xref: /openbsd-src/sbin/dump/tape.c (revision 1a8dbaac879b9f3335ad7fb25429ce63ac1d6bac)
1 /*	$OpenBSD: tape.c,v 1.46 2020/10/01 07:58:54 otto 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 int	writesize;		/* size of malloc()ed buffer for tape */
59 int64_t	lastspclrec = -1;	/* tape block number of last written header */
60 int	trecno = 0;		/* next record to write in current block */
61 extern	int64_t blocksperfile;	/* number of blocks per output file */
62 int64_t	blocksthisvol;		/* number of blocks on current output file */
63 extern	int ntrec;		/* blocking factor on tape */
64 extern	int cartridge;
65 extern	char *host;
66 char	*nexttape;
67 
68 static	ssize_t atomic(ssize_t (*)(int, void *, size_t), int, char *, int);
69 static	void doslave(int, int);
70 static	void enslave(void);
71 static	void flushtape(void);
72 static	void killall(void);
73 static	void rollforward(void);
74 
75 void	tperror(int signo);
76 void	sigpipe(int signo);
77 void	proceed(int signo);
78 
79 /*
80  * Concurrent dump mods (Caltech) - disk block reading and tape writing
81  * are exported to several slave processes.  While one slave writes the
82  * tape, the others read disk blocks; they pass control of the tape in
83  * a ring via signals. The parent process traverses the filesystem and
84  * sends writeheader()'s and lists of daddr's to the slaves via pipes.
85  * The following structure defines the instruction packets sent to slaves.
86  */
87 struct req {
88 	daddr_t dblk;
89 	int count;
90 };
91 int reqsiz;
92 
93 #define SLAVES 3		/* 1 slave writing, 1 reading, 1 for slack */
94 struct slave {
95 	int64_t tapea;		/* header number at start of this chunk */
96 	int64_t firstrec;	/* record number of this block */
97 	int count;		/* count to next header (used for TS_TAPE */
98 				/* after EOT) */
99 	int inode;		/* inode that we are currently dealing with */
100 	int fd;			/* FD for this slave */
101 	pid_t pid;		/* PID for this slave */
102 	int sent;		/* 1 == we've sent this slave requests */
103 	char (*tblock)[TP_BSIZE]; /* buffer for data blocks */
104 	struct req *req;	/* buffer for requests */
105 } slaves[SLAVES+1];
106 struct slave *slp;
107 
108 char	(*nextblock)[TP_BSIZE];
109 
110 static time_t tstart_volume;	/* time of volume start */
111 static int64_t tapea_volume;	/* value of spcl.c_tapea at volume start */
112 
113 pid_t master;		/* pid of master, for sending error signals */
114 int tenths;		/* length of tape used per block written */
115 static volatile sig_atomic_t caught;	/* have we caught the signal to proceed? */
116 
117 int
118 alloctape(void)
119 {
120 	int pgoff = getpagesize() - 1;
121 	char *buf;
122 	int i;
123 
124 	writesize = ntrec * TP_BSIZE;
125 	reqsiz = (ntrec + 1) * sizeof(struct req);
126 	/*
127 	 * CDC 92181's and 92185's make 0.8" gaps in 1600-bpi start/stop mode
128 	 * (see DEC TU80 User's Guide).  The shorter gaps of 6250-bpi require
129 	 * repositioning after stopping, i.e, streaming mode, where the gap is
130 	 * variable, 0.30" to 0.45".  The gap is maximal when the tape stops.
131 	 */
132 	if (blocksperfile == 0 && !unlimited)
133 		tenths = writesize / density +
134 		    (cartridge ? 16 : density == 625 ? 5 : 8);
135 	/*
136 	 * Allocate tape buffer contiguous with the array of instruction
137 	 * packets, so flushtape() can write them together with one write().
138 	 * Align tape buffer on page boundary to speed up tape write().
139 	 */
140 	for (i = 0; i <= SLAVES; i++) {
141 		buf = malloc((unsigned)(reqsiz + writesize + pgoff + TP_BSIZE));
142 		if (buf == NULL)
143 			return(0);
144 		slaves[i].tblock = (char (*)[TP_BSIZE])
145 		    (((long)&buf[ntrec + 1] + pgoff) &~ pgoff);
146 		slaves[i].req = (struct req *)slaves[i].tblock - ntrec - 1;
147 	}
148 	slp = &slaves[0];
149 	slp->count = 1;
150 	slp->tapea = 0;
151 	slp->firstrec = 0;
152 	nextblock = slp->tblock;
153 	return(1);
154 }
155 
156 void
157 writerec(char *dp, int isspcl)
158 {
159 
160 	slp->req[trecno].dblk = 0;
161 	slp->req[trecno].count = 1;
162 	*(union u_spcl *)(*(nextblock)++) = *(union u_spcl *)dp;
163 	if (isspcl)
164 		lastspclrec = spcl.c_tapea;
165 	trecno++;
166 	spcl.c_tapea++;
167 	if (trecno >= ntrec)
168 		flushtape();
169 }
170 
171 void
172 dumpblock(daddr_t blkno, int size)
173 {
174 	int avail, tpblks;
175 	daddr_t dblkno;
176 
177 	dblkno = fsbtodb(sblock, blkno);
178 	tpblks = size >> tp_bshift;
179 	while ((avail = MINIMUM(tpblks, ntrec - trecno)) > 0) {
180 		slp->req[trecno].dblk = dblkno;
181 		slp->req[trecno].count = avail;
182 		trecno += avail;
183 		spcl.c_tapea += avail;
184 		if (trecno >= ntrec)
185 			flushtape();
186 		dblkno += avail << (tp_bshift - (ffs(DEV_BSIZE) - 1));
187 		tpblks -= avail;
188 	}
189 }
190 
191 int	nogripe = 0;
192 
193 /* ARGSUSED */
194 void
195 tperror(int signo)
196 {
197 	/* XXX - signal races */
198 
199 	if (pipeout) {
200 		msg("write error on %s\n", tape);
201 		quit("Cannot recover\n");
202 		/* NOTREACHED */
203 	}
204 	msg("write error %lld blocks into volume %d\n",
205 	    (long long)blocksthisvol, tapeno);
206 	broadcast("DUMP WRITE ERROR!\n");
207 	if (!query("Do you want to restart?"))
208 		dumpabort(0);
209 	msg("Closing this volume.  Prepare to restart with new media;\n");
210 	msg("this dump volume will be rewritten.\n");
211 	killall();
212 	nogripe = 1;
213 	close_rewind();
214 	Exit(X_REWRITE);
215 }
216 
217 /* ARGSUSED */
218 void
219 sigpipe(int signo)
220 {
221 
222 	quit("Broken pipe\n");
223 }
224 
225 /*
226  * do_stats --
227  *	Update xferrate stats
228  */
229 time_t
230 do_stats(void)
231 {
232 	time_t tnow, ttaken;
233 	int64_t blocks;
234 
235 	(void)time(&tnow);
236 	ttaken = tnow - tstart_volume;
237 	blocks = spcl.c_tapea - tapea_volume;
238 	msg("Volume %d completed at: %s", tapeno, ctime(&tnow));
239 	if (ttaken > 0) {
240 		msg("Volume %d took %lld:%02lld:%02lld\n", tapeno,
241 		    (long long)ttaken / 3600, ((long long)ttaken % 3600) / 60,
242 		    (long long)ttaken % 60);
243 		blocks /= ttaken;
244 		msg("Volume %d transfer rate: %lld KB/s\n", tapeno, blocks);
245 		xferrate += blocks;
246 	}
247 	return(tnow);
248 }
249 
250 /*
251  * statussig --
252  *	information message upon receipt of SIGINFO
253  *	(derived from optr.c::timeest())
254  * XXX not safe
255  */
256 /* ARGSUSED */
257 void
258 statussig(int signo)
259 {
260 	time_t	tnow, deltat;
261 	int save_errno = errno;
262 
263 	if (blockswritten < 500)
264 		return;
265 	(void) time(&tnow);
266 	deltat = tstart_writing - tnow + (1.0 * (tnow - tstart_writing))
267 		/ blockswritten * tapesize;
268 	/* XXX not safe due to floating point printf */
269 	dprintf(STDERR_FILENO,
270 	    "dump: %s %3.2f%% done at %lld KB/s, finished in %d:%02d\n",
271 	    tape, (blockswritten * 100.0) / tapesize,
272 	    (spcl.c_tapea - tapea_volume) / (tnow - tstart_volume),
273 	    (int)(deltat / 3600), (int)((deltat % 3600) / 60));
274 	errno = save_errno;
275 }
276 
277 static void
278 flushtape(void)
279 {
280 	int i, blks, got;
281 	int64_t lastfirstrec;
282 
283 	int siz = (char *)nextblock - (char *)slp->req;
284 
285 	slp->req[trecno].count = 0;			/* Sentinel */
286 
287 	if (atomic((ssize_t (*)(int, void *, size_t))write, slp->fd,
288 	    (char *)slp->req, siz) != siz)
289 		quit("error writing command pipe: %s\n", strerror(errno));
290 	slp->sent = 1; /* we sent a request, read the response later */
291 
292 	lastfirstrec = slp->firstrec;
293 
294 	if (++slp >= &slaves[SLAVES])
295 		slp = &slaves[0];
296 
297 	/* Read results back from next slave */
298 	if (slp->sent) {
299 		if (atomic(read, slp->fd, (char *)&got, sizeof(got))
300 		    != sizeof(got)) {
301 			perror("  DUMP: error reading command pipe in master");
302 			dumpabort(0);
303 		}
304 		slp->sent = 0;
305 
306 		/* Check for end of tape */
307 		if (got < writesize) {
308 			msg("End of tape detected\n");
309 
310 			/*
311 			 * Drain the results, don't care what the values were.
312 			 * If we read them here then trewind won't...
313 			 */
314 			for (i = 0; i < SLAVES; i++) {
315 				if (slaves[i].sent) {
316 					if (atomic(read, slaves[i].fd,
317 					    (char *)&got, sizeof(got))
318 					    != sizeof(got)) {
319 						perror("  DUMP: error reading command pipe in master");
320 						dumpabort(0);
321 					}
322 					slaves[i].sent = 0;
323 				}
324 			}
325 
326 			close_rewind();
327 			rollforward();
328 			return;
329 		}
330 	}
331 
332 	blks = 0;
333 	if (spcl.c_type != TS_END && spcl.c_type != TS_CLRI &&
334 	    spcl.c_type != TS_BITS) {
335 		if (spcl.c_count > TP_NINDIR)
336 			quit("c_count too large\n");
337 		for (i = 0; i < spcl.c_count; i++)
338 			if (spcl.c_addr[i] != 0)
339 				blks++;
340 	}
341 	slp->count = lastspclrec + blks + 1 - spcl.c_tapea;
342 	slp->tapea = spcl.c_tapea;
343 	slp->firstrec = lastfirstrec + ntrec;
344 	slp->inode = curino;
345 	nextblock = slp->tblock;
346 	trecno = 0;
347 	asize += tenths;
348 	blockswritten += ntrec;
349 	blocksthisvol += ntrec;
350 	if (!pipeout && !unlimited && (blocksperfile ?
351 	    (blocksthisvol >= blocksperfile) : (asize > tsize))) {
352 		close_rewind();
353 		startnewtape(0);
354 	}
355 	timeest();
356 }
357 
358 void
359 trewind(void)
360 {
361 	struct stat sb;
362 	int f, got;
363 
364 	for (f = 0; f < SLAVES; f++) {
365 		/*
366 		 * Drain the results, but unlike EOT we DO (or should) care
367 		 * what the return values were, since if we detect EOT after
368 		 * we think we've written the last blocks to the tape anyway,
369 		 * we have to replay those blocks with rollforward.
370 		 *
371 		 * fixme: punt for now.
372 		 */
373 		if (slaves[f].sent) {
374 			if (atomic(read, slaves[f].fd, (char *)&got, sizeof(got))
375 			    != sizeof(got)) {
376 				perror("  DUMP: error reading command pipe in master");
377 				dumpabort(0);
378 			}
379 			slaves[f].sent = 0;
380 			if (got != writesize) {
381 				msg("EOT detected in last 2 tape records!\n");
382 				msg("Use a longer tape, decrease the size estimate\n");
383 				quit("or use no size estimate at all.\n");
384 			}
385 		}
386 		(void) close(slaves[f].fd);
387 	}
388 	while (wait((int *)NULL) >= 0)	/* wait for any signals from slaves */
389 		/* void */;
390 
391 	if (pipeout)
392 		return;
393 
394 	msg("Closing %s\n", tape);
395 
396 #ifdef RDUMP
397 	if (host) {
398 		rmtclose();
399 		while (rmtopen(tape, O_RDONLY) < 0)
400 			sleep(10);
401 		rmtclose();
402 		return;
403 	}
404 #endif
405 	/*
406 	 * st(4) says: "Bit 1 of the minor number specifies whether an eject is
407 	 * attempted when the device is closed.  When it is set, the device
408 	 * will attempt to eject its media on close ...".
409 	 *
410 	 * If the tape has been ejected, looping on open() will generate 'Media
411 	 * not present' errors until a tape is loaded. Once loaded the tape
412 	 * will be immediately ejected as a result of the second close().
413 	 *
414 	 * So if the tape will be ejected, just close and return.
415 	 */
416 	if ((fstat(tapefd, &sb) == 0) && (minor(sb.st_rdev) & 0x02)) {
417 		(void) close(tapefd);
418 		return;
419 	}
420 
421 	(void) close(tapefd);
422 	while ((f = open(tape, O_RDONLY)) == -1)
423 		sleep (10);
424 	(void) close(f);
425 }
426 
427 void
428 close_rewind(void)
429 {
430 	trewind();
431 	(void)do_stats();
432 	if (nexttape)
433 		return;
434 	if (!nogripe) {
435 		msg("Change Volumes: Mount volume #%d\n", tapeno+1);
436 		broadcast("CHANGE DUMP VOLUMES!\7\7\n");
437 	}
438 	while (!query("Is the new volume mounted and ready to go?"))
439 		if (query("Do you want to abort?")) {
440 			dumpabort(0);
441 			/*NOTREACHED*/
442 		}
443 }
444 
445 void
446 rollforward(void)
447 {
448 	struct req *p, *q, *prev;
449 	struct slave *tslp;
450 	int i, size, got;
451 	int64_t savedtapea;
452 	union u_spcl *ntb, *otb;
453 	tslp = &slaves[SLAVES];
454 	ntb = (union u_spcl *)tslp->tblock[1];
455 
456 	/*
457 	 * Each of the N slaves should have requests that need to
458 	 * be replayed on the next tape.  Use the extra slave buffers
459 	 * (slaves[SLAVES]) to construct request lists to be sent to
460 	 * each slave in turn.
461 	 */
462 	for (i = 0; i < SLAVES; i++) {
463 		q = &tslp->req[1];
464 		otb = (union u_spcl *)slp->tblock;
465 
466 		/*
467 		 * For each request in the current slave, copy it to tslp.
468 		 */
469 
470 		prev = NULL;
471 		for (p = slp->req; p->count > 0; p += p->count) {
472 			*q = *p;
473 			if (p->dblk == 0)
474 				*ntb++ = *otb++; /* copy the datablock also */
475 			prev = q;
476 			q += q->count;
477 		}
478 		if (prev == NULL)
479 			quit("rollforward: protocol botch\n");
480 		if (prev->dblk != 0)
481 			prev->count -= 1;
482 		else
483 			ntb--;
484 		q -= 1;
485 		q->count = 0;
486 		q = &tslp->req[0];
487 		if (i == 0) {
488 			q->dblk = 0;
489 			q->count = 1;
490 			trecno = 0;
491 			nextblock = tslp->tblock;
492 			savedtapea = spcl.c_tapea;
493 			spcl.c_tapea = slp->tapea;
494 			startnewtape(0);
495 			spcl.c_tapea = savedtapea;
496 			lastspclrec = savedtapea - 1;
497 		}
498 		size = (char *)ntb - (char *)q;
499 		if (atomic((ssize_t (*)(int, void *, size_t))write,
500 		    slp->fd, (char *)q, size) != size) {
501 			perror("  DUMP: error writing command pipe");
502 			dumpabort(0);
503 		}
504 		slp->sent = 1;
505 		if (++slp >= &slaves[SLAVES])
506 			slp = &slaves[0];
507 
508 		q->count = 1;
509 
510 		if (prev->dblk != 0) {
511 			/*
512 			 * If the last one was a disk block, make the
513 			 * first of this one be the last bit of that disk
514 			 * block...
515 			 */
516 			q->dblk = prev->dblk +
517 				prev->count * (TP_BSIZE / DEV_BSIZE);
518 			ntb = (union u_spcl *)tslp->tblock;
519 		} else {
520 			/*
521 			 * It wasn't a disk block.  Copy the data to its
522 			 * new location in the buffer.
523 			 */
524 			q->dblk = 0;
525 			*((union u_spcl *)tslp->tblock) = *ntb;
526 			ntb = (union u_spcl *)tslp->tblock[1];
527 		}
528 	}
529 	slp->req[0] = *q;
530 	nextblock = slp->tblock;
531 	if (q->dblk == 0)
532 		nextblock++;
533 	trecno = 1;
534 
535 	/*
536 	 * Clear the first slaves' response.  One hopes that it
537 	 * worked ok, otherwise the tape is much too short!
538 	 */
539 	if (slp->sent) {
540 		if (atomic(read, slp->fd, (char *)&got, sizeof(got))
541 		    != sizeof(got)) {
542 			perror("  DUMP: error reading command pipe in master");
543 			dumpabort(0);
544 		}
545 		slp->sent = 0;
546 
547 		if (got != writesize) {
548 			quit("EOT detected at start of the tape!\n");
549 		}
550 	}
551 }
552 
553 /*
554  * We implement taking and restoring checkpoints on the tape level.
555  * When each tape is opened, a new process is created by forking; this
556  * saves all of the necessary context in the parent.  The child
557  * continues the dump; the parent waits around, saving the context.
558  * If the child returns X_REWRITE, then it had problems writing that tape;
559  * this causes the parent to fork again, duplicating the context, and
560  * everything continues as if nothing had happened.
561  */
562 void
563 startnewtape(int top)
564 {
565 	pid_t	parentpid;
566 	pid_t	childpid;
567 	int	status;
568 	pid_t	waitingpid;
569 	char	*p;
570 	sig_t	interrupt_save;
571 
572 	interrupt_save = signal(SIGINT, SIG_IGN);
573 	parentpid = getpid();
574 	tapea_volume = spcl.c_tapea;
575 	(void)time(&tstart_volume);
576 
577 restore_check_point:
578 	(void)signal(SIGINT, interrupt_save);
579 	/*
580 	 *	All signals are inherited...
581 	 */
582 	childpid = fork();
583 	if (childpid == -1) {
584 		msg("Context save fork fails in parent %d\n", parentpid);
585 		Exit(X_ABORT);
586 	}
587 	if (childpid != 0) {
588 		/*
589 		 *	PARENT:
590 		 *	save the context by waiting
591 		 *	until the child doing all of the work returns.
592 		 *	don't catch the interrupt
593 		 */
594 		signal(SIGINT, SIG_IGN);
595 #ifdef TDEBUG
596 		msg("Tape: %d; parent process: %d child process %d\n",
597 			tapeno+1, parentpid, childpid);
598 #endif /* TDEBUG */
599 		while ((waitingpid = wait(&status)) != childpid)
600 			msg("Parent %d waiting for child %d has another child %d return\n",
601 				parentpid, childpid, waitingpid);
602 		if (status & 0xFF) {
603 			msg("Child %d returns LOB status %o\n",
604 				childpid, status&0xFF);
605 		}
606 		status = (status >> 8) & 0xFF;
607 #ifdef TDEBUG
608 		switch(status) {
609 			case X_FINOK:
610 				msg("Child %d finishes X_FINOK\n", childpid);
611 				break;
612 			case X_ABORT:
613 				msg("Child %d finishes X_ABORT\n", childpid);
614 				break;
615 			case X_REWRITE:
616 				msg("Child %d finishes X_REWRITE\n", childpid);
617 				break;
618 			default:
619 				msg("Child %d finishes unknown %d\n",
620 					childpid, status);
621 				break;
622 		}
623 #endif /* TDEBUG */
624 		switch(status) {
625 			case X_FINOK:
626 				Exit(X_FINOK);
627 				break;
628 			case X_ABORT:
629 				Exit(X_ABORT);
630 				break;
631 			case X_REWRITE:
632 				goto restore_check_point;
633 			default:
634 				msg("Bad return code from dump: %d\n", status);
635 				Exit(X_ABORT);
636 		}
637 		/*NOTREACHED*/
638 	} else {	/* we are the child; just continue */
639 #ifdef TDEBUG
640 		sleep(4);	/* allow time for parent's message to get out */
641 		msg("Child on Tape %d has parent %d, my pid = %d\n",
642 			tapeno+1, parentpid, getpid());
643 #endif /* TDEBUG */
644 		/*
645 		 * If we have a name like "/dev/rst0,/dev/rst1",
646 		 * use the name before the comma first, and save
647 		 * the remaining names for subsequent volumes.
648 		 */
649 		tapeno++;               /* current tape sequence */
650 		if (nexttape || strchr(tape, ',')) {
651 			if (nexttape && *nexttape)
652 				tape = nexttape;
653 			if ((p = strchr(tape, ',')) != NULL) {
654 				*p = '\0';
655 				nexttape = p + 1;
656 			} else
657 				nexttape = NULL;
658 			msg("Dumping volume %d on %s\n", tapeno, tape);
659 		}
660 #ifdef RDUMP
661 		while ((tapefd = (host ? rmtopen(tape, O_WRONLY|O_CREAT) :
662 			pipeout ? 1 : open(tape, O_WRONLY|O_CREAT, 0666))) == -1)
663 #else
664 		while ((tapefd = (pipeout ? 1 :
665 				  open(tape, O_WRONLY|O_CREAT, 0666))) == -1)
666 #endif
667 		    {
668 			msg("Cannot open output \"%s\".\n", tape);
669 			if (!query("Do you want to retry the open?"))
670 				dumpabort(0);
671 		}
672 
673 		enslave();  /* Share open tape file descriptor with slaves */
674 
675 		asize = 0;
676 		blocksthisvol = 0;
677 		if (top)
678 			newtape++;		/* new tape signal */
679 		spcl.c_count = slp->count;
680 		/*
681 		 * measure firstrec in TP_BSIZE units since restore doesn't
682 		 * know the correct ntrec value...
683 		 */
684 		spcl.c_firstrec = slp->firstrec;
685 		spcl.c_volume++;
686 		spcl.c_type = TS_TAPE;
687 		if (sblock->fs_magic != FS_UFS2_MAGIC)
688 			spcl.c_flags |= DR_NEWHEADER;
689 		writeheader((ino_t)slp->inode);
690 		if (sblock->fs_magic != FS_UFS2_MAGIC)
691 			spcl.c_flags &=~ DR_NEWHEADER;
692 		msg("Volume %d started at: %s", tapeno, ctime(&tstart_volume));
693 		if (tapeno > 1)
694 			msg("Volume %d begins with blocks from inode %llu\n",
695 			    tapeno, (unsigned long long)slp->inode);
696 	}
697 }
698 
699 /* ARGSUSED */
700 void
701 dumpabort(int signo)
702 {
703 
704 	if (master != 0 && master != getpid())
705 		/* Signals master to call dumpabort */
706 		(void) kill(master, SIGTERM);
707 	else {
708 		killall();
709 		msg("The ENTIRE dump is aborted.\n");
710 	}
711 #ifdef RDUMP
712 	rmtclose();
713 #endif
714 	Exit(X_ABORT);
715 }
716 
717 __dead void
718 Exit(int status)
719 {
720 
721 #ifdef TDEBUG
722 	msg("pid = %d exits with status %d\n", getpid(), status);
723 #endif /* TDEBUG */
724 	exit(status);
725 }
726 
727 /*
728  * proceed - handler for SIGUSR2, used to synchronize IO between the slaves.
729  */
730 /* ARGSUSED */
731 void
732 proceed(int signo)
733 {
734 	caught++;
735 }
736 
737 void
738 enslave(void)
739 {
740 	int cmd[2];
741 	int i, j;
742 
743 	master = getpid();
744 
745 	signal(SIGTERM, dumpabort);  /* Slave sends SIGTERM on dumpabort() */
746 	signal(SIGPIPE, sigpipe);
747 	signal(SIGUSR1, tperror);    /* Slave sends SIGUSR1 on tape errors */
748 	signal(SIGUSR2, proceed);    /* Slave sends SIGUSR2 to next slave */
749 
750 	for (i = 0; i < SLAVES; i++) {
751 		if (i == slp - &slaves[0]) {
752 			caught = 1;
753 		} else {
754 			caught = 0;
755 		}
756 
757 		if (socketpair(AF_UNIX, SOCK_STREAM, 0, cmd) == -1 ||
758 		    (slaves[i].pid = fork()) == -1)
759 			quit("too many slaves, %d (recompile smaller): %s\n",
760 			    i, strerror(errno));
761 
762 		slaves[i].fd = cmd[1];
763 		slaves[i].sent = 0;
764 		if (slaves[i].pid == 0) { 	    /* Slave starts up here */
765 			for (j = 0; j <= i; j++)
766 			        (void) close(slaves[j].fd);
767 			signal(SIGINT, SIG_IGN);    /* Master handles this */
768 			signal(SIGINFO, SIG_IGN);
769 			doslave(cmd[0], i);
770 			Exit(X_FINOK);
771 		}
772 	}
773 
774 	for (i = 0; i < SLAVES; i++)
775 		(void) atomic((ssize_t (*)(int, void *, size_t))write,
776 		    slaves[i].fd, (char *) &slaves[(i + 1) % SLAVES].pid,
777 		    sizeof(slaves[0].pid));
778 	master = 0;
779 }
780 
781 void
782 killall(void)
783 {
784 	int i;
785 
786 	for (i = 0; i < SLAVES; i++)
787 		if (slaves[i].pid > 0) {
788 			(void) kill(slaves[i].pid, SIGKILL);
789 			slaves[i].pid = 0;
790 		}
791 }
792 
793 /*
794  * Synchronization - each process has a lockfile, and shares file
795  * descriptors to the following process's lockfile.  When our write
796  * completes, we release our lock on the following process's lock-
797  * file, allowing the following process to lock it and proceed. We
798  * get the lock back for the next cycle by swapping descriptors.
799  */
800 static void
801 doslave(int cmd, int slave_number)
802 {
803 	int nread, nextslave, size, wrote = 0, eot_count;
804 	sigset_t nsigset, osigset;
805 
806 	/*
807 	 * Need our own seek pointer.
808 	 */
809 	(void) close(diskfd);
810 	if ((diskfd = open(disk, O_RDONLY)) == -1)
811 		quit("slave couldn't reopen disk: %s\n", strerror(errno));
812 
813 	/*
814 	 * Need the pid of the next slave in the loop...
815 	 */
816 	if ((nread = atomic(read, cmd, (char *)&nextslave, sizeof(nextslave)))
817 	    != sizeof(nextslave)) {
818 		quit("master/slave protocol botched - didn't get pid of next slave.\n");
819 	}
820 
821 	/*
822 	 * Get list of blocks to dump, read the blocks into tape buffer
823 	 */
824 	while ((nread = atomic(read, cmd, (char *)slp->req, reqsiz)) == reqsiz) {
825 		struct req *p = slp->req;
826 
827 		for (trecno = 0; trecno < ntrec;
828 		     trecno += p->count, p += p->count) {
829 			if (p->dblk) {
830 				bread(p->dblk, slp->tblock[trecno],
831 					p->count * TP_BSIZE);
832 			} else {
833 				if (p->count != 1 || atomic(read, cmd,
834 				    (char *)slp->tblock[trecno],
835 				    TP_BSIZE) != TP_BSIZE)
836 				       quit("master/slave protocol botched.\n");
837 			}
838 		}
839 
840 		sigemptyset(&nsigset);
841 		sigaddset(&nsigset, SIGUSR2);
842 		sigprocmask(SIG_BLOCK, &nsigset, &osigset);
843 		while (!caught)
844 			sigsuspend(&osigset);
845 		caught = 0;
846 		sigprocmask(SIG_SETMASK, &osigset, NULL);
847 
848 		/* Try to write the data... */
849 		eot_count = 0;
850 		size = 0;
851 
852 		while (eot_count < 10 && size < writesize) {
853 #ifdef RDUMP
854 			if (host)
855 				wrote = rmtwrite(slp->tblock[0]+size,
856 				    writesize-size);
857 			else
858 #endif
859 				wrote = write(tapefd, slp->tblock[0]+size,
860 				    writesize-size);
861 #ifdef WRITEDEBUG
862 			printf("slave %d wrote %d\n", slave_number, wrote);
863 #endif
864 			if (wrote < 0)
865 				break;
866 			if (wrote == 0)
867 				eot_count++;
868 			size += wrote;
869 		}
870 
871 #ifdef WRITEDEBUG
872 		if (size != writesize)
873 		 printf("slave %d only wrote %d out of %d bytes and gave up.\n",
874 		     slave_number, size, writesize);
875 #endif
876 
877 		if (eot_count > 0)
878 			size = 0;
879 
880 		/*
881 		 * Handle ENOSPC as an EOT condition
882 		 */
883 		if (wrote < 0 && errno == ENOSPC) {
884 			wrote = 0;
885 			eot_count++;
886 		}
887 
888 		if (size < 0) {
889 			(void) kill(master, SIGUSR1);
890 			sigemptyset(&nsigset);
891 			for (;;)
892 				sigsuspend(&nsigset);
893 		} else {
894 			/*
895 			 * pass size of write back to master
896 			 * (for EOT handling)
897 			 */
898 			(void) atomic((ssize_t (*)(int, void *, size_t))write,
899 			    cmd, (char *)&size, sizeof(size));
900 		}
901 
902 		/*
903 		 * If partial write, don't want next slave to go.
904 		 * Also jolts him awake.
905 		 */
906 		(void) kill(nextslave, SIGUSR2);
907 	}
908 	if (nread != 0)
909 		quit("error reading command pipe: %s\n", strerror(errno));
910 }
911 
912 /*
913  * Since a read from a pipe may not return all we asked for,
914  * or a write may not write all we ask if we get a signal,
915  * loop until the count is satisfied (or error).
916  */
917 static ssize_t
918 atomic(ssize_t (*func)(int, void *, size_t), int fd, char *buf, int count)
919 {
920 	ssize_t got, need = count;
921 
922 	while ((got = (*func)(fd, buf, need)) > 0 && (need -= got) > 0)
923 		buf += got;
924 	return (got < 0 ? got : count - need);
925 }
926