xref: /freebsd/sbin/dump/tape.c (revision 380a989b3223d455375b4fae70fd0b9bdd43bafb)
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 	"$Id: tape.c,v 1.9 1998/06/15 06:58:11 charnier Exp $";
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 	*(union u_spcl *)(*(nextblock)++) = *(union u_spcl *)dp;
178 	if (isspcl)
179 		lastspclrec = spcl.c_tapea;
180 	trecno++;
181 	spcl.c_tapea++;
182 	if (trecno >= ntrec)
183 		flushtape();
184 }
185 
186 void
187 dumpblock(blkno, size)
188 	daddr_t blkno;
189 	int size;
190 {
191 	int avail, tpblks, dblkno;
192 
193 	dblkno = fsbtodb(sblock, blkno);
194 	tpblks = size >> tp_bshift;
195 	while ((avail = MIN(tpblks, ntrec - trecno)) > 0) {
196 		slp->req[trecno].dblk = dblkno;
197 		slp->req[trecno].count = avail;
198 		trecno += avail;
199 		spcl.c_tapea += avail;
200 		if (trecno >= ntrec)
201 			flushtape();
202 		dblkno += avail << (tp_bshift - dev_bshift);
203 		tpblks -= avail;
204 	}
205 }
206 
207 int	nogripe = 0;
208 
209 void
210 tperror(signo)
211 	int signo;
212 {
213 
214 	if (pipeout) {
215 		msg("write error on %s\n", tape);
216 		quit("Cannot recover\n");
217 		/* NOTREACHED */
218 	}
219 	msg("write error %d blocks into volume %d\n", blocksthisvol, tapeno);
220 	broadcast("DUMP WRITE ERROR!\n");
221 	if (!query("Do you want to restart?"))
222 		dumpabort(0);
223 	msg("Closing this volume.  Prepare to restart with new media;\n");
224 	msg("this dump volume will be rewritten.\n");
225 	killall();
226 	nogripe = 1;
227 	close_rewind();
228 	Exit(X_REWRITE);
229 }
230 
231 void
232 sigpipe(signo)
233 	int signo;
234 {
235 
236 	quit("Broken pipe\n");
237 }
238 
239 static void
240 flushtape()
241 {
242 	int i, blks, got;
243 	long lastfirstrec;
244 
245 	int siz = (char *)nextblock - (char *)slp->req;
246 
247 	slp->req[trecno].count = 0;			/* Sentinel */
248 
249 	if (atomic(write, slp->fd, (char *)slp->req, siz) != siz)
250 		quit("error writing command pipe: %s\n", strerror(errno));
251 	slp->sent = 1; /* we sent a request, read the response later */
252 
253 	lastfirstrec = slp->firstrec;
254 
255 	if (++slp >= &slaves[SLAVES])
256 		slp = &slaves[0];
257 
258 	/* Read results back from next slave */
259 	if (slp->sent) {
260 		if (atomic(read, slp->fd, (char *)&got, sizeof got)
261 		    != sizeof got) {
262 			perror("  DUMP: error reading command pipe in master");
263 			dumpabort(0);
264 		}
265 		slp->sent = 0;
266 
267 		/* Check for end of tape */
268 		if (got < writesize) {
269 			msg("End of tape detected\n");
270 
271 			/*
272 			 * Drain the results, don't care what the values were.
273 			 * If we read them here then trewind won't...
274 			 */
275 			for (i = 0; i < SLAVES; i++) {
276 				if (slaves[i].sent) {
277 					if (atomic(read, slaves[i].fd,
278 					    (char *)&got, sizeof got)
279 					    != sizeof got) {
280 						perror("  DUMP: error reading command pipe in master");
281 						dumpabort(0);
282 					}
283 					slaves[i].sent = 0;
284 				}
285 			}
286 
287 			close_rewind();
288 			rollforward();
289 			return;
290 		}
291 	}
292 
293 	blks = 0;
294 	if (spcl.c_type != TS_END) {
295 		for (i = 0; i < spcl.c_count; i++)
296 			if (spcl.c_addr[i] != 0)
297 				blks++;
298 	}
299 	slp->count = lastspclrec + blks + 1 - spcl.c_tapea;
300 	slp->tapea = spcl.c_tapea;
301 	slp->firstrec = lastfirstrec + ntrec;
302 	slp->inode = curino;
303 	nextblock = slp->tblock;
304 	trecno = 0;
305 	asize += tenths;
306 	blockswritten += ntrec;
307 	blocksthisvol += ntrec;
308 	if (!pipeout && !unlimited && (blocksperfile ?
309 	    (blocksthisvol >= blocksperfile) : (asize > tsize))) {
310 		close_rewind();
311 		startnewtape(0);
312 	}
313 	timeest();
314 }
315 
316 void
317 trewind()
318 {
319 	int f;
320 	int got;
321 
322 	for (f = 0; f < SLAVES; f++) {
323 		/*
324 		 * Drain the results, but unlike EOT we DO (or should) care
325 		 * what the return values were, since if we detect EOT after
326 		 * we think we've written the last blocks to the tape anyway,
327 		 * we have to replay those blocks with rollforward.
328 		 *
329 		 * fixme: punt for now.
330 		 */
331 		if (slaves[f].sent) {
332 			if (atomic(read, slaves[f].fd, (char *)&got, sizeof got)
333 			    != sizeof got) {
334 				perror("  DUMP: error reading command pipe in master");
335 				dumpabort(0);
336 			}
337 			slaves[f].sent = 0;
338 			if (got != writesize) {
339 				msg("EOT detected in last 2 tape records!\n");
340 				msg("Use a longer tape, decrease the size estimate\n");
341 				quit("or use no size estimate at all.\n");
342 			}
343 		}
344 		(void) close(slaves[f].fd);
345 	}
346 	while (wait((int *)NULL) >= 0)	/* wait for any signals from slaves */
347 		/* void */;
348 
349 	if (pipeout)
350 		return;
351 
352 	msg("Closing %s\n", tape);
353 
354 #ifdef RDUMP
355 	if (host) {
356 		rmtclose();
357 		while (rmtopen(tape, 0) < 0)
358 			sleep(10);
359 		rmtclose();
360 		return;
361 	}
362 #endif
363 	(void) close(tapefd);
364 	while ((f = open(tape, 0)) < 0)
365 		sleep (10);
366 	(void) close(f);
367 }
368 
369 void
370 close_rewind()
371 {
372 	time_t tstart_changevol, tend_changevol;
373 
374 	trewind();
375 	if (nexttape)
376 		return;
377 	(void)time((time_t *)&(tstart_changevol));
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 	(void)time((time_t *)&(tend_changevol));
388 	if ((tstart_changevol != (time_t)-1) && (tend_changevol != (time_t)-1))
389 		tstart_writing += (tend_changevol - tstart_changevol);
390 }
391 
392 void
393 rollforward()
394 {
395 	register struct req *p, *q, *prev;
396 	register struct slave *tslp;
397 	int i, size, savedtapea, got;
398 	union u_spcl *ntb, *otb;
399 	tslp = &slaves[SLAVES];
400 	ntb = (union u_spcl *)tslp->tblock[1];
401 
402 	/*
403 	 * Each of the N slaves should have requests that need to
404 	 * be replayed on the next tape.  Use the extra slave buffers
405 	 * (slaves[SLAVES]) to construct request lists to be sent to
406 	 * each slave in turn.
407 	 */
408 	for (i = 0; i < SLAVES; i++) {
409 		q = &tslp->req[1];
410 		otb = (union u_spcl *)slp->tblock;
411 
412 		/*
413 		 * For each request in the current slave, copy it to tslp.
414 		 */
415 
416 		prev = NULL;
417 		for (p = slp->req; p->count > 0; p += p->count) {
418 			*q = *p;
419 			if (p->dblk == 0)
420 				*ntb++ = *otb++; /* copy the datablock also */
421 			prev = q;
422 			q += q->count;
423 		}
424 		if (prev == NULL)
425 			quit("rollforward: protocol botch");
426 		if (prev->dblk != 0)
427 			prev->count -= 1;
428 		else
429 			ntb--;
430 		q -= 1;
431 		q->count = 0;
432 		q = &tslp->req[0];
433 		if (i == 0) {
434 			q->dblk = 0;
435 			q->count = 1;
436 			trecno = 0;
437 			nextblock = tslp->tblock;
438 			savedtapea = spcl.c_tapea;
439 			spcl.c_tapea = slp->tapea;
440 			startnewtape(0);
441 			spcl.c_tapea = savedtapea;
442 			lastspclrec = savedtapea - 1;
443 		}
444 		size = (char *)ntb - (char *)q;
445 		if (atomic(write, slp->fd, (char *)q, size) != size) {
446 			perror("  DUMP: error writing command pipe");
447 			dumpabort(0);
448 		}
449 		slp->sent = 1;
450 		if (++slp >= &slaves[SLAVES])
451 			slp = &slaves[0];
452 
453 		q->count = 1;
454 
455 		if (prev->dblk != 0) {
456 			/*
457 			 * If the last one was a disk block, make the
458 			 * first of this one be the last bit of that disk
459 			 * block...
460 			 */
461 			q->dblk = prev->dblk +
462 				prev->count * (TP_BSIZE / DEV_BSIZE);
463 			ntb = (union u_spcl *)tslp->tblock;
464 		} else {
465 			/*
466 			 * It wasn't a disk block.  Copy the data to its
467 			 * new location in the buffer.
468 			 */
469 			q->dblk = 0;
470 			*((union u_spcl *)tslp->tblock) = *ntb;
471 			ntb = (union u_spcl *)tslp->tblock[1];
472 		}
473 	}
474 	slp->req[0] = *q;
475 	nextblock = slp->tblock;
476 	if (q->dblk == 0)
477 		nextblock++;
478 	trecno = 1;
479 
480 	/*
481 	 * Clear the first slaves' response.  One hopes that it
482 	 * worked ok, otherwise the tape is much too short!
483 	 */
484 	if (slp->sent) {
485 		if (atomic(read, slp->fd, (char *)&got, sizeof got)
486 		    != sizeof got) {
487 			perror("  DUMP: error reading command pipe in master");
488 			dumpabort(0);
489 		}
490 		slp->sent = 0;
491 
492 		if (got != writesize) {
493 			quit("EOT detected at start of the tape!\n");
494 		}
495 	}
496 }
497 
498 /*
499  * We implement taking and restoring checkpoints on the tape level.
500  * When each tape is opened, a new process is created by forking; this
501  * saves all of the necessary context in the parent.  The child
502  * continues the dump; the parent waits around, saving the context.
503  * If the child returns X_REWRITE, then it had problems writing that tape;
504  * this causes the parent to fork again, duplicating the context, and
505  * everything continues as if nothing had happened.
506  */
507 void
508 startnewtape(top)
509 	int top;
510 {
511 	int	parentpid;
512 	int	childpid;
513 	int	status;
514 	int	waitpid;
515 	char	*p;
516 #ifdef sunos
517 	void	(*interrupt_save)();
518 #else
519 	sig_t	interrupt_save;
520 #endif
521 
522 	interrupt_save = signal(SIGINT, SIG_IGN);
523 	parentpid = getpid();
524 
525 restore_check_point:
526 	(void)signal(SIGINT, interrupt_save);
527 	/*
528 	 *	All signals are inherited...
529 	 */
530 	childpid = fork();
531 	if (childpid < 0) {
532 		msg("Context save fork fails in parent %d\n", parentpid);
533 		Exit(X_ABORT);
534 	}
535 	if (childpid != 0) {
536 		/*
537 		 *	PARENT:
538 		 *	save the context by waiting
539 		 *	until the child doing all of the work returns.
540 		 *	don't catch the interrupt
541 		 */
542 		signal(SIGINT, SIG_IGN);
543 #ifdef TDEBUG
544 		msg("Tape: %d; parent process: %d child process %d\n",
545 			tapeno+1, parentpid, childpid);
546 #endif /* TDEBUG */
547 		while ((waitpid = wait(&status)) != childpid)
548 			msg("Parent %d waiting for child %d has another child %d return\n",
549 				parentpid, childpid, waitpid);
550 		if (status & 0xFF) {
551 			msg("Child %d returns LOB status %o\n",
552 				childpid, status&0xFF);
553 		}
554 		status = (status >> 8) & 0xFF;
555 #ifdef TDEBUG
556 		switch(status) {
557 			case X_FINOK:
558 				msg("Child %d finishes X_FINOK\n", childpid);
559 				break;
560 			case X_ABORT:
561 				msg("Child %d finishes X_ABORT\n", childpid);
562 				break;
563 			case X_REWRITE:
564 				msg("Child %d finishes X_REWRITE\n", childpid);
565 				break;
566 			default:
567 				msg("Child %d finishes unknown %d\n",
568 					childpid, status);
569 				break;
570 		}
571 #endif /* TDEBUG */
572 		switch(status) {
573 			case X_FINOK:
574 				Exit(X_FINOK);
575 			case X_ABORT:
576 				Exit(X_ABORT);
577 			case X_REWRITE:
578 				goto restore_check_point;
579 			default:
580 				msg("Bad return code from dump: %d\n", status);
581 				Exit(X_ABORT);
582 		}
583 		/*NOTREACHED*/
584 	} else {	/* we are the child; just continue */
585 #ifdef TDEBUG
586 		sleep(4);	/* allow time for parent's message to get out */
587 		msg("Child on Tape %d has parent %d, my pid = %d\n",
588 			tapeno+1, parentpid, getpid());
589 #endif /* TDEBUG */
590 		/*
591 		 * If we have a name like "/dev/rmt0,/dev/rmt1",
592 		 * use the name before the comma first, and save
593 		 * the remaining names for subsequent volumes.
594 		 */
595 		tapeno++;               /* current tape sequence */
596 		if (nexttape || strchr(tape, ',')) {
597 			if (nexttape && *nexttape)
598 				tape = nexttape;
599 			if ((p = strchr(tape, ',')) != NULL) {
600 				*p = '\0';
601 				nexttape = p + 1;
602 			} else
603 				nexttape = NULL;
604 			msg("Dumping volume %d on %s\n", tapeno, tape);
605 		}
606 #ifdef RDUMP
607 		while ((tapefd = (host ? rmtopen(tape, 2) :
608 			pipeout ? 1 : open(tape, O_WRONLY|O_CREAT, 0666))) < 0)
609 #else
610 		while ((tapefd = (pipeout ? 1 :
611 				  open(tape, O_WRONLY|O_CREAT, 0666))) < 0)
612 #endif
613 		    {
614 			msg("Cannot open output \"%s\".\n", tape);
615 			if (!query("Do you want to retry the open?"))
616 				dumpabort(0);
617 		}
618 
619 		enslave();  /* Share open tape file descriptor with slaves */
620 
621 		asize = 0;
622 		blocksthisvol = 0;
623 		if (top)
624 			newtape++;		/* new tape signal */
625 		spcl.c_count = slp->count;
626 		/*
627 		 * measure firstrec in TP_BSIZE units since restore doesn't
628 		 * know the correct ntrec value...
629 		 */
630 		spcl.c_firstrec = slp->firstrec;
631 		spcl.c_volume++;
632 		spcl.c_type = TS_TAPE;
633 		spcl.c_flags |= DR_NEWHEADER;
634 		writeheader((ino_t)slp->inode);
635 		spcl.c_flags &=~ DR_NEWHEADER;
636 		if (tapeno > 1)
637 			msg("Volume %d begins with blocks from inode %d\n",
638 				tapeno, slp->inode);
639 	}
640 }
641 
642 void
643 dumpabort(signo)
644 	int signo;
645 {
646 
647 	if (master != 0 && master != getpid())
648 		/* Signals master to call dumpabort */
649 		(void) kill(master, SIGTERM);
650 	else {
651 		killall();
652 		msg("The ENTIRE dump is aborted.\n");
653 	}
654 #ifdef RDUMP
655 	rmtclose();
656 #endif
657 	Exit(X_ABORT);
658 }
659 
660 void
661 Exit(status)
662 	int status;
663 {
664 
665 #ifdef TDEBUG
666 	msg("pid = %d exits with status %d\n", getpid(), status);
667 #endif /* TDEBUG */
668 	exit(status);
669 }
670 
671 /*
672  * proceed - handler for SIGUSR2, used to synchronize IO between the slaves.
673  */
674 void
675 proceed(signo)
676 	int signo;
677 {
678 
679 	if (ready)
680 		longjmp(jmpbuf, 1);
681 	caught++;
682 }
683 
684 void
685 enslave()
686 {
687 	int cmd[2];
688 	register int i, j;
689 
690 	master = getpid();
691 
692 	signal(SIGTERM, dumpabort);  /* Slave sends SIGTERM on dumpabort() */
693 	signal(SIGPIPE, sigpipe);
694 	signal(SIGUSR1, tperror);    /* Slave sends SIGUSR1 on tape errors */
695 	signal(SIGUSR2, proceed);    /* Slave sends SIGUSR2 to next slave */
696 
697 	for (i = 0; i < SLAVES; i++) {
698 		if (i == slp - &slaves[0]) {
699 			caught = 1;
700 		} else {
701 			caught = 0;
702 		}
703 
704 		if (socketpair(AF_UNIX, SOCK_STREAM, 0, cmd) < 0 ||
705 		    (slaves[i].pid = fork()) < 0)
706 			quit("too many slaves, %d (recompile smaller): %s\n",
707 			    i, strerror(errno));
708 
709 		slaves[i].fd = cmd[1];
710 		slaves[i].sent = 0;
711 		if (slaves[i].pid == 0) { 	    /* Slave starts up here */
712 			for (j = 0; j <= i; j++)
713 			        (void) close(slaves[j].fd);
714 			signal(SIGINT, SIG_IGN);    /* Master handles this */
715 			doslave(cmd[0], i);
716 			Exit(X_FINOK);
717 		}
718 	}
719 
720 	for (i = 0; i < SLAVES; i++)
721 		(void) atomic(write, slaves[i].fd,
722 			      (char *) &slaves[(i + 1) % SLAVES].pid,
723 		              sizeof slaves[0].pid);
724 
725 	master = 0;
726 }
727 
728 void
729 killall()
730 {
731 	register int i;
732 
733 	for (i = 0; i < SLAVES; i++)
734 		if (slaves[i].pid > 0) {
735 			(void) kill(slaves[i].pid, SIGKILL);
736 			slaves[i].sent = 0;
737 		}
738 }
739 
740 /*
741  * Synchronization - each process has a lockfile, and shares file
742  * descriptors to the following process's lockfile.  When our write
743  * completes, we release our lock on the following process's lock-
744  * file, allowing the following process to lock it and proceed. We
745  * get the lock back for the next cycle by swapping descriptors.
746  */
747 static void
748 doslave(cmd, slave_number)
749 	register int cmd;
750         int slave_number;
751 {
752 	register int nread;
753 	int nextslave, size, wrote, eot_count;
754 
755 	/*
756 	 * Need our own seek pointer.
757 	 */
758 	(void) close(diskfd);
759 	if ((diskfd = open(disk, O_RDONLY)) < 0)
760 		quit("slave couldn't reopen disk: %s\n", strerror(errno));
761 
762 	/*
763 	 * Need the pid of the next slave in the loop...
764 	 */
765 	if ((nread = atomic(read, cmd, (char *)&nextslave, sizeof nextslave))
766 	    != sizeof nextslave) {
767 		quit("master/slave protocol botched - didn't get pid of next slave.\n");
768 	}
769 
770 	/*
771 	 * Get list of blocks to dump, read the blocks into tape buffer
772 	 */
773 	while ((nread = atomic(read, cmd, (char *)slp->req, reqsiz)) == reqsiz) {
774 		register struct req *p = slp->req;
775 
776 		for (trecno = 0; trecno < ntrec;
777 		     trecno += p->count, p += p->count) {
778 			if (p->dblk) {
779 				bread(p->dblk, slp->tblock[trecno],
780 					p->count * TP_BSIZE);
781 			} else {
782 				if (p->count != 1 || atomic(read, cmd,
783 				    (char *)slp->tblock[trecno],
784 				    TP_BSIZE) != TP_BSIZE)
785 				       quit("master/slave protocol botched.\n");
786 			}
787 		}
788 		if (setjmp(jmpbuf) == 0) {
789 			ready = 1;
790 			if (!caught)
791 				(void) pause();
792 		}
793 		ready = 0;
794 		caught = 0;
795 
796 		/* Try to write the data... */
797 		eot_count = 0;
798 		size = 0;
799 
800 		while (eot_count < 10 && size < writesize) {
801 #ifdef RDUMP
802 			if (host)
803 				wrote = rmtwrite(slp->tblock[0]+size,
804 				    writesize-size);
805 			else
806 #endif
807 				wrote = write(tapefd, slp->tblock[0]+size,
808 				    writesize-size);
809 #ifdef WRITEDEBUG
810 			printf("slave %d wrote %d\n", slave_number, wrote);
811 #endif
812 			if (wrote < 0)
813 				break;
814 			if (wrote == 0)
815 				eot_count++;
816 			size += wrote;
817 		}
818 
819 #ifdef WRITEDEBUG
820 		if (size != writesize)
821 		 printf("slave %d only wrote %d out of %d bytes and gave up.\n",
822 		     slave_number, size, writesize);
823 #endif
824 
825 		/*
826 		 * Handle ENOSPC as an EOT condition.
827 		 */
828 		if (wrote < 0 && errno == ENOSPC) {
829 			wrote = 0;
830 			eot_count++;
831 		}
832 
833 		if (eot_count > 0)
834 			size = 0;
835 
836 		if (wrote < 0) {
837 			(void) kill(master, SIGUSR1);
838 			for (;;)
839 				(void) sigpause(0);
840 		} else {
841 			/*
842 			 * pass size of write back to master
843 			 * (for EOT handling)
844 			 */
845 			(void) atomic(write, cmd, (char *)&size, sizeof size);
846 		}
847 
848 		/*
849 		 * If partial write, don't want next slave to go.
850 		 * Also jolts him awake.
851 		 */
852 		(void) kill(nextslave, SIGUSR2);
853 	}
854 	if (nread != 0)
855 		quit("error reading command pipe: %s\n", strerror(errno));
856 }
857 
858 /*
859  * Since a read from a pipe may not return all we asked for,
860  * or a write may not write all we ask if we get a signal,
861  * loop until the count is satisfied (or error).
862  */
863 static int
864 atomic(func, fd, buf, count)
865 	ssize_t (*func)();
866 	int fd;
867 	char *buf;
868 	int count;
869 {
870 	int got, need = count;
871 
872 	while ((got = (*func)(fd, buf, need)) > 0 && (need -= got) > 0)
873 		buf += got;
874 	return (got < 0 ? got : count - need);
875 }
876