Home | History | Annotate | Line # | Download | only in dump
tape.c revision 1.2
      1 /*-
      2  * Copyright (c) 1980, 1991 The Regents of the University of California.
      3  * 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  * 3. All advertising materials mentioning features or use of this software
     14  *    must display the following acknowledgement:
     15  *	This product includes software developed by the University of
     16  *	California, Berkeley and its contributors.
     17  * 4. Neither the name of the University nor the names of its contributors
     18  *    may be used to endorse or promote products derived from this software
     19  *    without specific prior written permission.
     20  *
     21  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
     22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
     23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
     24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
     25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
     26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
     27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
     28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
     29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
     30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
     31  * SUCH DAMAGE.
     32  */
     33 
     34 #ifndef lint
     35 /* from: static char sccsid[] = "@(#)tape.c	5.25 (Berkeley) 7/16/92"; */
     36 static char *rcsid = "$Id: tape.c,v 1.2 1994/03/09 01:14:43 cgd Exp $";
     37 #endif /* not lint */
     38 
     39 #ifdef sunos
     40 #include <sys/param.h>
     41 #include <stdio.h>
     42 #include <ctype.h>
     43 #include <sys/stat.h>
     44 #include <ufs/fs.h>
     45 #else
     46 #include <sys/param.h>
     47 #include <sys/wait.h>
     48 #include <ufs/fs.h>
     49 #endif
     50 #include <sys/time.h>
     51 #include <ufs/dinode.h>
     52 #include <signal.h>
     53 #include <fcntl.h>
     54 #include <protocols/dumprestore.h>
     55 #include <errno.h>
     56 #include <setjmp.h>
     57 #ifdef __STDC__
     58 #include <unistd.h>
     59 #include <stdlib.h>
     60 #include <string.h>
     61 #endif
     62 #include <sys/socket.h>
     63 #include "dump.h"
     64 #include "pathnames.h"
     65 
     66 int	writesize;		/* size of malloc()ed buffer for tape */
     67 long	lastspclrec = -1;	/* tape block number of last written header */
     68 int	trecno = 0;		/* next record to write in current block */
     69 extern	long blocksperfile;	/* number of blocks per output file */
     70 long	blocksthisvol;		/* number of blocks on current output file */
     71 extern	int ntrec;		/* blocking factor on tape */
     72 extern	int cartridge;
     73 extern	char *host;
     74 char	*nexttape;
     75 #ifdef RDUMP
     76 int	rmtopen(), rmtwrite();
     77 void	rmtclose();
     78 #endif
     79 void	rollforward();
     80 int	atomic();
     81 void	doslave(), enslave(), flushtape(), killall();
     82 
     83 /*
     84  * Concurrent dump mods (Caltech) - disk block reading and tape writing
     85  * are exported to several slave processes.  While one slave writes the
     86  * tape, the others read disk blocks; they pass control of the tape in
     87  * a ring via signals. The parent process traverses the filesystem and
     88  * sends writeheader()'s and lists of daddr's to the slaves via pipes.
     89  * The following structure defines the instruction packets sent to slaves.
     90  */
     91 struct req {
     92 	daddr_t dblk;
     93 	int count;
     94 };
     95 int reqsiz;
     96 
     97 #define SLAVES 3		/* 1 slave writing, 1 reading, 1 for slack */
     98 struct slave {
     99 	int tapea;		/* header number at start of this chunk */
    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 	int firstrec;		/* record number of this block */
    107 	char (*tblock)[TP_BSIZE]; /* buffer for data blocks */
    108 	struct req *req;	/* buffer for requests */
    109 } slaves[SLAVES+1];
    110 struct slave *slp;
    111 
    112 char	(*nextblock)[TP_BSIZE];
    113 
    114 int master;		/* pid of master, for sending error signals */
    115 int tenths;		/* length of tape used per block written */
    116 static int caught;	/* have we caught the signal to proceed? */
    117 static int ready;	/* have we reached the lock point without having */
    118 			/* received the SIGUSR2 signal from the prev slave? */
    119 static jmp_buf jmpbuf;	/* where to jump to if we are ready when the */
    120 			/* SIGUSR2 arrives from the previous slave */
    121 
    122 int
    123 alloctape()
    124 {
    125 	int pgoff = getpagesize() - 1;
    126 	char *buf;
    127 	int i;
    128 
    129 	writesize = ntrec * TP_BSIZE;
    130 	reqsiz = (ntrec + 1) * sizeof(struct req);
    131 	/*
    132 	 * CDC 92181's and 92185's make 0.8" gaps in 1600-bpi start/stop mode
    133 	 * (see DEC TU80 User's Guide).  The shorter gaps of 6250-bpi require
    134 	 * repositioning after stopping, i.e, streaming mode, where the gap is
    135 	 * variable, 0.30" to 0.45".  The gap is maximal when the tape stops.
    136 	 */
    137 	if (blocksperfile == 0)
    138 		tenths = writesize / density +
    139 		    (cartridge ? 16 : density == 625 ? 5 : 8);
    140 	/*
    141 	 * Allocate tape buffer contiguous with the array of instruction
    142 	 * packets, so flushtape() can write them together with one write().
    143 	 * Align tape buffer on page boundary to speed up tape write().
    144 	 */
    145 	for (i = 0; i <= SLAVES; i++) {
    146 		buf = (char *)
    147 		    malloc((unsigned)(reqsiz + writesize + pgoff + TP_BSIZE));
    148 		if (buf == NULL)
    149 			return(0);
    150 		slaves[i].tblock = (char (*)[TP_BSIZE])
    151 		    (((long)&buf[ntrec + 1] + pgoff) &~ pgoff);
    152 		slaves[i].req = (struct req *)slaves[i].tblock - ntrec - 1;
    153 	}
    154 	slp = &slaves[0];
    155 	slp->count = 1;
    156 	slp->tapea = 0;
    157 	slp->firstrec = 0;
    158 	nextblock = slp->tblock;
    159 	return(1);
    160 }
    161 
    162 void
    163 writerec(dp, isspcl)
    164 	char *dp;
    165 	int isspcl;
    166 {
    167 
    168 	slp->req[trecno].dblk = (daddr_t)0;
    169 	slp->req[trecno].count = 1;
    170 	*(union u_spcl *)(*(nextblock)++) = *(union u_spcl *)dp;
    171 	if (isspcl)
    172 		lastspclrec = spcl.c_tapea;
    173 	trecno++;
    174 	spcl.c_tapea++;
    175 	if (trecno >= ntrec)
    176 		flushtape();
    177 }
    178 
    179 void
    180 dumpblock(blkno, size)
    181 	daddr_t blkno;
    182 	int size;
    183 {
    184 	int avail, tpblks, dblkno;
    185 
    186 	dblkno = fsbtodb(sblock, blkno);
    187 	tpblks = size >> tp_bshift;
    188 	while ((avail = MIN(tpblks, ntrec - trecno)) > 0) {
    189 		slp->req[trecno].dblk = dblkno;
    190 		slp->req[trecno].count = avail;
    191 		trecno += avail;
    192 		spcl.c_tapea += avail;
    193 		if (trecno >= ntrec)
    194 			flushtape();
    195 		dblkno += avail << (tp_bshift - dev_bshift);
    196 		tpblks -= avail;
    197 	}
    198 }
    199 
    200 int	nogripe = 0;
    201 
    202 void
    203 tperror(signo)
    204 	int signo;
    205 {
    206 
    207 	if (pipeout) {
    208 		msg("write error on %s\n", tape);
    209 		quit("Cannot recover\n");
    210 		/* NOTREACHED */
    211 	}
    212 	msg("write error %d blocks into volume %d\n", blocksthisvol, tapeno);
    213 	broadcast("DUMP WRITE ERROR!\n");
    214 	if (!query("Do you want to restart?"))
    215 		dumpabort(0);
    216 	msg("Closing this volume.  Prepare to restart with new media;\n");
    217 	msg("this dump volume will be rewritten.\n");
    218 	killall();
    219 	nogripe = 1;
    220 	close_rewind();
    221 	Exit(X_REWRITE);
    222 }
    223 
    224 void
    225 sigpipe(signo)
    226 	int signo;
    227 {
    228 
    229 	quit("Broken pipe\n");
    230 }
    231 
    232 void
    233 flushtape()
    234 {
    235 	int i, blks, got;
    236 	long lastfirstrec;
    237 #ifndef __STDC__
    238 	int write(), read();
    239 #endif
    240 
    241 	int siz = (char *)nextblock - (char *)slp->req;
    242 
    243 	slp->req[trecno].count = 0;			/* Sentinel */
    244 
    245 	if (atomic(write, slp->fd, (char *)slp->req, siz) != siz)
    246 		quit("error writing command pipe: %s\n", strerror(errno));
    247 	slp->sent = 1; /* we sent a request, read the response later */
    248 
    249 	lastfirstrec = slp->firstrec;
    250 
    251 	if (++slp >= &slaves[SLAVES])
    252 		slp = &slaves[0];
    253 
    254 	/* Read results back from next slave */
    255 	if (slp->sent) {
    256 		if (atomic(read, slp->fd, (char *)&got, sizeof got)
    257 		    != sizeof got) {
    258 			perror("  DUMP: error reading command pipe in master");
    259 			dumpabort(0);
    260 		}
    261 		slp->sent = 0;
    262 
    263 		/* Check for end of tape */
    264 		if (got < writesize) {
    265 			msg("End of tape detected\n");
    266 
    267 			/*
    268 			 * Drain the results, don't care what the values were.
    269 			 * If we read them here then trewind won't...
    270 			 */
    271 			for (i = 0; i < SLAVES; i++) {
    272 				if (slaves[i].sent) {
    273 					if (atomic(read, slaves[i].fd,
    274 					    (char *)&got, sizeof got)
    275 					    != sizeof got) {
    276 						perror("  DUMP: error reading command pipe in master");
    277 						dumpabort(0);
    278 					}
    279 					slaves[i].sent = 0;
    280 				}
    281 			}
    282 
    283 			close_rewind();
    284 			rollforward();
    285 			return;
    286 		}
    287 	}
    288 
    289 	blks = 0;
    290 	if (spcl.c_type != TS_END) {
    291 		for (i = 0; i < spcl.c_count; i++)
    292 			if (spcl.c_addr[i] != 0)
    293 				blks++;
    294 	}
    295 	slp->count = lastspclrec + blks + 1 - spcl.c_tapea;
    296 	slp->tapea = spcl.c_tapea;
    297 	slp->firstrec = lastfirstrec + ntrec;
    298 	slp->inode = curino;
    299 	nextblock = slp->tblock;
    300 	trecno = 0;
    301 	asize += tenths;
    302 	blockswritten += ntrec;
    303 	blocksthisvol += ntrec;
    304 	if (!pipeout && (blocksperfile ?
    305 	    (blocksthisvol >= blocksperfile) : (asize > tsize))) {
    306 		close_rewind();
    307 		startnewtape(0);
    308 	}
    309 	timeest();
    310 }
    311 
    312 void
    313 trewind()
    314 {
    315 	int f;
    316 	int got;
    317 
    318 	for (f = 0; f < SLAVES; f++) {
    319 		/*
    320 		 * Drain the results, but unlike EOT we DO (or should) care
    321 		 * what the return values were, since if we detect EOT after
    322 		 * we think we've written the last blocks to the tape anyway,
    323 		 * we have to replay those blocks with rollforward.
    324 		 *
    325 		 * fixme: punt for now.
    326 		 */
    327 		if (slaves[f].sent) {
    328 			if (atomic(read, slaves[f].fd, (char *)&got, sizeof got)
    329 			    != sizeof got) {
    330 				perror("  DUMP: error reading command pipe in master");
    331 				dumpabort(0);
    332 			}
    333 			slaves[f].sent = 0;
    334 			if (got != writesize) {
    335 				msg("EOT detected in last 2 tape records!\n");
    336 				msg("Use a longer tape, decrease the size estimate\n");
    337 				quit("or use no size estimate at all.\n");
    338 			}
    339 		}
    340 		(void) close(slaves[f].fd);
    341 	}
    342 	while (wait((int *)NULL) >= 0)	/* wait for any signals from slaves */
    343 		/* void */;
    344 
    345 	if (pipeout)
    346 		return;
    347 
    348 	msg("Closing %s\n", tape);
    349 
    350 #ifdef RDUMP
    351 	if (host) {
    352 		rmtclose();
    353 		while (rmtopen(tape, 0) < 0)
    354 			sleep(10);
    355 		rmtclose();
    356 		return;
    357 	}
    358 #endif
    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 	trewind();
    369 	if (nexttape)
    370 		return;
    371 	if (!nogripe) {
    372 		msg("Change Volumes: Mount volume #%d\n", tapeno+1);
    373 		broadcast("CHANGE DUMP VOLUMES!\7\7\n");
    374 	}
    375 	while (!query("Is the new volume mounted and ready to go?"))
    376 		if (query("Do you want to abort?")) {
    377 			dumpabort(0);
    378 			/*NOTREACHED*/
    379 		}
    380 }
    381 
    382 void
    383 rollforward()
    384 {
    385 	register struct req *p, *q, *prev;
    386 	register struct slave *tslp;
    387 	int i, size, savedtapea, got;
    388 	union u_spcl *ntb, *otb;
    389 	tslp = &slaves[SLAVES];
    390 	ntb = (union u_spcl *)tslp->tblock[1];
    391 
    392 	/*
    393 	 * Each of the N slaves should have requests that need to
    394 	 * be replayed on the next tape.  Use the extra slave buffers
    395 	 * (slaves[SLAVES]) to construct request lists to be sent to
    396 	 * each slave in turn.
    397 	 */
    398 	for (i = 0; i < SLAVES; i++) {
    399 		q = &tslp->req[1];
    400 		otb = (union u_spcl *)slp->tblock;
    401 
    402 		/*
    403 		 * For each request in the current slave, copy it to tslp.
    404 		 */
    405 
    406 		for (p = slp->req; p->count > 0; p += p->count) {
    407 			*q = *p;
    408 			if (p->dblk == 0)
    409 				*ntb++ = *otb++; /* copy the datablock also */
    410 			prev = q;
    411 			q += q->count;
    412 		}
    413 		if (prev->dblk != 0)
    414 			prev->count -= 1;
    415 		else
    416 			ntb--;
    417 		q -= 1;
    418 		q->count = 0;
    419 		q = &tslp->req[0];
    420 		if (i == 0) {
    421 			q->dblk = 0;
    422 			q->count = 1;
    423 			trecno = 0;
    424 			nextblock = tslp->tblock;
    425 			savedtapea = spcl.c_tapea;
    426 			spcl.c_tapea = slp->tapea;
    427 			startnewtape(0);
    428 			spcl.c_tapea = savedtapea;
    429 			lastspclrec = savedtapea - 1;
    430 		}
    431 		size = (char *)ntb - (char *)q;
    432 		if (atomic(write, slp->fd, (char *)q, size) != size) {
    433 			perror("  DUMP: error writing command pipe");
    434 			dumpabort(0);
    435 		}
    436 		slp->sent = 1;
    437 		if (++slp >= &slaves[SLAVES])
    438 			slp = &slaves[0];
    439 
    440 		q->count = 1;
    441 
    442 		if (prev->dblk != 0) {
    443 			/*
    444 			 * If the last one was a disk block, make the
    445 			 * first of this one be the last bit of that disk
    446 			 * block...
    447 			 */
    448 			q->dblk = prev->dblk +
    449 				prev->count * (TP_BSIZE / DEV_BSIZE);
    450 			ntb = (union u_spcl *)tslp->tblock;
    451 		} else {
    452 			/*
    453 			 * It wasn't a disk block.  Copy the data to its
    454 			 * new location in the buffer.
    455 			 */
    456 			q->dblk = 0;
    457 			*((union u_spcl *)tslp->tblock) = *ntb;
    458 			ntb = (union u_spcl *)tslp->tblock[1];
    459 		}
    460 	}
    461 	slp->req[0] = *q;
    462 	nextblock = slp->tblock;
    463 	if (q->dblk == 0)
    464 		nextblock++;
    465 	trecno = 1;
    466 
    467 	/*
    468 	 * Clear the first slaves' response.  One hopes that it
    469 	 * worked ok, otherwise the tape is much too short!
    470 	 */
    471 	if (slp->sent) {
    472 		if (atomic(read, slp->fd, (char *)&got, sizeof got)
    473 		    != sizeof got) {
    474 			perror("  DUMP: error reading command pipe in master");
    475 			dumpabort(0);
    476 		}
    477 		slp->sent = 0;
    478 
    479 		if (got != writesize) {
    480 			quit("EOT detected at start of the tape!\n");
    481 		}
    482 	}
    483 }
    484 
    485 /*
    486  * We implement taking and restoring checkpoints on the tape level.
    487  * When each tape is opened, a new process is created by forking; this
    488  * saves all of the necessary context in the parent.  The child
    489  * continues the dump; the parent waits around, saving the context.
    490  * If the child returns X_REWRITE, then it had problems writing that tape;
    491  * this causes the parent to fork again, duplicating the context, and
    492  * everything continues as if nothing had happened.
    493  */
    494 void
    495 startnewtape(top)
    496 	int top;
    497 {
    498 	int	parentpid;
    499 	int	childpid;
    500 	int	status;
    501 	int	waitpid;
    502 	char	*p;
    503 #ifdef sunos
    504 	void	(*interrupt_save)();
    505 	char	*index();
    506 #else
    507 	sig_t	interrupt_save;
    508 #endif
    509 
    510 	interrupt_save = signal(SIGINT, SIG_IGN);
    511 	parentpid = getpid();
    512 
    513     restore_check_point:
    514 	(void)signal(SIGINT, interrupt_save);
    515 	/*
    516 	 *	All signals are inherited...
    517 	 */
    518 	childpid = fork();
    519 	if (childpid < 0) {
    520 		msg("Context save fork fails in parent %d\n", parentpid);
    521 		Exit(X_ABORT);
    522 	}
    523 	if (childpid != 0) {
    524 		/*
    525 		 *	PARENT:
    526 		 *	save the context by waiting
    527 		 *	until the child doing all of the work returns.
    528 		 *	don't catch the interrupt
    529 		 */
    530 		signal(SIGINT, SIG_IGN);
    531 #ifdef TDEBUG
    532 		msg("Tape: %d; parent process: %d child process %d\n",
    533 			tapeno+1, parentpid, childpid);
    534 #endif TDEBUG
    535 		while ((waitpid = wait(&status)) != childpid)
    536 			msg("Parent %d waiting for child %d has another child %d return\n",
    537 				parentpid, childpid, waitpid);
    538 		if (status & 0xFF) {
    539 			msg("Child %d returns LOB status %o\n",
    540 				childpid, status&0xFF);
    541 		}
    542 		status = (status >> 8) & 0xFF;
    543 #ifdef TDEBUG
    544 		switch(status) {
    545 			case X_FINOK:
    546 				msg("Child %d finishes X_FINOK\n", childpid);
    547 				break;
    548 			case X_ABORT:
    549 				msg("Child %d finishes X_ABORT\n", childpid);
    550 				break;
    551 			case X_REWRITE:
    552 				msg("Child %d finishes X_REWRITE\n", childpid);
    553 				break;
    554 			default:
    555 				msg("Child %d finishes unknown %d\n",
    556 					childpid, status);
    557 				break;
    558 		}
    559 #endif TDEBUG
    560 		switch(status) {
    561 			case X_FINOK:
    562 				Exit(X_FINOK);
    563 			case X_ABORT:
    564 				Exit(X_ABORT);
    565 			case X_REWRITE:
    566 				goto restore_check_point;
    567 			default:
    568 				msg("Bad return code from dump: %d\n", status);
    569 				Exit(X_ABORT);
    570 		}
    571 		/*NOTREACHED*/
    572 	} else {	/* we are the child; just continue */
    573 #ifdef TDEBUG
    574 		sleep(4);	/* allow time for parent's message to get out */
    575 		msg("Child on Tape %d has parent %d, my pid = %d\n",
    576 			tapeno+1, parentpid, getpid());
    577 #endif TDEBUG
    578 		/*
    579 		 * If we have a name like "/dev/rmt0,/dev/rmt1",
    580 		 * use the name before the comma first, and save
    581 		 * the remaining names for subsequent volumes.
    582 		 */
    583 		tapeno++;               /* current tape sequence */
    584 		if (nexttape || index(tape, ',')) {
    585 			if (nexttape && *nexttape)
    586 				tape = nexttape;
    587 			if (p = index(tape, ',')) {
    588 				*p = '\0';
    589 				nexttape = p + 1;
    590 			} else
    591 				nexttape = NULL;
    592 			msg("Dumping volume %d on %s\n", tapeno, tape);
    593 		}
    594 #ifdef RDUMP
    595 		while ((tapefd = (host ? rmtopen(tape, 2) :
    596 			pipeout ? 1 : open(tape, O_WRONLY|O_CREAT, 0666))) < 0)
    597 #else
    598 		while ((tapefd = (pipeout ? 1 :
    599 				  open(tape, O_WRONLY|O_CREAT, 0666))) < 0)
    600 #endif
    601 		    {
    602 			msg("Cannot open output \"%s\".\n", tape);
    603 			if (!query("Do you want to retry the open?"))
    604 				dumpabort(0);
    605 		}
    606 
    607 		enslave();  /* Share open tape file descriptor with slaves */
    608 
    609 		asize = 0;
    610 		blocksthisvol = 0;
    611 		if (top)
    612 			newtape++;		/* new tape signal */
    613 		spcl.c_count = slp->count;
    614 		/*
    615 		 * measure firstrec in TP_BSIZE units since restore doesn't
    616 		 * know the correct ntrec value...
    617 		 */
    618 		spcl.c_firstrec = slp->firstrec;
    619 		spcl.c_volume++;
    620 		spcl.c_type = TS_TAPE;
    621 		spcl.c_flags |= DR_NEWHEADER;
    622 		writeheader((ino_t)slp->inode);
    623 		spcl.c_flags &=~ DR_NEWHEADER;
    624 		if (tapeno > 1)
    625 			msg("Volume %d begins with blocks from inode %d\n",
    626 				tapeno, slp->inode);
    627 	}
    628 }
    629 
    630 void
    631 dumpabort(signo)
    632 	int signo;
    633 {
    634 
    635 	if (master != 0 && master != getpid())
    636 		/* Signals master to call dumpabort */
    637 		(void) kill(master, SIGTERM);
    638 	else {
    639 		killall();
    640 		msg("The ENTIRE dump is aborted.\n");
    641 	}
    642 #ifdef RDUMP
    643 	rmtclose();
    644 #endif
    645 	Exit(X_ABORT);
    646 }
    647 
    648 void
    649 Exit(status)
    650 	int status;
    651 {
    652 
    653 #ifdef TDEBUG
    654 	msg("pid = %d exits with status %d\n", getpid(), status);
    655 #endif TDEBUG
    656 	(void) exit(status);
    657 }
    658 
    659 /*
    660  * proceed - handler for SIGUSR2, used to synchronize IO between the slaves.
    661  */
    662 void
    663 proceed(signo)
    664 	int signo;
    665 {
    666 
    667 	if (ready)
    668 		longjmp(jmpbuf, 1);
    669 	caught++;
    670 }
    671 
    672 void
    673 enslave()
    674 {
    675 	int cmd[2];
    676 	register int i, j;
    677 
    678 	master = getpid();
    679 
    680 	signal(SIGTERM, dumpabort);  /* Slave sends SIGTERM on dumpabort() */
    681 	signal(SIGPIPE, sigpipe);
    682 	signal(SIGUSR1, tperror);    /* Slave sends SIGUSR1 on tape errors */
    683 	signal(SIGUSR2, proceed);    /* Slave sends SIGUSR2 to next slave */
    684 
    685 	for (i = 0; i < SLAVES; i++) {
    686 		if (i == slp - &slaves[0]) {
    687 			caught = 1;
    688 		} else {
    689 			caught = 0;
    690 		}
    691 
    692 		if (socketpair(AF_UNIX, SOCK_STREAM, 0, cmd) < 0 ||
    693 		    (slaves[i].pid = fork()) < 0)
    694 			quit("too many slaves, %d (recompile smaller): %s\n",
    695 			    i, strerror(errno));
    696 
    697 		slaves[i].fd = cmd[1];
    698 		slaves[i].sent = 0;
    699 		if (slaves[i].pid == 0) { 	    /* Slave starts up here */
    700 			for (j = 0; j <= i; j++)
    701 			        (void) close(slaves[j].fd);
    702 			signal(SIGINT, SIG_IGN);    /* Master handles this */
    703 			doslave(cmd[0], i);
    704 			Exit(X_FINOK);
    705 		}
    706 	}
    707 
    708 	for (i = 0; i < SLAVES; i++)
    709 		(void) atomic(write, slaves[i].fd,
    710 			      (char *) &slaves[(i + 1) % SLAVES].pid,
    711 		              sizeof slaves[0].pid);
    712 
    713 	master = 0;
    714 }
    715 
    716 void
    717 killall()
    718 {
    719 	register int i;
    720 
    721 	for (i = 0; i < SLAVES; i++)
    722 		if (slaves[i].pid > 0)
    723 			(void) kill(slaves[i].pid, SIGKILL);
    724 }
    725 
    726 /*
    727  * Synchronization - each process has a lockfile, and shares file
    728  * descriptors to the following process's lockfile.  When our write
    729  * completes, we release our lock on the following process's lock-
    730  * file, allowing the following process to lock it and proceed. We
    731  * get the lock back for the next cycle by swapping descriptors.
    732  */
    733 void
    734 doslave(cmd, slave_number)
    735 	register int cmd;
    736         int slave_number;
    737 {
    738 	register int nread;
    739 	int nextslave, size, wrote, eot_count;
    740 #ifndef __STDC__
    741 	int read();
    742 #endif
    743 
    744 	/*
    745 	 * Need our own seek pointer.
    746 	 */
    747 	(void) close(diskfd);
    748 	if ((diskfd = open(disk, O_RDONLY)) < 0)
    749 		quit("slave couldn't reopen disk: %s\n", strerror(errno));
    750 
    751 	/*
    752 	 * Need the pid of the next slave in the loop...
    753 	 */
    754 	if ((nread = atomic(read, cmd, (char *)&nextslave, sizeof nextslave))
    755 	    != sizeof nextslave) {
    756 		quit("master/slave protocol botched - didn't get pid of next slave.\n");
    757 	}
    758 
    759 	/*
    760 	 * Get list of blocks to dump, read the blocks into tape buffer
    761 	 */
    762 	while ((nread = atomic(read, cmd, (char *)slp->req, reqsiz)) == reqsiz) {
    763 		register struct req *p = slp->req;
    764 
    765 		for (trecno = 0; trecno < ntrec;
    766 		     trecno += p->count, p += p->count) {
    767 			if (p->dblk) {
    768 				bread(p->dblk, slp->tblock[trecno],
    769 					p->count * TP_BSIZE);
    770 			} else {
    771 				if (p->count != 1 || atomic(read, cmd,
    772 				    (char *)slp->tblock[trecno],
    773 				    TP_BSIZE) != TP_BSIZE)
    774 				       quit("master/slave protocol botched.\n");
    775 			}
    776 		}
    777 		if (setjmp(jmpbuf) == 0) {
    778 			ready = 1;
    779 			if (!caught)
    780 				(void) pause();
    781 		}
    782 		ready = 0;
    783 		caught = 0;
    784 
    785 		/* Try to write the data... */
    786 		eot_count = 0;
    787 		size = 0;
    788 
    789 		while (eot_count < 10 && size < writesize) {
    790 #ifdef RDUMP
    791 			if (host)
    792 				wrote = rmtwrite(slp->tblock[0]+size,
    793 				    writesize-size);
    794 			else
    795 #endif
    796 				wrote = write(tapefd, slp->tblock[0]+size,
    797 				    writesize-size);
    798 #ifdef WRITEDEBUG
    799 			printf("slave %d wrote %d\n", slave_number, wrote);
    800 #endif
    801 			if (wrote < 0)
    802 				break;
    803 			if (wrote == 0)
    804 				eot_count++;
    805 			size += wrote;
    806 		}
    807 
    808 #ifdef WRITEDEBUG
    809 		if (size != writesize)
    810 		 printf("slave %d only wrote %d out of %d bytes and gave up.\n",
    811 		     slave_number, size, writesize);
    812 #endif
    813 
    814 		if (eot_count > 0)
    815 			size = 0;
    816 
    817 		/*
    818 		 * fixme: Pyramids running OSx return ENOSPC
    819 		 * at EOT on 1/2 inch drives.
    820 		 */
    821 		if (size < 0) {
    822 			(void) kill(master, SIGUSR1);
    823 			for (;;)
    824 				(void) sigpause(0);
    825 		} else {
    826 			/*
    827 			 * pass size of write back to master
    828 			 * (for EOT handling)
    829 			 */
    830 			(void) atomic(write, cmd, (char *)&size, sizeof size);
    831 		}
    832 
    833 		/*
    834 		 * If partial write, don't want next slave to go.
    835 		 * Also jolts him awake.
    836 		 */
    837 		(void) kill(nextslave, SIGUSR2);
    838 	}
    839 	if (nread != 0)
    840 		quit("error reading command pipe: %s\n", strerror(errno));
    841 }
    842 
    843 /*
    844  * Since a read from a pipe may not return all we asked for,
    845  * or a write may not write all we ask if we get a signal,
    846  * loop until the count is satisfied (or error).
    847  */
    848 int
    849 atomic(func, fd, buf, count)
    850 	int (*func)(), fd, count;
    851 	char *buf;
    852 {
    853 	int got, need = count;
    854 
    855 	while ((got = (*func)(fd, buf, need)) > 0 && (need -= got) > 0)
    856 		buf += got;
    857 	return (got < 0 ? got : count - need);
    858 }
    859