xref: /freebsd/sbin/dump/tape.c (revision 8aac90f18aef7c9eea906c3ff9a001ca7b94f375)
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 #include <sys/param.h>
33 #include <sys/socket.h>
34 #include <sys/wait.h>
35 #include <sys/stat.h>
36 
37 #include <ufs/ufs/dinode.h>
38 #include <ufs/ffs/fs.h>
39 
40 #include <protocols/dumprestore.h>
41 
42 #include <assert.h>
43 #include <errno.h>
44 #include <fcntl.h>
45 #include <limits.h>
46 #include <setjmp.h>
47 #include <signal.h>
48 #include <stdio.h>
49 #include <stdlib.h>
50 #include <string.h>
51 #include <time.h>
52 #include <unistd.h>
53 
54 #include "dump.h"
55 
56 ino_t	curino;			/* current inumber; used globally */
57 int	newtape;		/* new tape flag */
58 union	u_spcl u_spcl;		/* mapping of variables in a control block */
59 
60 static	int tapefd;		/* tape file descriptor */
61 static	long asize;		/* number of 0.1" units written on cur tape */
62 static	int writesize;		/* size of malloc()ed buffer for tape */
63 static	int64_t lastspclrec = -1; /* tape block number of last written header */
64 static	int trecno = 0;		/* next record to write in current block */
65 static	long blocksthisvol;	/* number of blocks on current output file */
66 static	char *nexttape;
67 static	FILE *popenfp = NULL;
68 
69 static	int atomic_read(int, void *, int);
70 static	int atomic_write(int, const void *, int);
71 static	void worker(int, int);
72 static	void create_workers(void);
73 static	void flushtape(void);
74 static	void killall(void);
75 static	void rollforward(void);
76 
77 /*
78  * Concurrent dump mods (Caltech) - disk block reading and tape writing
79  * are exported to several worker processes.  While one worker 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 file system and
82  * sends writeheader()'s and lists of daddr's to the workers via pipes.
83  * The following structure defines the instruction packets sent to workers.
84  */
85 struct req {
86 	ufs2_daddr_t dblk;
87 	int count;
88 };
89 static int reqsiz;
90 
91 #define WORKERS 3		/* 1 worker writing, 1 reading, 1 for slack */
92 static struct worker {
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 worker */
99 	int pid;		/* PID for this worker */
100 	int sent;		/* 1 == we've sent this worker requests */
101 	char (*tblock)[TP_BSIZE]; /* buffer for data blocks */
102 	struct req *req;	/* buffer for requests */
103 } workers[WORKERS+1];
104 static struct worker *wp;
105 
106 static char	(*nextblock)[TP_BSIZE];
107 
108 static int master;	/* pid of master, for sending error signals */
109 static int tenths;	/* length of tape used per block written */
110 static volatile sig_atomic_t caught; /* have we caught the signal to proceed? */
111 static volatile sig_atomic_t ready; /* reached the lock point without having */
112 			/* received the SIGUSR2 signal from the prev worker? */
113 static jmp_buf jmpbuf;	/* where to jump to if we are ready when the */
114 			/* SIGUSR2 arrives from the previous worker */
115 
116 int
117 alloctape(void)
118 {
119 	int pgoff = getpagesize() - 1;
120 	char *buf;
121 	int i;
122 
123 	writesize = ntrec * TP_BSIZE;
124 	reqsiz = (ntrec + 1) * sizeof(struct req);
125 	/*
126 	 * CDC 92181's and 92185's make 0.8" gaps in 1600-bpi start/stop mode
127 	 * (see DEC TU80 User's Guide).  The shorter gaps of 6250-bpi require
128 	 * repositioning after stopping, i.e, streaming mode, where the gap is
129 	 * variable, 0.30" to 0.45".  The gap is maximal when the tape stops.
130 	 */
131 	if (blocksperfile == 0 && !unlimited)
132 		tenths = writesize / density +
133 		    (cartridge ? 16 : density == 625 ? 5 : 8);
134 	/*
135 	 * Allocate tape buffer contiguous with the array of instruction
136 	 * packets, so flushtape() can write them together with one write().
137 	 * Align tape buffer on page boundary to speed up tape write().
138 	 */
139 	for (i = 0; i <= WORKERS; i++) {
140 		buf = (char *)
141 		    malloc((unsigned)(reqsiz + writesize + pgoff + TP_BSIZE));
142 		if (buf == NULL)
143 			return(0);
144 		workers[i].tblock = (char (*)[TP_BSIZE])
145 		    (((long)&buf[ntrec + 1] + pgoff) &~ pgoff);
146 		workers[i].req = (struct req *)workers[i].tblock - ntrec - 1;
147 	}
148 	wp = &workers[0];
149 	wp->count = 1;
150 	wp->tapea = 0;
151 	wp->firstrec = 0;
152 	nextblock = wp->tblock;
153 	return(1);
154 }
155 
156 void
157 writerec(char *dp, int isspcl)
158 {
159 
160 	wp->req[trecno].dblk = (ufs2_daddr_t)0;
161 	wp->req[trecno].count = 1;
162 	/* Can't do a structure assignment due to alignment problems */
163 	bcopy(dp, *(nextblock)++, sizeof (union u_spcl));
164 	if (isspcl)
165 		lastspclrec = spcl.c_tapea;
166 	trecno++;
167 	spcl.c_tapea++;
168 	if (trecno >= ntrec)
169 		flushtape();
170 }
171 
172 void
173 dumpblock(ufs2_daddr_t blkno, int size)
174 {
175 	int avail, tpblks;
176 	ufs2_daddr_t dblkno;
177 
178 	dblkno = fsbtodb(sblock, blkno);
179 	tpblks = size >> tp_bshift;
180 	while ((avail = MIN(tpblks, ntrec - trecno)) > 0) {
181 		wp->req[trecno].dblk = dblkno;
182 		wp->req[trecno].count = avail;
183 		trecno += avail;
184 		spcl.c_tapea += avail;
185 		if (trecno >= ntrec)
186 			flushtape();
187 		dblkno += avail << (tp_bshift - dev_bshift);
188 		tpblks -= avail;
189 	}
190 }
191 
192 int	nogripe = 0;
193 
194 void
195 tperror(int signo __unused)
196 {
197 
198 	if (pipeout) {
199 		msg("write error on %s\n", tape);
200 		quit("Cannot recover\n");
201 		/* NOTREACHED */
202 	}
203 	msg("write error %ld blocks into volume %d\n", blocksthisvol, tapeno);
204 	broadcast("DUMP WRITE ERROR!\n");
205 	if (!query("Do you want to restart?"))
206 		dumpabort(0);
207 	msg("Closing this volume.  Prepare to restart with new media;\n");
208 	msg("this dump volume will be rewritten.\n");
209 	killall();
210 	nogripe = 1;
211 	close_rewind();
212 	Exit(X_REWRITE);
213 }
214 
215 void
216 sigpipe(int signo __unused)
217 {
218 
219 	quit("Broken pipe\n");
220 }
221 
222 static void
223 flushtape(void)
224 {
225 	int i, blks, got;
226 	int64_t lastfirstrec;
227 
228 	int siz = (char *)nextblock - (char *)wp->req;
229 
230 	wp->req[trecno].count = 0;			/* Sentinel */
231 
232 	if (atomic_write(wp->fd, (const void *)wp->req, siz) != siz)
233 		quit("error writing command pipe: %s\n", strerror(errno));
234 	wp->sent = 1; /* we sent a request, read the response later */
235 
236 	lastfirstrec = wp->firstrec;
237 
238 	if (++wp >= &workers[WORKERS])
239 		wp = &workers[0];
240 
241 	/* Read results back from next worker */
242 	if (wp->sent) {
243 		if (atomic_read(wp->fd, (void *)&got, sizeof got)
244 		    != sizeof got) {
245 			perror("  DUMP: error reading command pipe in master");
246 			dumpabort(0);
247 		}
248 		wp->sent = 0;
249 
250 		/* Check for end of tape */
251 		if (got < writesize) {
252 			msg("End of tape detected\n");
253 
254 			/*
255 			 * Drain the results, don't care what the values were.
256 			 * If we read them here then trewind won't...
257 			 */
258 			for (i = 0; i < WORKERS; i++) {
259 				if (workers[i].sent) {
260 					if (atomic_read(workers[i].fd,
261 					    (void *)&got, sizeof got)
262 					    != sizeof got) {
263 						perror("  DUMP: error reading command pipe in master");
264 						dumpabort(0);
265 					}
266 					workers[i].sent = 0;
267 				}
268 			}
269 
270 			close_rewind();
271 			rollforward();
272 			return;
273 		}
274 	}
275 
276 	blks = 0;
277 	if (spcl.c_type != TS_END && spcl.c_type != TS_CLRI &&
278 	    spcl.c_type != TS_BITS) {
279 		assert(spcl.c_count <= TP_NINDIR);
280 		for (i = 0; i < spcl.c_count; i++)
281 			if (spcl.c_addr[i] != 0)
282 				blks++;
283 	}
284 	wp->count = lastspclrec + blks + 1 - spcl.c_tapea;
285 	wp->tapea = spcl.c_tapea;
286 	wp->firstrec = lastfirstrec + ntrec;
287 	wp->inode = curino;
288 	nextblock = wp->tblock;
289 	trecno = 0;
290 	asize += tenths;
291 	blockswritten += ntrec;
292 	blocksthisvol += ntrec;
293 	if (!pipeout && !unlimited && (blocksperfile ?
294 	    (blocksthisvol >= blocksperfile) : (asize > tsize))) {
295 		close_rewind();
296 		startnewtape(0);
297 	}
298 	timeest();
299 }
300 
301 void
302 trewind(void)
303 {
304 	struct stat sb;
305 	int f;
306 	int got;
307 
308 	for (f = 0; f < WORKERS; f++) {
309 		/*
310 		 * Drain the results, but unlike EOT we DO (or should) care
311 		 * what the return values were, since if we detect EOT after
312 		 * we think we've written the last blocks to the tape anyway,
313 		 * we have to replay those blocks with rollforward.
314 		 *
315 		 * fixme: punt for now.
316 		 */
317 		if (workers[f].sent) {
318 			if (atomic_read(workers[f].fd, (void *)&got, sizeof got)
319 			    != sizeof got) {
320 				perror("  DUMP: error reading command pipe in master");
321 				dumpabort(0);
322 			}
323 			workers[f].sent = 0;
324 			if (got != writesize) {
325 				msg("EOT detected in last 2 tape records!\n");
326 				msg("Use a longer tape, decrease the size estimate\n");
327 				quit("or use no size estimate at all.\n");
328 			}
329 		}
330 		(void) close(workers[f].fd);
331 	}
332 	while (wait((int *)NULL) >= 0)	/* wait for any signals from workers */
333 		/* void */;
334 
335 	if (pipeout)
336 		return;
337 
338 	msg("Closing %s\n", tape);
339 
340 	if (popenout) {
341 		tapefd = -1;
342 		(void)pclose(popenfp);
343 		popenfp = NULL;
344 		return;
345 	}
346 #ifdef RDUMP
347 	if (host) {
348 		rmtclose();
349 		while (rmtopen(tape, 0) < 0)
350 			sleep(10);
351 		rmtclose();
352 		return;
353 	}
354 #endif
355 	if (fstat(tapefd, &sb) == 0 && S_ISFIFO(sb.st_mode)) {
356 		(void)close(tapefd);
357 		return;
358 	}
359 	(void) close(tapefd);
360 	while ((f = open(tape, 0)) < 0)
361 		sleep (10);
362 	(void) close(f);
363 }
364 
365 void
366 close_rewind()
367 {
368 	time_t tstart_changevol, tend_changevol;
369 
370 	trewind();
371 	if (nexttape)
372 		return;
373 	(void)time((time_t *)&(tstart_changevol));
374 	if (!nogripe) {
375 		msg("Change Volumes: Mount volume #%d\n", tapeno+1);
376 		broadcast("CHANGE DUMP VOLUMES!\a\a\n");
377 	}
378 	while (!query("Is the new volume mounted and ready to go?"))
379 		if (query("Do you want to abort?")) {
380 			dumpabort(0);
381 			/*NOTREACHED*/
382 		}
383 	(void)time((time_t *)&(tend_changevol));
384 	if ((tstart_changevol != (time_t)-1) && (tend_changevol != (time_t)-1))
385 		tstart_writing += (tend_changevol - tstart_changevol);
386 }
387 
388 void
389 rollforward(void)
390 {
391 	struct req *p, *q, *prev;
392 	struct worker *twp;
393 	int i, size, got;
394 	int64_t savedtapea;
395 	union u_spcl *ntb, *otb;
396 	twp = &workers[WORKERS];
397 	ntb = (union u_spcl *)twp->tblock[1];
398 
399 	/*
400 	 * Each of the N workers should have requests that need to
401 	 * be replayed on the next tape.  Use the extra worker buffers
402 	 * (workers[WORKERS]) to construct request lists to be sent to
403 	 * each worker in turn.
404 	 */
405 	for (i = 0; i < WORKERS; i++) {
406 		q = &twp->req[1];
407 		otb = (union u_spcl *)wp->tblock;
408 
409 		/*
410 		 * For each request in the current worker, copy it to twp.
411 		 */
412 
413 		prev = NULL;
414 		for (p = wp->req; p->count > 0; p += p->count) {
415 			*q = *p;
416 			if (p->dblk == 0)
417 				*ntb++ = *otb++; /* copy the datablock also */
418 			prev = q;
419 			q += q->count;
420 		}
421 		if (prev == NULL)
422 			quit("rollforward: protocol botch");
423 		if (prev->dblk != 0)
424 			prev->count -= 1;
425 		else
426 			ntb--;
427 		q -= 1;
428 		q->count = 0;
429 		q = &twp->req[0];
430 		if (i == 0) {
431 			q->dblk = 0;
432 			q->count = 1;
433 			trecno = 0;
434 			nextblock = twp->tblock;
435 			savedtapea = spcl.c_tapea;
436 			spcl.c_tapea = wp->tapea;
437 			startnewtape(0);
438 			spcl.c_tapea = savedtapea;
439 			lastspclrec = savedtapea - 1;
440 		}
441 		size = (char *)ntb - (char *)q;
442 		if (atomic_write(wp->fd, (const void *)q, size) != size) {
443 			perror("  DUMP: error writing command pipe");
444 			dumpabort(0);
445 		}
446 		wp->sent = 1;
447 		if (++wp >= &workers[WORKERS])
448 			wp = &workers[0];
449 
450 		q->count = 1;
451 
452 		if (prev->dblk != 0) {
453 			/*
454 			 * If the last one was a disk block, make the
455 			 * first of this one be the last bit of that disk
456 			 * block...
457 			 */
458 			q->dblk = prev->dblk +
459 				prev->count * (TP_BSIZE / DEV_BSIZE);
460 			ntb = (union u_spcl *)twp->tblock;
461 		} else {
462 			/*
463 			 * It wasn't a disk block.  Copy the data to its
464 			 * new location in the buffer.
465 			 */
466 			q->dblk = 0;
467 			*((union u_spcl *)twp->tblock) = *ntb;
468 			ntb = (union u_spcl *)twp->tblock[1];
469 		}
470 	}
471 	wp->req[0] = *q;
472 	nextblock = wp->tblock;
473 	if (q->dblk == 0)
474 		nextblock++;
475 	trecno = 1;
476 
477 	/*
478 	 * Clear the first workers' response.  One hopes that it
479 	 * worked ok, otherwise the tape is much too short!
480 	 */
481 	if (wp->sent) {
482 		if (atomic_read(wp->fd, (void *)&got, sizeof got)
483 		    != sizeof got) {
484 			perror("  DUMP: error reading command pipe in master");
485 			dumpabort(0);
486 		}
487 		wp->sent = 0;
488 
489 		if (got != writesize) {
490 			quit("EOT detected at start of the tape!\n");
491 		}
492 	}
493 }
494 
495 /*
496  * We implement taking and restoring checkpoints on the tape level.
497  * When each tape is opened, a new process is created by forking; this
498  * saves all of the necessary context in the parent.  The child
499  * continues the dump; the parent waits around, saving the context.
500  * If the child returns X_REWRITE, then it had problems writing that tape;
501  * this causes the parent to fork again, duplicating the context, and
502  * everything continues as if nothing had happened.
503  */
504 void
505 startnewtape(int top)
506 {
507 	int	parentpid;
508 	int	childpid;
509 	int	status;
510 	char	*p;
511 	sig_t	interrupt_save;
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 	setproctitle(NULL);	/* Restore the proctitle. */
522 	childpid = fork();
523 	if (childpid < 0) {
524 		msg("Context save fork fails in parent %d\n", parentpid);
525 		Exit(X_ABORT);
526 	}
527 	if (childpid != 0) {
528 		/*
529 		 *	PARENT:
530 		 *	save the context by waiting
531 		 *	until the child doing all of the work returns.
532 		 *	don't catch the interrupt
533 		 */
534 		signal(SIGINT, SIG_IGN);
535 #ifdef TDEBUG
536 		msg("Tape: %d; parent process: %d child process %d\n",
537 			tapeno+1, parentpid, childpid);
538 #endif /* TDEBUG */
539 		if (waitpid(childpid, &status, 0) == -1)
540 			msg("Waiting for child %d: %s\n", childpid,
541 			    strerror(errno));
542 		if (status & 0xFF) {
543 			msg("Child %d returns LOB status %o\n",
544 				childpid, status&0xFF);
545 		}
546 		status = (status >> 8) & 0xFF;
547 #ifdef TDEBUG
548 		switch(status) {
549 			case X_FINOK:
550 				msg("Child %d finishes X_FINOK\n", childpid);
551 				break;
552 			case X_ABORT:
553 				msg("Child %d finishes X_ABORT\n", childpid);
554 				break;
555 			case X_REWRITE:
556 				msg("Child %d finishes X_REWRITE\n", childpid);
557 				break;
558 			default:
559 				msg("Child %d finishes unknown %d\n",
560 					childpid, status);
561 				break;
562 		}
563 #endif /* TDEBUG */
564 		switch(status) {
565 			case X_FINOK:
566 				Exit(X_FINOK);
567 			case X_ABORT:
568 				Exit(X_ABORT);
569 			case X_REWRITE:
570 				goto restore_check_point;
571 			default:
572 				msg("Bad return code from dump: %d\n", status);
573 				Exit(X_ABORT);
574 		}
575 		/*NOTREACHED*/
576 	} else {	/* we are the child; just continue */
577 #ifdef TDEBUG
578 		sleep(4);	/* allow time for parent's message to get out */
579 		msg("Child on Tape %d has parent %d, my pid = %d\n",
580 			tapeno+1, parentpid, getpid());
581 #endif /* TDEBUG */
582 		/*
583 		 * If we have a name like "/dev/rmt0,/dev/rmt1",
584 		 * use the name before the comma first, and save
585 		 * the remaining names for subsequent volumes.
586 		 */
587 		tapeno++;               /* current tape sequence */
588 		if (nexttape || strchr(tape, ',')) {
589 			if (nexttape && *nexttape)
590 				tape = nexttape;
591 			if ((p = strchr(tape, ',')) != NULL) {
592 				*p = '\0';
593 				nexttape = p + 1;
594 			} else
595 				nexttape = NULL;
596 			msg("Dumping volume %d on %s\n", tapeno, tape);
597 		}
598 		if (pipeout) {
599 			tapefd = STDOUT_FILENO;
600 		} else if (popenout) {
601 			char volno[sizeof("2147483647")];
602 
603 			(void)sprintf(volno, "%d", spcl.c_volume + 1);
604 			if (setenv("DUMP_VOLUME", volno, 1) == -1) {
605 				msg("Cannot set $DUMP_VOLUME.\n");
606 				dumpabort(0);
607 			}
608 			popenfp = popen(popenout, "w");
609 			if (popenfp == NULL) {
610 				msg("Cannot open output pipeline \"%s\".\n",
611 				    popenout);
612 				dumpabort(0);
613 			}
614 			tapefd = fileno(popenfp);
615 		} else {
616 #ifdef RDUMP
617 			while ((tapefd = (host ? rmtopen(tape, 2) :
618 				open(tape, O_WRONLY|O_CREAT, 0666))) < 0)
619 #else
620 			while ((tapefd =
621 			    open(tape, O_WRONLY|O_CREAT, 0666)) < 0)
622 #endif
623 			    {
624 				msg("Cannot open output \"%s\".\n", tape);
625 				if (!query("Do you want to retry the open?"))
626 					dumpabort(0);
627 			}
628 		}
629 
630 		create_workers();  /* Share open tape file descriptor with workers */
631 		if (popenout)
632 			close(tapefd);	/* Give up our copy of it. */
633 		signal(SIGINFO, infosch);
634 
635 		asize = 0;
636 		blocksthisvol = 0;
637 		if (top)
638 			newtape++;		/* new tape signal */
639 		spcl.c_count = wp->count;
640 		/*
641 		 * measure firstrec in TP_BSIZE units since restore doesn't
642 		 * know the correct ntrec value...
643 		 */
644 		spcl.c_firstrec = wp->firstrec;
645 		spcl.c_volume++;
646 		spcl.c_type = TS_TAPE;
647 		writeheader((ino_t)wp->inode);
648 		if (tapeno > 1)
649 			msg("Volume %d begins with blocks from inode %d\n",
650 				tapeno, wp->inode);
651 	}
652 }
653 
654 void
655 dumpabort(int signo __unused)
656 {
657 
658 	if (master != 0 && master != getpid())
659 		/* Signals master to call dumpabort */
660 		(void) kill(master, SIGTERM);
661 	else {
662 		killall();
663 		msg("The ENTIRE dump is aborted.\n");
664 	}
665 #ifdef RDUMP
666 	rmtclose();
667 #endif
668 	Exit(X_ABORT);
669 }
670 
671 void
672 Exit(int status)
673 {
674 
675 #ifdef TDEBUG
676 	msg("pid = %d exits with status %d\n", getpid(), status);
677 #endif /* TDEBUG */
678 	exit(status);
679 }
680 
681 /*
682  * proceed - handler for SIGUSR2, used to synchronize IO between the workers.
683  */
684 void
685 proceed(int signo __unused)
686 {
687 
688 	if (ready)
689 		longjmp(jmpbuf, 1);
690 	caught++;
691 }
692 
693 void
694 create_workers(void)
695 {
696 	int cmd[2];
697 	int i, j;
698 
699 	master = getpid();
700 
701 	signal(SIGTERM, dumpabort);  /* Worker sends SIGTERM on dumpabort() */
702 	signal(SIGPIPE, sigpipe);
703 	signal(SIGUSR1, tperror);    /* Worker sends SIGUSR1 on tape errors */
704 	signal(SIGUSR2, proceed);    /* Worker sends SIGUSR2 to next worker */
705 
706 	for (i = 0; i < WORKERS; i++) {
707 		if (i == wp - &workers[0]) {
708 			caught = 1;
709 		} else {
710 			caught = 0;
711 		}
712 
713 		if (socketpair(AF_UNIX, SOCK_STREAM, 0, cmd) < 0 ||
714 		    (workers[i].pid = fork()) < 0)
715 			quit("too many workers, %d (recompile smaller): %s\n",
716 			    i, strerror(errno));
717 
718 		workers[i].fd = cmd[1];
719 		workers[i].sent = 0;
720 		if (workers[i].pid == 0) { 	    /* Worker starts up here */
721 			for (j = 0; j <= i; j++)
722 			        (void) close(workers[j].fd);
723 			signal(SIGINT, SIG_IGN);    /* Master handles this */
724 			worker(cmd[0], i);
725 			Exit(X_FINOK);
726 		}
727 	}
728 
729 	for (i = 0; i < WORKERS; i++)
730 		(void) atomic_write(workers[i].fd,
731 			      (const void *) &workers[(i + 1) % WORKERS].pid,
732 		              sizeof workers[0].pid);
733 
734 	master = 0;
735 }
736 
737 void
738 killall(void)
739 {
740 	int i;
741 
742 	for (i = 0; i < WORKERS; i++)
743 		if (workers[i].pid > 0) {
744 			(void) kill(workers[i].pid, SIGKILL);
745 			workers[i].sent = 0;
746 		}
747 }
748 
749 /*
750  * Synchronization - each process has a lockfile, and shares file
751  * descriptors to the following process's lockfile.  When our write
752  * completes, we release our lock on the following process's lock-
753  * file, allowing the following process to lock it and proceed. We
754  * get the lock back for the next cycle by swapping descriptors.
755  */
756 static void
757 worker(int cmd, int worker_number)
758 {
759 	int nread;
760 	int nextworker, size, wrote, eot_count;
761 
762 	/*
763 	 * Need our own seek pointer.
764 	 */
765 	(void) close(diskfd);
766 	if ((diskfd = open(disk, O_RDONLY)) < 0)
767 		quit("worker couldn't reopen disk: %s\n", strerror(errno));
768 
769 	/*
770 	 * Need the pid of the next worker in the loop...
771 	 */
772 	if ((nread = atomic_read(cmd, (void *)&nextworker, sizeof nextworker))
773 	    != sizeof nextworker) {
774 		quit("master/worker protocol botched - didn't get pid of next worker.\n");
775 	}
776 
777 	/*
778 	 * Get list of blocks to dump, read the blocks into tape buffer
779 	 */
780 	while ((nread = atomic_read(cmd, (void *)wp->req, reqsiz)) == reqsiz) {
781 		struct req *p = wp->req;
782 
783 		for (trecno = 0; trecno < ntrec;
784 		     trecno += p->count, p += p->count) {
785 			if (p->dblk) {
786 				blkread(p->dblk, wp->tblock[trecno],
787 					p->count * TP_BSIZE);
788 			} else {
789 				if (p->count != 1 || atomic_read(cmd,
790 				    (void *)wp->tblock[trecno],
791 				    TP_BSIZE) != TP_BSIZE)
792 				       quit("master/worker protocol botched.\n");
793 			}
794 		}
795 		if (setjmp(jmpbuf) == 0) {
796 			ready = 1;
797 			if (!caught)
798 				(void) pause();
799 		}
800 		ready = 0;
801 		caught = 0;
802 
803 		/* Try to write the data... */
804 		eot_count = 0;
805 		size = 0;
806 
807 		wrote = 0;
808 		while (eot_count < 10 && size < writesize) {
809 #ifdef RDUMP
810 			if (host)
811 				wrote = rmtwrite(wp->tblock[0]+size,
812 				    writesize-size);
813 			else
814 #endif
815 				wrote = write(tapefd, wp->tblock[0]+size,
816 				    writesize-size);
817 #ifdef WRITEDEBUG
818 			printf("worker %d wrote %d\n", worker_number, wrote);
819 #endif
820 			if (wrote < 0)
821 				break;
822 			if (wrote == 0)
823 				eot_count++;
824 			size += wrote;
825 		}
826 
827 #ifdef WRITEDEBUG
828 		if (size != writesize)
829 		 printf("worker %d only wrote %d out of %d bytes and gave up.\n",
830 		     worker_number, size, writesize);
831 #endif
832 
833 		/*
834 		 * Handle ENOSPC as an EOT condition.
835 		 */
836 		if (wrote < 0 && errno == ENOSPC) {
837 			wrote = 0;
838 			eot_count++;
839 		}
840 
841 		if (eot_count > 0)
842 			size = 0;
843 
844 		if (wrote < 0) {
845 			(void) kill(master, SIGUSR1);
846 			for (;;)
847 				(void) sigpause(0);
848 		} else {
849 			/*
850 			 * pass size of write back to master
851 			 * (for EOT handling)
852 			 */
853 			(void)atomic_write(cmd, (const void *)&size,
854 			    sizeof size);
855 		}
856 
857 		/*
858 		 * If partial write, don't want next worker to go.
859 		 * Also jolts him awake.
860 		 */
861 		(void) kill(nextworker, SIGUSR2);
862 	}
863 	if (nread != 0)
864 		quit("error reading command pipe: %s\n", strerror(errno));
865 }
866 
867 /*
868  * Since a read from a pipe may not return all we asked for,
869  * loop until the count is satisfied (or error).
870  */
871 static int
872 atomic_read(int fd, void *buf, int count)
873 {
874 	int got, need = count;
875 
876 	while ((got = read(fd, buf, need)) > 0 && (need -= got) > 0)
877 		buf += got;
878 	return (got < 0 ? got : count - need);
879 }
880 
881 /*
882  * Since a write to a pipe may not write all we ask if we get a signal,
883  * loop until the count is satisfied (or error).
884  */
885 static int
886 atomic_write(int fd, const void *buf, int count)
887 {
888 	int got, need = count;
889 
890 	while ((got = write(fd, buf, need)) > 0 && (need -= got) > 0)
891 		buf += got;
892 	return (got < 0 ? got : count - need);
893 }
894