2 * This file is part of DisOrder.
3 * Copyright (C) 2004, 2005, 2007, 2008 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 3 of the License, or
8 * (at your option) any later version.
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU 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, see <http://www.gnu.org/licenses/>.
19 * @brief DisOrder event loop
27 #include <sys/types.h>
28 #include <sys/resource.h>
34 #include <sys/socket.h>
35 #include <netinet/in.h>
47 /** @brief A timeout */
51 ev_timeout_callback *callback;
56 /** @brief Comparison function for timeouts */
57 static int timeout_lt(const struct timeout *a,
58 const struct timeout *b) {
59 return tvlt(&a->when, &b->when);
62 HEAP_TYPE(timeout_heap, struct timeout *, timeout_lt);
63 HEAP_DEFINE(timeout_heap, struct timeout *, timeout_lt);
65 /** @brief A file descriptor in one mode */
68 ev_fd_callback *callback;
73 /** @brief All the file descriptors in a given mode */
75 /** @brief Mask of active file descriptors passed to @c select() */
78 /** @brief File descriptor mask returned from @c select() */
81 /** @brief Number of file descriptors in @p fds */
84 /** @brief Number of slots in @p fds */
87 /** @brief Array of all active file descriptors */
90 /** @brief Highest-numbered file descriptor or 0 */
94 /** @brief A signal handler */
96 struct sigaction oldsa;
97 ev_signal_callback *callback;
101 /** @brief A child process */
105 ev_child_callback *callback;
109 /** @brief An event loop */
111 /** @brief File descriptors, per mode */
112 struct fdmode mode[ev_nmodes];
114 /** @brief Heap of timeouts */
115 struct timeout_heap timeouts[1];
117 /** @brief Array of handled signals */
118 struct signal signals[NSIG];
120 /** @brief Mask of handled signals */
123 /** @brief Escape early from handling of @c select() results
125 * This is set if any of the file descriptor arrays are invalidated, since
126 * it's then not safe for processing of them to continue.
130 /** @brief Signal handling pipe
132 * The signal handle writes signal numbers down this pipe.
136 /** @brief Number of child processes in @p children */
139 /** @brief Number of slots in @p children */
142 /** @brief Array of child processes */
143 struct child *children;
146 /** @brief Names of file descriptor modes */
147 static const char *modenames[] = { "read", "write", "except" };
149 /* utilities ******************************************************************/
151 /* creation *******************************************************************/
153 /** @brief Create a new event loop */
154 ev_source *ev_new(void) {
155 ev_source *ev = xmalloc(sizeof *ev);
158 memset(ev, 0, sizeof *ev);
159 for(n = 0; n < ev_nmodes; ++n)
160 FD_ZERO(&ev->mode[n].enabled);
161 ev->sigpipe[0] = ev->sigpipe[1] = -1;
162 sigemptyset(&ev->sigmask);
163 timeout_heap_init(ev->timeouts);
167 /* event loop *****************************************************************/
169 /** @brief Run the event loop
170 * @return -1 on error, non-0 if any callback returned non-0
172 int ev_run(ev_source *ev) {
175 struct timeval delta;
179 struct timeout *timeouts, *t, **tt;
182 xgettimeofday(&now, 0);
183 /* Handle timeouts. We don't want to handle any timeouts that are added
184 * while we're handling them (otherwise we'd have to break out of infinite
185 * loops, preferrably without starving better-behaved subsystems). Hence
186 * the slightly complicated two-phase approach here. */
187 /* First we read those timeouts that have triggered out of the heap. We
188 * keep them in the same order they came out of the heap in. */
190 while(timeout_heap_count(ev->timeouts)
191 && tvle(&timeout_heap_first(ev->timeouts)->when, &now)) {
192 /* This timeout has reached its trigger time; provided it has not been
193 * cancelled we add it to the timeouts list. */
194 t = timeout_heap_remove(ev->timeouts);
201 /* Now we can run the callbacks for those timeouts. They might add further
202 * timeouts that are already in the past but they won't trigger until the
203 * next time round the event loop. */
204 for(t = timeouts; t; t = t->next) {
205 D(("calling timeout for %ld.%ld callback %p %p",
206 (long)t->when.tv_sec, (long)t->when.tv_usec,
207 (void *)t->callback, t->u));
208 ret = t->callback(ev, &now, t->u);
213 for(mode = 0; mode < ev_nmodes; ++mode) {
214 ev->mode[mode].tripped = ev->mode[mode].enabled;
215 if(ev->mode[mode].maxfd > maxfd)
216 maxfd = ev->mode[mode].maxfd;
218 xsigprocmask(SIG_UNBLOCK, &ev->sigmask, 0);
220 if(timeout_heap_count(ev->timeouts)) {
221 t = timeout_heap_first(ev->timeouts);
222 xgettimeofday(&now, 0);
223 delta.tv_sec = t->when.tv_sec - now.tv_sec;
224 delta.tv_usec = t->when.tv_usec - now.tv_usec;
225 if(delta.tv_usec < 0) {
226 delta.tv_usec += 1000000;
230 delta.tv_sec = delta.tv_usec = 0;
231 n = select(maxfd + 1,
232 &ev->mode[ev_read].tripped,
233 &ev->mode[ev_write].tripped,
234 &ev->mode[ev_except].tripped,
237 n = select(maxfd + 1,
238 &ev->mode[ev_read].tripped,
239 &ev->mode[ev_write].tripped,
240 &ev->mode[ev_except].tripped,
243 } while(n < 0 && errno == EINTR);
244 xsigprocmask(SIG_BLOCK, &ev->sigmask, 0);
246 disorder_error(errno, "error calling select");
248 /* If there's a bad FD in the mix then check them all and log what we
249 * find, to ease debugging */
250 for(mode = 0; mode < ev_nmodes; ++mode) {
251 for(n = 0; n < ev->mode[mode].nfds; ++n) {
252 const int fd = ev->mode[mode].fds[n].fd;
254 if(FD_ISSET(fd, &ev->mode[mode].enabled)
255 && fstat(fd, &sb) < 0)
256 disorder_error(errno, "mode %s fstat %d (%s)",
257 modenames[mode], fd, ev->mode[mode].fds[n].what);
259 for(n = 0; n <= maxfd; ++n)
260 if(FD_ISSET(n, &ev->mode[mode].enabled)
261 && fstat(n, &sb) < 0)
262 disorder_error(errno, "mode %s fstat %d", modenames[mode], n);
268 /* if anything deranges the meaning of an fd, or re-orders the
269 * fds[] tables, we'd better give up; such operations will
270 * therefore set @escape@. */
272 for(mode = 0; mode < ev_nmodes && !ev->escape; ++mode)
273 for(n = 0; n < ev->mode[mode].nfds && !ev->escape; ++n) {
274 int fd = ev->mode[mode].fds[n].fd;
275 if(FD_ISSET(fd, &ev->mode[mode].tripped)) {
276 D(("calling %s fd %d callback %p %p", modenames[mode], fd,
277 (void *)ev->mode[mode].fds[n].callback,
278 ev->mode[mode].fds[n].u));
279 ret = ev->mode[mode].fds[n].callback(ev, fd,
280 ev->mode[mode].fds[n].u);
286 /* we'll pick up timeouts back round the loop */
290 /* file descriptors ***********************************************************/
292 /** @brief Register a file descriptor
293 * @param ev Event loop
294 * @param mode @c ev_read or @c ev_write
295 * @param fd File descriptor
296 * @param callback Called when @p is readable/writable
297 * @param u Passed to @p callback
298 * @param what Text description
299 * @return 0 on success, non-0 on error
301 * Sets @ref ev_source::escape, so no further processing of file descriptors
302 * will occur this time round the event loop.
304 int ev_fd(ev_source *ev,
307 ev_fd_callback *callback,
312 D(("registering %s fd %d callback %p %p", modenames[mode], fd,
313 (void *)callback, u));
314 /* FreeBSD defines FD_SETSIZE as 1024u for some reason */
315 if((unsigned)fd >= FD_SETSIZE)
317 assert(mode < ev_nmodes);
318 if(ev->mode[mode].nfds >= ev->mode[mode].fdslots) {
319 ev->mode[mode].fdslots = (ev->mode[mode].fdslots
320 ? 2 * ev->mode[mode].fdslots : 16);
321 D(("expanding %s fd table to %d entries", modenames[mode],
322 ev->mode[mode].fdslots));
323 ev->mode[mode].fds = xrealloc(ev->mode[mode].fds,
324 ev->mode[mode].fdslots * sizeof (struct fd));
326 n = ev->mode[mode].nfds++;
327 FD_SET(fd, &ev->mode[mode].enabled);
328 ev->mode[mode].fds[n].fd = fd;
329 ev->mode[mode].fds[n].callback = callback;
330 ev->mode[mode].fds[n].u = u;
331 ev->mode[mode].fds[n].what = what;
332 if(fd > ev->mode[mode].maxfd)
333 ev->mode[mode].maxfd = fd;
338 /** @brief Cancel a file descriptor
339 * @param ev Event loop
340 * @param mode @c ev_read or @c ev_write
341 * @param fd File descriptor
342 * @return 0 on success, non-0 on error
344 * Sets @ref ev_source::escape, so no further processing of file descriptors
345 * will occur this time round the event loop.
347 int ev_fd_cancel(ev_source *ev, ev_fdmode mode, int fd) {
351 D(("cancelling mode %s fd %d", modenames[mode], fd));
352 /* find the right struct fd */
353 for(n = 0; n < ev->mode[mode].nfds && fd != ev->mode[mode].fds[n].fd; ++n)
355 assert(n < ev->mode[mode].nfds);
356 /* swap in the last fd and reduce the count */
357 if(n != ev->mode[mode].nfds - 1)
358 ev->mode[mode].fds[n] = ev->mode[mode].fds[ev->mode[mode].nfds - 1];
359 --ev->mode[mode].nfds;
360 /* if that was the biggest fd, find the new biggest one */
361 if(fd == ev->mode[mode].maxfd) {
363 for(n = 0; n < ev->mode[mode].nfds; ++n)
364 if(ev->mode[mode].fds[n].fd > maxfd)
365 maxfd = ev->mode[mode].fds[n].fd;
366 ev->mode[mode].maxfd = maxfd;
368 /* don't tell select about this fd any more */
369 FD_CLR(fd, &ev->mode[mode].enabled);
374 /** @brief Re-enable a file descriptor
375 * @param ev Event loop
376 * @param mode @c ev_read or @c ev_write
377 * @param fd File descriptor
378 * @return 0 on success, non-0 on error
380 * It is harmless if @p fd is currently disabled, but it must not have been
383 int ev_fd_enable(ev_source *ev, ev_fdmode mode, int fd) {
385 D(("enabling mode %s fd %d", modenames[mode], fd));
386 FD_SET(fd, &ev->mode[mode].enabled);
390 /** @brief Temporarily disable a file descriptor
391 * @param ev Event loop
392 * @param mode @c ev_read or @c ev_write
393 * @param fd File descriptor
394 * @return 0 on success, non-0 on error
396 * Re-enable with ev_fd_enable(). It is harmless if @p fd is already disabled,
397 * but it must not have been cancelled.
399 int ev_fd_disable(ev_source *ev, ev_fdmode mode, int fd) {
400 D(("disabling mode %s fd %d", modenames[mode], fd));
401 FD_CLR(fd, &ev->mode[mode].enabled);
402 FD_CLR(fd, &ev->mode[mode].tripped);
403 /* Suppress any pending callbacks */
408 /** @brief Log a report of file descriptor state */
409 void ev_report(ev_source *ev) {
418 for(mode = 0; mode < ev_nmodes; ++mode) {
419 D(("mode %s maxfd %d", modenames[mode], ev->mode[mode].maxfd));
420 for(n = 0; n < ev->mode[mode].nfds; ++n) {
421 fd = ev->mode[mode].fds[n].fd;
422 D(("fd %s %d%s%s (%s)", modenames[mode], fd,
423 FD_ISSET(fd, &ev->mode[mode].enabled) ? " enabled" : "",
424 FD_ISSET(fd, &ev->mode[mode].tripped) ? " tripped" : "",
425 ev->mode[mode].fds[n].what));
428 for(fd = 0; fd <= ev->mode[mode].maxfd; ++fd) {
429 if(!FD_ISSET(fd, &ev->mode[mode].enabled))
431 for(n = 0; n < ev->mode[mode].nfds; ++n) {
432 if(ev->mode[mode].fds[n].fd == fd)
435 if(n < ev->mode[mode].nfds)
436 snprintf(b, sizeof b, "%d(%s)", fd, ev->mode[mode].fds[n].what);
438 snprintf(b, sizeof b, "%d", fd);
439 dynstr_append(d, ' ');
440 dynstr_append_string(d, b);
443 D(("%s enabled:%s", modenames[mode], d->vec));
447 /* timeouts *******************************************************************/
449 /** @brief Register a timeout
450 * @param ev Event source
451 * @param handlep Where to store timeout handle, or @c NULL
452 * @param when Earliest time to call @p callback, or @c NULL
453 * @param callback Function to call at or after @p when
454 * @param u Passed to @p callback
455 * @return 0 on success, non-0 on error
457 * If @p when is a null pointer then a time of 0 is assumed. The effect is to
458 * call the timeout handler from ev_run() next time around the event loop.
459 * This is used internally to schedule various operations if it is not
460 * convenient to call them from the current place in the call stack, or
461 * externally to ensure that other clients of the event loop get a look in when
462 * performing some lengthy operation.
464 int ev_timeout(ev_source *ev,
465 ev_timeout_handle *handlep,
466 const struct timeval *when,
467 ev_timeout_callback *callback,
471 D(("registering timeout at %ld.%ld callback %p %p",
472 when ? (long)when->tv_sec : 0, when ? (long)when->tv_usec : 0,
473 (void *)callback, u));
474 t = xmalloc(sizeof *t);
477 t->callback = callback;
480 timeout_heap_insert(ev->timeouts, t);
486 /** @brief Cancel a timeout
487 * @param ev Event loop
488 * @param handle Handle returned from ev_timeout(), or 0
489 * @return 0 on success, non-0 on error
491 * If @p handle is 0 then this is a no-op.
493 int ev_timeout_cancel(ev_source attribute((unused)) *ev,
494 ev_timeout_handle handle) {
495 struct timeout *t = handle;
502 /* signals ********************************************************************/
504 /** @brief Mapping of signals to pipe write ends
506 * The pipes are per-event loop, it's possible in theory for there to be
507 * multiple event loops (e.g. in different threads), although in fact DisOrder
510 static int sigfd[NSIG];
512 /** @brief The signal handler
513 * @param s Signal number
515 * Writes to @c sigfd[s].
517 static void sighandler(int s) {
518 unsigned char sc = s;
519 static const char errmsg[] = "error writing to signal pipe";
521 /* probably the reader has stopped listening for some reason */
522 if(write(sigfd[s], &sc, 1) < 0) {
523 /* do the best we can as we're about to abort; shut _up_, gcc */
524 int _ignore = write(2, errmsg, sizeof errmsg - 1);
530 /** @brief Read callback for signals */
531 static int signal_read(ev_source *ev,
532 int attribute((unused)) fd,
533 void attribute((unused)) *u) {
538 if((n = read(ev->sigpipe[0], &s, 1)) == 1)
539 if((ret = ev->signals[s].callback(ev, s, ev->signals[s].u)))
542 if(n < 0 && (errno != EINTR && errno != EAGAIN)) {
543 disorder_error(errno, "error reading from signal pipe %d", ev->sigpipe[0]);
549 /** @brief Close the signal pipe */
550 static void close_sigpipe(ev_source *ev) {
551 int save_errno = errno;
553 xclose(ev->sigpipe[0]);
554 xclose(ev->sigpipe[1]);
555 ev->sigpipe[0] = ev->sigpipe[1] = -1;
559 /** @brief Register a signal handler
560 * @param ev Event loop
561 * @param sig Signal to handle
562 * @param callback Called when signal is delivered
563 * @param u Passed to @p callback
564 * @return 0 on success, non-0 on error
566 * Note that @p callback is called from inside ev_run(), not from inside the
567 * signal handler, so the usual restrictions on signal handlers do not apply.
569 int ev_signal(ev_source *ev,
571 ev_signal_callback *callback,
576 D(("registering signal %d handler callback %p %p", sig, (void *)callback, u));
579 assert(sig <= UCHAR_MAX);
580 if(ev->sigpipe[0] == -1) {
581 D(("creating signal pipe"));
583 D(("signal pipe is %d, %d", ev->sigpipe[0], ev->sigpipe[1]));
584 for(n = 0; n < 2; ++n) {
585 nonblock(ev->sigpipe[n]);
586 cloexec(ev->sigpipe[n]);
588 if(ev_fd(ev, ev_read, ev->sigpipe[0], signal_read, 0, "sigpipe read")) {
593 sigaddset(&ev->sigmask, sig);
594 xsigprocmask(SIG_BLOCK, &ev->sigmask, 0);
595 sigfd[sig] = ev->sigpipe[1];
596 ev->signals[sig].callback = callback;
597 ev->signals[sig].u = u;
598 sa.sa_handler = sighandler;
599 sigfillset(&sa.sa_mask);
600 sa.sa_flags = SA_RESTART;
601 xsigaction(sig, &sa, &ev->signals[sig].oldsa);
606 /** @brief Cancel a signal handler
607 * @param ev Event loop
608 * @param sig Signal to cancel
609 * @return 0 on success, non-0 on error
611 int ev_signal_cancel(ev_source *ev,
615 xsigaction(sig, &ev->signals[sig].oldsa, 0);
616 ev->signals[sig].callback = 0;
618 sigdelset(&ev->sigmask, sig);
621 xsigprocmask(SIG_UNBLOCK, &ss, 0);
625 /** @brief Clean up signal handling
626 * @param ev Event loop
628 * This function can be called from inside a fork. It restores signal
629 * handlers, unblocks the signals, and closes the signal pipe for @p ev.
631 void ev_signal_atfork(ev_source *ev) {
634 if(ev->sigpipe[0] != -1) {
635 /* revert any handled signals to their original state */
636 for(sig = 1; sig < NSIG; ++sig) {
637 if(ev->signals[sig].callback != 0)
638 xsigaction(sig, &ev->signals[sig].oldsa, 0);
640 /* and then unblock them */
641 xsigprocmask(SIG_UNBLOCK, &ev->sigmask, 0);
642 /* don't want a copy of the signal pipe open inside the fork */
643 xclose(ev->sigpipe[0]);
644 xclose(ev->sigpipe[1]);
648 /* child processes ************************************************************/
650 /** @brief Called on SIGCHLD */
651 static int sigchld_callback(ev_source *ev,
652 int attribute((unused)) sig,
653 void attribute((unused)) *u) {
656 int status, n, ret, revisit;
660 for(n = 0; n < ev->nchildren; ++n) {
661 r = wait4(ev->children[n].pid,
663 ev->children[n].options | WNOHANG,
666 ev_child_callback *c = ev->children[n].callback;
667 void *cu = ev->children[n].u;
669 if(WIFEXITED(status) || WIFSIGNALED(status))
670 ev_child_cancel(ev, r);
672 if((ret = c(ev, r, status, &ru, cu)))
675 /* We should "never" get an ECHILD but it can in fact happen. For
676 * instance on Linux 2.4.31, and probably other versions, if someone
677 * straces a child process and then a different child process
678 * terminates, when we wait4() the trace process we will get ECHILD
679 * because it has been reparented to strace. Obviously this is a
680 * hopeless design flaw in the tracing infrastructure, but we don't
681 * want the disorder server to bomb out because of it. So we just log
682 * the problem and ignore it.
684 disorder_error(errno, "error calling wait4 for PID %lu (broken ptrace?)",
685 (unsigned long)ev->children[n].pid);
694 /** @brief Configure event loop for child process handling
695 * @return 0 on success, non-0 on error
697 * Currently at most one event loop can handle child processes and it must be
698 * distinguished from others by calling this function on it. This could be
699 * fixed but since no process ever makes use of more than one event loop there
702 int ev_child_setup(ev_source *ev) {
703 D(("installing SIGCHLD handler"));
704 return ev_signal(ev, SIGCHLD, sigchld_callback, 0);
707 /** @brief Wait for a child process to terminate
708 * @param ev Event loop
709 * @param pid Process ID of child
710 * @param options Options to pass to @c wait4()
711 * @param callback Called when child terminates (or possibly when it stops)
712 * @param u Passed to @p callback
713 * @return 0 on success, non-0 on error
715 * You must have called ev_child_setup() on @p ev once first.
717 int ev_child(ev_source *ev,
720 ev_child_callback *callback,
724 D(("registering child handling %ld options %d callback %p %p",
725 (long)pid, options, (void *)callback, u));
726 assert(ev->signals[SIGCHLD].callback == sigchld_callback);
727 if(ev->nchildren >= ev->nchildslots) {
728 ev->nchildslots = ev->nchildslots ? 2 * ev->nchildslots : 16;
729 ev->children = xrealloc(ev->children,
730 ev->nchildslots * sizeof (struct child));
733 ev->children[n].pid = pid;
734 ev->children[n].options = options;
735 ev->children[n].callback = callback;
736 ev->children[n].u = u;
740 /** @brief Stop waiting for a child process
741 * @param ev Event loop
742 * @param pid Child process ID
743 * @return 0 on success, non-0 on error
745 int ev_child_cancel(ev_source *ev,
749 for(n = 0; n < ev->nchildren && ev->children[n].pid != pid; ++n)
751 assert(n < ev->nchildren);
752 if(n != ev->nchildren - 1)
753 ev->children[n] = ev->children[ev->nchildren - 1];
758 /** @brief Terminate and wait for all child processes
759 * @param ev Event loop
761 * Does *not* call the completion callbacks. Only used during teardown.
763 void ev_child_killall(ev_source *ev) {
766 for(n = 0; n < ev->nchildren; ++n) {
767 if(kill(ev->children[n].pid, SIGTERM) < 0) {
768 disorder_error(errno, "sending SIGTERM to pid %lu",
769 (unsigned long)ev->children[n].pid);
770 ev->children[n].pid = -1;
773 for(n = 0; n < ev->nchildren; ++n) {
774 if(ev->children[n].pid == -1)
777 rc = waitpid(ev->children[n].pid, &w, 0);
778 } while(rc < 0 && errno == EINTR);
780 disorder_error(errno, "waiting for pid %lu",
781 (unsigned long)ev->children[n].pid);
788 /* socket listeners ***********************************************************/
790 /** @brief State for a socket listener */
791 struct listen_state {
792 ev_listen_callback *callback;
796 /** @brief Called when a listenign socket is readable */
797 static int listen_callback(ev_source *ev, int fd, void *u) {
798 const struct listen_state *l = u;
801 struct sockaddr_in in;
802 #if HAVE_STRUCT_SOCKADDR_IN6
803 struct sockaddr_in6 in6;
805 struct sockaddr_un un;
811 D(("callback for listener fd %d", fd));
812 while((addrlen = sizeof addr),
813 (newfd = accept(fd, &addr.sa, &addrlen)) >= 0) {
814 if((ret = l->callback(ev, newfd, &addr.sa, addrlen, l->u)))
823 disorder_error(errno, "error calling accept");
828 /* XXX on some systems EPROTO should be fatal, but we don't know if
829 * we're running on one of them */
830 disorder_error(errno, "error calling accept");
834 disorder_fatal(errno, "error calling accept");
837 if(errno != EINTR && errno != EAGAIN)
838 disorder_error(errno, "error calling accept");
842 /** @brief Listen on a socket for inbound stream connections
843 * @param ev Event source
844 * @param fd File descriptor of socket
845 * @param callback Called when a new connection arrives
846 * @param u Passed to @p callback
847 * @param what Text description of socket
848 * @return 0 on success, non-0 on error
850 int ev_listen(ev_source *ev,
852 ev_listen_callback *callback,
855 struct listen_state *l = xmalloc(sizeof *l);
857 D(("registering listener fd %d callback %p %p", fd, (void *)callback, u));
858 l->callback = callback;
860 return ev_fd(ev, ev_read, fd, listen_callback, l, what);
863 /** @brief Stop listening on a socket
864 * @param ev Event loop
865 * @param fd File descriptor of socket
866 * @return 0 on success, non-0 on error
868 int ev_listen_cancel(ev_source *ev, int fd) {
869 D(("cancelling listener fd %d", fd));
870 return ev_fd_cancel(ev, ev_read, fd);
873 /* buffer *********************************************************************/
875 /** @brief Buffer structure */
877 char *base, *start, *end, *top;
880 /* @brief Make sure there is @p bytes available at @c b->end */
881 static void buffer_space(struct buffer *b, size_t bytes) {
882 D(("buffer_space %p %p %p %p want %lu",
883 (void *)b->base, (void *)b->start, (void *)b->end, (void *)b->top,
884 (unsigned long)bytes));
885 if(b->start == b->end)
886 b->start = b->end = b->base;
887 if((size_t)(b->top - b->end) < bytes) {
888 if((size_t)((b->top - b->end) + (b->start - b->base)) < bytes) {
889 size_t newspace = b->end - b->start + bytes, n;
892 for(n = 16; n < newspace; n *= 2)
894 newbase = xmalloc_noptr(n);
895 memcpy(newbase, b->start, b->end - b->start);
897 b->end = newbase + (b->end - b->start);
898 b->top = newbase + n;
899 b->start = newbase; /* must be last */
901 memmove(b->base, b->start, b->end - b->start);
902 b->end = b->base + (b->end - b->start);
906 D(("result %p %p %p %p",
907 (void *)b->base, (void *)b->start, (void *)b->end, (void *)b->top));
910 /* readers and writers *******************************************************/
912 /** @brief State structure for a buffered writer */
914 /** @brief Sink used for writing to the buffer */
917 /** @brief Output buffer */
920 /** @brief File descriptor to write to */
923 /** @brief Set if there'll be no more output */
926 /** @brief Error/termination callback */
927 ev_error_callback *callback;
929 /** @brief Passed to @p callback */
932 /** @brief Parent event source */
935 /** @brief Maximum amount of time between succesful writes, 0 = don't care */
937 /** @brief Maximum amount of data to buffer, 0 = don't care */
939 /** @brief Error code to pass to @p callback (see writer_shutdown()) */
941 /** @brief Timeout handle for @p timebound (or 0) */
942 ev_timeout_handle timeout;
944 /** @brief Description of this writer */
947 /** @brief Tied reader or 0 */
950 /** @brief Set when abandoned */
954 /** @brief State structure for a buffered reader */
956 /** @brief Input buffer */
958 /** @brief File descriptor read from */
960 /** @brief Called when new data is available */
961 ev_reader_callback *callback;
962 /** @brief Called on error and shutdown */
963 ev_error_callback *error_callback;
964 /** @brief Passed to @p callback and @p error_callback */
966 /** @brief Parent event loop */
968 /** @brief Set when EOF is detected */
970 /** @brief Error code to pass to error callback */
972 /** @brief Tied writer or NULL */
976 /* buffered writer ************************************************************/
978 /** @brief Shut down the writer
980 * This is called to shut down a writer. The error callback is not called
981 * through any other path. Also we do not cancel @p fd from anywhere else,
982 * though we might disable it.
984 * It has the signature of a timeout callback so that it can be called from a
987 * Calls @p callback with @p w->syntherr as the error code (which might be 0).
989 static int writer_shutdown(ev_source *ev,
990 const attribute((unused)) struct timeval *now,
995 return 0; /* already shut down */
996 D(("writer_shutdown fd=%d error=%d", w->fd, w->error));
997 ev_timeout_cancel(ev, w->timeout);
998 ev_fd_cancel(ev, ev_write, w->fd);
1001 D(("found a tied reader"));
1002 /* If there is a reader still around we just untie it */
1003 w->reader->writer = 0;
1004 shutdown(w->fd, SHUT_WR); /* there'll be no more writes */
1006 D(("no tied reader"));
1007 /* There's no reader so we are free to close the FD */
1011 return w->callback(ev, w->error, w->u);
1014 /** @brief Called when a writer's @p timebound expires */
1015 static int writer_timebound_exceeded(ev_source *ev,
1016 const struct timeval *now,
1018 ev_writer *const w = u;
1022 disorder_error(0, "abandoning writer '%s' because no writes within %ds",
1023 w->what, w->timebound);
1024 w->error = ETIMEDOUT;
1026 return writer_shutdown(ev, now, u);
1029 /** @brief Set the time bound callback (if not set already) */
1030 static void writer_set_timebound(ev_writer *w) {
1031 if(w->timebound && !w->timeout) {
1032 struct timeval when;
1033 ev_source *const ev = w->ev;
1035 xgettimeofday(&when, 0);
1036 when.tv_sec += w->timebound;
1037 ev_timeout(ev, &w->timeout, &when, writer_timebound_exceeded, w);
1041 /** @brief Called when a writer's file descriptor is writable */
1042 static int writer_callback(ev_source *ev, int fd, void *u) {
1043 ev_writer *const w = u;
1046 n = write(fd, w->b.start, w->b.end - w->b.start);
1047 D(("callback for writer fd %d, %ld bytes, n=%d, errno=%d",
1048 fd, (long)(w->b.end - w->b.start), n, errno));
1050 /* Consume bytes from the buffer */
1052 /* Suppress any outstanding timeout */
1053 ev_timeout_cancel(ev, w->timeout);
1055 if(w->b.start == w->b.end) {
1056 /* The buffer is empty */
1058 /* We're done, we can shut down this writer */
1060 return writer_shutdown(ev, 0, w);
1062 /* There might be more to come but we don't need writer_callback() to
1063 * be called for the time being */
1064 ev_fd_disable(ev, ev_write, fd);
1066 /* The buffer isn't empty, set a timeout so we give up if we don't manage
1067 * to write some more within a reasonable time */
1068 writer_set_timebound(w);
1076 return writer_shutdown(ev, 0, w);
1082 /** @brief Write bytes to a writer's buffer
1084 * This is the sink write callback.
1086 * Calls ev_fd_enable() if necessary (i.e. if the buffer was empty but
1089 static int ev_writer_write(struct sink *sk, const void *s, int n) {
1090 ev_writer *w = (ev_writer *)sk;
1093 return 0; /* avoid silliness */
1095 disorder_error(0, "ev_writer_write on %s after shutdown", w->what);
1096 if(w->spacebound && w->b.end - w->b.start + n > w->spacebound) {
1097 /* The new buffer contents will exceed the space bound. We assume that the
1098 * remote client has gone away and TCP hasn't noticed yet, or that it's got
1099 * hopelessly stuck. */
1102 disorder_error(0, "abandoning writer '%s' because buffer has reached %td bytes",
1103 w->what, w->b.end - w->b.start);
1104 ev_fd_disable(w->ev, ev_write, w->fd);
1106 return ev_timeout(w->ev, 0, 0, writer_shutdown, w);
1110 /* Make sure there is space */
1111 buffer_space(&w->b, n);
1112 /* If the buffer was formerly empty then we'll need to re-enable the FD */
1113 if(w->b.start == w->b.end)
1114 ev_fd_enable(w->ev, ev_write, w->fd);
1115 memcpy(w->b.end, s, n);
1117 /* Arrange a timeout if there wasn't one set already */
1118 writer_set_timebound(w);
1122 /** @brief Create a new buffered writer
1123 * @param ev Event loop
1124 * @param fd File descriptor to write to
1125 * @param callback Called if an error occurs and when finished
1126 * @param u Passed to @p callback
1127 * @param what Text description
1128 * @return New writer or @c NULL
1130 * Writers own their file descriptor and close it when they have finished with
1133 * If you pass the same fd to a reader and writer, you must tie them together
1136 ev_writer *ev_writer_new(ev_source *ev,
1138 ev_error_callback *callback,
1141 ev_writer *w = xmalloc(sizeof *w);
1143 D(("registering writer fd %d callback %p %p", fd, (void *)callback, u));
1144 w->s.write = ev_writer_write;
1146 w->callback = callback;
1149 w->timebound = 10 * 60;
1150 w->spacebound = 512 * 1024;
1152 if(ev_fd(ev, ev_write, fd, writer_callback, w, what))
1154 /* Buffer is initially empty so we don't want a callback */
1155 ev_fd_disable(ev, ev_write, fd);
1159 /** @brief Get/set the time bound
1161 * @param new_time_bound New bound or -1 for no change
1162 * @return Latest time bound
1164 * If @p new_time_bound is negative then the current time bound is returned.
1165 * Otherwise it is set and the new value returned.
1167 * The time bound is the number of seconds allowed between writes. If it takes
1168 * longer than this to flush a buffer then the peer will be assumed to be dead
1169 * and an error will be synthesized. 0 means "don't care". The default time
1170 * bound is 10 minutes.
1172 * Note that this value does not take into account kernel buffering and
1175 int ev_writer_time_bound(ev_writer *w,
1176 int new_time_bound) {
1177 if(new_time_bound >= 0)
1178 w->timebound = new_time_bound;
1179 return w->timebound;
1182 /** @brief Get/set the space bound
1184 * @param new_space_bound New bound or -1 for no change
1185 * @return Latest space bound
1187 * If @p new_space_bound is negative then the current space bound is returned.
1188 * Otherwise it is set and the new value returned.
1190 * The space bound is the number of bytes allowed between in the buffer. If
1191 * the buffer exceeds this size an error will be synthesized. 0 means "don't
1192 * care". The default space bound is 512Kbyte.
1194 * Note that this value does not take into account kernel buffering.
1196 int ev_writer_space_bound(ev_writer *w,
1197 int new_space_bound) {
1198 if(new_space_bound >= 0)
1199 w->spacebound = new_space_bound;
1200 return w->spacebound;
1203 /** @brief Return the sink associated with a writer
1205 * @return Pointer to sink
1207 * Writing to the sink will arrange for those bytes to be written to the file
1208 * descriptor as and when it is writable.
1210 struct sink *ev_writer_sink(ev_writer *w) {
1212 disorder_fatal(0, "ev_write_sink called with null writer");
1216 /** @brief Close a writer
1217 * @param w Writer to close
1218 * @return 0 on success, non-0 on error
1220 * Close a writer. No more bytes should be written to its sink.
1222 * When the last byte has been written the callback will be called with an
1223 * error code of 0. It is guaranteed that this will NOT happen before
1224 * ev_writer_close() returns (although the file descriptor for the writer might
1225 * be cancelled by the time it returns).
1227 int ev_writer_close(ev_writer *w) {
1228 D(("close writer fd %d", w->fd));
1230 return 0; /* already closed */
1232 if(w->b.start == w->b.end) {
1233 /* We're already finished */
1234 w->error = 0; /* no error */
1235 return ev_timeout(w->ev, 0, 0, writer_shutdown, w);
1240 /** @brief Attempt to flush a writer
1241 * @param w Writer to flush
1242 * @return 0 on success, non-0 on error
1244 * Does a speculative write of any buffered data. Does not block if it cannot
1247 int ev_writer_flush(ev_writer *w) {
1248 return writer_callback(w->ev, w->fd, w);
1251 /* buffered reader ************************************************************/
1253 /** @brief Shut down a reader
1255 * This is the only path through which we cancel and close the file descriptor.
1256 * As with the writer case it is given timeout signature to allow it be
1257 * deferred to the next iteration of the event loop.
1259 * We only call @p error_callback if @p error is nonzero (unlike the writer
1262 static int reader_shutdown(ev_source *ev,
1263 const attribute((unused)) struct timeval *now,
1265 ev_reader *const r = u;
1268 return 0; /* already shut down */
1269 D(("reader_shutdown fd=%d", r->fd));
1270 ev_fd_cancel(ev, ev_read, r->fd);
1273 D(("found a tied writer"));
1274 /* If there is a writer still around we just untie it */
1275 r->writer->reader = 0;
1276 shutdown(r->fd, SHUT_RD); /* there'll be no more reads */
1278 D(("no tied writer found"));
1279 /* There's no writer so we are free to close the FD */
1284 return r->error_callback(ev, r->error, r->u);
1289 /** @brief Called when a reader's @p fd is readable */
1290 static int reader_callback(ev_source *ev, int fd, void *u) {
1294 buffer_space(&r->b, 1);
1295 n = read(fd, r->b.end, r->b.top - r->b.end);
1296 D(("read fd %d buffer %d returned %d errno %d",
1297 fd, (int)(r->b.top - r->b.end), n, errno));
1300 return r->callback(ev, r, r->b.start, r->b.end - r->b.start, 0, r->u);
1302 /* No more read callbacks needed */
1303 ev_fd_disable(r->ev, ev_read, r->fd);
1304 ev_timeout(r->ev, 0, 0, reader_shutdown, r);
1305 /* Pass the remaining data and an eof indicator to the user */
1306 return r->callback(ev, r, r->b.start, r->b.end - r->b.start, 1, r->u);
1313 /* Fatal error, kill the reader now */
1315 return reader_shutdown(ev, 0, r);
1321 /** @brief Create a new buffered reader
1322 * @param ev Event loop
1323 * @param fd File descriptor to read from
1324 * @param callback Called when new data is available
1325 * @param error_callback Called if an error occurs
1326 * @param u Passed to callbacks
1327 * @param what Text description
1328 * @return New reader or @c NULL
1330 * Readers own their fd and close it when they are finished with it.
1332 * If you pass the same fd to a reader and writer, you must tie them together
1335 ev_reader *ev_reader_new(ev_source *ev,
1337 ev_reader_callback *callback,
1338 ev_error_callback *error_callback,
1341 ev_reader *r = xmalloc(sizeof *r);
1343 D(("registering reader fd %d callback %p %p %p",
1344 fd, (void *)callback, (void *)error_callback, u));
1346 r->callback = callback;
1347 r->error_callback = error_callback;
1350 if(ev_fd(ev, ev_read, fd, reader_callback, r, what))
1355 void ev_reader_buffer(ev_reader *r, size_t nbytes) {
1356 buffer_space(&r->b, nbytes - (r->b.end - r->b.start));
1359 /** @brief Consume @p n bytes from the reader's buffer
1361 * @param n Number of bytes to consume
1363 * Tells the reader than the next @p n bytes have been dealt with and can now
1366 void ev_reader_consume(ev_reader *r, size_t n) {
1370 /** @brief Cancel a reader
1372 * @return 0 on success, non-0 on error
1374 * No further callbacks will be made, and the FD will be closed (in a later
1375 * iteration of the event loop).
1377 int ev_reader_cancel(ev_reader *r) {
1378 D(("cancel reader fd %d", r->fd));
1380 return 0; /* already thoroughly cancelled */
1381 ev_fd_disable(r->ev, ev_read, r->fd);
1382 return ev_timeout(r->ev, 0, 0, reader_shutdown, r);
1385 /** @brief Temporarily disable a reader
1387 * @return 0 on success, non-0 on error
1389 * No further callbacks for this reader will be made. Re-enable with
1390 * ev_reader_enable().
1392 int ev_reader_disable(ev_reader *r) {
1393 D(("disable reader fd %d", r->fd));
1394 return ev_fd_disable(r->ev, ev_read, r->fd);
1397 /** @brief Called from ev_run() for ev_reader_incomplete() */
1398 static int reader_continuation(ev_source attribute((unused)) *ev,
1399 const attribute((unused)) struct timeval *now,
1403 D(("reader continuation callback fd %d", r->fd));
1404 /* If not at EOF turn the FD back on */
1406 if(ev_fd_enable(r->ev, ev_read, r->fd))
1408 /* We're already in a timeout callback so there's no reason we can't call the
1409 * user callback directly (compare ev_reader_enable()). */
1410 return r->callback(ev, r, r->b.start, r->b.end - r->b.start, r->eof, r->u);
1413 /** @brief Arrange another callback
1415 * @return 0 on success, non-0 on error
1417 * Indicates that the reader can process more input but would like to yield to
1418 * other clients of the event loop. Input will be disabled but it will be
1419 * re-enabled on the next iteration of the event loop and the read callback
1420 * will be called again (even if no further bytes are available).
1422 int ev_reader_incomplete(ev_reader *r) {
1423 if(ev_fd_disable(r->ev, ev_read, r->fd)) return -1;
1424 return ev_timeout(r->ev, 0, 0, reader_continuation, r);
1427 static int reader_enabled(ev_source *ev,
1428 const attribute((unused)) struct timeval *now,
1432 D(("reader enabled callback fd %d", r->fd));
1433 return r->callback(ev, r, r->b.start, r->b.end - r->b.start, r->eof, r->u);
1436 /** @brief Re-enable reading
1438 * @return 0 on success, non-0 on error
1440 * If there is unconsumed data then you get a callback next time round the
1441 * event loop even if nothing new has been read.
1443 * The idea is in your read callback you come across a line (or whatever) that
1444 * can't be processed immediately. So you set up processing and disable
1445 * reading with ev_reader_disable(). Later when you finish processing you
1446 * re-enable. You'll automatically get another callback directly from the
1447 * event loop (i.e. not from inside ev_reader_enable()) so you can handle the
1448 * next line (or whatever) if the whole thing has in fact already arrived.
1450 * The difference between this process and calling ev_reader_incomplete() is
1451 * ev_reader_incomplete() deals with the case where you can process now but
1452 * would rather yield to other clients of the event loop, while using
1453 * ev_reader_disable() and ev_reader_enable() deals with the case where you
1454 * cannot process input yet because some other process is actually not
1457 int ev_reader_enable(ev_reader *r) {
1458 D(("enable reader fd %d", r->fd));
1460 /* First if we're not at EOF then we re-enable reading */
1462 if(ev_fd_enable(r->ev, ev_read, r->fd))
1464 /* Arrange another callback next time round the event loop */
1465 return ev_timeout(r->ev, 0, 0, reader_enabled, r);
1468 /** @brief Tie a reader and a writer together
1471 * @return 0 on success, non-0 on error
1473 * This function must be called if @p r and @p w share a file descritptor.
1475 int ev_tie(ev_reader *r, ev_writer *w) {
1476 assert(r->writer == 0);
1477 assert(w->reader == 0);