xref: /netbsd-src/sbin/dump/tape.c (revision ae9172d6cd9432a6a1a56760d86b32c57a66c39c)
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 /*static char sccsid[] = "from: @(#)tape.c	8.2 (Berkeley) 3/17/94";*/
36 static char *rcsid = "$Id: tape.c,v 1.5 1994/10/31 04:21:35 cgd Exp $";
37 #endif /* not lint */
38 
39 #include <sys/param.h>
40 #include <sys/socket.h>
41 #include <sys/time.h>
42 #include <sys/wait.h>
43 #ifdef sunos
44 #include <sys/vnode.h>
45 
46 #include <ufs/fs.h>
47 #include <ufs/inode.h>
48 #else
49 #include <ufs/ffs/fs.h>
50 #include <ufs/ufs/dinode.h>
51 #endif
52 
53 #include <protocols/dumprestore.h>
54 
55 #include <errno.h>
56 #include <fcntl.h>
57 #include <setjmp.h>
58 #include <signal.h>
59 #include <stdio.h>
60 #ifdef __STDC__
61 #include <stdlib.h>
62 #include <string.h>
63 #include <unistd.h>
64 #else
65 int	write(), read();
66 #endif
67 
68 #include "dump.h"
69 #include "pathnames.h"
70 
71 int	writesize;		/* size of malloc()ed buffer for tape */
72 long	lastspclrec = -1;	/* tape block number of last written header */
73 int	trecno = 0;		/* next record to write in current block */
74 extern	long blocksperfile;	/* number of blocks per output file */
75 long	blocksthisvol;		/* number of blocks on current output file */
76 extern	int ntrec;		/* blocking factor on tape */
77 extern	int cartridge;
78 extern	char *host;
79 char	*nexttape;
80 
81 static	ssize_t atomic __P((ssize_t (*)(), int, char *, int));
82 static	void doslave __P((int, int));
83 static	void enslave __P((void));
84 static	void flushtape __P((void));
85 static	void killall __P((void));
86 static	void rollforward __P((void));
87 
88 /*
89  * Concurrent dump mods (Caltech) - disk block reading and tape writing
90  * are exported to several slave processes.  While one slave writes the
91  * tape, the others read disk blocks; they pass control of the tape in
92  * a ring via signals. The parent process traverses the filesystem and
93  * sends writeheader()'s and lists of daddr's to the slaves via pipes.
94  * The following structure defines the instruction packets sent to slaves.
95  */
96 struct req {
97 	daddr_t dblk;
98 	int count;
99 };
100 int reqsiz;
101 
102 #define SLAVES 3		/* 1 slave writing, 1 reading, 1 for slack */
103 struct slave {
104 	int tapea;		/* header number at start of this chunk */
105 	int count;		/* count to next header (used for TS_TAPE */
106 				/* after EOT) */
107 	int inode;		/* inode that we are currently dealing with */
108 	int fd;			/* FD for this slave */
109 	int pid;		/* PID for this slave */
110 	int sent;		/* 1 == we've sent this slave requests */
111 	int firstrec;		/* record number of this block */
112 	char (*tblock)[TP_BSIZE]; /* buffer for data blocks */
113 	struct req *req;	/* buffer for requests */
114 } slaves[SLAVES+1];
115 struct slave *slp;
116 
117 char	(*nextblock)[TP_BSIZE];
118 
119 int master;		/* pid of master, for sending error signals */
120 int tenths;		/* length of tape used per block written */
121 static int caught;	/* have we caught the signal to proceed? */
122 static int ready;	/* have we reached the lock point without having */
123 			/* received the SIGUSR2 signal from the prev slave? */
124 static jmp_buf jmpbuf;	/* where to jump to if we are ready when the */
125 			/* SIGUSR2 arrives from the previous slave */
126 
127 int
128 alloctape()
129 {
130 	int pgoff = getpagesize() - 1;
131 	char *buf;
132 	int i;
133 
134 	writesize = ntrec * TP_BSIZE;
135 	reqsiz = (ntrec + 1) * sizeof(struct req);
136 	/*
137 	 * CDC 92181's and 92185's make 0.8" gaps in 1600-bpi start/stop mode
138 	 * (see DEC TU80 User's Guide).  The shorter gaps of 6250-bpi require
139 	 * repositioning after stopping, i.e, streaming mode, where the gap is
140 	 * variable, 0.30" to 0.45".  The gap is maximal when the tape stops.
141 	 */
142 	if (blocksperfile == 0)
143 		tenths = writesize / density +
144 		    (cartridge ? 16 : density == 625 ? 5 : 8);
145 	/*
146 	 * Allocate tape buffer contiguous with the array of instruction
147 	 * packets, so flushtape() can write them together with one write().
148 	 * Align tape buffer on page boundary to speed up tape write().
149 	 */
150 	for (i = 0; i <= SLAVES; i++) {
151 		buf = (char *)
152 		    malloc((unsigned)(reqsiz + writesize + pgoff + TP_BSIZE));
153 		if (buf == NULL)
154 			return(0);
155 		slaves[i].tblock = (char (*)[TP_BSIZE])
156 		    (((long)&buf[ntrec + 1] + pgoff) &~ pgoff);
157 		slaves[i].req = (struct req *)slaves[i].tblock - ntrec - 1;
158 	}
159 	slp = &slaves[0];
160 	slp->count = 1;
161 	slp->tapea = 0;
162 	slp->firstrec = 0;
163 	nextblock = slp->tblock;
164 	return(1);
165 }
166 
167 void
168 writerec(dp, isspcl)
169 	char *dp;
170 	int isspcl;
171 {
172 
173 	slp->req[trecno].dblk = (daddr_t)0;
174 	slp->req[trecno].count = 1;
175 	*(union u_spcl *)(*(nextblock)++) = *(union u_spcl *)dp;
176 	if (isspcl)
177 		lastspclrec = 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 && (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 	trewind();
371 	if (nexttape)
372 		return;
373 	if (!nogripe) {
374 		msg("Change Volumes: Mount volume #%d\n", tapeno+1);
375 		broadcast("CHANGE DUMP VOLUMES!\7\7\n");
376 	}
377 	while (!query("Is the new volume mounted and ready to go?"))
378 		if (query("Do you want to abort?")) {
379 			dumpabort(0);
380 			/*NOTREACHED*/
381 		}
382 }
383 
384 void
385 rollforward()
386 {
387 	register struct req *p, *q, *prev;
388 	register struct slave *tslp;
389 	int i, size, savedtapea, got;
390 	union u_spcl *ntb, *otb;
391 	tslp = &slaves[SLAVES];
392 	ntb = (union u_spcl *)tslp->tblock[1];
393 
394 	/*
395 	 * Each of the N slaves should have requests that need to
396 	 * be replayed on the next tape.  Use the extra slave buffers
397 	 * (slaves[SLAVES]) to construct request lists to be sent to
398 	 * each slave in turn.
399 	 */
400 	for (i = 0; i < SLAVES; i++) {
401 		q = &tslp->req[1];
402 		otb = (union u_spcl *)slp->tblock;
403 
404 		/*
405 		 * For each request in the current slave, copy it to tslp.
406 		 */
407 
408 		prev = NULL;
409 		for (p = slp->req; p->count > 0; p += p->count) {
410 			*q = *p;
411 			if (p->dblk == 0)
412 				*ntb++ = *otb++; /* copy the datablock also */
413 			prev = q;
414 			q += q->count;
415 		}
416 		if (prev == NULL)
417 			quit("rollforward: protocol botch");
418 		if (prev->dblk != 0)
419 			prev->count -= 1;
420 		else
421 			ntb--;
422 		q -= 1;
423 		q->count = 0;
424 		q = &tslp->req[0];
425 		if (i == 0) {
426 			q->dblk = 0;
427 			q->count = 1;
428 			trecno = 0;
429 			nextblock = tslp->tblock;
430 			savedtapea = spcl.c_tapea;
431 			spcl.c_tapea = slp->tapea;
432 			startnewtape(0);
433 			spcl.c_tapea = savedtapea;
434 			lastspclrec = savedtapea - 1;
435 		}
436 		size = (char *)ntb - (char *)q;
437 		if (atomic(write, slp->fd, (char *)q, size) != size) {
438 			perror("  DUMP: error writing command pipe");
439 			dumpabort(0);
440 		}
441 		slp->sent = 1;
442 		if (++slp >= &slaves[SLAVES])
443 			slp = &slaves[0];
444 
445 		q->count = 1;
446 
447 		if (prev->dblk != 0) {
448 			/*
449 			 * If the last one was a disk block, make the
450 			 * first of this one be the last bit of that disk
451 			 * block...
452 			 */
453 			q->dblk = prev->dblk +
454 				prev->count * (TP_BSIZE / DEV_BSIZE);
455 			ntb = (union u_spcl *)tslp->tblock;
456 		} else {
457 			/*
458 			 * It wasn't a disk block.  Copy the data to its
459 			 * new location in the buffer.
460 			 */
461 			q->dblk = 0;
462 			*((union u_spcl *)tslp->tblock) = *ntb;
463 			ntb = (union u_spcl *)tslp->tblock[1];
464 		}
465 	}
466 	slp->req[0] = *q;
467 	nextblock = slp->tblock;
468 	if (q->dblk == 0)
469 		nextblock++;
470 	trecno = 1;
471 
472 	/*
473 	 * Clear the first slaves' response.  One hopes that it
474 	 * worked ok, otherwise the tape is much too short!
475 	 */
476 	if (slp->sent) {
477 		if (atomic(read, slp->fd, (char *)&got, sizeof got)
478 		    != sizeof got) {
479 			perror("  DUMP: error reading command pipe in master");
480 			dumpabort(0);
481 		}
482 		slp->sent = 0;
483 
484 		if (got != writesize) {
485 			quit("EOT detected at start of the tape!\n");
486 		}
487 	}
488 }
489 
490 /*
491  * We implement taking and restoring checkpoints on the tape level.
492  * When each tape is opened, a new process is created by forking; this
493  * saves all of the necessary context in the parent.  The child
494  * continues the dump; the parent waits around, saving the context.
495  * If the child returns X_REWRITE, then it had problems writing that tape;
496  * this causes the parent to fork again, duplicating the context, and
497  * everything continues as if nothing had happened.
498  */
499 void
500 startnewtape(top)
501 	int top;
502 {
503 	int	parentpid;
504 	int	childpid;
505 	int	status;
506 	int	waitpid;
507 	char	*p;
508 #ifdef sunos
509 	void	(*interrupt_save)();
510 #else
511 	sig_t	interrupt_save;
512 #endif
513 
514 	interrupt_save = signal(SIGINT, SIG_IGN);
515 	parentpid = getpid();
516 
517 restore_check_point:
518 	(void)signal(SIGINT, interrupt_save);
519 	/*
520 	 *	All signals are inherited...
521 	 */
522 	childpid = fork();
523 	if (childpid < 0) {
524 		msg("Context save fork fails in parent %d\n", parentpid);
525 		Exit(X_ABORT);
526 	}
527 	if (childpid != 0) {
528 		/*
529 		 *	PARENT:
530 		 *	save the context by waiting
531 		 *	until the child doing all of the work returns.
532 		 *	don't catch the interrupt
533 		 */
534 		signal(SIGINT, SIG_IGN);
535 #ifdef TDEBUG
536 		msg("Tape: %d; parent process: %d child process %d\n",
537 			tapeno+1, parentpid, childpid);
538 #endif /* TDEBUG */
539 		while ((waitpid = wait(&status)) != childpid)
540 			msg("Parent %d waiting for child %d has another child %d return\n",
541 				parentpid, childpid, waitpid);
542 		if (status & 0xFF) {
543 			msg("Child %d returns LOB status %o\n",
544 				childpid, status&0xFF);
545 		}
546 		status = (status >> 8) & 0xFF;
547 #ifdef TDEBUG
548 		switch(status) {
549 			case X_FINOK:
550 				msg("Child %d finishes X_FINOK\n", childpid);
551 				break;
552 			case X_ABORT:
553 				msg("Child %d finishes X_ABORT\n", childpid);
554 				break;
555 			case X_REWRITE:
556 				msg("Child %d finishes X_REWRITE\n", childpid);
557 				break;
558 			default:
559 				msg("Child %d finishes unknown %d\n",
560 					childpid, status);
561 				break;
562 		}
563 #endif /* TDEBUG */
564 		switch(status) {
565 			case X_FINOK:
566 				Exit(X_FINOK);
567 			case X_ABORT:
568 				Exit(X_ABORT);
569 			case X_REWRITE:
570 				goto restore_check_point;
571 			default:
572 				msg("Bad return code from dump: %d\n", status);
573 				Exit(X_ABORT);
574 		}
575 		/*NOTREACHED*/
576 	} else {	/* we are the child; just continue */
577 #ifdef TDEBUG
578 		sleep(4);	/* allow time for parent's message to get out */
579 		msg("Child on Tape %d has parent %d, my pid = %d\n",
580 			tapeno+1, parentpid, getpid());
581 #endif /* TDEBUG */
582 		/*
583 		 * If we have a name like "/dev/rmt0,/dev/rmt1",
584 		 * use the name before the comma first, and save
585 		 * the remaining names for subsequent volumes.
586 		 */
587 		tapeno++;               /* current tape sequence */
588 		if (nexttape || strchr(tape, ',')) {
589 			if (nexttape && *nexttape)
590 				tape = nexttape;
591 			if ((p = strchr(tape, ',')) != NULL) {
592 				*p = '\0';
593 				nexttape = p + 1;
594 			} else
595 				nexttape = NULL;
596 			msg("Dumping volume %d on %s\n", tapeno, tape);
597 		}
598 #ifdef RDUMP
599 		while ((tapefd = (host ? rmtopen(tape, 2) :
600 			pipeout ? 1 : open(tape, O_WRONLY|O_CREAT, 0666))) < 0)
601 #else
602 		while ((tapefd = (pipeout ? 1 :
603 				  open(tape, O_WRONLY|O_CREAT, 0666))) < 0)
604 #endif
605 		    {
606 			msg("Cannot open output \"%s\".\n", tape);
607 			if (!query("Do you want to retry the open?"))
608 				dumpabort(0);
609 		}
610 
611 		enslave();  /* Share open tape file descriptor with slaves */
612 
613 		asize = 0;
614 		blocksthisvol = 0;
615 		if (top)
616 			newtape++;		/* new tape signal */
617 		spcl.c_count = slp->count;
618 		/*
619 		 * measure firstrec in TP_BSIZE units since restore doesn't
620 		 * know the correct ntrec value...
621 		 */
622 		spcl.c_firstrec = slp->firstrec;
623 		spcl.c_volume++;
624 		spcl.c_type = TS_TAPE;
625 		spcl.c_flags |= DR_NEWHEADER;
626 		writeheader((ino_t)slp->inode);
627 		spcl.c_flags &=~ DR_NEWHEADER;
628 		if (tapeno > 1)
629 			msg("Volume %d begins with blocks from inode %d\n",
630 				tapeno, slp->inode);
631 	}
632 }
633 
634 void
635 dumpabort(signo)
636 	int signo;
637 {
638 
639 	if (master != 0 && master != getpid())
640 		/* Signals master to call dumpabort */
641 		(void) kill(master, SIGTERM);
642 	else {
643 		killall();
644 		msg("The ENTIRE dump is aborted.\n");
645 	}
646 #ifdef RDUMP
647 	rmtclose();
648 #endif
649 	Exit(X_ABORT);
650 }
651 
652 __dead void
653 Exit(status)
654 	int status;
655 {
656 
657 #ifdef TDEBUG
658 	msg("pid = %d exits with status %d\n", getpid(), status);
659 #endif /* TDEBUG */
660 	exit(status);
661 }
662 
663 /*
664  * proceed - handler for SIGUSR2, used to synchronize IO between the slaves.
665  */
666 void
667 proceed(signo)
668 	int signo;
669 {
670 
671 	if (ready)
672 		longjmp(jmpbuf, 1);
673 	caught++;
674 }
675 
676 void
677 enslave()
678 {
679 	int cmd[2];
680 	register int i, j;
681 
682 	master = getpid();
683 
684 	signal(SIGTERM, dumpabort);  /* Slave sends SIGTERM on dumpabort() */
685 	signal(SIGPIPE, sigpipe);
686 	signal(SIGUSR1, tperror);    /* Slave sends SIGUSR1 on tape errors */
687 	signal(SIGUSR2, proceed);    /* Slave sends SIGUSR2 to next slave */
688 
689 	for (i = 0; i < SLAVES; i++) {
690 		if (i == slp - &slaves[0]) {
691 			caught = 1;
692 		} else {
693 			caught = 0;
694 		}
695 
696 		if (socketpair(AF_UNIX, SOCK_STREAM, 0, cmd) < 0 ||
697 		    (slaves[i].pid = fork()) < 0)
698 			quit("too many slaves, %d (recompile smaller): %s\n",
699 			    i, strerror(errno));
700 
701 		slaves[i].fd = cmd[1];
702 		slaves[i].sent = 0;
703 		if (slaves[i].pid == 0) { 	    /* Slave starts up here */
704 			for (j = 0; j <= i; j++)
705 			        (void) close(slaves[j].fd);
706 			signal(SIGINT, SIG_IGN);    /* Master handles this */
707 			doslave(cmd[0], i);
708 			Exit(X_FINOK);
709 		}
710 	}
711 
712 	for (i = 0; i < SLAVES; i++)
713 		(void) atomic(write, slaves[i].fd,
714 			      (char *) &slaves[(i + 1) % SLAVES].pid,
715 		              sizeof slaves[0].pid);
716 
717 	master = 0;
718 }
719 
720 void
721 killall()
722 {
723 	register int i;
724 
725 	for (i = 0; i < SLAVES; i++)
726 		if (slaves[i].pid > 0)
727 			(void) kill(slaves[i].pid, SIGKILL);
728 }
729 
730 /*
731  * Synchronization - each process has a lockfile, and shares file
732  * descriptors to the following process's lockfile.  When our write
733  * completes, we release our lock on the following process's lock-
734  * file, allowing the following process to lock it and proceed. We
735  * get the lock back for the next cycle by swapping descriptors.
736  */
737 static void
738 doslave(cmd, slave_number)
739 	register int cmd;
740         int slave_number;
741 {
742 	register int nread;
743 	int nextslave, size, wrote, eot_count;
744 
745 	/*
746 	 * Need our own seek pointer.
747 	 */
748 	(void) close(diskfd);
749 	if ((diskfd = open(disk, O_RDONLY)) < 0)
750 		quit("slave couldn't reopen disk: %s\n", strerror(errno));
751 
752 	/*
753 	 * Need the pid of the next slave in the loop...
754 	 */
755 	if ((nread = atomic(read, cmd, (char *)&nextslave, sizeof nextslave))
756 	    != sizeof nextslave) {
757 		quit("master/slave protocol botched - didn't get pid of next slave.\n");
758 	}
759 
760 	/*
761 	 * Get list of blocks to dump, read the blocks into tape buffer
762 	 */
763 	while ((nread = atomic(read, cmd, (char *)slp->req, reqsiz)) == reqsiz) {
764 		register struct req *p = slp->req;
765 
766 		for (trecno = 0; trecno < ntrec;
767 		     trecno += p->count, p += p->count) {
768 			if (p->dblk) {
769 				bread(p->dblk, slp->tblock[trecno],
770 					p->count * TP_BSIZE);
771 			} else {
772 				if (p->count != 1 || atomic(read, cmd,
773 				    (char *)slp->tblock[trecno],
774 				    TP_BSIZE) != TP_BSIZE)
775 				       quit("master/slave protocol botched.\n");
776 			}
777 		}
778 		if (setjmp(jmpbuf) == 0) {
779 			ready = 1;
780 			if (!caught)
781 				(void) pause();
782 		}
783 		ready = 0;
784 		caught = 0;
785 
786 		/* Try to write the data... */
787 		eot_count = 0;
788 		size = 0;
789 
790 		while (eot_count < 10 && size < writesize) {
791 #ifdef RDUMP
792 			if (host)
793 				wrote = rmtwrite(slp->tblock[0]+size,
794 				    writesize-size);
795 			else
796 #endif
797 				wrote = write(tapefd, slp->tblock[0]+size,
798 				    writesize-size);
799 #ifdef WRITEDEBUG
800 			printf("slave %d wrote %d\n", slave_number, wrote);
801 #endif
802 			if (wrote < 0)
803 				break;
804 			if (wrote == 0)
805 				eot_count++;
806 			size += wrote;
807 		}
808 
809 #ifdef WRITEDEBUG
810 		if (size != writesize)
811 		 printf("slave %d only wrote %d out of %d bytes and gave up.\n",
812 		     slave_number, size, writesize);
813 #endif
814 
815 		if (eot_count > 0)
816 			size = 0;
817 
818 		/*
819 		 * fixme: Pyramids running OSx return ENOSPC
820 		 * at EOT on 1/2 inch drives.
821 		 */
822 		if (size < 0) {
823 			(void) kill(master, SIGUSR1);
824 			for (;;)
825 				(void) sigpause(0);
826 		} else {
827 			/*
828 			 * pass size of write back to master
829 			 * (for EOT handling)
830 			 */
831 			(void) atomic(write, cmd, (char *)&size, sizeof size);
832 		}
833 
834 		/*
835 		 * If partial write, don't want next slave to go.
836 		 * Also jolts him awake.
837 		 */
838 		(void) kill(nextslave, SIGUSR2);
839 	}
840 	if (nread != 0)
841 		quit("error reading command pipe: %s\n", strerror(errno));
842 }
843 
844 /*
845  * Since a read from a pipe may not return all we asked for,
846  * or a write may not write all we ask if we get a signal,
847  * loop until the count is satisfied (or error).
848  */
849 static ssize_t
850 atomic(func, fd, buf, count)
851 	ssize_t (*func)();
852 	int fd;
853 	char *buf;
854 	int count;
855 {
856 	ssize_t got, need = count;
857 
858 	while ((got = (*func)(fd, buf, need)) > 0 && (need -= got) > 0)
859 		buf += got;
860 	return (got < 0 ? got : count - need);
861 }
862