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 = 65536; 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_stop() 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 * 502 * NOTE: Even though the note is locked via KN_PROCSESING, filt_timerexpire() 503 * can still race us requeue the callout due to potential token cycling 504 * from various blocking conditions. If this situation arises, 505 * callout_stop_sync() will always return non-zero and we can simply 506 * retry the operation. 507 */ 508 static void 509 filt_timerdetach(struct knote *kn) 510 { 511 struct callout *calloutp; 512 513 calloutp = (struct callout *)kn->kn_hook; 514 while (callout_stop_sync(calloutp)) { 515 kprintf("debug: kqueue timer race fixed, pid %d %s\n", 516 (curthread->td_proc ? curthread->td_proc->p_pid : 0), 517 curthread->td_comm); 518 } 519 kn->kn_hook = NULL; 520 kfree(calloutp, M_KQUEUE); 521 atomic_subtract_int(&kq_ncallouts, 1); 522 } 523 524 static int 525 filt_timer(struct knote *kn, long hint) 526 { 527 return (kn->kn_data != 0); 528 } 529 530 /* 531 * EVFILT_USER 532 */ 533 static int 534 filt_userattach(struct knote *kn) 535 { 536 u_int ffctrl; 537 538 kn->kn_hook = NULL; 539 if (kn->kn_sfflags & NOTE_TRIGGER) 540 kn->kn_ptr.hookid = 1; 541 else 542 kn->kn_ptr.hookid = 0; 543 544 ffctrl = kn->kn_sfflags & NOTE_FFCTRLMASK; 545 kn->kn_sfflags &= NOTE_FFLAGSMASK; 546 switch (ffctrl) { 547 case NOTE_FFNOP: 548 break; 549 550 case NOTE_FFAND: 551 kn->kn_fflags &= kn->kn_sfflags; 552 break; 553 554 case NOTE_FFOR: 555 kn->kn_fflags |= kn->kn_sfflags; 556 break; 557 558 case NOTE_FFCOPY: 559 kn->kn_fflags = kn->kn_sfflags; 560 break; 561 562 default: 563 /* XXX Return error? */ 564 break; 565 } 566 /* We just happen to copy this value as well. Undocumented. */ 567 kn->kn_data = kn->kn_sdata; 568 569 return 0; 570 } 571 572 static void 573 filt_userdetach(struct knote *kn) 574 { 575 /* nothing to do */ 576 } 577 578 static int 579 filt_user(struct knote *kn, long hint) 580 { 581 return (kn->kn_ptr.hookid); 582 } 583 584 static void 585 filt_usertouch(struct knote *kn, struct kevent *kev, u_long type) 586 { 587 u_int ffctrl; 588 589 switch (type) { 590 case EVENT_REGISTER: 591 if (kev->fflags & NOTE_TRIGGER) 592 kn->kn_ptr.hookid = 1; 593 594 ffctrl = kev->fflags & NOTE_FFCTRLMASK; 595 kev->fflags &= NOTE_FFLAGSMASK; 596 switch (ffctrl) { 597 case NOTE_FFNOP: 598 break; 599 600 case NOTE_FFAND: 601 kn->kn_fflags &= kev->fflags; 602 break; 603 604 case NOTE_FFOR: 605 kn->kn_fflags |= kev->fflags; 606 break; 607 608 case NOTE_FFCOPY: 609 kn->kn_fflags = kev->fflags; 610 break; 611 612 default: 613 /* XXX Return error? */ 614 break; 615 } 616 /* We just happen to copy this value as well. Undocumented. */ 617 kn->kn_data = kev->data; 618 619 /* 620 * This is not the correct use of EV_CLEAR in an event 621 * modification, it should have been passed as a NOTE instead. 622 * But we need to maintain compatibility with Apple & FreeBSD. 623 * 624 * Note however that EV_CLEAR can still be used when doing 625 * the initial registration of the event and works as expected 626 * (clears the event on reception). 627 */ 628 if (kev->flags & EV_CLEAR) { 629 kn->kn_ptr.hookid = 0; 630 /* 631 * Clearing kn->kn_data is fine, since it gets set 632 * every time anyway. We just shouldn't clear 633 * kn->kn_fflags here, since that would limit the 634 * possible uses of this API. NOTE_FFAND or 635 * NOTE_FFCOPY should be used for explicitly clearing 636 * kn->kn_fflags. 637 */ 638 kn->kn_data = 0; 639 } 640 break; 641 642 case EVENT_PROCESS: 643 *kev = kn->kn_kevent; 644 kev->fflags = kn->kn_fflags; 645 kev->data = kn->kn_data; 646 if (kn->kn_flags & EV_CLEAR) { 647 kn->kn_ptr.hookid = 0; 648 /* kn_data, kn_fflags handled by parent */ 649 } 650 break; 651 652 default: 653 panic("filt_usertouch() - invalid type (%ld)", type); 654 break; 655 } 656 } 657 658 /* 659 * EVFILT_FS 660 */ 661 struct klist fs_klist = SLIST_HEAD_INITIALIZER(&fs_klist); 662 663 static int 664 filt_fsattach(struct knote *kn) 665 { 666 kn->kn_flags |= EV_CLEAR; 667 knote_insert(&fs_klist, kn); 668 669 return (0); 670 } 671 672 static void 673 filt_fsdetach(struct knote *kn) 674 { 675 knote_remove(&fs_klist, kn); 676 } 677 678 static int 679 filt_fs(struct knote *kn, long hint) 680 { 681 kn->kn_fflags |= hint; 682 return (kn->kn_fflags != 0); 683 } 684 685 /* 686 * Initialize a kqueue. 687 * 688 * NOTE: The lwp/proc code initializes a kqueue for select/poll ops. 689 * 690 * MPSAFE 691 */ 692 void 693 kqueue_init(struct kqueue *kq, struct filedesc *fdp) 694 { 695 TAILQ_INIT(&kq->kq_knpend); 696 TAILQ_INIT(&kq->kq_knlist); 697 kq->kq_count = 0; 698 kq->kq_fdp = fdp; 699 SLIST_INIT(&kq->kq_kqinfo.ki_note); 700 } 701 702 /* 703 * Terminate a kqueue. Freeing the actual kq itself is left up to the 704 * caller (it might be embedded in a lwp so we don't do it here). 705 * 706 * The kq's knlist must be completely eradicated so block on any 707 * processing races. 708 */ 709 void 710 kqueue_terminate(struct kqueue *kq) 711 { 712 struct knote *kn; 713 714 lwkt_getpooltoken(kq); 715 while ((kn = TAILQ_FIRST(&kq->kq_knlist)) != NULL) { 716 if (knote_acquire(kn)) 717 knote_detach_and_drop(kn); 718 } 719 lwkt_relpooltoken(kq); 720 721 if (kq->kq_knhash) { 722 hashdestroy(kq->kq_knhash, M_KQUEUE, kq->kq_knhashmask); 723 kq->kq_knhash = NULL; 724 kq->kq_knhashmask = 0; 725 } 726 } 727 728 /* 729 * MPSAFE 730 */ 731 int 732 sys_kqueue(struct kqueue_args *uap) 733 { 734 struct thread *td = curthread; 735 struct kqueue *kq; 736 struct file *fp; 737 int fd, error; 738 739 error = falloc(td->td_lwp, &fp, &fd); 740 if (error) 741 return (error); 742 fp->f_flag = FREAD | FWRITE; 743 fp->f_type = DTYPE_KQUEUE; 744 fp->f_ops = &kqueueops; 745 746 kq = kmalloc(sizeof(struct kqueue), M_KQUEUE, M_WAITOK | M_ZERO); 747 kqueue_init(kq, td->td_proc->p_fd); 748 fp->f_data = kq; 749 750 fsetfd(kq->kq_fdp, fp, fd); 751 uap->sysmsg_result = fd; 752 fdrop(fp); 753 return (error); 754 } 755 756 /* 757 * Copy 'count' items into the destination list pointed to by uap->eventlist. 758 */ 759 static int 760 kevent_copyout(void *arg, struct kevent *kevp, int count, int *res) 761 { 762 struct kevent_copyin_args *kap; 763 int error; 764 765 kap = (struct kevent_copyin_args *)arg; 766 767 error = copyout(kevp, kap->ka->eventlist, count * sizeof(*kevp)); 768 if (error == 0) { 769 kap->ka->eventlist += count; 770 *res += count; 771 } else { 772 *res = -1; 773 } 774 775 return (error); 776 } 777 778 /* 779 * Copy at most 'max' items from the list pointed to by kap->changelist, 780 * return number of items in 'events'. 781 */ 782 static int 783 kevent_copyin(void *arg, struct kevent *kevp, int max, int *events) 784 { 785 struct kevent_copyin_args *kap; 786 int error, count; 787 788 kap = (struct kevent_copyin_args *)arg; 789 790 count = min(kap->ka->nchanges - kap->pchanges, max); 791 error = copyin(kap->ka->changelist, kevp, count * sizeof *kevp); 792 if (error == 0) { 793 kap->ka->changelist += count; 794 kap->pchanges += count; 795 *events = count; 796 } 797 798 return (error); 799 } 800 801 /* 802 * MPSAFE 803 */ 804 int 805 kern_kevent(struct kqueue *kq, int nevents, int *res, void *uap, 806 k_copyin_fn kevent_copyinfn, k_copyout_fn kevent_copyoutfn, 807 struct timespec *tsp_in, int flags) 808 { 809 struct kevent *kevp; 810 struct timespec *tsp, ats; 811 int i, n, total, error, nerrors = 0; 812 int gobbled; 813 int lres; 814 int limit = kq_checkloop; 815 int closedcounter; 816 struct kevent kev[KQ_NEVENTS]; 817 struct knote marker; 818 struct lwkt_token *tok; 819 820 if (tsp_in == NULL || tsp_in->tv_sec || tsp_in->tv_nsec) 821 atomic_set_int(&curthread->td_mpflags, TDF_MP_BATCH_DEMARC); 822 823 tsp = tsp_in; 824 *res = 0; 825 826 closedcounter = kq->kq_fdp->fd_closedcounter; 827 828 for (;;) { 829 n = 0; 830 error = kevent_copyinfn(uap, kev, KQ_NEVENTS, &n); 831 if (error) 832 return error; 833 if (n == 0) 834 break; 835 for (i = 0; i < n; ++i) 836 kev[i].flags &= ~EV_SYSFLAGS; 837 for (i = 0; i < n; ++i) { 838 gobbled = n - i; 839 error = kqueue_register(kq, &kev[i], &gobbled); 840 i += gobbled - 1; 841 kevp = &kev[i]; 842 843 /* 844 * If a registration returns an error we 845 * immediately post the error. The kevent() 846 * call itself will fail with the error if 847 * no space is available for posting. 848 * 849 * Such errors normally bypass the timeout/blocking 850 * code. However, if the copyoutfn function refuses 851 * to post the error (see sys_poll()), then we 852 * ignore it too. 853 */ 854 if (error || (kevp->flags & EV_RECEIPT)) { 855 kevp->flags = EV_ERROR; 856 kevp->data = error; 857 lres = *res; 858 kevent_copyoutfn(uap, kevp, 1, res); 859 if (*res < 0) { 860 return error; 861 } else if (lres != *res) { 862 nevents--; 863 nerrors++; 864 } 865 } 866 } 867 } 868 if (nerrors) 869 return 0; 870 871 /* 872 * Acquire/wait for events - setup timeout 873 */ 874 if (tsp != NULL) { 875 if (tsp->tv_sec || tsp->tv_nsec) { 876 getnanouptime(&ats); 877 timespecadd(tsp, &ats); /* tsp = target time */ 878 } 879 } 880 881 /* 882 * Loop as required. 883 * 884 * Collect as many events as we can. Sleeping on successive 885 * loops is disabled if copyoutfn has incremented (*res). 886 * 887 * The loop stops if an error occurs, all events have been 888 * scanned (the marker has been reached), or fewer than the 889 * maximum number of events is found. 890 * 891 * The copyoutfn function does not have to increment (*res) in 892 * order for the loop to continue. 893 * 894 * NOTE: doselect() usually passes 0x7FFFFFFF for nevents. 895 */ 896 total = 0; 897 error = 0; 898 marker.kn_filter = EVFILT_MARKER; 899 marker.kn_status = KN_PROCESSING; 900 tok = lwkt_token_pool_lookup(kq); 901 lwkt_gettoken(tok); 902 TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker, kn_tqe); 903 lwkt_reltoken(tok); 904 while ((n = nevents - total) > 0) { 905 if (n > KQ_NEVENTS) 906 n = KQ_NEVENTS; 907 908 /* 909 * If no events are pending sleep until timeout (if any) 910 * or an event occurs. 911 * 912 * After the sleep completes the marker is moved to the 913 * end of the list, making any received events available 914 * to our scan. 915 */ 916 if (kq->kq_count == 0 && *res == 0) { 917 int timeout, ustimeout = 0; 918 919 if (tsp == NULL) { 920 timeout = 0; 921 } else if (tsp->tv_sec == 0 && tsp->tv_nsec == 0) { 922 error = EWOULDBLOCK; 923 break; 924 } else { 925 struct timespec atx = *tsp; 926 927 getnanouptime(&ats); 928 timespecsub(&atx, &ats); 929 if (atx.tv_sec < 0) { 930 error = EWOULDBLOCK; 931 break; 932 } else { 933 timeout = atx.tv_sec > 24 * 60 * 60 ? 934 24 * 60 * 60 * hz : 935 tstohz_high(&atx); 936 } 937 if (flags & KEVENT_TIMEOUT_PRECISE && 938 timeout != 0) { 939 if (atx.tv_sec == 0 && 940 atx.tv_nsec < kq_sleep_threshold) { 941 DELAY(atx.tv_nsec / 1000); 942 error = EWOULDBLOCK; 943 break; 944 } else if (atx.tv_sec < 2000) { 945 ustimeout = atx.tv_sec * 946 1000000 + atx.tv_nsec/1000; 947 } else { 948 ustimeout = 2000000000; 949 } 950 } 951 } 952 953 lwkt_gettoken(tok); 954 if (kq->kq_count == 0) { 955 kq->kq_sleep_cnt++; 956 if (__predict_false(kq->kq_sleep_cnt == 0)) { 957 /* 958 * Guard against possible wrapping. And 959 * set it to 2, so that kqueue_wakeup() 960 * can wake everyone up. 961 */ 962 kq->kq_sleep_cnt = 2; 963 } 964 if ((flags & KEVENT_TIMEOUT_PRECISE) && 965 timeout != 0) { 966 error = precise_sleep(kq, PCATCH, 967 "kqread", ustimeout); 968 } else { 969 error = tsleep(kq, PCATCH, "kqread", 970 timeout); 971 } 972 973 /* don't restart after signals... */ 974 if (error == ERESTART) 975 error = EINTR; 976 if (error) { 977 lwkt_reltoken(tok); 978 break; 979 } 980 981 TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe); 982 TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker, 983 kn_tqe); 984 } 985 lwkt_reltoken(tok); 986 } 987 988 /* 989 * Process all received events 990 * Account for all non-spurious events in our total 991 */ 992 i = kqueue_scan(kq, kev, n, &marker, closedcounter); 993 if (i) { 994 lres = *res; 995 error = kevent_copyoutfn(uap, kev, i, res); 996 total += *res - lres; 997 if (error) 998 break; 999 } 1000 if (limit && --limit == 0) 1001 panic("kqueue: checkloop failed i=%d", i); 1002 1003 /* 1004 * Normally when fewer events are returned than requested 1005 * we can stop. However, if only spurious events were 1006 * collected the copyout will not bump (*res) and we have 1007 * to continue. 1008 */ 1009 if (i < n && *res) 1010 break; 1011 1012 /* 1013 * Deal with an edge case where spurious events can cause 1014 * a loop to occur without moving the marker. This can 1015 * prevent kqueue_scan() from picking up new events which 1016 * race us. We must be sure to move the marker for this 1017 * case. 1018 * 1019 * NOTE: We do not want to move the marker if events 1020 * were scanned because normal kqueue operations 1021 * may reactivate events. Moving the marker in 1022 * that case could result in duplicates for the 1023 * same event. 1024 */ 1025 if (i == 0) { 1026 lwkt_gettoken(tok); 1027 TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe); 1028 TAILQ_INSERT_TAIL(&kq->kq_knpend, &marker, kn_tqe); 1029 lwkt_reltoken(tok); 1030 } 1031 } 1032 lwkt_gettoken(tok); 1033 TAILQ_REMOVE(&kq->kq_knpend, &marker, kn_tqe); 1034 lwkt_reltoken(tok); 1035 1036 /* Timeouts do not return EWOULDBLOCK. */ 1037 if (error == EWOULDBLOCK) 1038 error = 0; 1039 return error; 1040 } 1041 1042 /* 1043 * MPALMOSTSAFE 1044 */ 1045 int 1046 sys_kevent(struct kevent_args *uap) 1047 { 1048 struct thread *td = curthread; 1049 struct timespec ts, *tsp; 1050 struct kqueue *kq; 1051 struct file *fp = NULL; 1052 struct kevent_copyin_args *kap, ka; 1053 int error; 1054 1055 if (uap->timeout) { 1056 error = copyin(uap->timeout, &ts, sizeof(ts)); 1057 if (error) 1058 return (error); 1059 tsp = &ts; 1060 } else { 1061 tsp = NULL; 1062 } 1063 fp = holdfp(td, uap->fd, -1); 1064 if (fp == NULL) 1065 return (EBADF); 1066 if (fp->f_type != DTYPE_KQUEUE) { 1067 fdrop(fp); 1068 return (EBADF); 1069 } 1070 1071 kq = (struct kqueue *)fp->f_data; 1072 1073 kap = &ka; 1074 kap->ka = uap; 1075 kap->pchanges = 0; 1076 1077 error = kern_kevent(kq, uap->nevents, &uap->sysmsg_result, kap, 1078 kevent_copyin, kevent_copyout, tsp, 0); 1079 1080 dropfp(td, uap->fd, fp); 1081 1082 return (error); 1083 } 1084 1085 /* 1086 * Efficiently load multiple file pointers. This significantly reduces 1087 * threaded overhead. When doing simple polling we can depend on the 1088 * per-thread (fd,fp) cache. With more descriptors, we batch. 1089 */ 1090 static 1091 void 1092 floadkevfps(thread_t td, struct filedesc *fdp, struct kevent *kev, 1093 struct file **fp, int climit) 1094 { 1095 struct filterops *fops; 1096 int tdcache; 1097 1098 if (climit <= 2 && td->td_proc && td->td_proc->p_fd == fdp) { 1099 tdcache = 1; 1100 } else { 1101 tdcache = 0; 1102 spin_lock_shared(&fdp->fd_spin); 1103 } 1104 1105 while (climit) { 1106 *fp = NULL; 1107 if (kev->filter < 0 && 1108 kev->filter + EVFILT_SYSCOUNT >= 0) { 1109 fops = sysfilt_ops[~kev->filter]; 1110 if (fops->f_flags & FILTEROP_ISFD) { 1111 if (tdcache) { 1112 *fp = holdfp(td, kev->ident, -1); 1113 } else { 1114 *fp = holdfp_fdp_locked(fdp, 1115 kev->ident, -1); 1116 } 1117 } 1118 } 1119 --climit; 1120 ++fp; 1121 ++kev; 1122 } 1123 if (tdcache == 0) 1124 spin_unlock_shared(&fdp->fd_spin); 1125 } 1126 1127 /* 1128 * Register up to *countp kev's. Always registers at least 1. 1129 * 1130 * The number registered is returned in *countp. 1131 * 1132 * If an error occurs or a kev is flagged EV_RECEIPT, it is 1133 * processed and included in *countp, and processing then 1134 * stops. 1135 */ 1136 int 1137 kqueue_register(struct kqueue *kq, struct kevent *kev, int *countp) 1138 { 1139 struct filedesc *fdp = kq->kq_fdp; 1140 struct klist *list = NULL; 1141 struct filterops *fops; 1142 struct file *fp[KQ_NEVENTS]; 1143 struct knote *kn = NULL; 1144 struct thread *td; 1145 int error; 1146 int count; 1147 int climit; 1148 int closedcounter; 1149 struct knote_cache_list *cache_list; 1150 1151 td = curthread; 1152 climit = *countp; 1153 if (climit > KQ_NEVENTS) 1154 climit = KQ_NEVENTS; 1155 closedcounter = fdp->fd_closedcounter; 1156 floadkevfps(td, fdp, kev, fp, climit); 1157 1158 lwkt_getpooltoken(kq); 1159 count = 0; 1160 1161 /* 1162 * To avoid races, only one thread can register events on this 1163 * kqueue at a time. 1164 */ 1165 while (__predict_false(kq->kq_regtd != NULL && kq->kq_regtd != td)) { 1166 kq->kq_state |= KQ_REGWAIT; 1167 tsleep(&kq->kq_regtd, 0, "kqreg", 0); 1168 } 1169 if (__predict_false(kq->kq_regtd != NULL)) { 1170 /* Recursive calling of kqueue_register() */ 1171 td = NULL; 1172 } else { 1173 /* Owner of the kq_regtd, i.e. td != NULL */ 1174 kq->kq_regtd = td; 1175 } 1176 1177 loop: 1178 if (kev->filter < 0) { 1179 if (kev->filter + EVFILT_SYSCOUNT < 0) { 1180 error = EINVAL; 1181 ++count; 1182 goto done; 1183 } 1184 fops = sysfilt_ops[~kev->filter]; /* to 0-base index */ 1185 } else { 1186 /* 1187 * XXX 1188 * filter attach routine is responsible for insuring that 1189 * the identifier can be attached to it. 1190 */ 1191 error = EINVAL; 1192 ++count; 1193 goto done; 1194 } 1195 1196 if (fops->f_flags & FILTEROP_ISFD) { 1197 /* validate descriptor */ 1198 if (fp[count] == NULL) { 1199 error = EBADF; 1200 ++count; 1201 goto done; 1202 } 1203 } 1204 1205 cache_list = &knote_cache_lists[mycpuid]; 1206 if (SLIST_EMPTY(&cache_list->knote_cache)) { 1207 struct knote *new_kn; 1208 1209 new_kn = knote_alloc(); 1210 crit_enter(); 1211 SLIST_INSERT_HEAD(&cache_list->knote_cache, new_kn, kn_link); 1212 cache_list->knote_cache_cnt++; 1213 crit_exit(); 1214 } 1215 1216 if (fp[count] != NULL) { 1217 list = &fp[count]->f_klist; 1218 } else if (kq->kq_knhashmask) { 1219 list = &kq->kq_knhash[ 1220 KN_HASH((u_long)kev->ident, kq->kq_knhashmask)]; 1221 } 1222 if (list != NULL) { 1223 lwkt_getpooltoken(list); 1224 again: 1225 SLIST_FOREACH(kn, list, kn_link) { 1226 if (kn->kn_kq == kq && 1227 kn->kn_filter == kev->filter && 1228 kn->kn_id == kev->ident) { 1229 if (knote_acquire(kn) == 0) 1230 goto again; 1231 break; 1232 } 1233 } 1234 lwkt_relpooltoken(list); 1235 } 1236 1237 /* 1238 * NOTE: At this point if kn is non-NULL we will have acquired 1239 * it and set KN_PROCESSING. 1240 */ 1241 if (kn == NULL && ((kev->flags & EV_ADD) == 0)) { 1242 error = ENOENT; 1243 ++count; 1244 goto done; 1245 } 1246 1247 /* 1248 * kn now contains the matching knote, or NULL if no match 1249 */ 1250 if (kev->flags & EV_ADD) { 1251 if (kn == NULL) { 1252 crit_enter(); 1253 kn = SLIST_FIRST(&cache_list->knote_cache); 1254 if (kn == NULL) { 1255 crit_exit(); 1256 kn = knote_alloc(); 1257 } else { 1258 SLIST_REMOVE_HEAD(&cache_list->knote_cache, 1259 kn_link); 1260 cache_list->knote_cache_cnt--; 1261 crit_exit(); 1262 } 1263 kn->kn_fp = fp[count]; 1264 kn->kn_kq = kq; 1265 kn->kn_fop = fops; 1266 1267 /* 1268 * apply reference count to knote structure, and 1269 * do not release it at the end of this routine. 1270 */ 1271 fp[count] = NULL; /* safety */ 1272 1273 kn->kn_sfflags = kev->fflags; 1274 kn->kn_sdata = kev->data; 1275 kev->fflags = 0; 1276 kev->data = 0; 1277 kn->kn_kevent = *kev; 1278 1279 /* 1280 * KN_PROCESSING prevents the knote from getting 1281 * ripped out from under us while we are trying 1282 * to attach it, in case the attach blocks. 1283 */ 1284 kn->kn_status = KN_PROCESSING; 1285 knote_attach(kn); 1286 if ((error = filter_attach(kn)) != 0) { 1287 kn->kn_status |= KN_DELETING | KN_REPROCESS; 1288 knote_drop(kn); 1289 ++count; 1290 goto done; 1291 } 1292 1293 /* 1294 * Interlock against close races which either tried 1295 * to remove our knote while we were blocked or missed 1296 * it entirely prior to our attachment. We do not 1297 * want to end up with a knote on a closed descriptor. 1298 */ 1299 if ((fops->f_flags & FILTEROP_ISFD) && 1300 checkfdclosed(curthread, fdp, kev->ident, kn->kn_fp, 1301 closedcounter)) { 1302 kn->kn_status |= KN_DELETING | KN_REPROCESS; 1303 } 1304 } else { 1305 /* 1306 * The user may change some filter values after the 1307 * initial EV_ADD, but doing so will not reset any 1308 * filter which have already been triggered. 1309 */ 1310 KKASSERT(kn->kn_status & KN_PROCESSING); 1311 if (fops == &user_filtops) { 1312 filt_usertouch(kn, kev, EVENT_REGISTER); 1313 } else { 1314 kn->kn_sfflags = kev->fflags; 1315 kn->kn_sdata = kev->data; 1316 kn->kn_kevent.udata = kev->udata; 1317 } 1318 } 1319 1320 /* 1321 * Execute the filter event to immediately activate the 1322 * knote if necessary. If reprocessing events are pending 1323 * due to blocking above we do not run the filter here 1324 * but instead let knote_release() do it. Otherwise we 1325 * might run the filter on a deleted event. 1326 */ 1327 if ((kn->kn_status & KN_REPROCESS) == 0) { 1328 if (filter_event(kn, 0)) 1329 KNOTE_ACTIVATE(kn); 1330 } 1331 } else if (kev->flags & EV_DELETE) { 1332 /* 1333 * Delete the existing knote 1334 */ 1335 knote_detach_and_drop(kn); 1336 error = 0; 1337 ++count; 1338 goto done; 1339 } else { 1340 /* 1341 * Modify an existing event. 1342 * 1343 * The user may change some filter values after the 1344 * initial EV_ADD, but doing so will not reset any 1345 * filter which have already been triggered. 1346 */ 1347 KKASSERT(kn->kn_status & KN_PROCESSING); 1348 if (fops == &user_filtops) { 1349 filt_usertouch(kn, kev, EVENT_REGISTER); 1350 } else { 1351 kn->kn_sfflags = kev->fflags; 1352 kn->kn_sdata = kev->data; 1353 kn->kn_kevent.udata = kev->udata; 1354 } 1355 1356 /* 1357 * Execute the filter event to immediately activate the 1358 * knote if necessary. If reprocessing events are pending 1359 * due to blocking above we do not run the filter here 1360 * but instead let knote_release() do it. Otherwise we 1361 * might run the filter on a deleted event. 1362 */ 1363 if ((kn->kn_status & KN_REPROCESS) == 0) { 1364 if (filter_event(kn, 0)) 1365 KNOTE_ACTIVATE(kn); 1366 } 1367 } 1368 1369 /* 1370 * Disablement does not deactivate a knote here. 1371 */ 1372 if ((kev->flags & EV_DISABLE) && 1373 ((kn->kn_status & KN_DISABLED) == 0)) { 1374 kn->kn_status |= KN_DISABLED; 1375 } 1376 1377 /* 1378 * Re-enablement may have to immediately enqueue an active knote. 1379 */ 1380 if ((kev->flags & EV_ENABLE) && (kn->kn_status & KN_DISABLED)) { 1381 kn->kn_status &= ~KN_DISABLED; 1382 if ((kn->kn_status & KN_ACTIVE) && 1383 ((kn->kn_status & KN_QUEUED) == 0)) { 1384 knote_enqueue(kn); 1385 } 1386 } 1387 1388 /* 1389 * Handle any required reprocessing 1390 */ 1391 knote_release(kn); 1392 /* kn may be invalid now */ 1393 1394 /* 1395 * Loop control. We stop on errors (above), and also stop after 1396 * processing EV_RECEIPT, so the caller can process it. 1397 */ 1398 ++count; 1399 if (kev->flags & EV_RECEIPT) { 1400 error = 0; 1401 goto done; 1402 } 1403 ++kev; 1404 if (count < climit) { 1405 if (fp[count-1]) /* drop unprocessed fp */ 1406 fdrop(fp[count-1]); 1407 goto loop; 1408 } 1409 1410 /* 1411 * Cleanup 1412 */ 1413 done: 1414 if (td != NULL) { /* Owner of the kq_regtd */ 1415 kq->kq_regtd = NULL; 1416 if (__predict_false(kq->kq_state & KQ_REGWAIT)) { 1417 kq->kq_state &= ~KQ_REGWAIT; 1418 wakeup(&kq->kq_regtd); 1419 } 1420 } 1421 lwkt_relpooltoken(kq); 1422 1423 /* 1424 * Drop unprocessed file pointers 1425 */ 1426 *countp = count; 1427 if (count && fp[count-1]) 1428 fdrop(fp[count-1]); 1429 while (count < climit) { 1430 if (fp[count]) 1431 fdrop(fp[count]); 1432 ++count; 1433 } 1434 return (error); 1435 } 1436 1437 /* 1438 * Scan the kqueue, return the number of active events placed in kevp up 1439 * to count. 1440 * 1441 * Continuous mode events may get recycled, do not continue scanning past 1442 * marker unless no events have been collected. 1443 */ 1444 static int 1445 kqueue_scan(struct kqueue *kq, struct kevent *kevp, int count, 1446 struct knote *marker, int closedcounter) 1447 { 1448 struct knote *kn, local_marker; 1449 thread_t td = curthread; 1450 int total; 1451 1452 total = 0; 1453 local_marker.kn_filter = EVFILT_MARKER; 1454 local_marker.kn_status = KN_PROCESSING; 1455 1456 lwkt_getpooltoken(kq); 1457 1458 /* 1459 * Collect events. 1460 */ 1461 TAILQ_INSERT_HEAD(&kq->kq_knpend, &local_marker, kn_tqe); 1462 while (count) { 1463 kn = TAILQ_NEXT(&local_marker, kn_tqe); 1464 if (kn->kn_filter == EVFILT_MARKER) { 1465 /* Marker reached, we are done */ 1466 if (kn == marker) 1467 break; 1468 1469 /* Move local marker past some other threads marker */ 1470 kn = TAILQ_NEXT(kn, kn_tqe); 1471 TAILQ_REMOVE(&kq->kq_knpend, &local_marker, kn_tqe); 1472 TAILQ_INSERT_BEFORE(kn, &local_marker, kn_tqe); 1473 continue; 1474 } 1475 1476 /* 1477 * We can't skip a knote undergoing processing, otherwise 1478 * we risk not returning it when the user process expects 1479 * it should be returned. Sleep and retry. 1480 */ 1481 if (knote_acquire(kn) == 0) 1482 continue; 1483 1484 /* 1485 * Remove the event for processing. 1486 * 1487 * WARNING! We must leave KN_QUEUED set to prevent the 1488 * event from being KNOTE_ACTIVATE()d while 1489 * the queue state is in limbo, in case we 1490 * block. 1491 */ 1492 TAILQ_REMOVE(&kq->kq_knpend, kn, kn_tqe); 1493 kq->kq_count--; 1494 1495 /* 1496 * We have to deal with an extremely important race against 1497 * file descriptor close()s here. The file descriptor can 1498 * disappear MPSAFE, and there is a small window of 1499 * opportunity between that and the call to knote_fdclose(). 1500 * 1501 * If we hit that window here while doselect or dopoll is 1502 * trying to delete a spurious event they will not be able 1503 * to match up the event against a knote and will go haywire. 1504 */ 1505 if ((kn->kn_fop->f_flags & FILTEROP_ISFD) && 1506 checkfdclosed(td, kq->kq_fdp, kn->kn_kevent.ident, 1507 kn->kn_fp, closedcounter)) { 1508 kn->kn_status |= KN_DELETING | KN_REPROCESS; 1509 } 1510 1511 if (kn->kn_status & KN_DISABLED) { 1512 /* 1513 * If disabled we ensure the event is not queued 1514 * but leave its active bit set. On re-enablement 1515 * the event may be immediately triggered. 1516 */ 1517 kn->kn_status &= ~KN_QUEUED; 1518 } else if ((kn->kn_flags & EV_ONESHOT) == 0 && 1519 (kn->kn_status & KN_DELETING) == 0 && 1520 filter_event(kn, 0) == 0) { 1521 /* 1522 * If not running in one-shot mode and the event 1523 * is no longer present we ensure it is removed 1524 * from the queue and ignore it. 1525 */ 1526 kn->kn_status &= ~(KN_QUEUED | KN_ACTIVE); 1527 } else { 1528 /* 1529 * Post the event 1530 */ 1531 if (kn->kn_fop == &user_filtops) 1532 filt_usertouch(kn, kevp, EVENT_PROCESS); 1533 else 1534 *kevp = kn->kn_kevent; 1535 ++kevp; 1536 ++total; 1537 --count; 1538 1539 if (kn->kn_flags & EV_ONESHOT) { 1540 kn->kn_status &= ~KN_QUEUED; 1541 kn->kn_status |= KN_DELETING | KN_REPROCESS; 1542 } else { 1543 if (kn->kn_flags & (EV_CLEAR | EV_DISPATCH)) { 1544 if (kn->kn_flags & EV_CLEAR) { 1545 kn->kn_data = 0; 1546 kn->kn_fflags = 0; 1547 } 1548 if (kn->kn_flags & EV_DISPATCH) { 1549 kn->kn_status |= KN_DISABLED; 1550 } 1551 kn->kn_status &= ~(KN_QUEUED | 1552 KN_ACTIVE); 1553 } else { 1554 TAILQ_INSERT_TAIL(&kq->kq_knpend, kn, kn_tqe); 1555 kq->kq_count++; 1556 } 1557 } 1558 } 1559 1560 /* 1561 * Handle any post-processing states 1562 */ 1563 knote_release(kn); 1564 } 1565 TAILQ_REMOVE(&kq->kq_knpend, &local_marker, kn_tqe); 1566 1567 lwkt_relpooltoken(kq); 1568 return (total); 1569 } 1570 1571 /* 1572 * XXX 1573 * This could be expanded to call kqueue_scan, if desired. 1574 * 1575 * MPSAFE 1576 */ 1577 static int 1578 kqueue_read(struct file *fp, struct uio *uio, struct ucred *cred, int flags) 1579 { 1580 return (ENXIO); 1581 } 1582 1583 /* 1584 * MPSAFE 1585 */ 1586 static int 1587 kqueue_write(struct file *fp, struct uio *uio, struct ucred *cred, int flags) 1588 { 1589 return (ENXIO); 1590 } 1591 1592 /* 1593 * MPALMOSTSAFE 1594 */ 1595 static int 1596 kqueue_ioctl(struct file *fp, u_long com, caddr_t data, 1597 struct ucred *cred, struct sysmsg *msg) 1598 { 1599 struct kqueue *kq; 1600 int error; 1601 1602 kq = (struct kqueue *)fp->f_data; 1603 lwkt_getpooltoken(kq); 1604 switch(com) { 1605 case FIOASYNC: 1606 if (*(int *)data) 1607 kq->kq_state |= KQ_ASYNC; 1608 else 1609 kq->kq_state &= ~KQ_ASYNC; 1610 error = 0; 1611 break; 1612 case FIOSETOWN: 1613 error = fsetown(*(int *)data, &kq->kq_sigio); 1614 break; 1615 default: 1616 error = ENOTTY; 1617 break; 1618 } 1619 lwkt_relpooltoken(kq); 1620 return (error); 1621 } 1622 1623 /* 1624 * MPSAFE 1625 */ 1626 static int 1627 kqueue_stat(struct file *fp, struct stat *st, struct ucred *cred) 1628 { 1629 struct kqueue *kq = (struct kqueue *)fp->f_data; 1630 1631 bzero((void *)st, sizeof(*st)); 1632 st->st_size = kq->kq_count; 1633 st->st_blksize = sizeof(struct kevent); 1634 st->st_mode = S_IFIFO; 1635 return (0); 1636 } 1637 1638 /* 1639 * MPSAFE 1640 */ 1641 static int 1642 kqueue_close(struct file *fp) 1643 { 1644 struct kqueue *kq = (struct kqueue *)fp->f_data; 1645 1646 kqueue_terminate(kq); 1647 1648 fp->f_data = NULL; 1649 funsetown(&kq->kq_sigio); 1650 1651 kfree(kq, M_KQUEUE); 1652 return (0); 1653 } 1654 1655 static void 1656 kqueue_wakeup(struct kqueue *kq) 1657 { 1658 if (kq->kq_sleep_cnt) { 1659 u_int sleep_cnt = kq->kq_sleep_cnt; 1660 1661 kq->kq_sleep_cnt = 0; 1662 if (sleep_cnt == 1) 1663 wakeup_one(kq); 1664 else 1665 wakeup(kq); 1666 } 1667 KNOTE(&kq->kq_kqinfo.ki_note, 0); 1668 } 1669 1670 /* 1671 * Calls filterops f_attach function, acquiring mplock if filter is not 1672 * marked as FILTEROP_MPSAFE. 1673 * 1674 * Caller must be holding the related kq token 1675 */ 1676 static int 1677 filter_attach(struct knote *kn) 1678 { 1679 int ret; 1680 1681 if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) { 1682 ret = kn->kn_fop->f_attach(kn); 1683 } else { 1684 get_mplock(); 1685 ret = kn->kn_fop->f_attach(kn); 1686 rel_mplock(); 1687 } 1688 return (ret); 1689 } 1690 1691 /* 1692 * Detach the knote and drop it, destroying the knote. 1693 * 1694 * Calls filterops f_detach function, acquiring mplock if filter is not 1695 * marked as FILTEROP_MPSAFE. 1696 * 1697 * Caller must be holding the related kq token 1698 */ 1699 static void 1700 knote_detach_and_drop(struct knote *kn) 1701 { 1702 kn->kn_status |= KN_DELETING | KN_REPROCESS; 1703 if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) { 1704 kn->kn_fop->f_detach(kn); 1705 } else { 1706 get_mplock(); 1707 kn->kn_fop->f_detach(kn); 1708 rel_mplock(); 1709 } 1710 knote_drop(kn); 1711 } 1712 1713 /* 1714 * Calls filterops f_event function, acquiring mplock if filter is not 1715 * marked as FILTEROP_MPSAFE. 1716 * 1717 * If the knote is in the middle of being created or deleted we cannot 1718 * safely call the filter op. 1719 * 1720 * Caller must be holding the related kq token 1721 */ 1722 static int 1723 filter_event(struct knote *kn, long hint) 1724 { 1725 int ret; 1726 1727 if (kn->kn_fop->f_flags & FILTEROP_MPSAFE) { 1728 ret = kn->kn_fop->f_event(kn, hint); 1729 } else { 1730 get_mplock(); 1731 ret = kn->kn_fop->f_event(kn, hint); 1732 rel_mplock(); 1733 } 1734 return (ret); 1735 } 1736 1737 /* 1738 * Walk down a list of knotes, activating them if their event has triggered. 1739 * 1740 * If we encounter any knotes which are undergoing processing we just mark 1741 * them for reprocessing and do not try to [re]activate the knote. However, 1742 * if a hint is being passed we have to wait and that makes things a bit 1743 * sticky. 1744 */ 1745 void 1746 knote(struct klist *list, long hint) 1747 { 1748 struct kqueue *kq; 1749 struct knote *kn; 1750 struct knote *kntmp; 1751 1752 lwkt_getpooltoken(list); 1753 restart: 1754 SLIST_FOREACH(kn, list, kn_next) { 1755 kq = kn->kn_kq; 1756 lwkt_getpooltoken(kq); 1757 1758 /* temporary verification hack */ 1759 SLIST_FOREACH(kntmp, list, kn_next) { 1760 if (kn == kntmp) 1761 break; 1762 } 1763 if (kn != kntmp || kn->kn_kq != kq) { 1764 lwkt_relpooltoken(kq); 1765 goto restart; 1766 } 1767 1768 if (kn->kn_status & KN_PROCESSING) { 1769 /* 1770 * Someone else is processing the knote, ask the 1771 * other thread to reprocess it and don't mess 1772 * with it otherwise. 1773 */ 1774 if (hint == 0) { 1775 kn->kn_status |= KN_REPROCESS; 1776 lwkt_relpooltoken(kq); 1777 continue; 1778 } 1779 1780 /* 1781 * If the hint is non-zero we have to wait or risk 1782 * losing the state the caller is trying to update. 1783 * 1784 * XXX This is a real problem, certain process 1785 * and signal filters will bump kn_data for 1786 * already-processed notes more than once if 1787 * we restart the list scan. FIXME. 1788 */ 1789 kn->kn_status |= KN_WAITING | KN_REPROCESS; 1790 tsleep(kn, 0, "knotec", hz); 1791 lwkt_relpooltoken(kq); 1792 goto restart; 1793 } 1794 1795 /* 1796 * Become the reprocessing master ourselves. 1797 * 1798 * If hint is non-zero running the event is mandatory 1799 * when not deleting so do it whether reprocessing is 1800 * set or not. 1801 */ 1802 kn->kn_status |= KN_PROCESSING; 1803 if ((kn->kn_status & KN_DELETING) == 0) { 1804 if (filter_event(kn, hint)) 1805 KNOTE_ACTIVATE(kn); 1806 } 1807 if (knote_release(kn)) { 1808 lwkt_relpooltoken(kq); 1809 goto restart; 1810 } 1811 lwkt_relpooltoken(kq); 1812 } 1813 lwkt_relpooltoken(list); 1814 } 1815 1816 /* 1817 * Insert knote at head of klist. 1818 * 1819 * This function may only be called via a filter function and thus 1820 * kq_token should already be held and marked for processing. 1821 */ 1822 void 1823 knote_insert(struct klist *klist, struct knote *kn) 1824 { 1825 lwkt_getpooltoken(klist); 1826 KKASSERT(kn->kn_status & KN_PROCESSING); 1827 SLIST_INSERT_HEAD(klist, kn, kn_next); 1828 lwkt_relpooltoken(klist); 1829 } 1830 1831 /* 1832 * Remove knote from a klist 1833 * 1834 * This function may only be called via a filter function and thus 1835 * kq_token should already be held and marked for processing. 1836 */ 1837 void 1838 knote_remove(struct klist *klist, struct knote *kn) 1839 { 1840 lwkt_getpooltoken(klist); 1841 KKASSERT(kn->kn_status & KN_PROCESSING); 1842 SLIST_REMOVE(klist, kn, knote, kn_next); 1843 lwkt_relpooltoken(klist); 1844 } 1845 1846 void 1847 knote_assume_knotes(struct kqinfo *src, struct kqinfo *dst, 1848 struct filterops *ops, void *hook) 1849 { 1850 struct kqueue *kq; 1851 struct knote *kn; 1852 1853 lwkt_getpooltoken(&src->ki_note); 1854 lwkt_getpooltoken(&dst->ki_note); 1855 while ((kn = SLIST_FIRST(&src->ki_note)) != NULL) { 1856 kq = kn->kn_kq; 1857 lwkt_getpooltoken(kq); 1858 if (SLIST_FIRST(&src->ki_note) != kn || kn->kn_kq != kq) { 1859 lwkt_relpooltoken(kq); 1860 continue; 1861 } 1862 if (knote_acquire(kn)) { 1863 knote_remove(&src->ki_note, kn); 1864 kn->kn_fop = ops; 1865 kn->kn_hook = hook; 1866 knote_insert(&dst->ki_note, kn); 1867 knote_release(kn); 1868 /* kn may be invalid now */ 1869 } 1870 lwkt_relpooltoken(kq); 1871 } 1872 lwkt_relpooltoken(&dst->ki_note); 1873 lwkt_relpooltoken(&src->ki_note); 1874 } 1875 1876 /* 1877 * Remove all knotes referencing a specified fd 1878 */ 1879 void 1880 knote_fdclose(struct file *fp, struct filedesc *fdp, int fd) 1881 { 1882 struct kqueue *kq; 1883 struct knote *kn; 1884 struct knote *kntmp; 1885 1886 lwkt_getpooltoken(&fp->f_klist); 1887 restart: 1888 SLIST_FOREACH(kn, &fp->f_klist, kn_link) { 1889 if (kn->kn_kq->kq_fdp == fdp && kn->kn_id == fd) { 1890 kq = kn->kn_kq; 1891 lwkt_getpooltoken(kq); 1892 1893 /* temporary verification hack */ 1894 SLIST_FOREACH(kntmp, &fp->f_klist, kn_link) { 1895 if (kn == kntmp) 1896 break; 1897 } 1898 if (kn != kntmp || kn->kn_kq->kq_fdp != fdp || 1899 kn->kn_id != fd || kn->kn_kq != kq) { 1900 lwkt_relpooltoken(kq); 1901 goto restart; 1902 } 1903 if (knote_acquire(kn)) 1904 knote_detach_and_drop(kn); 1905 lwkt_relpooltoken(kq); 1906 goto restart; 1907 } 1908 } 1909 lwkt_relpooltoken(&fp->f_klist); 1910 } 1911 1912 /* 1913 * Low level attach function. 1914 * 1915 * The knote should already be marked for processing. 1916 * Caller must hold the related kq token. 1917 */ 1918 static void 1919 knote_attach(struct knote *kn) 1920 { 1921 struct klist *list; 1922 struct kqueue *kq = kn->kn_kq; 1923 1924 if (kn->kn_fop->f_flags & FILTEROP_ISFD) { 1925 KKASSERT(kn->kn_fp); 1926 list = &kn->kn_fp->f_klist; 1927 } else { 1928 if (kq->kq_knhashmask == 0) 1929 kq->kq_knhash = hashinit(KN_HASHSIZE, M_KQUEUE, 1930 &kq->kq_knhashmask); 1931 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)]; 1932 } 1933 lwkt_getpooltoken(list); 1934 SLIST_INSERT_HEAD(list, kn, kn_link); 1935 lwkt_relpooltoken(list); 1936 TAILQ_INSERT_HEAD(&kq->kq_knlist, kn, kn_kqlink); 1937 } 1938 1939 /* 1940 * Low level drop function. 1941 * 1942 * The knote should already be marked for processing. 1943 * Caller must hold the related kq token. 1944 */ 1945 static void 1946 knote_drop(struct knote *kn) 1947 { 1948 struct kqueue *kq; 1949 struct klist *list; 1950 1951 kq = kn->kn_kq; 1952 1953 if (kn->kn_fop->f_flags & FILTEROP_ISFD) 1954 list = &kn->kn_fp->f_klist; 1955 else 1956 list = &kq->kq_knhash[KN_HASH(kn->kn_id, kq->kq_knhashmask)]; 1957 1958 lwkt_getpooltoken(list); 1959 SLIST_REMOVE(list, kn, knote, kn_link); 1960 lwkt_relpooltoken(list); 1961 TAILQ_REMOVE(&kq->kq_knlist, kn, kn_kqlink); 1962 if (kn->kn_status & KN_QUEUED) 1963 knote_dequeue(kn); 1964 if (kn->kn_fop->f_flags & FILTEROP_ISFD) { 1965 fdrop(kn->kn_fp); 1966 kn->kn_fp = NULL; 1967 } 1968 knote_free(kn); 1969 } 1970 1971 /* 1972 * Low level enqueue function. 1973 * 1974 * The knote should already be marked for processing. 1975 * Caller must be holding the kq token 1976 */ 1977 static void 1978 knote_enqueue(struct knote *kn) 1979 { 1980 struct kqueue *kq = kn->kn_kq; 1981 1982 KASSERT((kn->kn_status & KN_QUEUED) == 0, ("knote already queued")); 1983 TAILQ_INSERT_TAIL(&kq->kq_knpend, kn, kn_tqe); 1984 kn->kn_status |= KN_QUEUED; 1985 ++kq->kq_count; 1986 1987 /* 1988 * Send SIGIO on request (typically set up as a mailbox signal) 1989 */ 1990 if (kq->kq_sigio && (kq->kq_state & KQ_ASYNC) && kq->kq_count == 1) 1991 pgsigio(kq->kq_sigio, SIGIO, 0); 1992 1993 kqueue_wakeup(kq); 1994 } 1995 1996 /* 1997 * Low level dequeue function. 1998 * 1999 * The knote should already be marked for processing. 2000 * Caller must be holding the kq token 2001 */ 2002 static void 2003 knote_dequeue(struct knote *kn) 2004 { 2005 struct kqueue *kq = kn->kn_kq; 2006 2007 KASSERT(kn->kn_status & KN_QUEUED, ("knote not queued")); 2008 TAILQ_REMOVE(&kq->kq_knpend, kn, kn_tqe); 2009 kn->kn_status &= ~KN_QUEUED; 2010 kq->kq_count--; 2011 } 2012 2013 static struct knote * 2014 knote_alloc(void) 2015 { 2016 return kmalloc(sizeof(struct knote), M_KQUEUE, M_WAITOK); 2017 } 2018 2019 static void 2020 knote_free(struct knote *kn) 2021 { 2022 struct knote_cache_list *cache_list; 2023 2024 cache_list = &knote_cache_lists[mycpuid]; 2025 if (cache_list->knote_cache_cnt < KNOTE_CACHE_MAX) { 2026 crit_enter(); 2027 SLIST_INSERT_HEAD(&cache_list->knote_cache, kn, kn_link); 2028 cache_list->knote_cache_cnt++; 2029 crit_exit(); 2030 return; 2031 } 2032 kfree(kn, M_KQUEUE); 2033 } 2034 2035 struct sleepinfo { 2036 void *ident; 2037 int timedout; 2038 }; 2039 2040 static void 2041 precise_sleep_intr(systimer_t info, int in_ipi, struct intrframe *frame) 2042 { 2043 struct sleepinfo *si; 2044 2045 si = info->data; 2046 si->timedout = 1; 2047 wakeup(si->ident); 2048 } 2049 2050 static int 2051 precise_sleep(void *ident, int flags, const char *wmesg, int us) 2052 { 2053 struct systimer info; 2054 struct sleepinfo si = { 2055 .ident = ident, 2056 .timedout = 0, 2057 }; 2058 int r; 2059 2060 tsleep_interlock(ident, flags); 2061 systimer_init_oneshot(&info, precise_sleep_intr, &si, 2062 us == 0 ? 1 : us); 2063 r = tsleep(ident, flags | PINTERLOCKED, wmesg, 0); 2064 systimer_del(&info); 2065 if (si.timedout) 2066 r = EWOULDBLOCK; 2067 2068 return r; 2069 } 2070