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