xref: /openbsd-src/sbin/dump/tape.c (revision f1dd7b858388b4a23f4f67a4957ec5ff656ebbe8)
1 /*	$OpenBSD: tape.c,v 1.47 2021/01/21 00:16:36 mortimer Exp $	*/
2 /*	$NetBSD: tape.c,v 1.11 1997/06/05 11:13:26 lukem Exp $	*/
3 
4 /*-
5  * Copyright (c) 1980, 1991, 1993
6  *	The Regents of the University of California.  All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. Neither the name of the University nor the names of its contributors
17  *    may be used to endorse or promote products derived from this software
18  *    without specific prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  */
32 
33 #include <sys/param.h>	/* MAXBSIZE DEV_BSIZE */
34 #include <sys/socket.h>
35 #include <sys/time.h>
36 #include <sys/wait.h>
37 #include <sys/stat.h>
38 #include <ufs/ffs/fs.h>
39 #include <ufs/ufs/dinode.h>
40 
41 #include <protocols/dumprestore.h>
42 
43 #include <errno.h>
44 #include <fcntl.h>
45 #include <signal.h>
46 #include <stdio.h>
47 #include <stdlib.h>
48 #include <string.h>
49 #include <time.h>
50 #include <unistd.h>
51 #include <limits.h>
52 
53 #include "dump.h"
54 #include "pathnames.h"
55 
56 #define MINIMUM(a, b)	(((a) < (b)) ? (a) : (b))
57 
58 ino_t   curino; 		/* current inumber; used globally */
59 int     newtape;		/* new tape flag */
60 union u_spcl u_spcl;		/* mapping of variables in a control block */
61 
62 static int tapefd;		/* tape file descriptor */
63 static int64_t asize;		/* number of 0.1" units written on cur tape */
64 static int writesize;		/* size of malloc()ed buffer for tape */
65 static int64_t	lastspclrec = -1; /* tape block number of last written header */
66 static int trecno = 0;		/* next record to write in current block */
67 static int64_t blocksthisvol;	/* number of blocks on current output file */
68 static char *nexttape;
69 
70 static	ssize_t atomic(ssize_t (*)(int, void *, size_t), int, char *, int);
71 static	void doslave(int, int);
72 static	void enslave(void);
73 static	void flushtape(void);
74 static	void killall(void);
75 static	void rollforward(void);
76 
77 void	tperror(int signo);
78 void	sigpipe(int signo);
79 void	proceed(int signo);
80 
81 /*
82  * Concurrent dump mods (Caltech) - disk block reading and tape writing
83  * are exported to several slave processes.  While one slave writes the
84  * tape, the others read disk blocks; they pass control of the tape in
85  * a ring via signals. The parent process traverses the filesystem and
86  * sends writeheader()'s and lists of daddr's to the slaves via pipes.
87  * The following structure defines the instruction packets sent to slaves.
88  */
89 struct req {
90 	daddr_t dblk;
91 	int count;
92 };
93 static int reqsiz;
94 
95 #define SLAVES 3		/* 1 slave writing, 1 reading, 1 for slack */
96 static struct slave {
97 	int64_t tapea;		/* header number at start of this chunk */
98 	int64_t firstrec;	/* record number of this block */
99 	int count;		/* count to next header (used for TS_TAPE */
100 				/* after EOT) */
101 	int inode;		/* inode that we are currently dealing with */
102 	int fd;			/* FD for this slave */
103 	pid_t pid;		/* PID for this slave */
104 	int sent;		/* 1 == we've sent this slave requests */
105 	char (*tblock)[TP_BSIZE]; /* buffer for data blocks */
106 	struct req *req;	/* buffer for requests */
107 } slaves[SLAVES+1];
108 static struct slave *slp;
109 
110 static char	(*nextblock)[TP_BSIZE];
111 
112 static time_t tstart_volume;	/* time of volume start */
113 static int64_t tapea_volume;	/* value of spcl.c_tapea at volume start */
114 
115 static pid_t master;		/* pid of master, for sending error signals */
116 static int tenths;		/* length of tape used per block written */
117 static volatile sig_atomic_t caught;	/* have we caught the signal to proceed? */
118 
119 int
120 alloctape(void)
121 {
122 	int pgoff = getpagesize() - 1;
123 	char *buf;
124 	int i;
125 
126 	writesize = ntrec * TP_BSIZE;
127 	reqsiz = (ntrec + 1) * sizeof(struct req);
128 	/*
129 	 * CDC 92181's and 92185's make 0.8" gaps in 1600-bpi start/stop mode
130 	 * (see DEC TU80 User's Guide).  The shorter gaps of 6250-bpi require
131 	 * repositioning after stopping, i.e, streaming mode, where the gap is
132 	 * variable, 0.30" to 0.45".  The gap is maximal when the tape stops.
133 	 */
134 	if (blocksperfile == 0 && !unlimited)
135 		tenths = writesize / density +
136 		    (cartridge ? 16 : density == 625 ? 5 : 8);
137 	/*
138 	 * Allocate tape buffer contiguous with the array of instruction
139 	 * packets, so flushtape() can write them together with one write().
140 	 * Align tape buffer on page boundary to speed up tape write().
141 	 */
142 	for (i = 0; i <= SLAVES; i++) {
143 		buf = malloc((unsigned)(reqsiz + writesize + pgoff + TP_BSIZE));
144 		if (buf == NULL)
145 			return(0);
146 		slaves[i].tblock = (char (*)[TP_BSIZE])
147 		    (((long)&buf[ntrec + 1] + pgoff) &~ pgoff);
148 		slaves[i].req = (struct req *)slaves[i].tblock - ntrec - 1;
149 	}
150 	slp = &slaves[0];
151 	slp->count = 1;
152 	slp->tapea = 0;
153 	slp->firstrec = 0;
154 	nextblock = slp->tblock;
155 	return(1);
156 }
157 
158 void
159 writerec(char *dp, int isspcl)
160 {
161 
162 	slp->req[trecno].dblk = 0;
163 	slp->req[trecno].count = 1;
164 	*(union u_spcl *)(*(nextblock)++) = *(union u_spcl *)dp;
165 	if (isspcl)
166 		lastspclrec = spcl.c_tapea;
167 	trecno++;
168 	spcl.c_tapea++;
169 	if (trecno >= ntrec)
170 		flushtape();
171 }
172 
173 void
174 dumpblock(daddr_t blkno, int size)
175 {
176 	int avail, tpblks;
177 	daddr_t dblkno;
178 
179 	dblkno = fsbtodb(sblock, blkno);
180 	tpblks = size >> tp_bshift;
181 	while ((avail = MINIMUM(tpblks, ntrec - trecno)) > 0) {
182 		slp->req[trecno].dblk = dblkno;
183 		slp->req[trecno].count = avail;
184 		trecno += avail;
185 		spcl.c_tapea += avail;
186 		if (trecno >= ntrec)
187 			flushtape();
188 		dblkno += avail << (tp_bshift - (ffs(DEV_BSIZE) - 1));
189 		tpblks -= avail;
190 	}
191 }
192 
193 int	nogripe = 0;
194 
195 /* ARGSUSED */
196 void
197 tperror(int signo)
198 {
199 	/* XXX - signal races */
200 
201 	if (pipeout) {
202 		msg("write error on %s\n", tape);
203 		quit("Cannot recover\n");
204 		/* NOTREACHED */
205 	}
206 	msg("write error %lld blocks into volume %d\n",
207 	    (long long)blocksthisvol, tapeno);
208 	broadcast("DUMP WRITE ERROR!\n");
209 	if (!query("Do you want to restart?"))
210 		dumpabort(0);
211 	msg("Closing this volume.  Prepare to restart with new media;\n");
212 	msg("this dump volume will be rewritten.\n");
213 	killall();
214 	nogripe = 1;
215 	close_rewind();
216 	Exit(X_REWRITE);
217 }
218 
219 /* ARGSUSED */
220 void
221 sigpipe(int signo)
222 {
223 
224 	quit("Broken pipe\n");
225 }
226 
227 /*
228  * do_stats --
229  *	Update xferrate stats
230  */
231 time_t
232 do_stats(void)
233 {
234 	time_t tnow, ttaken;
235 	int64_t blocks;
236 
237 	(void)time(&tnow);
238 	ttaken = tnow - tstart_volume;
239 	blocks = spcl.c_tapea - tapea_volume;
240 	msg("Volume %d completed at: %s", tapeno, ctime(&tnow));
241 	if (ttaken > 0) {
242 		msg("Volume %d took %lld:%02lld:%02lld\n", tapeno,
243 		    (long long)ttaken / 3600, ((long long)ttaken % 3600) / 60,
244 		    (long long)ttaken % 60);
245 		blocks /= ttaken;
246 		msg("Volume %d transfer rate: %lld KB/s\n", tapeno, blocks);
247 		xferrate += blocks;
248 	}
249 	return(tnow);
250 }
251 
252 /*
253  * statussig --
254  *	information message upon receipt of SIGINFO
255  *	(derived from optr.c::timeest())
256  * XXX not safe
257  */
258 /* ARGSUSED */
259 void
260 statussig(int signo)
261 {
262 	time_t	tnow, deltat;
263 	int save_errno = errno;
264 
265 	if (blockswritten < 500)
266 		return;
267 	(void) time(&tnow);
268 	deltat = tstart_writing - tnow + (1.0 * (tnow - tstart_writing))
269 		/ blockswritten * tapesize;
270 	/* XXX not safe due to floating point printf */
271 	dprintf(STDERR_FILENO,
272 	    "dump: %s %3.2f%% done at %lld KB/s, finished in %d:%02d\n",
273 	    tape, (blockswritten * 100.0) / tapesize,
274 	    (spcl.c_tapea - tapea_volume) / (tnow - tstart_volume),
275 	    (int)(deltat / 3600), (int)((deltat % 3600) / 60));
276 	errno = save_errno;
277 }
278 
279 static void
280 flushtape(void)
281 {
282 	int i, blks, got;
283 	int64_t lastfirstrec;
284 
285 	int siz = (char *)nextblock - (char *)slp->req;
286 
287 	slp->req[trecno].count = 0;			/* Sentinel */
288 
289 	if (atomic((ssize_t (*)(int, void *, size_t))write, slp->fd,
290 	    (char *)slp->req, siz) != siz)
291 		quit("error writing command pipe: %s\n", strerror(errno));
292 	slp->sent = 1; /* we sent a request, read the response later */
293 
294 	lastfirstrec = slp->firstrec;
295 
296 	if (++slp >= &slaves[SLAVES])
297 		slp = &slaves[0];
298 
299 	/* Read results back from next slave */
300 	if (slp->sent) {
301 		if (atomic(read, slp->fd, (char *)&got, sizeof(got))
302 		    != sizeof(got)) {
303 			perror("  DUMP: error reading command pipe in master");
304 			dumpabort(0);
305 		}
306 		slp->sent = 0;
307 
308 		/* Check for end of tape */
309 		if (got < writesize) {
310 			msg("End of tape detected\n");
311 
312 			/*
313 			 * Drain the results, don't care what the values were.
314 			 * If we read them here then trewind won't...
315 			 */
316 			for (i = 0; i < SLAVES; i++) {
317 				if (slaves[i].sent) {
318 					if (atomic(read, slaves[i].fd,
319 					    (char *)&got, sizeof(got))
320 					    != sizeof(got)) {
321 						perror("  DUMP: error reading command pipe in master");
322 						dumpabort(0);
323 					}
324 					slaves[i].sent = 0;
325 				}
326 			}
327 
328 			close_rewind();
329 			rollforward();
330 			return;
331 		}
332 	}
333 
334 	blks = 0;
335 	if (spcl.c_type != TS_END && spcl.c_type != TS_CLRI &&
336 	    spcl.c_type != TS_BITS) {
337 		if (spcl.c_count > TP_NINDIR)
338 			quit("c_count too large\n");
339 		for (i = 0; i < spcl.c_count; i++)
340 			if (spcl.c_addr[i] != 0)
341 				blks++;
342 	}
343 	slp->count = lastspclrec + blks + 1 - spcl.c_tapea;
344 	slp->tapea = spcl.c_tapea;
345 	slp->firstrec = lastfirstrec + ntrec;
346 	slp->inode = curino;
347 	nextblock = slp->tblock;
348 	trecno = 0;
349 	asize += tenths;
350 	blockswritten += ntrec;
351 	blocksthisvol += ntrec;
352 	if (!pipeout && !unlimited && (blocksperfile ?
353 	    (blocksthisvol >= blocksperfile) : (asize > tsize))) {
354 		close_rewind();
355 		startnewtape(0);
356 	}
357 	timeest();
358 }
359 
360 void
361 trewind(void)
362 {
363 	struct stat sb;
364 	int f, got;
365 
366 	for (f = 0; f < SLAVES; f++) {
367 		/*
368 		 * Drain the results, but unlike EOT we DO (or should) care
369 		 * what the return values were, since if we detect EOT after
370 		 * we think we've written the last blocks to the tape anyway,
371 		 * we have to replay those blocks with rollforward.
372 		 *
373 		 * fixme: punt for now.
374 		 */
375 		if (slaves[f].sent) {
376 			if (atomic(read, slaves[f].fd, (char *)&got, sizeof(got))
377 			    != sizeof(got)) {
378 				perror("  DUMP: error reading command pipe in master");
379 				dumpabort(0);
380 			}
381 			slaves[f].sent = 0;
382 			if (got != writesize) {
383 				msg("EOT detected in last 2 tape records!\n");
384 				msg("Use a longer tape, decrease the size estimate\n");
385 				quit("or use no size estimate at all.\n");
386 			}
387 		}
388 		(void) close(slaves[f].fd);
389 	}
390 	while (wait((int *)NULL) >= 0)	/* wait for any signals from slaves */
391 		/* void */;
392 
393 	if (pipeout)
394 		return;
395 
396 	msg("Closing %s\n", tape);
397 
398 #ifdef RDUMP
399 	if (host) {
400 		rmtclose();
401 		while (rmtopen(tape, O_RDONLY) < 0)
402 			sleep(10);
403 		rmtclose();
404 		return;
405 	}
406 #endif
407 	/*
408 	 * st(4) says: "Bit 1 of the minor number specifies whether an eject is
409 	 * attempted when the device is closed.  When it is set, the device
410 	 * will attempt to eject its media on close ...".
411 	 *
412 	 * If the tape has been ejected, looping on open() will generate 'Media
413 	 * not present' errors until a tape is loaded. Once loaded the tape
414 	 * will be immediately ejected as a result of the second close().
415 	 *
416 	 * So if the tape will be ejected, just close and return.
417 	 */
418 	if ((fstat(tapefd, &sb) == 0) && (minor(sb.st_rdev) & 0x02)) {
419 		(void) close(tapefd);
420 		return;
421 	}
422 
423 	(void) close(tapefd);
424 	while ((f = open(tape, O_RDONLY)) == -1)
425 		sleep (10);
426 	(void) close(f);
427 }
428 
429 void
430 close_rewind(void)
431 {
432 	trewind();
433 	(void)do_stats();
434 	if (nexttape)
435 		return;
436 	if (!nogripe) {
437 		msg("Change Volumes: Mount volume #%d\n", tapeno+1);
438 		broadcast("CHANGE DUMP VOLUMES!\7\7\n");
439 	}
440 	while (!query("Is the new volume mounted and ready to go?"))
441 		if (query("Do you want to abort?")) {
442 			dumpabort(0);
443 			/*NOTREACHED*/
444 		}
445 }
446 
447 void
448 rollforward(void)
449 {
450 	struct req *p, *q, *prev;
451 	struct slave *tslp;
452 	int i, size, got;
453 	int64_t savedtapea;
454 	union u_spcl *ntb, *otb;
455 	tslp = &slaves[SLAVES];
456 	ntb = (union u_spcl *)tslp->tblock[1];
457 
458 	/*
459 	 * Each of the N slaves should have requests that need to
460 	 * be replayed on the next tape.  Use the extra slave buffers
461 	 * (slaves[SLAVES]) to construct request lists to be sent to
462 	 * each slave in turn.
463 	 */
464 	for (i = 0; i < SLAVES; i++) {
465 		q = &tslp->req[1];
466 		otb = (union u_spcl *)slp->tblock;
467 
468 		/*
469 		 * For each request in the current slave, copy it to tslp.
470 		 */
471 
472 		prev = NULL;
473 		for (p = slp->req; p->count > 0; p += p->count) {
474 			*q = *p;
475 			if (p->dblk == 0)
476 				*ntb++ = *otb++; /* copy the datablock also */
477 			prev = q;
478 			q += q->count;
479 		}
480 		if (prev == NULL)
481 			quit("rollforward: protocol botch\n");
482 		if (prev->dblk != 0)
483 			prev->count -= 1;
484 		else
485 			ntb--;
486 		q -= 1;
487 		q->count = 0;
488 		q = &tslp->req[0];
489 		if (i == 0) {
490 			q->dblk = 0;
491 			q->count = 1;
492 			trecno = 0;
493 			nextblock = tslp->tblock;
494 			savedtapea = spcl.c_tapea;
495 			spcl.c_tapea = slp->tapea;
496 			startnewtape(0);
497 			spcl.c_tapea = savedtapea;
498 			lastspclrec = savedtapea - 1;
499 		}
500 		size = (char *)ntb - (char *)q;
501 		if (atomic((ssize_t (*)(int, void *, size_t))write,
502 		    slp->fd, (char *)q, size) != size) {
503 			perror("  DUMP: error writing command pipe");
504 			dumpabort(0);
505 		}
506 		slp->sent = 1;
507 		if (++slp >= &slaves[SLAVES])
508 			slp = &slaves[0];
509 
510 		q->count = 1;
511 
512 		if (prev->dblk != 0) {
513 			/*
514 			 * If the last one was a disk block, make the
515 			 * first of this one be the last bit of that disk
516 			 * block...
517 			 */
518 			q->dblk = prev->dblk +
519 				prev->count * (TP_BSIZE / DEV_BSIZE);
520 			ntb = (union u_spcl *)tslp->tblock;
521 		} else {
522 			/*
523 			 * It wasn't a disk block.  Copy the data to its
524 			 * new location in the buffer.
525 			 */
526 			q->dblk = 0;
527 			*((union u_spcl *)tslp->tblock) = *ntb;
528 			ntb = (union u_spcl *)tslp->tblock[1];
529 		}
530 	}
531 	slp->req[0] = *q;
532 	nextblock = slp->tblock;
533 	if (q->dblk == 0)
534 		nextblock++;
535 	trecno = 1;
536 
537 	/*
538 	 * Clear the first slaves' response.  One hopes that it
539 	 * worked ok, otherwise the tape is much too short!
540 	 */
541 	if (slp->sent) {
542 		if (atomic(read, slp->fd, (char *)&got, sizeof(got))
543 		    != sizeof(got)) {
544 			perror("  DUMP: error reading command pipe in master");
545 			dumpabort(0);
546 		}
547 		slp->sent = 0;
548 
549 		if (got != writesize) {
550 			quit("EOT detected at start of the tape!\n");
551 		}
552 	}
553 }
554 
555 /*
556  * We implement taking and restoring checkpoints on the tape level.
557  * When each tape is opened, a new process is created by forking; this
558  * saves all of the necessary context in the parent.  The child
559  * continues the dump; the parent waits around, saving the context.
560  * If the child returns X_REWRITE, then it had problems writing that tape;
561  * this causes the parent to fork again, duplicating the context, and
562  * everything continues as if nothing had happened.
563  */
564 void
565 startnewtape(int top)
566 {
567 	pid_t	parentpid;
568 	pid_t	childpid;
569 	int	status;
570 	pid_t	waitingpid;
571 	char	*p;
572 	sig_t	interrupt_save;
573 
574 	interrupt_save = signal(SIGINT, SIG_IGN);
575 	parentpid = getpid();
576 	tapea_volume = spcl.c_tapea;
577 	(void)time(&tstart_volume);
578 
579 restore_check_point:
580 	(void)signal(SIGINT, interrupt_save);
581 	/*
582 	 *	All signals are inherited...
583 	 */
584 	childpid = fork();
585 	if (childpid == -1) {
586 		msg("Context save fork fails in parent %d\n", parentpid);
587 		Exit(X_ABORT);
588 	}
589 	if (childpid != 0) {
590 		/*
591 		 *	PARENT:
592 		 *	save the context by waiting
593 		 *	until the child doing all of the work returns.
594 		 *	don't catch the interrupt
595 		 */
596 		signal(SIGINT, SIG_IGN);
597 #ifdef TDEBUG
598 		msg("Tape: %d; parent process: %d child process %d\n",
599 			tapeno+1, parentpid, childpid);
600 #endif /* TDEBUG */
601 		while ((waitingpid = wait(&status)) != childpid)
602 			msg("Parent %d waiting for child %d has another child %d return\n",
603 				parentpid, childpid, waitingpid);
604 		if (status & 0xFF) {
605 			msg("Child %d returns LOB status %o\n",
606 				childpid, status&0xFF);
607 		}
608 		status = (status >> 8) & 0xFF;
609 #ifdef TDEBUG
610 		switch(status) {
611 			case X_FINOK:
612 				msg("Child %d finishes X_FINOK\n", childpid);
613 				break;
614 			case X_ABORT:
615 				msg("Child %d finishes X_ABORT\n", childpid);
616 				break;
617 			case X_REWRITE:
618 				msg("Child %d finishes X_REWRITE\n", childpid);
619 				break;
620 			default:
621 				msg("Child %d finishes unknown %d\n",
622 					childpid, status);
623 				break;
624 		}
625 #endif /* TDEBUG */
626 		switch(status) {
627 			case X_FINOK:
628 				Exit(X_FINOK);
629 				break;
630 			case X_ABORT:
631 				Exit(X_ABORT);
632 				break;
633 			case X_REWRITE:
634 				goto restore_check_point;
635 			default:
636 				msg("Bad return code from dump: %d\n", status);
637 				Exit(X_ABORT);
638 		}
639 		/*NOTREACHED*/
640 	} else {	/* we are the child; just continue */
641 #ifdef TDEBUG
642 		sleep(4);	/* allow time for parent's message to get out */
643 		msg("Child on Tape %d has parent %d, my pid = %d\n",
644 			tapeno+1, parentpid, getpid());
645 #endif /* TDEBUG */
646 		/*
647 		 * If we have a name like "/dev/rst0,/dev/rst1",
648 		 * use the name before the comma first, and save
649 		 * the remaining names for subsequent volumes.
650 		 */
651 		tapeno++;               /* current tape sequence */
652 		if (nexttape || strchr(tape, ',')) {
653 			if (nexttape && *nexttape)
654 				tape = nexttape;
655 			if ((p = strchr(tape, ',')) != NULL) {
656 				*p = '\0';
657 				nexttape = p + 1;
658 			} else
659 				nexttape = NULL;
660 			msg("Dumping volume %d on %s\n", tapeno, tape);
661 		}
662 #ifdef RDUMP
663 		while ((tapefd = (host ? rmtopen(tape, O_WRONLY|O_CREAT) :
664 			pipeout ? 1 : open(tape, O_WRONLY|O_CREAT, 0666))) == -1)
665 #else
666 		while ((tapefd = (pipeout ? 1 :
667 				  open(tape, O_WRONLY|O_CREAT, 0666))) == -1)
668 #endif
669 		    {
670 			msg("Cannot open output \"%s\".\n", tape);
671 			if (!query("Do you want to retry the open?"))
672 				dumpabort(0);
673 		}
674 
675 		enslave();  /* Share open tape file descriptor with slaves */
676 
677 		asize = 0;
678 		blocksthisvol = 0;
679 		if (top)
680 			newtape++;		/* new tape signal */
681 		spcl.c_count = slp->count;
682 		/*
683 		 * measure firstrec in TP_BSIZE units since restore doesn't
684 		 * know the correct ntrec value...
685 		 */
686 		spcl.c_firstrec = slp->firstrec;
687 		spcl.c_volume++;
688 		spcl.c_type = TS_TAPE;
689 		if (sblock->fs_magic != FS_UFS2_MAGIC)
690 			spcl.c_flags |= DR_NEWHEADER;
691 		writeheader((ino_t)slp->inode);
692 		if (sblock->fs_magic != FS_UFS2_MAGIC)
693 			spcl.c_flags &=~ DR_NEWHEADER;
694 		msg("Volume %d started at: %s", tapeno, ctime(&tstart_volume));
695 		if (tapeno > 1)
696 			msg("Volume %d begins with blocks from inode %llu\n",
697 			    tapeno, (unsigned long long)slp->inode);
698 	}
699 }
700 
701 /* ARGSUSED */
702 void
703 dumpabort(int signo)
704 {
705 
706 	if (master != 0 && master != getpid())
707 		/* Signals master to call dumpabort */
708 		(void) kill(master, SIGTERM);
709 	else {
710 		killall();
711 		msg("The ENTIRE dump is aborted.\n");
712 	}
713 #ifdef RDUMP
714 	rmtclose();
715 #endif
716 	Exit(X_ABORT);
717 }
718 
719 __dead void
720 Exit(int status)
721 {
722 
723 #ifdef TDEBUG
724 	msg("pid = %d exits with status %d\n", getpid(), status);
725 #endif /* TDEBUG */
726 	exit(status);
727 }
728 
729 /*
730  * proceed - handler for SIGUSR2, used to synchronize IO between the slaves.
731  */
732 /* ARGSUSED */
733 void
734 proceed(int signo)
735 {
736 	caught++;
737 }
738 
739 void
740 enslave(void)
741 {
742 	int cmd[2];
743 	int i, j;
744 
745 	master = getpid();
746 
747 	signal(SIGTERM, dumpabort);  /* Slave sends SIGTERM on dumpabort() */
748 	signal(SIGPIPE, sigpipe);
749 	signal(SIGUSR1, tperror);    /* Slave sends SIGUSR1 on tape errors */
750 	signal(SIGUSR2, proceed);    /* Slave sends SIGUSR2 to next slave */
751 
752 	for (i = 0; i < SLAVES; i++) {
753 		if (i == slp - &slaves[0]) {
754 			caught = 1;
755 		} else {
756 			caught = 0;
757 		}
758 
759 		if (socketpair(AF_UNIX, SOCK_STREAM, 0, cmd) == -1 ||
760 		    (slaves[i].pid = fork()) == -1)
761 			quit("too many slaves, %d (recompile smaller): %s\n",
762 			    i, strerror(errno));
763 
764 		slaves[i].fd = cmd[1];
765 		slaves[i].sent = 0;
766 		if (slaves[i].pid == 0) { 	    /* Slave starts up here */
767 			for (j = 0; j <= i; j++)
768 			        (void) close(slaves[j].fd);
769 			signal(SIGINT, SIG_IGN);    /* Master handles this */
770 			signal(SIGINFO, SIG_IGN);
771 			doslave(cmd[0], i);
772 			Exit(X_FINOK);
773 		}
774 	}
775 
776 	for (i = 0; i < SLAVES; i++)
777 		(void) atomic((ssize_t (*)(int, void *, size_t))write,
778 		    slaves[i].fd, (char *) &slaves[(i + 1) % SLAVES].pid,
779 		    sizeof(slaves[0].pid));
780 	master = 0;
781 }
782 
783 void
784 killall(void)
785 {
786 	int i;
787 
788 	for (i = 0; i < SLAVES; i++)
789 		if (slaves[i].pid > 0) {
790 			(void) kill(slaves[i].pid, SIGKILL);
791 			slaves[i].pid = 0;
792 		}
793 }
794 
795 /*
796  * Synchronization - each process has a lockfile, and shares file
797  * descriptors to the following process's lockfile.  When our write
798  * completes, we release our lock on the following process's lock-
799  * file, allowing the following process to lock it and proceed. We
800  * get the lock back for the next cycle by swapping descriptors.
801  */
802 static void
803 doslave(int cmd, int slave_number)
804 {
805 	int nread, nextslave, size, wrote = 0, eot_count;
806 	sigset_t nsigset, osigset;
807 
808 	/*
809 	 * Need our own seek pointer.
810 	 */
811 	(void) close(diskfd);
812 	if ((diskfd = open(disk, O_RDONLY)) == -1)
813 		quit("slave couldn't reopen disk: %s\n", strerror(errno));
814 
815 	/*
816 	 * Need the pid of the next slave in the loop...
817 	 */
818 	if ((nread = atomic(read, cmd, (char *)&nextslave, sizeof(nextslave)))
819 	    != sizeof(nextslave)) {
820 		quit("master/slave protocol botched - didn't get pid of next slave.\n");
821 	}
822 
823 	/*
824 	 * Get list of blocks to dump, read the blocks into tape buffer
825 	 */
826 	while ((nread = atomic(read, cmd, (char *)slp->req, reqsiz)) == reqsiz) {
827 		struct req *p = slp->req;
828 
829 		for (trecno = 0; trecno < ntrec;
830 		     trecno += p->count, p += p->count) {
831 			if (p->dblk) {
832 				bread(p->dblk, slp->tblock[trecno],
833 					p->count * TP_BSIZE);
834 			} else {
835 				if (p->count != 1 || atomic(read, cmd,
836 				    (char *)slp->tblock[trecno],
837 				    TP_BSIZE) != TP_BSIZE)
838 				       quit("master/slave protocol botched.\n");
839 			}
840 		}
841 
842 		sigemptyset(&nsigset);
843 		sigaddset(&nsigset, SIGUSR2);
844 		sigprocmask(SIG_BLOCK, &nsigset, &osigset);
845 		while (!caught)
846 			sigsuspend(&osigset);
847 		caught = 0;
848 		sigprocmask(SIG_SETMASK, &osigset, NULL);
849 
850 		/* Try to write the data... */
851 		eot_count = 0;
852 		size = 0;
853 
854 		while (eot_count < 10 && size < writesize) {
855 #ifdef RDUMP
856 			if (host)
857 				wrote = rmtwrite(slp->tblock[0]+size,
858 				    writesize-size);
859 			else
860 #endif
861 				wrote = write(tapefd, slp->tblock[0]+size,
862 				    writesize-size);
863 #ifdef WRITEDEBUG
864 			printf("slave %d wrote %d\n", slave_number, wrote);
865 #endif
866 			if (wrote < 0)
867 				break;
868 			if (wrote == 0)
869 				eot_count++;
870 			size += wrote;
871 		}
872 
873 #ifdef WRITEDEBUG
874 		if (size != writesize)
875 		 printf("slave %d only wrote %d out of %d bytes and gave up.\n",
876 		     slave_number, size, writesize);
877 #endif
878 
879 		if (eot_count > 0)
880 			size = 0;
881 
882 		/*
883 		 * Handle ENOSPC as an EOT condition
884 		 */
885 		if (wrote < 0 && errno == ENOSPC) {
886 			wrote = 0;
887 			eot_count++;
888 		}
889 
890 		if (size < 0) {
891 			(void) kill(master, SIGUSR1);
892 			sigemptyset(&nsigset);
893 			for (;;)
894 				sigsuspend(&nsigset);
895 		} else {
896 			/*
897 			 * pass size of write back to master
898 			 * (for EOT handling)
899 			 */
900 			(void) atomic((ssize_t (*)(int, void *, size_t))write,
901 			    cmd, (char *)&size, sizeof(size));
902 		}
903 
904 		/*
905 		 * If partial write, don't want next slave to go.
906 		 * Also jolts him awake.
907 		 */
908 		(void) kill(nextslave, SIGUSR2);
909 	}
910 	if (nread != 0)
911 		quit("error reading command pipe: %s\n", strerror(errno));
912 }
913 
914 /*
915  * Since a read from a pipe may not return all we asked for,
916  * or a write may not write all we ask if we get a signal,
917  * loop until the count is satisfied (or error).
918  */
919 static ssize_t
920 atomic(ssize_t (*func)(int, void *, size_t), int fd, char *buf, int count)
921 {
922 	ssize_t got, need = count;
923 
924 	while ((got = (*func)(fd, buf, need)) > 0 && (need -= got) > 0)
925 		buf += got;
926 	return (got < 0 ? got : count - need);
927 }
928