1/*-
2 * Copyright (c) 1980, 1991, 1993
3 *	The Regents of the University of California.  All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
8 * 1. Redistributions of source code must retain the above copyright
9 *    notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 *    notice, this list of conditions and the following disclaimer in the
12 *    documentation and/or other materials provided with the distribution.
13 * 4. Neither the name of the University nor the names of its contributors
14 *    may be used to endorse or promote products derived from this software
15 *    without specific prior written permission.
16 *
17 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27 * SUCH DAMAGE.
28 */
29
30#ifndef lint
31#if 0
32static char sccsid[] = "@(#)tape.c	8.4 (Berkeley) 5/1/95";
33#endif
34static const char rcsid[] =
35  "$FreeBSD$";
36#endif /* not lint */
37
38#include <sys/param.h>
39#include <sys/socket.h>
40#include <sys/wait.h>
41#include <sys/stat.h>
42
43#include <ufs/ufs/dinode.h>
44#include <ufs/ffs/fs.h>
45
46#include <protocols/dumprestore.h>
47
48#include <errno.h>
49#include <fcntl.h>
50#include <limits.h>
51#include <setjmp.h>
52#include <signal.h>
53#include <stdio.h>
54#include <stdlib.h>
55#include <string.h>
56#include <time.h>
57#include <unistd.h>
58
59#include "dump.h"
60
61int	writesize;		/* size of malloc()ed buffer for tape */
62int64_t	lastspclrec = -1;	/* tape block number of last written header */
63int	trecno = 0;		/* next record to write in current block */
64extern	long blocksperfile;	/* number of blocks per output file */
65long	blocksthisvol;		/* number of blocks on current output file */
66extern	int ntrec;		/* blocking factor on tape */
67extern	int cartridge;
68extern	char *host;
69char	*nexttape;
70FILE	*popenfp = NULL;
71
72static	int atomic(ssize_t (*)(), int, char *, int);
73static	void doslave(int, int);
74static	void enslave(void);
75static	void flushtape(void);
76static	void killall(void);
77static	void rollforward(void);
78
79/*
80 * Concurrent dump mods (Caltech) - disk block reading and tape writing
81 * are exported to several slave processes.  While one slave writes the
82 * tape, the others read disk blocks; they pass control of the tape in
83 * a ring via signals. The parent process traverses the file system and
84 * sends writeheader()'s and lists of daddr's to the slaves via pipes.
85 * The following structure defines the instruction packets sent to slaves.
86 */
87struct req {
88	ufs2_daddr_t dblk;
89	int count;
90};
91int reqsiz;
92
93#define SLAVES 3		/* 1 slave writing, 1 reading, 1 for slack */
94struct slave {
95	int64_t tapea;		/* header number at start of this chunk */
96	int64_t firstrec;	/* record number of this block */
97	int count;		/* count to next header (used for TS_TAPE */
98				/* after EOT) */
99	int inode;		/* inode that we are currently dealing with */
100	int fd;			/* FD for this slave */
101	int pid;		/* PID for this slave */
102	int sent;		/* 1 == we've sent this slave requests */
103	char (*tblock)[TP_BSIZE]; /* buffer for data blocks */
104	struct req *req;	/* buffer for requests */
105} slaves[SLAVES+1];
106struct slave *slp;
107
108char	(*nextblock)[TP_BSIZE];
109
110int master;		/* pid of master, for sending error signals */
111int tenths;		/* length of tape used per block written */
112static volatile sig_atomic_t caught; /* have we caught the signal to proceed? */
113static volatile sig_atomic_t ready; /* reached the lock point without having */
114			/* received the SIGUSR2 signal from the prev slave? */
115static jmp_buf jmpbuf;	/* where to jump to if we are ready when the */
116			/* SIGUSR2 arrives from the previous slave */
117
118int
119alloctape(void)
120{
121	int pgoff = getpagesize() - 1;
122	char *buf;
123	int i;
124
125	writesize = ntrec * TP_BSIZE;
126	reqsiz = (ntrec + 1) * sizeof(struct req);
127	/*
128	 * CDC 92181's and 92185's make 0.8" gaps in 1600-bpi start/stop mode
129	 * (see DEC TU80 User's Guide).  The shorter gaps of 6250-bpi require
130	 * repositioning after stopping, i.e, streaming mode, where the gap is
131	 * variable, 0.30" to 0.45".  The gap is maximal when the tape stops.
132	 */
133	if (blocksperfile == 0 && !unlimited)
134		tenths = writesize / density +
135		    (cartridge ? 16 : density == 625 ? 5 : 8);
136	/*
137	 * Allocate tape buffer contiguous with the array of instruction
138	 * packets, so flushtape() can write them together with one write().
139	 * Align tape buffer on page boundary to speed up tape write().
140	 */
141	for (i = 0; i <= SLAVES; i++) {
142		buf = (char *)
143		    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
158void
159writerec(char *dp, int isspcl)
160{
161
162	slp->req[trecno].dblk = (ufs2_daddr_t)0;
163	slp->req[trecno].count = 1;
164	/* Can't do a structure assignment due to alignment problems */
165	bcopy(dp, *(nextblock)++, sizeof (union u_spcl));
166	if (isspcl)
167		lastspclrec = spcl.c_tapea;
168	trecno++;
169	spcl.c_tapea++;
170	if (trecno >= ntrec)
171		flushtape();
172}
173
174void
175dumpblock(ufs2_daddr_t blkno, int size)
176{
177	int avail, tpblks;
178	ufs2_daddr_t dblkno;
179
180	dblkno = fsbtodb(sblock, blkno);
181	tpblks = size >> tp_bshift;
182	while ((avail = MIN(tpblks, ntrec - trecno)) > 0) {
183		slp->req[trecno].dblk = dblkno;
184		slp->req[trecno].count = avail;
185		trecno += avail;
186		spcl.c_tapea += avail;
187		if (trecno >= ntrec)
188			flushtape();
189		dblkno += avail << (tp_bshift - dev_bshift);
190		tpblks -= avail;
191	}
192}
193
194int	nogripe = 0;
195
196void
197tperror(int signo __unused)
198{
199
200	if (pipeout) {
201		msg("write error on %s\n", tape);
202		quit("Cannot recover\n");
203		/* NOTREACHED */
204	}
205	msg("write error %ld blocks into volume %d\n", blocksthisvol, tapeno);
206	broadcast("DUMP WRITE ERROR!\n");
207	if (!query("Do you want to restart?"))
208		dumpabort(0);
209	msg("Closing this volume.  Prepare to restart with new media;\n");
210	msg("this dump volume will be rewritten.\n");
211	killall();
212	nogripe = 1;
213	close_rewind();
214	Exit(X_REWRITE);
215}
216
217void
218sigpipe(int signo __unused)
219{
220
221	quit("Broken pipe\n");
222}
223
224static void
225flushtape(void)
226{
227	int i, blks, got;
228	int64_t lastfirstrec;
229
230	int siz = (char *)nextblock - (char *)slp->req;
231
232	slp->req[trecno].count = 0;			/* Sentinel */
233
234	if (atomic(write, slp->fd, (char *)slp->req, siz) != siz)
235		quit("error writing command pipe: %s\n", strerror(errno));
236	slp->sent = 1; /* we sent a request, read the response later */
237
238	lastfirstrec = slp->firstrec;
239
240	if (++slp >= &slaves[SLAVES])
241		slp = &slaves[0];
242
243	/* Read results back from next slave */
244	if (slp->sent) {
245		if (atomic(read, slp->fd, (char *)&got, sizeof got)
246		    != sizeof got) {
247			perror("  DUMP: error reading command pipe in master");
248			dumpabort(0);
249		}
250		slp->sent = 0;
251
252		/* Check for end of tape */
253		if (got < writesize) {
254			msg("End of tape detected\n");
255
256			/*
257			 * Drain the results, don't care what the values were.
258			 * If we read them here then trewind won't...
259			 */
260			for (i = 0; i < SLAVES; i++) {
261				if (slaves[i].sent) {
262					if (atomic(read, slaves[i].fd,
263					    (char *)&got, sizeof got)
264					    != sizeof got) {
265						perror("  DUMP: error reading command pipe in master");
266						dumpabort(0);
267					}
268					slaves[i].sent = 0;
269				}
270			}
271
272			close_rewind();
273			rollforward();
274			return;
275		}
276	}
277
278	blks = 0;
279	if (spcl.c_type != TS_END) {
280		for (i = 0; i < spcl.c_count; i++)
281			if (spcl.c_addr[i] != 0)
282				blks++;
283	}
284	slp->count = lastspclrec + blks + 1 - spcl.c_tapea;
285	slp->tapea = spcl.c_tapea;
286	slp->firstrec = lastfirstrec + ntrec;
287	slp->inode = curino;
288	nextblock = slp->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
301void
302trewind(void)
303{
304	struct stat sb;
305	int f;
306	int got;
307
308	for (f = 0; f < SLAVES; 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 (slaves[f].sent) {
318			if (atomic(read, slaves[f].fd, (char *)&got, sizeof got)
319			    != sizeof got) {
320				perror("  DUMP: error reading command pipe in master");
321				dumpabort(0);
322			}
323			slaves[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(slaves[f].fd);
331	}
332	while (wait((int *)NULL) >= 0)	/* wait for any signals from slaves */
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
365void
366close_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
388void
389rollforward(void)
390{
391	struct req *p, *q, *prev;
392	struct slave *tslp;
393	int i, size, got;
394	int64_t savedtapea;
395	union u_spcl *ntb, *otb;
396	tslp = &slaves[SLAVES];
397	ntb = (union u_spcl *)tslp->tblock[1];
398
399	/*
400	 * Each of the N slaves should have requests that need to
401	 * be replayed on the next tape.  Use the extra slave buffers
402	 * (slaves[SLAVES]) to construct request lists to be sent to
403	 * each slave in turn.
404	 */
405	for (i = 0; i < SLAVES; i++) {
406		q = &tslp->req[1];
407		otb = (union u_spcl *)slp->tblock;
408
409		/*
410		 * For each request in the current slave, copy it to tslp.
411		 */
412
413		prev = NULL;
414		for (p = slp->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 = &tslp->req[0];
430		if (i == 0) {
431			q->dblk = 0;
432			q->count = 1;
433			trecno = 0;
434			nextblock = tslp->tblock;
435			savedtapea = spcl.c_tapea;
436			spcl.c_tapea = slp->tapea;
437			startnewtape(0);
438			spcl.c_tapea = savedtapea;
439			lastspclrec = savedtapea - 1;
440		}
441		size = (char *)ntb - (char *)q;
442		if (atomic(write, slp->fd, (char *)q, size) != size) {
443			perror("  DUMP: error writing command pipe");
444			dumpabort(0);
445		}
446		slp->sent = 1;
447		if (++slp >= &slaves[SLAVES])
448			slp = &slaves[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 *)tslp->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 *)tslp->tblock) = *ntb;
468			ntb = (union u_spcl *)tslp->tblock[1];
469		}
470	}
471	slp->req[0] = *q;
472	nextblock = slp->tblock;
473	if (q->dblk == 0)
474		nextblock++;
475	trecno = 1;
476
477	/*
478	 * Clear the first slaves' response.  One hopes that it
479	 * worked ok, otherwise the tape is much too short!
480	 */
481	if (slp->sent) {
482		if (atomic(read, slp->fd, (char *)&got, sizeof got)
483		    != sizeof got) {
484			perror("  DUMP: error reading command pipe in master");
485			dumpabort(0);
486		}
487		slp->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 */
504void
505startnewtape(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
516restore_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		enslave();  /* Share open tape file descriptor with slaves */
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 = slp->count;
640		/*
641		 * measure firstrec in TP_BSIZE units since restore doesn't
642		 * know the correct ntrec value...
643		 */
644		spcl.c_firstrec = slp->firstrec;
645		spcl.c_volume++;
646		spcl.c_type = TS_TAPE;
647		writeheader((ino_t)slp->inode);
648		if (tapeno > 1)
649			msg("Volume %d begins with blocks from inode %d\n",
650				tapeno, slp->inode);
651	}
652}
653
654void
655dumpabort(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
671void
672Exit(status)
673	int status;
674{
675
676#ifdef TDEBUG
677	msg("pid = %d exits with status %d\n", getpid(), status);
678#endif /* TDEBUG */
679	exit(status);
680}
681
682/*
683 * proceed - handler for SIGUSR2, used to synchronize IO between the slaves.
684 */
685void
686proceed(int signo __unused)
687{
688
689	if (ready)
690		longjmp(jmpbuf, 1);
691	caught++;
692}
693
694void
695enslave(void)
696{
697	int cmd[2];
698	int i, j;
699
700	master = getpid();
701
702	signal(SIGTERM, dumpabort);  /* Slave sends SIGTERM on dumpabort() */
703	signal(SIGPIPE, sigpipe);
704	signal(SIGUSR1, tperror);    /* Slave sends SIGUSR1 on tape errors */
705	signal(SIGUSR2, proceed);    /* Slave sends SIGUSR2 to next slave */
706
707	for (i = 0; i < SLAVES; i++) {
708		if (i == slp - &slaves[0]) {
709			caught = 1;
710		} else {
711			caught = 0;
712		}
713
714		if (socketpair(AF_UNIX, SOCK_STREAM, 0, cmd) < 0 ||
715		    (slaves[i].pid = fork()) < 0)
716			quit("too many slaves, %d (recompile smaller): %s\n",
717			    i, strerror(errno));
718
719		slaves[i].fd = cmd[1];
720		slaves[i].sent = 0;
721		if (slaves[i].pid == 0) { 	    /* Slave starts up here */
722			for (j = 0; j <= i; j++)
723			        (void) close(slaves[j].fd);
724			signal(SIGINT, SIG_IGN);    /* Master handles this */
725			doslave(cmd[0], i);
726			Exit(X_FINOK);
727		}
728	}
729
730	for (i = 0; i < SLAVES; i++)
731		(void) atomic(write, slaves[i].fd,
732			      (char *) &slaves[(i + 1) % SLAVES].pid,
733		              sizeof slaves[0].pid);
734
735	master = 0;
736}
737
738void
739killall(void)
740{
741	int i;
742
743	for (i = 0; i < SLAVES; i++)
744		if (slaves[i].pid > 0) {
745			(void) kill(slaves[i].pid, SIGKILL);
746			slaves[i].sent = 0;
747		}
748}
749
750/*
751 * Synchronization - each process has a lockfile, and shares file
752 * descriptors to the following process's lockfile.  When our write
753 * completes, we release our lock on the following process's lock-
754 * file, allowing the following process to lock it and proceed. We
755 * get the lock back for the next cycle by swapping descriptors.
756 */
757static void
758doslave(int cmd, int slave_number)
759{
760	int nread;
761	int nextslave, size, wrote, eot_count;
762
763	/*
764	 * Need our own seek pointer.
765	 */
766	(void) close(diskfd);
767	if ((diskfd = open(disk, O_RDONLY)) < 0)
768		quit("slave couldn't reopen disk: %s\n", strerror(errno));
769
770	/*
771	 * Need the pid of the next slave in the loop...
772	 */
773	if ((nread = atomic(read, cmd, (char *)&nextslave, sizeof nextslave))
774	    != sizeof nextslave) {
775		quit("master/slave protocol botched - didn't get pid of next slave.\n");
776	}
777
778	/*
779	 * Get list of blocks to dump, read the blocks into tape buffer
780	 */
781	while ((nread = atomic(read, cmd, (char *)slp->req, reqsiz)) == reqsiz) {
782		struct req *p = slp->req;
783
784		for (trecno = 0; trecno < ntrec;
785		     trecno += p->count, p += p->count) {
786			if (p->dblk) {
787				bread(p->dblk, slp->tblock[trecno],
788					p->count * TP_BSIZE);
789			} else {
790				if (p->count != 1 || atomic(read, cmd,
791				    (char *)slp->tblock[trecno],
792				    TP_BSIZE) != TP_BSIZE)
793				       quit("master/slave protocol botched.\n");
794			}
795		}
796		if (setjmp(jmpbuf) == 0) {
797			ready = 1;
798			if (!caught)
799				(void) pause();
800		}
801		ready = 0;
802		caught = 0;
803
804		/* Try to write the data... */
805		eot_count = 0;
806		size = 0;
807
808		wrote = 0;
809		while (eot_count < 10 && size < writesize) {
810#ifdef RDUMP
811			if (host)
812				wrote = rmtwrite(slp->tblock[0]+size,
813				    writesize-size);
814			else
815#endif
816				wrote = write(tapefd, slp->tblock[0]+size,
817				    writesize-size);
818#ifdef WRITEDEBUG
819			printf("slave %d wrote %d\n", slave_number, wrote);
820#endif
821			if (wrote < 0)
822				break;
823			if (wrote == 0)
824				eot_count++;
825			size += wrote;
826		}
827
828#ifdef WRITEDEBUG
829		if (size != writesize)
830		 printf("slave %d only wrote %d out of %d bytes and gave up.\n",
831		     slave_number, size, writesize);
832#endif
833
834		/*
835		 * Handle ENOSPC as an EOT condition.
836		 */
837		if (wrote < 0 && errno == ENOSPC) {
838			wrote = 0;
839			eot_count++;
840		}
841
842		if (eot_count > 0)
843			size = 0;
844
845		if (wrote < 0) {
846			(void) kill(master, SIGUSR1);
847			for (;;)
848				(void) sigpause(0);
849		} else {
850			/*
851			 * pass size of write back to master
852			 * (for EOT handling)
853			 */
854			(void) atomic(write, cmd, (char *)&size, sizeof size);
855		}
856
857		/*
858		 * If partial write, don't want next slave to go.
859		 * Also jolts him awake.
860		 */
861		(void) kill(nextslave, 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 * or a write may not write all we ask if we get a signal,
870 * loop until the count is satisfied (or error).
871 */
872static int
873atomic(ssize_t (*func)(), int fd, char *buf, int count)
874{
875	int got, need = count;
876
877	while ((got = (*func)(fd, buf, need)) > 0 && (need -= got) > 0)
878		buf += got;
879	return (got < 0 ? got : count - need);
880}
881