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 <sys/mman.h>
18 #include <sys/stat.h>
19 
20 #include <netdb.h>
21 
22 #include <assert.h>
23 #include <errno.h>
24 #include <fcntl.h>
25 #include <fnmatch.h>
26 #include <string.h>
27 
28 #include "gmid.h"
29 
30 int connected_clients;
31 
32 int
33 check_path(struct client *c, const char *path, int *fd)
34 {
35 	struct stat sb;
36 	const char *p;
37 
38 	assert(path != NULL);
39 
40 	if (*path == '\0')
41 		p = ".";
42 	else if (*path == '/')
43 		/* in send_dir we add an initial / (to be
44 		 * redirect-friendly), but here we want to skip it */
45 		p = path+1;
46 	else
47 		p = path;
48 
49 	if ((*fd = openat(c->host->dirfd, p, O_RDONLY | O_NOFOLLOW)) == -1) {
50 		return FILE_MISSING;
51 	}
52 
53 	if (fstat(*fd, &sb) == -1) {
54 		LOGN(c, "failed stat for %s: %s", path, strerror(errno));
55 		return FILE_MISSING;
56 	}
57 
58 	if (S_ISDIR(sb.st_mode))
59 		return FILE_DIRECTORY;
60 
61 	if (sb.st_mode & S_IXUSR)
62 		return FILE_EXECUTABLE;
63 
64 	return FILE_EXISTS;
65 }
66 
67 int
68 open_file(struct pollfd *fds, struct client *c)
69 {
70 	switch (check_path(c, c->iri.path, &c->fd)) {
71 	case FILE_EXECUTABLE:
72 		if (c->host->cgi != NULL && starts_with(c->iri.path, c->host->cgi))
73 			return start_cgi(c->iri.path, "", c->iri.query, fds, c);
74 
75 		/* fallthrough */
76 
77 	case FILE_EXISTS:
78 		if ((c->len = filesize(c->fd)) == -1) {
79 			LOGE(c, "failed to get file size for %s", c->iri.path);
80 			close_conn(fds, c);
81 			return 0;
82 		}
83 
84 		if ((c->buf = mmap(NULL, c->len, PROT_READ, MAP_PRIVATE,
85 			    c->fd, 0)) == MAP_FAILED) {
86 			LOGW(c, "mmap: %s: %s", c->iri.path, strerror(errno));
87 			close_conn(fds, c);
88 			return 0;
89 		}
90 		c->i = c->buf;
91 		if (!start_reply(fds, c, SUCCESS, mime(c->host, c->iri.path)))
92 			return 0;
93 		send_file(fds, c);
94 		return 0;
95 
96 	case FILE_DIRECTORY:
97 		close(c->fd);
98 		c->fd = -1;
99 		send_dir(fds, c);
100 		return 0;
101 
102 	case FILE_MISSING:
103 		if (c->host->cgi != NULL && starts_with(c->iri.path, c->host->cgi))
104 			return check_for_cgi(c->iri.path, c->iri.query, fds, c);
105 		goodbye(fds, c, NOT_FOUND, "not found");
106 		return 0;
107 
108 	default:
109 		/* unreachable */
110 		abort();
111 	}
112 }
113 
114 
115 /*
116  * the inverse of this algorithm, i.e. starting from the start of the
117  * path + strlen(cgi), and checking if each component, should be
118  * faster.  But it's tedious to write.  This does the opposite: starts
119  * from the end and strip one component at a time, until either an
120  * executable is found or we emptied the path.
121  */
122 int
123 check_for_cgi(char *path, char *query, struct pollfd *fds, struct client *c)
124 {
125 	char *end;
126 	end = strchr(path, '\0');
127 
128 	/* NB: assume CGI is enabled and path matches cgi */
129 
130 	while (end > path) {
131 		/* go up one level.  UNIX paths are simple and POSIX
132 		 * dirname, with its ambiguities on if the given path
133 		 * is changed or not, gives me headaches. */
134 		while (*end != '/')
135 			end--;
136 		*end = '\0';
137 
138 		switch (check_path(c, path, &c->fd)) {
139 		case FILE_EXECUTABLE:
140 			return start_cgi(path, end+1, query, fds,c);
141 		case FILE_MISSING:
142 			break;
143 		default:
144 			goto err;
145 		}
146 
147 		*end = '/';
148 		end--;
149 	}
150 
151 err:
152 	goodbye(fds, c, NOT_FOUND, "not found");
153 	return 0;
154 }
155 
156 void
157 mark_nonblock(int fd)
158 {
159 	int flags;
160 
161 	if ((flags = fcntl(fd, F_GETFL)) == -1)
162 		fatal("fcntl(F_GETFL): %s", strerror(errno));
163 	if (fcntl(fd, F_SETFL, flags | O_NONBLOCK) == -1)
164 		fatal("fcntl(F_SETFL): %s", strerror(errno));
165 }
166 
167 void
168 handle_handshake(struct pollfd *fds, struct client *c)
169 {
170 	struct vhost *h;
171 	const char *servname;
172 
173 	switch (tls_handshake(c->ctx)) {
174 	case 0:  /* success */
175 	case -1: /* already handshaked */
176 		break;
177 	case TLS_WANT_POLLIN:
178 		fds->events = POLLIN;
179 		return;
180 	case TLS_WANT_POLLOUT:
181 		fds->events = POLLOUT;
182 		return;
183 	default:
184 		/* unreachable */
185 		abort();
186 	}
187 
188 	servname = tls_conn_servername(c->ctx);
189 
190 	for (h = hosts; h->domain != NULL; ++h) {
191 		if (!strcmp(h->domain, "*"))
192 			break;
193 
194 		if (servname != NULL && !fnmatch(h->domain, servname, 0))
195 			break;
196 	}
197 
198 	if (h->domain != NULL) {
199 		c->state = S_OPEN;
200 		c->host = h;
201 		handle_open_conn(fds, c);
202 		return;
203 	}
204 
205 	if (servname != NULL)
206 		strncpy(c->req, servname, sizeof(c->req));
207 	else
208 		strncpy(c->req, "null", sizeof(c->req));
209 
210 	goodbye(fds, c, BAD_REQUEST, "Wrong host or missing SNI");
211 }
212 
213 void
214 handle_open_conn(struct pollfd *fds, struct client *c)
215 {
216 	const char *parse_err = "invalid request";
217 
218 	bzero(c->req, sizeof(c->req));
219 	bzero(&c->iri, sizeof(c->iri));
220 
221 	switch (tls_read(c->ctx, c->req, sizeof(c->req)-1)) {
222 	case -1:
223 		LOGE(c, "tls_read: %s", tls_error(c->ctx));
224 		close_conn(fds, c);
225 		return;
226 
227 	case TLS_WANT_POLLIN:
228 		fds->events = POLLIN;
229 		return;
230 
231 	case TLS_WANT_POLLOUT:
232 		fds->events = POLLOUT;
233 		return;
234 	}
235 
236 	if (!trim_req_iri(c->req) || !parse_iri(c->req, &c->iri, &parse_err)) {
237 		goodbye(fds, c, BAD_REQUEST, parse_err);
238 		return;
239 	}
240 
241 	/* XXX: we should check that the SNI matches the requested host */
242 	if (strcmp(c->iri.schema, "gemini") || c->iri.port_no != conf.port) {
243 		goodbye(fds, c, PROXY_REFUSED, "won't proxy request");
244 		return;
245 	}
246 
247 	open_file(fds, c);
248 }
249 
250 int
251 start_reply(struct pollfd *pfd, struct client *c, int code, const char *meta)
252 {
253 	char buf[1030];		/* status + ' ' + max reply len + \r\n\0 */
254 	size_t len;
255 
256 	c->code = code;
257 	c->meta = meta;
258 	c->state = S_INITIALIZING;
259 
260 	snprintf(buf, sizeof(buf), "%d ", code);
261 	strlcat(buf, meta, sizeof(buf));
262 	if (!strcmp(meta, "text/gemini") && c->host->lang != NULL) {
263 		strlcat(buf, "; lang=", sizeof(buf));
264 		strlcat(buf, c->host->lang, sizeof(buf));
265 	}
266 
267 	len = strlcat(buf, "\r\n", sizeof(buf));
268 	assert(len < sizeof(buf));
269 
270 	switch (tls_write(c->ctx, buf, len)) {
271 	case TLS_WANT_POLLIN:
272 		pfd->events = POLLIN;
273 		return 0;
274 	case TLS_WANT_POLLOUT:
275 		pfd->events = POLLOUT;
276 		return 0;
277 	default:
278 		log_request(c, buf, sizeof(buf));
279 		return 1;
280 	}
281 }
282 
283 int
284 start_cgi(const char *spath, const char *relpath, const char *query,
285     struct pollfd *fds, struct client *c)
286 {
287 	char addr[NI_MAXHOST];
288 	const char *ruser, *cissuer, *chash;
289 	int e;
290 
291 	e = getnameinfo((struct sockaddr*)&c->addr, sizeof(c->addr),
292 	    addr, sizeof(addr),
293 	    NULL, 0,
294 	    NI_NUMERICHOST);
295 	if (e != 0)
296 		goto err;
297 
298 	if (tls_peer_cert_provided(c->ctx)) {
299 		ruser = tls_peer_cert_subject(c->ctx);
300 		cissuer = tls_peer_cert_issuer(c->ctx);
301 		chash = tls_peer_cert_hash(c->ctx);
302 	} else {
303 		ruser = NULL;
304 		cissuer = NULL;
305 		chash = NULL;
306 	}
307 
308 	if (!send_string(exfd, spath)
309 	    || !send_string(exfd, relpath)
310 	    || !send_string(exfd, query)
311 	    || !send_string(exfd, addr)
312 	    || !send_string(exfd, ruser)
313 	    || !send_string(exfd, cissuer)
314 	    || !send_string(exfd, chash)
315 	    || !send_vhost(exfd, c->host))
316 		goto err;
317 
318 	close(c->fd);
319 	if ((c->fd = recv_fd(exfd)) == -1) {
320 		goodbye(fds, c, TEMP_FAILURE, "internal server error");
321 		return 0;
322 	}
323 	c->child = 1;
324 	c->state = S_SENDING;
325 	cgi_poll_on_child(fds, c);
326 	c->code = -1;
327 	/* handle_cgi(fds, c); */
328 	return 0;
329 
330 err:
331 	/* fatal("cannot talk to the executor process: %s", strerror(errno)); */
332 	fatal("cannot talk to the executor process");
333 }
334 
335 void
336 send_file(struct pollfd *fds, struct client *c)
337 {
338 	ssize_t ret, len;
339 
340 	/* ensure the correct state */
341 	c->state = S_SENDING;
342 
343 	len = (c->buf + c->len) - c->i;
344 
345 	while (len > 0) {
346 		switch (ret = tls_write(c->ctx, c->i, len)) {
347 		case -1:
348 			LOGE(c, "tls_write: %s", tls_error(c->ctx));
349 			close_conn(fds, c);
350 			return;
351 
352 		case TLS_WANT_POLLIN:
353 			fds->events = POLLIN;
354 			return;
355 
356 		case TLS_WANT_POLLOUT:
357 			fds->events = POLLOUT;
358 			return;
359 
360 		default:
361 			c->i += ret;
362 			len -= ret;
363 			break;
364 		}
365 	}
366 
367 	close_conn(fds, c);
368 }
369 
370 void
371 send_dir(struct pollfd *fds, struct client *c)
372 {
373 	size_t len;
374 
375 	/* guard against a re-entrant call:
376 	 *
377 	 *	open_file -> send_dir -> open_file -> send_dir
378 	 *
379 	 * this can happen only if:
380 	 *
381 	 *  - user requested a dir, say foo/
382 	 *  - we try to serve foo/index.gmi
383 	 *  - foo/index.gmi is a directory.
384 	 *
385 	 * It's an unlikely case, but can happen.  We then redirect
386 	 * to foo/index.gmi
387 	 */
388 	if (c->iri.path == c->sbuf) {
389 		goodbye(fds, c, TEMP_REDIRECT, c->sbuf);
390 		return;
391 	}
392 
393 	strlcpy(c->sbuf, "/", sizeof(c->sbuf));
394 
395 	len = strlen(c->iri.path);
396 	if (len > 0 && c->iri.path[len-1] != '/') {
397 		/* redirect to url with the trailing / */
398 		strlcat(c->sbuf, c->iri.path, sizeof(c->sbuf));
399 		strlcat(c->sbuf, "/", sizeof(c->sbuf));
400 		goodbye(fds, c, TEMP_REDIRECT, c->sbuf);
401 		return;
402 	}
403 
404 	strlcat(c->sbuf, c->iri.path, sizeof(c->sbuf));
405 
406 	if (!ends_with(c->sbuf, "/"))
407 		strlcat(c->sbuf, "/", sizeof(c->sbuf));
408 
409 	len = strlcat(c->sbuf, "index.gmi", sizeof(c->sbuf));
410 
411 	if (len >= sizeof(c->sbuf)) {
412 		goodbye(fds, c, TEMP_FAILURE, "internal server error");
413 		return;
414 	}
415 
416 	close(c->fd);
417 	c->iri.path = c->sbuf;
418 	open_file(fds, c);
419 }
420 
421 void
422 cgi_poll_on_child(struct pollfd *fds, struct client *c)
423 {
424 	int fd;
425 
426 	if (c->waiting_on_child)
427 		return;
428 	c->waiting_on_child = 1;
429 
430 	fds->events = POLLIN;
431 
432 	fd = fds->fd;
433 	fds->fd = c->fd;
434 	c->fd = fd;
435 }
436 
437 void
438 cgi_poll_on_client(struct pollfd *fds, struct client *c)
439 {
440 	int fd;
441 
442 	if (!c->waiting_on_child)
443 		return;
444 	c->waiting_on_child = 0;
445 
446 	fd = fds->fd;
447 	fds->fd = c->fd;
448 	c->fd = fd;
449 }
450 
451 /* handle the read from the child process.  Return like read(2) */
452 static ssize_t
453 read_from_cgi(struct client *c)
454 {
455 	void	*buf;
456 	size_t	 len;
457 	ssize_t	 r;
458 
459 	/* if we haven't read a whole response line, we want to
460 	 * continue reading. */
461 
462 	if (c->code == -1) {
463 		buf = c->sbuf + c->len;
464 		len = sizeof(c->sbuf) - c->len;
465 	} else {
466 		buf = c->sbuf;
467 		len = sizeof(c->sbuf);
468 	}
469 
470 	r = read(c->fd, buf, len);
471 	if (r == 0 || r == -1)
472 		return r;
473 
474 	c->len += r;
475 	c->off = 0;
476 
477 	if (c->code != -1)
478 		return r;
479 
480 	if (strchr(c->sbuf, '\n') || c->len == sizeof(c->sbuf)) {
481 		c->code = 0;
482 		log_request(c, c->sbuf, c->len);
483 	}
484 
485 	return r;
486 }
487 
488 void
489 handle_cgi(struct pollfd *fds, struct client *c)
490 {
491 	ssize_t r;
492 
493 	/* ensure c->fd is the child and fds->fd the client */
494 	cgi_poll_on_client(fds, c);
495 
496 	while (1) {
497 		if (c->code == -1 || c->len == 0) {
498 			switch (r = read_from_cgi(c)) {
499 			case 0:
500 				goto end;
501 
502 			case -1:
503 				if (errno == EAGAIN || errno == EWOULDBLOCK) {
504 					cgi_poll_on_child(fds, c);
505 					return;
506 				}
507 				goto end;
508 			}
509 		}
510 
511 		if (c->code == -1) {
512 			cgi_poll_on_child(fds, c);
513 			return;
514 		}
515 
516 		while (c->len > 0) {
517 			switch (r = tls_write(c->ctx, c->sbuf + c->off, c->len)) {
518 			case -1:
519 				goto end;
520 
521 			case TLS_WANT_POLLOUT:
522 				fds->events = POLLOUT;
523 				return;
524 
525 			case TLS_WANT_POLLIN:
526 				fds->events = POLLIN;
527 				return;
528 
529 			default:
530                                 c->off += r;
531 				c->len -= r;
532 				break;
533 			}
534 		}
535 	}
536 
537 end:
538 	close_conn(fds, c);
539 }
540 
541 void
542 close_conn(struct pollfd *pfd, struct client *c)
543 {
544 	c->state = S_CLOSING;
545 
546 	switch (tls_close(c->ctx)) {
547 	case TLS_WANT_POLLIN:
548 		pfd->events = POLLIN;
549 		return;
550 	case TLS_WANT_POLLOUT:
551 		pfd->events = POLLOUT;
552 		return;
553 	}
554 
555 	connected_clients--;
556 
557 	tls_free(c->ctx);
558 	c->ctx = NULL;
559 
560 	if (c->buf != MAP_FAILED)
561 		munmap(c->buf, c->len);
562 
563 	if (c->fd != -1)
564 		close(c->fd);
565 
566 	close(pfd->fd);
567 	pfd->fd = -1;
568 }
569 
570 void
571 goodbye(struct pollfd *fds, struct client *c, int code, const char *meta)
572 {
573 	if (!start_reply(fds, c, code, meta))
574 		return;
575 	close_conn(fds, c);
576 }
577 
578 void
579 do_accept(int sock, struct tls *ctx, struct pollfd *fds, struct client *clients)
580 {
581 	int i, fd;
582 	struct sockaddr_storage addr;
583 	socklen_t len;
584 
585 	len = sizeof(addr);
586 	if ((fd = accept(sock, (struct sockaddr*)&addr, &len)) == -1) {
587 		if (errno == EWOULDBLOCK)
588 			return;
589 		fatal("accept: %s", strerror(errno));
590 	}
591 
592 	mark_nonblock(fd);
593 
594 	for (i = 0; i < MAX_USERS; ++i) {
595 		if (fds[i].fd == -1) {
596 			bzero(&clients[i], sizeof(struct client));
597 			if (tls_accept_socket(ctx, &clients[i].ctx, fd) == -1)
598 				break; /* goodbye fd! */
599 
600 			fds[i].fd = fd;
601 			fds[i].events = POLLIN;
602 
603 			clients[i].state = S_HANDSHAKE;
604 			clients[i].fd = -1;
605 			clients[i].child = 0;
606 			clients[i].waiting_on_child = 0;
607 			clients[i].buf = MAP_FAILED;
608 			clients[i].addr = addr;
609 
610 			connected_clients++;
611 			return;
612 		}
613 	}
614 
615 	close(fd);
616 }
617 
618 void
619 handle(struct pollfd *fds, struct client *client)
620 {
621 	switch (client->state) {
622 	case S_HANDSHAKE:
623 		handle_handshake(fds, client);
624 		break;
625 
626 	case S_OPEN:
627                 handle_open_conn(fds, client);
628 		break;
629 
630 	case S_INITIALIZING:
631 		if (!start_reply(fds, client, client->code, client->meta))
632 			return;
633 
634 		if (client->code != SUCCESS) {
635 			/* we don't need a body */
636 			close_conn(fds, client);
637 			return;
638 		}
639 
640 		client->state = S_SENDING;
641 
642 		/* fallthrough */
643 
644 	case S_SENDING:
645 		if (client->child)
646 			handle_cgi(fds, client);
647 		else
648 			send_file(fds, client);
649 		break;
650 
651 	case S_CLOSING:
652 		close_conn(fds, client);
653 		break;
654 
655 	default:
656 		/* unreachable */
657 		abort();
658 	}
659 }
660 
661 void
662 loop(struct tls *ctx, int sock4, int sock6)
663 {
664 	int i;
665 	struct client clients[MAX_USERS];
666 	struct pollfd fds[MAX_USERS];
667 
668 	connected_clients = 0;
669 
670 	for (i = 0; i < MAX_USERS; ++i) {
671 		fds[i].fd = -1;
672 		fds[i].events = POLLIN;
673 		bzero(&clients[i], sizeof(struct client));
674 	}
675 
676 	fds[0].fd = sock4;
677 	fds[1].fd = sock6;
678 
679 	for (;;) {
680 		if (poll(fds, MAX_USERS, INFTIM) == -1) {
681 			if (errno == EINTR) {
682                                 fprintf(stderr, "connected clients: %d\n",
683 				    connected_clients);
684 				continue;
685 			}
686 			fatal("poll: %s", strerror(errno));
687 		}
688 
689 		for (i = 0; i < MAX_USERS; i++) {
690 			if (fds[i].revents == 0)
691 				continue;
692 
693 			if (fds[i].revents & (POLLERR|POLLNVAL))
694 				fatal("bad fd %d: %s", fds[i].fd,
695 				    strerror(errno));
696 
697 			if (fds[i].revents & POLLHUP) {
698 				/* fds[i] may be the fd of the stdin
699 				 * of a cgi script that has exited. */
700 				if (!clients[i].waiting_on_child) {
701 					close_conn(&fds[i], &clients[i]);
702 					continue;
703 				}
704 			}
705 
706 			if (fds[i].fd == sock4)
707 				do_accept(sock4, ctx, fds, clients);
708 			else if (fds[i].fd == sock6)
709 				do_accept(sock6, ctx, fds, clients);
710 			else
711 				handle(&fds[i], &clients[i]);
712 		}
713 	}
714 }