xref: /freebsd/sbin/dump/tape.c (revision 8fa113e5fc65fe6abc757f0089f477a87ee4d185)
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 
47 #include <ufs/ufs/dinode.h>
48 #include <ufs/ffs/fs.h>
49 
50 #include <protocols/dumprestore.h>
51 
52 #include <errno.h>
53 #include <fcntl.h>
54 #include <setjmp.h>
55 #include <signal.h>
56 #include <stdio.h>
57 #ifdef __STDC__
58 #include <stdlib.h>
59 #include <string.h>
60 #include <unistd.h>
61 #else
62 int	write(), read();
63 #endif
64 
65 #include "dump.h"
66 
67 int	writesize;		/* size of malloc()ed buffer for tape */
68 long	lastspclrec = -1;	/* tape block number of last written header */
69 int	trecno = 0;		/* next record to write in current block */
70 extern	long blocksperfile;	/* number of blocks per output file */
71 long	blocksthisvol;		/* number of blocks on current output file */
72 extern	int ntrec;		/* blocking factor on tape */
73 extern	int cartridge;
74 extern	char *host;
75 char	*nexttape;
76 
77 static	int atomic __P((ssize_t (*)(), int, char *, int));
78 static	void doslave __P((int, int));
79 static	void enslave __P((void));
80 static	void flushtape __P((void));
81 static	void killall __P((void));
82 static	void rollforward __P((void));
83 
84 /*
85  * Concurrent dump mods (Caltech) - disk block reading and tape writing
86  * are exported to several slave processes.  While one slave writes the
87  * tape, the others read disk blocks; they pass control of the tape in
88  * a ring via signals. The parent process traverses the filesystem and
89  * sends writeheader()'s and lists of daddr's to the slaves via pipes.
90  * The following structure defines the instruction packets sent to slaves.
91  */
92 struct req {
93 	daddr_t dblk;
94 	int count;
95 };
96 int reqsiz;
97 
98 #define SLAVES 3		/* 1 slave writing, 1 reading, 1 for slack */
99 struct slave {
100 	int tapea;		/* header number at start of this chunk */
101 	int count;		/* count to next header (used for TS_TAPE */
102 				/* after EOT) */
103 	int inode;		/* inode that we are currently dealing with */
104 	int fd;			/* FD for this slave */
105 	int pid;		/* PID for this slave */
106 	int sent;		/* 1 == we've sent this slave requests */
107 	int firstrec;		/* record number of this block */
108 	char (*tblock)[TP_BSIZE]; /* buffer for data blocks */
109 	struct req *req;	/* buffer for requests */
110 } slaves[SLAVES+1];
111 struct slave *slp;
112 
113 char	(*nextblock)[TP_BSIZE];
114 
115 int master;		/* pid of master, for sending error signals */
116 int tenths;		/* length of tape used per block written */
117 static int caught;	/* have we caught the signal to proceed? */
118 static int ready;	/* have we reached the lock point without having */
119 			/* received the SIGUSR2 signal from the prev slave? */
120 static jmp_buf jmpbuf;	/* where to jump to if we are ready when the */
121 			/* SIGUSR2 arrives from the previous slave */
122 
123 int
124 alloctape()
125 {
126 	int pgoff = getpagesize() - 1;
127 	char *buf;
128 	int i;
129 
130 	writesize = ntrec * TP_BSIZE;
131 	reqsiz = (ntrec + 1) * sizeof(struct req);
132 	/*
133 	 * CDC 92181's and 92185's make 0.8" gaps in 1600-bpi start/stop mode
134 	 * (see DEC TU80 User's Guide).  The shorter gaps of 6250-bpi require
135 	 * repositioning after stopping, i.e, streaming mode, where the gap is
136 	 * variable, 0.30" to 0.45".  The gap is maximal when the tape stops.
137 	 */
138 	if (blocksperfile == 0 && !unlimited)
139 		tenths = writesize / density +
140 		    (cartridge ? 16 : density == 625 ? 5 : 8);
141 	/*
142 	 * Allocate tape buffer contiguous with the array of instruction
143 	 * packets, so flushtape() can write them together with one write().
144 	 * Align tape buffer on page boundary to speed up tape write().
145 	 */
146 	for (i = 0; i <= SLAVES; i++) {
147 		buf = (char *)
148 		    malloc((unsigned)(reqsiz + writesize + pgoff + TP_BSIZE));
149 		if (buf == NULL)
150 			return(0);
151 		slaves[i].tblock = (char (*)[TP_BSIZE])
152 		    (((long)&buf[ntrec + 1] + pgoff) &~ pgoff);
153 		slaves[i].req = (struct req *)slaves[i].tblock - ntrec - 1;
154 	}
155 	slp = &slaves[0];
156 	slp->count = 1;
157 	slp->tapea = 0;
158 	slp->firstrec = 0;
159 	nextblock = slp->tblock;
160 	return(1);
161 }
162 
163 void
164 writerec(dp, isspcl)
165 	char *dp;
166 	int isspcl;
167 {
168 
169 	slp->req[trecno].dblk = (daddr_t)0;
170 	slp->req[trecno].count = 1;
171 #ifndef	__alpha__
172 	*(union u_spcl *)(*(nextblock)++) = *(union u_spcl *)dp;
173 #else
174 	bcopy(dp, *(nextblock)++, sizeof (union u_spcl));
175 #endif
176 	if (isspcl)
177 		lastspclrec = spcl.c_tapea;
178 	trecno++;
179 	spcl.c_tapea++;
180 	if (trecno >= ntrec)
181 		flushtape();
182 }
183 
184 void
185 dumpblock(blkno, size)
186 	daddr_t blkno;
187 	int size;
188 {
189 	int avail, tpblks, dblkno;
190 
191 	dblkno = fsbtodb(sblock, blkno);
192 	tpblks = size >> tp_bshift;
193 	while ((avail = MIN(tpblks, ntrec - trecno)) > 0) {
194 		slp->req[trecno].dblk = dblkno;
195 		slp->req[trecno].count = avail;
196 		trecno += avail;
197 		spcl.c_tapea += avail;
198 		if (trecno >= ntrec)
199 			flushtape();
200 		dblkno += avail << (tp_bshift - dev_bshift);
201 		tpblks -= avail;
202 	}
203 }
204 
205 int	nogripe = 0;
206 
207 void
208 tperror(signo)
209 	int signo;
210 {
211 
212 	if (pipeout) {
213 		msg("write error on %s\n", tape);
214 		quit("Cannot recover\n");
215 		/* NOTREACHED */
216 	}
217 	msg("write error %d blocks into volume %d\n", blocksthisvol, tapeno);
218 	broadcast("DUMP WRITE ERROR!\n");
219 	if (!query("Do you want to restart?"))
220 		dumpabort(0);
221 	msg("Closing this volume.  Prepare to restart with new media;\n");
222 	msg("this dump volume will be rewritten.\n");
223 	killall();
224 	nogripe = 1;
225 	close_rewind();
226 	Exit(X_REWRITE);
227 }
228 
229 void
230 sigpipe(signo)
231 	int signo;
232 {
233 
234 	quit("Broken pipe\n");
235 }
236 
237 static void
238 flushtape()
239 {
240 	int i, blks, got;
241 	long lastfirstrec;
242 
243 	int siz = (char *)nextblock - (char *)slp->req;
244 
245 	slp->req[trecno].count = 0;			/* Sentinel */
246 
247 	if (atomic(write, slp->fd, (char *)slp->req, siz) != siz)
248 		quit("error writing command pipe: %s\n", strerror(errno));
249 	slp->sent = 1; /* we sent a request, read the response later */
250 
251 	lastfirstrec = slp->firstrec;
252 
253 	if (++slp >= &slaves[SLAVES])
254 		slp = &slaves[0];
255 
256 	/* Read results back from next slave */
257 	if (slp->sent) {
258 		if (atomic(read, slp->fd, (char *)&got, sizeof got)
259 		    != sizeof got) {
260 			perror("  DUMP: error reading command pipe in master");
261 			dumpabort(0);
262 		}
263 		slp->sent = 0;
264 
265 		/* Check for end of tape */
266 		if (got < writesize) {
267 			msg("End of tape detected\n");
268 
269 			/*
270 			 * Drain the results, don't care what the values were.
271 			 * If we read them here then trewind won't...
272 			 */
273 			for (i = 0; i < SLAVES; i++) {
274 				if (slaves[i].sent) {
275 					if (atomic(read, slaves[i].fd,
276 					    (char *)&got, sizeof got)
277 					    != sizeof got) {
278 						perror("  DUMP: error reading command pipe in master");
279 						dumpabort(0);
280 					}
281 					slaves[i].sent = 0;
282 				}
283 			}
284 
285 			close_rewind();
286 			rollforward();
287 			return;
288 		}
289 	}
290 
291 	blks = 0;
292 	if (spcl.c_type != TS_END) {
293 		for (i = 0; i < spcl.c_count; i++)
294 			if (spcl.c_addr[i] != 0)
295 				blks++;
296 	}
297 	slp->count = lastspclrec + blks + 1 - spcl.c_tapea;
298 	slp->tapea = spcl.c_tapea;
299 	slp->firstrec = lastfirstrec + ntrec;
300 	slp->inode = curino;
301 	nextblock = slp->tblock;
302 	trecno = 0;
303 	asize += tenths;
304 	blockswritten += ntrec;
305 	blocksthisvol += ntrec;
306 	if (!pipeout && !unlimited && (blocksperfile ?
307 	    (blocksthisvol >= blocksperfile) : (asize > tsize))) {
308 		close_rewind();
309 		startnewtape(0);
310 	}
311 	timeest();
312 }
313 
314 void
315 trewind()
316 {
317 	int f;
318 	int got;
319 
320 	for (f = 0; f < SLAVES; f++) {
321 		/*
322 		 * Drain the results, but unlike EOT we DO (or should) care
323 		 * what the return values were, since if we detect EOT after
324 		 * we think we've written the last blocks to the tape anyway,
325 		 * we have to replay those blocks with rollforward.
326 		 *
327 		 * fixme: punt for now.
328 		 */
329 		if (slaves[f].sent) {
330 			if (atomic(read, slaves[f].fd, (char *)&got, sizeof got)
331 			    != sizeof got) {
332 				perror("  DUMP: error reading command pipe in master");
333 				dumpabort(0);
334 			}
335 			slaves[f].sent = 0;
336 			if (got != writesize) {
337 				msg("EOT detected in last 2 tape records!\n");
338 				msg("Use a longer tape, decrease the size estimate\n");
339 				quit("or use no size estimate at all.\n");
340 			}
341 		}
342 		(void) close(slaves[f].fd);
343 	}
344 	while (wait((int *)NULL) >= 0)	/* wait for any signals from slaves */
345 		/* void */;
346 
347 	if (pipeout)
348 		return;
349 
350 	msg("Closing %s\n", tape);
351 
352 #ifdef RDUMP
353 	if (host) {
354 		rmtclose();
355 		while (rmtopen(tape, 0) < 0)
356 			sleep(10);
357 		rmtclose();
358 		return;
359 	}
360 #endif
361 	(void) close(tapefd);
362 	while ((f = open(tape, 0)) < 0)
363 		sleep (10);
364 	(void) close(f);
365 }
366 
367 void
368 close_rewind()
369 {
370 	time_t tstart_changevol, tend_changevol;
371 
372 	trewind();
373 	if (nexttape)
374 		return;
375 	(void)time((time_t *)&(tstart_changevol));
376 	if (!nogripe) {
377 		msg("Change Volumes: Mount volume #%d\n", tapeno+1);
378 		broadcast("CHANGE DUMP VOLUMES!\a\a\n");
379 	}
380 	while (!query("Is the new volume mounted and ready to go?"))
381 		if (query("Do you want to abort?")) {
382 			dumpabort(0);
383 			/*NOTREACHED*/
384 		}
385 	(void)time((time_t *)&(tend_changevol));
386 	if ((tstart_changevol != (time_t)-1) && (tend_changevol != (time_t)-1))
387 		tstart_writing += (tend_changevol - tstart_changevol);
388 }
389 
390 void
391 rollforward()
392 {
393 	struct req *p, *q, *prev;
394 	struct slave *tslp;
395 	int i, size, savedtapea, got;
396 	union u_spcl *ntb, *otb;
397 	tslp = &slaves[SLAVES];
398 	ntb = (union u_spcl *)tslp->tblock[1];
399 
400 	/*
401 	 * Each of the N slaves should have requests that need to
402 	 * be replayed on the next tape.  Use the extra slave buffers
403 	 * (slaves[SLAVES]) to construct request lists to be sent to
404 	 * each slave in turn.
405 	 */
406 	for (i = 0; i < SLAVES; i++) {
407 		q = &tslp->req[1];
408 		otb = (union u_spcl *)slp->tblock;
409 
410 		/*
411 		 * For each request in the current slave, copy it to tslp.
412 		 */
413 
414 		prev = NULL;
415 		for (p = slp->req; p->count > 0; p += p->count) {
416 			*q = *p;
417 			if (p->dblk == 0)
418 				*ntb++ = *otb++; /* copy the datablock also */
419 			prev = q;
420 			q += q->count;
421 		}
422 		if (prev == NULL)
423 			quit("rollforward: protocol botch");
424 		if (prev->dblk != 0)
425 			prev->count -= 1;
426 		else
427 			ntb--;
428 		q -= 1;
429 		q->count = 0;
430 		q = &tslp->req[0];
431 		if (i == 0) {
432 			q->dblk = 0;
433 			q->count = 1;
434 			trecno = 0;
435 			nextblock = tslp->tblock;
436 			savedtapea = spcl.c_tapea;
437 			spcl.c_tapea = slp->tapea;
438 			startnewtape(0);
439 			spcl.c_tapea = savedtapea;
440 			lastspclrec = savedtapea - 1;
441 		}
442 		size = (char *)ntb - (char *)q;
443 		if (atomic(write, slp->fd, (char *)q, size) != size) {
444 			perror("  DUMP: error writing command pipe");
445 			dumpabort(0);
446 		}
447 		slp->sent = 1;
448 		if (++slp >= &slaves[SLAVES])
449 			slp = &slaves[0];
450 
451 		q->count = 1;
452 
453 		if (prev->dblk != 0) {
454 			/*
455 			 * If the last one was a disk block, make the
456 			 * first of this one be the last bit of that disk
457 			 * block...
458 			 */
459 			q->dblk = prev->dblk +
460 				prev->count * (TP_BSIZE / DEV_BSIZE);
461 			ntb = (union u_spcl *)tslp->tblock;
462 		} else {
463 			/*
464 			 * It wasn't a disk block.  Copy the data to its
465 			 * new location in the buffer.
466 			 */
467 			q->dblk = 0;
468 			*((union u_spcl *)tslp->tblock) = *ntb;
469 			ntb = (union u_spcl *)tslp->tblock[1];
470 		}
471 	}
472 	slp->req[0] = *q;
473 	nextblock = slp->tblock;
474 	if (q->dblk == 0)
475 		nextblock++;
476 	trecno = 1;
477 
478 	/*
479 	 * Clear the first slaves' response.  One hopes that it
480 	 * worked ok, otherwise the tape is much too short!
481 	 */
482 	if (slp->sent) {
483 		if (atomic(read, slp->fd, (char *)&got, sizeof got)
484 		    != sizeof got) {
485 			perror("  DUMP: error reading command pipe in master");
486 			dumpabort(0);
487 		}
488 		slp->sent = 0;
489 
490 		if (got != writesize) {
491 			quit("EOT detected at start of the tape!\n");
492 		}
493 	}
494 }
495 
496 /*
497  * We implement taking and restoring checkpoints on the tape level.
498  * When each tape is opened, a new process is created by forking; this
499  * saves all of the necessary context in the parent.  The child
500  * continues the dump; the parent waits around, saving the context.
501  * If the child returns X_REWRITE, then it had problems writing that tape;
502  * this causes the parent to fork again, duplicating the context, and
503  * everything continues as if nothing had happened.
504  */
505 void
506 startnewtape(top)
507 	int top;
508 {
509 	int	parentpid;
510 	int	childpid;
511 	int	status;
512 	int	waitpid;
513 	char	*p;
514 	sig_t	interrupt_save;
515 
516 	interrupt_save = signal(SIGINT, SIG_IGN);
517 	parentpid = getpid();
518 
519 restore_check_point:
520 	(void)signal(SIGINT, interrupt_save);
521 	/*
522 	 *	All signals are inherited...
523 	 */
524 	childpid = fork();
525 	if (childpid < 0) {
526 		msg("Context save fork fails in parent %d\n", parentpid);
527 		Exit(X_ABORT);
528 	}
529 	if (childpid != 0) {
530 		/*
531 		 *	PARENT:
532 		 *	save the context by waiting
533 		 *	until the child doing all of the work returns.
534 		 *	don't catch the interrupt
535 		 */
536 		signal(SIGINT, SIG_IGN);
537 #ifdef TDEBUG
538 		msg("Tape: %d; parent process: %d child process %d\n",
539 			tapeno+1, parentpid, childpid);
540 #endif /* TDEBUG */
541 		while ((waitpid = wait(&status)) != childpid)
542 			msg("Parent %d waiting for child %d has another child %d return\n",
543 				parentpid, childpid, waitpid);
544 		if (status & 0xFF) {
545 			msg("Child %d returns LOB status %o\n",
546 				childpid, status&0xFF);
547 		}
548 		status = (status >> 8) & 0xFF;
549 #ifdef TDEBUG
550 		switch(status) {
551 			case X_FINOK:
552 				msg("Child %d finishes X_FINOK\n", childpid);
553 				break;
554 			case X_ABORT:
555 				msg("Child %d finishes X_ABORT\n", childpid);
556 				break;
557 			case X_REWRITE:
558 				msg("Child %d finishes X_REWRITE\n", childpid);
559 				break;
560 			default:
561 				msg("Child %d finishes unknown %d\n",
562 					childpid, status);
563 				break;
564 		}
565 #endif /* TDEBUG */
566 		switch(status) {
567 			case X_FINOK:
568 				Exit(X_FINOK);
569 			case X_ABORT:
570 				Exit(X_ABORT);
571 			case X_REWRITE:
572 				goto restore_check_point;
573 			default:
574 				msg("Bad return code from dump: %d\n", status);
575 				Exit(X_ABORT);
576 		}
577 		/*NOTREACHED*/
578 	} else {	/* we are the child; just continue */
579 #ifdef TDEBUG
580 		sleep(4);	/* allow time for parent's message to get out */
581 		msg("Child on Tape %d has parent %d, my pid = %d\n",
582 			tapeno+1, parentpid, getpid());
583 #endif /* TDEBUG */
584 		/*
585 		 * If we have a name like "/dev/rmt0,/dev/rmt1",
586 		 * use the name before the comma first, and save
587 		 * the remaining names for subsequent volumes.
588 		 */
589 		tapeno++;               /* current tape sequence */
590 		if (nexttape || strchr(tape, ',')) {
591 			if (nexttape && *nexttape)
592 				tape = nexttape;
593 			if ((p = strchr(tape, ',')) != NULL) {
594 				*p = '\0';
595 				nexttape = p + 1;
596 			} else
597 				nexttape = NULL;
598 			msg("Dumping volume %d on %s\n", tapeno, tape);
599 		}
600 #ifdef RDUMP
601 		while ((tapefd = (host ? rmtopen(tape, 2) :
602 			pipeout ? 1 : open(tape, O_WRONLY|O_CREAT, 0666))) < 0)
603 #else
604 		while ((tapefd = (pipeout ? 1 :
605 				  open(tape, O_WRONLY|O_CREAT, 0666))) < 0)
606 #endif
607 		    {
608 			msg("Cannot open output \"%s\".\n", tape);
609 			if (!query("Do you want to retry the open?"))
610 				dumpabort(0);
611 		}
612 
613 		enslave();  /* Share open tape file descriptor with slaves */
614 
615 		asize = 0;
616 		blocksthisvol = 0;
617 		if (top)
618 			newtape++;		/* new tape signal */
619 		spcl.c_count = slp->count;
620 		/*
621 		 * measure firstrec in TP_BSIZE units since restore doesn't
622 		 * know the correct ntrec value...
623 		 */
624 		spcl.c_firstrec = slp->firstrec;
625 		spcl.c_volume++;
626 		spcl.c_type = TS_TAPE;
627 		spcl.c_flags |= DR_NEWHEADER;
628 		writeheader((ino_t)slp->inode);
629 		spcl.c_flags &=~ DR_NEWHEADER;
630 		if (tapeno > 1)
631 			msg("Volume %d begins with blocks from inode %d\n",
632 				tapeno, slp->inode);
633 	}
634 }
635 
636 void
637 dumpabort(signo)
638 	int signo;
639 {
640 
641 	if (master != 0 && master != getpid())
642 		/* Signals master to call dumpabort */
643 		(void) kill(master, SIGTERM);
644 	else {
645 		killall();
646 		msg("The ENTIRE dump is aborted.\n");
647 	}
648 #ifdef RDUMP
649 	rmtclose();
650 #endif
651 	Exit(X_ABORT);
652 }
653 
654 void
655 Exit(status)
656 	int status;
657 {
658 
659 #ifdef TDEBUG
660 	msg("pid = %d exits with status %d\n", getpid(), status);
661 #endif /* TDEBUG */
662 	exit(status);
663 }
664 
665 /*
666  * proceed - handler for SIGUSR2, used to synchronize IO between the slaves.
667  */
668 void
669 proceed(signo)
670 	int signo;
671 {
672 
673 	if (ready)
674 		longjmp(jmpbuf, 1);
675 	caught++;
676 }
677 
678 void
679 enslave()
680 {
681 	int cmd[2];
682 	int i, j;
683 
684 	master = getpid();
685 
686 	signal(SIGTERM, dumpabort);  /* Slave sends SIGTERM on dumpabort() */
687 	signal(SIGPIPE, sigpipe);
688 	signal(SIGUSR1, tperror);    /* Slave sends SIGUSR1 on tape errors */
689 	signal(SIGUSR2, proceed);    /* Slave sends SIGUSR2 to next slave */
690 
691 	for (i = 0; i < SLAVES; i++) {
692 		if (i == slp - &slaves[0]) {
693 			caught = 1;
694 		} else {
695 			caught = 0;
696 		}
697 
698 		if (socketpair(AF_UNIX, SOCK_STREAM, 0, cmd) < 0 ||
699 		    (slaves[i].pid = fork()) < 0)
700 			quit("too many slaves, %d (recompile smaller): %s\n",
701 			    i, strerror(errno));
702 
703 		slaves[i].fd = cmd[1];
704 		slaves[i].sent = 0;
705 		if (slaves[i].pid == 0) { 	    /* Slave starts up here */
706 			for (j = 0; j <= i; j++)
707 			        (void) close(slaves[j].fd);
708 			signal(SIGINT, SIG_IGN);    /* Master handles this */
709 			doslave(cmd[0], i);
710 			Exit(X_FINOK);
711 		}
712 	}
713 
714 	for (i = 0; i < SLAVES; i++)
715 		(void) atomic(write, slaves[i].fd,
716 			      (char *) &slaves[(i + 1) % SLAVES].pid,
717 		              sizeof slaves[0].pid);
718 
719 	master = 0;
720 }
721 
722 void
723 killall()
724 {
725 	int i;
726 
727 	for (i = 0; i < SLAVES; i++)
728 		if (slaves[i].pid > 0) {
729 			(void) kill(slaves[i].pid, SIGKILL);
730 			slaves[i].sent = 0;
731 		}
732 }
733 
734 /*
735  * Synchronization - each process has a lockfile, and shares file
736  * descriptors to the following process's lockfile.  When our write
737  * completes, we release our lock on the following process's lock-
738  * file, allowing the following process to lock it and proceed. We
739  * get the lock back for the next cycle by swapping descriptors.
740  */
741 static void
742 doslave(cmd, slave_number)
743 	int cmd;
744         int slave_number;
745 {
746 	int nread;
747 	int nextslave, size, wrote, eot_count;
748 
749 	/*
750 	 * Need our own seek pointer.
751 	 */
752 	(void) close(diskfd);
753 	if ((diskfd = open(disk, O_RDONLY)) < 0)
754 		quit("slave couldn't reopen disk: %s\n", strerror(errno));
755 
756 	/*
757 	 * Need the pid of the next slave in the loop...
758 	 */
759 	if ((nread = atomic(read, cmd, (char *)&nextslave, sizeof nextslave))
760 	    != sizeof nextslave) {
761 		quit("master/slave protocol botched - didn't get pid of next slave.\n");
762 	}
763 
764 	/*
765 	 * Get list of blocks to dump, read the blocks into tape buffer
766 	 */
767 	while ((nread = atomic(read, cmd, (char *)slp->req, reqsiz)) == reqsiz) {
768 		struct req *p = slp->req;
769 
770 		for (trecno = 0; trecno < ntrec;
771 		     trecno += p->count, p += p->count) {
772 			if (p->dblk) {
773 				bread(p->dblk, slp->tblock[trecno],
774 					p->count * TP_BSIZE);
775 			} else {
776 				if (p->count != 1 || atomic(read, cmd,
777 				    (char *)slp->tblock[trecno],
778 				    TP_BSIZE) != TP_BSIZE)
779 				       quit("master/slave protocol botched.\n");
780 			}
781 		}
782 		if (setjmp(jmpbuf) == 0) {
783 			ready = 1;
784 			if (!caught)
785 				(void) pause();
786 		}
787 		ready = 0;
788 		caught = 0;
789 
790 		/* Try to write the data... */
791 		eot_count = 0;
792 		size = 0;
793 
794 		while (eot_count < 10 && size < writesize) {
795 #ifdef RDUMP
796 			if (host)
797 				wrote = rmtwrite(slp->tblock[0]+size,
798 				    writesize-size);
799 			else
800 #endif
801 				wrote = write(tapefd, slp->tblock[0]+size,
802 				    writesize-size);
803 #ifdef WRITEDEBUG
804 			printf("slave %d wrote %d\n", slave_number, wrote);
805 #endif
806 			if (wrote < 0)
807 				break;
808 			if (wrote == 0)
809 				eot_count++;
810 			size += wrote;
811 		}
812 
813 #ifdef WRITEDEBUG
814 		if (size != writesize)
815 		 printf("slave %d only wrote %d out of %d bytes and gave up.\n",
816 		     slave_number, size, writesize);
817 #endif
818 
819 		/*
820 		 * Handle ENOSPC as an EOT condition.
821 		 */
822 		if (wrote < 0 && errno == ENOSPC) {
823 			wrote = 0;
824 			eot_count++;
825 		}
826 
827 		if (eot_count > 0)
828 			size = 0;
829 
830 		if (wrote < 0) {
831 			(void) kill(master, SIGUSR1);
832 			for (;;)
833 				(void) sigpause(0);
834 		} else {
835 			/*
836 			 * pass size of write back to master
837 			 * (for EOT handling)
838 			 */
839 			(void) atomic(write, cmd, (char *)&size, sizeof size);
840 		}
841 
842 		/*
843 		 * If partial write, don't want next slave to go.
844 		 * Also jolts him awake.
845 		 */
846 		(void) kill(nextslave, SIGUSR2);
847 	}
848 	if (nread != 0)
849 		quit("error reading command pipe: %s\n", strerror(errno));
850 }
851 
852 /*
853  * Since a read from a pipe may not return all we asked for,
854  * or a write may not write all we ask if we get a signal,
855  * loop until the count is satisfied (or error).
856  */
857 static int
858 atomic(func, fd, buf, count)
859 	ssize_t (*func)();
860 	int fd;
861 	char *buf;
862 	int count;
863 {
864 	int got, need = count;
865 
866 	while ((got = (*func)(fd, buf, need)) > 0 && (need -= got) > 0)
867 		buf += got;
868 	return (got < 0 ? got : count - need);
869 }
870