xref: /netbsd/lib/librumpclient/rumpclient.c (revision 47cc1f20)
1 /*      $NetBSD: rumpclient.c,v 1.70 2023/07/28 18:19:00 christos Exp $	*/
2 
3 /*
4  * Copyright (c) 2010, 2011 Antti Kantee.  All Rights Reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS
16  * OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
17  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
18  * DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
21  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 /*
29  * Client side routines for rump syscall proxy.
30  */
31 
32 #include <rump/rumpuser_port.h>
33 
34 /*
35  * We use kqueue on the BSDs, poll elsewhere.  We
36  * want to use kqueue because it will give us the ability to get signal
37  * notifications but defer their handling to a stage where we do not
38  * hold the communication lock.  Taking a signal while holding on to
39  * that lock may cause a deadlock.  Therefore, block signals throughout
40  * the RPC when using poll.  On Linux, we use signalfd in the same role
41  * as kqueue on NetBSD to be able to take signals while waiting for a
42  * response from the server.
43  */
44 
45 #if defined(__NetBSD__) || defined(__FreeBSD__) || \
46     defined(__DragonFly__) || defined(__OpenBSD__)
47 #define USE_KQUEUE
48 #endif
49 #if defined(__linux__)
50 #define USE_SIGNALFD
51 #endif
52 
53 __RCSID("$NetBSD: rumpclient.c,v 1.70 2023/07/28 18:19:00 christos Exp $");
54 
55 #include <sys/param.h>
56 #include <sys/mman.h>
57 #include <sys/socket.h>
58 #include <sys/time.h>
59 
60 #ifdef USE_KQUEUE
61 #include <sys/event.h>
62 #endif
63 
64 #include <arpa/inet.h>
65 #include <netinet/in.h>
66 #include <netinet/tcp.h>
67 
68 #include <assert.h>
69 #include <dlfcn.h>
70 #include <errno.h>
71 #include <fcntl.h>
72 #include <poll.h>
73 #include <pthread.h>
74 #include <signal.h>
75 #include <stdarg.h>
76 #include <stdbool.h>
77 #include <stdio.h>
78 #include <stdlib.h>
79 #include <string.h>
80 #include <unistd.h>
81 
82 #include <rump/rumpclient.h>
83 
84 #define HOSTOPS
85 int	(*host_socket)(int, int, int);
86 int	(*host_close)(int);
87 int	(*host_connect)(int, const struct sockaddr *, socklen_t);
88 int	(*host_fcntl)(int, int, ...);
89 int	(*host_poll)(struct pollfd *, nfds_t, int);
90 ssize_t	(*host_read)(int, void *, size_t);
91 ssize_t (*host_sendmsg)(int, const struct msghdr *, int);
92 int	(*host_setsockopt)(int, int, int, const void *, socklen_t);
93 int	(*host_dup)(int);
94 
95 #ifdef USE_KQUEUE
96 int	(*host_kqueue)(void);
97 #ifdef __NetBSD__
98 int	(*host_kevent)(int, const struct kevent *, size_t,
99 		       struct kevent *, size_t, const struct timespec *);
100 #else
101 int	(*host_kevent)(int, const struct kevent *, int,
102 		       struct kevent *, int, const struct timespec *);
103 #endif
104 #endif
105 
106 #ifdef USE_SIGNALFD
107 #include <sys/signalfd.h>
108 
109 int	(*host_signalfd)(int, const sigset_t *, int);
110 #endif
111 
112 int	(*host_execve)(const char *, char *const[], char *const[]);
113 
114 #include "sp_common.c"
115 #include "rumpuser_sigtrans.c"
116 
117 static struct spclient clispc = {
118 	.spc_fd = -1,
119 };
120 
121 static int holyfd = -1;
122 static sigset_t fullset;
123 
124 static int doconnect(void);
125 static int handshake_req(struct spclient *, int, void *, int, bool);
126 
127 /*
128  * Default: don't retry.  Most clients can't handle it
129  * (consider e.g. fds suddenly going missing).
130  */
131 static time_t retrytimo = 0;
132 
133 /* always defined to nothingness for now */
134 #define ERRLOG(a)
135 
136 static int
send_with_recon(struct spclient * spc,struct iovec * iov,size_t iovlen)137 send_with_recon(struct spclient *spc, struct iovec *iov, size_t iovlen)
138 {
139 	struct timeval starttime, curtime;
140 	time_t prevreconmsg;
141 	unsigned reconretries;
142 	int rv;
143 
144 	for (prevreconmsg = 0, reconretries = 0;;) {
145 		rv = dosend(spc, iov, iovlen);
146 		if (__predict_false(rv == ENOTCONN || rv == EBADF)) {
147 			/* no persistent connections */
148 			if (retrytimo == 0) {
149 				rv = ENOTCONN;
150 				break;
151 			}
152 			if (retrytimo == RUMPCLIENT_RETRYCONN_DIE)
153 				_exit(1);
154 
155 			if (!prevreconmsg) {
156 				prevreconmsg = time(NULL);
157 				gettimeofday(&starttime, NULL);
158 			}
159 			if (reconretries == 1) {
160 				if (retrytimo == RUMPCLIENT_RETRYCONN_ONCE) {
161 					rv = ENOTCONN;
162 					break;
163 				}
164 				fprintf(stderr, "rump_sp: connection to "
165 				    "kernel lost, trying to reconnect ...\n");
166 			} else if (time(NULL) - prevreconmsg > 120) {
167 				fprintf(stderr, "rump_sp: still trying to "
168 				    "reconnect ...\n");
169 				prevreconmsg = time(NULL);
170 			}
171 
172 			/* check that we aren't over the limit */
173 			if (retrytimo > 0) {
174 				time_t tdiff;
175 
176 				gettimeofday(&curtime, NULL);
177 				tdiff = curtime.tv_sec - starttime.tv_sec;
178 				if (starttime.tv_usec > curtime.tv_usec)
179 					tdiff--;
180 				if (tdiff >= retrytimo) {
181 					fprintf(stderr, "rump_sp: reconnect "
182 					    "failed, %lld second timeout\n",
183 					    (long long)retrytimo);
184 					return ENOTCONN;
185 				}
186 			}
187 
188 			/* adhoc backoff timer */
189 			if (reconretries < 10) {
190 				usleep(100000 * reconretries);
191 			} else {
192 				sleep(MIN(10, reconretries-9));
193 			}
194 			reconretries++;
195 
196 			if ((rv = doconnect()) != 0)
197 				continue;
198 			if ((rv = handshake_req(&clispc, HANDSHAKE_GUEST,
199 			    NULL, 0, true)) != 0)
200 				continue;
201 
202 			/*
203 			 * ok, reconnect successful.  we need to return to
204 			 * the upper layer to get the entire PDU resent.
205 			 */
206 			if (reconretries != 1)
207 				fprintf(stderr, "rump_sp: reconnected!\n");
208 			rv = EAGAIN;
209 			break;
210 		} else {
211 			_DIAGASSERT(errno != EAGAIN);
212 			break;
213 		}
214 	}
215 
216 	return rv;
217 }
218 
219 static int
cliwaitresp(struct spclient * spc,struct respwait * rw,sigset_t * mask,bool keeplock)220 cliwaitresp(struct spclient *spc, struct respwait *rw, sigset_t *mask,
221 	bool keeplock)
222 {
223 	uint64_t mygen;
224 	bool imalive = true;
225 
226 	pthread_mutex_lock(&spc->spc_mtx);
227 	if (!keeplock)
228 		sendunlockl(spc);
229 	mygen = spc->spc_generation;
230 
231 	rw->rw_error = 0;
232 	while (!rw->rw_done && rw->rw_error == 0) {
233 		if (__predict_false(spc->spc_generation != mygen || !imalive))
234 			break;
235 
236 		/* are we free to receive? */
237 		if (spc->spc_istatus == SPCSTATUS_FREE) {
238 			int gotresp, dosig, rv;
239 
240 			spc->spc_istatus = SPCSTATUS_BUSY;
241 			pthread_mutex_unlock(&spc->spc_mtx);
242 
243 			dosig = 0;
244 			for (gotresp = 0; !gotresp; ) {
245 #ifdef USE_KQUEUE
246 				struct kevent kev[8];
247 				int i;
248 
249 				/*
250 				 * typically we don't have a frame waiting
251 				 * when we come in here, so call kevent now
252 				 */
253 				rv = host_kevent(holyfd, NULL, 0,
254 				    kev, __arraycount(kev), NULL);
255 
256 				if (__predict_false(rv == -1)) {
257 					goto activity;
258 				}
259 
260 				/*
261 				 * XXX: don't know how this can happen
262 				 * (timeout cannot expire since there
263 				 * isn't one), but it does happen.
264 				 * treat it as an expectional condition
265 				 * and go through tryread to determine
266 				 * alive status.
267 				 */
268 				if (__predict_false(rv == 0))
269 					goto activity;
270 
271 				for (i = 0; i < rv; i++) {
272 					if (kev[i].filter == EVFILT_SIGNAL)
273 						dosig++;
274 				}
275 				if (dosig)
276 					goto cleanup;
277 
278 				/*
279 				 * ok, activity.  try to read a frame to
280 				 * determine what happens next.
281 				 */
282  activity:
283 #else /* !USE_KQUEUE */
284 				struct pollfd pfd[2];
285 
286 				pfd[0].fd = clispc.spc_fd;
287 				pfd[0].events = POLLIN;
288 				pfd[1].fd = holyfd;
289 				pfd[1].events = POLLIN;
290 
291 				rv = host_poll(pfd, 2, -1);
292 				if (rv >= 1 && pfd[1].revents & POLLIN) {
293 					dosig = 1;
294 					goto cleanup;
295 				}
296 #endif /* !USE_KQUEUE */
297 
298 				switch (readframe(spc)) {
299 				case 0:
300 					continue;
301 				case -1:
302 					imalive = false;
303 					goto cleanup;
304 				default:
305 					/* case 1 */
306 					break;
307 				}
308 
309 				switch (spc->spc_hdr.rsp_class) {
310 				case RUMPSP_RESP:
311 				case RUMPSP_ERROR:
312 					kickwaiter(spc);
313 					gotresp = spc->spc_hdr.rsp_reqno ==
314 					    rw->rw_reqno;
315 					break;
316 				case RUMPSP_REQ:
317 					handlereq(spc);
318 					break;
319 				default:
320 					/* panic */
321 					break;
322 				}
323 			}
324 
325  cleanup:
326 			pthread_mutex_lock(&spc->spc_mtx);
327 			if (spc->spc_istatus == SPCSTATUS_WANTED)
328 				kickall(spc);
329 			spc->spc_istatus = SPCSTATUS_FREE;
330 
331 			/* take one for the team */
332 			if (dosig) {
333 				pthread_mutex_unlock(&spc->spc_mtx);
334 				pthread_sigmask(SIG_SETMASK, mask, NULL);
335 				pthread_sigmask(SIG_SETMASK, &fullset, NULL);
336 				pthread_mutex_lock(&spc->spc_mtx);
337 			}
338 		} else {
339 			spc->spc_istatus = SPCSTATUS_WANTED;
340 			pthread_cond_wait(&rw->rw_cv, &spc->spc_mtx);
341 		}
342 	}
343 	TAILQ_REMOVE(&spc->spc_respwait, rw, rw_entries);
344 	pthread_mutex_unlock(&spc->spc_mtx);
345 	pthread_cond_destroy(&rw->rw_cv);
346 
347 	if (spc->spc_generation != mygen || !imalive) {
348 		return ENOTCONN;
349 	}
350 	return rw->rw_error;
351 }
352 
353 static int
syscall_req(struct spclient * spc,sigset_t * omask,int sysnum,const void * data,size_t dlen,void ** resp)354 syscall_req(struct spclient *spc, sigset_t *omask, int sysnum,
355 	const void *data, size_t dlen, void **resp)
356 {
357 	struct rsp_hdr rhdr;
358 	struct respwait rw;
359 	struct iovec iov[2];
360 	int rv;
361 
362 	rhdr.rsp_len = sizeof(rhdr) + dlen;
363 	rhdr.rsp_class = RUMPSP_REQ;
364 	rhdr.rsp_type = RUMPSP_SYSCALL;
365 	rhdr.rsp_sysnum = sysnum;
366 
367 	IOVPUT(iov[0], rhdr);
368 	IOVPUT_WITHSIZE(iov[1], __UNCONST(data), dlen);
369 
370 	do {
371 		putwait(spc, &rw, &rhdr);
372 		if ((rv = send_with_recon(spc, iov, __arraycount(iov))) != 0) {
373 			unputwait(spc, &rw);
374 			continue;
375 		}
376 
377 		rv = cliwaitresp(spc, &rw, omask, false);
378 		if (rv == ENOTCONN)
379 			rv = EAGAIN;
380 	} while (rv == EAGAIN);
381 
382 	*resp = rw.rw_data;
383 	return rv;
384 }
385 
386 static int
handshake_req(struct spclient * spc,int type,void * data,int cancel,bool haslock)387 handshake_req(struct spclient *spc, int type, void *data,
388 	int cancel, bool haslock)
389 {
390 	struct handshake_fork rf;
391 	const char *myprogname = NULL; /* XXXgcc */
392 	struct rsp_hdr rhdr;
393 	struct respwait rw;
394 	sigset_t omask;
395 	size_t bonus;
396 	struct iovec iov[2];
397 	int rv;
398 
399 	if (type == HANDSHAKE_FORK) {
400 		bonus = sizeof(rf);
401 	} else {
402 #ifdef __NetBSD__
403 		/* would procfs work on NetBSD too? */
404 		myprogname = getprogname();
405 #else
406 		int fd = open("/proc/self/comm", O_RDONLY);
407 		if (fd == -1) {
408 			myprogname = "???";
409 		} else {
410 			static char commname[128];
411 
412 			memset(commname, 0, sizeof(commname));
413 			if (read(fd, commname, sizeof(commname)) > 0) {
414 				char *n;
415 
416 				n = strrchr(commname, '\n');
417 				if (n)
418 					*n = '\0';
419 				myprogname = commname;
420 			} else {
421 				myprogname = "???";
422 			}
423 			close(fd);
424 		}
425 #endif
426 		bonus = strlen(myprogname)+1;
427 	}
428 
429 	/* performs server handshake */
430 	rhdr.rsp_len = sizeof(rhdr) + bonus;
431 	rhdr.rsp_class = RUMPSP_REQ;
432 	rhdr.rsp_type = RUMPSP_HANDSHAKE;
433 	rhdr.rsp_handshake = type;
434 
435 	IOVPUT(iov[0], rhdr);
436 
437 	pthread_sigmask(SIG_SETMASK, &fullset, &omask);
438 	if (haslock)
439 		putwait_locked(spc, &rw, &rhdr);
440 	else
441 		putwait(spc, &rw, &rhdr);
442 	if (type == HANDSHAKE_FORK) {
443 		memcpy(rf.rf_auth, data, sizeof(rf.rf_auth)); /* uh, why? */
444 		rf.rf_cancel = cancel;
445 		IOVPUT(iov[1], rf);
446 	} else {
447 		IOVPUT_WITHSIZE(iov[1], __UNCONST(myprogname), bonus);
448 	}
449 	rv = send_with_recon(spc, iov, __arraycount(iov));
450 	if (rv || cancel) {
451 		if (haslock)
452 			unputwait_locked(spc, &rw);
453 		else
454 			unputwait(spc, &rw);
455 		if (cancel) {
456 			goto out;
457 		}
458 	} else {
459 		rv = cliwaitresp(spc, &rw, &omask, haslock);
460 	}
461 	if (rv)
462 		goto out;
463 
464 	rv = *(int *)rw.rw_data;
465 	free(rw.rw_data);
466 
467  out:
468 	pthread_sigmask(SIG_SETMASK, &omask, NULL);
469 	return rv;
470 }
471 
472 static int
prefork_req(struct spclient * spc,sigset_t * omask,void ** resp)473 prefork_req(struct spclient *spc, sigset_t *omask, void **resp)
474 {
475 	struct rsp_hdr rhdr;
476 	struct respwait rw;
477 	struct iovec iov[1];
478 	int rv;
479 
480 	rhdr.rsp_len = sizeof(rhdr);
481 	rhdr.rsp_class = RUMPSP_REQ;
482 	rhdr.rsp_type = RUMPSP_PREFORK;
483 	rhdr.rsp_error = 0;
484 
485 	IOVPUT(iov[0], rhdr);
486 
487 	do {
488 		putwait(spc, &rw, &rhdr);
489 		rv = send_with_recon(spc, iov, __arraycount(iov));
490 		if (rv != 0) {
491 			unputwait(spc, &rw);
492 			continue;
493 		}
494 
495 		rv = cliwaitresp(spc, &rw, omask, false);
496 		if (rv == ENOTCONN)
497 			rv = EAGAIN;
498 	} while (rv == EAGAIN);
499 
500 	*resp = rw.rw_data;
501 	return rv;
502 }
503 
504 /*
505  * prevent response code from deadlocking with reconnect code
506  */
507 static int
resp_sendlock(struct spclient * spc)508 resp_sendlock(struct spclient *spc)
509 {
510 	int rv = 0;
511 
512 	pthread_mutex_lock(&spc->spc_mtx);
513 	while (spc->spc_ostatus != SPCSTATUS_FREE) {
514 		if (__predict_false(spc->spc_reconnecting)) {
515 			rv = EBUSY;
516 			goto out;
517 		}
518 		spc->spc_ostatus = SPCSTATUS_WANTED;
519 		pthread_cond_wait(&spc->spc_cv, &spc->spc_mtx);
520 	}
521 	spc->spc_ostatus = SPCSTATUS_BUSY;
522 
523  out:
524 	pthread_mutex_unlock(&spc->spc_mtx);
525 	return rv;
526 }
527 
528 static void
send_copyin_resp(struct spclient * spc,uint64_t reqno,void * data,size_t dlen,int wantstr)529 send_copyin_resp(struct spclient *spc, uint64_t reqno, void *data, size_t dlen,
530 	int wantstr)
531 {
532 	struct rsp_hdr rhdr;
533 	struct iovec iov[2];
534 
535 	if (wantstr)
536 		dlen = MIN(dlen, strlen(data)+1);
537 
538 	rhdr.rsp_len = sizeof(rhdr) + dlen;
539 	rhdr.rsp_reqno = reqno;
540 	rhdr.rsp_class = RUMPSP_RESP;
541 	rhdr.rsp_type = RUMPSP_COPYIN;
542 	rhdr.rsp_sysnum = 0;
543 
544 	IOVPUT(iov[0], rhdr);
545 	IOVPUT_WITHSIZE(iov[1], data, dlen);
546 
547 	if (resp_sendlock(spc) != 0)
548 		return;
549 	(void)SENDIOV(spc, iov);
550 	sendunlock(spc);
551 }
552 
553 static void
send_anonmmap_resp(struct spclient * spc,uint64_t reqno,void * addr)554 send_anonmmap_resp(struct spclient *spc, uint64_t reqno, void *addr)
555 {
556 	struct rsp_hdr rhdr;
557 	struct iovec iov[2];
558 
559 	rhdr.rsp_len = sizeof(rhdr) + sizeof(addr);
560 	rhdr.rsp_reqno = reqno;
561 	rhdr.rsp_class = RUMPSP_RESP;
562 	rhdr.rsp_type = RUMPSP_ANONMMAP;
563 	rhdr.rsp_sysnum = 0;
564 
565 	IOVPUT(iov[0], rhdr);
566 	IOVPUT(iov[1], addr);
567 
568 	if (resp_sendlock(spc) != 0)
569 		return;
570 	(void)SENDIOV(spc, iov);
571 	sendunlock(spc);
572 }
573 
574 int
rumpclient_syscall(int sysnum,const void * data,size_t dlen,register_t * retval)575 rumpclient_syscall(int sysnum, const void *data, size_t dlen,
576 	register_t *retval)
577 {
578 	struct rsp_sysresp *resp;
579 	sigset_t omask;
580 	void *rdata;
581 	int rv;
582 
583 	pthread_sigmask(SIG_SETMASK, &fullset, &omask);
584 
585 	DPRINTF(("rumpsp syscall_req: syscall %d with %p/%zu\n",
586 	    sysnum, data, dlen));
587 
588 	rv = syscall_req(&clispc, &omask, sysnum, data, dlen, &rdata);
589 	if (rv)
590 		goto out;
591 
592 	resp = rdata;
593 	DPRINTF(("rumpsp syscall_resp: syscall %d error %d, rv: %"
594 	    PRIxREGISTER"/%"PRIxREGISTER"\n",
595 	    sysnum, rv, resp->rsys_retval[0], resp->rsys_retval[1]));
596 
597 	memcpy(retval, &resp->rsys_retval, sizeof(resp->rsys_retval));
598 	rv = resp->rsys_error;
599 	free(rdata);
600 
601  out:
602 	pthread_sigmask(SIG_SETMASK, &omask, NULL);
603 	return rv;
604 }
605 
606 static void
handlereq(struct spclient * spc)607 handlereq(struct spclient *spc)
608 {
609 	struct rsp_copydata *copydata;
610 	struct rsp_hdr *rhdr = &spc->spc_hdr;
611 	void *mapaddr;
612 	size_t maplen;
613 	int reqtype = spc->spc_hdr.rsp_type;
614 	int sig;
615 
616 	switch (reqtype) {
617 	case RUMPSP_COPYIN:
618 	case RUMPSP_COPYINSTR:
619 		/*LINTED*/
620 		copydata = (struct rsp_copydata *)spc->spc_buf;
621 		DPRINTF(("rump_sp handlereq: copyin request: %p/%zu\n",
622 		    copydata->rcp_addr, copydata->rcp_len));
623 		send_copyin_resp(spc, spc->spc_hdr.rsp_reqno,
624 		    copydata->rcp_addr, copydata->rcp_len,
625 		    reqtype == RUMPSP_COPYINSTR);
626 		break;
627 	case RUMPSP_COPYOUT:
628 	case RUMPSP_COPYOUTSTR:
629 		/*LINTED*/
630 		copydata = (struct rsp_copydata *)spc->spc_buf;
631 		DPRINTF(("rump_sp handlereq: copyout request: %p/%zu\n",
632 		    copydata->rcp_addr, copydata->rcp_len));
633 		/*LINTED*/
634 		memcpy(copydata->rcp_addr, copydata->rcp_data,
635 		    copydata->rcp_len);
636 		break;
637 	case RUMPSP_ANONMMAP:
638 		/*LINTED*/
639 		maplen = *(size_t *)spc->spc_buf;
640 		mapaddr = mmap(NULL, maplen, PROT_READ|PROT_WRITE,
641 		    MAP_ANON|MAP_PRIVATE, -1, 0);
642 		if (mapaddr == MAP_FAILED)
643 			mapaddr = NULL;
644 		DPRINTF(("rump_sp handlereq: anonmmap: %p\n", mapaddr));
645 		send_anonmmap_resp(spc, spc->spc_hdr.rsp_reqno, mapaddr);
646 		break;
647 	case RUMPSP_RAISE:
648 		sig = rumpuser__sig_rump2host(rhdr->rsp_signo);
649 		DPRINTF(("rump_sp handlereq: raise sig %d\n", sig));
650 		raise(sig);
651 		/*
652 		 * We most likely have signals blocked, but the signal
653 		 * will be handled soon enough when we return.
654 		 */
655 		break;
656 	default:
657 		printf("PANIC: INVALID TYPE %d\n", reqtype);
658 		abort();
659 		break;
660 	}
661 
662 	spcfreebuf(spc);
663 }
664 
665 static unsigned ptab_idx;
666 static struct sockaddr *serv_sa;
667 
668 /* dup until we get a "good" fd which does not collide with stdio */
669 static int
dupgood(int myfd,int mustchange)670 dupgood(int myfd, int mustchange)
671 {
672 	int ofds[4];
673 	int sverrno;
674 	unsigned int i;
675 
676 	for (i = 0; (myfd <= 2 || mustchange) && myfd != -1; i++) {
677 		assert(i < __arraycount(ofds));
678 		ofds[i] = myfd;
679 		myfd = host_dup(myfd);
680 		if (mustchange) {
681 			i--; /* prevent closing old fd */
682 			mustchange = 0;
683 		}
684 	}
685 
686 	sverrno = 0;
687 	if (myfd == -1 && i > 0)
688 		sverrno = errno;
689 
690 	while (i-- > 0) {
691 		host_close(ofds[i]);
692 	}
693 
694 	if (sverrno)
695 		errno = sverrno;
696 
697 	return myfd;
698 }
699 
700 #if defined(USE_KQUEUE)
701 
702 static int
makeholyfd(void)703 makeholyfd(void)
704 {
705 	struct kevent kev[NSIG+1];
706 	int i, fd;
707 
708 	/* setup kqueue, we want all signals and the fd */
709 	if ((fd = dupgood(host_kqueue(), 0)) == -1) {
710 		ERRLOG(("rump_sp: cannot setup kqueue"));
711 		return -1;
712 	}
713 
714 	for (i = 0; i < NSIG; i++) {
715 		EV_SET(&kev[i], i+1, EVFILT_SIGNAL, EV_ADD|EV_ENABLE, 0, 0, 0);
716 	}
717 	EV_SET(&kev[NSIG], clispc.spc_fd,
718 	    EVFILT_READ, EV_ADD|EV_ENABLE, 0, 0, 0);
719 	if (host_kevent(fd, kev, NSIG+1, NULL, 0, NULL) == -1) {
720 		ERRLOG(("rump_sp: kevent() failed"));
721 		host_close(fd);
722 		return -1;
723 	}
724 
725 	return fd;
726 }
727 
728 #elif defined(USE_SIGNALFD) /* !USE_KQUEUE */
729 
730 static int
makeholyfd(void)731 makeholyfd(void)
732 {
733 
734 	return host_signalfd(-1, &fullset, 0);
735 }
736 
737 #else /* !USE_KQUEUE && !USE_SIGNALFD */
738 
739 static int
makeholyfd(void)740 makeholyfd(void)
741 {
742 
743 	return -1;
744 }
745 
746 #endif
747 
748 static int
doconnect(void)749 doconnect(void)
750 {
751 	struct respwait rw;
752 	struct rsp_hdr rhdr;
753 	char banner[MAXBANNER];
754 	int s, error, flags;
755 	ssize_t n;
756 
757 	if (holyfd != -1)
758 		host_close(holyfd);
759 	holyfd = -1;
760 	s = -1;
761 
762 	if (clispc.spc_fd != -1)
763 		host_close(clispc.spc_fd);
764 	clispc.spc_fd = -1;
765 
766 	/*
767 	 * for reconnect, gate everyone out of the receiver code
768 	 */
769 	putwait_locked(&clispc, &rw, &rhdr);
770 
771 	pthread_mutex_lock(&clispc.spc_mtx);
772 	clispc.spc_reconnecting = 1;
773 	pthread_cond_broadcast(&clispc.spc_cv);
774 	clispc.spc_generation++;
775 	while (clispc.spc_istatus != SPCSTATUS_FREE) {
776 		clispc.spc_istatus = SPCSTATUS_WANTED;
777 		pthread_cond_wait(&rw.rw_cv, &clispc.spc_mtx);
778 	}
779 	kickall(&clispc);
780 
781 	/*
782 	 * we can release it already since we hold the
783 	 * send lock during reconnect
784 	 * XXX: assert it
785 	 */
786 	clispc.spc_istatus = SPCSTATUS_FREE;
787 	pthread_mutex_unlock(&clispc.spc_mtx);
788 	unputwait_locked(&clispc, &rw);
789 
790 	free(clispc.spc_buf);
791 	clispc.spc_off = 0;
792 
793 	s = dupgood(host_socket(parsetab[ptab_idx].domain, SOCK_STREAM, 0), 0);
794 	if (s == -1)
795 		return -1;
796 
797 	while (host_connect(s, serv_sa, parsetab[ptab_idx].slen) == -1) {
798 		if (errno == EINTR)
799 			continue;
800 		ERRLOG(("rump_sp: client connect failed: %s\n",
801 		    strerror(errno)));
802 		return -1;
803 	}
804 
805 	if ((error = parsetab[ptab_idx].connhook(s)) != 0) {
806 		ERRLOG(("rump_sp: connect hook failed\n"));
807 		return -1;
808 	}
809 
810 	if ((n = host_read(s, banner, sizeof(banner)-1)) <= 0) {
811 		ERRLOG(("rump_sp: failed to read banner\n"));
812 		return -1;
813 	}
814 
815 	if (banner[n-1] != '\n') {
816 		ERRLOG(("rump_sp: invalid banner\n"));
817 		return -1;
818 	}
819 	banner[n] = '\0';
820 	/* XXX parse the banner some day */
821 
822 	flags = host_fcntl(s, F_GETFL, 0);
823 	if (host_fcntl(s, F_SETFL, flags | O_NONBLOCK) == -1) {
824 		ERRLOG(("rump_sp: socket fd NONBLOCK: %s\n", strerror(errno)));
825 		return -1;
826 	}
827 	clispc.spc_fd = s;
828 	clispc.spc_state = SPCSTATE_RUNNING;
829 	clispc.spc_reconnecting = 0;
830 	holyfd = makeholyfd();
831 
832 	return 0;
833 }
834 
835 static int
doinit(void)836 doinit(void)
837 {
838 
839 	TAILQ_INIT(&clispc.spc_respwait);
840 	pthread_mutex_init(&clispc.spc_mtx, NULL);
841 	pthread_cond_init(&clispc.spc_cv, NULL);
842 
843 	return 0;
844 }
845 
846 #ifdef RTLD_NEXT
847 void *rumpclient__dlsym(void *, const char *);
848 void *
rumpclient__dlsym(void * handle,const char * symbol)849 rumpclient__dlsym(void *handle, const char *symbol)
850 {
851 
852 	return dlsym(handle, symbol);
853 }
854 void *rumphijack_dlsym(void *, const char *)
855     __attribute__((__weak__, alias("rumpclient__dlsym")));
856 #endif
857 
858 static pid_t init_done = 0;
859 
860 int
rumpclient_init(void)861 rumpclient_init(void)
862 {
863 	char *p;
864 	int error;
865 	int rv = -1;
866 	int hstype;
867 	pid_t mypid;
868 
869 	/*
870 	 * Make sure we're not riding the context of a previous
871 	 * host fork.  Note: it's *possible* that after n>1 forks
872 	 * we have the same pid as one of our exited parents, but
873 	 * I'm pretty sure there are 0 practical implications, since
874 	 * it means generations would have to skip rumpclient init.
875 	 */
876 	if (init_done == (mypid = getpid()))
877 		return 0;
878 
879 #ifdef USE_KQUEUE
880 	/* kq does not traverse fork() */
881 	holyfd = -1;
882 #endif
883 	init_done = mypid;
884 
885 	sigfillset(&fullset);
886 
887 	/*
888 	 * sag mir, wo die symbols sind.  zogen fort, der krieg beginnt.
889 	 * wann wird man je verstehen?  wann wird man je verstehen?
890 	 */
891 #ifdef RTLD_NEXT
892 #define FINDSYM2(_name_,_syscall_)					\
893 	if ((host_##_name_ = rumphijack_dlsym(RTLD_NEXT,		\
894 	    #_syscall_)) == NULL) {					\
895 		if (rumphijack_dlsym == rumpclient__dlsym)		\
896 			host_##_name_ = _name_; /* static fallback */	\
897 		if (host_##_name_ == NULL) {				\
898 			fprintf(stderr,"cannot find %s: %s", #_syscall_,\
899 			    dlerror());					\
900 			exit(1);					\
901 		}							\
902 	}
903 #else
904 #define FINDSYM2(_name_,_syscall)					\
905 	host_##_name_ = _name_;
906 #endif
907 #define FINDSYM(_name_) FINDSYM2(_name_,_name_)
908 #ifdef __NetBSD__
909 	FINDSYM2(socket,__socket30)
910 #else
911 	FINDSYM(socket)
912 #endif
913 
914 	FINDSYM(close)
915 	FINDSYM(connect)
916 	FINDSYM(fcntl)
917 	FINDSYM(poll)
918 	FINDSYM(read)
919 	FINDSYM(sendmsg)
920 	FINDSYM(setsockopt)
921 	FINDSYM(dup)
922 	FINDSYM(execve)
923 
924 #ifdef USE_KQUEUE
925 	FINDSYM(kqueue)
926 #ifdef __NetBSD__
927 #if !__NetBSD_Prereq__(5,99,7)
928 	FINDSYM(kevent)
929 #elif !__NetBSD_Prereq__(10,99,4)
930 	FINDSYM2(kevent,_sys___kevent50)
931 #else
932 	FINDSYM2(kevent,_sys___kevent100)
933 #endif
934 #else
935 	FINDSYM(kevent)
936 #endif
937 #endif /* USE_KQUEUE */
938 
939 #ifdef USE_SIGNALFD
940 	FINDSYM(signalfd)
941 #endif
942 
943 #undef	FINDSYM
944 #undef	FINDSY2
945 
946 	if ((p = getenv("RUMP__PARSEDSERVER")) == NULL) {
947 		if ((p = getenv("RUMP_SERVER")) == NULL) {
948 			fprintf(stderr, "error: RUMP_SERVER not set\n");
949 			errno = ENOENT;
950 			goto out;
951 		}
952 	}
953 
954 	if ((error = parseurl(p, &serv_sa, &ptab_idx, 0)) != 0) {
955 		errno = error;
956 		goto out;
957 	}
958 
959 	if (doinit() == -1)
960 		goto out;
961 
962 	if ((p = getenv("RUMPCLIENT__EXECFD")) != NULL) {
963 		sscanf(p, "%d,%d", &clispc.spc_fd, &holyfd);
964 		unsetenv("RUMPCLIENT__EXECFD");
965 		hstype = HANDSHAKE_EXEC;
966 	} else {
967 		if (doconnect() == -1)
968 			goto out;
969 		hstype = HANDSHAKE_GUEST;
970 	}
971 
972 	error = handshake_req(&clispc, hstype, NULL, 0, false);
973 	if (error) {
974 		pthread_mutex_destroy(&clispc.spc_mtx);
975 		pthread_cond_destroy(&clispc.spc_cv);
976 		if (clispc.spc_fd != -1)
977 			host_close(clispc.spc_fd);
978 		errno = error;
979 		goto out;
980 	}
981 	rv = 0;
982 
983  out:
984 	if (rv == -1)
985 		init_done = 0;
986 	return rv;
987 }
988 
989 struct rumpclient_fork {
990 	uint32_t fork_auth[AUTHLEN];
991 	struct spclient fork_spc;
992 	int fork_holyfd;
993 };
994 
995 struct rumpclient_fork *
rumpclient_prefork(void)996 rumpclient_prefork(void)
997 {
998 	struct rumpclient_fork *rpf;
999 	sigset_t omask;
1000 	void *resp;
1001 	int rv;
1002 
1003 	pthread_sigmask(SIG_SETMASK, &fullset, &omask);
1004 	rpf = malloc(sizeof(*rpf));
1005 	if (rpf == NULL)
1006 		goto out;
1007 
1008 	if ((rv = prefork_req(&clispc, &omask, &resp)) != 0) {
1009 		free(rpf);
1010 		errno = rv;
1011 		rpf = NULL;
1012 		goto out;
1013 	}
1014 
1015 	memcpy(rpf->fork_auth, resp, sizeof(rpf->fork_auth));
1016 	free(resp);
1017 
1018 	rpf->fork_spc = clispc;
1019 	rpf->fork_holyfd = holyfd;
1020 
1021  out:
1022 	pthread_sigmask(SIG_SETMASK, &omask, NULL);
1023 	return rpf;
1024 }
1025 
1026 int
rumpclient_fork_init(struct rumpclient_fork * rpf)1027 rumpclient_fork_init(struct rumpclient_fork *rpf)
1028 {
1029 	int error;
1030 	int osock;
1031 
1032 	osock = clispc.spc_fd;
1033 	memset(&clispc, 0, sizeof(clispc));
1034 	clispc.spc_fd = osock;
1035 
1036 #ifdef USE_KQUEUE
1037 	holyfd = -1; /* kqueue descriptor is not copied over fork() */
1038 #else
1039 	if (holyfd != -1) {
1040 		host_close(holyfd);
1041 		holyfd = -1;
1042 	}
1043 #endif
1044 
1045 	if (doinit() == -1)
1046 		return -1;
1047 	if (doconnect() == -1)
1048 		return -1;
1049 
1050 	error = handshake_req(&clispc, HANDSHAKE_FORK, rpf->fork_auth,
1051 	    0, false);
1052 	if (error) {
1053 		pthread_mutex_destroy(&clispc.spc_mtx);
1054 		pthread_cond_destroy(&clispc.spc_cv);
1055 		errno = error;
1056 		return -1;
1057 	}
1058 
1059 	return 0;
1060 }
1061 
1062 /*ARGSUSED*/
1063 void
rumpclient_fork_cancel(struct rumpclient_fork * rpf)1064 rumpclient_fork_cancel(struct rumpclient_fork *rpf)
1065 {
1066 
1067 	/* EUNIMPL */
1068 }
1069 
1070 void
rumpclient_fork_vparent(struct rumpclient_fork * rpf)1071 rumpclient_fork_vparent(struct rumpclient_fork *rpf)
1072 {
1073 
1074 	clispc = rpf->fork_spc;
1075 	holyfd = rpf->fork_holyfd;
1076 }
1077 
1078 void
rumpclient_setconnretry(time_t timeout)1079 rumpclient_setconnretry(time_t timeout)
1080 {
1081 
1082 	if (timeout < RUMPCLIENT_RETRYCONN_DIE)
1083 		return; /* gigo */
1084 
1085 	retrytimo = timeout;
1086 }
1087 
1088 int
rumpclient__closenotify(int * fdp,enum rumpclient_closevariant variant)1089 rumpclient__closenotify(int *fdp, enum rumpclient_closevariant variant)
1090 {
1091 	int fd = *fdp;
1092 	int untilfd;
1093 	int newfd;
1094 
1095 	switch (variant) {
1096 	case RUMPCLIENT_CLOSE_FCLOSEM:
1097 		untilfd = MAX(clispc.spc_fd, holyfd);
1098 		for (; fd <= untilfd; fd++) {
1099 			if (fd == clispc.spc_fd || fd == holyfd)
1100 				continue;
1101 			(void)host_close(fd);
1102 		}
1103 		*fdp = fd;
1104 		break;
1105 
1106 	case RUMPCLIENT_CLOSE_CLOSE:
1107 	case RUMPCLIENT_CLOSE_DUP2:
1108 		if (fd == clispc.spc_fd) {
1109 			newfd = dupgood(clispc.spc_fd, 1);
1110 			if (newfd == -1)
1111 				return -1;
1112 
1113 #ifdef USE_KQUEUE
1114 			{
1115 			struct kevent kev[2];
1116 
1117 			/*
1118 			 * now, we have a new socket number, so change
1119 			 * the file descriptor that kqueue is
1120 			 * monitoring.  remove old and add new.
1121 			 */
1122 			EV_SET(&kev[0], clispc.spc_fd,
1123 			    EVFILT_READ, EV_DELETE, 0, 0, 0);
1124 			EV_SET(&kev[1], newfd,
1125 			    EVFILT_READ, EV_ADD|EV_ENABLE, 0, 0, 0);
1126 			if (host_kevent(holyfd, kev, 2, NULL, 0, NULL) == -1) {
1127 				int sverrno = errno;
1128 				host_close(newfd);
1129 				errno = sverrno;
1130 				return -1;
1131 			}}
1132 #endif /* !USE_KQUEUE */
1133 			clispc.spc_fd = newfd;
1134 		}
1135 		if (holyfd != -1 && fd == holyfd) {
1136 			newfd = dupgood(holyfd, 1);
1137 			if (newfd == -1)
1138 				return -1;
1139 			holyfd = newfd;
1140 		}
1141 		break;
1142 	}
1143 
1144 	return 0;
1145 }
1146 
1147 pid_t
rumpclient_fork(void)1148 rumpclient_fork(void)
1149 {
1150 
1151 	return rumpclient__dofork(fork);
1152 }
1153 
1154 /*
1155  * Process is about to exec.  Save info about our existing connection
1156  * in the env.  rumpclient will check for this info in init().
1157  * This is mostly for the benefit of rumphijack, but regular applications
1158  * may use it as well.
1159  */
1160 int
rumpclient_exec(const char * path,char * const argv[],char * const envp[])1161 rumpclient_exec(const char *path, char *const argv[], char *const envp[])
1162 {
1163 	char buf[4096];
1164 	char **newenv;
1165 	char *envstr, *envstr2;
1166 	size_t nelem;
1167 	int rv, sverrno;
1168 
1169 	snprintf(buf, sizeof(buf), "RUMPCLIENT__EXECFD=%d,%d",
1170 	    clispc.spc_fd, holyfd);
1171 	envstr = malloc(strlen(buf)+1);
1172 	if (envstr == NULL) {
1173 		return ENOMEM;
1174 	}
1175 	strcpy(envstr, buf);
1176 
1177 	/* do we have a fully parsed url we want to forward in the env? */
1178 	if (*parsedurl != '\0') {
1179 		snprintf(buf, sizeof(buf),
1180 		    "RUMP__PARSEDSERVER=%s", parsedurl);
1181 		envstr2 = malloc(strlen(buf)+1);
1182 		if (envstr2 == NULL) {
1183 			free(envstr);
1184 			return ENOMEM;
1185 		}
1186 		strcpy(envstr2, buf);
1187 	} else {
1188 		envstr2 = NULL;
1189 	}
1190 
1191 	for (nelem = 0; envp && envp[nelem]; nelem++)
1192 		continue;
1193 
1194 	newenv = malloc(sizeof(*newenv) * (nelem+3));
1195 	if (newenv == NULL) {
1196 		free(envstr2);
1197 		free(envstr);
1198 		return ENOMEM;
1199 	}
1200 	memcpy(&newenv[0], envp, nelem*sizeof(*envp));
1201 
1202 	newenv[nelem] = envstr;
1203 	newenv[nelem+1] = envstr2;
1204 	newenv[nelem+2] = NULL;
1205 
1206 	rv = host_execve(path, argv, newenv);
1207 
1208 	_DIAGASSERT(rv != 0);
1209 	sverrno = errno;
1210 	free(envstr2);
1211 	free(envstr);
1212 	free(newenv);
1213 	errno = sverrno;
1214 	return rv;
1215 }
1216 
1217 /*
1218  * daemon() is handwritten for the benefit of platforms which
1219  * do not support daemon().
1220  */
1221 int
rumpclient_daemon(int nochdir,int noclose)1222 rumpclient_daemon(int nochdir, int noclose)
1223 {
1224 	struct rumpclient_fork *rf;
1225 	int sverrno;
1226 
1227 	if ((rf = rumpclient_prefork()) == NULL)
1228 		return -1;
1229 
1230 	switch (fork()) {
1231 	case 0:
1232 		break;
1233 	case -1:
1234 		goto daemonerr;
1235 	default:
1236 		_exit(0);
1237 	}
1238 
1239 	if (setsid() == -1)
1240 		goto daemonerr;
1241 	if (!nochdir && chdir("/") == -1)
1242 		goto daemonerr;
1243 	if (!noclose) {
1244 		int fd = open("/dev/null", O_RDWR);
1245 		dup2(fd, 0);
1246 		dup2(fd, 1);
1247 		dup2(fd, 2);
1248 		if (fd > 2)
1249 			close(fd);
1250 	}
1251 
1252 	/* note: fork is either completed or cancelled by the call */
1253 	if (rumpclient_fork_init(rf) == -1)
1254 		return -1;
1255 
1256 	return 0;
1257 
1258  daemonerr:
1259 	sverrno = errno;
1260 	rumpclient_fork_cancel(rf);
1261 	errno = sverrno;
1262 	return -1;
1263 }
1264