xref: /openbsd-src/sbin/dump/tape.c (revision 50b7afb2c2c0993b0894d4e34bf857cb13ed9c80)
1 /*	$OpenBSD: tape.c,v 1.40 2014/06/13 20:43:06 naddy Exp $	*/
2 /*	$NetBSD: tape.c,v 1.11 1997/06/05 11:13:26 lukem Exp $	*/
3 
4 /*-
5  * Copyright (c) 1980, 1991, 1993
6  *	The Regents of the University of California.  All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. Neither the name of the University nor the names of its contributors
17  *    may be used to endorse or promote products derived from this software
18  *    without specific prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  */
32 
33 #include <sys/types.h>
34 #include <sys/param.h>
35 #include <sys/socket.h>
36 #include <sys/time.h>
37 #include <sys/wait.h>
38 #include <sys/stat.h>
39 #include <ufs/ffs/fs.h>
40 #include <ufs/ufs/dinode.h>
41 
42 #include <protocols/dumprestore.h>
43 
44 #include <errno.h>
45 #include <fcntl.h>
46 #include <signal.h>
47 #include <stdio.h>
48 #include <stdlib.h>
49 #include <string.h>
50 #include <time.h>
51 #include <unistd.h>
52 
53 #include "dump.h"
54 #include "pathnames.h"
55 
56 int	writesize;		/* size of malloc()ed buffer for tape */
57 int64_t	lastspclrec = -1;	/* tape block number of last written header */
58 int	trecno = 0;		/* next record to write in current block */
59 extern	int64_t blocksperfile;	/* number of blocks per output file */
60 int64_t	blocksthisvol;		/* number of blocks on current output file */
61 extern	int ntrec;		/* blocking factor on tape */
62 extern	int cartridge;
63 extern	char *host;
64 char	*nexttape;
65 
66 static	ssize_t atomic(ssize_t (*)(int, void *, size_t), int, char *, int);
67 static	void doslave(int, int);
68 static	void enslave(void);
69 static	void flushtape(void);
70 static	void killall(void);
71 static	void rollforward(void);
72 
73 void	tperror(int signo);
74 void	sigpipe(int signo);
75 void	proceed(int signo);
76 
77 /*
78  * Concurrent dump mods (Caltech) - disk block reading and tape writing
79  * are exported to several slave processes.  While one slave writes the
80  * tape, the others read disk blocks; they pass control of the tape in
81  * a ring via signals. The parent process traverses the filesystem and
82  * sends writeheader()'s and lists of daddr's to the slaves via pipes.
83  * The following structure defines the instruction packets sent to slaves.
84  */
85 struct req {
86 	daddr_t dblk;
87 	int count;
88 };
89 int reqsiz;
90 
91 #define SLAVES 3		/* 1 slave writing, 1 reading, 1 for slack */
92 struct slave {
93 	int64_t tapea;		/* header number at start of this chunk */
94 	int64_t firstrec;	/* record number of this block */
95 	int count;		/* count to next header (used for TS_TAPE */
96 				/* after EOT) */
97 	int inode;		/* inode that we are currently dealing with */
98 	int fd;			/* FD for this slave */
99 	pid_t pid;		/* PID for this slave */
100 	int sent;		/* 1 == we've sent this slave requests */
101 	char (*tblock)[TP_BSIZE]; /* buffer for data blocks */
102 	struct req *req;	/* buffer for requests */
103 } slaves[SLAVES+1];
104 struct slave *slp;
105 
106 char	(*nextblock)[TP_BSIZE];
107 
108 static time_t tstart_volume;	/* time of volume start */
109 static int64_t tapea_volume;	/* value of spcl.c_tapea at volume start */
110 
111 pid_t master;		/* pid of master, for sending error signals */
112 int tenths;		/* length of tape used per block written */
113 static volatile sig_atomic_t caught;	/* have we caught the signal to proceed? */
114 
115 int
116 alloctape(void)
117 {
118 	int pgoff = getpagesize() - 1;
119 	char *buf;
120 	int i;
121 
122 	writesize = ntrec * TP_BSIZE;
123 	reqsiz = (ntrec + 1) * sizeof(struct req);
124 	/*
125 	 * CDC 92181's and 92185's make 0.8" gaps in 1600-bpi start/stop mode
126 	 * (see DEC TU80 User's Guide).  The shorter gaps of 6250-bpi require
127 	 * repositioning after stopping, i.e, streaming mode, where the gap is
128 	 * variable, 0.30" to 0.45".  The gap is maximal when the tape stops.
129 	 */
130 	if (blocksperfile == 0 && !unlimited)
131 		tenths = writesize / density +
132 		    (cartridge ? 16 : density == 625 ? 5 : 8);
133 	/*
134 	 * Allocate tape buffer contiguous with the array of instruction
135 	 * packets, so flushtape() can write them together with one write().
136 	 * Align tape buffer on page boundary to speed up tape write().
137 	 */
138 	for (i = 0; i <= SLAVES; i++) {
139 		buf = (char *)
140 		    malloc((unsigned)(reqsiz + writesize + pgoff + TP_BSIZE));
141 		if (buf == NULL)
142 			return(0);
143 		slaves[i].tblock = (char (*)[TP_BSIZE])
144 		    (((long)&buf[ntrec + 1] + pgoff) &~ pgoff);
145 		slaves[i].req = (struct req *)slaves[i].tblock - ntrec - 1;
146 	}
147 	slp = &slaves[0];
148 	slp->count = 1;
149 	slp->tapea = 0;
150 	slp->firstrec = 0;
151 	nextblock = slp->tblock;
152 	return(1);
153 }
154 
155 void
156 writerec(char *dp, int isspcl)
157 {
158 
159 	slp->req[trecno].dblk = 0;
160 	slp->req[trecno].count = 1;
161 	*(union u_spcl *)(*(nextblock)++) = *(union u_spcl *)dp;
162 	if (isspcl)
163 		lastspclrec = spcl.c_tapea;
164 	trecno++;
165 	spcl.c_tapea++;
166 	if (trecno >= ntrec)
167 		flushtape();
168 }
169 
170 void
171 dumpblock(daddr_t blkno, int size)
172 {
173 	int avail, tpblks;
174 	daddr_t dblkno;
175 
176 	dblkno = fsbtodb(sblock, blkno);
177 	tpblks = size >> tp_bshift;
178 	while ((avail = MIN(tpblks, ntrec - trecno)) > 0) {
179 		slp->req[trecno].dblk = dblkno;
180 		slp->req[trecno].count = avail;
181 		trecno += avail;
182 		spcl.c_tapea += avail;
183 		if (trecno >= ntrec)
184 			flushtape();
185 		dblkno += avail << (tp_bshift - (ffs(DEV_BSIZE) - 1));
186 		tpblks -= avail;
187 	}
188 }
189 
190 int	nogripe = 0;
191 
192 /* ARGSUSED */
193 void
194 tperror(int signo)
195 {
196 	/* XXX - signal races */
197 
198 	if (pipeout) {
199 		msg("write error on %s\n", tape);
200 		quit("Cannot recover\n");
201 		/* NOTREACHED */
202 	}
203 	msg("write error %lld blocks into volume %d\n",
204 	    (long long)blocksthisvol, tapeno);
205 	broadcast("DUMP WRITE ERROR!\n");
206 	if (!query("Do you want to restart?"))
207 		dumpabort(0);
208 	msg("Closing this volume.  Prepare to restart with new media;\n");
209 	msg("this dump volume will be rewritten.\n");
210 	killall();
211 	nogripe = 1;
212 	close_rewind();
213 	Exit(X_REWRITE);
214 }
215 
216 /* ARGSUSED */
217 void
218 sigpipe(int signo)
219 {
220 
221 	quit("Broken pipe\n");
222 }
223 
224 /*
225  * do_stats --
226  *	Update xferrate stats
227  */
228 time_t
229 do_stats(void)
230 {
231 	time_t tnow, ttaken;
232 	int64_t blocks;
233 
234 	(void)time(&tnow);
235 	ttaken = tnow - tstart_volume;
236 	blocks = spcl.c_tapea - tapea_volume;
237 	msg("Volume %d completed at: %s", tapeno, ctime(&tnow));
238 	if (ttaken > 0) {
239 		msg("Volume %d took %lld:%02lld:%02lld\n", tapeno,
240 		    (long long)ttaken / 3600, ((long long)ttaken % 3600) / 60,
241 		    (long long)ttaken % 60);
242 		blocks /= ttaken;
243 		msg("Volume %d transfer rate: %lld KB/s\n", tapeno, blocks);
244 		xferrate += blocks;
245 	}
246 	return(tnow);
247 }
248 
249 /*
250  * statussig --
251  *	information message upon receipt of SIGINFO
252  *	(derived from optr.c::timeest())
253  * XXX not safe
254  */
255 /* ARGSUSED */
256 void
257 statussig(int signo)
258 {
259 	time_t	tnow, deltat;
260 	char	msgbuf[128];
261 	int save_errno = errno;
262 
263 	if (blockswritten < 500)
264 		return;
265 	(void) time(&tnow);
266 	deltat = tstart_writing - tnow + (1.0 * (tnow - tstart_writing))
267 		/ blockswritten * tapesize;
268 	(void)snprintf(msgbuf, sizeof(msgbuf),
269 	    "dump: %s %3.2f%% done at %lld KB/s, finished in %d:%02d\n",
270 	    tape, (blockswritten * 100.0) / tapesize,
271 	    (spcl.c_tapea - tapea_volume) / (tnow - tstart_volume),
272 	    (int)(deltat / 3600), (int)((deltat % 3600) / 60));
273 	write(STDERR_FILENO, msgbuf, strlen(msgbuf));
274 	errno = save_errno;
275 }
276 
277 static void
278 flushtape(void)
279 {
280 	int i, blks, got;
281 	int64_t lastfirstrec;
282 
283 	int siz = (char *)nextblock - (char *)slp->req;
284 
285 	slp->req[trecno].count = 0;			/* Sentinel */
286 
287 	if (atomic((ssize_t (*)(int, void *, size_t))write, slp->fd,
288 	    (char *)slp->req, siz) != siz)
289 		quit("error writing command pipe: %s\n", strerror(errno));
290 	slp->sent = 1; /* we sent a request, read the response later */
291 
292 	lastfirstrec = slp->firstrec;
293 
294 	if (++slp >= &slaves[SLAVES])
295 		slp = &slaves[0];
296 
297 	/* Read results back from next slave */
298 	if (slp->sent) {
299 		if (atomic(read, slp->fd, (char *)&got, sizeof(got))
300 		    != sizeof(got)) {
301 			perror("  DUMP: error reading command pipe in master");
302 			dumpabort(0);
303 		}
304 		slp->sent = 0;
305 
306 		/* Check for end of tape */
307 		if (got < writesize) {
308 			msg("End of tape detected\n");
309 
310 			/*
311 			 * Drain the results, don't care what the values were.
312 			 * If we read them here then trewind won't...
313 			 */
314 			for (i = 0; i < SLAVES; i++) {
315 				if (slaves[i].sent) {
316 					if (atomic(read, slaves[i].fd,
317 					    (char *)&got, sizeof(got))
318 					    != sizeof(got)) {
319 						perror("  DUMP: error reading command pipe in master");
320 						dumpabort(0);
321 					}
322 					slaves[i].sent = 0;
323 				}
324 			}
325 
326 			close_rewind();
327 			rollforward();
328 			return;
329 		}
330 	}
331 
332 	blks = 0;
333 	if (spcl.c_type != TS_END) {
334 		for (i = 0; i < spcl.c_count; i++)
335 			if (spcl.c_addr[i] != 0)
336 				blks++;
337 	}
338 	slp->count = lastspclrec + blks + 1 - spcl.c_tapea;
339 	slp->tapea = spcl.c_tapea;
340 	slp->firstrec = lastfirstrec + ntrec;
341 	slp->inode = curino;
342 	nextblock = slp->tblock;
343 	trecno = 0;
344 	asize += tenths;
345 	blockswritten += ntrec;
346 	blocksthisvol += ntrec;
347 	if (!pipeout && !unlimited && (blocksperfile ?
348 	    (blocksthisvol >= blocksperfile) : (asize > tsize))) {
349 		close_rewind();
350 		startnewtape(0);
351 	}
352 	timeest();
353 }
354 
355 void
356 trewind(void)
357 {
358 	struct stat sb;
359 	int f, got;
360 
361 	for (f = 0; f < SLAVES; f++) {
362 		/*
363 		 * Drain the results, but unlike EOT we DO (or should) care
364 		 * what the return values were, since if we detect EOT after
365 		 * we think we've written the last blocks to the tape anyway,
366 		 * we have to replay those blocks with rollforward.
367 		 *
368 		 * fixme: punt for now.
369 		 */
370 		if (slaves[f].sent) {
371 			if (atomic(read, slaves[f].fd, (char *)&got, sizeof(got))
372 			    != sizeof(got)) {
373 				perror("  DUMP: error reading command pipe in master");
374 				dumpabort(0);
375 			}
376 			slaves[f].sent = 0;
377 			if (got != writesize) {
378 				msg("EOT detected in last 2 tape records!\n");
379 				msg("Use a longer tape, decrease the size estimate\n");
380 				quit("or use no size estimate at all.\n");
381 			}
382 		}
383 		(void) close(slaves[f].fd);
384 	}
385 	while (wait((int *)NULL) >= 0)	/* wait for any signals from slaves */
386 		/* void */;
387 
388 	if (pipeout)
389 		return;
390 
391 	msg("Closing %s\n", tape);
392 
393 #ifdef RDUMP
394 	if (host) {
395 		rmtclose();
396 		while (rmtopen(tape, O_RDONLY) < 0)
397 			sleep(10);
398 		rmtclose();
399 		return;
400 	}
401 #endif
402 	/*
403 	 * st(4) says: "Bit 1 of the minor number specifies whether an eject is
404 	 * attempted when the device is closed.  When it is set, the device
405 	 * will attempt to eject its media on close ...".
406 	 *
407 	 * If the tape has been ejected, looping on open() will generate 'Media
408 	 * not present' errors until a tape is loaded. Once loaded the tape
409 	 * will be immediately ejected as a result of the second close().
410 	 *
411 	 * So if the tape will be ejected, just close and return.
412 	 */
413 	if ((fstat(tapefd, &sb) == 0) && (minor(sb.st_rdev) & 0x02)) {
414 		(void) close(tapefd);
415 		return;
416 	}
417 
418 	(void) close(tapefd);
419 	while ((f = open(tape, O_RDONLY)) < 0)
420 		sleep (10);
421 	(void) close(f);
422 }
423 
424 void
425 close_rewind(void)
426 {
427 	trewind();
428 	(void)do_stats();
429 	if (nexttape)
430 		return;
431 	if (!nogripe) {
432 		msg("Change Volumes: Mount volume #%d\n", tapeno+1);
433 		broadcast("CHANGE DUMP VOLUMES!\7\7\n");
434 	}
435 	while (!query("Is the new volume mounted and ready to go?"))
436 		if (query("Do you want to abort?")) {
437 			dumpabort(0);
438 			/*NOTREACHED*/
439 		}
440 }
441 
442 void
443 rollforward(void)
444 {
445 	struct req *p, *q, *prev;
446 	struct slave *tslp;
447 	int i, size, got;
448 	int64_t savedtapea;
449 	union u_spcl *ntb, *otb;
450 	tslp = &slaves[SLAVES];
451 	ntb = (union u_spcl *)tslp->tblock[1];
452 
453 	/*
454 	 * Each of the N slaves should have requests that need to
455 	 * be replayed on the next tape.  Use the extra slave buffers
456 	 * (slaves[SLAVES]) to construct request lists to be sent to
457 	 * each slave in turn.
458 	 */
459 	for (i = 0; i < SLAVES; i++) {
460 		q = &tslp->req[1];
461 		otb = (union u_spcl *)slp->tblock;
462 
463 		/*
464 		 * For each request in the current slave, copy it to tslp.
465 		 */
466 
467 		prev = NULL;
468 		for (p = slp->req; p->count > 0; p += p->count) {
469 			*q = *p;
470 			if (p->dblk == 0)
471 				*ntb++ = *otb++; /* copy the datablock also */
472 			prev = q;
473 			q += q->count;
474 		}
475 		if (prev == NULL)
476 			quit("rollforward: protocol botch\n");
477 		if (prev->dblk != 0)
478 			prev->count -= 1;
479 		else
480 			ntb--;
481 		q -= 1;
482 		q->count = 0;
483 		q = &tslp->req[0];
484 		if (i == 0) {
485 			q->dblk = 0;
486 			q->count = 1;
487 			trecno = 0;
488 			nextblock = tslp->tblock;
489 			savedtapea = spcl.c_tapea;
490 			spcl.c_tapea = slp->tapea;
491 			startnewtape(0);
492 			spcl.c_tapea = savedtapea;
493 			lastspclrec = savedtapea - 1;
494 		}
495 		size = (char *)ntb - (char *)q;
496 		if (atomic((ssize_t (*)(int, void *, size_t))write,
497 		    slp->fd, (char *)q, size) != size) {
498 			perror("  DUMP: error writing command pipe");
499 			dumpabort(0);
500 		}
501 		slp->sent = 1;
502 		if (++slp >= &slaves[SLAVES])
503 			slp = &slaves[0];
504 
505 		q->count = 1;
506 
507 		if (prev->dblk != 0) {
508 			/*
509 			 * If the last one was a disk block, make the
510 			 * first of this one be the last bit of that disk
511 			 * block...
512 			 */
513 			q->dblk = prev->dblk +
514 				prev->count * (TP_BSIZE / DEV_BSIZE);
515 			ntb = (union u_spcl *)tslp->tblock;
516 		} else {
517 			/*
518 			 * It wasn't a disk block.  Copy the data to its
519 			 * new location in the buffer.
520 			 */
521 			q->dblk = 0;
522 			*((union u_spcl *)tslp->tblock) = *ntb;
523 			ntb = (union u_spcl *)tslp->tblock[1];
524 		}
525 	}
526 	slp->req[0] = *q;
527 	nextblock = slp->tblock;
528 	if (q->dblk == 0)
529 		nextblock++;
530 	trecno = 1;
531 
532 	/*
533 	 * Clear the first slaves' response.  One hopes that it
534 	 * worked ok, otherwise the tape is much too short!
535 	 */
536 	if (slp->sent) {
537 		if (atomic(read, slp->fd, (char *)&got, sizeof(got))
538 		    != sizeof(got)) {
539 			perror("  DUMP: error reading command pipe in master");
540 			dumpabort(0);
541 		}
542 		slp->sent = 0;
543 
544 		if (got != writesize) {
545 			quit("EOT detected at start of the tape!\n");
546 		}
547 	}
548 }
549 
550 /*
551  * We implement taking and restoring checkpoints on the tape level.
552  * When each tape is opened, a new process is created by forking; this
553  * saves all of the necessary context in the parent.  The child
554  * continues the dump; the parent waits around, saving the context.
555  * If the child returns X_REWRITE, then it had problems writing that tape;
556  * this causes the parent to fork again, duplicating the context, and
557  * everything continues as if nothing had happened.
558  */
559 void
560 startnewtape(int top)
561 {
562 	pid_t	parentpid;
563 	pid_t	childpid;
564 	int	status;
565 	pid_t	waitingpid;
566 	char	*p;
567 	sig_t	interrupt_save;
568 
569 	interrupt_save = signal(SIGINT, SIG_IGN);
570 	parentpid = getpid();
571 	tapea_volume = spcl.c_tapea;
572 	(void)time(&tstart_volume);
573 
574 restore_check_point:
575 	(void)signal(SIGINT, interrupt_save);
576 	/*
577 	 *	All signals are inherited...
578 	 */
579 	childpid = fork();
580 	if (childpid < 0) {
581 		msg("Context save fork fails in parent %d\n", parentpid);
582 		Exit(X_ABORT);
583 	}
584 	if (childpid != 0) {
585 		/*
586 		 *	PARENT:
587 		 *	save the context by waiting
588 		 *	until the child doing all of the work returns.
589 		 *	don't catch the interrupt
590 		 */
591 		signal(SIGINT, SIG_IGN);
592 #ifdef TDEBUG
593 		msg("Tape: %d; parent process: %d child process %d\n",
594 			tapeno+1, parentpid, childpid);
595 #endif /* TDEBUG */
596 		while ((waitingpid = wait(&status)) != childpid)
597 			msg("Parent %d waiting for child %d has another child %d return\n",
598 				parentpid, childpid, waitingpid);
599 		if (status & 0xFF) {
600 			msg("Child %d returns LOB status %o\n",
601 				childpid, status&0xFF);
602 		}
603 		status = (status >> 8) & 0xFF;
604 #ifdef TDEBUG
605 		switch(status) {
606 			case X_FINOK:
607 				msg("Child %d finishes X_FINOK\n", childpid);
608 				break;
609 			case X_ABORT:
610 				msg("Child %d finishes X_ABORT\n", childpid);
611 				break;
612 			case X_REWRITE:
613 				msg("Child %d finishes X_REWRITE\n", childpid);
614 				break;
615 			default:
616 				msg("Child %d finishes unknown %d\n",
617 					childpid, status);
618 				break;
619 		}
620 #endif /* TDEBUG */
621 		switch(status) {
622 			case X_FINOK:
623 				Exit(X_FINOK);
624 				break;
625 			case X_ABORT:
626 				Exit(X_ABORT);
627 				break;
628 			case X_REWRITE:
629 				goto restore_check_point;
630 			default:
631 				msg("Bad return code from dump: %d\n", status);
632 				Exit(X_ABORT);
633 		}
634 		/*NOTREACHED*/
635 	} else {	/* we are the child; just continue */
636 #ifdef TDEBUG
637 		sleep(4);	/* allow time for parent's message to get out */
638 		msg("Child on Tape %d has parent %d, my pid = %d\n",
639 			tapeno+1, parentpid, getpid());
640 #endif /* TDEBUG */
641 		/*
642 		 * If we have a name like "/dev/rst0,/dev/rst1",
643 		 * use the name before the comma first, and save
644 		 * the remaining names for subsequent volumes.
645 		 */
646 		tapeno++;               /* current tape sequence */
647 		if (nexttape || strchr(tape, ',')) {
648 			if (nexttape && *nexttape)
649 				tape = nexttape;
650 			if ((p = strchr(tape, ',')) != NULL) {
651 				*p = '\0';
652 				nexttape = p + 1;
653 			} else
654 				nexttape = NULL;
655 			msg("Dumping volume %d on %s\n", tapeno, tape);
656 		}
657 #ifdef RDUMP
658 		while ((tapefd = (host ? rmtopen(tape, O_WRONLY|O_CREAT) :
659 			pipeout ? 1 : open(tape, O_WRONLY|O_CREAT, 0666))) < 0)
660 #else
661 		while ((tapefd = (pipeout ? 1 :
662 				  open(tape, O_WRONLY|O_CREAT, 0666))) < 0)
663 #endif
664 		    {
665 			msg("Cannot open output \"%s\".\n", tape);
666 			if (!query("Do you want to retry the open?"))
667 				dumpabort(0);
668 		}
669 
670 		enslave();  /* Share open tape file descriptor with slaves */
671 
672 		asize = 0;
673 		blocksthisvol = 0;
674 		if (top)
675 			newtape++;		/* new tape signal */
676 		spcl.c_count = slp->count;
677 		/*
678 		 * measure firstrec in TP_BSIZE units since restore doesn't
679 		 * know the correct ntrec value...
680 		 */
681 		spcl.c_firstrec = slp->firstrec;
682 		spcl.c_volume++;
683 		spcl.c_type = TS_TAPE;
684 		if (sblock->fs_magic != FS_UFS2_MAGIC)
685 			spcl.c_flags |= DR_NEWHEADER;
686 		writeheader((ino_t)slp->inode);
687 		if (sblock->fs_magic != FS_UFS2_MAGIC)
688 			spcl.c_flags &=~ DR_NEWHEADER;
689 		msg("Volume %d started at: %s", tapeno, ctime(&tstart_volume));
690 		if (tapeno > 1)
691 			msg("Volume %d begins with blocks from inode %llu\n",
692 			    tapeno, (unsigned long long)slp->inode);
693 	}
694 }
695 
696 /* ARGSUSED */
697 void
698 dumpabort(int signo)
699 {
700 
701 	if (master != 0 && master != getpid())
702 		/* Signals master to call dumpabort */
703 		(void) kill(master, SIGTERM);
704 	else {
705 		killall();
706 		msg("The ENTIRE dump is aborted.\n");
707 	}
708 #ifdef RDUMP
709 	rmtclose();
710 #endif
711 	Exit(X_ABORT);
712 }
713 
714 __dead void
715 Exit(int status)
716 {
717 
718 #ifdef TDEBUG
719 	msg("pid = %d exits with status %d\n", getpid(), status);
720 #endif /* TDEBUG */
721 	exit(status);
722 }
723 
724 /*
725  * proceed - handler for SIGUSR2, used to synchronize IO between the slaves.
726  */
727 /* ARGSUSED */
728 void
729 proceed(int signo)
730 {
731 	caught++;
732 }
733 
734 void
735 enslave(void)
736 {
737 	int cmd[2];
738 	int i, j;
739 
740 	master = getpid();
741 
742 	signal(SIGTERM, dumpabort);  /* Slave sends SIGTERM on dumpabort() */
743 	signal(SIGPIPE, sigpipe);
744 	signal(SIGUSR1, tperror);    /* Slave sends SIGUSR1 on tape errors */
745 	signal(SIGUSR2, proceed);    /* Slave sends SIGUSR2 to next slave */
746 
747 	for (i = 0; i < SLAVES; i++) {
748 		if (i == slp - &slaves[0]) {
749 			caught = 1;
750 		} else {
751 			caught = 0;
752 		}
753 
754 		if (socketpair(AF_UNIX, SOCK_STREAM, 0, cmd) < 0 ||
755 		    (slaves[i].pid = fork()) < 0)
756 			quit("too many slaves, %d (recompile smaller): %s\n",
757 			    i, strerror(errno));
758 
759 		slaves[i].fd = cmd[1];
760 		slaves[i].sent = 0;
761 		if (slaves[i].pid == 0) { 	    /* Slave starts up here */
762 			for (j = 0; j <= i; j++)
763 			        (void) close(slaves[j].fd);
764 			signal(SIGINT, SIG_IGN);    /* Master handles this */
765 			signal(SIGINFO, SIG_IGN);
766 			doslave(cmd[0], i);
767 			Exit(X_FINOK);
768 		}
769 	}
770 
771 	for (i = 0; i < SLAVES; i++)
772 		(void) atomic((ssize_t (*)(int, void *, size_t))write,
773 		    slaves[i].fd, (char *) &slaves[(i + 1) % SLAVES].pid,
774 		    sizeof(slaves[0].pid));
775 	master = 0;
776 }
777 
778 void
779 killall(void)
780 {
781 	int i;
782 
783 	for (i = 0; i < SLAVES; i++)
784 		if (slaves[i].pid > 0) {
785 			(void) kill(slaves[i].pid, SIGKILL);
786 			slaves[i].pid = 0;
787 		}
788 }
789 
790 /*
791  * Synchronization - each process has a lockfile, and shares file
792  * descriptors to the following process's lockfile.  When our write
793  * completes, we release our lock on the following process's lock-
794  * file, allowing the following process to lock it and proceed. We
795  * get the lock back for the next cycle by swapping descriptors.
796  */
797 static void
798 doslave(int cmd, int slave_number)
799 {
800 	int nread, nextslave, size, wrote = 0, eot_count;
801 	sigset_t nsigset, osigset;
802 
803 	/*
804 	 * Need our own seek pointer.
805 	 */
806 	(void) close(diskfd);
807 	if ((diskfd = open(disk, O_RDONLY)) < 0)
808 		quit("slave couldn't reopen disk: %s\n", strerror(errno));
809 
810 	/*
811 	 * Need the pid of the next slave in the loop...
812 	 */
813 	if ((nread = atomic(read, cmd, (char *)&nextslave, sizeof(nextslave)))
814 	    != sizeof(nextslave)) {
815 		quit("master/slave protocol botched - didn't get pid of next slave.\n");
816 	}
817 
818 	/*
819 	 * Get list of blocks to dump, read the blocks into tape buffer
820 	 */
821 	while ((nread = atomic(read, cmd, (char *)slp->req, reqsiz)) == reqsiz) {
822 		struct req *p = slp->req;
823 
824 		for (trecno = 0; trecno < ntrec;
825 		     trecno += p->count, p += p->count) {
826 			if (p->dblk) {
827 				bread(p->dblk, slp->tblock[trecno],
828 					p->count * TP_BSIZE);
829 			} else {
830 				if (p->count != 1 || atomic(read, cmd,
831 				    (char *)slp->tblock[trecno],
832 				    TP_BSIZE) != TP_BSIZE)
833 				       quit("master/slave protocol botched.\n");
834 			}
835 		}
836 
837 		sigemptyset(&nsigset);
838 		sigaddset(&nsigset, SIGUSR2);
839 		sigprocmask(SIG_BLOCK, &nsigset, &osigset);
840 		while (!caught)
841 			sigsuspend(&osigset);
842 		caught = 0;
843 		sigprocmask(SIG_SETMASK, &osigset, NULL);
844 
845 		/* Try to write the data... */
846 		eot_count = 0;
847 		size = 0;
848 
849 		while (eot_count < 10 && size < writesize) {
850 #ifdef RDUMP
851 			if (host)
852 				wrote = rmtwrite(slp->tblock[0]+size,
853 				    writesize-size);
854 			else
855 #endif
856 				wrote = write(tapefd, slp->tblock[0]+size,
857 				    writesize-size);
858 #ifdef WRITEDEBUG
859 			printf("slave %d wrote %d\n", slave_number, wrote);
860 #endif
861 			if (wrote < 0)
862 				break;
863 			if (wrote == 0)
864 				eot_count++;
865 			size += wrote;
866 		}
867 
868 #ifdef WRITEDEBUG
869 		if (size != writesize)
870 		 printf("slave %d only wrote %d out of %d bytes and gave up.\n",
871 		     slave_number, size, writesize);
872 #endif
873 
874 		if (eot_count > 0)
875 			size = 0;
876 
877 		/*
878 		 * Handle ENOSPC as an EOT condition
879 		 */
880 		if (wrote < 0 && errno == ENOSPC) {
881 			wrote = 0;
882 			eot_count++;
883 		}
884 
885 		if (size < 0) {
886 			(void) kill(master, SIGUSR1);
887 			sigemptyset(&nsigset);
888 			for (;;)
889 				sigsuspend(&nsigset);
890 		} else {
891 			/*
892 			 * pass size of write back to master
893 			 * (for EOT handling)
894 			 */
895 			(void) atomic((ssize_t (*)(int, void *, size_t))write,
896 			    cmd, (char *)&size, sizeof(size));
897 		}
898 
899 		/*
900 		 * If partial write, don't want next slave to go.
901 		 * Also jolts him awake.
902 		 */
903 		(void) kill(nextslave, SIGUSR2);
904 	}
905 	if (nread != 0)
906 		quit("error reading command pipe: %s\n", strerror(errno));
907 }
908 
909 /*
910  * Since a read from a pipe may not return all we asked for,
911  * or a write may not write all we ask if we get a signal,
912  * loop until the count is satisfied (or error).
913  */
914 static ssize_t
915 atomic(ssize_t (*func)(int, void *, size_t), int fd, char *buf, int count)
916 {
917 	ssize_t got, need = count;
918 
919 	while ((got = (*func)(fd, buf, need)) > 0 && (need -= got) > 0)
920 		buf += got;
921 	return (got < 0 ? got : count - need);
922 }
923