1 /*
2  * Copyright (c) 2021 Omar Polo <op@omarpolo.com>
3  *
4  * Permission to use, copy, modify, and distribute this software for any
5  * purpose with or without fee is hereby granted, provided that the above
6  * copyright notice and this permission notice appear in all copies.
7  *
8  * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
9  * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
10  * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
11  * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
12  * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
13  * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
14  * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
15  */
16 
17 #include "gmid.h"
18 
19 #include <sys/stat.h>
20 
21 #include <assert.h>
22 #include <ctype.h>
23 #include <errno.h>
24 #include <event.h>
25 #include <fcntl.h>
26 #include <fnmatch.h>
27 #include <limits.h>
28 #include <string.h>
29 
30 #define MIN(a, b)	((a) < (b) ? (a) : (b))
31 
32 int shutting_down;
33 
34 static struct tls	*ctx;
35 
36 static struct event e4, e6, imsgev, siginfo, sigusr2;
37 static int has_ipv6, has_siginfo;
38 
39 int connected_clients;
40 
41 static inline int matches(const char*, const char*);
42 
43 static int	 check_path(struct client*, const char*, int*);
44 static void	 open_file(struct client*);
45 static void	 check_for_cgi(struct client*);
46 static void	 handle_handshake(int, short, void*);
47 static const char *strip_path(const char*, int);
48 static void	 fmt_sbuf(const char*, struct client*, const char*);
49 static int	 apply_block_return(struct client*);
50 static int	 apply_fastcgi(struct client*);
51 static int	 apply_require_ca(struct client*);
52 static size_t	 host_nth(struct vhost*);
53 static void	 start_cgi(const char*, const char*, struct client*);
54 static void	 open_dir(struct client*);
55 static void	 redirect_canonical_dir(struct client*);
56 
57 static void	 client_tls_readcb(int, short, void *);
58 static void	 client_tls_writecb(int, short, void *);
59 
60 static void	 client_read(struct bufferevent *, void *);
61 void		 client_write(struct bufferevent *, void *);
62 static void	 client_error(struct bufferevent *, short, void *);
63 
64 static void	 client_close_ev(int, short, void *);
65 
66 static void	 cgi_read(struct bufferevent *, void *);
67 static void	 cgi_write(struct bufferevent *, void *);
68 static void	 cgi_error(struct bufferevent *, short, void *);
69 
70 static void	 do_accept(int, short, void*);
71 static struct client *client_by_id(int);
72 
73 static void	 handle_imsg_cgi_res(struct imsgbuf*, struct imsg*, size_t);
74 static void	 handle_imsg_fcgi_fd(struct imsgbuf*, struct imsg*, size_t);
75 static void	 handle_imsg_quit(struct imsgbuf*, struct imsg*, size_t);
76 static void	 handle_dispatch_imsg(int, short, void *);
77 static void	 handle_siginfo(int, short, void*);
78 
79 static imsg_handlerfn *handlers[] = {
80 	[IMSG_QUIT] = handle_imsg_quit,
81 	[IMSG_CGI_RES] = handle_imsg_cgi_res,
82 	[IMSG_FCGI_FD] = handle_imsg_fcgi_fd,
83 };
84 
85 static uint32_t server_client_id;
86 
87 struct client_tree_id clients;
88 
89 static inline int
90 matches(const char *pattern, const char *path)
91 {
92 	if (*path == '/')
93 		path++;
94 	return !fnmatch(pattern, path, 0);
95 }
96 
97 const char *
98 vhost_lang(struct vhost *v, const char *path)
99 {
100 	struct location *loc;
101 
102 	if (v == NULL || path == NULL)
103 		return NULL;
104 
105 	loc = TAILQ_FIRST(&v->locations);
106 	while ((loc = TAILQ_NEXT(loc, locations)) != NULL) {
107 		if (loc->lang != NULL) {
108 			if (matches(loc->match, path))
109 				return loc->lang;
110 		}
111 	}
112 
113 	return TAILQ_FIRST(&v->locations)->lang;
114 }
115 
116 const char *
117 vhost_default_mime(struct vhost *v, const char *path)
118 {
119 	struct location *loc;
120 	const char *default_mime = "application/octet-stream";
121 
122 	if (v == NULL || path == NULL)
123 		return default_mime;
124 
125 	loc = TAILQ_FIRST(&v->locations);
126 	while ((loc = TAILQ_NEXT(loc, locations)) != NULL) {
127 		if (loc->default_mime != NULL) {
128 			if (matches(loc->match, path))
129 				return loc->default_mime;
130 		}
131 	}
132 
133 	loc = TAILQ_FIRST(&v->locations);
134 	if (loc->default_mime != NULL)
135 		return loc->default_mime;
136 	return default_mime;
137 }
138 
139 const char *
140 vhost_index(struct vhost *v, const char *path)
141 {
142 	struct location *loc;
143 	const char *index = "index.gmi";
144 
145 	if (v == NULL || path == NULL)
146 		return index;
147 
148 	loc = TAILQ_FIRST(&v->locations);
149 	while ((loc = TAILQ_NEXT(loc, locations)) != NULL) {
150 		if (loc->index != NULL) {
151 			if (matches(loc->match, path))
152 				return loc->index;
153 		}
154 	}
155 
156 	loc = TAILQ_FIRST(&v->locations);
157 	if (loc->index != NULL)
158 		return loc->index;
159 	return index;
160 }
161 
162 int
163 vhost_auto_index(struct vhost *v, const char *path)
164 {
165 	struct location *loc;
166 
167 	if (v == NULL || path == NULL)
168 		return 0;
169 
170 	loc = TAILQ_FIRST(&v->locations);
171 	while ((loc = TAILQ_NEXT(loc, locations)) != NULL) {
172 		if (loc->auto_index != 0) {
173 			if (matches(loc->match, path))
174 				return loc->auto_index == 1;
175 		}
176 	}
177 
178 	loc = TAILQ_FIRST(&v->locations);
179 	return loc->auto_index == 1;
180 }
181 
182 int
183 vhost_block_return(struct vhost *v, const char *path, int *code, const char **fmt)
184 {
185 	struct location *loc;
186 
187 	if (v == NULL || path == NULL)
188 		return 0;
189 
190 	loc = TAILQ_FIRST(&v->locations);
191 	while ((loc = TAILQ_NEXT(loc, locations)) != NULL) {
192 		if (loc->block_code != 0) {
193 			if (matches(loc->match, path)) {
194 				*code = loc->block_code;
195 				*fmt = loc->block_fmt;
196 				return 1;
197 			}
198 		}
199 	}
200 
201 	loc = TAILQ_FIRST(&v->locations);
202 	*code = loc->block_code;
203 	*fmt = loc->block_fmt;
204 	return loc->block_code != 0;
205 }
206 
207 int
208 vhost_fastcgi(struct vhost *v, const char *path)
209 {
210 	struct location *loc;
211 
212 	if (v == NULL || path == NULL)
213 		return -1;
214 
215 	loc = TAILQ_FIRST(&v->locations);
216 	while ((loc = TAILQ_NEXT(loc, locations)) != NULL) {
217 		if (loc->fcgi != -1)
218 			if (matches(loc->match, path))
219 				return loc->fcgi;
220 	}
221 
222 	loc = TAILQ_FIRST(&v->locations);
223 	return loc->fcgi;
224 }
225 
226 int
227 vhost_dirfd(struct vhost *v, const char *path, size_t *retloc)
228 {
229 	struct location *loc;
230 	size_t		 l = 0;
231 
232 	if (v == NULL || path == NULL)
233 		return -1;
234 
235 	loc = TAILQ_FIRST(&v->locations);
236 	while ((loc = TAILQ_NEXT(loc, locations)) != NULL) {
237 		l++;
238 		if (loc->dirfd != -1)
239 			if (matches(loc->match, path)) {
240 				*retloc = l;
241 				return loc->dirfd;
242 			}
243 	}
244 
245 	*retloc = 0;
246 	loc = TAILQ_FIRST(&v->locations);
247 	return loc->dirfd;
248 }
249 
250 int
251 vhost_strip(struct vhost *v, const char *path)
252 {
253 	struct location *loc;
254 
255 	if (v == NULL || path == NULL)
256 		return 0;
257 
258 	loc = TAILQ_FIRST(&v->locations);
259 	while ((loc = TAILQ_NEXT(loc, locations)) != NULL) {
260 		if (loc->strip != 0) {
261 			if (matches(loc->match, path))
262 				return loc->strip;
263 		}
264 	}
265 
266 	loc = TAILQ_FIRST(&v->locations);
267 	return loc->strip;
268 }
269 
270 X509_STORE *
271 vhost_require_ca(struct vhost *v, const char *path)
272 {
273 	struct location *loc;
274 
275 	if (v == NULL || path == NULL)
276 		return NULL;
277 
278 	loc = TAILQ_FIRST(&v->locations);
279 	while ((loc = TAILQ_NEXT(loc, locations)) != NULL) {
280 		if (loc->reqca != NULL) {
281 			if (matches(loc->match, path))
282 				return loc->reqca;
283 		}
284 	}
285 
286 	loc = TAILQ_FIRST(&v->locations);
287 	return loc->reqca;
288 }
289 
290 int
291 vhost_disable_log(struct vhost *v, const char *path)
292 {
293 	struct location *loc;
294 
295 	if (v == NULL || path == NULL)
296 		return 0;
297 
298 	loc = TAILQ_FIRST(&v->locations);
299 	while ((loc = TAILQ_NEXT(loc, locations)) != NULL) {
300 		if (loc->disable_log && matches(loc->match, path))
301 				return 1;
302 	}
303 
304 	loc = TAILQ_FIRST(&v->locations);
305 	return loc->disable_log;
306 }
307 
308 static int
309 check_path(struct client *c, const char *path, int *fd)
310 {
311 	struct stat sb;
312 	const char *p;
313 	int dirfd, strip;
314 
315 	assert(path != NULL);
316 
317 	/*
318 	 * in send_dir we add an initial / (to be redirect-friendly),
319 	 * but here we want to skip it
320 	 */
321 	if (*path == '/')
322 		path++;
323 
324 	strip = vhost_strip(c->host, path);
325 	p = strip_path(path, strip);
326 
327 	if (*p == '/')
328 		p = p+1;
329 	if (*p == '\0')
330 		p = ".";
331 
332 	dirfd = vhost_dirfd(c->host, path, &c->loc);
333 	log_debug(c, "check_path: strip=%d path=%s original=%s",
334 	    strip, p, path);
335 	if (*fd == -1 && (*fd = openat(dirfd, p, O_RDONLY)) == -1)
336 		return FILE_MISSING;
337 
338 	if (fstat(*fd, &sb) == -1) {
339 		log_notice(c, "failed stat for %s: %s", path, strerror(errno));
340 		return FILE_MISSING;
341 	}
342 
343 	if (S_ISDIR(sb.st_mode))
344 		return FILE_DIRECTORY;
345 
346 	if (sb.st_mode & S_IXUSR)
347 		return FILE_EXECUTABLE;
348 
349 	return FILE_EXISTS;
350 }
351 
352 static void
353 open_file(struct client *c)
354 {
355 	switch (check_path(c, c->iri.path, &c->pfd)) {
356 	case FILE_EXECUTABLE:
357 		if (c->host->cgi != NULL && matches(c->host->cgi, c->iri.path)) {
358 			start_cgi(c->iri.path, "", c);
359 			return;
360 		}
361 
362 		/* fallthrough */
363 
364 	case FILE_EXISTS:
365 		c->type = REQUEST_FILE;
366 		start_reply(c, SUCCESS, mime(c->host, c->iri.path));
367 		return;
368 
369 	case FILE_DIRECTORY:
370 		open_dir(c);
371 		return;
372 
373 	case FILE_MISSING:
374 		if (c->host->cgi != NULL && matches(c->host->cgi, c->iri.path)) {
375 			check_for_cgi(c);
376 			return;
377 		}
378 		start_reply(c, NOT_FOUND, "not found");
379 		return;
380 
381 	default:
382 		/* unreachable */
383 		abort();
384 	}
385 }
386 
387 /*
388  * the inverse of this algorithm, i.e. starting from the start of the
389  * path + strlen(cgi), and checking if each component, should be
390  * faster.  But it's tedious to write.  This does the opposite: starts
391  * from the end and strip one component at a time, until either an
392  * executable is found or we emptied the path.
393  */
394 static void
395 check_for_cgi(struct client *c)
396 {
397 	char path[PATH_MAX];
398 	char *end;
399 
400 	strlcpy(path, c->iri.path, sizeof(path));
401 	end = strchr(path, '\0');
402 
403 	while (end > path) {
404 		/*
405 		 * go up one level.  UNIX paths are simple and POSIX
406 		 * dirname, with its ambiguities on if the given
407 		 * pointer is changed or not, gives me headaches.
408 		 */
409 		while (*end != '/' && end > path)
410 			end--;
411 
412 		if (end == path)
413 			break;
414 
415 		*end = '\0';
416 
417 		switch (check_path(c, path, &c->pfd)) {
418 		case FILE_EXECUTABLE:
419 			start_cgi(path, end+1, c);
420 			return;
421 		case FILE_MISSING:
422 			break;
423 		default:
424 			goto err;
425 		}
426 
427 		*end = '/';
428 		end--;
429 	}
430 
431 err:
432 	start_reply(c, NOT_FOUND, "not found");
433 	return;
434 }
435 
436 void
437 mark_nonblock(int fd)
438 {
439 	int flags;
440 
441 	if ((flags = fcntl(fd, F_GETFL)) == -1)
442 		fatal("fcntl(F_GETFL): %s", strerror(errno));
443 	if (fcntl(fd, F_SETFL, flags | O_NONBLOCK) == -1)
444 		fatal("fcntl(F_SETFL): %s", strerror(errno));
445 }
446 
447 static void
448 handle_handshake(int fd, short ev, void *d)
449 {
450 	struct client *c = d;
451 	struct vhost *h;
452 	struct alist *a;
453 	const char *servname;
454 	const char *parse_err = "unknown error";
455 
456 	switch (tls_handshake(c->ctx)) {
457 	case 0:  /* success */
458 	case -1: /* already handshaked */
459 		break;
460 	case TLS_WANT_POLLIN:
461 		event_once(c->fd, EV_READ, handle_handshake, c, NULL);
462 		return;
463 	case TLS_WANT_POLLOUT:
464 		event_once(c->fd, EV_WRITE, handle_handshake, c, NULL);
465 		return;
466 	default:
467 		/* unreachable */
468 		abort();
469 	}
470 
471 	c->bev = bufferevent_new(fd, client_read, client_write,
472 	    client_error, c);
473 	if (c->bev == NULL)
474 		fatal("%s: failed to allocate client buffer: %s",
475 		    __func__, strerror(errno));
476 
477 	event_set(&c->bev->ev_read, c->fd, EV_READ,
478 	    client_tls_readcb, c->bev);
479 	event_set(&c->bev->ev_write, c->fd, EV_WRITE,
480 	    client_tls_writecb, c->bev);
481 
482 #if HAVE_LIBEVENT2
483 	evbuffer_unfreeze(c->bev->input, 0);
484 	evbuffer_unfreeze(c->bev->output, 1);
485 #endif
486 
487 	if ((servname = tls_conn_servername(c->ctx)) == NULL) {
488 		log_debug(c, "handshake: missing SNI");
489 		goto err;
490 	}
491 
492 	if (!puny_decode(servname, c->domain, sizeof(c->domain), &parse_err)) {
493 		log_info(c, "puny_decode: %s", parse_err);
494 		goto err;
495 	}
496 
497 	TAILQ_FOREACH(h, &hosts, vhosts) {
498 		if (matches(h->domain, c->domain))
499 			goto found;
500 		TAILQ_FOREACH(a, &h->aliases, aliases) {
501 			if (matches(a->alias, c->domain))
502 				goto found;
503 		}
504 	}
505 
506 found:
507 	log_debug(c, "handshake: SNI: \"%s\"; decoded: \"%s\"; matched: \"%s\"",
508 	    servname != NULL ? servname : "(null)",
509 	    c->domain,
510 	    h != NULL ? h->domain : "(null)");
511 
512 	if (h != NULL) {
513 		c->host = h;
514 		bufferevent_enable(c->bev, EV_READ);
515 		return;
516 	}
517 
518 err:
519 	start_reply(c, BAD_REQUEST, "Wrong/malformed host or missing SNI");
520 }
521 
522 static const char *
523 strip_path(const char *path, int strip)
524 {
525 	char *t;
526 
527 	while (strip > 0) {
528 		if ((t = strchr(path, '/')) == NULL) {
529 			path = strchr(path, '\0');
530 			break;
531 		}
532 		path = t;
533 		strip--;
534 	}
535 
536 	return path;
537 }
538 
539 static void
540 fmt_sbuf(const char *fmt, struct client *c, const char *path)
541 {
542 	size_t i;
543 	char buf[32];
544 
545 	memset(buf, 0, sizeof(buf));
546 	for (i = 0; *fmt; ++fmt) {
547 		if (i == sizeof(buf)-1 || *fmt == '%') {
548 			strlcat(c->sbuf, buf, sizeof(c->sbuf));
549 			memset(buf, 0, sizeof(buf));
550 			i = 0;
551 		}
552 
553 		if (*fmt != '%') {
554 			buf[i++] = *fmt;
555 			continue;
556 		}
557 
558 		switch (*++fmt) {
559 		case '%':
560 			strlcat(c->sbuf, "%", sizeof(c->sbuf));
561 			break;
562 		case 'p':
563 			if (*path != '/')
564 				strlcat(c->sbuf, "/", sizeof(c->sbuf));
565 			strlcat(c->sbuf, path, sizeof(c->sbuf));
566 			break;
567 		case 'q':
568 			strlcat(c->sbuf, c->iri.query, sizeof(c->sbuf));
569 			break;
570 		case 'P':
571 			snprintf(buf, sizeof(buf), "%d", conf.port);
572 			strlcat(c->sbuf, buf, sizeof(c->sbuf));
573 			memset(buf, 0, sizeof(buf));
574 			break;
575 		case 'N':
576 			strlcat(c->sbuf, c->domain, sizeof(c->sbuf));
577 			break;
578 		default:
579 			fatal("%s: unknown fmt specifier %c",
580 			    __func__, *fmt);
581 		}
582 	}
583 
584 	if (i != 0)
585 		strlcat(c->sbuf, buf, sizeof(c->sbuf));
586 }
587 
588 /* 1 if a matching `block return' (and apply it), 0 otherwise */
589 static int
590 apply_block_return(struct client *c)
591 {
592 	const char *fmt, *path;
593 	int code;
594 
595 	if (!vhost_block_return(c->host, c->iri.path, &code, &fmt))
596 		return 0;
597 
598 	path = strip_path(c->iri.path, vhost_strip(c->host, c->iri.path));
599 	fmt_sbuf(fmt, c, path);
600 
601 	start_reply(c, code, c->sbuf);
602 	return 1;
603 }
604 
605 /* 1 if matching `fcgi' (and apply it), 0 otherwise */
606 static int
607 apply_fastcgi(struct client *c)
608 {
609 	int		 id;
610 	struct fcgi	*f;
611 
612 	if ((id = vhost_fastcgi(c->host, c->iri.path)) == -1)
613 		return 0;
614 
615 	f = &fcgi[id];
616 
617 	log_debug(c, "opening fastcgi connection for (%s,%s,%s)",
618 	    f->path, f->port, f->prog);
619 
620 	imsg_compose(&exibuf, IMSG_FCGI_REQ, c->id, 0, -1,
621 	    &id, sizeof(id));
622 	imsg_flush(&exibuf);
623 	return 1;
624 }
625 
626 /* 1 if matching `require client ca' fails (and apply it), 0 otherwise */
627 static int
628 apply_require_ca(struct client *c)
629 {
630 	X509_STORE	*store;
631 	const uint8_t	*cert;
632 	size_t		 len;
633 
634 	if ((store = vhost_require_ca(c->host, c->iri.path)) == NULL)
635 		return 0;
636 
637 	if (!tls_peer_cert_provided(c->ctx)) {
638 		start_reply(c, CLIENT_CERT_REQ, "client certificate required");
639 		return 1;
640 	}
641 
642 	cert = tls_peer_cert_chain_pem(c->ctx, &len);
643 	if (!validate_against_ca(store, cert, len)) {
644 		start_reply(c, CERT_NOT_AUTH, "certificate not authorised");
645 		return 1;
646 	}
647 
648 	return 0;
649 }
650 
651 static size_t
652 host_nth(struct vhost *h)
653 {
654 	struct vhost	*v;
655 	size_t		 i = 0;
656 
657 	TAILQ_FOREACH(v, &hosts, vhosts) {
658 		if (v == h)
659 			return i;
660 		i++;
661 	}
662 
663 	abort();
664 }
665 
666 static void
667 start_cgi(const char *spath, const char *relpath, struct client *c)
668 {
669 	char addr[NI_MAXHOST];
670 	const char *t;
671 	struct cgireq req;
672 	int e;
673 
674 	c->type = REQUEST_CGI;
675 
676 	e = getnameinfo((struct sockaddr*)&c->addr, sizeof(c->addr),
677 	    addr, sizeof(addr),
678 	    NULL, 0,
679 	    NI_NUMERICHOST);
680 	if (e != 0)
681 		fatal("getnameinfo failed");
682 
683 	memset(&req, 0, sizeof(req));
684 
685 	memcpy(req.buf, c->req, sizeof(req.buf));
686 
687 	req.iri_schema_off = c->iri.schema - c->req;
688 	req.iri_host_off = c->iri.host - c->req;
689 	req.iri_port_off = c->iri.port - c->req;
690 	req.iri_path_off = c->iri.path - c->req;
691 	req.iri_query_off = c->iri.query - c->req;
692 	req.iri_fragment_off = c->iri.fragment - c->req;
693 
694 	req.iri_portno = c->iri.port_no;
695 
696 	strlcpy(req.spath, spath, sizeof(req.spath));
697 	strlcpy(req.relpath, relpath, sizeof(req.relpath));
698 	strlcpy(req.addr, addr, sizeof(req.addr));
699 
700 	if ((t = tls_peer_cert_subject(c->ctx)) != NULL)
701 		strlcpy(req.subject, t, sizeof(req.subject));
702 	if ((t = tls_peer_cert_issuer(c->ctx)) != NULL)
703 		strlcpy(req.issuer, t, sizeof(req.issuer));
704 	if ((t = tls_peer_cert_hash(c->ctx)) != NULL)
705 		strlcpy(req.hash, t, sizeof(req.hash));
706 	if ((t = tls_conn_version(c->ctx)) != NULL)
707 		strlcpy(req.version, t, sizeof(req.version));
708 	if ((t = tls_conn_cipher(c->ctx)) != NULL)
709 		strlcpy(req.cipher, t, sizeof(req.cipher));
710 
711 	req.cipher_strength = tls_conn_cipher_strength(c->ctx);
712 	req.notbefore = tls_peer_cert_notbefore(c->ctx);
713 	req.notafter = tls_peer_cert_notafter(c->ctx);
714 
715 	req.host_off = host_nth(c->host);
716 	req.loc_off = c->loc;
717 
718 	imsg_compose(&exibuf, IMSG_CGI_REQ, c->id, 0, -1, &req, sizeof(req));
719 	imsg_flush(&exibuf);
720 
721 	close(c->pfd);
722 }
723 
724 static void
725 open_dir(struct client *c)
726 {
727 	size_t len;
728 	int dirfd, root;
729 	char *before_file;
730 
731 	root = !strcmp(c->iri.path, "/") || *c->iri.path == '\0';
732 
733 	len = strlen(c->iri.path);
734 	if (len > 0 && !ends_with(c->iri.path, "/")) {
735 		redirect_canonical_dir(c);
736 		return;
737 	}
738 
739 	strlcpy(c->sbuf, "/", sizeof(c->sbuf));
740 	strlcat(c->sbuf, c->iri.path, sizeof(c->sbuf));
741 	if (!ends_with(c->sbuf, "/"))
742 		strlcat(c->sbuf, "/", sizeof(c->sbuf));
743 	before_file = strchr(c->sbuf, '\0');
744 	len = strlcat(c->sbuf, vhost_index(c->host, c->iri.path),
745 	    sizeof(c->sbuf));
746 	if (len >= sizeof(c->sbuf)) {
747 		start_reply(c, TEMP_FAILURE, "internal server error");
748 		return;
749 	}
750 
751 	c->iri.path = c->sbuf;
752 
753 	/* close later unless we have to generate the dir listing */
754 	dirfd = c->pfd;
755 	c->pfd = -1;
756 
757 	switch (check_path(c, c->iri.path, &c->pfd)) {
758 	case FILE_EXECUTABLE:
759 		if (c->host->cgi != NULL && matches(c->host->cgi, c->iri.path)) {
760 			start_cgi(c->iri.path, "", c);
761 			break;
762 		}
763 
764 		/* fallthrough */
765 
766 	case FILE_EXISTS:
767 		c->type = REQUEST_FILE;
768 		start_reply(c, SUCCESS, mime(c->host, c->iri.path));
769 		break;
770 
771 	case FILE_DIRECTORY:
772 		start_reply(c, TEMP_REDIRECT, c->sbuf);
773 		break;
774 
775 	case FILE_MISSING:
776 		*before_file = '\0';
777 
778 		if (!vhost_auto_index(c->host, c->iri.path)) {
779 			start_reply(c, NOT_FOUND, "not found");
780 			break;
781 		}
782 
783 		c->type = REQUEST_DIR;
784 
785 		c->dirlen = scandir_fd(dirfd, &c->dir,
786 		    root ? select_non_dotdot : select_non_dot,
787 		    alphasort);
788 		if (c->dirlen == -1) {
789 			log_err(c, "scandir_fd(%d) (vhost:%s) %s: %s",
790 			    c->pfd, c->host->domain, c->iri.path, strerror(errno));
791 			start_reply(c, TEMP_FAILURE, "internal server error");
792 			return;
793 		}
794 		c->diroff = 0;
795 		c->off = 0;
796 
797 		start_reply(c, SUCCESS, "text/gemini");
798 		evbuffer_add_printf(EVBUFFER_OUTPUT(c->bev),
799 		    "# Index of %s\n\n", c->iri.path);
800 		return;
801 
802 	default:
803 		/* unreachable */
804 		abort();
805 	}
806 
807 	close(dirfd);
808 }
809 
810 static void
811 redirect_canonical_dir(struct client *c)
812 {
813 	size_t len;
814 
815 	strlcpy(c->sbuf, "/", sizeof(c->sbuf));
816 	strlcat(c->sbuf, c->iri.path, sizeof(c->sbuf));
817 	len = strlcat(c->sbuf, "/", sizeof(c->sbuf));
818 
819 	if (len >= sizeof(c->sbuf)) {
820 		start_reply(c, TEMP_FAILURE, "internal server error");
821 		return;
822 	}
823 
824 	start_reply(c, TEMP_REDIRECT, c->sbuf);
825 }
826 
827 static void
828 client_tls_readcb(int fd, short event, void *d)
829 {
830 	struct bufferevent	*bufev = d;
831 	struct client		*client = bufev->cbarg;
832 	ssize_t			 ret;
833 	size_t			 len;
834 	int			 what = EVBUFFER_READ;
835 	int			 howmuch = IBUF_READ_SIZE;
836 	char			 buf[IBUF_READ_SIZE];
837 
838 	if (event == EV_TIMEOUT) {
839 		what |= EVBUFFER_TIMEOUT;
840 		goto err;
841 	}
842 
843 	if (bufev->wm_read.high != 0)
844 		howmuch = MIN(sizeof(buf), bufev->wm_read.high);
845 
846 	switch (ret = tls_read(client->ctx, buf, howmuch)) {
847 	case TLS_WANT_POLLIN:
848 	case TLS_WANT_POLLOUT:
849 		goto retry;
850 	case -1:
851 		what |= EVBUFFER_ERROR;
852 		goto err;
853 	}
854 	len = ret;
855 
856 	if (len == 0) {
857 		what |= EVBUFFER_EOF;
858 		goto err;
859 	}
860 
861 	if (evbuffer_add(bufev->input, buf, len) == -1) {
862 		what |= EVBUFFER_ERROR;
863 		goto err;
864 	}
865 
866 	event_add(&bufev->ev_read, NULL);
867 	if (bufev->wm_read.low != 0 && len < bufev->wm_read.low)
868 		return;
869 	if (bufev->wm_read.high != 0 && len > bufev->wm_read.high) {
870 		/*
871 		 * here we could implement a read pressure policy.
872 		 */
873 	}
874 
875 	if (bufev->readcb != NULL)
876 		(*bufev->readcb)(bufev, bufev->cbarg);
877 
878 	return;
879 
880 retry:
881 	event_add(&bufev->ev_read, NULL);
882 	return;
883 
884 err:
885 	(*bufev->errorcb)(bufev, what, bufev->cbarg);
886 }
887 
888 static void
889 client_tls_writecb(int fd, short event, void *d)
890 {
891 	struct bufferevent	*bufev = d;
892 	struct client		*client = bufev->cbarg;
893 	ssize_t			 ret;
894 	size_t			 len;
895 	short			 what = EVBUFFER_WRITE;
896 
897 	if (event == EV_TIMEOUT) {
898 		what |= EVBUFFER_TIMEOUT;
899 		goto err;
900 	}
901 
902 	if (EVBUFFER_LENGTH(bufev->output) != 0) {
903 		ret = tls_write(client->ctx,
904 		    EVBUFFER_DATA(bufev->output),
905 		    EVBUFFER_LENGTH(bufev->output));
906 		switch (ret) {
907 		case TLS_WANT_POLLIN:
908 		case TLS_WANT_POLLOUT:
909 			goto retry;
910 		case -1:
911 			what |= EVBUFFER_ERROR;
912 			goto err;
913 		}
914 		len = ret;
915 		evbuffer_drain(bufev->output, len);
916 	}
917 
918 	if (EVBUFFER_LENGTH(bufev->output) != 0)
919 		event_add(&bufev->ev_write, NULL);
920 
921 	if (bufev->writecb != NULL &&
922 	    EVBUFFER_LENGTH(bufev->output) <= bufev->wm_write.low)
923 		(*bufev->writecb)(bufev, bufev->cbarg);
924 	return;
925 
926 retry:
927 	event_add(&bufev->ev_write, NULL);
928 	return;
929 err:
930 	log_err(client, "tls error: %s", tls_error(client->ctx));
931 	(*bufev->errorcb)(bufev, what, bufev->cbarg);
932 }
933 
934 static void
935 client_read(struct bufferevent *bev, void *d)
936 {
937 	struct client	*c = d;
938 	struct evbuffer	*src = EVBUFFER_INPUT(bev);
939 	const char	*parse_err = "invalid request";
940 	char		 decoded[DOMAIN_NAME_LEN];
941 	size_t		 len;
942 
943 	bufferevent_disable(bev, EVBUFFER_READ);
944 
945 	/* max url len + \r\n */
946 	if (EVBUFFER_LENGTH(src) > 1024 + 2) {
947 		log_err(c, "too much data received");
948 		start_reply(c, BAD_REQUEST, "bad request");
949 		return;
950 	}
951 
952 	c->req = evbuffer_readln(src, &len, EVBUFFER_EOL_CRLF_STRICT);
953 	if (c->req == NULL) {
954 		/* not enough data yet. */
955 		bufferevent_enable(bev, EVBUFFER_READ);
956 		return;
957 	}
958 
959 	if (!parse_iri(c->req, &c->iri, &parse_err) ||
960 	    !puny_decode(c->iri.host, decoded, sizeof(decoded), &parse_err)) {
961 		log_err(c, "IRI parse error: %s", parse_err);
962 		start_reply(c, BAD_REQUEST, "bad request");
963 		return;
964 	}
965 
966 	/* ignore the port number */
967 	if (strcmp(c->iri.schema, "gemini") ||
968 	    strcmp(decoded, c->domain)) {
969 		start_reply(c, PROXY_REFUSED, "won't proxy request");
970 		return;
971 	}
972 
973 	if (apply_require_ca(c) ||
974 	    apply_block_return(c)||
975 	    apply_fastcgi(c))
976 		return;
977 
978 	if (c->host->entrypoint != NULL) {
979 		c->loc = 0;
980 		start_cgi(c->host->entrypoint, c->iri.path, c);
981 		return;
982 	}
983 
984 	open_file(c);
985 }
986 
987 void
988 client_write(struct bufferevent *bev, void *d)
989 {
990 	struct client	*c = d;
991 	struct evbuffer	*out = EVBUFFER_OUTPUT(bev);
992 	char		 buf[BUFSIZ];
993 	ssize_t		 r;
994 
995 	switch (c->type) {
996 	case REQUEST_UNDECIDED:
997 		/*
998 		 * Ignore spurious calls when we still don't have idea
999 		 * what to do with the request.
1000 		 */
1001 		break;
1002 
1003 	case REQUEST_FILE:
1004 		if ((r = read(c->pfd, buf, sizeof(buf))) == -1) {
1005 			log_warn(c, "read: %s", strerror(errno));
1006 			client_error(bev, EVBUFFER_ERROR, c);
1007 			return;
1008 		} else if (r == 0) {
1009 			client_close(c);
1010 			return;
1011 		} else if (r != sizeof(buf))
1012 			c->type = REQUEST_DONE;
1013 		bufferevent_write(bev, buf, r);
1014 		break;
1015 
1016 	case REQUEST_DIR:
1017 		/* TODO: handle big big directories better */
1018 		for (c->diroff = 0; c->diroff < c->dirlen; ++c->diroff) {
1019 			evbuffer_add_printf(out, "=> %s\n",
1020 			    c->dir[c->diroff]->d_name);
1021 			free(c->dir[c->diroff]);
1022 		}
1023 		free(c->dir);
1024 		c->dir = NULL;
1025 
1026 		c->type = REQUEST_DONE;
1027 
1028 		event_add(&c->bev->ev_write, NULL);
1029 		break;
1030 
1031 	case REQUEST_CGI:
1032 	case REQUEST_FCGI:
1033 		/*
1034 		 *  Here we depend on on the cgi script or fastcgi
1035 		 *  connection to provide data.
1036 		 */
1037 		break;
1038 
1039 	case REQUEST_DONE:
1040 		if (EVBUFFER_LENGTH(out) == 0)
1041 			client_close(c);
1042 		break;
1043 	}
1044 }
1045 
1046 static void
1047 client_error(struct bufferevent *bev, short error, void *d)
1048 {
1049 	struct client	*c = d;
1050 
1051 	c->type = REQUEST_DONE;
1052 
1053 	if (error & EVBUFFER_TIMEOUT) {
1054 		log_warn(c, "timeout reached, "
1055 		    "forcefully closing the connection");
1056 		if (c->code == 0)
1057 			start_reply(c, BAD_REQUEST, "timeout");
1058 		else
1059 			client_close(c);
1060 		return;
1061 	}
1062 
1063 	if (error & EVBUFFER_EOF) {
1064 		client_close(c);
1065 		return;
1066 	}
1067 
1068 	log_err(c, "unknown bufferevent error: %s", strerror(errno));
1069 	client_close(c);
1070 }
1071 
1072 void
1073 start_reply(struct client *c, int code, const char *meta)
1074 {
1075 	struct evbuffer	*evb = EVBUFFER_OUTPUT(c->bev);
1076 	const char	*lang;
1077 	int		 r, rr;
1078 
1079 	bufferevent_enable(c->bev, EVBUFFER_WRITE);
1080 
1081 	c->code = code;
1082 	c->meta = meta;
1083 
1084 	r = evbuffer_add_printf(evb, "%d %s", code, meta);
1085 	if (r == -1)
1086 		goto err;
1087 
1088 	/* 2 digit status + space + 1024 max reply */
1089 	if (r > 1027)
1090 		goto overflow;
1091 
1092 	if (c->type != REQUEST_CGI &&
1093 	    c->type != REQUEST_FCGI &&
1094 	    !strcmp(meta, "text/gemini") &&
1095 	    (lang = vhost_lang(c->host, c->iri.path)) != NULL) {
1096 		rr = evbuffer_add_printf(evb, ";lang=%s", lang);
1097 		if (rr == -1)
1098 			goto err;
1099 		if (r + rr > 1027)
1100 			goto overflow;
1101 	}
1102 
1103 	bufferevent_write(c->bev, "\r\n", 2);
1104 
1105 	if (!vhost_disable_log(c->host, c->iri.path))
1106 		log_request(c, EVBUFFER_DATA(evb), EVBUFFER_LENGTH(evb));
1107 
1108 	if (code != 20 && IS_INTERNAL_REQUEST(c->type))
1109 		c->type = REQUEST_DONE;
1110 
1111 	return;
1112 
1113 err:
1114 	log_err(c, "evbuffer_add_printf error: no memory");
1115 	evbuffer_drain(evb, EVBUFFER_LENGTH(evb));
1116 	client_close(c);
1117 	return;
1118 
1119 overflow:
1120 	log_warn(c, "reply header overflow");
1121 	evbuffer_drain(evb, EVBUFFER_LENGTH(evb));
1122 	start_reply(c, TEMP_FAILURE, "internal error");
1123 }
1124 
1125 static void
1126 client_close_ev(int fd, short event, void *d)
1127 {
1128 	struct client	*c = d;
1129 
1130 	switch (tls_close(c->ctx)) {
1131 	case TLS_WANT_POLLIN:
1132 		event_once(c->fd, EV_READ, client_close_ev, c, NULL);
1133 		break;
1134 	case TLS_WANT_POLLOUT:
1135 		event_once(c->fd, EV_WRITE, client_close_ev, c, NULL);
1136 		break;
1137 	}
1138 
1139 	connected_clients--;
1140 
1141 	free(c->req);
1142 
1143 	tls_free(c->ctx);
1144 	c->ctx = NULL;
1145 
1146 	free(c->header);
1147 
1148 	if (c->pfd != -1)
1149 		close(c->pfd);
1150 
1151 	if (c->dir != NULL)
1152 		free(c->dir);
1153 
1154 	close(c->fd);
1155 	c->fd = -1;
1156 }
1157 
1158 void
1159 client_close(struct client *c)
1160 {
1161 	/*
1162 	 * We may end up calling client_close in various situations
1163 	 * and for the most unexpected reasons.  Therefore, we need to
1164 	 * ensure that everything is properly released once we reach
1165 	 * this point.
1166 	 */
1167 
1168 	SPLAY_REMOVE(client_tree_id, &clients, c);
1169 
1170 	if (c->cgibev != NULL) {
1171 		bufferevent_disable(c->cgibev, EVBUFFER_READ|EVBUFFER_WRITE);
1172 		bufferevent_free(c->cgibev);
1173 		c->cgibev = NULL;
1174 		close(c->pfd);
1175 		c->pfd = -1;
1176 	}
1177 
1178 	bufferevent_disable(c->bev, EVBUFFER_READ|EVBUFFER_WRITE);
1179 	bufferevent_free(c->bev);
1180 	c->bev = NULL;
1181 
1182 	client_close_ev(c->fd, 0, c);
1183 }
1184 
1185 static void
1186 cgi_read(struct bufferevent *bev, void *d)
1187 {
1188 	struct client	*client = d;
1189 	struct evbuffer	*src = EVBUFFER_INPUT(bev);
1190 	char		*header;
1191 	size_t		 len;
1192 	int		 code;
1193 
1194 	/* intercept the header */
1195 	if (client->code == 0) {
1196 		header = evbuffer_readln(src, &len, EVBUFFER_EOL_CRLF_STRICT);
1197 		if (header == NULL) {
1198 			/* max reply + \r\n */
1199 			if (EVBUFFER_LENGTH(src) > 1026) {
1200 				log_warn(client, "CGI script is trying to "
1201 				    "send a header too long.");
1202 				cgi_error(bev, EVBUFFER_READ, client);
1203 			}
1204 
1205 			/* wait a bit */
1206 			return;
1207 		}
1208 
1209 		if (len < 3 || len > 1029 ||
1210 		    !isdigit(header[0]) ||
1211 		    !isdigit(header[1]) ||
1212 		    !isspace(header[2])) {
1213 			free(header);
1214 			log_warn(client, "CGI script is trying to send a "
1215 			    "malformed header");
1216 			cgi_error(bev, EVBUFFER_READ, client);
1217 			return;
1218 		}
1219 
1220 		client->header = header;
1221 		code = (header[0] - '0') * 10 + (header[1] - '0');
1222 
1223 		if (code < 10 || code >= 70) {
1224 			log_warn(client, "CGI script is trying to send an "
1225 			    "invalid reply code (%d)", code);
1226 			cgi_error(bev, EVBUFFER_READ, client);
1227 			return;
1228 		}
1229 
1230 		start_reply(client, code, header + 3);
1231 
1232 		if (client->code < 20 || client->code > 29) {
1233 			cgi_error(client->cgibev, EVBUFFER_EOF, client);
1234 			return;
1235 		}
1236 	}
1237 
1238 	bufferevent_write_buffer(client->bev, src);
1239 }
1240 
1241 static void
1242 cgi_write(struct bufferevent *bev, void *d)
1243 {
1244 	/*
1245 	 * Never called.  We don't send data to a CGI script.
1246 	 */
1247 	abort();
1248 }
1249 
1250 static void
1251 cgi_error(struct bufferevent *bev, short error, void *d)
1252 {
1253 	struct client	*client = d;
1254 
1255 	if (error & EVBUFFER_ERROR)
1256 		log_err(client, "%s: evbuffer error (%x): %s",
1257 		    __func__, error, strerror(errno));
1258 
1259 	bufferevent_disable(bev, EVBUFFER_READ|EVBUFFER_WRITE);
1260 	bufferevent_free(bev);
1261 	client->cgibev = NULL;
1262 
1263 	close(client->pfd);
1264 	client->pfd = -1;
1265 
1266 	client->type = REQUEST_DONE;
1267 	if (client->code != 0)
1268 		client_write(client->bev, client);
1269 	else
1270 		start_reply(client, CGI_ERROR, "CGI error");
1271 }
1272 
1273 static void
1274 do_accept(int sock, short et, void *d)
1275 {
1276 	struct client *c;
1277 	struct sockaddr_storage addr;
1278 	struct sockaddr *saddr;
1279 	socklen_t len;
1280 	int fd;
1281 
1282 	saddr = (struct sockaddr*)&addr;
1283 	len = sizeof(addr);
1284 	if ((fd = accept(sock, saddr, &len)) == -1) {
1285 		if (errno == EWOULDBLOCK || errno == EAGAIN ||
1286 		    errno == ECONNABORTED)
1287 			return;
1288 		fatal("accept: %s", strerror(errno));
1289 	}
1290 
1291 	mark_nonblock(fd);
1292 
1293 	c = xcalloc(1, sizeof(*c));
1294 	c->id = ++server_client_id;
1295 	c->fd = fd;
1296 	c->pfd = -1;
1297 	c->addr = addr;
1298 
1299 	if (tls_accept_socket(ctx, &c->ctx, fd) == -1) {
1300 		log_warn(c, "failed to accept socket: %s", tls_error(c->ctx));
1301 		close(c->fd);
1302 		free(c);
1303 		return;
1304 	}
1305 
1306 	SPLAY_INSERT(client_tree_id, &clients, c);
1307 	event_once(c->fd, EV_READ|EV_WRITE, handle_handshake, c, NULL);
1308 	connected_clients++;
1309 }
1310 
1311 static struct client *
1312 client_by_id(int id)
1313 {
1314 	struct client *c;
1315 
1316 	if ((c = try_client_by_id(id)) == NULL)
1317 		fatal("in client_by_id: invalid id %d", id);
1318 	return c;
1319 }
1320 
1321 struct client *
1322 try_client_by_id(int id)
1323 {
1324 	struct client find;
1325 
1326 	find.id = id;
1327 	return SPLAY_FIND(client_tree_id, &clients, &find);
1328 }
1329 
1330 static void
1331 handle_imsg_cgi_res(struct imsgbuf *ibuf, struct imsg *imsg, size_t len)
1332 {
1333 	struct client *c;
1334 
1335 	c = client_by_id(imsg->hdr.peerid);
1336 
1337 	if ((c->pfd = imsg->fd) == -1) {
1338 		start_reply(c, TEMP_FAILURE, "internal server error");
1339 		return;
1340 	}
1341 
1342 	c->type = REQUEST_CGI;
1343 
1344 	c->cgibev = bufferevent_new(c->pfd, cgi_read, cgi_write,
1345 	    cgi_error, c);
1346 
1347 	bufferevent_enable(c->cgibev, EV_READ);
1348 }
1349 
1350 static void
1351 handle_imsg_fcgi_fd(struct imsgbuf *ibuf, struct imsg *imsg, size_t len)
1352 {
1353 	struct client	*c;
1354 	int		 id;
1355 
1356 	id = imsg->hdr.peerid;
1357 
1358 	if ((c = try_client_by_id(id)) == NULL) {
1359 		if (imsg->fd != -1)
1360 			close(imsg->fd);
1361 		return;
1362 	}
1363 
1364 	if ((c->pfd = imsg->fd) == -1) {
1365 		start_reply(c, CGI_ERROR, "CGI error");
1366 		return;
1367 	}
1368 
1369 	mark_nonblock(c->pfd);
1370 
1371 	c->cgibev = bufferevent_new(c->pfd, fcgi_read, fcgi_write,
1372 	    fcgi_error, c);
1373 	if (c->cgibev == NULL) {
1374 		start_reply(c, TEMP_FAILURE, "internal server error");
1375 		return;
1376 	}
1377 
1378 	bufferevent_enable(c->cgibev, EV_READ|EV_WRITE);
1379 	fcgi_req(c);
1380 }
1381 
1382 static void
1383 handle_imsg_quit(struct imsgbuf *ibuf, struct imsg *imsg, size_t len)
1384 {
1385 	/*
1386 	 * don't call event_loopbreak since we want to finish to
1387 	 * handle the ongoing connections.
1388 	 */
1389 
1390 	shutting_down = 1;
1391 
1392 	event_del(&e4);
1393 	if (has_ipv6)
1394 		event_del(&e6);
1395 	if (has_siginfo)
1396 		signal_del(&siginfo);
1397 	event_del(&imsgev);
1398 	signal_del(&sigusr2);
1399 }
1400 
1401 static void
1402 handle_dispatch_imsg(int fd, short ev, void *d)
1403 {
1404 	struct imsgbuf *ibuf = d;
1405 	dispatch_imsg(ibuf, handlers, sizeof(handlers));
1406 }
1407 
1408 static void
1409 handle_siginfo(int fd, short ev, void *d)
1410 {
1411 	log_info(NULL, "%d connected clients", connected_clients);
1412 }
1413 
1414 void
1415 loop(struct tls *ctx_, int sock4, int sock6, struct imsgbuf *ibuf)
1416 {
1417 	ctx = ctx_;
1418 
1419 	SPLAY_INIT(&clients);
1420 
1421 	event_init();
1422 
1423 	event_set(&e4, sock4, EV_READ | EV_PERSIST, &do_accept, NULL);
1424 	event_add(&e4, NULL);
1425 
1426 	if (sock6 != -1) {
1427 		has_ipv6 = 1;
1428 		event_set(&e6, sock6, EV_READ | EV_PERSIST, &do_accept, NULL);
1429 		event_add(&e6, NULL);
1430 	}
1431 
1432 	event_set(&imsgev, ibuf->fd, EV_READ | EV_PERSIST, handle_dispatch_imsg, ibuf);
1433 	event_add(&imsgev, NULL);
1434 
1435 #ifdef SIGINFO
1436 	has_siginfo = 1;
1437 	signal_set(&siginfo, SIGINFO, &handle_siginfo, NULL);
1438 	signal_add(&siginfo, NULL);
1439 #endif
1440 	signal_set(&sigusr2, SIGUSR2, &handle_siginfo, NULL);
1441 	signal_add(&sigusr2, NULL);
1442 
1443 	sandbox_server_process();
1444 	event_dispatch();
1445 	_exit(0);
1446 }
1447 
1448 int
1449 client_tree_cmp(struct client *a, struct client *b)
1450 {
1451 	if (a->id == b->id)
1452 		return 0;
1453 	else if (a->id < b->id)
1454 		return -1;
1455 	else
1456 		return +1;
1457 }
1458 
1459 SPLAY_GENERATE(client_tree_id, client, entry, client_tree_cmp)