1. 1135 * 1136 * The number registered is returned in *countp. 1137 * 1138 * If an error occurs or a kev is flagged EV_RECEIPT, it is 1139 * processed and included in *countp, and processing then 1140 * stops. 1141 */ 1142 int 1143 kqueue_register(struct kqueue *kq, struct kevent *kev, int *countp) 1144 { 1145 struct filedesc *fdp = kq->kq_fdp; 1146 struct klist *list = NULL; 1147 struct filterops *fops; 1148 struct file *fp[KQ_NEVENTS]; 1149 struct knote *kn = NULL; 1150 struct thread *td; 1151 int error; 1152 int count; 1153 int climit; 1154 int closedcounter; 1155 struct knote_cache_list *cache_list; 1156 1157 td = curthread; 1158 climit = *countp; 1159 if (climit > KQ_NEVENTS) 1160 climit = KQ_NEVENTS; 1161 closedcounter = fdp->fd_closedcounter; 1162 floadkevfps(td, fdp, kev, fp, climit); 1163 1164 lwkt_getpooltoken(kq); 1165 count = 0; 1166 error = 0; 1167 1168 /* 1169 * To avoid races, only one thread can register events on this 1170 * kqueue at a time. 1171 */ 1172 while (__predict_false(kq->kq_regtd != NULL && kq->kq_regtd != td)) { 1173 kq->kq_state |= KQ_REGWAIT; 1174 tsleep(&kq->kq_regtd, 0, "kqreg", 0); 1175 } 1176 if (__predict_false(kq->kq_regtd != NULL)) { 1177 /* Recursive calling of kqueue_register() */ 1178 td = NULL; 1179 } else { 1180 /* Owner of the kq_regtd, i.e. td != NULL */ 1181 kq->kq_regtd = td; 1182 } 1183 1184 loop: 1185 if (kev->filter < 0) { 1186 if (kev->filter + EVFILT_SYSCOUNT < 0) { 1187 error = EINVAL; 1188 ++count; 1189 goto done; 1190 } 1191 fops = sysfilt_ops[~kev->filter]; /* to 0-base index */ 1192 } else { 1193 /* 1194 * XXX 1195 * filter attach routine is responsible for insuring that 1196 * the identifier can be attached to it. 1197 */ 1198 error = EINVAL; 1199 ++count; 1200 goto done; 1201 } 1202 1203 if (fops->f_flags & FILTEROP_ISFD) { 1204 /* validate descriptor */ 1205 if (fp[count] == NULL) { 1206 error = EBADF; 1207 ++count; 1208 goto done; 1209 } 1210 } 1211 1212 cache_list = &knote_cache_lists[mycpuid]; 1213 if (SLIST_EMPTY(&cache_list->knote_cache)) { 1214 struct knote *new_kn; 1215 1216 new_kn = knote_alloc(); 1217 crit_enter(); 1218 SLIST_INSERT_HEAD(&cache_list->knote_cache, new_kn, kn_link); 1219 cache_list->knote_cache_cnt++; 1220 crit_exit(); 1221 } 1222 1223 if (fp[count] != NULL) { 1224 list = &fp[count]->f_klist; 1225 } else if (kq->kq_knhashmask) { 1226 list = &kq->kq_knhash[ 1227 KN_HASH((u_long)kev->ident, kq->kq_knhashmask)]; 1228 } 1229 if (list != NULL) { 1230 lwkt_getpooltoken(list); 1231 again: 1232 SLIST_FOREACH(kn, list, kn_link) { 1233 if (kn->kn_kq == kq && 1234 kn->kn_filter == kev->filter && 1235 kn->kn_id == kev->ident) { 1236 if (knote_acquire(kn) == 0) 1237 goto again; 1238 break; 1239 } 1240 } 1241 lwkt_relpooltoken(list); 1242 } 1243 1244 /* 1245 * NOTE: At this point if kn is non-NULL we will have acquired 1246 * it and set KN_PROCESSING. 1247 */ 1248 if (kn == NULL && ((kev->flags & EV_ADD) == 0)) { 1249 error = ENOENT; 1250 ++count; 1251 goto done; 1252 } 1253 1254 /* 1255 * kn now contains the matching knote, or NULL if no match 1256 */ 1257 if (kev->flags & EV_ADD) { 1258 if (kn == NULL) { 1259 crit_enter(); 1260 kn = SLIST_FIRST(&cache_list->knote_cache); 1261 if (kn == NULL) { 1262 crit_exit(); 1263 kn = knote_alloc(); 1264 } else { 1265 SLIST_REMOVE_HEAD(&cache_list->knote_cache, 1266 kn_link); 1267 cache_list->knote_cache_cnt--; 1268 crit_exit(); 1269 } 1270 kn->kn_fp = fp[count]; 1271 kn->kn_kq = kq; 1272 kn->kn_fop = fops; 1273 1274 /* 1275 * apply reference count to knote structure, and 1276 * do not release it at the end of this routine. 1277 */ 1278 fp[count] = NULL; /* safety */ 1279 1280 kn->kn_sfflags = kev->fflags; 1281 kn->kn_sdata = kev->data; 1282 kev->fflags = 0; 1283 kev->data = 0; 1284 kn->kn_kevent = *kev; 1285 1286 /* 1287 * KN_PROCESSING prevents the knote from getting 1288 * ripped out from under us while we are trying 1289 * to attach it, in case the attach blocks. 1290 */ 1291 kn->kn_status = KN_PROCESSING; 1292 knote_attach(kn); 1293 if ((error = filter_attach(kn)) != 0) { 1294 kn->kn_status |= KN_DELETING | KN_REPROCESS; 1295 knote_drop(kn); 1296 ++count; 1297 goto done; 1298 } 1299 1300 /* 1301 * Interlock against close races which either tried 1302 * to remove our knote while we were blocked or missed 1303 * it entirely prior to our attachment. We do not 1304 * want to end up with a knote on a closed descriptor. 1305 */ 1306 if ((fops->f_flags & FILTEROP_ISFD) && 1307 checkfdclosed(curthread, fdp, kev->ident, kn->kn_fp, 1308 closedcounter)) { 1309 kn->kn_status |= KN_DELETING | KN_REPROCESS; 1310 } 1311 } else { 1312 /* 1313 * The user may change some filter values after the 1314 * initial EV_ADD, but doing so will not reset any 1315 * filter which have already been triggered. 1316 */ 1317 KKASSERT(kn->kn_status & KN_PROCESSING); 1318 if (fops == &user_filtops) { 1319 filt_usertouch(kn, kev, EVENT_REGISTER); 1320 } else { 1321 kn->kn_sfflags = kev->fflags; 1322 kn->kn_sdata = kev->data; 1323 kn->kn_kevent.udata = kev->udata; 1324 } 1325 } 1326 1327 /* 1328 * Execute the filter event to immediately activate the 1329 * knote if necessary. If reprocessing events are pending 1330 * due to blocking above we do not run the filter here 1331 * but instead let knote_release() do it. Otherwise we 1332 * might run the filter on a deleted event. 1333 */ 1334 if ((kn->kn_status & KN_REPROCESS) == 0) { 1335 if (filter_event(kn, 0)) 1336 KNOTE_ACTIVATE(kn); 1337 } 1338 } else if (kev->flags & EV_DELETE) { 1339 /* 1340 * Delete the existing knote 1341 */ 1342 knote_detach_and_drop(kn); 1343 error = 0; 1344 ++count; 1345 goto done; 1346 } else { 1347 /* 1348 * Modify an existing event. 1349 * 1350 * The user may change some filter values after the 1351 * initial EV_ADD, but doing so will not reset any 1352 * filter which have already been triggered. 1353 */ 1354 KKASSERT(kn->kn_status & KN_PROCESSING); 1355 if (fops == &user_filtops) { 1356 filt_usertouch(kn, kev, EVENT_REGISTER); 1357 } else { 1358 kn->kn_sfflags = kev->fflags; 1359 kn->kn_sdata = kev->data; 1360 kn->kn_kevent.udata = kev->udata; 1361 } 1362 1363 /* 1364 * Execute the filter event to immediately activate the 1365 * knote if necessary. If reprocessing events are pending 1366 * due to blocking above we do not run the filter here 1367 * but instead let knote_release() do it. Otherwise we 1368 * might run the filter on a deleted event. 1369 */ 1370 if ((kn->kn_status & KN_REPROCESS) == 0) { 1371 if (filter_event(kn, 0)) 1372 KNOTE_ACTIVATE(kn); 1373 } 1374 } 1375 1376 /* 1377 * Disablement does not deactivate a knote here. 1378 */ 1379 if ((kev->flags & EV_DISABLE) && 1380 ((kn->kn_status & KN_DISABLED) == 0)) { 1381 kn->kn_status |= KN_DISABLED; 1382 } 1383 1384 /* 1385 * Re-enablement may have to immediately enqueue an active knote. 1386 */ 1387 if ((kev->flags & EV_ENABLE) && (kn->kn_status & KN_DISABLED)) { 1388 kn->kn_status &= ~KN_DISABLED; 1389 if ((kn->kn_status & KN_ACTIVE) && 1390 ((kn->kn_status & KN_QUEUED) == 0)) { 1391 knote_enqueue(kn); 1392 } 1393 } 1394 1395 /* 1396 * Handle any required reprocessing 1397 */ 1398 knote_release(kn); 1399 /* kn may be invalid now */ 1400 1401 /* 1402 * Loop control. We stop on errors (above), and also stop after 1403 * processing EV_RECEIPT, so the caller can process it. 1404 */ 1405 ++count; 1406 if (kev->flags & EV_RECEIPT) { 1407 error = 0; 1408 goto done; 1409 } 1410 ++kev; 1411 if (count < climit) { 1412 if (fp[count-1]) /* drop unprocessed fp */ 1413 fdrop(fp[count-1]); 1414 goto loop; 1415 } 1416 1417 /* 1418 * Cleanup 1419 */ 1420 done: 1421 if (td != NULL) { /* Owner of the kq_regtd */ 1422 kq->kq_regtd = NULL; 1423 if (__predict_false(kq->kq_state & KQ_REGWAIT)) { 1424 kq->kq_state &= ~KQ_REGWAIT; 1425 wakeup(&kq->kq_regtd); 1426 } 1427 } 1428 lwkt_relpooltoken(kq); 1429 1430 /* 1431 * Drop unprocessed file pointers 1432 */ 1433 *countp = count; 1434 if (count && fp[count-1]) 1435 fdrop(fp[count-1]); 1436 while (count < climit) { 1437 if (fp[count]) 1438 fdrop(fp[count]); 1439 ++count; 1440 } 1441 return (error); 1442 } 1443 1444 /* 1445 * Scan the kqueue, return the number of active events placed in kevp up 1446 * to count. 1447 * 1448 * Continuous mode events may get recycled, do not continue scanning past 1449 * marker unless no events have been collected. 1450 */ 1451 static int 1452 kqueue_scan(struct kqueue *kq, struct kevent *kevp, int count, 1453 struct knote *marker, int closedcounter, int scan_flags) 1454 { 1455 struct knote *kn, local_marker; 1456 thread_t td = curthread; 1457 int total; 1458 1459 total = 0; 1460 local_marker.kn_filter = EVFILT_MARKER; 1461 local_marker.kn_status = KN_PROCESSING; 1462 1463 lwkt_getpooltoken(kq); 1464 1465 /* 1466 * Adjust marker, insert initial marker, or leave the marker alone. 1467 * 1468 * Also setup our local_marker. 1469 */ 1470 switch(scan_flags) { 1471 case KEVENT_SCAN_RELOAD_MARKER: 1472 TAILQ_REMOVE(&kq->kq_knpend, marker, kn_tqe); 1473 /* fall through */ 1474 case KEVENT_SCAN_INSERT_MARKER: 1475 TAILQ_INSERT_TAIL(&kq->kq_knpend, marker, kn_tqe); 1476 break; 1477 } 1478 TAILQ_INSERT_HEAD(&kq->kq_knpend, &local_marker, kn_tqe); 1479 1480 /* 1481 * Collect events. 1482 */ 1483 while (count) { 1484 kn = TAILQ_NEXT(&local_marker, kn_tqe); 1485 if (kn->kn_filter == EVFILT_MARKER) { 1486 /* Marker reached, we are done */ 1487 if (kn == marker) 1488 break; 1489 1490 /* Move local marker past some other threads marker */ 1491 kn = TAILQ_NEXT(kn, kn_tqe); 1492 TAILQ_REMOVE(&kq->kq_knpend, &local_marker, kn_tqe); 1493 TAILQ_INSERT_BEFORE(kn, &local_marker, kn_tqe); 1494 continue; 1495 } 1496 1497 /* 1498 * We can't skip a knote undergoing processing, otherwise 1499 * we risk not returning it when the user process expects 1500 * it should be returned. Sleep and retry. 1501 */ 1502 if (knote_acquire(kn) == 0) 1503 continue; 1504 1505 /* 1506 * Remove the event for processing. 1507 * 1508 * WARNING! We must leave KN_QUEUED set to prevent the 1509 * event from being KNOTE_ACTIVATE()d while 1510 * the queue state is in limbo, in case we 1511 * block. 1512 */ 1513 TAILQ_REMOVE(&kq->kq_knpend, kn, kn_tqe); 1514 kq->kq_count--; 1515 1516 /* 1517 * We have to deal with an extremely important race against 1518 * file descriptor close()s here. The file descriptor can 1519 * disappear MPSAFE, and there is a small window of 1520 * opportunity between that and the call to knote_fdclose(). 1521 * 1522 * If we hit that window here while doselect or dopoll is 1523 * trying to delete a spurious event they will not be able 1524 * to match up the event against a knote and will go haywire. 1525 */ 1526 if ((kn->kn_fop->f_flags & FILTEROP_ISFD) && 1527 checkfdclosed(td, kq->kq_fdp, kn->kn_kevent.ident, 1528 kn->kn_fp, closedcounter)) { 1529 kn->kn_status |= KN_DELETING | KN_REPROCESS; 1530 } 1531 1532 if (kn->kn_status & KN_DISABLED) { 1533 /* 1534 * If disabled we ensure the event is not queued 1535 * but leave its active bit set. On re-enablement 1536 * the event may be immediately triggered. 1537 */ 1538 kn->kn_status &= ~KN_QUEUED; 1539 } else if ((kn->kn_flags & EV_ONESHOT) == 0 && 1540 (kn->kn_status & KN_DELETING) == 0 && 1541 filter_event(kn, 0) == 0) { 1542 /* 1543 * If not running in one-shot mode and the event 1544 * is no longer present we ensure it is removed 1545 * from the queue and ignore it. 1546 */ 1547 kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE); 1548 } else { 1549 /* 1550 * Post the event 1551 */ 1552 if (kn->kn_fop == &user_filtops) 1553 filt_usertouch(kn, kevp, EVENT_PROCESS); 1554 else 1555 *kevp = kn->kn_kevent; 1556 ++kevp; 1557 ++total; 1558 --count; 1559 1560 if (kn->kn_flags & EV_ONESHOT) { 1561 kn->kn_status &= ~KN_QUEUED; 1562 kn->kn_status |= KN_DELETING | KN_REPROCESS; 1563 } else { 1564 if (kn->kn_flags & (EV_CLEAR | EV_DISPATCH)) { 1565 if (kn->kn_flags & EV_CLEAR) { 1566 kn->kn_data = 0; 1567 kn->kn_fflags = 0; 1568 } 1569 if (kn->kn_flags & EV_DISPATCH) { 1570 kn->kn_status |= KN_DISABLED; 1571 } 1572 kn->kn_status &= ~(KN_QUEUED | 1573 KN_ACTIVE); 1574 } else { 1575 TAILQ_INSERT_TAIL(&kq->kq_knpend, kn, kn_tqe); 1576 kq->kq_count++; 1577 } 1578 } 1579 } 1580 1581 /* 1582 * Handle any post-processing states 1583 */ 1584 knote_release(kn); 1585 } 1586 TAILQ_REMOVE(&kq->kq_knpend, &local_marker, kn_tqe); 1587 1588 lwkt_relpooltoken(kq); 1589 return (total); 1590 } 1591 1592 /* 1593 * XXX 1594 * This could be expanded to call kqueue_scan, if desired. 1595 * 1596 * MPSAFE 1597 */ 1598 static int 1599 kqueue_read(struct file *fp, struct uio *uio, struct ucred *cred, int flags) 1600 { 1601 return (ENXIO); 1602 } 1603 1604 /* 1605 * MPSAFE 1606 */ 1607 static int 1608 kqueue_write(struct file *fp, struct uio *uio, struct ucred *cred, int flags) 1609 { 1610 return (ENXIO); 1611 } 1612 1613 /* 1614 * MPALMOSTSAFE 1615 */ 1616 static int 1617 kqueue_ioctl(struct file *fp, u_long com, caddr_t data, 1618 struct ucred *cred, struct sysmsg *msg) 1619 { 1620 struct kqueue *kq; 1621 int error; 1622 1623 kq = (struct kqueue *)fp->f_data; 1624 lwkt_getpooltoken(kq); 1625 switch(com) { 1626 case FIOASYNC: 1627 if (*(int *)data) 1628 kq->kq_state |= KQ_ASYNC; 1629 else 1630 kq->kq_state &= ~KQ_ASYNC; 1631 error = 0; 1632 break; 1633 case FIOSETOWN: 1634 error = fsetown(*(int *)data, &kq->kq_sigio); 1635 break; 1636 default: 1637 error = ENOTTY; 1638 break; 1639 } 1640 lwkt_relpooltoken(kq); 1641 return (error); 1642 } 1643 1644 /* 1645 * MPSAFE 1646 */ 1647 static int 1648 kqueue_stat(struct file *fp, struct stat *st, struct ucred *cred) 1649 { 1650 struct kqueue *kq = (struct kqueue *)fp->f_data; 1651 1652 bzero((void *)st, sizeof(*st)); 1653 st->st_size = kq->kq_count; 1654 st->st_blksize = sizeof(struct kevent); 1655 st->st_mode = S_IFIFO; 1656 return (0); 1657 } 1658 1659 /* 1660 * MPSAFE 1661 */ 1662 static int 1663 kqueue_close(struct file *fp) 1664 { 1665 struct kqueue *kq = (struct kqueue *)fp->f_data; 1666 1667 kqueue_terminate(kq); 1668 1669 fp->f_data = NULL; 1670 funsetown(&kq->kq_sigio); 1671 1672 kfree(kq, M_KQUEUE); 1673 return (0); 1674 } 1675 1676 static void 1677 kqueue_wakeup(struct kqueue *kq) 1678 { 1679 if (kq->kq_sleep_cnt) { 1680 u_int sleep_cnt = kq->kq_sleep_cnt; 1681 1682 kq->kq_sleep_cnt = 0; 1683 if (sleep_cnt == 1) 1684 wakeup_one(kq); 1685 else 1686 wakeup(kq); 1687 } 1688 KNOTE(&kq->kq_kqinfo.ki_note, 0); 1689 } 1690 1691 /* 1692 * Calls filterops f_attach function, acquiring mplock if filter is not 1693 * marked as FILTEROP_MPSAFE. 1694 * 1695 * Caller must be holding the related kq token 1696 */ 1697 static int 1698 filter_attach(struct knote *kn) 1699 { 1700 int ret; 1701 1702 if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) { 1703 ret = kn->kn_fop->f_attach(kn); 1704 } else { 1705 get_mplock(); 1706 ret = kn->kn_fop->f_attach(kn); 1707 rel_mplock(); 1708 } 1709 return (ret); 1710 } 1711 1712 /* 1713 * Detach the knote and drop it, destroying the knote. 1714 * 1715 * Calls filterops f_detach function, acquiring mplock if filter is not 1716 * marked as FILTEROP_MPSAFE. 1717 * 1718 * Caller must be holding the related kq token 1719 */ 1720 static void 1721 knote_detach_and_drop(struct knote *kn) 1722 { 1723 kn->kn_status |= KN_DELETING | KN_REPROCESS; 1724 if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) { 1725 kn->kn_fop->f_detach(kn); 1726 } else { 1727 get_mplock(); 1728 kn->kn_fop->f_detach(kn); 1729 rel_mplock(); 1730 } 1731 knote_drop(kn); 1732 } 1733 1734 /* 1735 * Calls filterops f_event function, acquiring mplock if filter is not 1736 * marked as FILTEROP_MPSAFE. 1737 * 1738 * If the knote is in the middle of being created or deleted we cannot 1739 * safely call the filter op. 1740 * 1741 * Caller must be holding the related kq token 1742 */ 1743 static int 1744 filter_event(struct knote *kn, long hint) 1745 { 1746 int ret; 1747 1748 if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) { 1749 ret = kn->kn_fop->f_event(kn, hint); 1750 } else { 1751 get_mplock(); 1752 ret = kn->kn_fop->f_event(kn, hint); 1753 rel_mplock(); 1754 } 1755 return (ret); 1756 } 1757 1758 /* 1759 * Walk down a list of knotes, activating them if their event has triggered. 1760 * 1761 * If we encounter any knotes which are undergoing processing we just mark 1762 * them for reprocessing and do not try to [re]activate the knote. However, 1763 * if a hint is being passed we have to wait and that makes things a bit 1764 * sticky. 1765 */ 1766 void 1767 knote(struct klist *list, long hint) 1768 { 1769 struct kqueue *kq; 1770 struct knote *kn; 1771 struct knote *kntmp; 1772 1773 lwkt_getpooltoken(list); 1774 restart: 1775 SLIST_FOREACH(kn, list, kn_next) { 1776 kq = kn->kn_kq; 1777 lwkt_getpooltoken(kq); 1778 1779 /* temporary verification hack */ 1780 SLIST_FOREACH(kntmp, list, kn_next) { 1781 if (kn == kntmp) 1782 break; 1783 } 1784 if (kn != kntmp || kn->kn_kq != kq) { 1785 lwkt_relpooltoken(kq); 1786 goto restart; 1787 } 1788 1789 if (kn->kn_status & KN_PROCESSING) { 1790 /* 1791 * Someone else is processing the knote, ask the 1792 * other thread to reprocess it and don't mess 1793 * with it otherwise. 1794 */ 1795 if (hint == 0) { 1796 kn->kn_status |= KN_REPROCESS; 1797 lwkt_relpooltoken(kq); 1798 continue; 1799 } 1800 1801 /* 1802 * If the hint is non-zero we have to wait or risk 1803 * losing the state the caller is trying to update. 1804 * 1805 * XXX This is a real problem, certain process 1806 * and signal filters will bump kn_data for 1807 * already-processed notes more than once if 1808 * we restart the list scan. FIXME. 1809 */ 1810 kn->kn_status |= KN_WAITING | KN_REPROCESS; 1811 tsleep(kn, 0, "knotec", hz); 1812 lwkt_relpooltoken(kq); 1813 goto restart; 1814 } 1815 1816 /* 1817 * Become the reprocessing master ourselves. 1818 * 1819 * If hint is non-zero running the event is mandatory 1820 * when not deleting so do it whether reprocessing is 1821 * set or not. 1822 */ 1823 kn->kn_status |= KN_PROCESSING; 1824 if ((kn->kn_status & KN_DELETING) == 0) { 1825 if (filter_event(kn, hint)) 1826 KNOTE_ACTIVATE(kn); 1827 } 1828 if (knote_release(kn)) { 1829 lwkt_relpooltoken(kq); 1830 goto restart; 1831 } 1832 lwkt_relpooltoken(kq); 1833 } 1834 lwkt_relpooltoken(list); 1835 } 1836 1837 /* 1838 * Insert knote at head of klist. 1839 * 1840 * This function may only be called via a filter function and thus 1841 * kq_token should already be held and marked for processing. 1842 */ 1843 void 1844 knote_insert(struct klist *klist, struct knote *kn) 1845 { 1846 lwkt_getpooltoken(klist); 1847 KKASSERT(kn->kn_status & KN_PROCESSING); 1848 SLIST_INSERT_HEAD(klist, kn, kn_next); 1849 lwkt_relpooltoken(klist); 1850 } 1851 1852 /* 1853 * Remove knote from a klist 1854 * 1855 * This function may only be called via a filter function and thus 1856 * kq_token should already be held and marked for processing. 1857 */ 1858 void 1859 knote_remove(struct klist *klist, struct knote *kn) 1860 { 1861 lwkt_getpooltoken(klist); 1862 KKASSERT(kn->kn_status & KN_PROCESSING); 1863 SLIST_REMOVE(klist, kn, knote, kn_next); 1864 lwkt_relpooltoken(klist); 1865 } 1866 1867 void 1868 knote_assume_knotes(struct kqinfo *src, struct kqinfo *dst, 1869 struct filterops *ops, void *hook) 1870 { 1871 struct kqueue *kq; 1872 struct knote *kn; 1873 1874 lwkt_getpooltoken(&src->ki_note); 1875 lwkt_getpooltoken(&dst->ki_note); 1876 while ((kn = SLIST_FIRST(&src->ki_note)) != NULL) { 1877 kq = kn->kn_kq; 1878 lwkt_getpooltoken(kq); 1879 if (SLIST_FIRST(&src->ki_note) != kn || kn->kn_kq != kq) { 1880 lwkt_relpooltoken(kq); 1881 continue; 1882 } 1883 if (knote_acquire(kn)) { 1884 knote_remove(&src->ki_note, kn); 1885 kn->kn_fop = ops; 1886 kn->kn_hook = hook; 1887 knote_insert(&dst->ki_note, kn); 1888 knote_release(kn); 1889 /* kn may be invalid now */ 1890 } 1891 lwkt_relpooltoken(kq); 1892 } 1893 lwkt_relpooltoken(&dst->ki_note); 1894 lwkt_relpooltoken(&src->ki_note); 1895 } 1896 1897 /* 1898 * Remove all knotes referencing a specified fd 1899 */ 1900 void 1901 knote_fdclose(struct file *fp, struct filedesc *fdp, int fd) 1902 { 1903 struct kqueue *kq; 1904 struct knote *kn; 1905 struct knote *kntmp; 1906 1907 lwkt_getpooltoken(&fp->f_klist); 1908 restart: 1909 SLIST_FOREACH(kn, &fp->f_klist, kn_link) { 1910 if (kn->kn_kq->kq_fdp == fdp && kn->kn_id == fd) { 1911 kq = kn->kn_kq; 1912 lwkt_getpooltoken(kq); 1913 1914 /* temporary verification hack */ 1915 SLIST_FOREACH(kntmp, &fp->f_klist, kn_link) { 1916 if (kn == kntmp) 1917 break; 1918 } 1919 if (kn != kntmp || kn->kn_kq->kq_fdp != fdp || 1920 kn->kn_id != fd || kn->kn_kq != kq) { 1921 lwkt_relpooltoken(kq); 1922 goto restart; 1923 } 1924 if (knote_acquire(kn)) 1925 knote_detach_and_drop(kn); 1926 lwkt_relpooltoken(kq); 1927 goto restart; 1928 } 1929 } 1930 lwkt_relpooltoken(&fp->f_klist); 1931 } 1932 1933 /* 1934 * Low level attach function. 1935 * 1936 * The knote should already be marked for processing. 1937 * Caller must hold the related kq token. 1938 */ 1939 static void 1940 knote_attach(struct knote *kn) 1941 { 1942 struct klist *list; 1943 struct kqueue *kq = kn->kn_kq; 1944 1945 if (kn->kn_fop->f_flags & FILTEROP_ISFD) { 1946 KKASSERT(kn->kn_fp); 1947 list = &kn->kn_fp->f_klist; 1948 } else { 1949 if (kq->kq_knhashmask == 0) 1950 kq->kq_knhash = hashinit(KN_HASHSIZE, M_KQUEUE, 1951 &kq->kq_knhashmask); 1952 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)]; 1953 } 1954 lwkt_getpooltoken(list); 1955 SLIST_INSERT_HEAD(list, kn, kn_link); 1956 lwkt_relpooltoken(list); 1957 TAILQ_INSERT_HEAD(&kq->kq_knlist, kn, kn_kqlink); 1958 } 1959 1960 /* 1961 * Low level drop function. 1962 * 1963 * The knote should already be marked for processing. 1964 * Caller must hold the related kq token. 1965 */ 1966 static void 1967 knote_drop(struct knote *kn) 1968 { 1969 struct kqueue *kq; 1970 struct klist *list; 1971 1972 kq = kn->kn_kq; 1973 1974 if (kn->kn_fop->f_flags & FILTEROP_ISFD) 1975 list = &kn->kn_fp->f_klist; 1976 else 1977 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)]; 1978 1979 lwkt_getpooltoken(list); 1980 SLIST_REMOVE(list, kn, knote, kn_link); 1981 lwkt_relpooltoken(list); 1982 TAILQ_REMOVE(&kq->kq_knlist, kn, kn_kqlink); 1983 if (kn->kn_status & KN_QUEUED) 1984 knote_dequeue(kn); 1985 if (kn->kn_fop->f_flags & FILTEROP_ISFD) { 1986 fdrop(kn->kn_fp); 1987 kn->kn_fp = NULL; 1988 } 1989 knote_free(kn); 1990 } 1991 1992 /* 1993 * Low level enqueue function. 1994 * 1995 * The knote should already be marked for processing. 1996 * Caller must be holding the kq token 1997 */ 1998 static void 1999 knote_enqueue(struct knote *kn) 2000 { 2001 struct kqueue *kq = kn->kn_kq; 2002 2003 KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued")); 2004 TAILQ_INSERT_TAIL(&kq->kq_knpend, kn, kn_tqe); 2005 kn->kn_status |= KN_QUEUED; 2006 ++kq->kq_count; 2007 2008 /* 2009 * Send SIGIO on request (typically set up as a mailbox signal) 2010 */ 2011 if (kq->kq_sigio && (kq->kq_state & KQ_ASYNC) && kq->kq_count == 1) 2012 pgsigio(kq->kq_sigio, SIGIO, 0); 2013 2014 kqueue_wakeup(kq); 2015 } 2016 2017 /* 2018 * Low level dequeue function. 2019 * 2020 * The knote should already be marked for processing. 2021 * Caller must be holding the kq token 2022 */ 2023 static void 2024 knote_dequeue(struct knote *kn) 2025 { 2026 struct kqueue *kq = kn->kn_kq; 2027 2028 KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued")); 2029 TAILQ_REMOVE(&kq->kq_knpend, kn, kn_tqe); 2030 kn->kn_status &= ~KN_QUEUED; 2031 kq->kq_count--; 2032 } 2033 2034 static struct knote * 2035 knote_alloc(void) 2036 { 2037 return kmalloc(sizeof(struct knote), M_KQUEUE, M_WAITOK); 2038 } 2039 2040 static void 2041 knote_free(struct knote *kn) 2042 { 2043 struct knote_cache_list *cache_list; 2044 2045 cache_list = &knote_cache_lists[mycpuid]; 2046 if (cache_list->knote_cache_cnt < KNOTE_CACHE_MAX) { 2047 crit_enter(); 2048 SLIST_INSERT_HEAD(&cache_list->knote_cache, kn, kn_link); 2049 cache_list->knote_cache_cnt++; 2050 crit_exit(); 2051 return; 2052 } 2053 kfree(kn, M_KQUEUE); 2054 } 2055 2056 struct sleepinfo { 2057 void *ident; 2058 int timedout; 2059 }; 2060 2061 static void 2062 precise_sleep_intr(systimer_t info, int in_ipi, struct intrframe *frame) 2063 { 2064 struct sleepinfo *si; 2065 2066 si = info->data; 2067 si->timedout = 1; 2068 wakeup(si->ident); 2069 } 2070 2071 static int 2072 precise_sleep(void *ident, int flags, const char *wmesg, int us) 2073 { 2074 struct systimer info; 2075 struct sleepinfo si = { 2076 .ident = ident, 2077 .timedout = 0, 2078 }; 2079 int r; 2080 2081 tsleep_interlock(ident, flags); 2082 systimer_init_oneshot(&info, precise_sleep_intr, &si, us); 2083 r = tsleep(ident, flags | PINTERLOCKED, wmesg, 0); 2084 systimer_del(&info); 2085 if (si.timedout) 2086 r = EWOULDBLOCK; 2087 2088 return r; 2089 } 2090