1 /* $NetBSD: kqueue.c,v 1.6 2020/05/25 20:47:33 christos Exp $ */ 2 3 /* $OpenBSD: kqueue.c,v 1.5 2002/07/10 14:41:31 art Exp $ */ 4 5 /* 6 * Copyright 2000-2007 Niels Provos <provos@citi.umich.edu> 7 * Copyright 2007-2012 Niels Provos and Nick Mathewson 8 * 9 * Redistribution and use in source and binary forms, with or without 10 * modification, are permitted provided that the following conditions 11 * are met: 12 * 1. Redistributions of source code must retain the above copyright 13 * notice, this list of conditions and the following disclaimer. 14 * 2. Redistributions in binary form must reproduce the above copyright 15 * notice, this list of conditions and the following disclaimer in the 16 * documentation and/or other materials provided with the distribution. 17 * 3. The name of the author may not be used to endorse or promote products 18 * derived from this software without specific prior written permission. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 21 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 22 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 23 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 24 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 25 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 26 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 27 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 28 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 29 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 30 */ 31 #include "event2/event-config.h" 32 #include "evconfig-private.h" 33 34 #ifdef EVENT__HAVE_KQUEUE 35 36 #include <sys/types.h> 37 #ifdef EVENT__HAVE_SYS_TIME_H 38 #include <sys/time.h> 39 #endif 40 #include <sys/queue.h> 41 #include <sys/event.h> 42 #include <signal.h> 43 #include <stdio.h> 44 #include <stdlib.h> 45 #include <string.h> 46 #include <unistd.h> 47 #include <errno.h> 48 #ifdef EVENT__HAVE_INTTYPES_H 49 #include <inttypes.h> 50 #endif 51 52 /* Some platforms apparently define the udata field of struct kevent as 53 * intptr_t, whereas others define it as void*. There doesn't seem to be an 54 * easy way to tell them apart via autoconf, so we need to use OS macros. */ 55 #if defined(EVENT__HAVE_INTTYPES_H) && !defined(__OpenBSD__) && !defined(__FreeBSD__) && !defined(__darwin__) && !defined(__APPLE__) 56 #define PTR_TO_UDATA(x) ((intptr_t)(x)) 57 #define INT_TO_UDATA(x) ((intptr_t)(x)) 58 #else 59 #define PTR_TO_UDATA(x) (x) 60 #define INT_TO_UDATA(x) ((void*)(x)) 61 #endif 62 63 #include "event-internal.h" 64 #include "log-internal.h" 65 #include "evmap-internal.h" 66 #include "event2/thread.h" 67 #include "evthread-internal.h" 68 #include "changelist-internal.h" 69 70 #include "kqueue-internal.h" 71 72 #define NEVENT 64 73 74 struct kqop { 75 struct kevent *changes; 76 int changes_size; 77 78 struct kevent *events; 79 int events_size; 80 int kq; 81 int notify_event_added; 82 pid_t pid; 83 }; 84 85 static void kqop_free(struct kqop *kqop); 86 87 static void *kq_init(struct event_base *); 88 static int kq_sig_add(struct event_base *, int, short, short, void *); 89 static int kq_sig_del(struct event_base *, int, short, short, void *); 90 static int kq_dispatch(struct event_base *, struct timeval *); 91 static void kq_dealloc(struct event_base *); 92 93 const struct eventop kqops = { 94 "kqueue", 95 kq_init, 96 event_changelist_add_, 97 event_changelist_del_, 98 kq_dispatch, 99 kq_dealloc, 100 1 /* need reinit */, 101 EV_FEATURE_ET|EV_FEATURE_O1|EV_FEATURE_FDS, 102 EVENT_CHANGELIST_FDINFO_SIZE 103 }; 104 105 static const struct eventop kqsigops = { 106 "kqueue_signal", 107 NULL, 108 kq_sig_add, 109 kq_sig_del, 110 NULL, 111 NULL, 112 1 /* need reinit */, 113 0, 114 0 115 }; 116 117 static void * 118 kq_init(struct event_base *base) 119 { 120 int kq = -1; 121 struct kqop *kqueueop = NULL; 122 123 if (!(kqueueop = mm_calloc(1, sizeof(struct kqop)))) 124 return (NULL); 125 126 /* Initialize the kernel queue */ 127 128 if ((kq = kqueue()) == -1) { 129 event_warn("kqueue"); 130 goto err; 131 } 132 133 kqueueop->kq = kq; 134 135 kqueueop->pid = getpid(); 136 137 /* Initialize fields */ 138 kqueueop->changes = mm_calloc(NEVENT, sizeof(struct kevent)); 139 if (kqueueop->changes == NULL) 140 goto err; 141 kqueueop->events = mm_calloc(NEVENT, sizeof(struct kevent)); 142 if (kqueueop->events == NULL) 143 goto err; 144 kqueueop->events_size = kqueueop->changes_size = NEVENT; 145 146 /* Check for Mac OS X kqueue bug. */ 147 memset(&kqueueop->changes[0], 0, sizeof kqueueop->changes[0]); 148 kqueueop->changes[0].ident = -1; 149 kqueueop->changes[0].filter = EVFILT_READ; 150 kqueueop->changes[0].flags = EV_ADD; 151 /* 152 * If kqueue works, then kevent will succeed, and it will 153 * stick an error in events[0]. If kqueue is broken, then 154 * kevent will fail. 155 */ 156 if (kevent(kq, 157 kqueueop->changes, 1, kqueueop->events, NEVENT, NULL) != 1 || 158 (int)kqueueop->events[0].ident != -1 || 159 kqueueop->events[0].flags != EV_ERROR) { 160 event_warn("%s: detected broken kqueue; not using.", __func__); 161 goto err; 162 } 163 164 base->evsigsel = &kqsigops; 165 166 return (kqueueop); 167 err: 168 if (kqueueop) 169 kqop_free(kqueueop); 170 171 return (NULL); 172 } 173 174 #define ADD_UDATA 0x30303 175 176 static void 177 kq_setup_kevent(struct kevent *out, evutil_socket_t fd, int filter, short change) 178 { 179 memset(out, 0, sizeof(struct kevent)); 180 out->ident = fd; 181 out->filter = filter; 182 183 if (change & EV_CHANGE_ADD) { 184 out->flags = EV_ADD; 185 /* We set a magic number here so that we can tell 'add' 186 * errors from 'del' errors. */ 187 out->udata = INT_TO_UDATA(ADD_UDATA); 188 if (change & EV_ET) 189 out->flags |= EV_CLEAR; 190 #ifdef NOTE_EOF 191 /* Make it behave like select() and poll() */ 192 if (filter == EVFILT_READ) 193 out->fflags = NOTE_EOF; 194 #endif 195 } else { 196 EVUTIL_ASSERT(change & EV_CHANGE_DEL); 197 out->flags = EV_DELETE; 198 } 199 } 200 201 static int 202 kq_build_changes_list(const struct event_changelist *changelist, 203 struct kqop *kqop) 204 { 205 int i; 206 int n_changes = 0; 207 208 for (i = 0; i < changelist->n_changes; ++i) { 209 struct event_change *in_ch = &changelist->changes[i]; 210 struct kevent *out_ch; 211 if (n_changes >= kqop->changes_size - 1) { 212 int newsize = kqop->changes_size * 2; 213 struct kevent *newchanges; 214 215 newchanges = mm_realloc(kqop->changes, 216 newsize * sizeof(struct kevent)); 217 if (newchanges == NULL) { 218 event_warn("%s: realloc", __func__); 219 return (-1); 220 } 221 kqop->changes = newchanges; 222 kqop->changes_size = newsize; 223 } 224 if (in_ch->read_change) { 225 out_ch = &kqop->changes[n_changes++]; 226 kq_setup_kevent(out_ch, in_ch->fd, EVFILT_READ, 227 in_ch->read_change); 228 } 229 if (in_ch->write_change) { 230 out_ch = &kqop->changes[n_changes++]; 231 kq_setup_kevent(out_ch, in_ch->fd, EVFILT_WRITE, 232 in_ch->write_change); 233 } 234 } 235 return n_changes; 236 } 237 238 static int 239 kq_grow_events(struct kqop *kqop, size_t new_size) 240 { 241 struct kevent *newresult; 242 243 newresult = mm_realloc(kqop->events, 244 new_size * sizeof(struct kevent)); 245 246 if (newresult) { 247 kqop->events = newresult; 248 kqop->events_size = new_size; 249 return 0; 250 } else { 251 return -1; 252 } 253 } 254 255 static int 256 kq_dispatch(struct event_base *base, struct timeval *tv) 257 { 258 struct kqop *kqop = base->evbase; 259 struct kevent *events = kqop->events; 260 struct kevent *changes; 261 struct timespec ts, *ts_p = NULL; 262 int i, n_changes, res; 263 264 if (tv != NULL) { 265 TIMEVAL_TO_TIMESPEC(tv, &ts); 266 ts_p = &ts; 267 } 268 269 /* Build "changes" from "base->changes" */ 270 EVUTIL_ASSERT(kqop->changes); 271 n_changes = kq_build_changes_list(&base->changelist, kqop); 272 if (n_changes < 0) 273 return -1; 274 275 event_changelist_remove_all_(&base->changelist, base); 276 277 /* steal the changes array in case some broken code tries to call 278 * dispatch twice at once. */ 279 changes = kqop->changes; 280 kqop->changes = NULL; 281 282 /* Make sure that 'events' is at least as long as the list of changes: 283 * otherwise errors in the changes can get reported as a -1 return 284 * value from kevent() rather than as EV_ERROR events in the events 285 * array. 286 * 287 * (We could instead handle -1 return values from kevent() by 288 * retrying with a smaller changes array or a larger events array, 289 * but this approach seems less risky for now.) 290 */ 291 if (kqop->events_size < n_changes) { 292 int new_size = kqop->events_size; 293 do { 294 new_size *= 2; 295 } while (new_size < n_changes); 296 297 kq_grow_events(kqop, new_size); 298 events = kqop->events; 299 } 300 301 EVBASE_RELEASE_LOCK(base, th_base_lock); 302 303 res = kevent(kqop->kq, changes, n_changes, 304 events, kqop->events_size, ts_p); 305 306 EVBASE_ACQUIRE_LOCK(base, th_base_lock); 307 308 EVUTIL_ASSERT(kqop->changes == NULL); 309 kqop->changes = changes; 310 311 if (res == -1) { 312 if (errno != EINTR) { 313 event_warn("kevent"); 314 return (-1); 315 } 316 317 return (0); 318 } 319 320 event_debug(("%s: kevent reports %d", __func__, res)); 321 322 for (i = 0; i < res; i++) { 323 int which = 0; 324 325 if (events[i].flags & EV_ERROR) { 326 switch (events[i].data) { 327 328 /* Can occur on delete if we are not currently 329 * watching any events on this fd. That can 330 * happen when the fd was closed and another 331 * file was opened with that fd. */ 332 case ENOENT: 333 /* Can occur for reasons not fully understood 334 * on FreeBSD. */ 335 case EINVAL: 336 continue; 337 #if defined(__FreeBSD__) && defined(ENOTCAPABLE) 338 /* 339 * This currently occurs if an FD is closed 340 * before the EV_DELETE makes it out via kevent(). 341 * The FreeBSD capabilities code sees the blank 342 * capability set and rejects the request to 343 * modify an event. 344 * 345 * To be strictly correct - when an FD is closed, 346 * all the registered events are also removed. 347 * Queuing EV_DELETE to a closed FD is wrong. 348 * The event(s) should just be deleted from 349 * the pending changelist. 350 */ 351 case ENOTCAPABLE: 352 continue; 353 #endif 354 355 /* Can occur on a delete if the fd is closed. */ 356 case EBADF: 357 /* XXXX On NetBSD, we can also get EBADF if we 358 * try to add the write side of a pipe, but 359 * the read side has already been closed. 360 * Other BSDs call this situation 'EPIPE'. It 361 * would be good if we had a way to report 362 * this situation. */ 363 continue; 364 /* These two can occur on an add if the fd was one side 365 * of a pipe, and the other side was closed. */ 366 case EPERM: 367 case EPIPE: 368 /* Report read events, if we're listening for 369 * them, so that the user can learn about any 370 * add errors. (If the operation was a 371 * delete, then udata should be cleared.) */ 372 if (events[i].udata) { 373 /* The operation was an add: 374 * report the error as a read. */ 375 which |= EV_READ; 376 break; 377 } else { 378 /* The operation was a del: 379 * report nothing. */ 380 continue; 381 } 382 383 /* Other errors shouldn't occur. */ 384 default: 385 errno = events[i].data; 386 return (-1); 387 } 388 } else if (events[i].filter == EVFILT_READ) { 389 which |= EV_READ; 390 } else if (events[i].filter == EVFILT_WRITE) { 391 which |= EV_WRITE; 392 } else if (events[i].filter == EVFILT_SIGNAL) { 393 which |= EV_SIGNAL; 394 #ifdef EVFILT_USER 395 } else if (events[i].filter == EVFILT_USER) { 396 base->is_notify_pending = 0; 397 #endif 398 } 399 400 if (!which) 401 continue; 402 403 if (events[i].filter == EVFILT_SIGNAL) { 404 evmap_signal_active_(base, events[i].ident, 1); 405 } else { 406 evmap_io_active_(base, events[i].ident, which | EV_ET); 407 } 408 } 409 410 if (res == kqop->events_size) { 411 /* We used all the events space that we have. Maybe we should 412 make it bigger. */ 413 kq_grow_events(kqop, kqop->events_size * 2); 414 } 415 416 return (0); 417 } 418 419 static void 420 kqop_free(struct kqop *kqop) 421 { 422 if (kqop->changes) 423 mm_free(kqop->changes); 424 if (kqop->events) 425 mm_free(kqop->events); 426 if (kqop->kq >= 0 && kqop->pid == getpid()) 427 close(kqop->kq); 428 memset(kqop, 0, sizeof(struct kqop)); 429 mm_free(kqop); 430 } 431 432 static void 433 kq_dealloc(struct event_base *base) 434 { 435 struct kqop *kqop = base->evbase; 436 evsig_dealloc_(base); 437 kqop_free(kqop); 438 } 439 440 /* signal handling */ 441 static int 442 kq_sig_add(struct event_base *base, int nsignal, short old, short events, void *p) 443 { 444 struct kqop *kqop = base->evbase; 445 struct kevent kev; 446 struct timespec timeout = { 0, 0 }; 447 (void)p; 448 449 EVUTIL_ASSERT(nsignal >= 0 && nsignal < NSIG); 450 451 memset(&kev, 0, sizeof(kev)); 452 kev.ident = nsignal; 453 kev.filter = EVFILT_SIGNAL; 454 kev.flags = EV_ADD; 455 456 /* Be ready for the signal if it is sent any 457 * time between now and the next call to 458 * kq_dispatch. */ 459 if (kevent(kqop->kq, &kev, 1, NULL, 0, &timeout) == -1) 460 return (-1); 461 462 /* We can set the handler for most signals to SIG_IGN and 463 * still have them reported to us in the queue. However, 464 * if the handler for SIGCHLD is SIG_IGN, the system reaps 465 * zombie processes for us, and we don't get any notification. 466 * This appears to be the only signal with this quirk. */ 467 if (evsig_set_handler_(base, nsignal, 468 nsignal == SIGCHLD ? SIG_DFL : SIG_IGN) == -1) 469 return (-1); 470 471 return (0); 472 } 473 474 static int 475 kq_sig_del(struct event_base *base, int nsignal, short old, short events, void *p) 476 { 477 struct kqop *kqop = base->evbase; 478 struct kevent kev; 479 480 struct timespec timeout = { 0, 0 }; 481 (void)p; 482 483 EVUTIL_ASSERT(nsignal >= 0 && nsignal < NSIG); 484 485 memset(&kev, 0, sizeof(kev)); 486 kev.ident = nsignal; 487 kev.filter = EVFILT_SIGNAL; 488 kev.flags = EV_DELETE; 489 490 /* Because we insert signal events 491 * immediately, we need to delete them 492 * immediately, too */ 493 if (kevent(kqop->kq, &kev, 1, NULL, 0, &timeout) == -1) 494 return (-1); 495 496 if (evsig_restore_handler_(base, nsignal) == -1) 497 return (-1); 498 499 return (0); 500 } 501 502 503 /* OSX 10.6 and FreeBSD 8.1 add support for EVFILT_USER, which we can use 504 * to wake up the event loop from another thread. */ 505 506 /* Magic number we use for our filter ID. */ 507 #define NOTIFY_IDENT 42 508 509 int 510 event_kq_add_notify_event_(struct event_base *base) 511 { 512 struct kqop *kqop = base->evbase; 513 #if defined(EVFILT_USER) && defined(NOTE_TRIGGER) 514 struct kevent kev; 515 struct timespec timeout = { 0, 0 }; 516 #endif 517 518 if (kqop->notify_event_added) 519 return 0; 520 521 #if defined(EVFILT_USER) && defined(NOTE_TRIGGER) 522 memset(&kev, 0, sizeof(kev)); 523 kev.ident = NOTIFY_IDENT; 524 kev.filter = EVFILT_USER; 525 kev.flags = EV_ADD | EV_CLEAR; 526 527 if (kevent(kqop->kq, &kev, 1, NULL, 0, &timeout) == -1) { 528 event_warn("kevent: adding EVFILT_USER event"); 529 return -1; 530 } 531 532 kqop->notify_event_added = 1; 533 534 return 0; 535 #else 536 return -1; 537 #endif 538 } 539 540 int 541 event_kq_notify_base_(struct event_base *base) 542 { 543 struct kqop *kqop = base->evbase; 544 #if defined(EVFILT_USER) && defined(NOTE_TRIGGER) 545 struct kevent kev; 546 struct timespec timeout = { 0, 0 }; 547 #endif 548 if (! kqop->notify_event_added) 549 return -1; 550 551 #if defined(EVFILT_USER) && defined(NOTE_TRIGGER) 552 memset(&kev, 0, sizeof(kev)); 553 kev.ident = NOTIFY_IDENT; 554 kev.filter = EVFILT_USER; 555 kev.fflags = NOTE_TRIGGER; 556 557 if (kevent(kqop->kq, &kev, 1, NULL, 0, &timeout) == -1) { 558 event_warn("kevent: triggering EVFILT_USER event"); 559 return -1; 560 } 561 562 return 0; 563 #else 564 return -1; 565 #endif 566 } 567 568 #endif /* EVENT__HAVE_KQUEUE */ 569