xref: /netbsd-src/sbin/dump/tape.c (revision d710132b4b8ce7f7cccaaf660cb16aa16b4077a0)
1 /*	$NetBSD: tape.c,v 1.40 2003/04/02 10:39:24 fvdl Exp $	*/
2 
3 /*-
4  * Copyright (c) 1980, 1991, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. All advertising materials mentioning features or use of this software
16  *    must display the following acknowledgement:
17  *	This product includes software developed by the University of
18  *	California, Berkeley and its contributors.
19  * 4. Neither the name of the University nor the names of its contributors
20  *    may be used to endorse or promote products derived from this software
21  *    without specific prior written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
27  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33  * SUCH DAMAGE.
34  */
35 
36 #include <sys/cdefs.h>
37 #ifndef lint
38 #if 0
39 static char sccsid[] = "@(#)tape.c	8.4 (Berkeley) 5/1/95";
40 #else
41 __RCSID("$NetBSD: tape.c,v 1.40 2003/04/02 10:39:24 fvdl Exp $");
42 #endif
43 #endif /* not lint */
44 
45 #include <sys/param.h>
46 #include <sys/socket.h>
47 #include <sys/time.h>
48 #include <sys/wait.h>
49 #include <ufs/ufs/dinode.h>
50 #include <sys/ioctl.h>
51 #include <sys/mtio.h>
52 
53 #include <protocols/dumprestore.h>
54 
55 #include <errno.h>
56 #include <fcntl.h>
57 #include <setjmp.h>
58 #include <signal.h>
59 #include <stdio.h>
60 #include <stdlib.h>
61 #include <string.h>
62 #include <time.h>
63 #include <unistd.h>
64 
65 #include "dump.h"
66 #include "pathnames.h"
67 
68 int	writesize;		/* size of malloc()ed buffer for tape */
69 int64_t	lastspclrec = -1;	/* tape block number of last written header */
70 int	trecno = 0;		/* next record to write in current block */
71 extern	long blocksperfile;	/* number of blocks per output file */
72 long	blocksthisvol;		/* number of blocks on current output file */
73 extern	int ntrec;		/* blocking factor on tape */
74 extern	int cartridge;
75 extern	char *host;
76 char	*nexttape;
77 
78 static	ssize_t atomic_read(int, char *, int);
79 static	ssize_t atomic_write(int, char *, int);
80 static	void doslave(int, int);
81 static	void enslave(void);
82 static	void flushtape(void);
83 static	void killall(void);
84 static	void proceed(int);
85 static	void rollforward(void);
86 static	void sigpipe(int);
87 static	void tperror(int);
88 
89 /*
90  * Concurrent dump mods (Caltech) - disk block reading and tape writing
91  * are exported to several slave processes.  While one slave writes the
92  * tape, the others read disk blocks; they pass control of the tape in
93  * a ring via signals. The parent process traverses the file system and
94  * sends writeheader()'s and lists of daddr's to the slaves via pipes.
95  * The following structure defines the instruction packets sent to slaves.
96  */
97 struct req {
98 	daddr_t dblk;
99 	int count;
100 };
101 int reqsiz;
102 
103 #define SLAVES 3		/* 1 slave writing, 1 reading, 1 for slack */
104 struct slave {
105 	int64_t tapea;		/* header number at start of this chunk */
106 	int64_t firstrec;	/* record number of this block */
107 	int count;		/* count to next header (used for TS_TAPE */
108 				/* after EOT) */
109 	int inode;		/* inode that we are currently dealing with */
110 	int fd;			/* FD for this slave */
111 	int pid;		/* PID for this slave */
112 	int sent;		/* 1 == we've sent this slave requests */
113 	char (*tblock)[TP_BSIZE]; /* buffer for data blocks */
114 	struct req *req;	/* buffer for requests */
115 } slaves[SLAVES+1];
116 struct slave *slp;
117 
118 char	(*nextblock)[TP_BSIZE];
119 
120 static time_t tstart_volume;	/* time of volume start */
121 static int64_t tapea_volume;	/* value of spcl.c_tapea at volume start */
122 
123 int master;		/* pid of master, for sending error signals */
124 int tenths;		/* length of tape used per block written */
125 static int caught;	/* have we caught the signal to proceed? */
126 static int ready;	/* have we reached the lock point without having */
127 			/* received the SIGUSR2 signal from the prev slave? */
128 static jmp_buf jmpbuf;	/* where to jump to if we are ready when the */
129 			/* SIGUSR2 arrives from the previous slave */
130 
131 int
132 alloctape(void)
133 {
134 	int pgoff = getpagesize() - 1;
135 	char *buf;
136 	int i;
137 
138 	writesize = ntrec * TP_BSIZE;
139 	reqsiz = (ntrec + 1) * sizeof(struct req);
140 	/*
141 	 * CDC 92181's and 92185's make 0.8" gaps in 1600-bpi start/stop mode
142 	 * (see DEC TU80 User's Guide).  The shorter gaps of 6250-bpi require
143 	 * repositioning after stopping, i.e, streaming mode, where the gap is
144 	 * variable, 0.30" to 0.45".  The gap is maximal when the tape stops.
145 	 */
146 	if (blocksperfile == 0 && !unlimited)
147 		tenths = writesize / density +
148 		    (cartridge ? 16 : density == 625 ? 5 : 8);
149 	/*
150 	 * Allocate tape buffer contiguous with the array of instruction
151 	 * packets, so flushtape() can write them together with one write().
152 	 * Align tape buffer on page boundary to speed up tape write().
153 	 */
154 	for (i = 0; i <= SLAVES; i++) {
155 		buf = (char *)
156 		    xmalloc((unsigned)(reqsiz + writesize + pgoff + TP_BSIZE));
157 		slaves[i].tblock = (char (*)[TP_BSIZE])
158 		    (((long)&buf[ntrec + 1] + pgoff) &~ pgoff);
159 		slaves[i].req = (struct req *)slaves[i].tblock - ntrec - 1;
160 	}
161 	slp = &slaves[0];
162 	slp->count = 1;
163 	slp->tapea = 0;
164 	slp->firstrec = 0;
165 	nextblock = slp->tblock;
166 	return(1);
167 }
168 
169 void
170 writerec(char *dp, int isspcl)
171 {
172 
173 	slp->req[trecno].dblk = (daddr_t)0;
174 	slp->req[trecno].count = 1;
175 	*(union u_spcl *)(*(nextblock)++) = *(union u_spcl *)dp;
176 	if (isspcl)
177 		lastspclrec = iswap64(spcl.c_tapea);
178 	trecno++;
179 	spcl.c_tapea = iswap64(iswap64(spcl.c_tapea) +1);
180 	if (trecno >= ntrec)
181 		flushtape();
182 }
183 
184 void
185 dumpblock(daddr_t blkno, int size)
186 {
187 	int avail, tpblks;
188 	daddr_t dblkno;
189 
190 	dblkno = fsatoda(ufsib, blkno);
191 	tpblks = size >> tp_bshift;
192 	while ((avail = MIN(tpblks, ntrec - trecno)) > 0) {
193 		slp->req[trecno].dblk = dblkno;
194 		slp->req[trecno].count = avail;
195 		trecno += avail;
196 		spcl.c_tapea = iswap64(iswap64(spcl.c_tapea) + avail);
197 		if (trecno >= ntrec)
198 			flushtape();
199 		dblkno += avail << (tp_bshift - dev_bshift);
200 		tpblks -= avail;
201 	}
202 }
203 
204 int	nogripe = 0;
205 
206 static void
207 tperror(int signo)
208 {
209 
210 	if (pipeout) {
211 		msg("write error on %s\n", tape);
212 		quit("Cannot recover\n");
213 		/* NOTREACHED */
214 	}
215 	msg("write error %ld blocks into volume %d\n", blocksthisvol, tapeno);
216 	broadcast("DUMP WRITE ERROR!\n");
217 	if (!query("Do you want to restart?"))
218 		dumpabort(0);
219 	msg("Closing this volume.  Prepare to restart with new media;\n");
220 	msg("this dump volume will be rewritten.\n");
221 	killall();
222 	nogripe = 1;
223 	close_rewind();
224 	Exit(X_REWRITE);
225 }
226 
227 static void
228 sigpipe(int signo)
229 {
230 
231 	quit("Broken pipe\n");
232 }
233 
234 /*
235  * do_stats --
236  *	Update xferrate stats
237  */
238 time_t
239 do_stats(void)
240 {
241 	time_t tnow, ttaken;
242 	int64_t blocks;
243 
244 	(void)time(&tnow);
245 	ttaken = tnow - tstart_volume;
246 	blocks = iswap64(spcl.c_tapea) - tapea_volume;
247 	msg("Volume %d completed at: %s", tapeno, ctime(&tnow));
248 	if (ttaken > 0) {
249 		msg("Volume %d took %d:%02d:%02d\n", tapeno,
250 		    (int) (ttaken / 3600), (int) ((ttaken % 3600) / 60),
251 		    (int) (ttaken % 60));
252 		msg("Volume %d transfer rate: %d KB/s\n", tapeno,
253 		    (int) (blocks / ttaken));
254 		xferrate += blocks / ttaken;
255 	}
256 	return(tnow);
257 }
258 
259 /*
260  * statussig --
261  *	information message upon receipt of SIGINFO
262  *	(derived from optr.c::timeest())
263  */
264 void
265 statussig(int notused)
266 {
267 	time_t	tnow, deltat;
268 	char	msgbuf[128];
269 	int	errno_save;
270 
271 	if (blockswritten < 500)
272 		return;
273 	errno_save = errno;
274 	(void) time((time_t *) &tnow);
275 	if (tnow <= tstart_volume)
276 		return;
277 	deltat = tstart_writing - tnow +
278 	    (1.0 * (tnow - tstart_writing)) / blockswritten * tapesize;
279 	(void)snprintf(msgbuf, sizeof(msgbuf),
280 	    "%3.2f%% done at %ld KB/s, finished in %d:%02d\n",
281 	    (blockswritten * 100.0) / tapesize,
282 	    (long)((iswap64(spcl.c_tapea) - tapea_volume) /
283 	      (tnow - tstart_volume)),
284 	    (int)(deltat / 3600), (int)((deltat % 3600) / 60));
285 	write(STDERR_FILENO, msgbuf, strlen(msgbuf));
286 	errno = errno_save;
287 }
288 
289 static void
290 flushtape(void)
291 {
292 	int i, blks, got;
293 	int64_t lastfirstrec;
294 
295 	int siz = (char *)nextblock - (char *)slp->req;
296 
297 	slp->req[trecno].count = 0;			/* Sentinel */
298 
299 	if (atomic_write(slp->fd, (char *)slp->req, siz) != siz)
300 		quit("error writing command pipe: %s\n", strerror(errno));
301 	slp->sent = 1; /* we sent a request, read the response later */
302 
303 	lastfirstrec = slp->firstrec;
304 
305 	if (++slp >= &slaves[SLAVES])
306 		slp = &slaves[0];
307 
308 	/* Read results back from next slave */
309 	if (slp->sent) {
310 		if (atomic_read(slp->fd, (char *)&got, sizeof got)
311 		    != sizeof got) {
312 			perror("  DUMP: error reading command pipe in master");
313 			dumpabort(0);
314 		}
315 		slp->sent = 0;
316 
317 		/* Check for end of tape */
318 		if (got < writesize) {
319 			msg("End of tape detected\n");
320 
321 			/*
322 			 * Drain the results, don't care what the values were.
323 			 * If we read them here then trewind won't...
324 			 */
325 			for (i = 0; i < SLAVES; i++) {
326 				if (slaves[i].sent) {
327 					if (atomic_read(slaves[i].fd,
328 					    (char *)&got, sizeof got)
329 					    != sizeof got) {
330 						perror("  DUMP: error reading command pipe in master");
331 						dumpabort(0);
332 					}
333 					slaves[i].sent = 0;
334 				}
335 			}
336 
337 			close_rewind();
338 			rollforward();
339 			return;
340 		}
341 	}
342 
343 	blks = 0;
344 	if (iswap32(spcl.c_type) != TS_END) {
345 		for (i = 0; i < iswap32(spcl.c_count); i++)
346 			if (spcl.c_addr[i] != 0)
347 				blks++;
348 	}
349 	slp->count = lastspclrec + blks + 1 - iswap32(spcl.c_tapea);
350 	slp->tapea = iswap64(spcl.c_tapea);
351 	slp->firstrec = lastfirstrec + ntrec;
352 	slp->inode = curino;
353 	nextblock = slp->tblock;
354 	trecno = 0;
355 	asize += tenths;
356 	blockswritten += ntrec;
357 	blocksthisvol += ntrec;
358 	if (!pipeout && !unlimited && (blocksperfile ?
359 	    (blocksthisvol >= blocksperfile) : (asize > tsize))) {
360 		close_rewind();
361 		startnewtape(0);
362 	}
363 	timeest();
364 }
365 
366 void
367 trewind(int eject)
368 {
369 	int f;
370 	int got;
371 
372 	for (f = 0; f < SLAVES; f++) {
373 		/*
374 		 * Drain the results, but unlike EOT we DO (or should) care
375 		 * what the return values were, since if we detect EOT after
376 		 * we think we've written the last blocks to the tape anyway,
377 		 * we have to replay those blocks with rollforward.
378 		 *
379 		 * fixme: punt for now.
380 		 */
381 		if (slaves[f].sent) {
382 			if (atomic_read(slaves[f].fd, (char *)&got, sizeof got)
383 			    != sizeof got) {
384 				perror("  DUMP: error reading command pipe in master");
385 				dumpabort(0);
386 			}
387 			slaves[f].sent = 0;
388 			if (got != writesize) {
389 				msg("EOT detected in last 2 tape records!\n");
390 				msg("Use a longer tape, decrease the size estimate\n");
391 				quit("or use no size estimate at all.\n");
392 			}
393 		}
394 		(void) close(slaves[f].fd);
395 	}
396 	while (wait((int *)NULL) >= 0)	/* wait for any signals from slaves */
397 		/* void */;
398 
399 	if (pipeout)
400 		return;
401 
402 	msg("Closing %s\n", tape);
403 
404 #ifdef RDUMP
405 	if (host) {
406 		rmtclose();
407 		while (rmtopen(tape, 0, 0) < 0)
408 			sleep(10);
409 		if (eflag && eject) {
410 			msg("Ejecting %s\n", tape);
411 			(void) rmtioctl(MTOFFL, 0);
412 		}
413 		rmtclose();
414 		return;
415 	}
416 #endif
417 	(void) close(tapefd);
418 	while ((f = open(tape, 0)) < 0)
419 		sleep (10);
420 	if (eflag && eject) {
421 		struct mtop offl;
422 
423 		msg("Ejecting %s\n", tape);
424 		offl.mt_op = MTOFFL;
425 		offl.mt_count = 0;
426 		(void) ioctl(f, MTIOCTOP, &offl);
427 	}
428 	(void) close(f);
429 }
430 
431 void
432 close_rewind(void)
433 {
434 	int i, f;
435 
436 	trewind(1);
437 	(void)do_stats();
438 	if (nexttape)
439 		return;
440 	if (!nogripe) {
441 		msg("Change Volumes: Mount volume #%d\n", tapeno+1);
442 		broadcast("CHANGE DUMP VOLUMES!\a\a\n");
443 	}
444 	if (lflag) {
445 		for (i = 0; i < lflag / 10; i++) { /* wait lflag seconds */
446 			if (host) {
447 				if (rmtopen(tape, 0, 0) >= 0) {
448 					rmtclose();
449 					return;
450 				}
451 			} else {
452 				if ((f = open(tape, 0)) >= 0) {
453 					close(f);
454 					return;
455 				}
456 			}
457 			sleep (10);
458 		}
459 	}
460 
461 	while (!query("Is the new volume mounted and ready to go?"))
462 		if (query("Do you want to abort?")) {
463 			dumpabort(0);
464 			/*NOTREACHED*/
465 		}
466 }
467 
468 void
469 rollforward(void)
470 {
471 	struct req *p, *q, *prev;
472 	struct slave *tslp;
473 	int i, size, savedtapea, got;
474 	union u_spcl *ntb, *otb;
475 	tslp = &slaves[SLAVES];
476 	ntb = (union u_spcl *)tslp->tblock[1];
477 
478 	/*
479 	 * Each of the N slaves should have requests that need to
480 	 * be replayed on the next tape.  Use the extra slave buffers
481 	 * (slaves[SLAVES]) to construct request lists to be sent to
482 	 * each slave in turn.
483 	 */
484 	for (i = 0; i < SLAVES; i++) {
485 		q = &tslp->req[1];
486 		otb = (union u_spcl *)slp->tblock;
487 
488 		/*
489 		 * For each request in the current slave, copy it to tslp.
490 		 */
491 
492 		prev = NULL;
493 		for (p = slp->req; p->count > 0; p += p->count) {
494 			*q = *p;
495 			if (p->dblk == 0)
496 				*ntb++ = *otb++; /* copy the datablock also */
497 			prev = q;
498 			q += q->count;
499 		}
500 		if (prev == NULL)
501 			quit("rollforward: protocol botch");
502 		if (prev->dblk != 0)
503 			prev->count -= 1;
504 		else
505 			ntb--;
506 		q -= 1;
507 		q->count = 0;
508 		q = &tslp->req[0];
509 		if (i == 0) {
510 			q->dblk = 0;
511 			q->count = 1;
512 			trecno = 0;
513 			nextblock = tslp->tblock;
514 			savedtapea = iswap32(spcl.c_tapea);
515 			spcl.c_tapea = iswap32(slp->tapea);
516 			startnewtape(0);
517 			spcl.c_tapea = iswap32(savedtapea);
518 			lastspclrec = savedtapea - 1;
519 		}
520 		size = (char *)ntb - (char *)q;
521 		if (atomic_write(slp->fd, (char *)q, size) != size) {
522 			perror("  DUMP: error writing command pipe");
523 			dumpabort(0);
524 		}
525 		slp->sent = 1;
526 		if (++slp >= &slaves[SLAVES])
527 			slp = &slaves[0];
528 
529 		q->count = 1;
530 
531 		if (prev->dblk != 0) {
532 			/*
533 			 * If the last one was a disk block, make the
534 			 * first of this one be the last bit of that disk
535 			 * block...
536 			 */
537 			q->dblk = prev->dblk +
538 				prev->count * (TP_BSIZE / DEV_BSIZE);
539 			ntb = (union u_spcl *)tslp->tblock;
540 		} else {
541 			/*
542 			 * It wasn't a disk block.  Copy the data to its
543 			 * new location in the buffer.
544 			 */
545 			q->dblk = 0;
546 			*((union u_spcl *)tslp->tblock) = *ntb;
547 			ntb = (union u_spcl *)tslp->tblock[1];
548 		}
549 	}
550 	slp->req[0] = *q;
551 	nextblock = slp->tblock;
552 	if (q->dblk == 0)
553 		nextblock++;
554 	trecno = 1;
555 
556 	/*
557 	 * Clear the first slaves' response.  One hopes that it
558 	 * worked ok, otherwise the tape is much too short!
559 	 */
560 	if (slp->sent) {
561 		if (atomic_read(slp->fd, (char *)&got, sizeof got)
562 		    != sizeof got) {
563 			perror("  DUMP: error reading command pipe in master");
564 			dumpabort(0);
565 		}
566 		slp->sent = 0;
567 
568 		if (got != writesize) {
569 			quit("EOT detected at start of the tape!\n");
570 		}
571 	}
572 }
573 
574 /*
575  * We implement taking and restoring checkpoints on the tape level.
576  * When each tape is opened, a new process is created by forking; this
577  * saves all of the necessary context in the parent.  The child
578  * continues the dump; the parent waits around, saving the context.
579  * If the child returns X_REWRITE, then it had problems writing that tape;
580  * this causes the parent to fork again, duplicating the context, and
581  * everything continues as if nothing had happened.
582  */
583 void
584 startnewtape(int top)
585 {
586 	int	parentpid;
587 	int	childpid;
588 	int	status;
589 	int	waitforpid;
590 	char	*p;
591 	sig_t	interrupt_save;
592 
593 	interrupt_save = signal(SIGINT, SIG_IGN);
594 	parentpid = getpid();
595 	tapea_volume = iswap32(spcl.c_tapea);
596 	(void)time(&tstart_volume);
597 
598 restore_check_point:
599 	(void)signal(SIGINT, interrupt_save);
600 	/*
601 	 *	All signals are inherited...
602 	 */
603 	childpid = fork();
604 	if (childpid < 0) {
605 		msg("Context save fork fails in parent %d\n", parentpid);
606 		Exit(X_ABORT);
607 	}
608 	if (childpid != 0) {
609 		/*
610 		 *	PARENT:
611 		 *	save the context by waiting
612 		 *	until the child doing all of the work returns.
613 		 *	don't catch the interrupt
614 		 */
615 		signal(SIGINT, SIG_IGN);
616 		signal(SIGINFO, SIG_IGN);	/* only want child's stats */
617 #ifdef TDEBUG
618 		msg("Tape: %d; parent process: %d child process %d\n",
619 			tapeno+1, parentpid, childpid);
620 #endif /* TDEBUG */
621 		while ((waitforpid = wait(&status)) != childpid)
622 			msg("Parent %d waiting for child %d has another child %d return\n",
623 				parentpid, childpid, waitforpid);
624 		if (status & 0xFF) {
625 			msg("Child %d returns LOB status %o\n",
626 				childpid, status&0xFF);
627 		}
628 		status = (status >> 8) & 0xFF;
629 #ifdef TDEBUG
630 		switch(status) {
631 			case X_FINOK:
632 				msg("Child %d finishes X_FINOK\n", childpid);
633 				break;
634 			case X_ABORT:
635 				msg("Child %d finishes X_ABORT\n", childpid);
636 				break;
637 			case X_REWRITE:
638 				msg("Child %d finishes X_REWRITE\n", childpid);
639 				break;
640 			default:
641 				msg("Child %d finishes unknown %d\n",
642 					childpid, status);
643 				break;
644 		}
645 #endif /* TDEBUG */
646 		switch(status) {
647 			case X_FINOK:
648 				Exit(X_FINOK);
649 			case X_ABORT:
650 				Exit(X_ABORT);
651 			case X_REWRITE:
652 				goto restore_check_point;
653 			default:
654 				msg("Bad return code from dump: %d\n", status);
655 				Exit(X_ABORT);
656 		}
657 		/*NOTREACHED*/
658 	} else {	/* we are the child; just continue */
659 		signal(SIGINFO, statussig);	/* now want child's stats */
660 #ifdef TDEBUG
661 		sleep(4);	/* allow time for parent's message to get out */
662 		msg("Child on Tape %d has parent %d, my pid = %d\n",
663 			tapeno+1, parentpid, getpid());
664 #endif /* TDEBUG */
665 		/*
666 		 * If we have a name like "/dev/rst0,/dev/rst1",
667 		 * use the name before the comma first, and save
668 		 * the remaining names for subsequent volumes.
669 		 */
670 		tapeno++;			/* current tape sequence */
671 		if (nexttape || strchr(tape, ',')) {
672 			if (nexttape && *nexttape)
673 				tape = nexttape;
674 			if ((p = strchr(tape, ',')) != NULL) {
675 				*p = '\0';
676 				nexttape = p + 1;
677 			} else
678 				nexttape = NULL;
679 			msg("Dumping volume %d on %s\n", tapeno, tape);
680 		}
681 #ifdef RDUMP
682 		while ((tapefd = (host ? rmtopen(tape, 2, 1) :
683 			pipeout ? 1 : open(tape, O_WRONLY|O_CREAT, 0666))) < 0)
684 #else
685 		while ((tapefd = (pipeout ? 1 :
686 				  open(tape, O_WRONLY|O_CREAT, 0666))) < 0)
687 #endif
688 		    {
689 			msg("Cannot open output \"%s\".\n", tape);
690 			if (!query("Do you want to retry the open?"))
691 				dumpabort(0);
692 		}
693 
694 		enslave();  /* Share open tape file descriptor with slaves */
695 
696 		asize = 0;
697 		blocksthisvol = 0;
698 		if (top)
699 			newtape++;		/* new tape signal */
700 		spcl.c_count = iswap32(slp->count);
701 		/*
702 		 * measure firstrec in TP_BSIZE units since restore doesn't
703 		 * know the correct ntrec value...
704 		 */
705 		spcl.c_firstrec = iswap32(slp->firstrec);
706 		spcl.c_volume = iswap32(iswap32(spcl.c_volume) + 1);
707 		spcl.c_type = iswap32(TS_TAPE);
708 		if (!is_ufs2)
709 			spcl.c_flags = iswap32(iswap32(spcl.c_flags)
710 			    | DR_NEWHEADER);
711 		writeheader((ino_t)slp->inode);
712 		if (!is_ufs2)
713 			spcl.c_flags  = iswap32(iswap32(spcl.c_flags) &
714 			    ~ DR_NEWHEADER);
715 		msg("Volume %d started at: %s", tapeno, ctime(&tstart_volume));
716 		if (tapeno > 1)
717 			msg("Volume %d begins with blocks from inode %d\n",
718 				tapeno, slp->inode);
719 	}
720 }
721 
722 void
723 dumpabort(int signo)
724 {
725 
726 	if (master != 0 && master != getpid())
727 		/* Signals master to call dumpabort */
728 		(void) kill(master, SIGTERM);
729 	else {
730 		killall();
731 		msg("The ENTIRE dump is aborted.\n");
732 	}
733 #ifdef RDUMP
734 	rmtclose();
735 #endif
736 	Exit(X_ABORT);
737 }
738 
739 void
740 Exit(int status)
741 {
742 
743 #ifdef TDEBUG
744 	msg("pid = %d exits with status %d\n", getpid(), status);
745 #endif /* TDEBUG */
746 	exit(status);
747 }
748 
749 /*
750  * proceed - handler for SIGUSR2, used to synchronize IO between the slaves.
751  */
752 static void
753 proceed(int signo)
754 {
755 
756 	if (ready)
757 		longjmp(jmpbuf, 1);
758 	caught++;
759 }
760 
761 void
762 enslave(void)
763 {
764 	int cmd[2];
765 	int i, j;
766 
767 	master = getpid();
768 
769 	signal(SIGTERM, dumpabort);  /* Slave sends SIGTERM on dumpabort() */
770 	signal(SIGPIPE, sigpipe);
771 	signal(SIGUSR1, tperror);    /* Slave sends SIGUSR1 on tape errors */
772 	signal(SIGUSR2, proceed);    /* Slave sends SIGUSR2 to next slave */
773 
774 	for (i = 0; i < SLAVES; i++) {
775 		if (i == slp - &slaves[0]) {
776 			caught = 1;
777 		} else {
778 			caught = 0;
779 		}
780 
781 		if (socketpair(AF_LOCAL, SOCK_STREAM, 0, cmd) < 0 ||
782 		    (slaves[i].pid = fork()) < 0)
783 			quit("too many slaves, %d (recompile smaller): %s\n",
784 			    i, strerror(errno));
785 
786 		slaves[i].fd = cmd[1];
787 		slaves[i].sent = 0;
788 		if (slaves[i].pid == 0) { 	    /* Slave starts up here */
789 			for (j = 0; j <= i; j++)
790 				(void) close(slaves[j].fd);
791 			signal(SIGINT, SIG_IGN);    /* Master handles this */
792 			signal(SIGINFO, SIG_IGN);
793 			doslave(cmd[0], i);
794 			Exit(X_FINOK);
795 		}
796 	}
797 
798 	for (i = 0; i < SLAVES; i++)
799 		(void) atomic_write(slaves[i].fd,
800 				(char *) &slaves[(i + 1) % SLAVES].pid,
801 				sizeof slaves[0].pid);
802 
803 	master = 0;
804 }
805 
806 void
807 killall(void)
808 {
809 	int i;
810 
811 	for (i = 0; i < SLAVES; i++)
812 		if (slaves[i].pid > 0) {
813 			(void) kill(slaves[i].pid, SIGKILL);
814 			slaves[i].sent = 0;
815 		}
816 }
817 
818 /*
819  * Synchronization - each process has a lockfile, and shares file
820  * descriptors to the following process's lockfile.  When our write
821  * completes, we release our lock on the following process's lock-
822  * file, allowing the following process to lock it and proceed. We
823  * get the lock back for the next cycle by swapping descriptors.
824  */
825 static void
826 doslave(int cmd, int slave_number)
827 {
828 	int nread, nextslave, size, wrote, eot_count, werror;
829 	sigset_t nsigset;
830 
831 	/*
832 	 * Need our own seek pointer.
833 	 */
834 	(void) close(diskfd);
835 	if ((diskfd = open(disk, O_RDONLY)) < 0)
836 		quit("slave couldn't reopen disk: %s\n", strerror(errno));
837 
838 	/*
839 	 * Need the pid of the next slave in the loop...
840 	 */
841 	if ((nread = atomic_read(cmd, (char *)&nextslave, sizeof nextslave))
842 	    != sizeof nextslave) {
843 		quit("master/slave protocol botched - didn't get pid of next slave.\n");
844 	}
845 
846 	/*
847 	 * Get list of blocks to dump, read the blocks into tape buffer
848 	 */
849 	while ((nread = atomic_read(cmd, (char *)slp->req, reqsiz)) == reqsiz) {
850 		struct req *p = slp->req;
851 
852 		for (trecno = 0; trecno < ntrec;
853 		     trecno += p->count, p += p->count) {
854 			if (p->dblk) {
855 				bread(p->dblk, slp->tblock[trecno],
856 					p->count * TP_BSIZE);
857 			} else {
858 				if (p->count != 1 || atomic_read(cmd,
859 				    (char *)slp->tblock[trecno],
860 				    TP_BSIZE) != TP_BSIZE)
861 				       quit("master/slave protocol botched.\n");
862 			}
863 		}
864 		if (setjmp(jmpbuf) == 0) {
865 			ready = 1;
866 			if (!caught)
867 				(void) pause();
868 		}
869 		ready = 0;
870 		caught = 0;
871 
872 		/* Try to write the data... */
873 		eot_count = 0;
874 		size = 0;
875 		werror = 0;
876 
877 		while (eot_count < 10 && size < writesize) {
878 #ifdef RDUMP
879 			if (host)
880 				wrote = rmtwrite(slp->tblock[0]+size,
881 				    writesize-size);
882 			else
883 #endif
884 				wrote = write(tapefd, slp->tblock[0]+size,
885 				    writesize-size);
886 			werror = errno;
887 #ifdef WRITEDEBUG
888 			fprintf(stderr, "slave %d wrote %d werror %d\n",
889 			    slave_number, wrote, werror);
890 #endif
891 			if (wrote < 0)
892 				break;
893 			if (wrote == 0)
894 				eot_count++;
895 			size += wrote;
896 		}
897 
898 #ifdef WRITEDEBUG
899 		if (size != writesize)
900 			fprintf(stderr,
901 		    "slave %d only wrote %d out of %d bytes and gave up.\n",
902 			    slave_number, size, writesize);
903 #endif
904 
905 		/*
906 		 * Handle ENOSPC as an EOT condition.
907 		 */
908 		if (wrote < 0 && werror == ENOSPC) {
909 			wrote = 0;
910 			eot_count++;
911 		}
912 
913 		if (eot_count > 0)
914 			size = 0;
915 
916 		if (wrote < 0) {
917 			(void) kill(master, SIGUSR1);
918 			sigemptyset(&nsigset);
919 			for (;;)
920 				sigsuspend(&nsigset);
921 		} else {
922 			/*
923 			 * pass size of write back to master
924 			 * (for EOT handling)
925 			 */
926 			(void) atomic_write(cmd, (char *)&size, sizeof size);
927 		}
928 
929 		/*
930 		 * If partial write, don't want next slave to go.
931 		 * Also jolts him awake.
932 		 */
933 		(void) kill(nextslave, SIGUSR2);
934 	}
935 	printcachestats();
936 	if (nread != 0)
937 		quit("error reading command pipe: %s\n", strerror(errno));
938 }
939 
940 /*
941  * Since a read from a pipe may not return all we asked for,
942  * loop until the count is satisfied (or error).
943  */
944 static ssize_t
945 atomic_read(int fd, char *buf, int count)
946 {
947 	ssize_t got, need = count;
948 
949 	while ((got = read(fd, buf, need)) > 0 && (need -= got) > 0)
950 		buf += got;
951 	return (got < 0 ? got : count - need);
952 }
953 
954 /*
955  * Since a write may not write all we ask if we get a signal,
956  * loop until the count is satisfied (or error).
957  */
958 static ssize_t
959 atomic_write(int fd, char *buf, int count)
960 {
961 	ssize_t got, need = count;
962 
963 	while ((got = write(fd, buf, need)) > 0 && (need -= got) > 0)
964 		buf += got;
965 	return (got < 0 ? got : count - need);
966 }
967