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 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 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 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 assert(mode
< ev_nmodes
);
315 if(ev
->mode
[mode
].nfds
>= ev
->mode
[mode
].fdslots
) {
316 ev
->mode
[mode
].fdslots
= (ev
->mode
[mode
].fdslots
317 ?
2 * ev
->mode
[mode
].fdslots
: 16);
318 D(("expanding %s fd table to %d entries", modenames
[mode
],
319 ev
->mode
[mode
].fdslots
));
320 ev
->mode
[mode
].fds
= xrealloc(ev
->mode
[mode
].fds
,
321 ev
->mode
[mode
].fdslots
* sizeof (struct fd
));
323 n
= ev
->mode
[mode
].nfds
++;
324 FD_SET(fd
, &ev
->mode
[mode
].enabled
);
325 ev
->mode
[mode
].fds
[n
].fd
= fd
;
326 ev
->mode
[mode
].fds
[n
].callback
= callback
;
327 ev
->mode
[mode
].fds
[n
].u
= u
;
328 ev
->mode
[mode
].fds
[n
].what
= what
;
329 if(fd
> ev
->mode
[mode
].maxfd
)
330 ev
->mode
[mode
].maxfd
= fd
;
335 /** @brief Cancel a file descriptor
336 * @param ev Event loop
337 * @param mode @c ev_read or @c ev_write
338 * @param fd File descriptor
339 * @return 0 on success, non-0 on error
341 * Sets @ref ev_source::escape, so no further processing of file descriptors
342 * will occur this time round the event loop.
344 int ev_fd_cancel(ev_source
*ev
, ev_fdmode mode
, int fd
) {
348 D(("cancelling mode %s fd %d", modenames
[mode
], fd
));
349 /* find the right struct fd */
350 for(n
= 0; n
< ev
->mode
[mode
].nfds
&& fd
!= ev
->mode
[mode
].fds
[n
].fd
; ++n
)
352 assert(n
< ev
->mode
[mode
].nfds
);
353 /* swap in the last fd and reduce the count */
354 if(n
!= ev
->mode
[mode
].nfds
- 1)
355 ev
->mode
[mode
].fds
[n
] = ev
->mode
[mode
].fds
[ev
->mode
[mode
].nfds
- 1];
356 --ev
->mode
[mode
].nfds
;
357 /* if that was the biggest fd, find the new biggest one */
358 if(fd
== ev
->mode
[mode
].maxfd
) {
360 for(n
= 0; n
< ev
->mode
[mode
].nfds
; ++n
)
361 if(ev
->mode
[mode
].fds
[n
].fd
> maxfd
)
362 maxfd
= ev
->mode
[mode
].fds
[n
].fd
;
363 ev
->mode
[mode
].maxfd
= maxfd
;
365 /* don't tell select about this fd any more */
366 FD_CLR(fd
, &ev
->mode
[mode
].enabled
);
371 /** @brief Re-enable a file descriptor
372 * @param ev Event loop
373 * @param mode @c ev_read or @c ev_write
374 * @param fd File descriptor
375 * @return 0 on success, non-0 on error
377 * It is harmless if @p fd is currently disabled, but it must not have been
380 int ev_fd_enable(ev_source
*ev
, ev_fdmode mode
, int fd
) {
382 D(("enabling mode %s fd %d", modenames
[mode
], fd
));
383 FD_SET(fd
, &ev
->mode
[mode
].enabled
);
387 /** @brief Temporarily disable a file descriptor
388 * @param ev Event loop
389 * @param mode @c ev_read or @c ev_write
390 * @param fd File descriptor
391 * @return 0 on success, non-0 on error
393 * Re-enable with ev_fd_enable(). It is harmless if @p fd is already disabled,
394 * but it must not have been cancelled.
396 int ev_fd_disable(ev_source
*ev
, ev_fdmode mode
, int fd
) {
397 D(("disabling mode %s fd %d", modenames
[mode
], fd
));
398 FD_CLR(fd
, &ev
->mode
[mode
].enabled
);
399 FD_CLR(fd
, &ev
->mode
[mode
].tripped
);
400 /* Suppress any pending callbacks */
405 /** @brief Log a report of file descriptor state */
406 void ev_report(ev_source
*ev
) {
415 for(mode
= 0; mode
< ev_nmodes
; ++mode
) {
416 D(("mode %s maxfd %d", modenames
[mode
], ev
->mode
[mode
].maxfd
));
417 for(n
= 0; n
< ev
->mode
[mode
].nfds
; ++n
) {
418 fd
= ev
->mode
[mode
].fds
[n
].fd
;
419 D(("fd %s %d%s%s (%s)", modenames
[mode
], fd
,
420 FD_ISSET(fd
, &ev
->mode
[mode
].enabled
) ?
" enabled" : "",
421 FD_ISSET(fd
, &ev
->mode
[mode
].tripped
) ?
" tripped" : "",
422 ev
->mode
[mode
].fds
[n
].what
));
425 for(fd
= 0; fd
<= ev
->mode
[mode
].maxfd
; ++fd
) {
426 if(!FD_ISSET(fd
, &ev
->mode
[mode
].enabled
))
428 for(n
= 0; n
< ev
->mode
[mode
].nfds
; ++n
) {
429 if(ev
->mode
[mode
].fds
[n
].fd
== fd
)
432 if(n
< ev
->mode
[mode
].nfds
)
433 snprintf(b
, sizeof b
, "%d(%s)", fd
, ev
->mode
[mode
].fds
[n
].what
);
435 snprintf(b
, sizeof b
, "%d", fd
);
436 dynstr_append(d
, ' ');
437 dynstr_append_string(d
, b
);
440 D(("%s enabled:%s", modenames
[mode
], d
->vec
));
444 /* timeouts *******************************************************************/
446 /** @brief Register a timeout
447 * @param ev Event source
448 * @param handlep Where to store timeout handle, or @c NULL
449 * @param when Earliest time to call @p callback, or @c NULL
450 * @param callback Function to call at or after @p when
451 * @param u Passed to @p callback
452 * @return 0 on success, non-0 on error
454 * If @p when is a null pointer then a time of 0 is assumed. The effect is to
455 * call the timeout handler from ev_run() next time around the event loop.
456 * This is used internally to schedule various operations if it is not
457 * convenient to call them from the current place in the call stack, or
458 * externally to ensure that other clients of the event loop get a look in when
459 * performing some lengthy operation.
461 int ev_timeout(ev_source
*ev
,
462 ev_timeout_handle
*handlep
,
463 const struct timeval
*when
,
464 ev_timeout_callback
*callback
,
468 D(("registering timeout at %ld.%ld callback %p %p",
469 when ?
(long)when
->tv_sec
: 0, when ?
(long)when
->tv_usec
: 0,
470 (void *)callback
, u
));
471 t
= xmalloc(sizeof *t
);
474 t
->callback
= callback
;
477 timeout_heap_insert(ev
->timeouts
, t
);
483 /** @brief Cancel a timeout
484 * @param ev Event loop
485 * @param handle Handle returned from ev_timeout(), or 0
486 * @return 0 on success, non-0 on error
488 * If @p handle is 0 then this is a no-op.
490 int ev_timeout_cancel(ev_source
attribute((unused
)) *ev
,
491 ev_timeout_handle handle
) {
492 struct timeout
*t
= handle
;
499 /* signals ********************************************************************/
501 /** @brief Mapping of signals to pipe write ends
503 * The pipes are per-event loop, it's possible in theory for there to be
504 * multiple event loops (e.g. in different threads), although in fact DisOrder
507 static int sigfd
[NSIG
];
509 /** @brief The signal handler
510 * @param s Signal number
512 * Writes to @c sigfd[s].
514 static void sighandler(int s
) {
515 unsigned char sc
= s
;
516 static const char errmsg
[] = "error writing to signal pipe";
518 /* probably the reader has stopped listening for some reason */
519 if(write(sigfd
[s
], &sc
, 1) < 0) {
520 write(2, errmsg
, sizeof errmsg
- 1);
525 /** @brief Read callback for signals */
526 static int signal_read(ev_source
*ev
,
527 int attribute((unused
)) fd
,
528 void attribute((unused
)) *u
) {
533 if((n
= read(ev
->sigpipe
[0], &s
, 1)) == 1)
534 if((ret
= ev
->signals
[s
].callback(ev
, s
, ev
->signals
[s
].u
)))
537 if(n
< 0 && (errno
!= EINTR
&& errno
!= EAGAIN
)) {
538 error(errno
, "error reading from signal pipe %d", ev
->sigpipe
[0]);
544 /** @brief Close the signal pipe */
545 static void close_sigpipe(ev_source
*ev
) {
546 int save_errno
= errno
;
548 xclose(ev
->sigpipe
[0]);
549 xclose(ev
->sigpipe
[1]);
550 ev
->sigpipe
[0] = ev
->sigpipe
[1] = -1;
554 /** @brief Register a signal handler
555 * @param ev Event loop
556 * @param sig Signal to handle
557 * @param callback Called when signal is delivered
558 * @param u Passed to @p callback
559 * @return 0 on success, non-0 on error
561 * Note that @p callback is called from inside ev_run(), not from inside the
562 * signal handler, so the usual restrictions on signal handlers do not apply.
564 int ev_signal(ev_source
*ev
,
566 ev_signal_callback
*callback
,
571 D(("registering signal %d handler callback %p %p", sig
, (void *)callback
, u
));
574 assert(sig
<= UCHAR_MAX
);
575 if(ev
->sigpipe
[0] == -1) {
576 D(("creating signal pipe"));
578 D(("signal pipe is %d, %d", ev
->sigpipe
[0], ev
->sigpipe
[1]));
579 for(n
= 0; n
< 2; ++n
) {
580 nonblock(ev
->sigpipe
[n
]);
581 cloexec(ev
->sigpipe
[n
]);
583 if(ev_fd(ev
, ev_read
, ev
->sigpipe
[0], signal_read
, 0, "sigpipe read")) {
588 sigaddset(&ev
->sigmask
, sig
);
589 xsigprocmask(SIG_BLOCK
, &ev
->sigmask
, 0);
590 sigfd
[sig
] = ev
->sigpipe
[1];
591 ev
->signals
[sig
].callback
= callback
;
592 ev
->signals
[sig
].u
= u
;
593 sa
.sa_handler
= sighandler
;
594 sigfillset(&sa
.sa_mask
);
595 sa
.sa_flags
= SA_RESTART
;
596 xsigaction(sig
, &sa
, &ev
->signals
[sig
].oldsa
);
601 /** @brief Cancel a signal handler
602 * @param ev Event loop
603 * @param sig Signal to cancel
604 * @return 0 on success, non-0 on error
606 int ev_signal_cancel(ev_source
*ev
,
610 xsigaction(sig
, &ev
->signals
[sig
].oldsa
, 0);
611 ev
->signals
[sig
].callback
= 0;
613 sigdelset(&ev
->sigmask
, sig
);
616 xsigprocmask(SIG_UNBLOCK
, &ss
, 0);
620 /** @brief Clean up signal handling
621 * @param ev Event loop
623 * This function can be called from inside a fork. It restores signal
624 * handlers, unblocks the signals, and closes the signal pipe for @p ev.
626 void ev_signal_atfork(ev_source
*ev
) {
629 if(ev
->sigpipe
[0] != -1) {
630 /* revert any handled signals to their original state */
631 for(sig
= 1; sig
< NSIG
; ++sig
) {
632 if(ev
->signals
[sig
].callback
!= 0)
633 xsigaction(sig
, &ev
->signals
[sig
].oldsa
, 0);
635 /* and then unblock them */
636 xsigprocmask(SIG_UNBLOCK
, &ev
->sigmask
, 0);
637 /* don't want a copy of the signal pipe open inside the fork */
638 xclose(ev
->sigpipe
[0]);
639 xclose(ev
->sigpipe
[1]);
643 /* child processes ************************************************************/
645 /** @brief Called on SIGCHLD */
646 static int sigchld_callback(ev_source
*ev
,
647 int attribute((unused
)) sig
,
648 void attribute((unused
)) *u
) {
651 int status
, n
, ret
, revisit
;
655 for(n
= 0; n
< ev
->nchildren
; ++n
) {
656 r
= wait4(ev
->children
[n
].pid
,
658 ev
->children
[n
].options
| WNOHANG
,
661 ev_child_callback
*c
= ev
->children
[n
].callback
;
662 void *cu
= ev
->children
[n
].u
;
664 if(WIFEXITED(status
) || WIFSIGNALED(status
))
665 ev_child_cancel(ev
, r
);
667 if((ret
= c(ev
, r
, status
, &ru
, cu
)))
670 /* We should "never" get an ECHILD but it can in fact happen. For
671 * instance on Linux 2.4.31, and probably other versions, if someone
672 * straces a child process and then a different child process
673 * terminates, when we wait4() the trace process we will get ECHILD
674 * because it has been reparented to strace. Obviously this is a
675 * hopeless design flaw in the tracing infrastructure, but we don't
676 * want the disorder server to bomb out because of it. So we just log
677 * the problem and ignore it.
679 error(errno
, "error calling wait4 for PID %lu (broken ptrace?)",
680 (unsigned long)ev
->children
[n
].pid
);
689 /** @brief Configure event loop for child process handling
690 * @return 0 on success, non-0 on error
692 * Currently at most one event loop can handle child processes and it must be
693 * distinguished from others by calling this function on it. This could be
694 * fixed but since no process ever makes use of more than one event loop there
697 int ev_child_setup(ev_source
*ev
) {
698 D(("installing SIGCHLD handler"));
699 return ev_signal(ev
, SIGCHLD
, sigchld_callback
, 0);
702 /** @brief Wait for a child process to terminate
703 * @param ev Event loop
704 * @param pid Process ID of child
705 * @param options Options to pass to @c wait4()
706 * @param callback Called when child terminates (or possibly when it stops)
707 * @param u Passed to @p callback
708 * @return 0 on success, non-0 on error
710 * You must have called ev_child_setup() on @p ev once first.
712 int ev_child(ev_source
*ev
,
715 ev_child_callback
*callback
,
719 D(("registering child handling %ld options %d callback %p %p",
720 (long)pid
, options
, (void *)callback
, u
));
721 assert(ev
->signals
[SIGCHLD
].callback
== sigchld_callback
);
722 if(ev
->nchildren
>= ev
->nchildslots
) {
723 ev
->nchildslots
= ev
->nchildslots ?
2 * ev
->nchildslots
: 16;
724 ev
->children
= xrealloc(ev
->children
,
725 ev
->nchildslots
* sizeof (struct child
));
728 ev
->children
[n
].pid
= pid
;
729 ev
->children
[n
].options
= options
;
730 ev
->children
[n
].callback
= callback
;
731 ev
->children
[n
].u
= u
;
735 /** @brief Stop waiting for a child process
736 * @param ev Event loop
737 * @param pid Child process ID
738 * @return 0 on success, non-0 on error
740 int ev_child_cancel(ev_source
*ev
,
744 for(n
= 0; n
< ev
->nchildren
&& ev
->children
[n
].pid
!= pid
; ++n
)
746 assert(n
< ev
->nchildren
);
747 if(n
!= ev
->nchildren
- 1)
748 ev
->children
[n
] = ev
->children
[ev
->nchildren
- 1];
753 /* socket listeners ***********************************************************/
755 /** @brief State for a socket listener */
756 struct listen_state
{
757 ev_listen_callback
*callback
;
761 /** @brief Called when a listenign socket is readable */
762 static int listen_callback(ev_source
*ev
, int fd
, void *u
) {
763 const struct listen_state
*l
= u
;
766 struct sockaddr_in in
;
767 #if HAVE_STRUCT_SOCKADDR_IN6
768 struct sockaddr_in6 in6
;
770 struct sockaddr_un un
;
776 D(("callback for listener fd %d", fd
));
777 while((addrlen
= sizeof addr
),
778 (newfd
= accept(fd
, &addr
.sa
, &addrlen
)) >= 0) {
779 if((ret
= l
->callback(ev
, newfd
, &addr
.sa
, addrlen
, l
->u
)))
788 error(errno
, "error calling accept");
793 /* XXX on some systems EPROTO should be fatal, but we don't know if
794 * we're running on one of them */
795 error(errno
, "error calling accept");
799 fatal(errno
, "error calling accept");
802 if(errno
!= EINTR
&& errno
!= EAGAIN
)
803 error(errno
, "error calling accept");
807 /** @brief Listen on a socket for inbound stream connections
808 * @param ev Event source
809 * @param fd File descriptor of socket
810 * @param callback Called when a new connection arrives
811 * @param u Passed to @p callback
812 * @param what Text description of socket
813 * @return 0 on success, non-0 on error
815 int ev_listen(ev_source
*ev
,
817 ev_listen_callback
*callback
,
820 struct listen_state
*l
= xmalloc(sizeof *l
);
822 D(("registering listener fd %d callback %p %p", fd
, (void *)callback
, u
));
823 l
->callback
= callback
;
825 return ev_fd(ev
, ev_read
, fd
, listen_callback
, l
, what
);
828 /** @brief Stop listening on a socket
829 * @param ev Event loop
830 * @param fd File descriptor of socket
831 * @return 0 on success, non-0 on error
833 int ev_listen_cancel(ev_source
*ev
, int fd
) {
834 D(("cancelling listener fd %d", fd
));
835 return ev_fd_cancel(ev
, ev_read
, fd
);
838 /* buffer *********************************************************************/
840 /** @brief Buffer structure */
842 char *base
, *start
, *end
, *top
;
845 /* @brief Make sure there is @p bytes available at @c b->end */
846 static void buffer_space(struct buffer
*b
, size_t bytes
) {
847 D(("buffer_space %p %p %p %p want %lu",
848 (void *)b
->base
, (void *)b
->start
, (void *)b
->end
, (void *)b
->top
,
849 (unsigned long)bytes
));
850 if(b
->start
== b
->end
)
851 b
->start
= b
->end
= b
->base
;
852 if((size_t)(b
->top
- b
->end
) < bytes
) {
853 if((size_t)((b
->top
- b
->end
) + (b
->start
- b
->base
)) < bytes
) {
854 size_t newspace
= b
->end
- b
->start
+ bytes
, n
;
857 for(n
= 16; n
< newspace
; n
*= 2)
859 newbase
= xmalloc_noptr(n
);
860 memcpy(newbase
, b
->start
, b
->end
- b
->start
);
862 b
->end
= newbase
+ (b
->end
- b
->start
);
863 b
->top
= newbase
+ n
;
864 b
->start
= newbase
; /* must be last */
866 memmove(b
->base
, b
->start
, b
->end
- b
->start
);
867 b
->end
= b
->base
+ (b
->end
- b
->start
);
871 D(("result %p %p %p %p",
872 (void *)b
->base
, (void *)b
->start
, (void *)b
->end
, (void *)b
->top
));
875 /* readers and writers *******************************************************/
877 /** @brief State structure for a buffered writer */
879 /** @brief Sink used for writing to the buffer */
882 /** @brief Output buffer */
885 /** @brief File descriptor to write to */
888 /** @brief Set if there'll be no more output */
891 /** @brief Error/termination callback */
892 ev_error_callback
*callback
;
894 /** @brief Passed to @p callback */
897 /** @brief Parent event source */
900 /** @brief Maximum amount of time between succesful writes, 0 = don't care */
902 /** @brief Maximum amount of data to buffer, 0 = don't care */
904 /** @brief Error code to pass to @p callback (see writer_shutdown()) */
906 /** @brief Timeout handle for @p timebound (or 0) */
907 ev_timeout_handle timeout
;
909 /** @brief Description of this writer */
912 /** @brief Tied reader or 0 */
915 /** @brief Set when abandoned */
919 /** @brief State structure for a buffered reader */
921 /** @brief Input buffer */
923 /** @brief File descriptor read from */
925 /** @brief Called when new data is available */
926 ev_reader_callback
*callback
;
927 /** @brief Called on error and shutdown */
928 ev_error_callback
*error_callback
;
929 /** @brief Passed to @p callback and @p error_callback */
931 /** @brief Parent event loop */
933 /** @brief Set when EOF is detected */
935 /** @brief Error code to pass to error callback */
937 /** @brief Tied writer or NULL */
941 /* buffered writer ************************************************************/
943 /** @brief Shut down the writer
945 * This is called to shut down a writer. The error callback is not called
946 * through any other path. Also we do not cancel @p fd from anywhere else,
947 * though we might disable it.
949 * It has the signature of a timeout callback so that it can be called from a
952 * Calls @p callback with @p w->syntherr as the error code (which might be 0).
954 static int writer_shutdown(ev_source
*ev
,
955 const attribute((unused
)) struct timeval
*now
,
960 return 0; /* already shut down */
961 D(("writer_shutdown fd=%d error=%d", w
->fd
, w
->error
));
962 ev_timeout_cancel(ev
, w
->timeout
);
963 ev_fd_cancel(ev
, ev_write
, w
->fd
);
966 D(("found a tied reader"));
967 /* If there is a reader still around we just untie it */
968 w
->reader
->writer
= 0;
969 shutdown(w
->fd
, SHUT_WR
); /* there'll be no more writes */
971 D(("no tied reader"));
972 /* There's no reader so we are free to close the FD */
976 return w
->callback(ev
, w
->error
, w
->u
);
979 /** @brief Called when a writer's @p timebound expires */
980 static int writer_timebound_exceeded(ev_source
*ev
,
981 const struct timeval
*now
,
983 ev_writer
*const w
= u
;
987 error(0, "abandoning writer '%s' because no writes within %ds",
988 w
->what
, w
->timebound
);
989 w
->error
= ETIMEDOUT
;
991 return writer_shutdown(ev
, now
, u
);
994 /** @brief Set the time bound callback (if not set already) */
995 static void writer_set_timebound(ev_writer
*w
) {
996 if(w
->timebound
&& !w
->timeout
) {
998 ev_source
*const ev
= w
->ev
;
1000 xgettimeofday(&when
, 0);
1001 when
.tv_sec
+= w
->timebound
;
1002 ev_timeout(ev
, &w
->timeout
, &when
, writer_timebound_exceeded
, w
);
1006 /** @brief Called when a writer's file descriptor is writable */
1007 static int writer_callback(ev_source
*ev
, int fd
, void *u
) {
1008 ev_writer
*const w
= u
;
1011 n
= write(fd
, w
->b
.start
, w
->b
.end
- w
->b
.start
);
1012 D(("callback for writer fd %d, %ld bytes, n=%d, errno=%d",
1013 fd
, (long)(w
->b
.end
- w
->b
.start
), n
, errno
));
1015 /* Consume bytes from the buffer */
1017 /* Suppress any outstanding timeout */
1018 ev_timeout_cancel(ev
, w
->timeout
);
1020 if(w
->b
.start
== w
->b
.end
) {
1021 /* The buffer is empty */
1023 /* We're done, we can shut down this writer */
1025 return writer_shutdown(ev
, 0, w
);
1027 /* There might be more to come but we don't need writer_callback() to
1028 * be called for the time being */
1029 ev_fd_disable(ev
, ev_write
, fd
);
1031 /* The buffer isn't empty, set a timeout so we give up if we don't manage
1032 * to write some more within a reasonable time */
1033 writer_set_timebound(w
);
1041 return writer_shutdown(ev
, 0, w
);
1047 /** @brief Write bytes to a writer's buffer
1049 * This is the sink write callback.
1051 * Calls ev_fd_enable() if necessary (i.e. if the buffer was empty but
1054 static int ev_writer_write(struct sink
*sk
, const void *s
, int n
) {
1055 ev_writer
*w
= (ev_writer
*)sk
;
1058 return 0; /* avoid silliness */
1060 error(0, "ev_writer_write on %s after shutdown", w
->what
);
1061 if(w
->spacebound
&& w
->b
.end
- w
->b
.start
+ n
> w
->spacebound
) {
1062 /* The new buffer contents will exceed the space bound. We assume that the
1063 * remote client has gone away and TCP hasn't noticed yet, or that it's got
1064 * hopelessly stuck. */
1067 error(0, "abandoning writer '%s' because buffer has reached %td bytes",
1068 w
->what
, w
->b
.end
- w
->b
.start
);
1069 ev_fd_disable(w
->ev
, ev_write
, w
->fd
);
1071 return ev_timeout(w
->ev
, 0, 0, writer_shutdown
, w
);
1075 /* Make sure there is space */
1076 buffer_space(&w
->b
, n
);
1077 /* If the buffer was formerly empty then we'll need to re-enable the FD */
1078 if(w
->b
.start
== w
->b
.end
)
1079 ev_fd_enable(w
->ev
, ev_write
, w
->fd
);
1080 memcpy(w
->b
.end
, s
, n
);
1082 /* Arrange a timeout if there wasn't one set already */
1083 writer_set_timebound(w
);
1087 /** @brief Create a new buffered writer
1088 * @param ev Event loop
1089 * @param fd File descriptor to write to
1090 * @param callback Called if an error occurs and when finished
1091 * @param u Passed to @p callback
1092 * @param what Text description
1093 * @return New writer or @c NULL
1095 * Writers own their file descriptor and close it when they have finished with
1098 * If you pass the same fd to a reader and writer, you must tie them together
1101 ev_writer
*ev_writer_new(ev_source
*ev
,
1103 ev_error_callback
*callback
,
1106 ev_writer
*w
= xmalloc(sizeof *w
);
1108 D(("registering writer fd %d callback %p %p", fd
, (void *)callback
, u
));
1109 w
->s
.write
= ev_writer_write
;
1111 w
->callback
= callback
;
1114 w
->timebound
= 10 * 60;
1115 w
->spacebound
= 512 * 1024;
1117 if(ev_fd(ev
, ev_write
, fd
, writer_callback
, w
, what
))
1119 /* Buffer is initially empty so we don't want a callback */
1120 ev_fd_disable(ev
, ev_write
, fd
);
1124 /** @brief Get/set the time bound
1126 * @param new_time_bound New bound or -1 for no change
1127 * @return Latest time bound
1129 * If @p new_time_bound is negative then the current time bound is returned.
1130 * Otherwise it is set and the new value returned.
1132 * The time bound is the number of seconds allowed between writes. If it takes
1133 * longer than this to flush a buffer then the peer will be assumed to be dead
1134 * and an error will be synthesized. 0 means "don't care". The default time
1135 * bound is 10 minutes.
1137 * Note that this value does not take into account kernel buffering and
1140 int ev_writer_time_bound(ev_writer
*w
,
1141 int new_time_bound
) {
1142 if(new_time_bound
>= 0)
1143 w
->timebound
= new_time_bound
;
1144 return w
->timebound
;
1147 /** @brief Get/set the space bound
1149 * @param new_space_bound New bound or -1 for no change
1150 * @return Latest space bound
1152 * If @p new_space_bound is negative then the current space bound is returned.
1153 * Otherwise it is set and the new value returned.
1155 * The space bound is the number of bytes allowed between in the buffer. If
1156 * the buffer exceeds this size an error will be synthesized. 0 means "don't
1157 * care". The default space bound is 512Kbyte.
1159 * Note that this value does not take into account kernel buffering.
1161 int ev_writer_space_bound(ev_writer
*w
,
1162 int new_space_bound
) {
1163 if(new_space_bound
>= 0)
1164 w
->spacebound
= new_space_bound
;
1165 return w
->spacebound
;
1168 /** @brief Return the sink associated with a writer
1170 * @return Pointer to sink
1172 * Writing to the sink will arrange for those bytes to be written to the file
1173 * descriptor as and when it is writable.
1175 struct sink
*ev_writer_sink(ev_writer
*w
) {
1177 fatal(0, "ev_write_sink called with null writer");
1181 /** @brief Close a writer
1182 * @param w Writer to close
1183 * @return 0 on success, non-0 on error
1185 * Close a writer. No more bytes should be written to its sink.
1187 * When the last byte has been written the callback will be called with an
1188 * error code of 0. It is guaranteed that this will NOT happen before
1189 * ev_writer_close() returns (although the file descriptor for the writer might
1190 * be cancelled by the time it returns).
1192 int ev_writer_close(ev_writer
*w
) {
1193 D(("close writer fd %d", w
->fd
));
1195 return 0; /* already closed */
1197 if(w
->b
.start
== w
->b
.end
) {
1198 /* We're already finished */
1199 w
->error
= 0; /* no error */
1200 return ev_timeout(w
->ev
, 0, 0, writer_shutdown
, w
);
1205 /** @brief Attempt to flush a writer
1206 * @param w Writer to flush
1207 * @return 0 on success, non-0 on error
1209 * Does a speculative write of any buffered data. Does not block if it cannot
1212 int ev_writer_flush(ev_writer
*w
) {
1213 return writer_callback(w
->ev
, w
->fd
, w
);
1216 /* buffered reader ************************************************************/
1218 /** @brief Shut down a reader
1220 * This is the only path through which we cancel and close the file descriptor.
1221 * As with the writer case it is given timeout signature to allow it be
1222 * deferred to the next iteration of the event loop.
1224 * We only call @p error_callback if @p error is nonzero (unlike the writer
1227 static int reader_shutdown(ev_source
*ev
,
1228 const attribute((unused
)) struct timeval
*now
,
1230 ev_reader
*const r
= u
;
1233 return 0; /* already shut down */
1234 D(("reader_shutdown fd=%d", r
->fd
));
1235 ev_fd_cancel(ev
, ev_read
, r
->fd
);
1238 D(("found a tied writer"));
1239 /* If there is a writer still around we just untie it */
1240 r
->writer
->reader
= 0;
1241 shutdown(r
->fd
, SHUT_RD
); /* there'll be no more reads */
1243 D(("no tied writer found"));
1244 /* There's no writer so we are free to close the FD */
1249 return r
->error_callback(ev
, r
->error
, r
->u
);
1254 /** @brief Called when a reader's @p fd is readable */
1255 static int reader_callback(ev_source
*ev
, int fd
, void *u
) {
1259 buffer_space(&r
->b
, 1);
1260 n
= read(fd
, r
->b
.end
, r
->b
.top
- r
->b
.end
);
1261 D(("read fd %d buffer %d returned %d errno %d",
1262 fd
, (int)(r
->b
.top
- r
->b
.end
), n
, errno
));
1265 return r
->callback(ev
, r
, r
->b
.start
, r
->b
.end
- r
->b
.start
, 0, r
->u
);
1267 /* No more read callbacks needed */
1268 ev_fd_disable(r
->ev
, ev_read
, r
->fd
);
1269 ev_timeout(r
->ev
, 0, 0, reader_shutdown
, r
);
1270 /* Pass the remaining data and an eof indicator to the user */
1271 return r
->callback(ev
, r
, r
->b
.start
, r
->b
.end
- r
->b
.start
, 1, r
->u
);
1278 /* Fatal error, kill the reader now */
1280 return reader_shutdown(ev
, 0, r
);
1286 /** @brief Create a new buffered reader
1287 * @param ev Event loop
1288 * @param fd File descriptor to read from
1289 * @param callback Called when new data is available
1290 * @param error_callback Called if an error occurs
1291 * @param u Passed to callbacks
1292 * @param what Text description
1293 * @return New reader or @c NULL
1295 * Readers own their fd and close it when they are finished with it.
1297 * If you pass the same fd to a reader and writer, you must tie them together
1300 ev_reader
*ev_reader_new(ev_source
*ev
,
1302 ev_reader_callback
*callback
,
1303 ev_error_callback
*error_callback
,
1306 ev_reader
*r
= xmalloc(sizeof *r
);
1308 D(("registering reader fd %d callback %p %p %p",
1309 fd
, (void *)callback
, (void *)error_callback
, u
));
1311 r
->callback
= callback
;
1312 r
->error_callback
= error_callback
;
1315 if(ev_fd(ev
, ev_read
, fd
, reader_callback
, r
, what
))
1320 void ev_reader_buffer(ev_reader
*r
, size_t nbytes
) {
1321 buffer_space(&r
->b
, nbytes
- (r
->b
.end
- r
->b
.start
));
1324 /** @brief Consume @p n bytes from the reader's buffer
1326 * @param n Number of bytes to consume
1328 * Tells the reader than the next @p n bytes have been dealt with and can now
1331 void ev_reader_consume(ev_reader
*r
, size_t n
) {
1335 /** @brief Cancel a reader
1337 * @return 0 on success, non-0 on error
1339 * No further callbacks will be made, and the FD will be closed (in a later
1340 * iteration of the event loop).
1342 int ev_reader_cancel(ev_reader
*r
) {
1343 D(("cancel reader fd %d", r
->fd
));
1345 return 0; /* already thoroughly cancelled */
1346 ev_fd_disable(r
->ev
, ev_read
, r
->fd
);
1347 return ev_timeout(r
->ev
, 0, 0, reader_shutdown
, r
);
1350 /** @brief Temporarily disable a reader
1352 * @return 0 on success, non-0 on error
1354 * No further callbacks for this reader will be made. Re-enable with
1355 * ev_reader_enable().
1357 int ev_reader_disable(ev_reader
*r
) {
1358 D(("disable reader fd %d", r
->fd
));
1359 return ev_fd_disable(r
->ev
, ev_read
, r
->fd
);
1362 /** @brief Called from ev_run() for ev_reader_incomplete() */
1363 static int reader_continuation(ev_source
attribute((unused
)) *ev
,
1364 const attribute((unused
)) struct timeval
*now
,
1368 D(("reader continuation callback fd %d", r
->fd
));
1369 /* If not at EOF turn the FD back on */
1371 if(ev_fd_enable(r
->ev
, ev_read
, r
->fd
))
1373 /* We're already in a timeout callback so there's no reason we can't call the
1374 * user callback directly (compare ev_reader_enable()). */
1375 return r
->callback(ev
, r
, r
->b
.start
, r
->b
.end
- r
->b
.start
, r
->eof
, r
->u
);
1378 /** @brief Arrange another callback
1380 * @return 0 on success, non-0 on error
1382 * Indicates that the reader can process more input but would like to yield to
1383 * other clients of the event loop. Input will be disabled but it will be
1384 * re-enabled on the next iteration of the event loop and the read callback
1385 * will be called again (even if no further bytes are available).
1387 int ev_reader_incomplete(ev_reader
*r
) {
1388 if(ev_fd_disable(r
->ev
, ev_read
, r
->fd
)) return -1;
1389 return ev_timeout(r
->ev
, 0, 0, reader_continuation
, r
);
1392 static int reader_enabled(ev_source
*ev
,
1393 const attribute((unused
)) struct timeval
*now
,
1397 D(("reader enabled callback fd %d", r
->fd
));
1398 return r
->callback(ev
, r
, r
->b
.start
, r
->b
.end
- r
->b
.start
, r
->eof
, r
->u
);
1401 /** @brief Re-enable reading
1403 * @return 0 on success, non-0 on error
1405 * If there is unconsumed data then you get a callback next time round the
1406 * event loop even if nothing new has been read.
1408 * The idea is in your read callback you come across a line (or whatever) that
1409 * can't be processed immediately. So you set up processing and disable
1410 * reading with ev_reader_disable(). Later when you finish processing you
1411 * re-enable. You'll automatically get another callback directly from the
1412 * event loop (i.e. not from inside ev_reader_enable()) so you can handle the
1413 * next line (or whatever) if the whole thing has in fact already arrived.
1415 * The difference between this process and calling ev_reader_incomplete() is
1416 * ev_reader_incomplete() deals with the case where you can process now but
1417 * would rather yield to other clients of the event loop, while using
1418 * ev_reader_disable() and ev_reader_enable() deals with the case where you
1419 * cannot process input yet because some other process is actually not
1422 int ev_reader_enable(ev_reader
*r
) {
1423 D(("enable reader fd %d", r
->fd
));
1425 /* First if we're not at EOF then we re-enable reading */
1427 if(ev_fd_enable(r
->ev
, ev_read
, r
->fd
))
1429 /* Arrange another callback next time round the event loop */
1430 return ev_timeout(r
->ev
, 0, 0, reader_enabled
, r
);
1433 /** @brief Tie a reader and a writer together
1436 * @return 0 on success, non-0 on error
1438 * This function must be called if @p r and @p w share a file descritptor.
1440 int ev_tie(ev_reader
*r
, ev_writer
*w
) {
1441 assert(r
->writer
== 0);
1442 assert(w
->reader
== 0);