xref: /freebsd/sbin/dump/tape.c (revision 6990ffd8a95caaba6858ad44ff1b3157d1efba8f)
1 /*-
2  * Copyright (c) 1980, 1991, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  * 3. All advertising materials mentioning features or use of this software
14  *    must display the following acknowledgement:
15  *	This product includes software developed by the University of
16  *	California, Berkeley and its contributors.
17  * 4. Neither the name of the University nor the names of its contributors
18  *    may be used to endorse or promote products derived from this software
19  *    without specific prior written permission.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  */
33 
34 #ifndef lint
35 #if 0
36 static char sccsid[] = "@(#)tape.c	8.4 (Berkeley) 5/1/95";
37 #endif
38 static const char rcsid[] =
39   "$FreeBSD$";
40 #endif /* not lint */
41 
42 #include <sys/param.h>
43 #include <sys/socket.h>
44 #include <sys/time.h>
45 #include <sys/wait.h>
46 #ifdef sunos
47 #include <sys/vnode.h>
48 
49 #include <ufs/fs.h>
50 #include <ufs/inode.h>
51 #else
52 #include <ufs/ufs/dinode.h>
53 #include <ufs/ffs/fs.h>
54 #endif
55 
56 #include <protocols/dumprestore.h>
57 
58 #include <errno.h>
59 #include <fcntl.h>
60 #include <setjmp.h>
61 #include <signal.h>
62 #include <stdio.h>
63 #ifdef __STDC__
64 #include <stdlib.h>
65 #include <string.h>
66 #include <unistd.h>
67 #else
68 int	write(), read();
69 #endif
70 
71 #include "dump.h"
72 
73 int	writesize;		/* size of malloc()ed buffer for tape */
74 long	lastspclrec = -1;	/* tape block number of last written header */
75 int	trecno = 0;		/* next record to write in current block */
76 extern	long blocksperfile;	/* number of blocks per output file */
77 long	blocksthisvol;		/* number of blocks on current output file */
78 extern	int ntrec;		/* blocking factor on tape */
79 extern	int cartridge;
80 extern	char *host;
81 char	*nexttape;
82 
83 static	int atomic __P((ssize_t (*)(), int, char *, int));
84 static	void doslave __P((int, int));
85 static	void enslave __P((void));
86 static	void flushtape __P((void));
87 static	void killall __P((void));
88 static	void rollforward __P((void));
89 
90 /*
91  * Concurrent dump mods (Caltech) - disk block reading and tape writing
92  * are exported to several slave processes.  While one slave writes the
93  * tape, the others read disk blocks; they pass control of the tape in
94  * a ring via signals. The parent process traverses the filesystem and
95  * sends writeheader()'s and lists of daddr's to the slaves via pipes.
96  * The following structure defines the instruction packets sent to slaves.
97  */
98 struct req {
99 	daddr_t dblk;
100 	int count;
101 };
102 int reqsiz;
103 
104 #define SLAVES 3		/* 1 slave writing, 1 reading, 1 for slack */
105 struct slave {
106 	int tapea;		/* header number at start of this chunk */
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 	int firstrec;		/* record number of this block */
114 	char (*tblock)[TP_BSIZE]; /* buffer for data blocks */
115 	struct req *req;	/* buffer for requests */
116 } slaves[SLAVES+1];
117 struct slave *slp;
118 
119 char	(*nextblock)[TP_BSIZE];
120 
121 int master;		/* pid of master, for sending error signals */
122 int tenths;		/* length of tape used per block written */
123 static int caught;	/* have we caught the signal to proceed? */
124 static int ready;	/* have we reached the lock point without having */
125 			/* received the SIGUSR2 signal from the prev slave? */
126 static jmp_buf jmpbuf;	/* where to jump to if we are ready when the */
127 			/* SIGUSR2 arrives from the previous slave */
128 
129 int
130 alloctape()
131 {
132 	int pgoff = getpagesize() - 1;
133 	char *buf;
134 	int i;
135 
136 	writesize = ntrec * TP_BSIZE;
137 	reqsiz = (ntrec + 1) * sizeof(struct req);
138 	/*
139 	 * CDC 92181's and 92185's make 0.8" gaps in 1600-bpi start/stop mode
140 	 * (see DEC TU80 User's Guide).  The shorter gaps of 6250-bpi require
141 	 * repositioning after stopping, i.e, streaming mode, where the gap is
142 	 * variable, 0.30" to 0.45".  The gap is maximal when the tape stops.
143 	 */
144 	if (blocksperfile == 0 && !unlimited)
145 		tenths = writesize / density +
146 		    (cartridge ? 16 : density == 625 ? 5 : 8);
147 	/*
148 	 * Allocate tape buffer contiguous with the array of instruction
149 	 * packets, so flushtape() can write them together with one write().
150 	 * Align tape buffer on page boundary to speed up tape write().
151 	 */
152 	for (i = 0; i <= SLAVES; i++) {
153 		buf = (char *)
154 		    malloc((unsigned)(reqsiz + writesize + pgoff + TP_BSIZE));
155 		if (buf == NULL)
156 			return(0);
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(dp, isspcl)
171 	char *dp;
172 	int isspcl;
173 {
174 
175 	slp->req[trecno].dblk = (daddr_t)0;
176 	slp->req[trecno].count = 1;
177 #ifndef	__alpha__
178 	*(union u_spcl *)(*(nextblock)++) = *(union u_spcl *)dp;
179 #else
180 	bcopy(dp, *(nextblock)++, sizeof (union u_spcl));
181 #endif
182 	if (isspcl)
183 		lastspclrec = spcl.c_tapea;
184 	trecno++;
185 	spcl.c_tapea++;
186 	if (trecno >= ntrec)
187 		flushtape();
188 }
189 
190 void
191 dumpblock(blkno, size)
192 	daddr_t blkno;
193 	int size;
194 {
195 	int avail, tpblks, dblkno;
196 
197 	dblkno = fsbtodb(sblock, blkno);
198 	tpblks = size >> tp_bshift;
199 	while ((avail = MIN(tpblks, ntrec - trecno)) > 0) {
200 		slp->req[trecno].dblk = dblkno;
201 		slp->req[trecno].count = avail;
202 		trecno += avail;
203 		spcl.c_tapea += avail;
204 		if (trecno >= ntrec)
205 			flushtape();
206 		dblkno += avail << (tp_bshift - dev_bshift);
207 		tpblks -= avail;
208 	}
209 }
210 
211 int	nogripe = 0;
212 
213 void
214 tperror(signo)
215 	int signo;
216 {
217 
218 	if (pipeout) {
219 		msg("write error on %s\n", tape);
220 		quit("Cannot recover\n");
221 		/* NOTREACHED */
222 	}
223 	msg("write error %d blocks into volume %d\n", blocksthisvol, tapeno);
224 	broadcast("DUMP WRITE ERROR!\n");
225 	if (!query("Do you want to restart?"))
226 		dumpabort(0);
227 	msg("Closing this volume.  Prepare to restart with new media;\n");
228 	msg("this dump volume will be rewritten.\n");
229 	killall();
230 	nogripe = 1;
231 	close_rewind();
232 	Exit(X_REWRITE);
233 }
234 
235 void
236 sigpipe(signo)
237 	int signo;
238 {
239 
240 	quit("Broken pipe\n");
241 }
242 
243 static void
244 flushtape()
245 {
246 	int i, blks, got;
247 	long lastfirstrec;
248 
249 	int siz = (char *)nextblock - (char *)slp->req;
250 
251 	slp->req[trecno].count = 0;			/* Sentinel */
252 
253 	if (atomic(write, slp->fd, (char *)slp->req, siz) != siz)
254 		quit("error writing command pipe: %s\n", strerror(errno));
255 	slp->sent = 1; /* we sent a request, read the response later */
256 
257 	lastfirstrec = slp->firstrec;
258 
259 	if (++slp >= &slaves[SLAVES])
260 		slp = &slaves[0];
261 
262 	/* Read results back from next slave */
263 	if (slp->sent) {
264 		if (atomic(read, slp->fd, (char *)&got, sizeof got)
265 		    != sizeof got) {
266 			perror("  DUMP: error reading command pipe in master");
267 			dumpabort(0);
268 		}
269 		slp->sent = 0;
270 
271 		/* Check for end of tape */
272 		if (got < writesize) {
273 			msg("End of tape detected\n");
274 
275 			/*
276 			 * Drain the results, don't care what the values were.
277 			 * If we read them here then trewind won't...
278 			 */
279 			for (i = 0; i < SLAVES; i++) {
280 				if (slaves[i].sent) {
281 					if (atomic(read, slaves[i].fd,
282 					    (char *)&got, sizeof got)
283 					    != sizeof got) {
284 						perror("  DUMP: error reading command pipe in master");
285 						dumpabort(0);
286 					}
287 					slaves[i].sent = 0;
288 				}
289 			}
290 
291 			close_rewind();
292 			rollforward();
293 			return;
294 		}
295 	}
296 
297 	blks = 0;
298 	if (spcl.c_type != TS_END) {
299 		for (i = 0; i < spcl.c_count; i++)
300 			if (spcl.c_addr[i] != 0)
301 				blks++;
302 	}
303 	slp->count = lastspclrec + blks + 1 - spcl.c_tapea;
304 	slp->tapea = spcl.c_tapea;
305 	slp->firstrec = lastfirstrec + ntrec;
306 	slp->inode = curino;
307 	nextblock = slp->tblock;
308 	trecno = 0;
309 	asize += tenths;
310 	blockswritten += ntrec;
311 	blocksthisvol += ntrec;
312 	if (!pipeout && !unlimited && (blocksperfile ?
313 	    (blocksthisvol >= blocksperfile) : (asize > tsize))) {
314 		close_rewind();
315 		startnewtape(0);
316 	}
317 	timeest();
318 }
319 
320 void
321 trewind()
322 {
323 	int f;
324 	int got;
325 
326 	for (f = 0; f < SLAVES; f++) {
327 		/*
328 		 * Drain the results, but unlike EOT we DO (or should) care
329 		 * what the return values were, since if we detect EOT after
330 		 * we think we've written the last blocks to the tape anyway,
331 		 * we have to replay those blocks with rollforward.
332 		 *
333 		 * fixme: punt for now.
334 		 */
335 		if (slaves[f].sent) {
336 			if (atomic(read, slaves[f].fd, (char *)&got, sizeof got)
337 			    != sizeof got) {
338 				perror("  DUMP: error reading command pipe in master");
339 				dumpabort(0);
340 			}
341 			slaves[f].sent = 0;
342 			if (got != writesize) {
343 				msg("EOT detected in last 2 tape records!\n");
344 				msg("Use a longer tape, decrease the size estimate\n");
345 				quit("or use no size estimate at all.\n");
346 			}
347 		}
348 		(void) close(slaves[f].fd);
349 	}
350 	while (wait((int *)NULL) >= 0)	/* wait for any signals from slaves */
351 		/* void */;
352 
353 	if (pipeout)
354 		return;
355 
356 	msg("Closing %s\n", tape);
357 
358 #ifdef RDUMP
359 	if (host) {
360 		rmtclose();
361 		while (rmtopen(tape, 0) < 0)
362 			sleep(10);
363 		rmtclose();
364 		return;
365 	}
366 #endif
367 	(void) close(tapefd);
368 	while ((f = open(tape, 0)) < 0)
369 		sleep (10);
370 	(void) close(f);
371 }
372 
373 void
374 close_rewind()
375 {
376 	time_t tstart_changevol, tend_changevol;
377 
378 	trewind();
379 	if (nexttape)
380 		return;
381 	(void)time((time_t *)&(tstart_changevol));
382 	if (!nogripe) {
383 		msg("Change Volumes: Mount volume #%d\n", tapeno+1);
384 		broadcast("CHANGE DUMP VOLUMES!\a\a\n");
385 	}
386 	while (!query("Is the new volume mounted and ready to go?"))
387 		if (query("Do you want to abort?")) {
388 			dumpabort(0);
389 			/*NOTREACHED*/
390 		}
391 	(void)time((time_t *)&(tend_changevol));
392 	if ((tstart_changevol != (time_t)-1) && (tend_changevol != (time_t)-1))
393 		tstart_writing += (tend_changevol - tstart_changevol);
394 }
395 
396 void
397 rollforward()
398 {
399 	register struct req *p, *q, *prev;
400 	register struct slave *tslp;
401 	int i, size, savedtapea, got;
402 	union u_spcl *ntb, *otb;
403 	tslp = &slaves[SLAVES];
404 	ntb = (union u_spcl *)tslp->tblock[1];
405 
406 	/*
407 	 * Each of the N slaves should have requests that need to
408 	 * be replayed on the next tape.  Use the extra slave buffers
409 	 * (slaves[SLAVES]) to construct request lists to be sent to
410 	 * each slave in turn.
411 	 */
412 	for (i = 0; i < SLAVES; i++) {
413 		q = &tslp->req[1];
414 		otb = (union u_spcl *)slp->tblock;
415 
416 		/*
417 		 * For each request in the current slave, copy it to tslp.
418 		 */
419 
420 		prev = NULL;
421 		for (p = slp->req; p->count > 0; p += p->count) {
422 			*q = *p;
423 			if (p->dblk == 0)
424 				*ntb++ = *otb++; /* copy the datablock also */
425 			prev = q;
426 			q += q->count;
427 		}
428 		if (prev == NULL)
429 			quit("rollforward: protocol botch");
430 		if (prev->dblk != 0)
431 			prev->count -= 1;
432 		else
433 			ntb--;
434 		q -= 1;
435 		q->count = 0;
436 		q = &tslp->req[0];
437 		if (i == 0) {
438 			q->dblk = 0;
439 			q->count = 1;
440 			trecno = 0;
441 			nextblock = tslp->tblock;
442 			savedtapea = spcl.c_tapea;
443 			spcl.c_tapea = slp->tapea;
444 			startnewtape(0);
445 			spcl.c_tapea = savedtapea;
446 			lastspclrec = savedtapea - 1;
447 		}
448 		size = (char *)ntb - (char *)q;
449 		if (atomic(write, slp->fd, (char *)q, size) != size) {
450 			perror("  DUMP: error writing command pipe");
451 			dumpabort(0);
452 		}
453 		slp->sent = 1;
454 		if (++slp >= &slaves[SLAVES])
455 			slp = &slaves[0];
456 
457 		q->count = 1;
458 
459 		if (prev->dblk != 0) {
460 			/*
461 			 * If the last one was a disk block, make the
462 			 * first of this one be the last bit of that disk
463 			 * block...
464 			 */
465 			q->dblk = prev->dblk +
466 				prev->count * (TP_BSIZE / DEV_BSIZE);
467 			ntb = (union u_spcl *)tslp->tblock;
468 		} else {
469 			/*
470 			 * It wasn't a disk block.  Copy the data to its
471 			 * new location in the buffer.
472 			 */
473 			q->dblk = 0;
474 			*((union u_spcl *)tslp->tblock) = *ntb;
475 			ntb = (union u_spcl *)tslp->tblock[1];
476 		}
477 	}
478 	slp->req[0] = *q;
479 	nextblock = slp->tblock;
480 	if (q->dblk == 0)
481 		nextblock++;
482 	trecno = 1;
483 
484 	/*
485 	 * Clear the first slaves' response.  One hopes that it
486 	 * worked ok, otherwise the tape is much too short!
487 	 */
488 	if (slp->sent) {
489 		if (atomic(read, slp->fd, (char *)&got, sizeof got)
490 		    != sizeof got) {
491 			perror("  DUMP: error reading command pipe in master");
492 			dumpabort(0);
493 		}
494 		slp->sent = 0;
495 
496 		if (got != writesize) {
497 			quit("EOT detected at start of the tape!\n");
498 		}
499 	}
500 }
501 
502 /*
503  * We implement taking and restoring checkpoints on the tape level.
504  * When each tape is opened, a new process is created by forking; this
505  * saves all of the necessary context in the parent.  The child
506  * continues the dump; the parent waits around, saving the context.
507  * If the child returns X_REWRITE, then it had problems writing that tape;
508  * this causes the parent to fork again, duplicating the context, and
509  * everything continues as if nothing had happened.
510  */
511 void
512 startnewtape(top)
513 	int top;
514 {
515 	int	parentpid;
516 	int	childpid;
517 	int	status;
518 	int	waitpid;
519 	char	*p;
520 #ifdef sunos
521 	void	(*interrupt_save)();
522 #else
523 	sig_t	interrupt_save;
524 #endif
525 
526 	interrupt_save = signal(SIGINT, SIG_IGN);
527 	parentpid = getpid();
528 
529 restore_check_point:
530 	(void)signal(SIGINT, interrupt_save);
531 	/*
532 	 *	All signals are inherited...
533 	 */
534 	childpid = fork();
535 	if (childpid < 0) {
536 		msg("Context save fork fails in parent %d\n", parentpid);
537 		Exit(X_ABORT);
538 	}
539 	if (childpid != 0) {
540 		/*
541 		 *	PARENT:
542 		 *	save the context by waiting
543 		 *	until the child doing all of the work returns.
544 		 *	don't catch the interrupt
545 		 */
546 		signal(SIGINT, SIG_IGN);
547 #ifdef TDEBUG
548 		msg("Tape: %d; parent process: %d child process %d\n",
549 			tapeno+1, parentpid, childpid);
550 #endif /* TDEBUG */
551 		while ((waitpid = wait(&status)) != childpid)
552 			msg("Parent %d waiting for child %d has another child %d return\n",
553 				parentpid, childpid, waitpid);
554 		if (status & 0xFF) {
555 			msg("Child %d returns LOB status %o\n",
556 				childpid, status&0xFF);
557 		}
558 		status = (status >> 8) & 0xFF;
559 #ifdef TDEBUG
560 		switch(status) {
561 			case X_FINOK:
562 				msg("Child %d finishes X_FINOK\n", childpid);
563 				break;
564 			case X_ABORT:
565 				msg("Child %d finishes X_ABORT\n", childpid);
566 				break;
567 			case X_REWRITE:
568 				msg("Child %d finishes X_REWRITE\n", childpid);
569 				break;
570 			default:
571 				msg("Child %d finishes unknown %d\n",
572 					childpid, status);
573 				break;
574 		}
575 #endif /* TDEBUG */
576 		switch(status) {
577 			case X_FINOK:
578 				Exit(X_FINOK);
579 			case X_ABORT:
580 				Exit(X_ABORT);
581 			case X_REWRITE:
582 				goto restore_check_point;
583 			default:
584 				msg("Bad return code from dump: %d\n", status);
585 				Exit(X_ABORT);
586 		}
587 		/*NOTREACHED*/
588 	} else {	/* we are the child; just continue */
589 #ifdef TDEBUG
590 		sleep(4);	/* allow time for parent's message to get out */
591 		msg("Child on Tape %d has parent %d, my pid = %d\n",
592 			tapeno+1, parentpid, getpid());
593 #endif /* TDEBUG */
594 		/*
595 		 * If we have a name like "/dev/rmt0,/dev/rmt1",
596 		 * use the name before the comma first, and save
597 		 * the remaining names for subsequent volumes.
598 		 */
599 		tapeno++;               /* current tape sequence */
600 		if (nexttape || strchr(tape, ',')) {
601 			if (nexttape && *nexttape)
602 				tape = nexttape;
603 			if ((p = strchr(tape, ',')) != NULL) {
604 				*p = '\0';
605 				nexttape = p + 1;
606 			} else
607 				nexttape = NULL;
608 			msg("Dumping volume %d on %s\n", tapeno, tape);
609 		}
610 #ifdef RDUMP
611 		while ((tapefd = (host ? rmtopen(tape, 2) :
612 			pipeout ? 1 : open(tape, O_WRONLY|O_CREAT, 0666))) < 0)
613 #else
614 		while ((tapefd = (pipeout ? 1 :
615 				  open(tape, O_WRONLY|O_CREAT, 0666))) < 0)
616 #endif
617 		    {
618 			msg("Cannot open output \"%s\".\n", tape);
619 			if (!query("Do you want to retry the open?"))
620 				dumpabort(0);
621 		}
622 
623 		enslave();  /* Share open tape file descriptor with slaves */
624 
625 		asize = 0;
626 		blocksthisvol = 0;
627 		if (top)
628 			newtape++;		/* new tape signal */
629 		spcl.c_count = slp->count;
630 		/*
631 		 * measure firstrec in TP_BSIZE units since restore doesn't
632 		 * know the correct ntrec value...
633 		 */
634 		spcl.c_firstrec = slp->firstrec;
635 		spcl.c_volume++;
636 		spcl.c_type = TS_TAPE;
637 		spcl.c_flags |= DR_NEWHEADER;
638 		writeheader((ino_t)slp->inode);
639 		spcl.c_flags &=~ DR_NEWHEADER;
640 		if (tapeno > 1)
641 			msg("Volume %d begins with blocks from inode %d\n",
642 				tapeno, slp->inode);
643 	}
644 }
645 
646 void
647 dumpabort(signo)
648 	int signo;
649 {
650 
651 	if (master != 0 && master != getpid())
652 		/* Signals master to call dumpabort */
653 		(void) kill(master, SIGTERM);
654 	else {
655 		killall();
656 		msg("The ENTIRE dump is aborted.\n");
657 	}
658 #ifdef RDUMP
659 	rmtclose();
660 #endif
661 	Exit(X_ABORT);
662 }
663 
664 void
665 Exit(status)
666 	int status;
667 {
668 
669 #ifdef TDEBUG
670 	msg("pid = %d exits with status %d\n", getpid(), status);
671 #endif /* TDEBUG */
672 	exit(status);
673 }
674 
675 /*
676  * proceed - handler for SIGUSR2, used to synchronize IO between the slaves.
677  */
678 void
679 proceed(signo)
680 	int signo;
681 {
682 
683 	if (ready)
684 		longjmp(jmpbuf, 1);
685 	caught++;
686 }
687 
688 void
689 enslave()
690 {
691 	int cmd[2];
692 	register int i, j;
693 
694 	master = getpid();
695 
696 	signal(SIGTERM, dumpabort);  /* Slave sends SIGTERM on dumpabort() */
697 	signal(SIGPIPE, sigpipe);
698 	signal(SIGUSR1, tperror);    /* Slave sends SIGUSR1 on tape errors */
699 	signal(SIGUSR2, proceed);    /* Slave sends SIGUSR2 to next slave */
700 
701 	for (i = 0; i < SLAVES; i++) {
702 		if (i == slp - &slaves[0]) {
703 			caught = 1;
704 		} else {
705 			caught = 0;
706 		}
707 
708 		if (socketpair(AF_UNIX, SOCK_STREAM, 0, cmd) < 0 ||
709 		    (slaves[i].pid = fork()) < 0)
710 			quit("too many slaves, %d (recompile smaller): %s\n",
711 			    i, strerror(errno));
712 
713 		slaves[i].fd = cmd[1];
714 		slaves[i].sent = 0;
715 		if (slaves[i].pid == 0) { 	    /* Slave starts up here */
716 			for (j = 0; j <= i; j++)
717 			        (void) close(slaves[j].fd);
718 			signal(SIGINT, SIG_IGN);    /* Master handles this */
719 			doslave(cmd[0], i);
720 			Exit(X_FINOK);
721 		}
722 	}
723 
724 	for (i = 0; i < SLAVES; i++)
725 		(void) atomic(write, slaves[i].fd,
726 			      (char *) &slaves[(i + 1) % SLAVES].pid,
727 		              sizeof slaves[0].pid);
728 
729 	master = 0;
730 }
731 
732 void
733 killall()
734 {
735 	register int i;
736 
737 	for (i = 0; i < SLAVES; i++)
738 		if (slaves[i].pid > 0) {
739 			(void) kill(slaves[i].pid, SIGKILL);
740 			slaves[i].sent = 0;
741 		}
742 }
743 
744 /*
745  * Synchronization - each process has a lockfile, and shares file
746  * descriptors to the following process's lockfile.  When our write
747  * completes, we release our lock on the following process's lock-
748  * file, allowing the following process to lock it and proceed. We
749  * get the lock back for the next cycle by swapping descriptors.
750  */
751 static void
752 doslave(cmd, slave_number)
753 	register int cmd;
754         int slave_number;
755 {
756 	register int nread;
757 	int nextslave, size, wrote, eot_count;
758 
759 	/*
760 	 * Need our own seek pointer.
761 	 */
762 	(void) close(diskfd);
763 	if ((diskfd = open(disk, O_RDONLY)) < 0)
764 		quit("slave couldn't reopen disk: %s\n", strerror(errno));
765 
766 	/*
767 	 * Need the pid of the next slave in the loop...
768 	 */
769 	if ((nread = atomic(read, cmd, (char *)&nextslave, sizeof nextslave))
770 	    != sizeof nextslave) {
771 		quit("master/slave protocol botched - didn't get pid of next slave.\n");
772 	}
773 
774 	/*
775 	 * Get list of blocks to dump, read the blocks into tape buffer
776 	 */
777 	while ((nread = atomic(read, cmd, (char *)slp->req, reqsiz)) == reqsiz) {
778 		register struct req *p = slp->req;
779 
780 		for (trecno = 0; trecno < ntrec;
781 		     trecno += p->count, p += p->count) {
782 			if (p->dblk) {
783 				bread(p->dblk, slp->tblock[trecno],
784 					p->count * TP_BSIZE);
785 			} else {
786 				if (p->count != 1 || atomic(read, cmd,
787 				    (char *)slp->tblock[trecno],
788 				    TP_BSIZE) != TP_BSIZE)
789 				       quit("master/slave protocol botched.\n");
790 			}
791 		}
792 		if (setjmp(jmpbuf) == 0) {
793 			ready = 1;
794 			if (!caught)
795 				(void) pause();
796 		}
797 		ready = 0;
798 		caught = 0;
799 
800 		/* Try to write the data... */
801 		eot_count = 0;
802 		size = 0;
803 
804 		while (eot_count < 10 && size < writesize) {
805 #ifdef RDUMP
806 			if (host)
807 				wrote = rmtwrite(slp->tblock[0]+size,
808 				    writesize-size);
809 			else
810 #endif
811 				wrote = write(tapefd, slp->tblock[0]+size,
812 				    writesize-size);
813 #ifdef WRITEDEBUG
814 			printf("slave %d wrote %d\n", slave_number, wrote);
815 #endif
816 			if (wrote < 0)
817 				break;
818 			if (wrote == 0)
819 				eot_count++;
820 			size += wrote;
821 		}
822 
823 #ifdef WRITEDEBUG
824 		if (size != writesize)
825 		 printf("slave %d only wrote %d out of %d bytes and gave up.\n",
826 		     slave_number, size, writesize);
827 #endif
828 
829 		/*
830 		 * Handle ENOSPC as an EOT condition.
831 		 */
832 		if (wrote < 0 && errno == ENOSPC) {
833 			wrote = 0;
834 			eot_count++;
835 		}
836 
837 		if (eot_count > 0)
838 			size = 0;
839 
840 		if (wrote < 0) {
841 			(void) kill(master, SIGUSR1);
842 			for (;;)
843 				(void) sigpause(0);
844 		} else {
845 			/*
846 			 * pass size of write back to master
847 			 * (for EOT handling)
848 			 */
849 			(void) atomic(write, cmd, (char *)&size, sizeof size);
850 		}
851 
852 		/*
853 		 * If partial write, don't want next slave to go.
854 		 * Also jolts him awake.
855 		 */
856 		(void) kill(nextslave, SIGUSR2);
857 	}
858 	if (nread != 0)
859 		quit("error reading command pipe: %s\n", strerror(errno));
860 }
861 
862 /*
863  * Since a read from a pipe may not return all we asked for,
864  * or a write may not write all we ask if we get a signal,
865  * loop until the count is satisfied (or error).
866  */
867 static int
868 atomic(func, fd, buf, count)
869 	ssize_t (*func)();
870 	int fd;
871 	char *buf;
872 	int count;
873 {
874 	int got, need = count;
875 
876 	while ((got = (*func)(fd, buf, need)) > 0 && (need -= got) > 0)
877 		buf += got;
878 	return (got < 0 ? got : count - need);
879 }
880