xref: /csrg-svn/sbin/dump/tape.c (revision 50504)
1 /*-
2  * Copyright (c) 1980, 1991 The Regents of the University of California.
3  * All rights reserved.
4  *
5  * %sccs.include.redist.c%
6  */
7 
8 #ifndef lint
9 static char sccsid[] = "@(#)tape.c	5.19 (Berkeley) 07/23/91";
10 #endif /* not lint */
11 
12 #include <sys/param.h>
13 #include <sys/wait.h>
14 #include <ufs/dinode.h>
15 #include <ufs/fs.h>
16 #include <signal.h>
17 #include <fcntl.h>
18 #include <protocols/dumprestore.h>
19 #include <errno.h>
20 #include <setjmp.h>
21 #ifdef __STDC__
22 #include <unistd.h>
23 #include <stdlib.h>
24 #include <string.h>
25 #endif
26 #include <sys/socket.h>
27 #include "dump.h"
28 #include "pathnames.h"
29 
30 int	writesize;		/* size of malloc()ed buffer for tape */
31 long	lastspclrec = -1;	/* tape block number of last written header */
32 int	trecno = 0;		/* next record to write in current block */
33 extern	long blocksperfile;	/* number of blocks per output file */
34 long	blocksthisvol;		/* number of blocks on current output file */
35 extern	int ntrec;		/* blocking factor on tape */
36 extern	int cartridge;
37 extern	char *host;
38 char	*nexttape;
39 #ifdef RDUMP
40 int	rmtopen(), rmtwrite();
41 void	rmtclose();
42 #endif RDUMP
43 void	rollforward();
44 int	atomic();
45 void	doslave(), enslave(), flushtape(), killall();
46 
47 /*
48  * Concurrent dump mods (Caltech) - disk block reading and tape writing
49  * are exported to several slave processes.  While one slave writes the
50  * tape, the others read disk blocks; they pass control of the tape in
51  * a ring via signals. The parent process traverses the filesystem and
52  * sends writeheader()'s and lists of daddr's to the slaves via pipes.
53  * The following structure defines the instruction packets sent to slaves.
54  */
55 struct req {
56 	daddr_t dblk;
57 	int count;
58 };
59 int reqsiz;
60 
61 #define SLAVES 3		/* 1 slave writing, 1 reading, 1 for slack */
62 struct slave {
63 	int tapea;		/* header number at start of this chunk */
64 	int count;		/* count to next header (used for TS_TAPE */
65 				/* after EOT) */
66 	int inode;		/* inode that we are currently dealing with */
67 	int fd;			/* FD for this slave */
68 	int pid;		/* PID for this slave */
69 	int sent;		/* 1 == we've sent this slave requests */
70 	int firstrec;		/* record number of this block */
71 	char (*tblock)[TP_BSIZE]; /* buffer for data blocks */
72 	struct req *req;	/* buffer for requests */
73 } slaves[SLAVES+1];
74 struct slave *slp;
75 
76 char	(*nextblock)[TP_BSIZE];
77 
78 int master;		/* pid of master, for sending error signals */
79 int tenths;		/* length of tape used per block written */
80 static int caught;	/* have we caught the signal to proceed? */
81 static int ready;	/* have we reached the lock point without having */
82 			/* received the SIGUSR2 signal from the prev slave? */
83 static jmp_buf jmpbuf;	/* where to jump to if we are ready when the */
84 			/* SIGUSR2 arrives from the previous slave */
85 
86 int
87 alloctape()
88 {
89 	int pgoff = getpagesize() - 1;
90 	char *buf;
91 	int i;
92 
93 	writesize = ntrec * TP_BSIZE;
94 	reqsiz = (ntrec + 1) * sizeof(struct req);
95 	/*
96 	 * CDC 92181's and 92185's make 0.8" gaps in 1600-bpi start/stop mode
97 	 * (see DEC TU80 User's Guide).  The shorter gaps of 6250-bpi require
98 	 * repositioning after stopping, i.e, streaming mode, where the gap is
99 	 * variable, 0.30" to 0.45".  The gap is maximal when the tape stops.
100 	 */
101 	if (blocksperfile == 0)
102 		tenths = writesize / density +
103 		    (cartridge ? 16 : density == 625 ? 5 : 8);
104 	/*
105 	 * Allocate tape buffer contiguous with the array of instruction
106 	 * packets, so flushtape() can write them together with one write().
107 	 * Align tape buffer on page boundary to speed up tape write().
108 	 */
109 	for (i = 0; i <= SLAVES; i++) {
110 		buf = (char *) malloc(reqsiz + writesize + pgoff + TP_BSIZE);
111 		if (buf == NULL)
112 		  return(0);
113 		slaves[i].tblock = (char (*)[TP_BSIZE])
114 		    (((long)&buf[ntrec + 1] + pgoff) &~ pgoff);
115 		slaves[i].req = (struct req *)slaves[i].tblock - ntrec - 1;
116 	}
117 	slp = &slaves[0];
118 	slp->count = 1;
119 	slp->tapea = 0;
120 	slp->firstrec = 0;
121 	nextblock = slp->tblock;
122 	return(1);
123 }
124 
125 void
126 writerec(dp)
127 	char *dp;
128 {
129 
130 	slp->req[trecno].dblk = (daddr_t)0;
131 	slp->req[trecno].count = 1;
132 	*(union u_spcl *)(*(nextblock)++) = *(union u_spcl *)dp;
133 	lastspclrec = spcl.c_tapea;
134 	trecno++;
135 	spcl.c_tapea++;
136 	if (trecno >= ntrec)
137 		flushtape();
138 }
139 
140 void
141 dumpblock(blkno, size)
142 	daddr_t blkno;
143 	int size;
144 {
145 	int avail, tpblks, dblkno;
146 
147 	dblkno = fsbtodb(sblock, blkno);
148 	tpblks = size >> tp_bshift;
149 	while ((avail = MIN(tpblks, ntrec - trecno)) > 0) {
150 		slp->req[trecno].dblk = dblkno;
151 		slp->req[trecno].count = avail;
152 		trecno += avail;
153 		spcl.c_tapea += avail;
154 		if (trecno >= ntrec)
155 			flushtape();
156 		dblkno += avail << (tp_bshift - dev_bshift);
157 		tpblks -= avail;
158 	}
159 }
160 
161 int	nogripe = 0;
162 
163 void
164 tperror()
165 {
166 
167 	if (pipeout) {
168 		msg("write error on %s\n", tape);
169 		quit("Cannot recover\n");
170 		/* NOTREACHED */
171 	}
172 	msg("write error %d blocks into volume %d\n", blocksthisvol, tapeno);
173 	broadcast("DUMP WRITE ERROR!\n");
174 	if (!query("Do you want to restart?"))
175 		dumpabort();
176 	msg("Closing this volume.  Prepare to restart with new media;\n");
177 	msg("this dump volume will be rewritten.\n");
178 	killall();
179 	nogripe = 1;
180 	close_rewind();
181 	Exit(X_REWRITE);
182 }
183 
184 void
185 sigpipe()
186 {
187 
188 	quit("Broken pipe\n");
189 }
190 
191 void
192 flushtape()
193 {
194 	int i, blks, got;
195 	long lastfirstrec;
196 #ifndef __STDC__
197 	int write(), read();
198 #endif
199 
200 	int siz = (char *)nextblock - (char *)slp->req;
201 
202 	slp->req[trecno].count = 0;			/* Sentinel */
203 
204 	if (atomic(write, slp->fd, slp->req, siz) != siz)
205 		quit("error writing command pipe: %s\n", strerror(errno));
206 	slp->sent = 1; /* we sent a request, read the response later */
207 
208 	lastfirstrec = slp->firstrec;
209 
210 	if (++slp >= &slaves[SLAVES])
211 		slp = &slaves[0];
212 
213 	/* Read results back from next slave */
214 	if (slp->sent) {
215 		if (atomic(read, slp->fd, &got, sizeof got) != sizeof got) {
216 			perror("  DUMP: error reading command pipe in master");
217 			dumpabort();
218 		}
219 		slp->sent = 0;
220 
221 		/* Check for end of tape */
222 		if (got < writesize) {
223 			msg("End of tape detected\n");
224 
225 			/*
226 			 * Drain the results, don't care what the values were.
227 			 * If we read them here then trewind won't...
228 			 */
229 			for (i = 0; i < SLAVES; i++) {
230 				if (slaves[i].sent) {
231 					if (atomic(read, slaves[i].fd, &got,
232 					    sizeof got) != sizeof got) {
233 						perror("  DUMP: error reading command pipe in master");
234 						dumpabort();
235 					}
236 					slaves[i].sent = 0;
237 				}
238 			}
239 
240 			close_rewind();
241 			rollforward();
242 			return;
243 		}
244 	}
245 
246 	blks = 0;
247 	if (spcl.c_type != TS_END) {
248 		for (i = 0; i < spcl.c_count; i++)
249 			if (spcl.c_addr[i] != 0)
250 				blks++;
251 	}
252 	slp->count = lastspclrec + blks + 1 - spcl.c_tapea;
253 	slp->tapea = spcl.c_tapea;
254 	slp->firstrec = lastfirstrec + ntrec;
255 	slp->inode = curino;
256 	nextblock = slp->tblock;
257 	trecno = 0;
258 	asize += tenths;
259 	blockswritten += ntrec;
260 	blocksthisvol += ntrec;
261 	if (!pipeout && (blocksperfile ?
262 	    (blocksthisvol >= blocksperfile) : (asize > tsize))) {
263 		close_rewind();
264 		startnewtape(0);
265 	}
266 	timeest();
267 }
268 
269 void
270 trewind()
271 {
272 	int f;
273 	int got;
274 
275 	if (pipeout)
276 		return;
277 	for (f = 0; f < SLAVES; f++) {
278 		/*
279 		 * Drain the results, but unlike EOT we DO (or should) care
280 		 * what the return values were, since if we detect EOT after
281 		 * we think we've written the last blocks to the tape anyway,
282 		 * we have to replay those blocks with rollforward.
283 		 *
284 		 * fixme: punt for now.
285 		 */
286 		if (slaves[f].sent) {
287 			if (atomic(read, slaves[f].fd, &got, sizeof got)
288 			    != sizeof got) {
289 				perror("  DUMP: error reading command pipe in master");
290 				dumpabort();
291 			}
292 			slaves[f].sent = 0;
293 			if (got != writesize) {
294 				msg("EOT detected in last 2 tape records!\n");
295 				msg("Use a longer tape, decrease the size estimate\n");
296 				quit("or use no size estimate at all.\n");
297 			}
298 		}
299 		close(slaves[f].fd);
300 	}
301 	while (wait((int *)NULL) >= 0)	/* wait for any signals from slaves */
302 		/* void */;
303 	msg("Closing %s\n", tape);
304 
305 #ifdef RDUMP
306 	if (host) {
307 		rmtclose();
308 		while (rmtopen(tape, 0) < 0)
309 			sleep(10);
310 		rmtclose();
311 		return;
312 	}
313 #endif
314 	close(tapefd);
315 	while ((f = open(tape, 0)) < 0)
316 		sleep (10);
317 	close(f);
318 }
319 
320 void
321 close_rewind()
322 {
323 	trewind();
324 	if (nexttape)
325 		return;
326 	if (!nogripe) {
327 		msg("Change Volumes: Mount volume #%d\n", tapeno+1);
328 		broadcast("CHANGE DUMP VOLUMES!\7\7\n");
329 	}
330 	while (!query("Is the new volume mounted and ready to go?"))
331 		if (query("Do you want to abort?")) {
332 			dumpabort();
333 			/*NOTREACHED*/
334 		}
335 }
336 
337 #ifdef ROLLDEBUG
338 int do_sum(block)
339      union u_spcl *block;
340 
341 {
342 	char sum = 0;
343 	int i;
344 
345 	for (i = 0; i < TP_BSIZE; i++) {
346 		sum = sum ^ block->dummy[i];
347 	}
348 	return(sum);
349 }
350 #endif
351 
352 void
353 rollforward()
354 {
355 	register struct req *p, *q, *prev;
356 	register struct slave *tslp;
357 	int i, next, size, savedtapea, got;
358 	union u_spcl *ntb, *otb;
359 #ifdef ROLLDEBUG
360 	int j;
361 #endif
362 	tslp = &slaves[SLAVES];
363 	ntb = (union u_spcl *)tslp->tblock[1];
364 
365 	/*
366 	 * Each of the N slaves should have requests that need to
367 	 * be replayed on the next tape.  Use the extra slave buffers
368 	 * (slaves[SLAVES]) to construct request lists to be sent to
369 	 * each slave in turn.
370 	 */
371 	for (i = 0; i < SLAVES; i++) {
372 		q = &tslp->req[1];
373 		otb = (union u_spcl *)slp->tblock;
374 
375 		/*
376 		 * For each request in the current slave, copy it to tslp.
377 		 */
378 #ifdef ROLLDEBUG
379 		printf("replaying reqs to slave %d (%d)\n", slp - &slaves[0],
380 		    slp->pid);
381 		j = 0;
382 #endif
383 
384 		for (p = slp->req; p->count > 0; p += p->count) {
385 #ifdef ROLLDEBUG
386 			printf("    req %d count %d dblk %d\n",
387 			       j++, p->count, p->dblk);
388 			if (p->dblk == 0)
389 				printf("\tsum %x\n", do_sum(otb));
390 #endif
391 			*q = *p;
392 			if (p->dblk == 0)
393 				*ntb++ = *otb++; /* copy the datablock also */
394 			prev = q;
395 			q += q->count;
396 		}
397 		if (prev->dblk != 0)
398 			prev->count -= 1;
399 		else
400 			ntb--;
401 		q -= 1;
402 		q->count = 0;
403 		q = &tslp->req[0];
404 		if (i == 0) {
405 			q->dblk = 0;
406 			q->count = 1;
407 			trecno = 0;
408 			nextblock = tslp->tblock;
409 			savedtapea = spcl.c_tapea;
410 			spcl.c_tapea = slp->tapea;
411 			startnewtape(0);
412 			spcl.c_tapea = savedtapea;
413 			lastspclrec = savedtapea - 1;
414 		}
415 		size = (char *)ntb - (char *)q;
416 		if (atomic(write, slp->fd, q, size) != size) {
417 			perror("  DUMP: error writing command pipe");
418 			dumpabort();
419 		}
420 		slp->sent = 1;
421 #ifdef ROLLDEBUG
422 		printf("after the shift:\n");
423 		j = 0;
424 		for (p = tslp->req; p->count > 0; p += p->count) {
425 			printf("    req %d count %d dblk %d\n",
426 			       j++, p->count, p->dblk);
427 			if (p->dblk == 0) {
428 				/* dump block also */
429 			}
430 		}
431 #endif
432 		if (++slp >= &slaves[SLAVES])
433 			slp = &slaves[0];
434 
435 		q->count = 1;
436 
437 		if (prev->dblk != 0) {
438 			/*
439 			 * If the last one was a disk block, make the
440 			 * first of this one be the last bit of that disk
441 			 * block...
442 			 */
443 			q->dblk = prev->dblk +
444 				prev->count * (TP_BSIZE / DEV_BSIZE);
445 			ntb = (union u_spcl *)tslp->tblock;
446 		} else {
447 			/*
448 			 * It wasn't a disk block.  Copy the data to its
449 			 * new location in the buffer.
450 			 */
451 			q->dblk = 0;
452 			*((union u_spcl *)tslp->tblock) = *ntb;
453 			ntb = (union u_spcl *)tslp->tblock[1];
454 		}
455 	}
456 	slp->req[0] = *q;
457 	nextblock = slp->tblock;
458 	if (q->dblk == 0)
459 		nextblock++;
460 	trecno = 1;
461 
462 	/*
463 	 * Clear the first slaves' response.  One hopes that it
464 	 * worked ok, otherwise the tape is much too short!
465 	 */
466 	if (slp->sent) {
467 		if (atomic(read, slp->fd, &got, sizeof got) != sizeof got) {
468 			perror("  DUMP: error reading command pipe in master");
469 			dumpabort();
470 		}
471 		slp->sent = 0;
472 
473 		if (got != writesize) {
474 			quit("EOT detected at start of the tape!\n");
475 		}
476 	}
477 }
478 
479 /*
480  * We implement taking and restoring checkpoints on the tape level.
481  * When each tape is opened, a new process is created by forking; this
482  * saves all of the necessary context in the parent.  The child
483  * continues the dump; the parent waits around, saving the context.
484  * If the child returns X_REWRITE, then it had problems writing that tape;
485  * this causes the parent to fork again, duplicating the context, and
486  * everything continues as if nothing had happened.
487  */
488 void
489 startnewtape(top)
490 	int top;
491 {
492 	int	parentpid;
493 	int	childpid;
494 	int	status;
495 	int	waitpid;
496 	sig_t	interrupt;
497 	int	i;
498 	char	*p;
499 
500 	interrupt = signal(SIGINT, SIG_IGN);
501 	parentpid = getpid();
502 
503     restore_check_point:
504 	(void)signal(SIGINT, interrupt);
505 	/*
506 	 *	All signals are inherited...
507 	 */
508 	childpid = fork();
509 	if (childpid < 0) {
510 		msg("Context save fork fails in parent %d\n", parentpid);
511 		Exit(X_ABORT);
512 	}
513 	if (childpid != 0) {
514 		/*
515 		 *	PARENT:
516 		 *	save the context by waiting
517 		 *	until the child doing all of the work returns.
518 		 *	don't catch the interrupt
519 		 */
520 		signal(SIGINT, SIG_IGN);
521 #ifdef TDEBUG
522 		msg("Tape: %d; parent process: %d child process %d\n",
523 			tapeno+1, parentpid, childpid);
524 #endif TDEBUG
525 		while ((waitpid = wait(&status)) != childpid)
526 			msg("Parent %d waiting for child %d has another child %d return\n",
527 				parentpid, childpid, waitpid);
528 		if (status & 0xFF) {
529 			msg("Child %d returns LOB status %o\n",
530 				childpid, status&0xFF);
531 		}
532 		status = (status >> 8) & 0xFF;
533 #ifdef TDEBUG
534 		switch(status) {
535 			case X_FINOK:
536 				msg("Child %d finishes X_FINOK\n", childpid);
537 				break;
538 			case X_ABORT:
539 				msg("Child %d finishes X_ABORT\n", childpid);
540 				break;
541 			case X_REWRITE:
542 				msg("Child %d finishes X_REWRITE\n", childpid);
543 				break;
544 			default:
545 				msg("Child %d finishes unknown %d\n",
546 					childpid, status);
547 				break;
548 		}
549 #endif TDEBUG
550 		switch(status) {
551 			case X_FINOK:
552 				Exit(X_FINOK);
553 			case X_ABORT:
554 				Exit(X_ABORT);
555 			case X_REWRITE:
556 				goto restore_check_point;
557 			default:
558 				msg("Bad return code from dump: %d\n", status);
559 				Exit(X_ABORT);
560 		}
561 		/*NOTREACHED*/
562 	} else {	/* we are the child; just continue */
563 #ifdef TDEBUG
564 		sleep(4);	/* allow time for parent's message to get out */
565 		msg("Child on Tape %d has parent %d, my pid = %d\n",
566 			tapeno+1, parentpid, getpid());
567 #endif TDEBUG
568 		/*
569 		 * If we have a name like "/dev/rmt0,/dev/rmt1",
570 		 * use the name before the comma first, and save
571 		 * the remaining names for subsequent volumes.
572 		 */
573 		tapeno++;               /* current tape sequence */
574 		if (nexttape || index(tape, ',')) {
575 			if (nexttape && *nexttape)
576 				tape = nexttape;
577 			if (p = index(tape, ',')) {
578 				*p = '\0';
579 				nexttape = p + 1;
580 			} else
581 				nexttape = NULL;
582 			msg("Dumping volume %d on %s\n", tapeno, tape);
583 		}
584 #ifdef RDUMP
585 		while ((tapefd = (host ? rmtopen(tape, 2) :
586 			pipeout ? 1 : open(tape, O_WRONLY|O_CREAT, 0666))) < 0)
587 #else
588 		while ((tapefd = (pipeout ? 1 :
589 				  open(tape, O_WRONLY|O_CREAT, 0666))) < 0)
590 #endif
591 		    {
592 			msg("Cannot open output \"%s\".\n", tape);
593 			if (!query("Do you want to retry the open?"))
594 				dumpabort();
595 		}
596 
597 		enslave();  /* Share open tape file descriptor with slaves */
598 
599 		asize = 0;
600 		blocksthisvol = 0;
601 		if (top)
602 			newtape++;		/* new tape signal */
603 		spcl.c_count = slp->count;
604 		/*
605 		 * measure firstrec in TP_BSIZE units since restore doesn't
606 		 * know the correct ntrec value...
607 		 */
608 		spcl.c_firstrec = slp->firstrec;
609 		spcl.c_volume++;
610 		spcl.c_type = TS_TAPE;
611 		spcl.c_flags |= DR_NEWHEADER;
612 		writeheader(slp->inode);
613 		spcl.c_flags &=~ DR_NEWHEADER;
614 		if (tapeno > 1)
615 			msg("Volume %d begins with blocks from inode %d\n",
616 				tapeno, slp->inode);
617 	}
618 }
619 
620 void
621 dumpabort()
622 {
623 
624 	if (master != 0 && master != getpid())
625 		kill(master, SIGTERM);	/* Signals master to call dumpabort */
626 	else {
627 		killall();
628 		msg("The ENTIRE dump is aborted.\n");
629 	}
630 	Exit(X_ABORT);
631 }
632 
633 void
634 Exit(status)
635 	int status;
636 {
637 
638 #ifdef TDEBUG
639 	msg("pid = %d exits with status %d\n", getpid(), status);
640 #endif TDEBUG
641 	exit(status);
642 }
643 
644 /*
645  * proceed - handler for SIGUSR2, used to synchronize IO between the slaves.
646  */
647 void
648 proceed()
649 {
650 
651 	if (ready)
652 		longjmp(jmpbuf, 1);
653 	caught++;
654 }
655 
656 void
657 enslave()
658 {
659 	int cmd[2];
660 	register int i, j;
661 
662 	master = getpid();
663 
664 	signal(SIGTERM, dumpabort);  /* Slave sends SIGTERM on dumpabort() */
665 	signal(SIGPIPE, sigpipe);
666 	signal(SIGUSR1, tperror);    /* Slave sends SIGUSR1 on tape errors */
667 	signal(SIGUSR2, proceed);    /* Slave sends SIGUSR2 to next slave */
668 
669 	for (i = 0; i < SLAVES; i++) {
670 		if (i == slp - &slaves[0]) {
671 			caught = 1;
672 		} else {
673 			caught = 0;
674 		}
675 
676 		if (socketpair(AF_UNIX, SOCK_STREAM, 0, cmd) < 0 ||
677 		    (slaves[i].pid = fork()) < 0)
678 			quit("too many slaves, %d (recompile smaller): %s\n",
679 			    i, strerror(errno));
680 
681 		slaves[i].fd = cmd[1];
682 		slaves[i].sent = 0;
683 		if (slaves[i].pid == 0) { 	    /* Slave starts up here */
684 			for (j = 0; j <= i; j++)
685 			        close(slaves[j].fd);
686 			signal(SIGINT, SIG_IGN);    /* Master handles this */
687 			doslave(cmd[0], i);
688 			Exit(X_FINOK);
689 		}
690 	}
691 
692 	for (i = 0; i < SLAVES; i++)
693 		atomic(write, slaves[i].fd, &slaves[(i + 1) % SLAVES].pid,
694 		       sizeof slaves[0].pid);
695 
696 	master = 0;
697 }
698 
699 void
700 killall()
701 {
702 	register int i;
703 
704 	for (i = 0; i < SLAVES; i++)
705 		if (slaves[i].pid > 0)
706 			kill(slaves[i].pid, SIGKILL);
707 }
708 
709 /*
710  * Synchronization - each process has a lockfile, and shares file
711  * descriptors to the following process's lockfile.  When our write
712  * completes, we release our lock on the following process's lock-
713  * file, allowing the following process to lock it and proceed. We
714  * get the lock back for the next cycle by swapping descriptors.
715  */
716 void
717 doslave(cmd, slave_number)
718 	register int cmd;
719         int slave_number;
720 {
721 	register int nread;
722 	int nextslave, size, wrote, eot_count;
723 #ifndef __STDC__
724 	int read();
725 #endif
726 #ifdef ROLLDEBUG
727 	int dodump = 2;
728 	FILE *out;
729 	char name[64];
730 #endif
731 
732 	/*
733 	 * Need our own seek pointer.
734 	 */
735 	close(diskfd);
736 	if ((diskfd = open(disk, O_RDONLY)) < 0)
737 		quit("slave couldn't reopen disk: %s\n", strerror(errno));
738 
739 	/*
740 	 * Need the pid of the next slave in the loop...
741 	 */
742 	if ((nread = atomic(read, cmd, &nextslave, sizeof nextslave))
743 	    != sizeof nextslave) {
744 		quit("master/slave protocol botched - didn't get pid of next slave.\n");
745 	}
746 
747 #ifdef ROLLDEBUG
748 	sprintf(name, "slave.%d", slave_number);
749 	out = fopen(name, "w");
750 #endif
751 	/*
752 	 * Get list of blocks to dump, read the blocks into tape buffer
753 	 */
754 	while ((nread = atomic(read, cmd, slp->req, reqsiz)) == reqsiz) {
755 		register struct req *p = slp->req;
756 		int j;
757 		struct req *rover;
758 		char (*orover)[TP_BSIZE];
759 
760 		j = 0;
761 		for (trecno = 0; trecno < ntrec;
762 		     trecno += p->count, p += p->count) {
763 			if (p->dblk) {
764 				bread(p->dblk, slp->tblock[trecno],
765 					p->count * TP_BSIZE);
766 			} else {
767 				if (p->count != 1 || atomic(read, cmd,
768 				    slp->tblock[trecno], TP_BSIZE) != TP_BSIZE)
769 					quit("master/slave protocol botched.\n");
770 			}
771 #ifdef ROLLDEBUG
772 			if (dodump) {
773 				fprintf(out, "    req %d count %d dblk %d\n",
774 					j++, p->count, p->dblk);
775 				if (p->dblk == 0) {
776 					fprintf(out, "\tsum %x\n",
777 						do_sum(slp->tblock[trecno]));
778 				}
779 			}
780 #endif
781 		}
782 #ifdef ROLLDEBUG
783 		if (dodump) {
784 			fprintf(out, "\n");
785 		}
786 		if (--dodump == 0) {
787 			fclose(out);
788 		}
789 #endif
790 		if (setjmp(jmpbuf) == 0) {
791 			ready = 1;
792 			if (!caught)
793 				pause();
794 		}
795 		ready = 0;
796 		caught = 0;
797 
798 		/* Try to write the data... */
799 		eot_count = 0;
800 		size = 0;
801 
802 		while (eot_count < 10 && size < writesize) {
803 #ifdef RDUMP
804 			if (host)
805 				wrote = rmtwrite(slp->tblock[0]+size,
806 				    writesize-size);
807 			else
808 #endif
809 				wrote = write(tapefd, slp->tblock[0]+size,
810 				    writesize-size);
811 #ifdef WRITEDEBUG
812 			printf("slave %d wrote %d\n", slave_number, wrote);
813 #endif
814 			if (wrote < 0)
815 				break;
816 			if (wrote == 0)
817 				eot_count++;
818 			size += wrote;
819 		}
820 
821 #ifdef WRITEDEBUG
822 		if (size != writesize)
823 		 printf("slave %d only wrote %d out of %d bytes and gave up.\n",
824 		     slave_number, size, writesize);
825 #endif
826 
827 		if (eot_count > 0)
828 			size = 0;
829 
830 		/*
831 		 * fixme: Pyramids running OSx return ENOSPC
832 		 * at EOT on 1/2 inch drives.
833 		 */
834 		if (size < 0) {
835 			kill(master, SIGUSR1);
836 			for (;;)
837 				sigpause(0);
838 		} else {
839 			/*
840 			 * pass size of write back to master
841 			 * (for EOT handling)
842 			 */
843 			atomic(write, cmd, &size, sizeof size);
844 		}
845 
846 		/*
847 		 * If partial write, don't want next slave to go.
848 		 * Also jolts him awake.
849 		 */
850 		kill(nextslave, SIGUSR2);
851 	}
852 	if (nread != 0)
853 		quit("error reading command pipe: %s\n", strerror(errno));
854 }
855 
856 /*
857  * Since a read from a pipe may not return all we asked for,
858  * or a write may not write all we ask if we get a signal,
859  * loop until the count is satisfied (or error).
860  */
861 int
862 atomic(func, fd, buf, count)
863 	int (*func)(), fd, count;
864 	char *buf;
865 {
866 	int got, need = count;
867 
868 	while ((got = (*func)(fd, buf, need)) > 0 && (need -= got) > 0)
869 		buf += got;
870 	return (got < 0 ? got : count - need);
871 }
872