1 /* $NetBSD: kqueue.c,v 1.1.1.2 2013/04/11 16:43:24 christos Exp $ */ 2 /* $OpenBSD: kqueue.c,v 1.5 2002/07/10 14:41:31 art Exp $ */ 3 4 /* 5 * Copyright 2000-2007 Niels Provos <provos@citi.umich.edu> 6 * Copyright 2007-2012 Niels Provos and Nick Mathewson 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 3. The name of the author may not be used to endorse or promote products 17 * derived from this software without specific prior written permission. 18 * 19 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 20 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 21 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 22 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 23 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 24 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 25 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 26 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 27 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 28 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 29 */ 30 #include "event2/event-config.h" 31 #include <sys/cdefs.h> 32 __RCSID("$NetBSD: kqueue.c,v 1.1.1.2 2013/04/11 16:43:24 christos Exp $"); 33 34 #define _GNU_SOURCE 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 #define NEVENT 64 71 72 struct kqop { 73 struct kevent *changes; 74 int changes_size; 75 76 struct kevent *events; 77 int events_size; 78 int kq; 79 pid_t pid; 80 }; 81 82 static void kqop_free(struct kqop *kqop); 83 84 static void *kq_init(struct event_base *); 85 static int kq_sig_add(struct event_base *, int, short, short, void *); 86 static int kq_sig_del(struct event_base *, int, short, short, void *); 87 static int kq_dispatch(struct event_base *, struct timeval *); 88 static void kq_dealloc(struct event_base *); 89 90 const struct eventop kqops = { 91 "kqueue", 92 kq_init, 93 event_changelist_add, 94 event_changelist_del, 95 kq_dispatch, 96 kq_dealloc, 97 1 /* need reinit */, 98 EV_FEATURE_ET|EV_FEATURE_O1|EV_FEATURE_FDS, 99 EVENT_CHANGELIST_FDINFO_SIZE 100 }; 101 102 static const struct eventop kqsigops = { 103 "kqueue_signal", 104 NULL, 105 kq_sig_add, 106 kq_sig_del, 107 NULL, 108 NULL, 109 1 /* need reinit */, 110 0, 111 0 112 }; 113 114 static void * 115 kq_init(struct event_base *base) 116 { 117 int kq = -1; 118 struct kqop *kqueueop = NULL; 119 120 if (!(kqueueop = mm_calloc(1, sizeof(struct kqop)))) 121 return (NULL); 122 123 /* Initialize the kernel queue */ 124 125 if ((kq = kqueue()) == -1) { 126 event_warn("kqueue"); 127 goto err; 128 } 129 130 kqueueop->kq = kq; 131 132 kqueueop->pid = getpid(); 133 134 /* Initialize fields */ 135 kqueueop->changes = mm_calloc(NEVENT, sizeof(struct kevent)); 136 if (kqueueop->changes == NULL) 137 goto err; 138 kqueueop->events = mm_calloc(NEVENT, sizeof(struct kevent)); 139 if (kqueueop->events == NULL) 140 goto err; 141 kqueueop->events_size = kqueueop->changes_size = NEVENT; 142 143 /* Check for Mac OS X kqueue bug. */ 144 memset(&kqueueop->changes[0], 0, sizeof kqueueop->changes[0]); 145 kqueueop->changes[0].ident = -1; 146 kqueueop->changes[0].filter = EVFILT_READ; 147 kqueueop->changes[0].flags = EV_ADD; 148 /* 149 * If kqueue works, then kevent will succeed, and it will 150 * stick an error in events[0]. If kqueue is broken, then 151 * kevent will fail. 152 */ 153 if (kevent(kq, 154 kqueueop->changes, 1, kqueueop->events, NEVENT, NULL) != 1 || 155 (int)kqueueop->events[0].ident != -1 || 156 kqueueop->events[0].flags != EV_ERROR) { 157 event_warn("%s: detected broken kqueue; not using.", __func__); 158 goto err; 159 } 160 161 base->evsigsel = &kqsigops; 162 163 return (kqueueop); 164 err: 165 if (kqueueop) 166 kqop_free(kqueueop); 167 168 return (NULL); 169 } 170 171 static void 172 kq_sighandler(int sig) 173 { 174 /* Do nothing here */ 175 } 176 177 #define ADD_UDATA 0x30303 178 179 static void 180 kq_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 } 203 204 static int 205 kq_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 = kqop->changes_size * 2; 216 struct kevent *newchanges; 217 218 newchanges = mm_realloc(kqop->changes, 219 newsize * sizeof(struct kevent)); 220 if (newchanges == NULL) { 221 event_warn("%s: realloc", __func__); 222 return (-1); 223 } 224 kqop->changes = newchanges; 225 kqop->changes_size = newsize; 226 } 227 if (in_ch->read_change) { 228 out_ch = &kqop->changes[n_changes++]; 229 kq_setup_kevent(out_ch, in_ch->fd, EVFILT_READ, 230 in_ch->read_change); 231 } 232 if (in_ch->write_change) { 233 out_ch = &kqop->changes[n_changes++]; 234 kq_setup_kevent(out_ch, in_ch->fd, EVFILT_WRITE, 235 in_ch->write_change); 236 } 237 } 238 return n_changes; 239 } 240 241 static int 242 kq_grow_events(struct kqop *kqop, size_t new_size) 243 { 244 struct kevent *newresult; 245 246 newresult = mm_realloc(kqop->events, 247 new_size * sizeof(struct kevent)); 248 249 if (newresult) { 250 kqop->events = newresult; 251 kqop->events_size = new_size; 252 return 0; 253 } else { 254 return -1; 255 } 256 } 257 258 static int 259 kq_dispatch(struct event_base *base, struct timeval *tv) 260 { 261 struct kqop *kqop = base->evbase; 262 struct kevent *events = kqop->events; 263 struct kevent *changes; 264 struct timespec ts, *ts_p = NULL; 265 int i, n_changes, res; 266 267 if (tv != NULL) { 268 TIMEVAL_TO_TIMESPEC(tv, &ts); 269 ts_p = &ts; 270 } 271 272 /* Build "changes" from "base->changes" */ 273 EVUTIL_ASSERT(kqop->changes); 274 n_changes = kq_build_changes_list(&base->changelist, kqop); 275 if (n_changes < 0) 276 return -1; 277 278 event_changelist_remove_all(&base->changelist, base); 279 280 /* steal the changes array in case some broken code tries to call 281 * dispatch twice at once. */ 282 changes = kqop->changes; 283 kqop->changes = NULL; 284 285 /* Make sure that 'events' is at least as long as the list of changes: 286 * otherwise errors in the changes can get reported as a -1 return 287 * value from kevent() rather than as EV_ERROR events in the events 288 * array. 289 * 290 * (We could instead handle -1 return values from kevent() by 291 * retrying with a smaller changes array or a larger events array, 292 * but this approach seems less risky for now.) 293 */ 294 if (kqop->events_size < n_changes) { 295 int new_size = kqop->events_size; 296 do { 297 new_size *= 2; 298 } while (new_size < n_changes); 299 300 kq_grow_events(kqop, new_size); 301 events = kqop->events; 302 } 303 304 EVBASE_RELEASE_LOCK(base, th_base_lock); 305 306 res = kevent(kqop->kq, changes, n_changes, 307 events, kqop->events_size, ts_p); 308 309 EVBASE_ACQUIRE_LOCK(base, th_base_lock); 310 311 EVUTIL_ASSERT(kqop->changes == NULL); 312 kqop->changes = changes; 313 314 if (res == -1) { 315 if (errno != EINTR) { 316 event_warn("kevent"); 317 return (-1); 318 } 319 320 return (0); 321 } 322 323 event_debug(("%s: kevent reports %d", __func__, res)); 324 325 for (i = 0; i < res; i++) { 326 int which = 0; 327 328 if (events[i].flags & EV_ERROR) { 329 switch (events[i].data) { 330 331 /* Can occur on delete if we are not currently 332 * watching any events on this fd. That can 333 * happen when the fd was closed and another 334 * file was opened with that fd. */ 335 case ENOENT: 336 /* Can occur for reasons not fully understood 337 * on FreeBSD. */ 338 case EINVAL: 339 continue; 340 341 /* Can occur on a delete if the fd is closed. */ 342 case EBADF: 343 /* XXXX On NetBSD, we can also get EBADF if we 344 * try to add the write side of a pipe, but 345 * the read side has already been closed. 346 * Other BSDs call this situation 'EPIPE'. It 347 * would be good if we had a way to report 348 * this situation. */ 349 continue; 350 /* These two can occur on an add if the fd was one side 351 * of a pipe, and the other side was closed. */ 352 case EPERM: 353 case EPIPE: 354 /* Report read events, if we're listening for 355 * them, so that the user can learn about any 356 * add errors. (If the operation was a 357 * delete, then udata should be cleared.) */ 358 if (events[i].udata) { 359 /* The operation was an add: 360 * report the error as a read. */ 361 which |= EV_READ; 362 break; 363 } else { 364 /* The operation was a del: 365 * report nothing. */ 366 continue; 367 } 368 369 /* Other errors shouldn't occur. */ 370 default: 371 errno = events[i].data; 372 return (-1); 373 } 374 } else if (events[i].filter == EVFILT_READ) { 375 which |= EV_READ; 376 } else if (events[i].filter == EVFILT_WRITE) { 377 which |= EV_WRITE; 378 } else if (events[i].filter == EVFILT_SIGNAL) { 379 which |= EV_SIGNAL; 380 } 381 382 if (!which) 383 continue; 384 385 if (events[i].filter == EVFILT_SIGNAL) { 386 evmap_signal_active(base, events[i].ident, 1); 387 } else { 388 evmap_io_active(base, events[i].ident, which | EV_ET); 389 } 390 } 391 392 if (res == kqop->events_size) { 393 /* We used all the events space that we have. Maybe we should 394 make it bigger. */ 395 kq_grow_events(kqop, kqop->events_size * 2); 396 } 397 398 return (0); 399 } 400 401 static void 402 kqop_free(struct kqop *kqop) 403 { 404 if (kqop->changes) 405 mm_free(kqop->changes); 406 if (kqop->events) 407 mm_free(kqop->events); 408 if (kqop->kq >= 0 && kqop->pid == getpid()) 409 close(kqop->kq); 410 memset(kqop, 0, sizeof(struct kqop)); 411 mm_free(kqop); 412 } 413 414 static void 415 kq_dealloc(struct event_base *base) 416 { 417 struct kqop *kqop = base->evbase; 418 evsig_dealloc(base); 419 kqop_free(kqop); 420 } 421 422 /* signal handling */ 423 static int 424 kq_sig_add(struct event_base *base, int nsignal, short old, short events, void *p) 425 { 426 struct kqop *kqop = base->evbase; 427 struct kevent kev; 428 struct timespec timeout = { 0, 0 }; 429 (void)p; 430 431 EVUTIL_ASSERT(nsignal >= 0 && nsignal < NSIG); 432 433 memset(&kev, 0, sizeof(kev)); 434 kev.ident = nsignal; 435 kev.filter = EVFILT_SIGNAL; 436 kev.flags = EV_ADD; 437 438 /* Be ready for the signal if it is sent any 439 * time between now and the next call to 440 * kq_dispatch. */ 441 if (kevent(kqop->kq, &kev, 1, NULL, 0, &timeout) == -1) 442 return (-1); 443 444 /* XXXX The manpage suggest we could use SIG_IGN instead of a 445 * do-nothing handler */ 446 if (_evsig_set_handler(base, nsignal, kq_sighandler) == -1) 447 return (-1); 448 449 return (0); 450 } 451 452 static int 453 kq_sig_del(struct event_base *base, int nsignal, short old, short events, void *p) 454 { 455 struct kqop *kqop = base->evbase; 456 struct kevent kev; 457 458 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_DELETE; 467 468 /* Because we insert signal events 469 * immediately, we need to delete them 470 * immediately, too */ 471 if (kevent(kqop->kq, &kev, 1, NULL, 0, &timeout) == -1) 472 return (-1); 473 474 if (_evsig_restore_handler(base, nsignal) == -1) 475 return (-1); 476 477 return (0); 478 } 479