1/* $OpenBSD: kqueue.c,v 1.5 2002/07/10 14:41:31 art Exp $ */ 2 3/* 4 * Copyright 2000-2007 Niels Provos <provos@citi.umich.edu> 5 * Copyright 2007-2012 Niels Provos and Nick Mathewson 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice, this list of conditions and the following disclaimer. 12 * 2. Redistributions in binary form must reproduce the above copyright 13 * notice, this list of conditions and the following disclaimer in the 14 * documentation and/or other materials provided with the distribution. 15 * 3. The name of the author may not be used to endorse or promote products 16 * derived from this software without specific prior written permission. 17 * 18 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 19 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 20 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 21 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 22 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 23 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 24 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 25 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 26 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 27 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 28 */ 29#include "event2/event-config.h" 30#include "evconfig-private.h" 31 32#ifdef EVENT__HAVE_KQUEUE 33 34#include <sys/types.h> 35#ifdef EVENT__HAVE_SYS_TIME_H 36#include <sys/time.h> 37#endif 38#include <sys/queue.h> 39#include <sys/event.h> 40#include <limits.h> 41#include <signal.h> 42#include <stdio.h> 43#include <stdlib.h> 44#include <string.h> 45#include <unistd.h> 46#include <errno.h> 47#ifdef EVENT__HAVE_INTTYPES_H 48#include <inttypes.h> 49#endif 50 51/* Some platforms apparently define the udata field of struct kevent as 52 * intptr_t, whereas others define it as void*. There doesn't seem to be an 53 * easy way to tell them apart via autoconf, so we need to use OS macros. */ 54#if defined(__NetBSD__) 55#define PTR_TO_UDATA(x) ((typeof(((struct kevent *)0)->udata))(x)) 56#define INT_TO_UDATA(x) ((typeof(((struct kevent *)0)->udata))(intptr_t)(x)) 57#elif defined(EVENT__HAVE_INTTYPES_H) && !defined(__OpenBSD__) && !defined(__FreeBSD__) && !defined(__darwin__) && !defined(__APPLE__) && !defined(__CloudABI__) 58#define PTR_TO_UDATA(x) ((intptr_t)(x)) 59#define INT_TO_UDATA(x) ((intptr_t)(x)) 60#else 61#define PTR_TO_UDATA(x) (x) 62#define INT_TO_UDATA(x) ((void*)(x)) 63#endif 64 65#include "event-internal.h" 66#include "log-internal.h" 67#include "evmap-internal.h" 68#include "event2/thread.h" 69#include "event2/util.h" 70#include "evthread-internal.h" 71#include "changelist-internal.h" 72 73#include "kqueue-internal.h" 74 75#define NEVENT 64 76 77struct kqop { 78 struct kevent *changes; 79 int changes_size; 80 81 struct kevent *events; 82 int events_size; 83 int kq; 84 int notify_event_added; 85 pid_t pid; 86}; 87 88static void kqop_free(struct kqop *kqop); 89 90static void *kq_init(struct event_base *); 91static int kq_sig_add(struct event_base *, int, short, short, void *); 92static int kq_sig_del(struct event_base *, int, short, short, void *); 93static int kq_dispatch(struct event_base *, struct timeval *); 94static void kq_dealloc(struct event_base *); 95 96const struct eventop kqops = { 97 "kqueue", 98 kq_init, 99 event_changelist_add_, 100 event_changelist_del_, 101 kq_dispatch, 102 kq_dealloc, 103 1 /* need reinit */, 104 EV_FEATURE_ET|EV_FEATURE_O1|EV_FEATURE_FDS, 105 EVENT_CHANGELIST_FDINFO_SIZE 106}; 107 108static const struct eventop kqsigops = { 109 "kqueue_signal", 110 NULL, 111 kq_sig_add, 112 kq_sig_del, 113 NULL, 114 NULL, 115 1 /* need reinit */, 116 0, 117 0 118}; 119 120static void * 121kq_init(struct event_base *base) 122{ 123 int kq = -1; 124 struct kqop *kqueueop = NULL; 125 126 if (!(kqueueop = mm_calloc(1, sizeof(struct kqop)))) 127 return (NULL); 128 129/* Initialize the kernel queue */ 130 131 if ((kq = kqueue()) == -1) { 132 event_warn("kqueue"); 133 goto err; 134 } 135 136 kqueueop->kq = kq; 137 138 kqueueop->pid = getpid(); 139 140 /* Initialize fields */ 141 kqueueop->changes = mm_calloc(NEVENT, sizeof(struct kevent)); 142 if (kqueueop->changes == NULL) 143 goto err; 144 kqueueop->events = mm_calloc(NEVENT, sizeof(struct kevent)); 145 if (kqueueop->events == NULL) 146 goto err; 147 kqueueop->events_size = kqueueop->changes_size = NEVENT; 148 149 /* Check for Mac OS X kqueue bug. */ 150 memset(&kqueueop->changes[0], 0, sizeof kqueueop->changes[0]); 151 kqueueop->changes[0].ident = -1; 152 kqueueop->changes[0].filter = EVFILT_READ; 153 kqueueop->changes[0].flags = EV_ADD; 154 /* 155 * If kqueue works, then kevent will succeed, and it will 156 * stick an error in events[0]. If kqueue is broken, then 157 * kevent will fail. 158 */ 159 if (kevent(kq, 160 kqueueop->changes, 1, kqueueop->events, NEVENT, NULL) != 1 || 161 (int)kqueueop->events[0].ident != -1 || 162 !(kqueueop->events[0].flags & EV_ERROR)) { 163 event_warn("%s: detected broken kqueue; not using.", __func__); 164 goto err; 165 } 166 167 base->evsigsel = &kqsigops; 168 169 return (kqueueop); 170err: 171 if (kqueueop) 172 kqop_free(kqueueop); 173 174 return (NULL); 175} 176 177#define ADD_UDATA 0x30303 178 179static void 180kq_setup_kevent(struct kevent *out, evutil_socket_t fd, int filter, short change) 181{ 182 memset(out, 0, sizeof(struct kevent)); 183 out->ident = fd; 184 out->filter = filter; 185 186 if (change & EV_CHANGE_ADD) { 187 out->flags = EV_ADD; 188 /* We set a magic number here so that we can tell 'add' 189 * errors from 'del' errors. */ 190 out->udata = INT_TO_UDATA(ADD_UDATA); 191 if (change & EV_ET) 192 out->flags |= EV_CLEAR; 193#ifdef NOTE_EOF 194 /* Make it behave like select() and poll() */ 195 if (filter == EVFILT_READ) 196 out->fflags = NOTE_EOF; 197#endif 198 } else { 199 EVUTIL_ASSERT(change & EV_CHANGE_DEL); 200 out->flags = EV_DELETE; 201 } 202} 203 204static int 205kq_build_changes_list(const struct event_changelist *changelist, 206 struct kqop *kqop) 207{ 208 int i; 209 int n_changes = 0; 210 211 for (i = 0; i < changelist->n_changes; ++i) { 212 struct event_change *in_ch = &changelist->changes[i]; 213 struct kevent *out_ch; 214 if (n_changes >= kqop->changes_size - 1) { 215 int newsize; 216 struct kevent *newchanges; 217 218 if (kqop->changes_size > INT_MAX / 2 || 219 (size_t)kqop->changes_size * 2 > EV_SIZE_MAX / 220 sizeof(struct kevent)) { 221 event_warnx("%s: int overflow", __func__); 222 return (-1); 223 } 224 225 newsize = kqop->changes_size * 2; 226 newchanges = mm_realloc(kqop->changes, 227 newsize * sizeof(struct kevent)); 228 if (newchanges == NULL) { 229 event_warn("%s: realloc", __func__); 230 return (-1); 231 } 232 kqop->changes = newchanges; 233 kqop->changes_size = newsize; 234 } 235 if (in_ch->read_change) { 236 out_ch = &kqop->changes[n_changes++]; 237 kq_setup_kevent(out_ch, in_ch->fd, EVFILT_READ, 238 in_ch->read_change); 239 } 240 if (in_ch->write_change) { 241 out_ch = &kqop->changes[n_changes++]; 242 kq_setup_kevent(out_ch, in_ch->fd, EVFILT_WRITE, 243 in_ch->write_change); 244 } 245 } 246 return n_changes; 247} 248 249static int 250kq_grow_events(struct kqop *kqop, size_t new_size) 251{ 252 struct kevent *newresult; 253 254 newresult = mm_realloc(kqop->events, 255 new_size * sizeof(struct kevent)); 256 257 if (newresult) { 258 kqop->events = newresult; 259 kqop->events_size = new_size; 260 return 0; 261 } else { 262 return -1; 263 } 264} 265 266static int 267kq_dispatch(struct event_base *base, struct timeval *tv) 268{ 269 struct kqop *kqop = base->evbase; 270 struct kevent *events = kqop->events; 271 struct kevent *changes; 272 struct timespec ts, *ts_p = NULL; 273 int i, n_changes, res; 274 275 if (tv != NULL) { 276 ts.tv_sec = tv->tv_sec; 277 ts.tv_nsec = tv->tv_usec * 1000; 278 ts_p = &ts; 279 } 280 281 /* Build "changes" from "base->changes" */ 282 EVUTIL_ASSERT(kqop->changes); 283 n_changes = kq_build_changes_list(&base->changelist, kqop); 284 if (n_changes < 0) 285 return -1; 286 287 event_changelist_remove_all_(&base->changelist, base); 288 289 /* steal the changes array in case some broken code tries to call 290 * dispatch twice at once. */ 291 changes = kqop->changes; 292 kqop->changes = NULL; 293 294 /* Make sure that 'events' is at least as long as the list of changes: 295 * otherwise errors in the changes can get reported as a -1 return 296 * value from kevent() rather than as EV_ERROR events in the events 297 * array. 298 * 299 * (We could instead handle -1 return values from kevent() by 300 * retrying with a smaller changes array or a larger events array, 301 * but this approach seems less risky for now.) 302 */ 303 if (kqop->events_size < n_changes) { 304 int new_size = kqop->events_size; 305 do { 306 new_size *= 2; 307 } while (new_size < n_changes); 308 309 kq_grow_events(kqop, new_size); 310 events = kqop->events; 311 } 312 313 EVBASE_RELEASE_LOCK(base, th_base_lock); 314 315 res = kevent(kqop->kq, changes, n_changes, 316 events, kqop->events_size, ts_p); 317 318 EVBASE_ACQUIRE_LOCK(base, th_base_lock); 319 320 EVUTIL_ASSERT(kqop->changes == NULL); 321 kqop->changes = changes; 322 323 if (res == -1) { 324 if (errno != EINTR) { 325 event_warn("kevent"); 326 return (-1); 327 } 328 329 return (0); 330 } 331 332 event_debug(("%s: kevent reports %d", __func__, res)); 333 334 for (i = 0; i < res; i++) { 335 int which = 0; 336 337 if (events[i].flags & EV_ERROR) { 338 switch (events[i].data) { 339 340 /* Can occur on delete if we are not currently 341 * watching any events on this fd. That can 342 * happen when the fd was closed and another 343 * file was opened with that fd. */ 344 case ENOENT: 345 /* Can occur for reasons not fully understood 346 * on FreeBSD. */ 347 case EINVAL: 348 continue; 349#if defined(__FreeBSD__) 350 /* 351 * This currently occurs if an FD is closed 352 * before the EV_DELETE makes it out via kevent(). 353 * The FreeBSD capabilities code sees the blank 354 * capability set and rejects the request to 355 * modify an event. 356 * 357 * To be strictly correct - when an FD is closed, 358 * all the registered events are also removed. 359 * Queuing EV_DELETE to a closed FD is wrong. 360 * The event(s) should just be deleted from 361 * the pending changelist. 362 */ 363 case ENOTCAPABLE: 364 continue; 365#endif 366 367 /* Can occur on a delete if the fd is closed. */ 368 case EBADF: 369 /* XXXX On NetBSD, we can also get EBADF if we 370 * try to add the write side of a pipe, but 371 * the read side has already been closed. 372 * Other BSDs call this situation 'EPIPE'. It 373 * would be good if we had a way to report 374 * this situation. */ 375 continue; 376 /* These two can occur on an add if the fd was one side 377 * of a pipe, and the other side was closed. */ 378 case EPERM: 379 case EPIPE: 380 /* Report read events, if we're listening for 381 * them, so that the user can learn about any 382 * add errors. (If the operation was a 383 * delete, then udata should be cleared.) */ 384 if (events[i].udata) { 385 /* The operation was an add: 386 * report the error as a read. */ 387 which |= EV_READ; 388 break; 389 } else { 390 /* The operation was a del: 391 * report nothing. */ 392 continue; 393 } 394 395 /* Other errors shouldn't occur. */ 396 default: 397 errno = events[i].data; 398 return (-1); 399 } 400 } else if (events[i].filter == EVFILT_READ) { 401 which |= EV_READ; 402 } else if (events[i].filter == EVFILT_WRITE) { 403 which |= EV_WRITE; 404 } else if (events[i].filter == EVFILT_SIGNAL) { 405 which |= EV_SIGNAL; 406#ifdef EVFILT_USER 407 } else if (events[i].filter == EVFILT_USER) { 408 base->is_notify_pending = 0; 409#endif 410 } 411 412 if (!which) 413 continue; 414 415 if (events[i].filter == EVFILT_SIGNAL) { 416 evmap_signal_active_(base, events[i].ident, 1); 417 } else { 418 evmap_io_active_(base, events[i].ident, which | EV_ET); 419 } 420 } 421 422 if (res == kqop->events_size) { 423 /* We used all the events space that we have. Maybe we should 424 make it bigger. */ 425 kq_grow_events(kqop, kqop->events_size * 2); 426 } 427 428 return (0); 429} 430 431static void 432kqop_free(struct kqop *kqop) 433{ 434 if (kqop->changes) 435 mm_free(kqop->changes); 436 if (kqop->events) 437 mm_free(kqop->events); 438 if (kqop->kq >= 0 && kqop->pid == getpid()) 439 close(kqop->kq); 440 memset(kqop, 0, sizeof(struct kqop)); 441 mm_free(kqop); 442} 443 444static void 445kq_dealloc(struct event_base *base) 446{ 447 struct kqop *kqop = base->evbase; 448 evsig_dealloc_(base); 449 kqop_free(kqop); 450} 451 452/* signal handling */ 453static int 454kq_sig_add(struct event_base *base, int nsignal, short old, short events, void *p) 455{ 456 struct kqop *kqop = base->evbase; 457 struct kevent kev; 458 struct timespec timeout = { 0, 0 }; 459 (void)p; 460 461 EVUTIL_ASSERT(nsignal >= 0 && nsignal < NSIG); 462 463 memset(&kev, 0, sizeof(kev)); 464 kev.ident = nsignal; 465 kev.filter = EVFILT_SIGNAL; 466 kev.flags = EV_ADD; 467 468 /* Be ready for the signal if it is sent any 469 * time between now and the next call to 470 * kq_dispatch. */ 471 if (kevent(kqop->kq, &kev, 1, NULL, 0, &timeout) == -1) 472 return (-1); 473 474 /* We can set the handler for most signals to SIG_IGN and 475 * still have them reported to us in the queue. However, 476 * if the handler for SIGCHLD is SIG_IGN, the system reaps 477 * zombie processes for us, and we don't get any notification. 478 * This appears to be the only signal with this quirk. */ 479 if (evsig_set_handler_(base, nsignal, 480 nsignal == SIGCHLD ? SIG_DFL : SIG_IGN) == -1) 481 return (-1); 482 483 return (0); 484} 485 486static int 487kq_sig_del(struct event_base *base, int nsignal, short old, short events, void *p) 488{ 489 struct kqop *kqop = base->evbase; 490 struct kevent kev; 491 492 struct timespec timeout = { 0, 0 }; 493 (void)p; 494 495 EVUTIL_ASSERT(nsignal >= 0 && nsignal < NSIG); 496 497 memset(&kev, 0, sizeof(kev)); 498 kev.ident = nsignal; 499 kev.filter = EVFILT_SIGNAL; 500 kev.flags = EV_DELETE; 501 502 /* Because we insert signal events 503 * immediately, we need to delete them 504 * immediately, too */ 505 if (kevent(kqop->kq, &kev, 1, NULL, 0, &timeout) == -1) 506 return (-1); 507 508 if (evsig_restore_handler_(base, nsignal) == -1) 509 return (-1); 510 511 return (0); 512} 513 514 515/* OSX 10.6 and FreeBSD 8.1 add support for EVFILT_USER, which we can use 516 * to wake up the event loop from another thread. */ 517 518/* Magic number we use for our filter ID. */ 519#define NOTIFY_IDENT 42 520 521int 522event_kq_add_notify_event_(struct event_base *base) 523{ 524 struct kqop *kqop = base->evbase; 525#if defined(EVFILT_USER) && defined(NOTE_TRIGGER) 526 struct kevent kev; 527 struct timespec timeout = { 0, 0 }; 528#endif 529 530 if (kqop->notify_event_added) 531 return 0; 532 533#if defined(EVFILT_USER) && defined(NOTE_TRIGGER) 534 memset(&kev, 0, sizeof(kev)); 535 kev.ident = NOTIFY_IDENT; 536 kev.filter = EVFILT_USER; 537 kev.flags = EV_ADD | EV_CLEAR; 538 539 if (kevent(kqop->kq, &kev, 1, NULL, 0, &timeout) == -1) { 540 event_warn("kevent: adding EVFILT_USER event"); 541 return -1; 542 } 543 544 kqop->notify_event_added = 1; 545 546 return 0; 547#else 548 return -1; 549#endif 550} 551 552int 553event_kq_notify_base_(struct event_base *base) 554{ 555 struct kqop *kqop = base->evbase; 556#if defined(EVFILT_USER) && defined(NOTE_TRIGGER) 557 struct kevent kev; 558 struct timespec timeout = { 0, 0 }; 559#endif 560 if (! kqop->notify_event_added) 561 return -1; 562 563#if defined(EVFILT_USER) && defined(NOTE_TRIGGER) 564 memset(&kev, 0, sizeof(kev)); 565 kev.ident = NOTIFY_IDENT; 566 kev.filter = EVFILT_USER; 567 kev.fflags = NOTE_TRIGGER; 568 569 if (kevent(kqop->kq, &kev, 1, NULL, 0, &timeout) == -1) { 570 event_warn("kevent: triggering EVFILT_USER event"); 571 return -1; 572 } 573 574 return 0; 575#else 576 return -1; 577#endif 578} 579 580#endif /* EVENT__HAVE_KQUEUE */ 581