2 * This file is part of DisOrder.
3 * Copyright (C) 2004, 2005, 2007 Richard Kettlewell
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License for more details.
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307
21 * @brief DisOrder event loop
29 #include <sys/types.h>
30 #include <sys/resource.h>
39 #include <sys/socket.h>
40 #include <netinet/in.h>
51 /** @brief A timeout */
55 ev_timeout_callback *callback;
60 /** @brief A file descriptor in one mode */
63 ev_fd_callback *callback;
68 /** @brief All the file descriptors in a given mode */
70 /** @brief Mask of active file descriptors passed to @c select() */
73 /** @brief File descriptor mask returned from @c select() */
76 /** @brief Number of file descriptors in @p fds */
79 /** @brief Number of slots in @p fds */
82 /** @brief Array of all active file descriptors */
85 /** @brief Highest-numbered file descriptor or 0 */
89 /** @brief A signal handler */
91 struct sigaction oldsa;
92 ev_signal_callback *callback;
96 /** @brief A child process */
100 ev_child_callback *callback;
104 /** @brief An event loop */
106 /** @brief File descriptors, per mode */
107 struct fdmode mode[ev_nmodes];
109 /** @brief Sorted linked list of timeouts
111 * We could use @ref HEAP_TYPE now, but there aren't many timeouts.
113 struct timeout *timeouts;
115 /** @brief Array of handled signals */
116 struct signal signals[NSIG];
118 /** @brief Mask of handled signals */
121 /** @brief Escape early from handling of @c select() results
123 * This is set if any of the file descriptor arrays are invalidated, since
124 * it's then not safe for processing of them to continue.
128 /** @brief Signal handling pipe
130 * The signal handle writes signal numbers down this pipe.
134 /** @brief Number of child processes in @p children */
137 /** @brief Number of slots in @p children */
140 /** @brief Array of child processes */
141 struct child *children;
144 /** @brief Names of file descriptor modes */
145 static const char *modenames[] = { "read", "write", "except" };
147 /* utilities ******************************************************************/
149 /** @brief Great-than comparison for timevals
151 * Ought to be in @file lib/timeval.h
153 static inline int gt(const struct timeval *a, const struct timeval *b) {
154 if(a->tv_sec > b->tv_sec)
156 if(a->tv_sec == b->tv_sec
157 && a->tv_usec > b->tv_usec)
162 /** @brief Greater-than-or-equal comparison for timevals
164 * Ought to be in @file lib/timeval.h
166 static inline int ge(const struct timeval *a, const struct timeval *b) {
170 /* creation *******************************************************************/
172 /** @brief Create a new event loop */
173 ev_source *ev_new(void) {
174 ev_source *ev = xmalloc(sizeof *ev);
177 memset(ev, 0, sizeof *ev);
178 for(n = 0; n < ev_nmodes; ++n)
179 FD_ZERO(&ev->mode[n].enabled);
180 ev->sigpipe[0] = ev->sigpipe[1] = -1;
181 sigemptyset(&ev->sigmask);
185 /* event loop *****************************************************************/
187 /** @brief Run the event loop
188 * @return -1 on error, non-0 if any callback returned non-0
190 int ev_run(ev_source *ev) {
193 struct timeval delta;
197 struct timeout *t, **tt;
200 xgettimeofday(&now, 0);
201 /* Handle timeouts. We don't want to handle any timeouts that are added
202 * while we're handling them (otherwise we'd have to break out of infinite
203 * loops, preferrably without starving better-behaved subsystems). Hence
204 * the slightly complicated two-phase approach here. */
205 for(t = ev->timeouts;
206 t && ge(&now, &t->when);
209 D(("calling timeout for %ld.%ld callback %p %p",
210 (long)t->when.tv_sec, (long)t->when.tv_usec,
211 (void *)t->callback, t->u));
212 ret = t->callback(ev, &now, t->u);
224 for(mode = 0; mode < ev_nmodes; ++mode) {
225 ev->mode[mode].tripped = ev->mode[mode].enabled;
226 if(ev->mode[mode].maxfd > maxfd)
227 maxfd = ev->mode[mode].maxfd;
229 xsigprocmask(SIG_UNBLOCK, &ev->sigmask, 0);
232 xgettimeofday(&now, 0);
233 delta.tv_sec = ev->timeouts->when.tv_sec - now.tv_sec;
234 delta.tv_usec = ev->timeouts->when.tv_usec - now.tv_usec;
235 if(delta.tv_usec < 0) {
236 delta.tv_usec += 1000000;
240 delta.tv_sec = delta.tv_usec = 0;
241 n = select(maxfd + 1,
242 &ev->mode[ev_read].tripped,
243 &ev->mode[ev_write].tripped,
244 &ev->mode[ev_except].tripped,
247 n = select(maxfd + 1,
248 &ev->mode[ev_read].tripped,
249 &ev->mode[ev_write].tripped,
250 &ev->mode[ev_except].tripped,
253 } while(n < 0 && errno == EINTR);
254 xsigprocmask(SIG_BLOCK, &ev->sigmask, 0);
256 error(errno, "error calling select");
258 /* If there's a bad FD in the mix then check them all and log what we
259 * find, to ease debugging */
260 for(mode = 0; mode < ev_nmodes; ++mode) {
261 for(n = 0; n < ev->mode[mode].nfds; ++n) {
262 const int fd = ev->mode[mode].fds[n].fd;
264 if(FD_ISSET(fd, &ev->mode[mode].enabled)
265 && fstat(fd, &sb) < 0)
266 error(errno, "mode %s fstat %d (%s)",
267 modenames[mode], fd, ev->mode[mode].fds[n].what);
269 for(n = 0; n <= maxfd; ++n)
270 if(FD_ISSET(n, &ev->mode[mode].enabled)
271 && fstat(n, &sb) < 0)
272 error(errno, "mode %s fstat %d", modenames[mode], n);
278 /* if anything deranges the meaning of an fd, or re-orders the
279 * fds[] tables, we'd better give up; such operations will
280 * therefore set @escape@. */
282 for(mode = 0; mode < ev_nmodes && !ev->escape; ++mode)
283 for(n = 0; n < ev->mode[mode].nfds && !ev->escape; ++n) {
284 int fd = ev->mode[mode].fds[n].fd;
285 if(FD_ISSET(fd, &ev->mode[mode].tripped)) {
286 D(("calling %s fd %d callback %p %p", modenames[mode], fd,
287 (void *)ev->mode[mode].fds[n].callback,
288 ev->mode[mode].fds[n].u));
289 ret = ev->mode[mode].fds[n].callback(ev, fd,
290 ev->mode[mode].fds[n].u);
296 /* we'll pick up timeouts back round the loop */
300 /* file descriptors ***********************************************************/
302 /** @brief Register a file descriptor
303 * @param ev Event loop
304 * @param mode @c ev_read or @c ev_write
305 * @param fd File descriptor
306 * @param callback Called when @p is readable/writable
307 * @param u Passed to @p callback
308 * @param what Text description
309 * @return 0 on success, non-0 on error
311 * Sets @ref ev_source::escape, so no further processing of file descriptors
312 * will occur this time round the event loop.
314 int ev_fd(ev_source *ev,
317 ev_fd_callback *callback,
322 D(("registering %s fd %d callback %p %p", modenames[mode], fd,
323 (void *)callback, u));
324 assert(mode < ev_nmodes);
325 if(ev->mode[mode].nfds >= ev->mode[mode].fdslots) {
326 ev->mode[mode].fdslots = (ev->mode[mode].fdslots
327 ? 2 * ev->mode[mode].fdslots : 16);
328 D(("expanding %s fd table to %d entries", modenames[mode],
329 ev->mode[mode].fdslots));
330 ev->mode[mode].fds = xrealloc(ev->mode[mode].fds,
331 ev->mode[mode].fdslots * sizeof (struct fd));
333 n = ev->mode[mode].nfds++;
334 FD_SET(fd, &ev->mode[mode].enabled);
335 ev->mode[mode].fds[n].fd = fd;
336 ev->mode[mode].fds[n].callback = callback;
337 ev->mode[mode].fds[n].u = u;
338 ev->mode[mode].fds[n].what = what;
339 if(fd > ev->mode[mode].maxfd)
340 ev->mode[mode].maxfd = fd;
345 /** @brief Cancel a file descriptor
346 * @param ev Event loop
347 * @param mode @c ev_read or @c ev_write
348 * @param fd File descriptor
349 * @return 0 on success, non-0 on error
351 * Sets @ref ev_source::escape, so no further processing of file descriptors
352 * will occur this time round the event loop.
354 int ev_fd_cancel(ev_source *ev, ev_fdmode mode, int fd) {
358 D(("cancelling mode %s fd %d", modenames[mode], fd));
359 /* find the right struct fd */
360 for(n = 0; n < ev->mode[mode].nfds && fd != ev->mode[mode].fds[n].fd; ++n)
362 assert(n < ev->mode[mode].nfds);
363 /* swap in the last fd and reduce the count */
364 if(n != ev->mode[mode].nfds - 1)
365 ev->mode[mode].fds[n] = ev->mode[mode].fds[ev->mode[mode].nfds - 1];
366 --ev->mode[mode].nfds;
367 /* if that was the biggest fd, find the new biggest one */
368 if(fd == ev->mode[mode].maxfd) {
370 for(n = 0; n < ev->mode[mode].nfds; ++n)
371 if(ev->mode[mode].fds[n].fd > maxfd)
372 maxfd = ev->mode[mode].fds[n].fd;
373 ev->mode[mode].maxfd = maxfd;
375 /* don't tell select about this fd any more */
376 FD_CLR(fd, &ev->mode[mode].enabled);
381 /** @brief Re-enable a file descriptor
382 * @param ev Event loop
383 * @param mode @c ev_read or @c ev_write
384 * @param fd File descriptor
385 * @return 0 on success, non-0 on error
387 * It is harmless if @p fd is currently disabled, but it must not have been
390 int ev_fd_enable(ev_source *ev, ev_fdmode mode, int fd) {
392 D(("enabling mode %s fd %d", modenames[mode], fd));
393 FD_SET(fd, &ev->mode[mode].enabled);
397 /** @brief Temporarily disable a file descriptor
398 * @param ev Event loop
399 * @param mode @c ev_read or @c ev_write
400 * @param fd File descriptor
401 * @return 0 on success, non-0 on error
403 * Re-enable with ev_fd_enable(). It is harmless if @p fd is already disabled,
404 * but it must not have been cancelled.
406 int ev_fd_disable(ev_source *ev, ev_fdmode mode, int fd) {
407 D(("disabling mode %s fd %d", modenames[mode], fd));
408 FD_CLR(fd, &ev->mode[mode].enabled);
409 FD_CLR(fd, &ev->mode[mode].tripped);
410 /* Suppress any pending callbacks */
415 /** @brief Log a report of file descriptor state */
416 void ev_report(ev_source *ev) {
423 for(mode = 0; mode < ev_nmodes; ++mode) {
424 info("mode %s maxfd %d", modenames[mode], ev->mode[mode].maxfd);
425 for(n = 0; n < ev->mode[mode].nfds; ++n) {
426 fd = ev->mode[mode].fds[n].fd;
427 info("fd %s %d%s%s (%s)", modenames[mode], fd,
428 FD_ISSET(fd, &ev->mode[mode].enabled) ? " enabled" : "",
429 FD_ISSET(fd, &ev->mode[mode].tripped) ? " tripped" : "",
430 ev->mode[mode].fds[n].what);
433 for(fd = 0; fd <= ev->mode[mode].maxfd; ++fd) {
434 if(!FD_ISSET(fd, &ev->mode[mode].enabled))
436 for(n = 0; n < ev->mode[mode].nfds; ++n) {
437 if(ev->mode[mode].fds[n].fd == fd)
440 if(n < ev->mode[mode].nfds)
441 snprintf(b, sizeof b, "%d(%s)", fd, ev->mode[mode].fds[n].what);
443 snprintf(b, sizeof b, "%d", fd);
444 dynstr_append(d, ' ');
445 dynstr_append_string(d, b);
448 info("%s enabled:%s", modenames[mode], d->vec);
452 /* timeouts *******************************************************************/
454 /** @brief Register a timeout
455 * @param ev Event source
456 * @param handle Where to store timeout handle, or @c NULL
457 * @param when Earliest time to call @p callback, or @c NULL
458 * @param callback Function to call at or after @p when
459 * @param u Passed to @p callback
460 * @return 0 on success, non-0 on error
462 * If @p when is a null pointer then a time of 0 is assumed. The effect is to
463 * call the timeout handler from ev_run() next time around the event loop.
464 * This is used internally to schedule various operations if it is not
465 * convenient to call them from the current place in the call stack, or
466 * externally to ensure that other clients of the event loop get a look in when
467 * performing some lengthy operation.
469 int ev_timeout(ev_source *ev,
470 ev_timeout_handle *handlep,
471 const struct timeval *when,
472 ev_timeout_callback *callback,
474 struct timeout *t, *p, **pp;
476 D(("registering timeout at %ld.%ld callback %p %p",
477 when ? (long)when->tv_sec : 0, when ? (long)when->tv_usec : 0,
478 (void *)callback, u));
479 t = xmalloc(sizeof *t);
482 t->callback = callback;
485 while((p = *pp) && gt(&t->when, &p->when))
494 /** @brief Cancel a timeout
495 * @param ev Event loop
496 * @param handle Handle returned from ev_timeout(), or 0
497 * @return 0 on success, non-0 on error
499 * If @p handle is 0 then this is a no-op.
501 int ev_timeout_cancel(ev_source *ev,
502 ev_timeout_handle handle) {
503 struct timeout *t = handle, *p, **pp;
507 for(pp = &ev->timeouts; (p = *pp) && p != t; pp = &p->next)
516 /* signals ********************************************************************/
518 /** @brief Mapping of signals to pipe write ends
520 * The pipes are per-event loop, it's possible in theory for there to be
521 * multiple event loops (e.g. in different threads), although in fact DisOrder
524 static int sigfd[NSIG];
526 /** @brief The signal handler
527 * @param s Signal number
529 * Writes to @c sigfd[s].
531 static void sighandler(int s) {
532 unsigned char sc = s;
533 static const char errmsg[] = "error writing to signal pipe";
535 /* probably the reader has stopped listening for some reason */
536 if(write(sigfd[s], &sc, 1) < 0) {
537 write(2, errmsg, sizeof errmsg - 1);
542 /** @brief Read callback for signals */
543 static int signal_read(ev_source *ev,
544 int attribute((unused)) fd,
545 void attribute((unused)) *u) {
550 if((n = read(ev->sigpipe[0], &s, 1)) == 1)
551 if((ret = ev->signals[s].callback(ev, s, ev->signals[s].u)))
554 if(n < 0 && (errno != EINTR && errno != EAGAIN)) {
555 error(errno, "error reading from signal pipe %d", ev->sigpipe[0]);
561 /** @brief Close the signal pipe */
562 static void close_sigpipe(ev_source *ev) {
563 int save_errno = errno;
565 xclose(ev->sigpipe[0]);
566 xclose(ev->sigpipe[1]);
567 ev->sigpipe[0] = ev->sigpipe[1] = -1;
571 /** @brief Register a signal handler
572 * @param ev Event loop
573 * @param sig Signal to handle
574 * @param callback Called when signal is delivered
575 * @param u Passed to @p callback
576 * @return 0 on success, non-0 on error
578 * Note that @p callback is called from inside ev_run(), not from inside the
579 * signal handler, so the usual restrictions on signal handlers do not apply.
581 int ev_signal(ev_source *ev,
583 ev_signal_callback *callback,
588 D(("registering signal %d handler callback %p %p", sig, (void *)callback, u));
591 assert(sig <= UCHAR_MAX);
592 if(ev->sigpipe[0] == -1) {
593 D(("creating signal pipe"));
595 D(("signal pipe is %d, %d", ev->sigpipe[0], ev->sigpipe[1]));
596 for(n = 0; n < 2; ++n) {
597 nonblock(ev->sigpipe[n]);
598 cloexec(ev->sigpipe[n]);
600 if(ev_fd(ev, ev_read, ev->sigpipe[0], signal_read, 0, "sigpipe read")) {
605 sigaddset(&ev->sigmask, sig);
606 xsigprocmask(SIG_BLOCK, &ev->sigmask, 0);
607 sigfd[sig] = ev->sigpipe[1];
608 ev->signals[sig].callback = callback;
609 ev->signals[sig].u = u;
610 sa.sa_handler = sighandler;
611 sigfillset(&sa.sa_mask);
612 sa.sa_flags = SA_RESTART;
613 xsigaction(sig, &sa, &ev->signals[sig].oldsa);
618 /** @brief Cancel a signal handler
619 * @param ev Event loop
620 * @param sig Signal to cancel
621 * @return 0 on success, non-0 on error
623 int ev_signal_cancel(ev_source *ev,
627 xsigaction(sig, &ev->signals[sig].oldsa, 0);
628 ev->signals[sig].callback = 0;
630 sigdelset(&ev->sigmask, sig);
633 xsigprocmask(SIG_UNBLOCK, &ss, 0);
637 /** @brief Clean up signal handling
638 * @param ev Event loop
640 * This function can be called from inside a fork. It restores signal
641 * handlers, unblocks the signals, and closes the signal pipe for @p ev.
643 void ev_signal_atfork(ev_source *ev) {
646 if(ev->sigpipe[0] != -1) {
647 /* revert any handled signals to their original state */
648 for(sig = 1; sig < NSIG; ++sig) {
649 if(ev->signals[sig].callback != 0)
650 xsigaction(sig, &ev->signals[sig].oldsa, 0);
652 /* and then unblock them */
653 xsigprocmask(SIG_UNBLOCK, &ev->sigmask, 0);
654 /* don't want a copy of the signal pipe open inside the fork */
655 xclose(ev->sigpipe[0]);
656 xclose(ev->sigpipe[1]);
660 /* child processes ************************************************************/
662 /** @brief Called on SIGCHLD */
663 static int sigchld_callback(ev_source *ev,
664 int attribute((unused)) sig,
665 void attribute((unused)) *u) {
668 int status, n, ret, revisit;
672 for(n = 0; n < ev->nchildren; ++n) {
673 r = wait4(ev->children[n].pid,
675 ev->children[n].options | WNOHANG,
678 ev_child_callback *c = ev->children[n].callback;
679 void *cu = ev->children[n].u;
681 if(WIFEXITED(status) || WIFSIGNALED(status))
682 ev_child_cancel(ev, r);
684 if((ret = c(ev, r, status, &ru, cu)))
687 /* We should "never" get an ECHILD but it can in fact happen. For
688 * instance on Linux 2.4.31, and probably other versions, if someone
689 * straces a child process and then a different child process
690 * terminates, when we wait4() the trace process we will get ECHILD
691 * because it has been reparented to strace. Obviously this is a
692 * hopeless design flaw in the tracing infrastructure, but we don't
693 * want the disorder server to bomb out because of it. So we just log
694 * the problem and ignore it.
696 error(errno, "error calling wait4 for PID %lu (broken ptrace?)",
697 (unsigned long)ev->children[n].pid);
706 /** @brief Configure event loop for child process handling
707 * @return 0 on success, non-0 on error
709 * Currently at most one event loop can handle child processes and it must be
710 * distinguished from others by calling this function on it. This could be
711 * fixed but since no process ever makes use of more than one event loop there
714 int ev_child_setup(ev_source *ev) {
715 D(("installing SIGCHLD handler"));
716 return ev_signal(ev, SIGCHLD, sigchld_callback, 0);
719 /** @brief Wait for a child process to terminate
720 * @param ev Event loop
721 * @param pid Process ID of child
722 * @param options Options to pass to @c wait4()
723 * @param callback Called when child terminates (or possibly when it stops)
724 * @param u Passed to @p callback
725 * @return 0 on success, non-0 on error
727 * You must have called ev_child_setup() on @p ev once first.
729 int ev_child(ev_source *ev,
732 ev_child_callback *callback,
736 D(("registering child handling %ld options %d callback %p %p",
737 (long)pid, options, (void *)callback, u));
738 assert(ev->signals[SIGCHLD].callback == sigchld_callback);
739 if(ev->nchildren >= ev->nchildslots) {
740 ev->nchildslots = ev->nchildslots ? 2 * ev->nchildslots : 16;
741 ev->children = xrealloc(ev->children,
742 ev->nchildslots * sizeof (struct child));
745 ev->children[n].pid = pid;
746 ev->children[n].options = options;
747 ev->children[n].callback = callback;
748 ev->children[n].u = u;
752 /** @brief Stop waiting for a child process
753 * @param ev Event loop
754 * @param pid Child process ID
755 * @return 0 on success, non-0 on error
757 int ev_child_cancel(ev_source *ev,
761 for(n = 0; n < ev->nchildren && ev->children[n].pid != pid; ++n)
763 assert(n < ev->nchildren);
764 if(n != ev->nchildren - 1)
765 ev->children[n] = ev->children[ev->nchildren - 1];
770 /* socket listeners ***********************************************************/
772 /** @brief State for a socket listener */
773 struct listen_state {
774 ev_listen_callback *callback;
778 /** @brief Called when a listenign socket is readable */
779 static int listen_callback(ev_source *ev, int fd, void *u) {
780 const struct listen_state *l = u;
783 struct sockaddr_in in;
784 #if HAVE_STRUCT_SOCKADDR_IN6
785 struct sockaddr_in6 in6;
787 struct sockaddr_un un;
793 D(("callback for listener fd %d", fd));
794 while((addrlen = sizeof addr),
795 (newfd = accept(fd, &addr.sa, &addrlen)) >= 0) {
796 if((ret = l->callback(ev, newfd, &addr.sa, addrlen, l->u)))
805 error(errno, "error calling accept");
810 /* XXX on some systems EPROTO should be fatal, but we don't know if
811 * we're running on one of them */
812 error(errno, "error calling accept");
816 fatal(errno, "error calling accept");
819 if(errno != EINTR && errno != EAGAIN)
820 error(errno, "error calling accept");
824 /** @brief Listen on a socket for inbound stream connections
825 * @param ev Event source
826 * @param fd File descriptor of socket
827 * @param callback Called when a new connection arrives
828 * @param u Passed to @p callback
829 * @param what Text description of socket
830 * @return 0 on success, non-0 on error
832 int ev_listen(ev_source *ev,
834 ev_listen_callback *callback,
837 struct listen_state *l = xmalloc(sizeof *l);
839 D(("registering listener fd %d callback %p %p", fd, (void *)callback, u));
840 l->callback = callback;
842 return ev_fd(ev, ev_read, fd, listen_callback, l, what);
845 /** @brief Stop listening on a socket
846 * @param ev Event loop
847 * @param fd File descriptor of socket
848 * @return 0 on success, non-0 on error
850 int ev_listen_cancel(ev_source *ev, int fd) {
851 D(("cancelling listener fd %d", fd));
852 return ev_fd_cancel(ev, ev_read, fd);
855 /* buffer *********************************************************************/
857 /** @brief Buffer structure */
859 char *base, *start, *end, *top;
862 /* @brief Make sure there is @p bytes available at @c b->end */
863 static void buffer_space(struct buffer *b, size_t bytes) {
864 D(("buffer_space %p %p %p %p want %lu",
865 (void *)b->base, (void *)b->start, (void *)b->end, (void *)b->top,
866 (unsigned long)bytes));
867 if(b->start == b->end)
868 b->start = b->end = b->base;
869 if((size_t)(b->top - b->end) < bytes) {
870 if((size_t)((b->top - b->end) + (b->start - b->base)) < bytes) {
871 size_t newspace = b->end - b->start + bytes, n;
874 for(n = 16; n < newspace; n *= 2)
876 newbase = xmalloc_noptr(n);
877 memcpy(newbase, b->start, b->end - b->start);
879 b->end = newbase + (b->end - b->start);
880 b->top = newbase + n;
881 b->start = newbase; /* must be last */
883 memmove(b->base, b->start, b->end - b->start);
884 b->end = b->base + (b->end - b->start);
888 D(("result %p %p %p %p",
889 (void *)b->base, (void *)b->start, (void *)b->end, (void *)b->top));
892 /* readers and writers *******************************************************/
894 /** @brief State structure for a buffered writer */
896 /** @brief Sink used for writing to the buffer */
899 /** @brief Output buffer */
902 /** @brief File descriptor to write to */
905 /** @brief Set if there'll be no more output */
908 /** @brief Error/termination callback */
909 ev_error_callback *callback;
911 /** @brief Passed to @p callback */
914 /** @brief Parent event source */
917 /** @brief Maximum amount of time between succesful writes, 0 = don't care */
919 /** @brief Maximum amount of data to buffer, 0 = don't care */
921 /** @brief Error code to pass to @p callback (see writer_shutdown()) */
923 /** @brief Timeout handle for @p timebound (or 0) */
924 ev_timeout_handle timeout;
926 /** @brief Description of this writer */
929 /** @brief Tied reader or 0 */
933 /** @brief State structure for a buffered reader */
935 /** @brief Input buffer */
937 /** @brief File descriptor read from */
939 /** @brief Called when new data is available */
940 ev_reader_callback *callback;
941 /** @brief Called on error and shutdown */
942 ev_error_callback *error_callback;
943 /** @brief Passed to @p callback and @p error_callback */
945 /** @brief Parent event loop */
947 /** @brief Set when EOF is detected */
949 /** @brief Error code to pass to error callback */
951 /** @brief Tied writer or NULL */
955 /* buffered writer ************************************************************/
957 /** @brief Shut down the writer
959 * This is called to shut down a writer. The error callback is not called
960 * through any other path. Also we do not cancel @p fd from anywhere else,
961 * though we might disable it.
963 * It has the signature of a timeout callback so that it can be called from a
966 * Calls @p callback with @p w->syntherr as the error code (which might be 0).
968 static int writer_shutdown(ev_source *ev,
969 const attribute((unused)) struct timeval *now,
974 return 0; /* already shut down */
975 info("writer_shutdown fd=%d error=%d", w->fd, w->error);
976 ev_timeout_cancel(ev, w->timeout);
977 ev_fd_cancel(ev, ev_write, w->fd);
980 info("found a tied reader");
981 /* If there is a reader still around we just untie it */
982 w->reader->writer = 0;
983 shutdown(w->fd, SHUT_WR); /* there'll be no more writes */
985 info("no tied reader");
986 /* There's no reader so we are free to close the FD */
990 return w->callback(ev, w->error, w->u);
993 /** @brief Called when a writer's @p timebound expires */
994 static int writer_timebound_exceeded(ev_source *ev,
995 const struct timeval *now,
997 ev_writer *const w = u;
999 error(0, "abandoning writer %s because no writes within %ds",
1000 w->what, w->timebound);
1001 w->error = ETIMEDOUT;
1002 return writer_shutdown(ev, now, u);
1005 /** @brief Set the time bound callback (if not set already) */
1006 static void writer_set_timebound(ev_writer *w) {
1007 if(w->timebound && !w->timeout) {
1008 struct timeval when;
1009 ev_source *const ev = w->ev;
1011 xgettimeofday(&when, 0);
1012 when.tv_sec += w->timebound;
1013 ev_timeout(ev, &w->timeout, &when, writer_timebound_exceeded, w);
1017 /** @brief Called when a writer's file descriptor is writable */
1018 static int writer_callback(ev_source *ev, int fd, void *u) {
1019 ev_writer *const w = u;
1022 n = write(fd, w->b.start, w->b.end - w->b.start);
1023 D(("callback for writer fd %d, %ld bytes, n=%d, errno=%d",
1024 fd, (long)(w->b.end - w->b.start), n, errno));
1026 /* Consume bytes from the buffer */
1028 /* Suppress any outstanding timeout */
1029 ev_timeout_cancel(ev, w->timeout);
1031 if(w->b.start == w->b.end) {
1032 /* The buffer is empty */
1034 /* We're done, we can shut down this writer */
1036 return writer_shutdown(ev, 0, w);
1038 /* There might be more to come but we don't need writer_callback() to
1039 * be called for the time being */
1040 ev_fd_disable(ev, ev_write, fd);
1042 /* The buffer isn't empty, set a timeout so we give up if we don't manage
1043 * to write some more within a reasonable time */
1044 writer_set_timebound(w);
1052 return writer_shutdown(ev, 0, w);
1058 /** @brief Write bytes to a writer's buffer
1060 * This is the sink write callback.
1062 * Calls ev_fd_enable() if necessary (i.e. if the buffer was empty but
1065 static int ev_writer_write(struct sink *sk, const void *s, int n) {
1066 ev_writer *w = (ev_writer *)sk;
1069 return 0; /* avoid silliness */
1071 error(0, "ev_writer_write on %s after shutdown", w->what);
1072 if(w->spacebound && w->b.end - w->b.start + n > w->spacebound) {
1073 /* The new buffer contents will exceed the space bound. We assume that the
1074 * remote client has gone away and TCP hasn't noticed yet, or that it's got
1075 * hopelessly stuck. */
1076 error(0, "abandoning writer %s because buffer has reached %td bytes",
1077 w->what, w->b.end - w->b.start);
1078 ev_fd_disable(w->ev, ev_write, w->fd);
1080 return ev_timeout(w->ev, 0, 0, writer_shutdown, w);
1082 /* Make sure there is space */
1083 buffer_space(&w->b, n);
1084 /* If the buffer was formerly empty then we'll need to re-enable the FD */
1085 if(w->b.start == w->b.end)
1086 ev_fd_enable(w->ev, ev_write, w->fd);
1087 memcpy(w->b.end, s, n);
1089 /* Arrange a timeout if there wasn't one set already */
1090 writer_set_timebound(w);
1094 /** @brief Create a new buffered writer
1095 * @param ev Event loop
1096 * @param fd File descriptor to write to
1097 * @param callback Called if an error occurs and when finished
1098 * @param u Passed to @p callback
1099 * @param what Text description
1100 * @return New writer or @c NULL
1102 * Writers own their file descriptor and close it when they have finished with
1105 * If you pass the same fd to a reader and writer, you must tie them together
1108 ev_writer *ev_writer_new(ev_source *ev,
1110 ev_error_callback *callback,
1113 ev_writer *w = xmalloc(sizeof *w);
1115 D(("registering writer fd %d callback %p %p", fd, (void *)callback, u));
1116 w->s.write = ev_writer_write;
1118 w->callback = callback;
1121 w->timebound = 10 * 60;
1122 w->spacebound = 512 * 1024;
1124 if(ev_fd(ev, ev_write, fd, writer_callback, w, what))
1126 /* Buffer is initially empty so we don't want a callback */
1127 ev_fd_disable(ev, ev_write, fd);
1131 /** @brief Get/set the time bound
1133 * @param new_time_bound New bound or -1 for no change
1134 * @return Latest time bound
1136 * If @p new_time_bound is negative then the current time bound is returned.
1137 * Otherwise it is set and the new value returned.
1139 * The time bound is the number of seconds allowed between writes. If it takes
1140 * longer than this to flush a buffer then the peer will be assumed to be dead
1141 * and an error will be synthesized. 0 means "don't care". The default time
1142 * bound is 10 minutes.
1144 * Note that this value does not take into account kernel buffering and
1147 int ev_writer_time_bound(ev_writer *w,
1148 int new_time_bound) {
1149 if(new_time_bound >= 0)
1150 w->timebound = new_time_bound;
1151 return w->timebound;
1154 /** @brief Get/set the space bound
1156 * @param new_space_bound New bound or -1 for no change
1157 * @return Latest space bound
1159 * If @p new_space_bound is negative then the current space bound is returned.
1160 * Otherwise it is set and the new value returned.
1162 * The space bound is the number of bytes allowed between in the buffer. If
1163 * the buffer exceeds this size an error will be synthesized. 0 means "don't
1164 * care". The default space bound is 512Kbyte.
1166 * Note that this value does not take into account kernel buffering.
1168 int ev_writer_space_bound(ev_writer *w,
1169 int new_space_bound) {
1170 if(new_space_bound >= 0)
1171 w->spacebound = new_space_bound;
1172 return w->spacebound;
1175 /** @brief Return the sink associated with a writer
1177 * @return Pointer to sink
1179 * Writing to the sink will arrange for those bytes to be written to the file
1180 * descriptor as and when it is writable.
1182 struct sink *ev_writer_sink(ev_writer *w) {
1184 fatal(0, "ev_write_sink called with null writer");
1188 /** @brief Close a writer
1189 * @param w Writer to close
1190 * @return 0 on success, non-0 on error
1192 * Close a writer. No more bytes should be written to its sink.
1194 * When the last byte has been written the callback will be called with an
1195 * error code of 0. It is guaranteed that this will NOT happen before
1196 * ev_writer_close() returns (although the file descriptor for the writer might
1197 * be cancelled by the time it returns).
1199 int ev_writer_close(ev_writer *w) {
1200 D(("close writer fd %d", w->fd));
1202 return 0; /* already closed */
1204 if(w->b.start == w->b.end) {
1205 /* We're already finished */
1206 w->error = 0; /* no error */
1207 return ev_timeout(w->ev, 0, 0, writer_shutdown, w);
1212 /** @brief Attempt to flush a writer
1213 * @param w Writer to flush
1214 * @return 0 on success, non-0 on error
1216 * Does a speculative write of any buffered data. Does not block if it cannot
1219 int ev_writer_flush(ev_writer *w) {
1220 return writer_callback(w->ev, w->fd, w);
1223 /* buffered reader ************************************************************/
1225 /** @brief Shut down a reader*
1227 * This is the only path through which we cancel and close the file descriptor.
1228 * As with the writer case it is given timeout signature to allow it be
1229 * deferred to the next iteration of the event loop.
1231 * We only call @p error_callback if @p error is nonzero (unlike the writer
1234 static int reader_shutdown(ev_source *ev,
1235 const attribute((unused)) struct timeval *now,
1237 ev_reader *const r = u;
1240 return 0; /* already shut down */
1241 info("reader_shutdown fd=%d", r->fd);
1242 ev_fd_cancel(ev, ev_read, r->fd);
1245 info("found a tied writer");
1246 /* If there is a writer still around we just untie it */
1247 r->writer->reader = 0;
1248 shutdown(r->fd, SHUT_RD); /* there'll be no more reads */
1250 info("no tied writer found");
1251 /* There's no writer so we are free to close the FD */
1256 return r->error_callback(ev, r->error, r->u);
1261 /** @brief Called when a reader's @p fd is readable */
1262 static int reader_callback(ev_source *ev, int fd, void *u) {
1266 buffer_space(&r->b, 1);
1267 n = read(fd, r->b.end, r->b.top - r->b.end);
1268 D(("read fd %d buffer %d returned %d errno %d",
1269 fd, (int)(r->b.top - r->b.end), n, errno));
1272 return r->callback(ev, r, r->b.start, r->b.end - r->b.start, 0, r->u);
1274 /* No more read callbacks needed */
1275 ev_fd_disable(r->ev, ev_read, r->fd);
1276 ev_timeout(r->ev, 0, 0, reader_shutdown, r);
1277 /* Pass the remaining data and an eof indicator to the user */
1278 return r->callback(ev, r, r->b.start, r->b.end - r->b.start, 1, r->u);
1285 /* Fatal error, kill the reader now */
1287 return reader_shutdown(ev, 0, r);
1293 /** @brief Create a new buffered reader
1294 * @param ev Event loop
1295 * @param fd File descriptor to read from
1296 * @param callback Called when new data is available
1297 * @param error_callback Called if an error occurs
1298 * @param u Passed to callbacks
1299 * @param what Text description
1300 * @return New reader or @c NULL
1302 * Readers own their fd and close it when they are finished with it.
1304 * If you pass the same fd to a reader and writer, you must tie them together
1307 ev_reader *ev_reader_new(ev_source *ev,
1309 ev_reader_callback *callback,
1310 ev_error_callback *error_callback,
1313 ev_reader *r = xmalloc(sizeof *r);
1315 D(("registering reader fd %d callback %p %p %p",
1316 fd, (void *)callback, (void *)error_callback, u));
1318 r->callback = callback;
1319 r->error_callback = error_callback;
1322 if(ev_fd(ev, ev_read, fd, reader_callback, r, what))
1327 void ev_reader_buffer(ev_reader *r, size_t nbytes) {
1328 buffer_space(&r->b, nbytes - (r->b.end - r->b.start));
1331 /** @brief Consume @p n bytes from the reader's buffer
1333 * @param n Number of bytes to consume
1335 * Tells the reader than the next @p n bytes have been dealt with and can now
1338 void ev_reader_consume(ev_reader *r, size_t n) {
1342 /** @brief Cancel a reader
1344 * @return 0 on success, non-0 on error
1346 * No further callbacks will be made, and the FD will be closed (in a later
1347 * iteration of the event loop).
1349 int ev_reader_cancel(ev_reader *r) {
1350 D(("cancel reader fd %d", r->fd));
1352 return 0; /* already thoroughly cancelled */
1353 ev_fd_disable(r->ev, ev_read, r->fd);
1354 return ev_timeout(r->ev, 0, 0, reader_shutdown, r);
1357 /** @brief Temporarily disable a reader
1359 * @return 0 on success, non-0 on error
1361 * No further callbacks for this reader will be made. Re-enable with
1362 * ev_reader_enable().
1364 int ev_reader_disable(ev_reader *r) {
1365 D(("disable reader fd %d", r->fd));
1366 return ev_fd_disable(r->ev, ev_read, r->fd);
1369 /** @brief Called from ev_run() for ev_reader_incomplete() */
1370 static int reader_continuation(ev_source attribute((unused)) *ev,
1371 const attribute((unused)) struct timeval *now,
1375 D(("reader continuation callback fd %d", r->fd));
1376 /* If not at EOF turn the FD back on */
1378 if(ev_fd_enable(r->ev, ev_read, r->fd))
1380 /* We're already in a timeout callback so there's no reason we can't call the
1381 * user callback directly (compare ev_reader_enable()). */
1382 return r->callback(ev, r, r->b.start, r->b.end - r->b.start, r->eof, r->u);
1385 /** @brief Arrange another callback
1387 * @return 0 on success, non-0 on error
1389 * Indicates that the reader can process more input but would like to yield to
1390 * other clients of the event loop. Input will be disabled but it will be
1391 * re-enabled on the next iteration of the event loop and the read callback
1392 * will be called again (even if no further bytes are available).
1394 int ev_reader_incomplete(ev_reader *r) {
1395 if(ev_fd_disable(r->ev, ev_read, r->fd)) return -1;
1396 return ev_timeout(r->ev, 0, 0, reader_continuation, r);
1399 static int reader_enabled(ev_source *ev,
1400 const attribute((unused)) struct timeval *now,
1404 D(("reader enabled callback fd %d", r->fd));
1405 return r->callback(ev, r, r->b.start, r->b.end - r->b.start, r->eof, r->u);
1408 /** @brief Re-enable reading
1410 * @return 0 on success, non-0 on error
1412 * If there is unconsumed data then you get a callback next time round the
1413 * event loop even if nothing new has been read.
1415 * The idea is in your read callback you come across a line (or whatever) that
1416 * can't be processed immediately. So you set up processing and disable
1417 * reading with ev_reader_disable(). Later when you finish processing you
1418 * re-enable. You'll automatically get another callback directly from the
1419 * event loop (i.e. not from inside ev_reader_enable()) so you can handle the
1420 * next line (or whatever) if the whole thing has in fact already arrived.
1422 * The difference between this process and calling ev_reader_incomplete() is
1423 * ev_reader_incomplete() deals with the case where you can process now but
1424 * would rather yield to other clients of the event loop, while using
1425 * ev_reader_disable() and ev_reader_enable() deals with the case where you
1426 * cannot process input yet because some other process is actually not
1429 int ev_reader_enable(ev_reader *r) {
1430 D(("enable reader fd %d", r->fd));
1432 /* First if we're not at EOF then we re-enable reading */
1434 if(ev_fd_enable(r->ev, ev_read, r->fd))
1436 /* Arrange another callback next time round the event loop */
1437 return ev_timeout(r->ev, 0, 0, reader_enabled, r);
1440 /** @brief Tie a reader and a writer together
1443 * @return 0 on success, non-0 on error
1445 * This function must be called if @p r and @p w share a file descritptor.
1447 int ev_tie(ev_reader *r, ev_writer *w) {
1448 assert(r->writer == 0);
1449 assert(w->reader == 0);