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