xref: /netbsd-src/sbin/dump/tape.c (revision 1f2744e6e4915c9da2a3f980279398c4cf7d5e6d)
1 /*	$NetBSD: tape.c,v 1.7 1995/03/21 18:48:47 mycroft 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 #ifndef lint
37 #if 0
38 static char sccsid[] = "@(#)tape.c	8.2 (Berkeley) 3/17/94";
39 #else
40 static char rcsid[] = "$NetBSD: tape.c,v 1.7 1995/03/21 18:48:47 mycroft Exp $";
41 #endif
42 #endif /* not lint */
43 
44 #include <sys/param.h>
45 #include <sys/socket.h>
46 #include <sys/time.h>
47 #include <sys/wait.h>
48 #ifdef sunos
49 #include <sys/vnode.h>
50 
51 #include <ufs/fs.h>
52 #include <ufs/inode.h>
53 #else
54 #include <ufs/ffs/fs.h>
55 #include <ufs/ufs/dinode.h>
56 #endif
57 
58 #include <protocols/dumprestore.h>
59 
60 #include <errno.h>
61 #include <fcntl.h>
62 #include <setjmp.h>
63 #include <signal.h>
64 #include <stdio.h>
65 #ifdef __STDC__
66 #include <stdlib.h>
67 #include <string.h>
68 #include <unistd.h>
69 #else
70 int	write(), read();
71 #endif
72 
73 #include "dump.h"
74 #include "pathnames.h"
75 
76 int	writesize;		/* size of malloc()ed buffer for tape */
77 long	lastspclrec = -1;	/* tape block number of last written header */
78 int	trecno = 0;		/* next record to write in current block */
79 extern	long blocksperfile;	/* number of blocks per output file */
80 long	blocksthisvol;		/* number of blocks on current output file */
81 extern	int ntrec;		/* blocking factor on tape */
82 extern	int cartridge;
83 extern	char *host;
84 char	*nexttape;
85 
86 static	ssize_t atomic __P((ssize_t (*)(), int, char *, int));
87 static	void doslave __P((int, int));
88 static	void enslave __P((void));
89 static	void flushtape __P((void));
90 static	void killall __P((void));
91 static	void rollforward __P((void));
92 
93 /*
94  * Concurrent dump mods (Caltech) - disk block reading and tape writing
95  * are exported to several slave processes.  While one slave writes the
96  * tape, the others read disk blocks; they pass control of the tape in
97  * a ring via signals. The parent process traverses the filesystem and
98  * sends writeheader()'s and lists of daddr's to the slaves via pipes.
99  * The following structure defines the instruction packets sent to slaves.
100  */
101 struct req {
102 	daddr_t dblk;
103 	int count;
104 };
105 int reqsiz;
106 
107 #define SLAVES 3		/* 1 slave writing, 1 reading, 1 for slack */
108 struct slave {
109 	int tapea;		/* header number at start of this chunk */
110 	int count;		/* count to next header (used for TS_TAPE */
111 				/* after EOT) */
112 	int inode;		/* inode that we are currently dealing with */
113 	int fd;			/* FD for this slave */
114 	int pid;		/* PID for this slave */
115 	int sent;		/* 1 == we've sent this slave requests */
116 	int firstrec;		/* record number of this block */
117 	char (*tblock)[TP_BSIZE]; /* buffer for data blocks */
118 	struct req *req;	/* buffer for requests */
119 } slaves[SLAVES+1];
120 struct slave *slp;
121 
122 char	(*nextblock)[TP_BSIZE];
123 
124 int master;		/* pid of master, for sending error signals */
125 int tenths;		/* length of tape used per block written */
126 static int caught;	/* have we caught the signal to proceed? */
127 static int ready;	/* have we reached the lock point without having */
128 			/* received the SIGUSR2 signal from the prev slave? */
129 static jmp_buf jmpbuf;	/* where to jump to if we are ready when the */
130 			/* SIGUSR2 arrives from the previous slave */
131 
132 int
133 alloctape()
134 {
135 	int pgoff = getpagesize() - 1;
136 	char *buf;
137 	int i;
138 
139 	writesize = ntrec * TP_BSIZE;
140 	reqsiz = (ntrec + 1) * sizeof(struct req);
141 	/*
142 	 * CDC 92181's and 92185's make 0.8" gaps in 1600-bpi start/stop mode
143 	 * (see DEC TU80 User's Guide).  The shorter gaps of 6250-bpi require
144 	 * repositioning after stopping, i.e, streaming mode, where the gap is
145 	 * variable, 0.30" to 0.45".  The gap is maximal when the tape stops.
146 	 */
147 	if (blocksperfile == 0)
148 		tenths = writesize / density +
149 		    (cartridge ? 16 : density == 625 ? 5 : 8);
150 	/*
151 	 * Allocate tape buffer contiguous with the array of instruction
152 	 * packets, so flushtape() can write them together with one write().
153 	 * Align tape buffer on page boundary to speed up tape write().
154 	 */
155 	for (i = 0; i <= SLAVES; i++) {
156 		buf = (char *)
157 		    malloc((unsigned)(reqsiz + writesize + pgoff + TP_BSIZE));
158 		if (buf == NULL)
159 			return(0);
160 		slaves[i].tblock = (char (*)[TP_BSIZE])
161 		    (((long)&buf[ntrec + 1] + pgoff) &~ pgoff);
162 		slaves[i].req = (struct req *)slaves[i].tblock - ntrec - 1;
163 	}
164 	slp = &slaves[0];
165 	slp->count = 1;
166 	slp->tapea = 0;
167 	slp->firstrec = 0;
168 	nextblock = slp->tblock;
169 	return(1);
170 }
171 
172 void
173 writerec(dp, isspcl)
174 	char *dp;
175 	int isspcl;
176 {
177 
178 	slp->req[trecno].dblk = (daddr_t)0;
179 	slp->req[trecno].count = 1;
180 	*(union u_spcl *)(*(nextblock)++) = *(union u_spcl *)dp;
181 	if (isspcl)
182 		lastspclrec = spcl.c_tapea;
183 	trecno++;
184 	spcl.c_tapea++;
185 	if (trecno >= ntrec)
186 		flushtape();
187 }
188 
189 void
190 dumpblock(blkno, size)
191 	daddr_t blkno;
192 	int size;
193 {
194 	int avail, tpblks, dblkno;
195 
196 	dblkno = fsbtodb(sblock, blkno);
197 	tpblks = size >> tp_bshift;
198 	while ((avail = MIN(tpblks, ntrec - trecno)) > 0) {
199 		slp->req[trecno].dblk = dblkno;
200 		slp->req[trecno].count = avail;
201 		trecno += avail;
202 		spcl.c_tapea += avail;
203 		if (trecno >= ntrec)
204 			flushtape();
205 		dblkno += avail << (tp_bshift - dev_bshift);
206 		tpblks -= avail;
207 	}
208 }
209 
210 int	nogripe = 0;
211 
212 void
213 tperror(signo)
214 	int signo;
215 {
216 
217 	if (pipeout) {
218 		msg("write error on %s\n", tape);
219 		quit("Cannot recover\n");
220 		/* NOTREACHED */
221 	}
222 	msg("write error %d blocks into volume %d\n", blocksthisvol, tapeno);
223 	broadcast("DUMP WRITE ERROR!\n");
224 	if (!query("Do you want to restart?"))
225 		dumpabort(0);
226 	msg("Closing this volume.  Prepare to restart with new media;\n");
227 	msg("this dump volume will be rewritten.\n");
228 	killall();
229 	nogripe = 1;
230 	close_rewind();
231 	Exit(X_REWRITE);
232 }
233 
234 void
235 sigpipe(signo)
236 	int signo;
237 {
238 
239 	quit("Broken pipe\n");
240 }
241 
242 static void
243 flushtape()
244 {
245 	int i, blks, got;
246 	long lastfirstrec;
247 
248 	int siz = (char *)nextblock - (char *)slp->req;
249 
250 	slp->req[trecno].count = 0;			/* Sentinel */
251 
252 	if (atomic(write, slp->fd, (char *)slp->req, siz) != siz)
253 		quit("error writing command pipe: %s\n", strerror(errno));
254 	slp->sent = 1; /* we sent a request, read the response later */
255 
256 	lastfirstrec = slp->firstrec;
257 
258 	if (++slp >= &slaves[SLAVES])
259 		slp = &slaves[0];
260 
261 	/* Read results back from next slave */
262 	if (slp->sent) {
263 		if (atomic(read, slp->fd, (char *)&got, sizeof got)
264 		    != sizeof got) {
265 			perror("  DUMP: error reading command pipe in master");
266 			dumpabort(0);
267 		}
268 		slp->sent = 0;
269 
270 		/* Check for end of tape */
271 		if (got < writesize) {
272 			msg("End of tape detected\n");
273 
274 			/*
275 			 * Drain the results, don't care what the values were.
276 			 * If we read them here then trewind won't...
277 			 */
278 			for (i = 0; i < SLAVES; i++) {
279 				if (slaves[i].sent) {
280 					if (atomic(read, slaves[i].fd,
281 					    (char *)&got, sizeof got)
282 					    != sizeof got) {
283 						perror("  DUMP: error reading command pipe in master");
284 						dumpabort(0);
285 					}
286 					slaves[i].sent = 0;
287 				}
288 			}
289 
290 			close_rewind();
291 			rollforward();
292 			return;
293 		}
294 	}
295 
296 	blks = 0;
297 	if (spcl.c_type != TS_END) {
298 		for (i = 0; i < spcl.c_count; i++)
299 			if (spcl.c_addr[i] != 0)
300 				blks++;
301 	}
302 	slp->count = lastspclrec + blks + 1 - spcl.c_tapea;
303 	slp->tapea = spcl.c_tapea;
304 	slp->firstrec = lastfirstrec + ntrec;
305 	slp->inode = curino;
306 	nextblock = slp->tblock;
307 	trecno = 0;
308 	asize += tenths;
309 	blockswritten += ntrec;
310 	blocksthisvol += ntrec;
311 	if (!pipeout && (blocksperfile ?
312 	    (blocksthisvol >= blocksperfile) : (asize > tsize))) {
313 		close_rewind();
314 		startnewtape(0);
315 	}
316 	timeest();
317 }
318 
319 void
320 trewind()
321 {
322 	int f;
323 	int got;
324 
325 	for (f = 0; f < SLAVES; f++) {
326 		/*
327 		 * Drain the results, but unlike EOT we DO (or should) care
328 		 * what the return values were, since if we detect EOT after
329 		 * we think we've written the last blocks to the tape anyway,
330 		 * we have to replay those blocks with rollforward.
331 		 *
332 		 * fixme: punt for now.
333 		 */
334 		if (slaves[f].sent) {
335 			if (atomic(read, slaves[f].fd, (char *)&got, sizeof got)
336 			    != sizeof got) {
337 				perror("  DUMP: error reading command pipe in master");
338 				dumpabort(0);
339 			}
340 			slaves[f].sent = 0;
341 			if (got != writesize) {
342 				msg("EOT detected in last 2 tape records!\n");
343 				msg("Use a longer tape, decrease the size estimate\n");
344 				quit("or use no size estimate at all.\n");
345 			}
346 		}
347 		(void) close(slaves[f].fd);
348 	}
349 	while (wait((int *)NULL) >= 0)	/* wait for any signals from slaves */
350 		/* void */;
351 
352 	if (pipeout)
353 		return;
354 
355 	msg("Closing %s\n", tape);
356 
357 #ifdef RDUMP
358 	if (host) {
359 		rmtclose();
360 		while (rmtopen(tape, 0) < 0)
361 			sleep(10);
362 		rmtclose();
363 		return;
364 	}
365 #endif
366 	(void) close(tapefd);
367 	while ((f = open(tape, 0)) < 0)
368 		sleep (10);
369 	(void) close(f);
370 }
371 
372 void
373 close_rewind()
374 {
375 	trewind();
376 	if (nexttape)
377 		return;
378 	if (!nogripe) {
379 		msg("Change Volumes: Mount volume #%d\n", tapeno+1);
380 		broadcast("CHANGE DUMP VOLUMES!\7\7\n");
381 	}
382 	while (!query("Is the new volume mounted and ready to go?"))
383 		if (query("Do you want to abort?")) {
384 			dumpabort(0);
385 			/*NOTREACHED*/
386 		}
387 }
388 
389 void
390 rollforward()
391 {
392 	register struct req *p, *q, *prev;
393 	register struct slave *tslp;
394 	int i, size, savedtapea, got;
395 	union u_spcl *ntb, *otb;
396 	tslp = &slaves[SLAVES];
397 	ntb = (union u_spcl *)tslp->tblock[1];
398 
399 	/*
400 	 * Each of the N slaves should have requests that need to
401 	 * be replayed on the next tape.  Use the extra slave buffers
402 	 * (slaves[SLAVES]) to construct request lists to be sent to
403 	 * each slave in turn.
404 	 */
405 	for (i = 0; i < SLAVES; i++) {
406 		q = &tslp->req[1];
407 		otb = (union u_spcl *)slp->tblock;
408 
409 		/*
410 		 * For each request in the current slave, copy it to tslp.
411 		 */
412 
413 		prev = NULL;
414 		for (p = slp->req; p->count > 0; p += p->count) {
415 			*q = *p;
416 			if (p->dblk == 0)
417 				*ntb++ = *otb++; /* copy the datablock also */
418 			prev = q;
419 			q += q->count;
420 		}
421 		if (prev == NULL)
422 			quit("rollforward: protocol botch");
423 		if (prev->dblk != 0)
424 			prev->count -= 1;
425 		else
426 			ntb--;
427 		q -= 1;
428 		q->count = 0;
429 		q = &tslp->req[0];
430 		if (i == 0) {
431 			q->dblk = 0;
432 			q->count = 1;
433 			trecno = 0;
434 			nextblock = tslp->tblock;
435 			savedtapea = spcl.c_tapea;
436 			spcl.c_tapea = slp->tapea;
437 			startnewtape(0);
438 			spcl.c_tapea = savedtapea;
439 			lastspclrec = savedtapea - 1;
440 		}
441 		size = (char *)ntb - (char *)q;
442 		if (atomic(write, slp->fd, (char *)q, size) != size) {
443 			perror("  DUMP: error writing command pipe");
444 			dumpabort(0);
445 		}
446 		slp->sent = 1;
447 		if (++slp >= &slaves[SLAVES])
448 			slp = &slaves[0];
449 
450 		q->count = 1;
451 
452 		if (prev->dblk != 0) {
453 			/*
454 			 * If the last one was a disk block, make the
455 			 * first of this one be the last bit of that disk
456 			 * block...
457 			 */
458 			q->dblk = prev->dblk +
459 				prev->count * (TP_BSIZE / DEV_BSIZE);
460 			ntb = (union u_spcl *)tslp->tblock;
461 		} else {
462 			/*
463 			 * It wasn't a disk block.  Copy the data to its
464 			 * new location in the buffer.
465 			 */
466 			q->dblk = 0;
467 			*((union u_spcl *)tslp->tblock) = *ntb;
468 			ntb = (union u_spcl *)tslp->tblock[1];
469 		}
470 	}
471 	slp->req[0] = *q;
472 	nextblock = slp->tblock;
473 	if (q->dblk == 0)
474 		nextblock++;
475 	trecno = 1;
476 
477 	/*
478 	 * Clear the first slaves' response.  One hopes that it
479 	 * worked ok, otherwise the tape is much too short!
480 	 */
481 	if (slp->sent) {
482 		if (atomic(read, slp->fd, (char *)&got, sizeof got)
483 		    != sizeof got) {
484 			perror("  DUMP: error reading command pipe in master");
485 			dumpabort(0);
486 		}
487 		slp->sent = 0;
488 
489 		if (got != writesize) {
490 			quit("EOT detected at start of the tape!\n");
491 		}
492 	}
493 }
494 
495 /*
496  * We implement taking and restoring checkpoints on the tape level.
497  * When each tape is opened, a new process is created by forking; this
498  * saves all of the necessary context in the parent.  The child
499  * continues the dump; the parent waits around, saving the context.
500  * If the child returns X_REWRITE, then it had problems writing that tape;
501  * this causes the parent to fork again, duplicating the context, and
502  * everything continues as if nothing had happened.
503  */
504 void
505 startnewtape(top)
506 	int top;
507 {
508 	int	parentpid;
509 	int	childpid;
510 	int	status;
511 	int	waitpid;
512 	char	*p;
513 #ifdef sunos
514 	void	(*interrupt_save)();
515 #else
516 	sig_t	interrupt_save;
517 #endif
518 
519 	interrupt_save = signal(SIGINT, SIG_IGN);
520 	parentpid = getpid();
521 
522 restore_check_point:
523 	(void)signal(SIGINT, interrupt_save);
524 	/*
525 	 *	All signals are inherited...
526 	 */
527 	childpid = fork();
528 	if (childpid < 0) {
529 		msg("Context save fork fails in parent %d\n", parentpid);
530 		Exit(X_ABORT);
531 	}
532 	if (childpid != 0) {
533 		/*
534 		 *	PARENT:
535 		 *	save the context by waiting
536 		 *	until the child doing all of the work returns.
537 		 *	don't catch the interrupt
538 		 */
539 		signal(SIGINT, SIG_IGN);
540 #ifdef TDEBUG
541 		msg("Tape: %d; parent process: %d child process %d\n",
542 			tapeno+1, parentpid, childpid);
543 #endif /* TDEBUG */
544 		while ((waitpid = wait(&status)) != childpid)
545 			msg("Parent %d waiting for child %d has another child %d return\n",
546 				parentpid, childpid, waitpid);
547 		if (status & 0xFF) {
548 			msg("Child %d returns LOB status %o\n",
549 				childpid, status&0xFF);
550 		}
551 		status = (status >> 8) & 0xFF;
552 #ifdef TDEBUG
553 		switch(status) {
554 			case X_FINOK:
555 				msg("Child %d finishes X_FINOK\n", childpid);
556 				break;
557 			case X_ABORT:
558 				msg("Child %d finishes X_ABORT\n", childpid);
559 				break;
560 			case X_REWRITE:
561 				msg("Child %d finishes X_REWRITE\n", childpid);
562 				break;
563 			default:
564 				msg("Child %d finishes unknown %d\n",
565 					childpid, status);
566 				break;
567 		}
568 #endif /* TDEBUG */
569 		switch(status) {
570 			case X_FINOK:
571 				Exit(X_FINOK);
572 			case X_ABORT:
573 				Exit(X_ABORT);
574 			case X_REWRITE:
575 				goto restore_check_point;
576 			default:
577 				msg("Bad return code from dump: %d\n", status);
578 				Exit(X_ABORT);
579 		}
580 		/*NOTREACHED*/
581 	} else {	/* we are the child; just continue */
582 #ifdef TDEBUG
583 		sleep(4);	/* allow time for parent's message to get out */
584 		msg("Child on Tape %d has parent %d, my pid = %d\n",
585 			tapeno+1, parentpid, getpid());
586 #endif /* TDEBUG */
587 		/*
588 		 * If we have a name like "/dev/rmt0,/dev/rmt1",
589 		 * use the name before the comma first, and save
590 		 * the remaining names for subsequent volumes.
591 		 */
592 		tapeno++;               /* current tape sequence */
593 		if (nexttape || strchr(tape, ',')) {
594 			if (nexttape && *nexttape)
595 				tape = nexttape;
596 			if ((p = strchr(tape, ',')) != NULL) {
597 				*p = '\0';
598 				nexttape = p + 1;
599 			} else
600 				nexttape = NULL;
601 			msg("Dumping volume %d on %s\n", tapeno, tape);
602 		}
603 #ifdef RDUMP
604 		while ((tapefd = (host ? rmtopen(tape, 2) :
605 			pipeout ? 1 : open(tape, O_WRONLY|O_CREAT, 0666))) < 0)
606 #else
607 		while ((tapefd = (pipeout ? 1 :
608 				  open(tape, O_WRONLY|O_CREAT, 0666))) < 0)
609 #endif
610 		    {
611 			msg("Cannot open output \"%s\".\n", tape);
612 			if (!query("Do you want to retry the open?"))
613 				dumpabort(0);
614 		}
615 
616 		enslave();  /* Share open tape file descriptor with slaves */
617 
618 		asize = 0;
619 		blocksthisvol = 0;
620 		if (top)
621 			newtape++;		/* new tape signal */
622 		spcl.c_count = slp->count;
623 		/*
624 		 * measure firstrec in TP_BSIZE units since restore doesn't
625 		 * know the correct ntrec value...
626 		 */
627 		spcl.c_firstrec = slp->firstrec;
628 		spcl.c_volume++;
629 		spcl.c_type = TS_TAPE;
630 		spcl.c_flags |= DR_NEWHEADER;
631 		writeheader((ino_t)slp->inode);
632 		spcl.c_flags &=~ DR_NEWHEADER;
633 		if (tapeno > 1)
634 			msg("Volume %d begins with blocks from inode %d\n",
635 				tapeno, slp->inode);
636 	}
637 }
638 
639 void
640 dumpabort(signo)
641 	int signo;
642 {
643 
644 	if (master != 0 && master != getpid())
645 		/* Signals master to call dumpabort */
646 		(void) kill(master, SIGTERM);
647 	else {
648 		killall();
649 		msg("The ENTIRE dump is aborted.\n");
650 	}
651 #ifdef RDUMP
652 	rmtclose();
653 #endif
654 	Exit(X_ABORT);
655 }
656 
657 __dead void
658 Exit(status)
659 	int status;
660 {
661 
662 #ifdef TDEBUG
663 	msg("pid = %d exits with status %d\n", getpid(), status);
664 #endif /* TDEBUG */
665 	exit(status);
666 }
667 
668 /*
669  * proceed - handler for SIGUSR2, used to synchronize IO between the slaves.
670  */
671 void
672 proceed(signo)
673 	int signo;
674 {
675 
676 	if (ready)
677 		longjmp(jmpbuf, 1);
678 	caught++;
679 }
680 
681 void
682 enslave()
683 {
684 	int cmd[2];
685 	register int i, j;
686 
687 	master = getpid();
688 
689 	signal(SIGTERM, dumpabort);  /* Slave sends SIGTERM on dumpabort() */
690 	signal(SIGPIPE, sigpipe);
691 	signal(SIGUSR1, tperror);    /* Slave sends SIGUSR1 on tape errors */
692 	signal(SIGUSR2, proceed);    /* Slave sends SIGUSR2 to next slave */
693 
694 	for (i = 0; i < SLAVES; i++) {
695 		if (i == slp - &slaves[0]) {
696 			caught = 1;
697 		} else {
698 			caught = 0;
699 		}
700 
701 		if (socketpair(AF_UNIX, SOCK_STREAM, 0, cmd) < 0 ||
702 		    (slaves[i].pid = fork()) < 0)
703 			quit("too many slaves, %d (recompile smaller): %s\n",
704 			    i, strerror(errno));
705 
706 		slaves[i].fd = cmd[1];
707 		slaves[i].sent = 0;
708 		if (slaves[i].pid == 0) { 	    /* Slave starts up here */
709 			for (j = 0; j <= i; j++)
710 			        (void) close(slaves[j].fd);
711 			signal(SIGINT, SIG_IGN);    /* Master handles this */
712 			doslave(cmd[0], i);
713 			Exit(X_FINOK);
714 		}
715 	}
716 
717 	for (i = 0; i < SLAVES; i++)
718 		(void) atomic(write, slaves[i].fd,
719 			      (char *) &slaves[(i + 1) % SLAVES].pid,
720 		              sizeof slaves[0].pid);
721 
722 	master = 0;
723 }
724 
725 void
726 killall()
727 {
728 	register int i;
729 
730 	for (i = 0; i < SLAVES; i++)
731 		if (slaves[i].pid > 0)
732 			(void) kill(slaves[i].pid, SIGKILL);
733 }
734 
735 /*
736  * Synchronization - each process has a lockfile, and shares file
737  * descriptors to the following process's lockfile.  When our write
738  * completes, we release our lock on the following process's lock-
739  * file, allowing the following process to lock it and proceed. We
740  * get the lock back for the next cycle by swapping descriptors.
741  */
742 static void
743 doslave(cmd, slave_number)
744 	register int cmd;
745         int slave_number;
746 {
747 	register int nread;
748 	int nextslave, size, wrote, eot_count;
749 	sigset_t sigset;
750 
751 	/*
752 	 * Need our own seek pointer.
753 	 */
754 	(void) close(diskfd);
755 	if ((diskfd = open(disk, O_RDONLY)) < 0)
756 		quit("slave couldn't reopen disk: %s\n", strerror(errno));
757 
758 	/*
759 	 * Need the pid of the next slave in the loop...
760 	 */
761 	if ((nread = atomic(read, cmd, (char *)&nextslave, sizeof nextslave))
762 	    != sizeof nextslave) {
763 		quit("master/slave protocol botched - didn't get pid of next slave.\n");
764 	}
765 
766 	/*
767 	 * Get list of blocks to dump, read the blocks into tape buffer
768 	 */
769 	while ((nread = atomic(read, cmd, (char *)slp->req, reqsiz)) == reqsiz) {
770 		register struct req *p = slp->req;
771 
772 		for (trecno = 0; trecno < ntrec;
773 		     trecno += p->count, p += p->count) {
774 			if (p->dblk) {
775 				bread(p->dblk, slp->tblock[trecno],
776 					p->count * TP_BSIZE);
777 			} else {
778 				if (p->count != 1 || atomic(read, cmd,
779 				    (char *)slp->tblock[trecno],
780 				    TP_BSIZE) != TP_BSIZE)
781 				       quit("master/slave protocol botched.\n");
782 			}
783 		}
784 		if (setjmp(jmpbuf) == 0) {
785 			ready = 1;
786 			if (!caught)
787 				(void) pause();
788 		}
789 		ready = 0;
790 		caught = 0;
791 
792 		/* Try to write the data... */
793 		eot_count = 0;
794 		size = 0;
795 
796 		while (eot_count < 10 && size < writesize) {
797 #ifdef RDUMP
798 			if (host)
799 				wrote = rmtwrite(slp->tblock[0]+size,
800 				    writesize-size);
801 			else
802 #endif
803 				wrote = write(tapefd, slp->tblock[0]+size,
804 				    writesize-size);
805 #ifdef WRITEDEBUG
806 			printf("slave %d wrote %d\n", slave_number, wrote);
807 #endif
808 			if (wrote < 0)
809 				break;
810 			if (wrote == 0)
811 				eot_count++;
812 			size += wrote;
813 		}
814 
815 #ifdef WRITEDEBUG
816 		if (size != writesize)
817 		 printf("slave %d only wrote %d out of %d bytes and gave up.\n",
818 		     slave_number, size, writesize);
819 #endif
820 
821 		if (eot_count > 0)
822 			size = 0;
823 
824 		/*
825 		 * fixme: Pyramids running OSx return ENOSPC
826 		 * at EOT on 1/2 inch drives.
827 		 */
828 		if (size < 0) {
829 			(void) kill(master, SIGUSR1);
830 			sigemptyset(&sigset);
831 			for (;;)
832 				sigsuspend(&sigset);
833 		} else {
834 			/*
835 			 * pass size of write back to master
836 			 * (for EOT handling)
837 			 */
838 			(void) atomic(write, cmd, (char *)&size, sizeof size);
839 		}
840 
841 		/*
842 		 * If partial write, don't want next slave to go.
843 		 * Also jolts him awake.
844 		 */
845 		(void) kill(nextslave, SIGUSR2);
846 	}
847 	if (nread != 0)
848 		quit("error reading command pipe: %s\n", strerror(errno));
849 }
850 
851 /*
852  * Since a read from a pipe may not return all we asked for,
853  * or a write may not write all we ask if we get a signal,
854  * loop until the count is satisfied (or error).
855  */
856 static ssize_t
857 atomic(func, fd, buf, count)
858 	ssize_t (*func)();
859 	int fd;
860 	char *buf;
861 	int count;
862 {
863 	ssize_t got, need = count;
864 
865 	while ((got = (*func)(fd, buf, need)) > 0 && (need -= got) > 0)
866 		buf += got;
867 	return (got < 0 ? got : count - need);
868 }
869