xref: /freebsd/sbin/dump/tape.c (revision 17ee9d00bc1ae1e598c38f25826f861e4bc6c3ce)
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[] = "@(#)tape.c	8.2 (Berkeley) 3/17/94";
36 #endif /* not lint */
37 
38 #include <sys/param.h>
39 #include <sys/socket.h>
40 #include <sys/time.h>
41 #include <sys/wait.h>
42 #ifdef sunos
43 #include <sys/vnode.h>
44 
45 #include <ufs/fs.h>
46 #include <ufs/inode.h>
47 #else
48 #include <ufs/ffs/fs.h>
49 #include <ufs/ufs/dinode.h>
50 #endif
51 
52 #include <protocols/dumprestore.h>
53 
54 #include <errno.h>
55 #include <fcntl.h>
56 #include <setjmp.h>
57 #include <signal.h>
58 #include <stdio.h>
59 #ifdef __STDC__
60 #include <stdlib.h>
61 #include <string.h>
62 #include <unistd.h>
63 #else
64 int	write(), read();
65 #endif
66 
67 #include "dump.h"
68 #include "pathnames.h"
69 
70 int	writesize;		/* size of malloc()ed buffer for tape */
71 long	lastspclrec = -1;	/* tape block number of last written header */
72 int	trecno = 0;		/* next record to write in current block */
73 extern	long blocksperfile;	/* number of blocks per output file */
74 long	blocksthisvol;		/* number of blocks on current output file */
75 extern	int ntrec;		/* blocking factor on tape */
76 extern	int cartridge;
77 extern	char *host;
78 char	*nexttape;
79 
80 static	int atomic __P((int (*)(), int, char *, int));
81 static	void doslave __P((int, int));
82 static	void enslave __P((void));
83 static	void flushtape __P((void));
84 static	void killall __P((void));
85 static	void rollforward __P((void));
86 
87 /*
88  * Concurrent dump mods (Caltech) - disk block reading and tape writing
89  * are exported to several slave processes.  While one slave writes the
90  * tape, the others read disk blocks; they pass control of the tape in
91  * a ring via signals. The parent process traverses the filesystem and
92  * sends writeheader()'s and lists of daddr's to the slaves via pipes.
93  * The following structure defines the instruction packets sent to slaves.
94  */
95 struct req {
96 	daddr_t dblk;
97 	int count;
98 };
99 int reqsiz;
100 
101 #define SLAVES 3		/* 1 slave writing, 1 reading, 1 for slack */
102 struct slave {
103 	int tapea;		/* header number at start of this chunk */
104 	int count;		/* count to next header (used for TS_TAPE */
105 				/* after EOT) */
106 	int inode;		/* inode that we are currently dealing with */
107 	int fd;			/* FD for this slave */
108 	int pid;		/* PID for this slave */
109 	int sent;		/* 1 == we've sent this slave requests */
110 	int firstrec;		/* record number of this block */
111 	char (*tblock)[TP_BSIZE]; /* buffer for data blocks */
112 	struct req *req;	/* buffer for requests */
113 } slaves[SLAVES+1];
114 struct slave *slp;
115 
116 char	(*nextblock)[TP_BSIZE];
117 
118 int master;		/* pid of master, for sending error signals */
119 int tenths;		/* length of tape used per block written */
120 static int caught;	/* have we caught the signal to proceed? */
121 static int ready;	/* have we reached the lock point without having */
122 			/* received the SIGUSR2 signal from the prev slave? */
123 static jmp_buf jmpbuf;	/* where to jump to if we are ready when the */
124 			/* SIGUSR2 arrives from the previous slave */
125 
126 int
127 alloctape()
128 {
129 	int pgoff = getpagesize() - 1;
130 	char *buf;
131 	int i;
132 
133 	writesize = ntrec * TP_BSIZE;
134 	reqsiz = (ntrec + 1) * sizeof(struct req);
135 	/*
136 	 * CDC 92181's and 92185's make 0.8" gaps in 1600-bpi start/stop mode
137 	 * (see DEC TU80 User's Guide).  The shorter gaps of 6250-bpi require
138 	 * repositioning after stopping, i.e, streaming mode, where the gap is
139 	 * variable, 0.30" to 0.45".  The gap is maximal when the tape stops.
140 	 */
141 	if (blocksperfile == 0)
142 		tenths = writesize / density +
143 		    (cartridge ? 16 : density == 625 ? 5 : 8);
144 	/*
145 	 * Allocate tape buffer contiguous with the array of instruction
146 	 * packets, so flushtape() can write them together with one write().
147 	 * Align tape buffer on page boundary to speed up tape write().
148 	 */
149 	for (i = 0; i <= SLAVES; i++) {
150 		buf = (char *)
151 		    malloc((unsigned)(reqsiz + writesize + pgoff + TP_BSIZE));
152 		if (buf == NULL)
153 			return(0);
154 		slaves[i].tblock = (char (*)[TP_BSIZE])
155 		    (((long)&buf[ntrec + 1] + pgoff) &~ pgoff);
156 		slaves[i].req = (struct req *)slaves[i].tblock - ntrec - 1;
157 	}
158 	slp = &slaves[0];
159 	slp->count = 1;
160 	slp->tapea = 0;
161 	slp->firstrec = 0;
162 	nextblock = slp->tblock;
163 	return(1);
164 }
165 
166 void
167 writerec(dp, isspcl)
168 	char *dp;
169 	int isspcl;
170 {
171 
172 	slp->req[trecno].dblk = (daddr_t)0;
173 	slp->req[trecno].count = 1;
174 	*(union u_spcl *)(*(nextblock)++) = *(union u_spcl *)dp;
175 	if (isspcl)
176 		lastspclrec = spcl.c_tapea;
177 	trecno++;
178 	spcl.c_tapea++;
179 	if (trecno >= ntrec)
180 		flushtape();
181 }
182 
183 void
184 dumpblock(blkno, size)
185 	daddr_t blkno;
186 	int size;
187 {
188 	int avail, tpblks, dblkno;
189 
190 	dblkno = fsbtodb(sblock, blkno);
191 	tpblks = size >> tp_bshift;
192 	while ((avail = MIN(tpblks, ntrec - trecno)) > 0) {
193 		slp->req[trecno].dblk = dblkno;
194 		slp->req[trecno].count = avail;
195 		trecno += avail;
196 		spcl.c_tapea += avail;
197 		if (trecno >= ntrec)
198 			flushtape();
199 		dblkno += avail << (tp_bshift - dev_bshift);
200 		tpblks -= avail;
201 	}
202 }
203 
204 int	nogripe = 0;
205 
206 void
207 tperror(signo)
208 	int signo;
209 {
210 
211 	if (pipeout) {
212 		msg("write error on %s\n", tape);
213 		quit("Cannot recover\n");
214 		/* NOTREACHED */
215 	}
216 	msg("write error %d blocks into volume %d\n", blocksthisvol, tapeno);
217 	broadcast("DUMP WRITE ERROR!\n");
218 	if (!query("Do you want to restart?"))
219 		dumpabort(0);
220 	msg("Closing this volume.  Prepare to restart with new media;\n");
221 	msg("this dump volume will be rewritten.\n");
222 	killall();
223 	nogripe = 1;
224 	close_rewind();
225 	Exit(X_REWRITE);
226 }
227 
228 void
229 sigpipe(signo)
230 	int signo;
231 {
232 
233 	quit("Broken pipe\n");
234 }
235 
236 static void
237 flushtape()
238 {
239 	int i, blks, got;
240 	long lastfirstrec;
241 
242 	int siz = (char *)nextblock - (char *)slp->req;
243 
244 	slp->req[trecno].count = 0;			/* Sentinel */
245 
246 	if (atomic(write, slp->fd, (char *)slp->req, siz) != siz)
247 		quit("error writing command pipe: %s\n", strerror(errno));
248 	slp->sent = 1; /* we sent a request, read the response later */
249 
250 	lastfirstrec = slp->firstrec;
251 
252 	if (++slp >= &slaves[SLAVES])
253 		slp = &slaves[0];
254 
255 	/* Read results back from next slave */
256 	if (slp->sent) {
257 		if (atomic(read, slp->fd, (char *)&got, sizeof got)
258 		    != sizeof got) {
259 			perror("  DUMP: error reading command pipe in master");
260 			dumpabort(0);
261 		}
262 		slp->sent = 0;
263 
264 		/* Check for end of tape */
265 		if (got < writesize) {
266 			msg("End of tape detected\n");
267 
268 			/*
269 			 * Drain the results, don't care what the values were.
270 			 * If we read them here then trewind won't...
271 			 */
272 			for (i = 0; i < SLAVES; i++) {
273 				if (slaves[i].sent) {
274 					if (atomic(read, slaves[i].fd,
275 					    (char *)&got, sizeof got)
276 					    != sizeof got) {
277 						perror("  DUMP: error reading command pipe in master");
278 						dumpabort(0);
279 					}
280 					slaves[i].sent = 0;
281 				}
282 			}
283 
284 			close_rewind();
285 			rollforward();
286 			return;
287 		}
288 	}
289 
290 	blks = 0;
291 	if (spcl.c_type != TS_END) {
292 		for (i = 0; i < spcl.c_count; i++)
293 			if (spcl.c_addr[i] != 0)
294 				blks++;
295 	}
296 	slp->count = lastspclrec + blks + 1 - spcl.c_tapea;
297 	slp->tapea = spcl.c_tapea;
298 	slp->firstrec = lastfirstrec + ntrec;
299 	slp->inode = curino;
300 	nextblock = slp->tblock;
301 	trecno = 0;
302 	asize += tenths;
303 	blockswritten += ntrec;
304 	blocksthisvol += ntrec;
305 	if (!pipeout && (blocksperfile ?
306 	    (blocksthisvol >= blocksperfile) : (asize > tsize))) {
307 		close_rewind();
308 		startnewtape(0);
309 	}
310 	timeest();
311 }
312 
313 void
314 trewind()
315 {
316 	int f;
317 	int got;
318 
319 	for (f = 0; f < SLAVES; f++) {
320 		/*
321 		 * Drain the results, but unlike EOT we DO (or should) care
322 		 * what the return values were, since if we detect EOT after
323 		 * we think we've written the last blocks to the tape anyway,
324 		 * we have to replay those blocks with rollforward.
325 		 *
326 		 * fixme: punt for now.
327 		 */
328 		if (slaves[f].sent) {
329 			if (atomic(read, slaves[f].fd, (char *)&got, sizeof got)
330 			    != sizeof got) {
331 				perror("  DUMP: error reading command pipe in master");
332 				dumpabort(0);
333 			}
334 			slaves[f].sent = 0;
335 			if (got != writesize) {
336 				msg("EOT detected in last 2 tape records!\n");
337 				msg("Use a longer tape, decrease the size estimate\n");
338 				quit("or use no size estimate at all.\n");
339 			}
340 		}
341 		(void) close(slaves[f].fd);
342 	}
343 	while (wait((int *)NULL) >= 0)	/* wait for any signals from slaves */
344 		/* void */;
345 
346 	if (pipeout)
347 		return;
348 
349 	msg("Closing %s\n", tape);
350 
351 #ifdef RDUMP
352 	if (host) {
353 		rmtclose();
354 		while (rmtopen(tape, 0) < 0)
355 			sleep(10);
356 		rmtclose();
357 		return;
358 	}
359 #endif
360 	(void) close(tapefd);
361 	while ((f = open(tape, 0)) < 0)
362 		sleep (10);
363 	(void) close(f);
364 }
365 
366 void
367 close_rewind()
368 {
369 	trewind();
370 	if (nexttape)
371 		return;
372 	if (!nogripe) {
373 		msg("Change Volumes: Mount volume #%d\n", tapeno+1);
374 		broadcast("CHANGE DUMP VOLUMES!\7\7\n");
375 	}
376 	while (!query("Is the new volume mounted and ready to go?"))
377 		if (query("Do you want to abort?")) {
378 			dumpabort(0);
379 			/*NOTREACHED*/
380 		}
381 }
382 
383 void
384 rollforward()
385 {
386 	register struct req *p, *q, *prev;
387 	register struct slave *tslp;
388 	int i, size, savedtapea, got;
389 	union u_spcl *ntb, *otb;
390 	tslp = &slaves[SLAVES];
391 	ntb = (union u_spcl *)tslp->tblock[1];
392 
393 	/*
394 	 * Each of the N slaves should have requests that need to
395 	 * be replayed on the next tape.  Use the extra slave buffers
396 	 * (slaves[SLAVES]) to construct request lists to be sent to
397 	 * each slave in turn.
398 	 */
399 	for (i = 0; i < SLAVES; i++) {
400 		q = &tslp->req[1];
401 		otb = (union u_spcl *)slp->tblock;
402 
403 		/*
404 		 * For each request in the current slave, copy it to tslp.
405 		 */
406 
407 		prev = NULL;
408 		for (p = slp->req; p->count > 0; p += p->count) {
409 			*q = *p;
410 			if (p->dblk == 0)
411 				*ntb++ = *otb++; /* copy the datablock also */
412 			prev = q;
413 			q += q->count;
414 		}
415 		if (prev == NULL)
416 			quit("rollforward: protocol botch");
417 		if (prev->dblk != 0)
418 			prev->count -= 1;
419 		else
420 			ntb--;
421 		q -= 1;
422 		q->count = 0;
423 		q = &tslp->req[0];
424 		if (i == 0) {
425 			q->dblk = 0;
426 			q->count = 1;
427 			trecno = 0;
428 			nextblock = tslp->tblock;
429 			savedtapea = spcl.c_tapea;
430 			spcl.c_tapea = slp->tapea;
431 			startnewtape(0);
432 			spcl.c_tapea = savedtapea;
433 			lastspclrec = savedtapea - 1;
434 		}
435 		size = (char *)ntb - (char *)q;
436 		if (atomic(write, slp->fd, (char *)q, size) != size) {
437 			perror("  DUMP: error writing command pipe");
438 			dumpabort(0);
439 		}
440 		slp->sent = 1;
441 		if (++slp >= &slaves[SLAVES])
442 			slp = &slaves[0];
443 
444 		q->count = 1;
445 
446 		if (prev->dblk != 0) {
447 			/*
448 			 * If the last one was a disk block, make the
449 			 * first of this one be the last bit of that disk
450 			 * block...
451 			 */
452 			q->dblk = prev->dblk +
453 				prev->count * (TP_BSIZE / DEV_BSIZE);
454 			ntb = (union u_spcl *)tslp->tblock;
455 		} else {
456 			/*
457 			 * It wasn't a disk block.  Copy the data to its
458 			 * new location in the buffer.
459 			 */
460 			q->dblk = 0;
461 			*((union u_spcl *)tslp->tblock) = *ntb;
462 			ntb = (union u_spcl *)tslp->tblock[1];
463 		}
464 	}
465 	slp->req[0] = *q;
466 	nextblock = slp->tblock;
467 	if (q->dblk == 0)
468 		nextblock++;
469 	trecno = 1;
470 
471 	/*
472 	 * Clear the first slaves' response.  One hopes that it
473 	 * worked ok, otherwise the tape is much too short!
474 	 */
475 	if (slp->sent) {
476 		if (atomic(read, slp->fd, (char *)&got, sizeof got)
477 		    != sizeof got) {
478 			perror("  DUMP: error reading command pipe in master");
479 			dumpabort(0);
480 		}
481 		slp->sent = 0;
482 
483 		if (got != writesize) {
484 			quit("EOT detected at start of the tape!\n");
485 		}
486 	}
487 }
488 
489 /*
490  * We implement taking and restoring checkpoints on the tape level.
491  * When each tape is opened, a new process is created by forking; this
492  * saves all of the necessary context in the parent.  The child
493  * continues the dump; the parent waits around, saving the context.
494  * If the child returns X_REWRITE, then it had problems writing that tape;
495  * this causes the parent to fork again, duplicating the context, and
496  * everything continues as if nothing had happened.
497  */
498 void
499 startnewtape(top)
500 	int top;
501 {
502 	int	parentpid;
503 	int	childpid;
504 	int	status;
505 	int	waitpid;
506 	char	*p;
507 #ifdef sunos
508 	void	(*interrupt_save)();
509 #else
510 	sig_t	interrupt_save;
511 #endif
512 
513 	interrupt_save = signal(SIGINT, SIG_IGN);
514 	parentpid = getpid();
515 
516 restore_check_point:
517 	(void)signal(SIGINT, interrupt_save);
518 	/*
519 	 *	All signals are inherited...
520 	 */
521 	childpid = fork();
522 	if (childpid < 0) {
523 		msg("Context save fork fails in parent %d\n", parentpid);
524 		Exit(X_ABORT);
525 	}
526 	if (childpid != 0) {
527 		/*
528 		 *	PARENT:
529 		 *	save the context by waiting
530 		 *	until the child doing all of the work returns.
531 		 *	don't catch the interrupt
532 		 */
533 		signal(SIGINT, SIG_IGN);
534 #ifdef TDEBUG
535 		msg("Tape: %d; parent process: %d child process %d\n",
536 			tapeno+1, parentpid, childpid);
537 #endif /* TDEBUG */
538 		while ((waitpid = wait(&status)) != childpid)
539 			msg("Parent %d waiting for child %d has another child %d return\n",
540 				parentpid, childpid, waitpid);
541 		if (status & 0xFF) {
542 			msg("Child %d returns LOB status %o\n",
543 				childpid, status&0xFF);
544 		}
545 		status = (status >> 8) & 0xFF;
546 #ifdef TDEBUG
547 		switch(status) {
548 			case X_FINOK:
549 				msg("Child %d finishes X_FINOK\n", childpid);
550 				break;
551 			case X_ABORT:
552 				msg("Child %d finishes X_ABORT\n", childpid);
553 				break;
554 			case X_REWRITE:
555 				msg("Child %d finishes X_REWRITE\n", childpid);
556 				break;
557 			default:
558 				msg("Child %d finishes unknown %d\n",
559 					childpid, status);
560 				break;
561 		}
562 #endif /* TDEBUG */
563 		switch(status) {
564 			case X_FINOK:
565 				Exit(X_FINOK);
566 			case X_ABORT:
567 				Exit(X_ABORT);
568 			case X_REWRITE:
569 				goto restore_check_point;
570 			default:
571 				msg("Bad return code from dump: %d\n", status);
572 				Exit(X_ABORT);
573 		}
574 		/*NOTREACHED*/
575 	} else {	/* we are the child; just continue */
576 #ifdef TDEBUG
577 		sleep(4);	/* allow time for parent's message to get out */
578 		msg("Child on Tape %d has parent %d, my pid = %d\n",
579 			tapeno+1, parentpid, getpid());
580 #endif /* TDEBUG */
581 		/*
582 		 * If we have a name like "/dev/rmt0,/dev/rmt1",
583 		 * use the name before the comma first, and save
584 		 * the remaining names for subsequent volumes.
585 		 */
586 		tapeno++;               /* current tape sequence */
587 		if (nexttape || index(tape, ',')) {
588 			if (nexttape && *nexttape)
589 				tape = nexttape;
590 			if ((p = index(tape, ',')) != NULL) {
591 				*p = '\0';
592 				nexttape = p + 1;
593 			} else
594 				nexttape = NULL;
595 			msg("Dumping volume %d on %s\n", tapeno, tape);
596 		}
597 #ifdef RDUMP
598 		while ((tapefd = (host ? rmtopen(tape, 2) :
599 			pipeout ? 1 : open(tape, O_WRONLY|O_CREAT, 0666))) < 0)
600 #else
601 		while ((tapefd = (pipeout ? 1 :
602 				  open(tape, O_WRONLY|O_CREAT, 0666))) < 0)
603 #endif
604 		    {
605 			msg("Cannot open output \"%s\".\n", tape);
606 			if (!query("Do you want to retry the open?"))
607 				dumpabort(0);
608 		}
609 
610 		enslave();  /* Share open tape file descriptor with slaves */
611 
612 		asize = 0;
613 		blocksthisvol = 0;
614 		if (top)
615 			newtape++;		/* new tape signal */
616 		spcl.c_count = slp->count;
617 		/*
618 		 * measure firstrec in TP_BSIZE units since restore doesn't
619 		 * know the correct ntrec value...
620 		 */
621 		spcl.c_firstrec = slp->firstrec;
622 		spcl.c_volume++;
623 		spcl.c_type = TS_TAPE;
624 		spcl.c_flags |= DR_NEWHEADER;
625 		writeheader((ino_t)slp->inode);
626 		spcl.c_flags &=~ DR_NEWHEADER;
627 		if (tapeno > 1)
628 			msg("Volume %d begins with blocks from inode %d\n",
629 				tapeno, slp->inode);
630 	}
631 }
632 
633 void
634 dumpabort(signo)
635 	int signo;
636 {
637 
638 	if (master != 0 && master != getpid())
639 		/* Signals master to call dumpabort */
640 		(void) kill(master, SIGTERM);
641 	else {
642 		killall();
643 		msg("The ENTIRE dump is aborted.\n");
644 	}
645 #ifdef RDUMP
646 	rmtclose();
647 #endif
648 	Exit(X_ABORT);
649 }
650 
651 __dead void
652 Exit(status)
653 	int status;
654 {
655 
656 #ifdef TDEBUG
657 	msg("pid = %d exits with status %d\n", getpid(), status);
658 #endif /* TDEBUG */
659 	exit(status);
660 }
661 
662 /*
663  * proceed - handler for SIGUSR2, used to synchronize IO between the slaves.
664  */
665 void
666 proceed(signo)
667 	int signo;
668 {
669 
670 	if (ready)
671 		longjmp(jmpbuf, 1);
672 	caught++;
673 }
674 
675 void
676 enslave()
677 {
678 	int cmd[2];
679 	register int i, j;
680 
681 	master = getpid();
682 
683 	signal(SIGTERM, dumpabort);  /* Slave sends SIGTERM on dumpabort() */
684 	signal(SIGPIPE, sigpipe);
685 	signal(SIGUSR1, tperror);    /* Slave sends SIGUSR1 on tape errors */
686 	signal(SIGUSR2, proceed);    /* Slave sends SIGUSR2 to next slave */
687 
688 	for (i = 0; i < SLAVES; i++) {
689 		if (i == slp - &slaves[0]) {
690 			caught = 1;
691 		} else {
692 			caught = 0;
693 		}
694 
695 		if (socketpair(AF_UNIX, SOCK_STREAM, 0, cmd) < 0 ||
696 		    (slaves[i].pid = fork()) < 0)
697 			quit("too many slaves, %d (recompile smaller): %s\n",
698 			    i, strerror(errno));
699 
700 		slaves[i].fd = cmd[1];
701 		slaves[i].sent = 0;
702 		if (slaves[i].pid == 0) { 	    /* Slave starts up here */
703 			for (j = 0; j <= i; j++)
704 			        (void) close(slaves[j].fd);
705 			signal(SIGINT, SIG_IGN);    /* Master handles this */
706 			doslave(cmd[0], i);
707 			Exit(X_FINOK);
708 		}
709 	}
710 
711 	for (i = 0; i < SLAVES; i++)
712 		(void) atomic(write, slaves[i].fd,
713 			      (char *) &slaves[(i + 1) % SLAVES].pid,
714 		              sizeof slaves[0].pid);
715 
716 	master = 0;
717 }
718 
719 void
720 killall()
721 {
722 	register int i;
723 
724 	for (i = 0; i < SLAVES; i++)
725 		if (slaves[i].pid > 0)
726 			(void) kill(slaves[i].pid, SIGKILL);
727 }
728 
729 /*
730  * Synchronization - each process has a lockfile, and shares file
731  * descriptors to the following process's lockfile.  When our write
732  * completes, we release our lock on the following process's lock-
733  * file, allowing the following process to lock it and proceed. We
734  * get the lock back for the next cycle by swapping descriptors.
735  */
736 static void
737 doslave(cmd, slave_number)
738 	register int cmd;
739         int slave_number;
740 {
741 	register int nread;
742 	int nextslave, size, wrote, eot_count;
743 
744 	/*
745 	 * Need our own seek pointer.
746 	 */
747 	(void) close(diskfd);
748 	if ((diskfd = open(disk, O_RDONLY)) < 0)
749 		quit("slave couldn't reopen disk: %s\n", strerror(errno));
750 
751 	/*
752 	 * Need the pid of the next slave in the loop...
753 	 */
754 	if ((nread = atomic(read, cmd, (char *)&nextslave, sizeof nextslave))
755 	    != sizeof nextslave) {
756 		quit("master/slave protocol botched - didn't get pid of next slave.\n");
757 	}
758 
759 	/*
760 	 * Get list of blocks to dump, read the blocks into tape buffer
761 	 */
762 	while ((nread = atomic(read, cmd, (char *)slp->req, reqsiz)) == reqsiz) {
763 		register struct req *p = slp->req;
764 
765 		for (trecno = 0; trecno < ntrec;
766 		     trecno += p->count, p += p->count) {
767 			if (p->dblk) {
768 				bread(p->dblk, slp->tblock[trecno],
769 					p->count * TP_BSIZE);
770 			} else {
771 				if (p->count != 1 || atomic(read, cmd,
772 				    (char *)slp->tblock[trecno],
773 				    TP_BSIZE) != TP_BSIZE)
774 				       quit("master/slave protocol botched.\n");
775 			}
776 		}
777 		if (setjmp(jmpbuf) == 0) {
778 			ready = 1;
779 			if (!caught)
780 				(void) pause();
781 		}
782 		ready = 0;
783 		caught = 0;
784 
785 		/* Try to write the data... */
786 		eot_count = 0;
787 		size = 0;
788 
789 		while (eot_count < 10 && size < writesize) {
790 #ifdef RDUMP
791 			if (host)
792 				wrote = rmtwrite(slp->tblock[0]+size,
793 				    writesize-size);
794 			else
795 #endif
796 				wrote = write(tapefd, slp->tblock[0]+size,
797 				    writesize-size);
798 #ifdef WRITEDEBUG
799 			printf("slave %d wrote %d\n", slave_number, wrote);
800 #endif
801 			if (wrote < 0)
802 				break;
803 			if (wrote == 0)
804 				eot_count++;
805 			size += wrote;
806 		}
807 
808 #ifdef WRITEDEBUG
809 		if (size != writesize)
810 		 printf("slave %d only wrote %d out of %d bytes and gave up.\n",
811 		     slave_number, size, writesize);
812 #endif
813 
814 		if (eot_count > 0)
815 			size = 0;
816 
817 		/*
818 		 * fixme: Pyramids running OSx return ENOSPC
819 		 * at EOT on 1/2 inch drives.
820 		 */
821 		if (size < 0) {
822 			(void) kill(master, SIGUSR1);
823 			for (;;)
824 				(void) sigpause(0);
825 		} else {
826 			/*
827 			 * pass size of write back to master
828 			 * (for EOT handling)
829 			 */
830 			(void) atomic(write, cmd, (char *)&size, sizeof size);
831 		}
832 
833 		/*
834 		 * If partial write, don't want next slave to go.
835 		 * Also jolts him awake.
836 		 */
837 		(void) kill(nextslave, SIGUSR2);
838 	}
839 	if (nread != 0)
840 		quit("error reading command pipe: %s\n", strerror(errno));
841 }
842 
843 /*
844  * Since a read from a pipe may not return all we asked for,
845  * or a write may not write all we ask if we get a signal,
846  * loop until the count is satisfied (or error).
847  */
848 static int
849 atomic(func, fd, buf, count)
850 	int (*func)(), fd;
851 	char *buf;
852 	int count;
853 {
854 	int got, need = count;
855 
856 	while ((got = (*func)(fd, buf, need)) > 0 && (need -= got) > 0)
857 		buf += got;
858 	return (got < 0 ? got : count - need);
859 }
860