xref: /openbsd/sbin/dump/tape.c (revision a7b9eedc)
1 /*	$OpenBSD: tape.c,v 1.49 2024/05/09 08:35:40 florian 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/param.h>	/* MAXBSIZE DEV_BSIZE */
34 #include <sys/socket.h>
35 #include <sys/time.h>
36 #include <sys/wait.h>
37 #include <sys/stat.h>
38 #include <ufs/ffs/fs.h>
39 #include <ufs/ufs/dinode.h>
40 
41 #include <protocols/dumprestore.h>
42 
43 #include <errno.h>
44 #include <fcntl.h>
45 #include <signal.h>
46 #include <stdio.h>
47 #include <stdlib.h>
48 #include <string.h>
49 #include <time.h>
50 #include <unistd.h>
51 #include <limits.h>
52 
53 #include "dump.h"
54 #include "pathnames.h"
55 
56 #define MINIMUM(a, b)	(((a) < (b)) ? (a) : (b))
57 
58 ino_t   curino; 		/* current inumber; used globally */
59 int     newtape;		/* new tape flag */
60 union u_spcl u_spcl;		/* mapping of variables in a control block */
61 
62 static int tapefd;		/* tape file descriptor */
63 static int64_t asize;		/* number of 0.1" units written on cur tape */
64 static int writesize;		/* size of malloc()ed buffer for tape */
65 static int64_t	lastspclrec = -1; /* tape block number of last written header */
66 static int trecno = 0;		/* next record to write in current block */
67 static int64_t blocksthisvol;	/* number of blocks on current output file */
68 static char *nexttape;
69 
70 static	ssize_t atomic(ssize_t (*)(int, void *, size_t), int, char *, int);
71 static	void doslave(int, int);
72 static	void enslave(void);
73 static	void flushtape(void);
74 static	void killall(void);
75 static	void rollforward(void);
76 
77 void	tperror(int signo);
78 void	sigpipe(int signo);
79 void	proceed(int signo);
80 
81 /*
82  * Concurrent dump mods (Caltech) - disk block reading and tape writing
83  * are exported to several slave processes.  While one slave writes the
84  * tape, the others read disk blocks; they pass control of the tape in
85  * a ring via signals. The parent process traverses the filesystem and
86  * sends writeheader()'s and lists of daddr's to the slaves via pipes.
87  * The following structure defines the instruction packets sent to slaves.
88  */
89 struct req {
90 	daddr_t dblk;
91 	int count;
92 };
93 static int reqsiz;
94 
95 #define SLAVES 3		/* 1 slave writing, 1 reading, 1 for slack */
96 static struct slave {
97 	int64_t tapea;		/* header number at start of this chunk */
98 	int64_t firstrec;	/* record number of this block */
99 	int count;		/* count to next header (used for TS_TAPE */
100 				/* after EOT) */
101 	int inode;		/* inode that we are currently dealing with */
102 	int fd;			/* FD for this slave */
103 	pid_t pid;		/* PID for this slave */
104 	int sent;		/* 1 == we've sent this slave requests */
105 	char (*tblock)[TP_BSIZE]; /* buffer for data blocks */
106 	struct req *req;	/* buffer for requests */
107 } slaves[SLAVES+1];
108 static struct slave *slp;
109 
110 static char	(*nextblock)[TP_BSIZE];
111 
112 static time_t tstart_volume;	/* time of volume start */
113 static int64_t tapea_volume;	/* value of spcl.c_tapea at volume start */
114 
115 static pid_t master;		/* pid of master, for sending error signals */
116 static int tenths;		/* length of tape used per block written */
117 static volatile sig_atomic_t caught;	/* have we caught the signal to proceed? */
118 
119 int
alloctape(void)120 alloctape(void)
121 {
122 	int pgoff = getpagesize() - 1;
123 	char *buf;
124 	int i;
125 
126 	writesize = ntrec * TP_BSIZE;
127 	reqsiz = (ntrec + 1) * sizeof(struct req);
128 	/*
129 	 * CDC 92181's and 92185's make 0.8" gaps in 1600-bpi start/stop mode
130 	 * (see DEC TU80 User's Guide).  The shorter gaps of 6250-bpi require
131 	 * repositioning after stopping, i.e, streaming mode, where the gap is
132 	 * variable, 0.30" to 0.45".  The gap is maximal when the tape stops.
133 	 */
134 	if (blocksperfile == 0 && !unlimited)
135 		tenths = writesize / density +
136 		    (cartridge ? 16 : density == 625 ? 5 : 8);
137 	/*
138 	 * Allocate tape buffer contiguous with the array of instruction
139 	 * packets, so flushtape() can write them together with one write().
140 	 * Align tape buffer on page boundary to speed up tape write().
141 	 */
142 	for (i = 0; i <= SLAVES; i++) {
143 		buf = malloc((unsigned)(reqsiz + writesize + pgoff + TP_BSIZE));
144 		if (buf == NULL)
145 			return(0);
146 		slaves[i].tblock = (char (*)[TP_BSIZE])
147 		    (((long)&buf[ntrec + 1] + pgoff) &~ pgoff);
148 		slaves[i].req = (struct req *)slaves[i].tblock - ntrec - 1;
149 	}
150 	slp = &slaves[0];
151 	slp->count = 1;
152 	slp->tapea = 0;
153 	slp->firstrec = 0;
154 	nextblock = slp->tblock;
155 	return(1);
156 }
157 
158 void
writerec(char * dp,int isspcl)159 writerec(char *dp, int isspcl)
160 {
161 
162 	slp->req[trecno].dblk = 0;
163 	slp->req[trecno].count = 1;
164 	*(union u_spcl *)(*(nextblock)++) = *(union u_spcl *)dp;
165 	if (isspcl)
166 		lastspclrec = spcl.c_tapea;
167 	trecno++;
168 	spcl.c_tapea++;
169 	if (trecno >= ntrec)
170 		flushtape();
171 }
172 
173 void
dumpblock(daddr_t blkno,int size)174 dumpblock(daddr_t blkno, int size)
175 {
176 	int avail, tpblks;
177 	daddr_t dblkno;
178 
179 	dblkno = fsbtodb(sblock, blkno);
180 	tpblks = size >> tp_bshift;
181 	while ((avail = MINIMUM(tpblks, ntrec - trecno)) > 0) {
182 		slp->req[trecno].dblk = dblkno;
183 		slp->req[trecno].count = avail;
184 		trecno += avail;
185 		spcl.c_tapea += avail;
186 		if (trecno >= ntrec)
187 			flushtape();
188 		dblkno += avail << (tp_bshift - (ffs(DEV_BSIZE) - 1));
189 		tpblks -= avail;
190 	}
191 }
192 
193 int	nogripe = 0;
194 
195 void
tperror(int signo)196 tperror(int signo)
197 {
198 	/* XXX - signal races */
199 
200 	if (pipeout) {
201 		msg("write error on %s\n", tape);
202 		quit("Cannot recover\n");
203 		/* NOTREACHED */
204 	}
205 	msg("write error %lld blocks into volume %d\n",
206 	    (long long)blocksthisvol, tapeno);
207 	broadcast("DUMP WRITE ERROR!\n");
208 	if (!query("Do you want to restart?"))
209 		dumpabort(0);
210 	msg("Closing this volume.  Prepare to restart with new media;\n");
211 	msg("this dump volume will be rewritten.\n");
212 	killall();
213 	nogripe = 1;
214 	close_rewind();
215 	Exit(X_REWRITE);
216 }
217 
218 void
sigpipe(int signo)219 sigpipe(int signo)
220 {
221 
222 	quit("Broken pipe\n");
223 }
224 
225 /*
226  * do_stats --
227  *	Update xferrate stats
228  */
229 time_t
do_stats(void)230 do_stats(void)
231 {
232 	time_t tnow, ttaken;
233 	int64_t blocks;
234 	char *ct;
235 
236 	(void)time(&tnow);
237 	ttaken = tnow - tstart_volume;
238 	blocks = spcl.c_tapea - tapea_volume;
239 	ct = ctime(&tnow);
240 	msg("Volume %d completed at: %s", tapeno, ct ? ct : "?\n");
241 	if (ttaken > 0) {
242 		msg("Volume %d took %lld:%02lld:%02lld\n", tapeno,
243 		    (long long)ttaken / 3600, ((long long)ttaken % 3600) / 60,
244 		    (long long)ttaken % 60);
245 		blocks /= ttaken;
246 		msg("Volume %d transfer rate: %lld KB/s\n", tapeno, blocks);
247 		xferrate += blocks;
248 	}
249 	return(tnow);
250 }
251 
252 /*
253  * statussig --
254  *	information message upon receipt of SIGINFO
255  *	(derived from optr.c::timeest())
256  * XXX not safe
257  */
258 void
statussig(int signo)259 statussig(int signo)
260 {
261 	time_t	tnow, deltat;
262 	int save_errno = errno;
263 
264 	if (blockswritten < 500)
265 		return;
266 	(void) time(&tnow);
267 	deltat = tstart_writing - tnow + (1.0 * (tnow - tstart_writing))
268 		/ blockswritten * tapesize;
269 	/* XXX not safe due to floating point printf */
270 	dprintf(STDERR_FILENO,
271 	    "dump: %s %3.2f%% done at %lld KB/s, finished in %d:%02d\n",
272 	    tape, (blockswritten * 100.0) / tapesize,
273 	    (spcl.c_tapea - tapea_volume) / (tnow - tstart_volume),
274 	    (int)(deltat / 3600), (int)((deltat % 3600) / 60));
275 	errno = save_errno;
276 }
277 
278 static void
flushtape(void)279 flushtape(void)
280 {
281 	int i, blks, got;
282 	int64_t lastfirstrec;
283 
284 	int siz = (char *)nextblock - (char *)slp->req;
285 
286 	slp->req[trecno].count = 0;			/* Sentinel */
287 
288 	if (atomic((ssize_t (*)(int, void *, size_t))write, slp->fd,
289 	    (char *)slp->req, siz) != siz)
290 		quit("error writing command pipe: %s\n", strerror(errno));
291 	slp->sent = 1; /* we sent a request, read the response later */
292 
293 	lastfirstrec = slp->firstrec;
294 
295 	if (++slp >= &slaves[SLAVES])
296 		slp = &slaves[0];
297 
298 	/* Read results back from next slave */
299 	if (slp->sent) {
300 		if (atomic(read, slp->fd, (char *)&got, sizeof(got))
301 		    != sizeof(got)) {
302 			perror("  DUMP: error reading command pipe in master");
303 			dumpabort(0);
304 		}
305 		slp->sent = 0;
306 
307 		/* Check for end of tape */
308 		if (got < writesize) {
309 			msg("End of tape detected\n");
310 
311 			/*
312 			 * Drain the results, don't care what the values were.
313 			 * If we read them here then trewind won't...
314 			 */
315 			for (i = 0; i < SLAVES; i++) {
316 				if (slaves[i].sent) {
317 					if (atomic(read, slaves[i].fd,
318 					    (char *)&got, sizeof(got))
319 					    != sizeof(got)) {
320 						perror("  DUMP: error reading command pipe in master");
321 						dumpabort(0);
322 					}
323 					slaves[i].sent = 0;
324 				}
325 			}
326 
327 			close_rewind();
328 			rollforward();
329 			return;
330 		}
331 	}
332 
333 	blks = 0;
334 	if (spcl.c_type != TS_END && spcl.c_type != TS_CLRI &&
335 	    spcl.c_type != TS_BITS) {
336 		if (spcl.c_count > TP_NINDIR)
337 			quit("c_count too large\n");
338 		for (i = 0; i < spcl.c_count; i++)
339 			if (spcl.c_addr[i] != 0)
340 				blks++;
341 	}
342 	slp->count = lastspclrec + blks + 1 - spcl.c_tapea;
343 	slp->tapea = spcl.c_tapea;
344 	slp->firstrec = lastfirstrec + ntrec;
345 	slp->inode = curino;
346 	nextblock = slp->tblock;
347 	trecno = 0;
348 	asize += tenths;
349 	blockswritten += ntrec;
350 	blocksthisvol += ntrec;
351 	if (!pipeout && !unlimited && (blocksperfile ?
352 	    (blocksthisvol >= blocksperfile) : (asize > tsize))) {
353 		close_rewind();
354 		startnewtape(0);
355 	}
356 	timeest();
357 }
358 
359 void
trewind(void)360 trewind(void)
361 {
362 	struct stat sb;
363 	int f, got;
364 
365 	for (f = 0; f < SLAVES; f++) {
366 		/*
367 		 * Drain the results, but unlike EOT we DO (or should) care
368 		 * what the return values were, since if we detect EOT after
369 		 * we think we've written the last blocks to the tape anyway,
370 		 * we have to replay those blocks with rollforward.
371 		 *
372 		 * fixme: punt for now.
373 		 */
374 		if (slaves[f].sent) {
375 			if (atomic(read, slaves[f].fd, (char *)&got, sizeof(got))
376 			    != sizeof(got)) {
377 				perror("  DUMP: error reading command pipe in master");
378 				dumpabort(0);
379 			}
380 			slaves[f].sent = 0;
381 			if (got != writesize) {
382 				msg("EOT detected in last 2 tape records!\n");
383 				msg("Use a longer tape, decrease the size estimate\n");
384 				quit("or use no size estimate at all.\n");
385 			}
386 		}
387 		(void) close(slaves[f].fd);
388 	}
389 	while (wait((int *)NULL) >= 0)	/* wait for any signals from slaves */
390 		/* void */;
391 
392 	if (pipeout)
393 		return;
394 
395 	msg("Closing %s\n", tape);
396 
397 #ifdef RDUMP
398 	if (host) {
399 		rmtclose();
400 		while (rmtopen(tape, O_RDONLY) < 0)
401 			sleep(10);
402 		rmtclose();
403 		return;
404 	}
405 #endif
406 	/*
407 	 * st(4) says: "Bit 1 of the minor number specifies whether an eject is
408 	 * attempted when the device is closed.  When it is set, the device
409 	 * will attempt to eject its media on close ...".
410 	 *
411 	 * If the tape has been ejected, looping on open() will generate 'Media
412 	 * not present' errors until a tape is loaded. Once loaded the tape
413 	 * will be immediately ejected as a result of the second close().
414 	 *
415 	 * So if the tape will be ejected, just close and return.
416 	 */
417 	if ((fstat(tapefd, &sb) == 0) && (minor(sb.st_rdev) & 0x02)) {
418 		(void) close(tapefd);
419 		return;
420 	}
421 
422 	(void) close(tapefd);
423 	while ((f = open(tape, O_RDONLY)) == -1)
424 		sleep (10);
425 	(void) close(f);
426 }
427 
428 void
close_rewind(void)429 close_rewind(void)
430 {
431 	trewind();
432 	(void)do_stats();
433 	if (nexttape)
434 		return;
435 	if (!nogripe) {
436 		msg("Change Volumes: Mount volume #%d\n", tapeno+1);
437 		broadcast("CHANGE DUMP VOLUMES!\7\7\n");
438 	}
439 	while (!query("Is the new volume mounted and ready to go?"))
440 		if (query("Do you want to abort?")) {
441 			dumpabort(0);
442 			/*NOTREACHED*/
443 		}
444 }
445 
446 void
rollforward(void)447 rollforward(void)
448 {
449 	struct req *p, *q, *prev;
450 	struct slave *tslp;
451 	int i, size, got;
452 	int64_t savedtapea;
453 	union u_spcl *ntb, *otb;
454 	tslp = &slaves[SLAVES];
455 	ntb = (union u_spcl *)tslp->tblock[1];
456 
457 	/*
458 	 * Each of the N slaves should have requests that need to
459 	 * be replayed on the next tape.  Use the extra slave buffers
460 	 * (slaves[SLAVES]) to construct request lists to be sent to
461 	 * each slave in turn.
462 	 */
463 	for (i = 0; i < SLAVES; i++) {
464 		q = &tslp->req[1];
465 		otb = (union u_spcl *)slp->tblock;
466 
467 		/*
468 		 * For each request in the current slave, copy it to tslp.
469 		 */
470 
471 		prev = NULL;
472 		for (p = slp->req; p->count > 0; p += p->count) {
473 			*q = *p;
474 			if (p->dblk == 0)
475 				*ntb++ = *otb++; /* copy the datablock also */
476 			prev = q;
477 			q += q->count;
478 		}
479 		if (prev == NULL)
480 			quit("rollforward: protocol botch\n");
481 		if (prev->dblk != 0)
482 			prev->count -= 1;
483 		else
484 			ntb--;
485 		q -= 1;
486 		q->count = 0;
487 		q = &tslp->req[0];
488 		if (i == 0) {
489 			q->dblk = 0;
490 			q->count = 1;
491 			trecno = 0;
492 			nextblock = tslp->tblock;
493 			savedtapea = spcl.c_tapea;
494 			spcl.c_tapea = slp->tapea;
495 			startnewtape(0);
496 			spcl.c_tapea = savedtapea;
497 			lastspclrec = savedtapea - 1;
498 		}
499 		size = (char *)ntb - (char *)q;
500 		if (atomic((ssize_t (*)(int, void *, size_t))write,
501 		    slp->fd, (char *)q, size) != size) {
502 			perror("  DUMP: error writing command pipe");
503 			dumpabort(0);
504 		}
505 		slp->sent = 1;
506 		if (++slp >= &slaves[SLAVES])
507 			slp = &slaves[0];
508 
509 		q->count = 1;
510 
511 		if (prev->dblk != 0) {
512 			/*
513 			 * If the last one was a disk block, make the
514 			 * first of this one be the last bit of that disk
515 			 * block...
516 			 */
517 			q->dblk = prev->dblk +
518 				prev->count * (TP_BSIZE / DEV_BSIZE);
519 			ntb = (union u_spcl *)tslp->tblock;
520 		} else {
521 			/*
522 			 * It wasn't a disk block.  Copy the data to its
523 			 * new location in the buffer.
524 			 */
525 			q->dblk = 0;
526 			*((union u_spcl *)tslp->tblock) = *ntb;
527 			ntb = (union u_spcl *)tslp->tblock[1];
528 		}
529 	}
530 	slp->req[0] = *q;
531 	nextblock = slp->tblock;
532 	if (q->dblk == 0)
533 		nextblock++;
534 	trecno = 1;
535 
536 	/*
537 	 * Clear the first slaves' response.  One hopes that it
538 	 * worked ok, otherwise the tape is much too short!
539 	 */
540 	if (slp->sent) {
541 		if (atomic(read, slp->fd, (char *)&got, sizeof(got))
542 		    != sizeof(got)) {
543 			perror("  DUMP: error reading command pipe in master");
544 			dumpabort(0);
545 		}
546 		slp->sent = 0;
547 
548 		if (got != writesize) {
549 			quit("EOT detected at start of the tape!\n");
550 		}
551 	}
552 }
553 
554 /*
555  * We implement taking and restoring checkpoints on the tape level.
556  * When each tape is opened, a new process is created by forking; this
557  * saves all of the necessary context in the parent.  The child
558  * continues the dump; the parent waits around, saving the context.
559  * If the child returns X_REWRITE, then it had problems writing that tape;
560  * this causes the parent to fork again, duplicating the context, and
561  * everything continues as if nothing had happened.
562  */
563 void
startnewtape(int top)564 startnewtape(int top)
565 {
566 	pid_t	parentpid;
567 	pid_t	childpid;
568 	int	status;
569 	pid_t	waitingpid;
570 	char	*p, *ct;
571 	sig_t	interrupt_save;
572 
573 	interrupt_save = signal(SIGINT, SIG_IGN);
574 	parentpid = getpid();
575 	tapea_volume = spcl.c_tapea;
576 	(void)time(&tstart_volume);
577 
578 restore_check_point:
579 	(void)signal(SIGINT, interrupt_save);
580 	/*
581 	 *	All signals are inherited...
582 	 */
583 	childpid = fork();
584 	if (childpid == -1) {
585 		msg("Context save fork fails in parent %d\n", parentpid);
586 		Exit(X_ABORT);
587 	}
588 	if (childpid != 0) {
589 		/*
590 		 *	PARENT:
591 		 *	save the context by waiting
592 		 *	until the child doing all of the work returns.
593 		 *	don't catch the interrupt
594 		 */
595 		signal(SIGINT, SIG_IGN);
596 #ifdef TDEBUG
597 		msg("Tape: %d; parent process: %d child process %d\n",
598 			tapeno+1, parentpid, childpid);
599 #endif /* TDEBUG */
600 		while ((waitingpid = wait(&status)) != childpid)
601 			msg("Parent %d waiting for child %d has another child %d return\n",
602 				parentpid, childpid, waitingpid);
603 		if (status & 0xFF) {
604 			msg("Child %d returns LOB status %o\n",
605 				childpid, status&0xFF);
606 		}
607 		status = (status >> 8) & 0xFF;
608 #ifdef TDEBUG
609 		switch(status) {
610 			case X_FINOK:
611 				msg("Child %d finishes X_FINOK\n", childpid);
612 				break;
613 			case X_ABORT:
614 				msg("Child %d finishes X_ABORT\n", childpid);
615 				break;
616 			case X_REWRITE:
617 				msg("Child %d finishes X_REWRITE\n", childpid);
618 				break;
619 			default:
620 				msg("Child %d finishes unknown %d\n",
621 					childpid, status);
622 				break;
623 		}
624 #endif /* TDEBUG */
625 		switch(status) {
626 			case X_FINOK:
627 				Exit(X_FINOK);
628 				break;
629 			case X_ABORT:
630 				Exit(X_ABORT);
631 				break;
632 			case X_REWRITE:
633 				goto restore_check_point;
634 			default:
635 				msg("Bad return code from dump: %d\n", status);
636 				Exit(X_ABORT);
637 		}
638 		/*NOTREACHED*/
639 	} else {	/* we are the child; just continue */
640 #ifdef TDEBUG
641 		sleep(4);	/* allow time for parent's message to get out */
642 		msg("Child on Tape %d has parent %d, my pid = %d\n",
643 			tapeno+1, parentpid, getpid());
644 #endif /* TDEBUG */
645 		/*
646 		 * If we have a name like "/dev/rst0,/dev/rst1",
647 		 * use the name before the comma first, and save
648 		 * the remaining names for subsequent volumes.
649 		 */
650 		tapeno++;               /* current tape sequence */
651 		if (nexttape || strchr(tape, ',')) {
652 			if (nexttape && *nexttape)
653 				tape = nexttape;
654 			if ((p = strchr(tape, ',')) != NULL) {
655 				*p = '\0';
656 				nexttape = p + 1;
657 			} else
658 				nexttape = NULL;
659 			msg("Dumping volume %d on %s\n", tapeno, tape);
660 		}
661 #ifdef RDUMP
662 		while ((tapefd = (host ? rmtopen(tape, O_WRONLY|O_CREAT) :
663 			pipeout ? 1 : open(tape, O_WRONLY|O_CREAT, 0666))) == -1)
664 #else
665 		while ((tapefd = (pipeout ? 1 :
666 				  open(tape, O_WRONLY|O_CREAT, 0666))) == -1)
667 #endif
668 		    {
669 			msg("Cannot open output \"%s\".\n", tape);
670 			if (!query("Do you want to retry the open?"))
671 				dumpabort(0);
672 		}
673 
674 		enslave();  /* Share open tape file descriptor with slaves */
675 
676 		asize = 0;
677 		blocksthisvol = 0;
678 		if (top)
679 			newtape++;		/* new tape signal */
680 		spcl.c_count = slp->count;
681 		/*
682 		 * measure firstrec in TP_BSIZE units since restore doesn't
683 		 * know the correct ntrec value...
684 		 */
685 		spcl.c_firstrec = slp->firstrec;
686 		spcl.c_volume++;
687 		spcl.c_type = TS_TAPE;
688 		if (sblock->fs_magic != FS_UFS2_MAGIC)
689 			spcl.c_flags |= DR_NEWHEADER;
690 		writeheader((ino_t)slp->inode);
691 		if (sblock->fs_magic != FS_UFS2_MAGIC)
692 			spcl.c_flags &=~ DR_NEWHEADER;
693 		ct = ctime(&tstart_volume);
694 		msg("Volume %d started at: %s", tapeno, ct ? ct : "?\n");
695 		if (tapeno > 1)
696 			msg("Volume %d begins with blocks from inode %llu\n",
697 			    tapeno, (unsigned long long)slp->inode);
698 	}
699 }
700 
701 void
dumpabort(int signo)702 dumpabort(int signo)
703 {
704 
705 	if (master != 0 && master != getpid())
706 		/* Signals master to call dumpabort */
707 		(void) kill(master, SIGTERM);
708 	else {
709 		killall();
710 		msg("The ENTIRE dump is aborted.\n");
711 	}
712 #ifdef RDUMP
713 	rmtclose();
714 #endif
715 	Exit(X_ABORT);
716 }
717 
718 __dead void
Exit(int status)719 Exit(int status)
720 {
721 
722 #ifdef TDEBUG
723 	msg("pid = %d exits with status %d\n", getpid(), status);
724 #endif /* TDEBUG */
725 	exit(status);
726 }
727 
728 /*
729  * proceed - handler for SIGUSR2, used to synchronize IO between the slaves.
730  */
731 void
proceed(int signo)732 proceed(int signo)
733 {
734 	caught++;
735 }
736 
737 void
enslave(void)738 enslave(void)
739 {
740 	int cmd[2];
741 	int i, j;
742 
743 	master = getpid();
744 
745 	signal(SIGTERM, dumpabort);  /* Slave sends SIGTERM on dumpabort() */
746 	signal(SIGPIPE, sigpipe);
747 	signal(SIGUSR1, tperror);    /* Slave sends SIGUSR1 on tape errors */
748 	signal(SIGUSR2, proceed);    /* Slave sends SIGUSR2 to next slave */
749 
750 	for (i = 0; i < SLAVES; i++) {
751 		if (i == slp - &slaves[0]) {
752 			caught = 1;
753 		} else {
754 			caught = 0;
755 		}
756 
757 		if (socketpair(AF_UNIX, SOCK_STREAM, 0, cmd) == -1 ||
758 		    (slaves[i].pid = fork()) == -1)
759 			quit("too many slaves, %d (recompile smaller): %s\n",
760 			    i, strerror(errno));
761 
762 		slaves[i].fd = cmd[1];
763 		slaves[i].sent = 0;
764 		if (slaves[i].pid == 0) { 	    /* Slave starts up here */
765 			for (j = 0; j <= i; j++)
766 			        (void) close(slaves[j].fd);
767 			signal(SIGINT, SIG_IGN);    /* Master handles this */
768 			signal(SIGINFO, SIG_IGN);
769 			doslave(cmd[0], i);
770 			Exit(X_FINOK);
771 		}
772 	}
773 
774 	for (i = 0; i < SLAVES; i++)
775 		(void) atomic((ssize_t (*)(int, void *, size_t))write,
776 		    slaves[i].fd, (char *) &slaves[(i + 1) % SLAVES].pid,
777 		    sizeof(slaves[0].pid));
778 	master = 0;
779 }
780 
781 void
killall(void)782 killall(void)
783 {
784 	int i;
785 
786 	for (i = 0; i < SLAVES; i++)
787 		if (slaves[i].pid > 0) {
788 			(void) kill(slaves[i].pid, SIGKILL);
789 			slaves[i].pid = 0;
790 		}
791 }
792 
793 /*
794  * Synchronization - each process has a lockfile, and shares file
795  * descriptors to the following process's lockfile.  When our write
796  * completes, we release our lock on the following process's lock-
797  * file, allowing the following process to lock it and proceed. We
798  * get the lock back for the next cycle by swapping descriptors.
799  */
800 static void
doslave(int cmd,int slave_number)801 doslave(int cmd, int slave_number)
802 {
803 	int nread, nextslave, size, wrote = 0, eot_count;
804 	sigset_t nsigset, osigset;
805 
806 	/*
807 	 * Need our own seek pointer.
808 	 */
809 	(void) close(diskfd);
810 	if ((diskfd = open(disk, O_RDONLY)) == -1)
811 		quit("slave couldn't reopen disk: %s\n", strerror(errno));
812 
813 	/*
814 	 * Need the pid of the next slave in the loop...
815 	 */
816 	if ((nread = atomic(read, cmd, (char *)&nextslave, sizeof(nextslave)))
817 	    != sizeof(nextslave)) {
818 		quit("master/slave protocol botched - didn't get pid of next slave.\n");
819 	}
820 
821 	/*
822 	 * Get list of blocks to dump, read the blocks into tape buffer
823 	 */
824 	while ((nread = atomic(read, cmd, (char *)slp->req, reqsiz)) == reqsiz) {
825 		struct req *p = slp->req;
826 
827 		for (trecno = 0; trecno < ntrec;
828 		     trecno += p->count, p += p->count) {
829 			if (p->dblk) {
830 				bread(p->dblk, slp->tblock[trecno],
831 					p->count * TP_BSIZE);
832 			} else {
833 				if (p->count != 1 || atomic(read, cmd,
834 				    (char *)slp->tblock[trecno],
835 				    TP_BSIZE) != TP_BSIZE)
836 				       quit("master/slave protocol botched.\n");
837 			}
838 		}
839 
840 		sigemptyset(&nsigset);
841 		sigaddset(&nsigset, SIGUSR2);
842 		sigprocmask(SIG_BLOCK, &nsigset, &osigset);
843 		while (!caught)
844 			sigsuspend(&osigset);
845 		caught = 0;
846 		sigprocmask(SIG_SETMASK, &osigset, NULL);
847 
848 		/* Try to write the data... */
849 		eot_count = 0;
850 		size = 0;
851 
852 		while (eot_count < 10 && size < writesize) {
853 #ifdef RDUMP
854 			if (host)
855 				wrote = rmtwrite(slp->tblock[0]+size,
856 				    writesize-size);
857 			else
858 #endif
859 				wrote = write(tapefd, slp->tblock[0]+size,
860 				    writesize-size);
861 #ifdef WRITEDEBUG
862 			printf("slave %d wrote %d\n", slave_number, wrote);
863 #endif
864 			if (wrote < 0)
865 				break;
866 			if (wrote == 0)
867 				eot_count++;
868 			size += wrote;
869 		}
870 
871 #ifdef WRITEDEBUG
872 		if (size != writesize)
873 		 printf("slave %d only wrote %d out of %d bytes and gave up.\n",
874 		     slave_number, size, writesize);
875 #endif
876 
877 		if (eot_count > 0)
878 			size = 0;
879 
880 		/*
881 		 * Handle ENOSPC as an EOT condition
882 		 */
883 		if (wrote < 0 && errno == ENOSPC) {
884 			wrote = 0;
885 			eot_count++;
886 		}
887 
888 		if (size < 0) {
889 			(void) kill(master, SIGUSR1);
890 			sigemptyset(&nsigset);
891 			for (;;)
892 				sigsuspend(&nsigset);
893 		} else {
894 			/*
895 			 * pass size of write back to master
896 			 * (for EOT handling)
897 			 */
898 			(void) atomic((ssize_t (*)(int, void *, size_t))write,
899 			    cmd, (char *)&size, sizeof(size));
900 		}
901 
902 		/*
903 		 * If partial write, don't want next slave to go.
904 		 * Also jolts him awake.
905 		 */
906 		(void) kill(nextslave, SIGUSR2);
907 	}
908 	if (nread != 0)
909 		quit("error reading command pipe: %s\n", strerror(errno));
910 }
911 
912 /*
913  * Since a read from a pipe may not return all we asked for,
914  * or a write may not write all we ask if we get a signal,
915  * loop until the count is satisfied (or error).
916  */
917 static ssize_t
atomic(ssize_t (* func)(int,void *,size_t),int fd,char * buf,int count)918 atomic(ssize_t (*func)(int, void *, size_t), int fd, char *buf, int count)
919 {
920 	ssize_t got, need = count;
921 
922 	while ((got = (*func)(fd, buf, need)) > 0 && (need -= got) > 0)
923 		buf += got;
924 	return (got < 0 ? got : count - need);
925 }
926