xref: /freebsd/sbin/dump/tape.c (revision 6829dae12bb055451fa467da4589c43bd03b1e64)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1980, 1991, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. Neither the name of the University nor the names of its contributors
16  *    may be used to endorse or promote products derived from this software
17  *    without specific prior written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  */
31 
32 #ifndef lint
33 #if 0
34 static char sccsid[] = "@(#)tape.c	8.4 (Berkeley) 5/1/95";
35 #endif
36 static const char rcsid[] =
37   "$FreeBSD$";
38 #endif /* not lint */
39 
40 #include <sys/param.h>
41 #include <sys/socket.h>
42 #include <sys/wait.h>
43 #include <sys/stat.h>
44 
45 #include <ufs/ufs/dinode.h>
46 #include <ufs/ffs/fs.h>
47 
48 #include <protocols/dumprestore.h>
49 
50 #include <assert.h>
51 #include <errno.h>
52 #include <fcntl.h>
53 #include <limits.h>
54 #include <setjmp.h>
55 #include <signal.h>
56 #include <stdio.h>
57 #include <stdlib.h>
58 #include <string.h>
59 #include <time.h>
60 #include <unistd.h>
61 
62 #include "dump.h"
63 
64 int	writesize;		/* size of malloc()ed buffer for tape */
65 int64_t	lastspclrec = -1;	/* tape block number of last written header */
66 int	trecno = 0;		/* next record to write in current block */
67 extern	long blocksperfile;	/* number of blocks per output file */
68 long	blocksthisvol;		/* number of blocks on current output file */
69 extern	int ntrec;		/* blocking factor on tape */
70 extern	int cartridge;
71 extern	char *host;
72 char	*nexttape;
73 FILE	*popenfp = NULL;
74 
75 static	int atomic(ssize_t (*)(), int, char *, int);
76 static	void doslave(int, int);
77 static	void enslave(void);
78 static	void flushtape(void);
79 static	void killall(void);
80 static	void rollforward(void);
81 
82 /*
83  * Concurrent dump mods (Caltech) - disk block reading and tape writing
84  * are exported to several slave processes.  While one slave writes the
85  * tape, the others read disk blocks; they pass control of the tape in
86  * a ring via signals. The parent process traverses the file system and
87  * sends writeheader()'s and lists of daddr's to the slaves via pipes.
88  * The following structure defines the instruction packets sent to slaves.
89  */
90 struct req {
91 	ufs2_daddr_t dblk;
92 	int count;
93 };
94 int reqsiz;
95 
96 #define SLAVES 3		/* 1 slave writing, 1 reading, 1 for slack */
97 struct slave {
98 	int64_t tapea;		/* header number at start of this chunk */
99 	int64_t firstrec;	/* record number of this block */
100 	int count;		/* count to next header (used for TS_TAPE */
101 				/* after EOT) */
102 	int inode;		/* inode that we are currently dealing with */
103 	int fd;			/* FD for this slave */
104 	int pid;		/* PID for this slave */
105 	int sent;		/* 1 == we've sent this slave requests */
106 	char (*tblock)[TP_BSIZE]; /* buffer for data blocks */
107 	struct req *req;	/* buffer for requests */
108 } slaves[SLAVES+1];
109 struct slave *slp;
110 
111 char	(*nextblock)[TP_BSIZE];
112 
113 int master;		/* pid of master, for sending error signals */
114 int tenths;		/* length of tape used per block written */
115 static volatile sig_atomic_t caught; /* have we caught the signal to proceed? */
116 static volatile sig_atomic_t ready; /* reached the lock point without having */
117 			/* received the SIGUSR2 signal from the prev slave? */
118 static jmp_buf jmpbuf;	/* where to jump to if we are ready when the */
119 			/* SIGUSR2 arrives from the previous slave */
120 
121 int
122 alloctape(void)
123 {
124 	int pgoff = getpagesize() - 1;
125 	char *buf;
126 	int i;
127 
128 	writesize = ntrec * TP_BSIZE;
129 	reqsiz = (ntrec + 1) * sizeof(struct req);
130 	/*
131 	 * CDC 92181's and 92185's make 0.8" gaps in 1600-bpi start/stop mode
132 	 * (see DEC TU80 User's Guide).  The shorter gaps of 6250-bpi require
133 	 * repositioning after stopping, i.e, streaming mode, where the gap is
134 	 * variable, 0.30" to 0.45".  The gap is maximal when the tape stops.
135 	 */
136 	if (blocksperfile == 0 && !unlimited)
137 		tenths = writesize / density +
138 		    (cartridge ? 16 : density == 625 ? 5 : 8);
139 	/*
140 	 * Allocate tape buffer contiguous with the array of instruction
141 	 * packets, so flushtape() can write them together with one write().
142 	 * Align tape buffer on page boundary to speed up tape write().
143 	 */
144 	for (i = 0; i <= SLAVES; i++) {
145 		buf = (char *)
146 		    malloc((unsigned)(reqsiz + writesize + pgoff + TP_BSIZE));
147 		if (buf == NULL)
148 			return(0);
149 		slaves[i].tblock = (char (*)[TP_BSIZE])
150 		    (((long)&buf[ntrec + 1] + pgoff) &~ pgoff);
151 		slaves[i].req = (struct req *)slaves[i].tblock - ntrec - 1;
152 	}
153 	slp = &slaves[0];
154 	slp->count = 1;
155 	slp->tapea = 0;
156 	slp->firstrec = 0;
157 	nextblock = slp->tblock;
158 	return(1);
159 }
160 
161 void
162 writerec(char *dp, int isspcl)
163 {
164 
165 	slp->req[trecno].dblk = (ufs2_daddr_t)0;
166 	slp->req[trecno].count = 1;
167 	/* Can't do a structure assignment due to alignment problems */
168 	bcopy(dp, *(nextblock)++, sizeof (union u_spcl));
169 	if (isspcl)
170 		lastspclrec = spcl.c_tapea;
171 	trecno++;
172 	spcl.c_tapea++;
173 	if (trecno >= ntrec)
174 		flushtape();
175 }
176 
177 void
178 dumpblock(ufs2_daddr_t blkno, int size)
179 {
180 	int avail, tpblks;
181 	ufs2_daddr_t dblkno;
182 
183 	dblkno = fsbtodb(sblock, blkno);
184 	tpblks = size >> tp_bshift;
185 	while ((avail = MIN(tpblks, ntrec - trecno)) > 0) {
186 		slp->req[trecno].dblk = dblkno;
187 		slp->req[trecno].count = avail;
188 		trecno += avail;
189 		spcl.c_tapea += avail;
190 		if (trecno >= ntrec)
191 			flushtape();
192 		dblkno += avail << (tp_bshift - dev_bshift);
193 		tpblks -= avail;
194 	}
195 }
196 
197 int	nogripe = 0;
198 
199 void
200 tperror(int signo __unused)
201 {
202 
203 	if (pipeout) {
204 		msg("write error on %s\n", tape);
205 		quit("Cannot recover\n");
206 		/* NOTREACHED */
207 	}
208 	msg("write error %ld blocks into volume %d\n", blocksthisvol, tapeno);
209 	broadcast("DUMP WRITE ERROR!\n");
210 	if (!query("Do you want to restart?"))
211 		dumpabort(0);
212 	msg("Closing this volume.  Prepare to restart with new media;\n");
213 	msg("this dump volume will be rewritten.\n");
214 	killall();
215 	nogripe = 1;
216 	close_rewind();
217 	Exit(X_REWRITE);
218 }
219 
220 void
221 sigpipe(int signo __unused)
222 {
223 
224 	quit("Broken pipe\n");
225 }
226 
227 static void
228 flushtape(void)
229 {
230 	int i, blks, got;
231 	int64_t lastfirstrec;
232 
233 	int siz = (char *)nextblock - (char *)slp->req;
234 
235 	slp->req[trecno].count = 0;			/* Sentinel */
236 
237 	if (atomic(write, slp->fd, (char *)slp->req, siz) != siz)
238 		quit("error writing command pipe: %s\n", strerror(errno));
239 	slp->sent = 1; /* we sent a request, read the response later */
240 
241 	lastfirstrec = slp->firstrec;
242 
243 	if (++slp >= &slaves[SLAVES])
244 		slp = &slaves[0];
245 
246 	/* Read results back from next slave */
247 	if (slp->sent) {
248 		if (atomic(read, slp->fd, (char *)&got, sizeof got)
249 		    != sizeof got) {
250 			perror("  DUMP: error reading command pipe in master");
251 			dumpabort(0);
252 		}
253 		slp->sent = 0;
254 
255 		/* Check for end of tape */
256 		if (got < writesize) {
257 			msg("End of tape detected\n");
258 
259 			/*
260 			 * Drain the results, don't care what the values were.
261 			 * If we read them here then trewind won't...
262 			 */
263 			for (i = 0; i < SLAVES; i++) {
264 				if (slaves[i].sent) {
265 					if (atomic(read, slaves[i].fd,
266 					    (char *)&got, sizeof got)
267 					    != sizeof got) {
268 						perror("  DUMP: error reading command pipe in master");
269 						dumpabort(0);
270 					}
271 					slaves[i].sent = 0;
272 				}
273 			}
274 
275 			close_rewind();
276 			rollforward();
277 			return;
278 		}
279 	}
280 
281 	blks = 0;
282 	if (spcl.c_type != TS_END && spcl.c_type != TS_CLRI &&
283 	    spcl.c_type != TS_BITS) {
284 		assert(spcl.c_count <= TP_NINDIR);
285 		for (i = 0; i < spcl.c_count; i++)
286 			if (spcl.c_addr[i] != 0)
287 				blks++;
288 	}
289 	slp->count = lastspclrec + blks + 1 - spcl.c_tapea;
290 	slp->tapea = spcl.c_tapea;
291 	slp->firstrec = lastfirstrec + ntrec;
292 	slp->inode = curino;
293 	nextblock = slp->tblock;
294 	trecno = 0;
295 	asize += tenths;
296 	blockswritten += ntrec;
297 	blocksthisvol += ntrec;
298 	if (!pipeout && !unlimited && (blocksperfile ?
299 	    (blocksthisvol >= blocksperfile) : (asize > tsize))) {
300 		close_rewind();
301 		startnewtape(0);
302 	}
303 	timeest();
304 }
305 
306 void
307 trewind(void)
308 {
309 	struct stat sb;
310 	int f;
311 	int got;
312 
313 	for (f = 0; f < SLAVES; f++) {
314 		/*
315 		 * Drain the results, but unlike EOT we DO (or should) care
316 		 * what the return values were, since if we detect EOT after
317 		 * we think we've written the last blocks to the tape anyway,
318 		 * we have to replay those blocks with rollforward.
319 		 *
320 		 * fixme: punt for now.
321 		 */
322 		if (slaves[f].sent) {
323 			if (atomic(read, slaves[f].fd, (char *)&got, sizeof got)
324 			    != sizeof got) {
325 				perror("  DUMP: error reading command pipe in master");
326 				dumpabort(0);
327 			}
328 			slaves[f].sent = 0;
329 			if (got != writesize) {
330 				msg("EOT detected in last 2 tape records!\n");
331 				msg("Use a longer tape, decrease the size estimate\n");
332 				quit("or use no size estimate at all.\n");
333 			}
334 		}
335 		(void) close(slaves[f].fd);
336 	}
337 	while (wait((int *)NULL) >= 0)	/* wait for any signals from slaves */
338 		/* void */;
339 
340 	if (pipeout)
341 		return;
342 
343 	msg("Closing %s\n", tape);
344 
345 	if (popenout) {
346 		tapefd = -1;
347 		(void)pclose(popenfp);
348 		popenfp = NULL;
349 		return;
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 	if (fstat(tapefd, &sb) == 0 && S_ISFIFO(sb.st_mode)) {
361 		(void)close(tapefd);
362 		return;
363 	}
364 	(void) close(tapefd);
365 	while ((f = open(tape, 0)) < 0)
366 		sleep (10);
367 	(void) close(f);
368 }
369 
370 void
371 close_rewind()
372 {
373 	time_t tstart_changevol, tend_changevol;
374 
375 	trewind();
376 	if (nexttape)
377 		return;
378 	(void)time((time_t *)&(tstart_changevol));
379 	if (!nogripe) {
380 		msg("Change Volumes: Mount volume #%d\n", tapeno+1);
381 		broadcast("CHANGE DUMP VOLUMES!\a\a\n");
382 	}
383 	while (!query("Is the new volume mounted and ready to go?"))
384 		if (query("Do you want to abort?")) {
385 			dumpabort(0);
386 			/*NOTREACHED*/
387 		}
388 	(void)time((time_t *)&(tend_changevol));
389 	if ((tstart_changevol != (time_t)-1) && (tend_changevol != (time_t)-1))
390 		tstart_writing += (tend_changevol - tstart_changevol);
391 }
392 
393 void
394 rollforward(void)
395 {
396 	struct req *p, *q, *prev;
397 	struct slave *tslp;
398 	int i, size, got;
399 	int64_t savedtapea;
400 	union u_spcl *ntb, *otb;
401 	tslp = &slaves[SLAVES];
402 	ntb = (union u_spcl *)tslp->tblock[1];
403 
404 	/*
405 	 * Each of the N slaves should have requests that need to
406 	 * be replayed on the next tape.  Use the extra slave buffers
407 	 * (slaves[SLAVES]) to construct request lists to be sent to
408 	 * each slave in turn.
409 	 */
410 	for (i = 0; i < SLAVES; i++) {
411 		q = &tslp->req[1];
412 		otb = (union u_spcl *)slp->tblock;
413 
414 		/*
415 		 * For each request in the current slave, copy it to tslp.
416 		 */
417 
418 		prev = NULL;
419 		for (p = slp->req; p->count > 0; p += p->count) {
420 			*q = *p;
421 			if (p->dblk == 0)
422 				*ntb++ = *otb++; /* copy the datablock also */
423 			prev = q;
424 			q += q->count;
425 		}
426 		if (prev == NULL)
427 			quit("rollforward: protocol botch");
428 		if (prev->dblk != 0)
429 			prev->count -= 1;
430 		else
431 			ntb--;
432 		q -= 1;
433 		q->count = 0;
434 		q = &tslp->req[0];
435 		if (i == 0) {
436 			q->dblk = 0;
437 			q->count = 1;
438 			trecno = 0;
439 			nextblock = tslp->tblock;
440 			savedtapea = spcl.c_tapea;
441 			spcl.c_tapea = slp->tapea;
442 			startnewtape(0);
443 			spcl.c_tapea = savedtapea;
444 			lastspclrec = savedtapea - 1;
445 		}
446 		size = (char *)ntb - (char *)q;
447 		if (atomic(write, slp->fd, (char *)q, size) != size) {
448 			perror("  DUMP: error writing command pipe");
449 			dumpabort(0);
450 		}
451 		slp->sent = 1;
452 		if (++slp >= &slaves[SLAVES])
453 			slp = &slaves[0];
454 
455 		q->count = 1;
456 
457 		if (prev->dblk != 0) {
458 			/*
459 			 * If the last one was a disk block, make the
460 			 * first of this one be the last bit of that disk
461 			 * block...
462 			 */
463 			q->dblk = prev->dblk +
464 				prev->count * (TP_BSIZE / DEV_BSIZE);
465 			ntb = (union u_spcl *)tslp->tblock;
466 		} else {
467 			/*
468 			 * It wasn't a disk block.  Copy the data to its
469 			 * new location in the buffer.
470 			 */
471 			q->dblk = 0;
472 			*((union u_spcl *)tslp->tblock) = *ntb;
473 			ntb = (union u_spcl *)tslp->tblock[1];
474 		}
475 	}
476 	slp->req[0] = *q;
477 	nextblock = slp->tblock;
478 	if (q->dblk == 0)
479 		nextblock++;
480 	trecno = 1;
481 
482 	/*
483 	 * Clear the first slaves' response.  One hopes that it
484 	 * worked ok, otherwise the tape is much too short!
485 	 */
486 	if (slp->sent) {
487 		if (atomic(read, slp->fd, (char *)&got, sizeof got)
488 		    != sizeof got) {
489 			perror("  DUMP: error reading command pipe in master");
490 			dumpabort(0);
491 		}
492 		slp->sent = 0;
493 
494 		if (got != writesize) {
495 			quit("EOT detected at start of the tape!\n");
496 		}
497 	}
498 }
499 
500 /*
501  * We implement taking and restoring checkpoints on the tape level.
502  * When each tape is opened, a new process is created by forking; this
503  * saves all of the necessary context in the parent.  The child
504  * continues the dump; the parent waits around, saving the context.
505  * If the child returns X_REWRITE, then it had problems writing that tape;
506  * this causes the parent to fork again, duplicating the context, and
507  * everything continues as if nothing had happened.
508  */
509 void
510 startnewtape(int top)
511 {
512 	int	parentpid;
513 	int	childpid;
514 	int	status;
515 	char	*p;
516 	sig_t	interrupt_save;
517 
518 	interrupt_save = signal(SIGINT, SIG_IGN);
519 	parentpid = getpid();
520 
521 restore_check_point:
522 	(void)signal(SIGINT, interrupt_save);
523 	/*
524 	 *	All signals are inherited...
525 	 */
526 	setproctitle(NULL);	/* Restore the proctitle. */
527 	childpid = fork();
528 	if (childpid < 0) {
529 		msg("Context save fork fails in parent %d\n", parentpid);
530 		Exit(X_ABORT);
531 	}
532 	if (childpid != 0) {
533 		/*
534 		 *	PARENT:
535 		 *	save the context by waiting
536 		 *	until the child doing all of the work returns.
537 		 *	don't catch the interrupt
538 		 */
539 		signal(SIGINT, SIG_IGN);
540 #ifdef TDEBUG
541 		msg("Tape: %d; parent process: %d child process %d\n",
542 			tapeno+1, parentpid, childpid);
543 #endif /* TDEBUG */
544 		if (waitpid(childpid, &status, 0) == -1)
545 			msg("Waiting for child %d: %s\n", childpid,
546 			    strerror(errno));
547 		if (status & 0xFF) {
548 			msg("Child %d returns LOB status %o\n",
549 				childpid, status&0xFF);
550 		}
551 		status = (status >> 8) & 0xFF;
552 #ifdef TDEBUG
553 		switch(status) {
554 			case X_FINOK:
555 				msg("Child %d finishes X_FINOK\n", childpid);
556 				break;
557 			case X_ABORT:
558 				msg("Child %d finishes X_ABORT\n", childpid);
559 				break;
560 			case X_REWRITE:
561 				msg("Child %d finishes X_REWRITE\n", childpid);
562 				break;
563 			default:
564 				msg("Child %d finishes unknown %d\n",
565 					childpid, status);
566 				break;
567 		}
568 #endif /* TDEBUG */
569 		switch(status) {
570 			case X_FINOK:
571 				Exit(X_FINOK);
572 			case X_ABORT:
573 				Exit(X_ABORT);
574 			case X_REWRITE:
575 				goto restore_check_point;
576 			default:
577 				msg("Bad return code from dump: %d\n", status);
578 				Exit(X_ABORT);
579 		}
580 		/*NOTREACHED*/
581 	} else {	/* we are the child; just continue */
582 #ifdef TDEBUG
583 		sleep(4);	/* allow time for parent's message to get out */
584 		msg("Child on Tape %d has parent %d, my pid = %d\n",
585 			tapeno+1, parentpid, getpid());
586 #endif /* TDEBUG */
587 		/*
588 		 * If we have a name like "/dev/rmt0,/dev/rmt1",
589 		 * use the name before the comma first, and save
590 		 * the remaining names for subsequent volumes.
591 		 */
592 		tapeno++;               /* current tape sequence */
593 		if (nexttape || strchr(tape, ',')) {
594 			if (nexttape && *nexttape)
595 				tape = nexttape;
596 			if ((p = strchr(tape, ',')) != NULL) {
597 				*p = '\0';
598 				nexttape = p + 1;
599 			} else
600 				nexttape = NULL;
601 			msg("Dumping volume %d on %s\n", tapeno, tape);
602 		}
603 		if (pipeout) {
604 			tapefd = STDOUT_FILENO;
605 		} else if (popenout) {
606 			char volno[sizeof("2147483647")];
607 
608 			(void)sprintf(volno, "%d", spcl.c_volume + 1);
609 			if (setenv("DUMP_VOLUME", volno, 1) == -1) {
610 				msg("Cannot set $DUMP_VOLUME.\n");
611 				dumpabort(0);
612 			}
613 			popenfp = popen(popenout, "w");
614 			if (popenfp == NULL) {
615 				msg("Cannot open output pipeline \"%s\".\n",
616 				    popenout);
617 				dumpabort(0);
618 			}
619 			tapefd = fileno(popenfp);
620 		} else {
621 #ifdef RDUMP
622 			while ((tapefd = (host ? rmtopen(tape, 2) :
623 				open(tape, O_WRONLY|O_CREAT, 0666))) < 0)
624 #else
625 			while ((tapefd =
626 			    open(tape, O_WRONLY|O_CREAT, 0666)) < 0)
627 #endif
628 			    {
629 				msg("Cannot open output \"%s\".\n", tape);
630 				if (!query("Do you want to retry the open?"))
631 					dumpabort(0);
632 			}
633 		}
634 
635 		enslave();  /* Share open tape file descriptor with slaves */
636 		if (popenout)
637 			close(tapefd);	/* Give up our copy of it. */
638 		signal(SIGINFO, infosch);
639 
640 		asize = 0;
641 		blocksthisvol = 0;
642 		if (top)
643 			newtape++;		/* new tape signal */
644 		spcl.c_count = slp->count;
645 		/*
646 		 * measure firstrec in TP_BSIZE units since restore doesn't
647 		 * know the correct ntrec value...
648 		 */
649 		spcl.c_firstrec = slp->firstrec;
650 		spcl.c_volume++;
651 		spcl.c_type = TS_TAPE;
652 		writeheader((ino_t)slp->inode);
653 		if (tapeno > 1)
654 			msg("Volume %d begins with blocks from inode %d\n",
655 				tapeno, slp->inode);
656 	}
657 }
658 
659 void
660 dumpabort(int signo __unused)
661 {
662 
663 	if (master != 0 && master != getpid())
664 		/* Signals master to call dumpabort */
665 		(void) kill(master, SIGTERM);
666 	else {
667 		killall();
668 		msg("The ENTIRE dump is aborted.\n");
669 	}
670 #ifdef RDUMP
671 	rmtclose();
672 #endif
673 	Exit(X_ABORT);
674 }
675 
676 void
677 Exit(status)
678 	int status;
679 {
680 
681 #ifdef TDEBUG
682 	msg("pid = %d exits with status %d\n", getpid(), status);
683 #endif /* TDEBUG */
684 	exit(status);
685 }
686 
687 /*
688  * proceed - handler for SIGUSR2, used to synchronize IO between the slaves.
689  */
690 void
691 proceed(int signo __unused)
692 {
693 
694 	if (ready)
695 		longjmp(jmpbuf, 1);
696 	caught++;
697 }
698 
699 void
700 enslave(void)
701 {
702 	int cmd[2];
703 	int i, j;
704 
705 	master = getpid();
706 
707 	signal(SIGTERM, dumpabort);  /* Slave sends SIGTERM on dumpabort() */
708 	signal(SIGPIPE, sigpipe);
709 	signal(SIGUSR1, tperror);    /* Slave sends SIGUSR1 on tape errors */
710 	signal(SIGUSR2, proceed);    /* Slave sends SIGUSR2 to next slave */
711 
712 	for (i = 0; i < SLAVES; i++) {
713 		if (i == slp - &slaves[0]) {
714 			caught = 1;
715 		} else {
716 			caught = 0;
717 		}
718 
719 		if (socketpair(AF_UNIX, SOCK_STREAM, 0, cmd) < 0 ||
720 		    (slaves[i].pid = fork()) < 0)
721 			quit("too many slaves, %d (recompile smaller): %s\n",
722 			    i, strerror(errno));
723 
724 		slaves[i].fd = cmd[1];
725 		slaves[i].sent = 0;
726 		if (slaves[i].pid == 0) { 	    /* Slave starts up here */
727 			for (j = 0; j <= i; j++)
728 			        (void) close(slaves[j].fd);
729 			signal(SIGINT, SIG_IGN);    /* Master handles this */
730 			doslave(cmd[0], i);
731 			Exit(X_FINOK);
732 		}
733 	}
734 
735 	for (i = 0; i < SLAVES; i++)
736 		(void) atomic(write, slaves[i].fd,
737 			      (char *) &slaves[(i + 1) % SLAVES].pid,
738 		              sizeof slaves[0].pid);
739 
740 	master = 0;
741 }
742 
743 void
744 killall(void)
745 {
746 	int i;
747 
748 	for (i = 0; i < SLAVES; i++)
749 		if (slaves[i].pid > 0) {
750 			(void) kill(slaves[i].pid, SIGKILL);
751 			slaves[i].sent = 0;
752 		}
753 }
754 
755 /*
756  * Synchronization - each process has a lockfile, and shares file
757  * descriptors to the following process's lockfile.  When our write
758  * completes, we release our lock on the following process's lock-
759  * file, allowing the following process to lock it and proceed. We
760  * get the lock back for the next cycle by swapping descriptors.
761  */
762 static void
763 doslave(int cmd, int slave_number)
764 {
765 	int nread;
766 	int nextslave, size, wrote, eot_count;
767 
768 	/*
769 	 * Need our own seek pointer.
770 	 */
771 	(void) close(diskfd);
772 	if ((diskfd = open(disk, O_RDONLY)) < 0)
773 		quit("slave couldn't reopen disk: %s\n", strerror(errno));
774 
775 	/*
776 	 * Need the pid of the next slave in the loop...
777 	 */
778 	if ((nread = atomic(read, cmd, (char *)&nextslave, sizeof nextslave))
779 	    != sizeof nextslave) {
780 		quit("master/slave protocol botched - didn't get pid of next slave.\n");
781 	}
782 
783 	/*
784 	 * Get list of blocks to dump, read the blocks into tape buffer
785 	 */
786 	while ((nread = atomic(read, cmd, (char *)slp->req, reqsiz)) == reqsiz) {
787 		struct req *p = slp->req;
788 
789 		for (trecno = 0; trecno < ntrec;
790 		     trecno += p->count, p += p->count) {
791 			if (p->dblk) {
792 				blkread(p->dblk, slp->tblock[trecno],
793 					p->count * TP_BSIZE);
794 			} else {
795 				if (p->count != 1 || atomic(read, cmd,
796 				    (char *)slp->tblock[trecno],
797 				    TP_BSIZE) != TP_BSIZE)
798 				       quit("master/slave protocol botched.\n");
799 			}
800 		}
801 		if (setjmp(jmpbuf) == 0) {
802 			ready = 1;
803 			if (!caught)
804 				(void) pause();
805 		}
806 		ready = 0;
807 		caught = 0;
808 
809 		/* Try to write the data... */
810 		eot_count = 0;
811 		size = 0;
812 
813 		wrote = 0;
814 		while (eot_count < 10 && size < writesize) {
815 #ifdef RDUMP
816 			if (host)
817 				wrote = rmtwrite(slp->tblock[0]+size,
818 				    writesize-size);
819 			else
820 #endif
821 				wrote = write(tapefd, slp->tblock[0]+size,
822 				    writesize-size);
823 #ifdef WRITEDEBUG
824 			printf("slave %d wrote %d\n", slave_number, wrote);
825 #endif
826 			if (wrote < 0)
827 				break;
828 			if (wrote == 0)
829 				eot_count++;
830 			size += wrote;
831 		}
832 
833 #ifdef WRITEDEBUG
834 		if (size != writesize)
835 		 printf("slave %d only wrote %d out of %d bytes and gave up.\n",
836 		     slave_number, size, writesize);
837 #endif
838 
839 		/*
840 		 * Handle ENOSPC as an EOT condition.
841 		 */
842 		if (wrote < 0 && errno == ENOSPC) {
843 			wrote = 0;
844 			eot_count++;
845 		}
846 
847 		if (eot_count > 0)
848 			size = 0;
849 
850 		if (wrote < 0) {
851 			(void) kill(master, SIGUSR1);
852 			for (;;)
853 				(void) sigpause(0);
854 		} else {
855 			/*
856 			 * pass size of write back to master
857 			 * (for EOT handling)
858 			 */
859 			(void) atomic(write, cmd, (char *)&size, sizeof size);
860 		}
861 
862 		/*
863 		 * If partial write, don't want next slave to go.
864 		 * Also jolts him awake.
865 		 */
866 		(void) kill(nextslave, SIGUSR2);
867 	}
868 	if (nread != 0)
869 		quit("error reading command pipe: %s\n", strerror(errno));
870 }
871 
872 /*
873  * Since a read from a pipe may not return all we asked for,
874  * or a write may not write all we ask if we get a signal,
875  * loop until the count is satisfied (or error).
876  */
877 static int
878 atomic(ssize_t (*func)(), int fd, char *buf, int count)
879 {
880 	int got, need = count;
881 
882 	while ((got = (*func)(fd, buf, need)) > 0 && (need -= got) > 0)
883 		buf += got;
884 	return (got < 0 ? got : count - need);
885 }
886