xref: /openbsd-src/sbin/dump/tape.c (revision fc405d53b73a2d73393cb97f684863d17b583e38)
1 /*	$OpenBSD: tape.c,v 1.48 2023/03/08 04:43:06 guenther 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 void
196 tperror(int signo)
197 {
198 	/* XXX - signal races */
199 
200 	if (pipeout) {
201 		msg("write error on %s\n", tape);
202 		quit("Cannot recover\n");
203 		/* NOTREACHED */
204 	}
205 	msg("write error %lld blocks into volume %d\n",
206 	    (long long)blocksthisvol, tapeno);
207 	broadcast("DUMP WRITE ERROR!\n");
208 	if (!query("Do you want to restart?"))
209 		dumpabort(0);
210 	msg("Closing this volume.  Prepare to restart with new media;\n");
211 	msg("this dump volume will be rewritten.\n");
212 	killall();
213 	nogripe = 1;
214 	close_rewind();
215 	Exit(X_REWRITE);
216 }
217 
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 void
257 statussig(int signo)
258 {
259 	time_t	tnow, deltat;
260 	int save_errno = errno;
261 
262 	if (blockswritten < 500)
263 		return;
264 	(void) time(&tnow);
265 	deltat = tstart_writing - tnow + (1.0 * (tnow - tstart_writing))
266 		/ blockswritten * tapesize;
267 	/* XXX not safe due to floating point printf */
268 	dprintf(STDERR_FILENO,
269 	    "dump: %s %3.2f%% done at %lld KB/s, finished in %d:%02d\n",
270 	    tape, (blockswritten * 100.0) / tapesize,
271 	    (spcl.c_tapea - tapea_volume) / (tnow - tstart_volume),
272 	    (int)(deltat / 3600), (int)((deltat % 3600) / 60));
273 	errno = save_errno;
274 }
275 
276 static void
277 flushtape(void)
278 {
279 	int i, blks, got;
280 	int64_t lastfirstrec;
281 
282 	int siz = (char *)nextblock - (char *)slp->req;
283 
284 	slp->req[trecno].count = 0;			/* Sentinel */
285 
286 	if (atomic((ssize_t (*)(int, void *, size_t))write, slp->fd,
287 	    (char *)slp->req, siz) != siz)
288 		quit("error writing command pipe: %s\n", strerror(errno));
289 	slp->sent = 1; /* we sent a request, read the response later */
290 
291 	lastfirstrec = slp->firstrec;
292 
293 	if (++slp >= &slaves[SLAVES])
294 		slp = &slaves[0];
295 
296 	/* Read results back from next slave */
297 	if (slp->sent) {
298 		if (atomic(read, slp->fd, (char *)&got, sizeof(got))
299 		    != sizeof(got)) {
300 			perror("  DUMP: error reading command pipe in master");
301 			dumpabort(0);
302 		}
303 		slp->sent = 0;
304 
305 		/* Check for end of tape */
306 		if (got < writesize) {
307 			msg("End of tape detected\n");
308 
309 			/*
310 			 * Drain the results, don't care what the values were.
311 			 * If we read them here then trewind won't...
312 			 */
313 			for (i = 0; i < SLAVES; i++) {
314 				if (slaves[i].sent) {
315 					if (atomic(read, slaves[i].fd,
316 					    (char *)&got, sizeof(got))
317 					    != sizeof(got)) {
318 						perror("  DUMP: error reading command pipe in master");
319 						dumpabort(0);
320 					}
321 					slaves[i].sent = 0;
322 				}
323 			}
324 
325 			close_rewind();
326 			rollforward();
327 			return;
328 		}
329 	}
330 
331 	blks = 0;
332 	if (spcl.c_type != TS_END && spcl.c_type != TS_CLRI &&
333 	    spcl.c_type != TS_BITS) {
334 		if (spcl.c_count > TP_NINDIR)
335 			quit("c_count too large\n");
336 		for (i = 0; i < spcl.c_count; i++)
337 			if (spcl.c_addr[i] != 0)
338 				blks++;
339 	}
340 	slp->count = lastspclrec + blks + 1 - spcl.c_tapea;
341 	slp->tapea = spcl.c_tapea;
342 	slp->firstrec = lastfirstrec + ntrec;
343 	slp->inode = curino;
344 	nextblock = slp->tblock;
345 	trecno = 0;
346 	asize += tenths;
347 	blockswritten += ntrec;
348 	blocksthisvol += ntrec;
349 	if (!pipeout && !unlimited && (blocksperfile ?
350 	    (blocksthisvol >= blocksperfile) : (asize > tsize))) {
351 		close_rewind();
352 		startnewtape(0);
353 	}
354 	timeest();
355 }
356 
357 void
358 trewind(void)
359 {
360 	struct stat sb;
361 	int f, got;
362 
363 	for (f = 0; f < SLAVES; f++) {
364 		/*
365 		 * Drain the results, but unlike EOT we DO (or should) care
366 		 * what the return values were, since if we detect EOT after
367 		 * we think we've written the last blocks to the tape anyway,
368 		 * we have to replay those blocks with rollforward.
369 		 *
370 		 * fixme: punt for now.
371 		 */
372 		if (slaves[f].sent) {
373 			if (atomic(read, slaves[f].fd, (char *)&got, sizeof(got))
374 			    != sizeof(got)) {
375 				perror("  DUMP: error reading command pipe in master");
376 				dumpabort(0);
377 			}
378 			slaves[f].sent = 0;
379 			if (got != writesize) {
380 				msg("EOT detected in last 2 tape records!\n");
381 				msg("Use a longer tape, decrease the size estimate\n");
382 				quit("or use no size estimate at all.\n");
383 			}
384 		}
385 		(void) close(slaves[f].fd);
386 	}
387 	while (wait((int *)NULL) >= 0)	/* wait for any signals from slaves */
388 		/* void */;
389 
390 	if (pipeout)
391 		return;
392 
393 	msg("Closing %s\n", tape);
394 
395 #ifdef RDUMP
396 	if (host) {
397 		rmtclose();
398 		while (rmtopen(tape, O_RDONLY) < 0)
399 			sleep(10);
400 		rmtclose();
401 		return;
402 	}
403 #endif
404 	/*
405 	 * st(4) says: "Bit 1 of the minor number specifies whether an eject is
406 	 * attempted when the device is closed.  When it is set, the device
407 	 * will attempt to eject its media on close ...".
408 	 *
409 	 * If the tape has been ejected, looping on open() will generate 'Media
410 	 * not present' errors until a tape is loaded. Once loaded the tape
411 	 * will be immediately ejected as a result of the second close().
412 	 *
413 	 * So if the tape will be ejected, just close and return.
414 	 */
415 	if ((fstat(tapefd, &sb) == 0) && (minor(sb.st_rdev) & 0x02)) {
416 		(void) close(tapefd);
417 		return;
418 	}
419 
420 	(void) close(tapefd);
421 	while ((f = open(tape, O_RDONLY)) == -1)
422 		sleep (10);
423 	(void) close(f);
424 }
425 
426 void
427 close_rewind(void)
428 {
429 	trewind();
430 	(void)do_stats();
431 	if (nexttape)
432 		return;
433 	if (!nogripe) {
434 		msg("Change Volumes: Mount volume #%d\n", tapeno+1);
435 		broadcast("CHANGE DUMP VOLUMES!\7\7\n");
436 	}
437 	while (!query("Is the new volume mounted and ready to go?"))
438 		if (query("Do you want to abort?")) {
439 			dumpabort(0);
440 			/*NOTREACHED*/
441 		}
442 }
443 
444 void
445 rollforward(void)
446 {
447 	struct req *p, *q, *prev;
448 	struct slave *tslp;
449 	int i, size, got;
450 	int64_t savedtapea;
451 	union u_spcl *ntb, *otb;
452 	tslp = &slaves[SLAVES];
453 	ntb = (union u_spcl *)tslp->tblock[1];
454 
455 	/*
456 	 * Each of the N slaves should have requests that need to
457 	 * be replayed on the next tape.  Use the extra slave buffers
458 	 * (slaves[SLAVES]) to construct request lists to be sent to
459 	 * each slave in turn.
460 	 */
461 	for (i = 0; i < SLAVES; i++) {
462 		q = &tslp->req[1];
463 		otb = (union u_spcl *)slp->tblock;
464 
465 		/*
466 		 * For each request in the current slave, copy it to tslp.
467 		 */
468 
469 		prev = NULL;
470 		for (p = slp->req; p->count > 0; p += p->count) {
471 			*q = *p;
472 			if (p->dblk == 0)
473 				*ntb++ = *otb++; /* copy the datablock also */
474 			prev = q;
475 			q += q->count;
476 		}
477 		if (prev == NULL)
478 			quit("rollforward: protocol botch\n");
479 		if (prev->dblk != 0)
480 			prev->count -= 1;
481 		else
482 			ntb--;
483 		q -= 1;
484 		q->count = 0;
485 		q = &tslp->req[0];
486 		if (i == 0) {
487 			q->dblk = 0;
488 			q->count = 1;
489 			trecno = 0;
490 			nextblock = tslp->tblock;
491 			savedtapea = spcl.c_tapea;
492 			spcl.c_tapea = slp->tapea;
493 			startnewtape(0);
494 			spcl.c_tapea = savedtapea;
495 			lastspclrec = savedtapea - 1;
496 		}
497 		size = (char *)ntb - (char *)q;
498 		if (atomic((ssize_t (*)(int, void *, size_t))write,
499 		    slp->fd, (char *)q, size) != size) {
500 			perror("  DUMP: error writing command pipe");
501 			dumpabort(0);
502 		}
503 		slp->sent = 1;
504 		if (++slp >= &slaves[SLAVES])
505 			slp = &slaves[0];
506 
507 		q->count = 1;
508 
509 		if (prev->dblk != 0) {
510 			/*
511 			 * If the last one was a disk block, make the
512 			 * first of this one be the last bit of that disk
513 			 * block...
514 			 */
515 			q->dblk = prev->dblk +
516 				prev->count * (TP_BSIZE / DEV_BSIZE);
517 			ntb = (union u_spcl *)tslp->tblock;
518 		} else {
519 			/*
520 			 * It wasn't a disk block.  Copy the data to its
521 			 * new location in the buffer.
522 			 */
523 			q->dblk = 0;
524 			*((union u_spcl *)tslp->tblock) = *ntb;
525 			ntb = (union u_spcl *)tslp->tblock[1];
526 		}
527 	}
528 	slp->req[0] = *q;
529 	nextblock = slp->tblock;
530 	if (q->dblk == 0)
531 		nextblock++;
532 	trecno = 1;
533 
534 	/*
535 	 * Clear the first slaves' response.  One hopes that it
536 	 * worked ok, otherwise the tape is much too short!
537 	 */
538 	if (slp->sent) {
539 		if (atomic(read, slp->fd, (char *)&got, sizeof(got))
540 		    != sizeof(got)) {
541 			perror("  DUMP: error reading command pipe in master");
542 			dumpabort(0);
543 		}
544 		slp->sent = 0;
545 
546 		if (got != writesize) {
547 			quit("EOT detected at start of the tape!\n");
548 		}
549 	}
550 }
551 
552 /*
553  * We implement taking and restoring checkpoints on the tape level.
554  * When each tape is opened, a new process is created by forking; this
555  * saves all of the necessary context in the parent.  The child
556  * continues the dump; the parent waits around, saving the context.
557  * If the child returns X_REWRITE, then it had problems writing that tape;
558  * this causes the parent to fork again, duplicating the context, and
559  * everything continues as if nothing had happened.
560  */
561 void
562 startnewtape(int top)
563 {
564 	pid_t	parentpid;
565 	pid_t	childpid;
566 	int	status;
567 	pid_t	waitingpid;
568 	char	*p;
569 	sig_t	interrupt_save;
570 
571 	interrupt_save = signal(SIGINT, SIG_IGN);
572 	parentpid = getpid();
573 	tapea_volume = spcl.c_tapea;
574 	(void)time(&tstart_volume);
575 
576 restore_check_point:
577 	(void)signal(SIGINT, interrupt_save);
578 	/*
579 	 *	All signals are inherited...
580 	 */
581 	childpid = fork();
582 	if (childpid == -1) {
583 		msg("Context save fork fails in parent %d\n", parentpid);
584 		Exit(X_ABORT);
585 	}
586 	if (childpid != 0) {
587 		/*
588 		 *	PARENT:
589 		 *	save the context by waiting
590 		 *	until the child doing all of the work returns.
591 		 *	don't catch the interrupt
592 		 */
593 		signal(SIGINT, SIG_IGN);
594 #ifdef TDEBUG
595 		msg("Tape: %d; parent process: %d child process %d\n",
596 			tapeno+1, parentpid, childpid);
597 #endif /* TDEBUG */
598 		while ((waitingpid = wait(&status)) != childpid)
599 			msg("Parent %d waiting for child %d has another child %d return\n",
600 				parentpid, childpid, waitingpid);
601 		if (status & 0xFF) {
602 			msg("Child %d returns LOB status %o\n",
603 				childpid, status&0xFF);
604 		}
605 		status = (status >> 8) & 0xFF;
606 #ifdef TDEBUG
607 		switch(status) {
608 			case X_FINOK:
609 				msg("Child %d finishes X_FINOK\n", childpid);
610 				break;
611 			case X_ABORT:
612 				msg("Child %d finishes X_ABORT\n", childpid);
613 				break;
614 			case X_REWRITE:
615 				msg("Child %d finishes X_REWRITE\n", childpid);
616 				break;
617 			default:
618 				msg("Child %d finishes unknown %d\n",
619 					childpid, status);
620 				break;
621 		}
622 #endif /* TDEBUG */
623 		switch(status) {
624 			case X_FINOK:
625 				Exit(X_FINOK);
626 				break;
627 			case X_ABORT:
628 				Exit(X_ABORT);
629 				break;
630 			case X_REWRITE:
631 				goto restore_check_point;
632 			default:
633 				msg("Bad return code from dump: %d\n", status);
634 				Exit(X_ABORT);
635 		}
636 		/*NOTREACHED*/
637 	} else {	/* we are the child; just continue */
638 #ifdef TDEBUG
639 		sleep(4);	/* allow time for parent's message to get out */
640 		msg("Child on Tape %d has parent %d, my pid = %d\n",
641 			tapeno+1, parentpid, getpid());
642 #endif /* TDEBUG */
643 		/*
644 		 * If we have a name like "/dev/rst0,/dev/rst1",
645 		 * use the name before the comma first, and save
646 		 * the remaining names for subsequent volumes.
647 		 */
648 		tapeno++;               /* current tape sequence */
649 		if (nexttape || strchr(tape, ',')) {
650 			if (nexttape && *nexttape)
651 				tape = nexttape;
652 			if ((p = strchr(tape, ',')) != NULL) {
653 				*p = '\0';
654 				nexttape = p + 1;
655 			} else
656 				nexttape = NULL;
657 			msg("Dumping volume %d on %s\n", tapeno, tape);
658 		}
659 #ifdef RDUMP
660 		while ((tapefd = (host ? rmtopen(tape, O_WRONLY|O_CREAT) :
661 			pipeout ? 1 : open(tape, O_WRONLY|O_CREAT, 0666))) == -1)
662 #else
663 		while ((tapefd = (pipeout ? 1 :
664 				  open(tape, O_WRONLY|O_CREAT, 0666))) == -1)
665 #endif
666 		    {
667 			msg("Cannot open output \"%s\".\n", tape);
668 			if (!query("Do you want to retry the open?"))
669 				dumpabort(0);
670 		}
671 
672 		enslave();  /* Share open tape file descriptor with slaves */
673 
674 		asize = 0;
675 		blocksthisvol = 0;
676 		if (top)
677 			newtape++;		/* new tape signal */
678 		spcl.c_count = slp->count;
679 		/*
680 		 * measure firstrec in TP_BSIZE units since restore doesn't
681 		 * know the correct ntrec value...
682 		 */
683 		spcl.c_firstrec = slp->firstrec;
684 		spcl.c_volume++;
685 		spcl.c_type = TS_TAPE;
686 		if (sblock->fs_magic != FS_UFS2_MAGIC)
687 			spcl.c_flags |= DR_NEWHEADER;
688 		writeheader((ino_t)slp->inode);
689 		if (sblock->fs_magic != FS_UFS2_MAGIC)
690 			spcl.c_flags &=~ DR_NEWHEADER;
691 		msg("Volume %d started at: %s", tapeno, ctime(&tstart_volume));
692 		if (tapeno > 1)
693 			msg("Volume %d begins with blocks from inode %llu\n",
694 			    tapeno, (unsigned long long)slp->inode);
695 	}
696 }
697 
698 void
699 dumpabort(int signo)
700 {
701 
702 	if (master != 0 && master != getpid())
703 		/* Signals master to call dumpabort */
704 		(void) kill(master, SIGTERM);
705 	else {
706 		killall();
707 		msg("The ENTIRE dump is aborted.\n");
708 	}
709 #ifdef RDUMP
710 	rmtclose();
711 #endif
712 	Exit(X_ABORT);
713 }
714 
715 __dead void
716 Exit(int status)
717 {
718 
719 #ifdef TDEBUG
720 	msg("pid = %d exits with status %d\n", getpid(), status);
721 #endif /* TDEBUG */
722 	exit(status);
723 }
724 
725 /*
726  * proceed - handler for SIGUSR2, used to synchronize IO between the slaves.
727  */
728 void
729 proceed(int signo)
730 {
731 	caught++;
732 }
733 
734 void
735 enslave(void)
736 {
737 	int cmd[2];
738 	int i, j;
739 
740 	master = getpid();
741 
742 	signal(SIGTERM, dumpabort);  /* Slave sends SIGTERM on dumpabort() */
743 	signal(SIGPIPE, sigpipe);
744 	signal(SIGUSR1, tperror);    /* Slave sends SIGUSR1 on tape errors */
745 	signal(SIGUSR2, proceed);    /* Slave sends SIGUSR2 to next slave */
746 
747 	for (i = 0; i < SLAVES; i++) {
748 		if (i == slp - &slaves[0]) {
749 			caught = 1;
750 		} else {
751 			caught = 0;
752 		}
753 
754 		if (socketpair(AF_UNIX, SOCK_STREAM, 0, cmd) == -1 ||
755 		    (slaves[i].pid = fork()) == -1)
756 			quit("too many slaves, %d (recompile smaller): %s\n",
757 			    i, strerror(errno));
758 
759 		slaves[i].fd = cmd[1];
760 		slaves[i].sent = 0;
761 		if (slaves[i].pid == 0) { 	    /* Slave starts up here */
762 			for (j = 0; j <= i; j++)
763 			        (void) close(slaves[j].fd);
764 			signal(SIGINT, SIG_IGN);    /* Master handles this */
765 			signal(SIGINFO, SIG_IGN);
766 			doslave(cmd[0], i);
767 			Exit(X_FINOK);
768 		}
769 	}
770 
771 	for (i = 0; i < SLAVES; i++)
772 		(void) atomic((ssize_t (*)(int, void *, size_t))write,
773 		    slaves[i].fd, (char *) &slaves[(i + 1) % SLAVES].pid,
774 		    sizeof(slaves[0].pid));
775 	master = 0;
776 }
777 
778 void
779 killall(void)
780 {
781 	int i;
782 
783 	for (i = 0; i < SLAVES; i++)
784 		if (slaves[i].pid > 0) {
785 			(void) kill(slaves[i].pid, SIGKILL);
786 			slaves[i].pid = 0;
787 		}
788 }
789 
790 /*
791  * Synchronization - each process has a lockfile, and shares file
792  * descriptors to the following process's lockfile.  When our write
793  * completes, we release our lock on the following process's lock-
794  * file, allowing the following process to lock it and proceed. We
795  * get the lock back for the next cycle by swapping descriptors.
796  */
797 static void
798 doslave(int cmd, int slave_number)
799 {
800 	int nread, nextslave, size, wrote = 0, eot_count;
801 	sigset_t nsigset, osigset;
802 
803 	/*
804 	 * Need our own seek pointer.
805 	 */
806 	(void) close(diskfd);
807 	if ((diskfd = open(disk, O_RDONLY)) == -1)
808 		quit("slave couldn't reopen disk: %s\n", strerror(errno));
809 
810 	/*
811 	 * Need the pid of the next slave in the loop...
812 	 */
813 	if ((nread = atomic(read, cmd, (char *)&nextslave, sizeof(nextslave)))
814 	    != sizeof(nextslave)) {
815 		quit("master/slave protocol botched - didn't get pid of next slave.\n");
816 	}
817 
818 	/*
819 	 * Get list of blocks to dump, read the blocks into tape buffer
820 	 */
821 	while ((nread = atomic(read, cmd, (char *)slp->req, reqsiz)) == reqsiz) {
822 		struct req *p = slp->req;
823 
824 		for (trecno = 0; trecno < ntrec;
825 		     trecno += p->count, p += p->count) {
826 			if (p->dblk) {
827 				bread(p->dblk, slp->tblock[trecno],
828 					p->count * TP_BSIZE);
829 			} else {
830 				if (p->count != 1 || atomic(read, cmd,
831 				    (char *)slp->tblock[trecno],
832 				    TP_BSIZE) != TP_BSIZE)
833 				       quit("master/slave protocol botched.\n");
834 			}
835 		}
836 
837 		sigemptyset(&nsigset);
838 		sigaddset(&nsigset, SIGUSR2);
839 		sigprocmask(SIG_BLOCK, &nsigset, &osigset);
840 		while (!caught)
841 			sigsuspend(&osigset);
842 		caught = 0;
843 		sigprocmask(SIG_SETMASK, &osigset, NULL);
844 
845 		/* Try to write the data... */
846 		eot_count = 0;
847 		size = 0;
848 
849 		while (eot_count < 10 && size < writesize) {
850 #ifdef RDUMP
851 			if (host)
852 				wrote = rmtwrite(slp->tblock[0]+size,
853 				    writesize-size);
854 			else
855 #endif
856 				wrote = write(tapefd, slp->tblock[0]+size,
857 				    writesize-size);
858 #ifdef WRITEDEBUG
859 			printf("slave %d wrote %d\n", slave_number, wrote);
860 #endif
861 			if (wrote < 0)
862 				break;
863 			if (wrote == 0)
864 				eot_count++;
865 			size += wrote;
866 		}
867 
868 #ifdef WRITEDEBUG
869 		if (size != writesize)
870 		 printf("slave %d only wrote %d out of %d bytes and gave up.\n",
871 		     slave_number, size, writesize);
872 #endif
873 
874 		if (eot_count > 0)
875 			size = 0;
876 
877 		/*
878 		 * Handle ENOSPC as an EOT condition
879 		 */
880 		if (wrote < 0 && errno == ENOSPC) {
881 			wrote = 0;
882 			eot_count++;
883 		}
884 
885 		if (size < 0) {
886 			(void) kill(master, SIGUSR1);
887 			sigemptyset(&nsigset);
888 			for (;;)
889 				sigsuspend(&nsigset);
890 		} else {
891 			/*
892 			 * pass size of write back to master
893 			 * (for EOT handling)
894 			 */
895 			(void) atomic((ssize_t (*)(int, void *, size_t))write,
896 			    cmd, (char *)&size, sizeof(size));
897 		}
898 
899 		/*
900 		 * If partial write, don't want next slave to go.
901 		 * Also jolts him awake.
902 		 */
903 		(void) kill(nextslave, SIGUSR2);
904 	}
905 	if (nread != 0)
906 		quit("error reading command pipe: %s\n", strerror(errno));
907 }
908 
909 /*
910  * Since a read from a pipe may not return all we asked for,
911  * or a write may not write all we ask if we get a signal,
912  * loop until the count is satisfied (or error).
913  */
914 static ssize_t
915 atomic(ssize_t (*func)(int, void *, size_t), int fd, char *buf, int count)
916 {
917 	ssize_t got, need = count;
918 
919 	while ((got = (*func)(fd, buf, need)) > 0 && (need -= got) > 0)
920 		buf += got;
921 	return (got < 0 ? got : count - need);
922 }
923