xref: /csrg-svn/sbin/dump/tape.c (revision 46585)
1 /*
2  * Copyright (c) 1980 Regents of the University of California.
3  * All rights reserved.  The Berkeley software License Agreement
4  * specifies the terms and conditions for redistribution.
5  */
6 
7 #ifndef lint
8 static char sccsid[] = "@(#)tape.c	5.12 (Berkeley) 02/23/91";
9 #endif /* not lint */
10 
11 #include "dump.h"
12 #include <sys/types.h>
13 #include <sys/wait.h>
14 #include <errno.h>
15 #include <fcntl.h>
16 #include "pathnames.h"
17 
18 char	(*tblock)[TP_BSIZE];	/* pointer to malloc()ed buffer for tape */
19 int	writesize;		/* size of malloc()ed buffer for tape */
20 long	lastspclrec = -1;	/* tape block number of last written header */
21 int	trecno = 0;		/* next record to write in current block */
22 extern int ntrec;		/* blocking factor on tape */
23 extern int cartridge;
24 extern int read(), write();
25 #ifdef RDUMP
26 extern char *host;
27 int	rmtopen(), rmtwrite();
28 void	rmtclose();
29 #endif RDUMP
30 
31 int	atomic();
32 void	doslave(), enslave(), flusht(), killall();
33 
34 /*
35  * Concurrent dump mods (Caltech) - disk block reading and tape writing
36  * are exported to several slave processes.  While one slave writes the
37  * tape, the others read disk blocks; they pass control of the tape in
38  * a ring via flock().	The parent process traverses the filesystem and
39  * sends spclrec()'s and lists of daddr's to the slaves via pipes.
40  */
41 struct req {			/* instruction packets sent to slaves */
42 	daddr_t dblk;
43 	int count;
44 } *req;
45 int reqsiz;
46 
47 #define SLAVES 3		/* 1 slave writing, 1 reading, 1 for slack */
48 int slavefd[SLAVES];		/* pipes from master to each slave */
49 int slavepid[SLAVES];		/* used by killall() */
50 int rotor;			/* next slave to be instructed */
51 int master;			/* pid of master, for sending error signals */
52 int tenths;			/* length of tape used per block written */
53 
54 int
55 alloctape()
56 {
57 	int pgoff = getpagesize() - 1;
58 
59 	writesize = ntrec * TP_BSIZE;
60 	reqsiz = ntrec * sizeof(struct req);
61 	/*
62 	 * CDC 92181's and 92185's make 0.8" gaps in 1600-bpi start/stop mode
63 	 * (see DEC TU80 User's Guide).  The shorter gaps of 6250-bpi require
64 	 * repositioning after stopping, i.e, streaming mode, where the gap is
65 	 * variable, 0.30" to 0.45".  The gap is maximal when the tape stops.
66 	 */
67 	tenths = writesize/density + (cartridge ? 16 : density == 625 ? 5 : 8);
68 	/*
69 	 * Allocate tape buffer contiguous with the array of instruction
70 	 * packets, so flusht() can write them together with one write().
71 	 * Align tape buffer on page boundary to speed up tape write().
72 	 */
73 	req = (struct req *)malloc(reqsiz + writesize + pgoff);
74 	if (req == NULL)
75 		return(0);
76 	tblock = (char (*)[TP_BSIZE]) (((long)&req[ntrec] + pgoff) &~ pgoff);
77 	req = (struct req *)tblock - ntrec;
78 	return(1);
79 }
80 
81 
82 void
83 taprec(dp)
84 	char *dp;
85 {
86 	req[trecno].dblk = (daddr_t)0;
87 	req[trecno].count = 1;
88 	*(union u_spcl *)(*tblock++) = *(union u_spcl *)dp;	/* movc3 */
89 	lastspclrec = spcl.c_tapea;
90 	trecno++;
91 	spcl.c_tapea++;
92 	if (trecno >= ntrec)
93 		flusht();
94 }
95 
96 void
97 dmpblk(blkno, size)
98 	daddr_t blkno;
99 	int size;
100 {
101 	int avail, tpblks, dblkno;
102 
103 	dblkno = fsbtodb(sblock, blkno);
104 	tpblks = size >> tp_bshift;
105 	while ((avail = MIN(tpblks, ntrec - trecno)) > 0) {
106 		req[trecno].dblk = dblkno;
107 		req[trecno].count = avail;
108 		trecno += avail;
109 		spcl.c_tapea += avail;
110 		if (trecno >= ntrec)
111 			flusht();
112 		dblkno += avail << (tp_bshift - dev_bshift);
113 		tpblks -= avail;
114 	}
115 }
116 
117 int	nogripe = 0;
118 
119 void
120 tperror()
121 {
122 	if (pipeout) {
123 		msg("Tape write error on %s\n", tape);
124 		quit("Cannot recover\n");
125 		/* NOTREACHED */
126 	}
127 	msg("Tape write error %d feet into tape %d\n", asize/120L, tapeno);
128 	broadcast("TAPE ERROR!\n");
129 	if (!query("Do you want to restart?"))
130 		dumpabort();
131 	msg("This tape will rewind.  After it is rewound,\n");
132 	msg("replace the faulty tape with a new one;\n");
133 	msg("this dump volume will be rewritten.\n");
134 	killall();
135 	nogripe = 1;
136 	close_rewind();
137 	Exit(X_REWRITE);
138 }
139 
140 void
141 sigpipe()
142 {
143 
144 	quit("Broken pipe\n");
145 }
146 
147 #ifdef RDUMP
148 /*
149  * compatibility routine
150  */
151 void
152 tflush(i)
153 	int i;
154 {
155 
156 	for (i = 0; i < ntrec; i++)
157 		spclrec();
158 }
159 #endif RDUMP
160 
161 void
162 flusht()
163 {
164 	int siz = (char *)tblock - (char *)req;
165 
166 	if (atomic(write, slavefd[rotor], req, siz) != siz)
167 		quit("error writing command pipe: %s\n", strerror(errno));
168 	if (++rotor >= SLAVES)
169 		rotor = 0;
170 	tblock = (char (*)[TP_BSIZE]) &req[ntrec];
171 	trecno = 0;
172 	asize += tenths;
173 	blockswritten += ntrec;
174 	if (!pipeout && asize > tsize) {
175 		close_rewind();
176 		otape();
177 	}
178 	timeest();
179 }
180 
181 void
182 trewind()
183 {
184 	int f;
185 
186 	if (pipeout)
187 		return;
188 	for (f = 0; f < SLAVES; f++)
189 		close(slavefd[f]);
190 	while (wait((int *)NULL) >= 0)	/* wait for any signals from slaves */
191 		/* void */;
192 	msg("Tape rewinding\n");
193 #ifdef RDUMP
194 	if (host) {
195 		rmtclose();
196 		while (rmtopen(tape, 0) < 0)
197 			sleep(10);
198 		rmtclose();
199 		return;
200 	}
201 #endif RDUMP
202 	close(to);
203 	while ((f = open(tape, 0)) < 0)
204 		sleep (10);
205 	close(f);
206 }
207 
208 void
209 close_rewind()
210 {
211 	trewind();
212 	if (!nogripe) {
213 		msg("Change Tapes: Mount tape #%d\n", tapeno+1);
214 		broadcast("CHANGE TAPES!\7\7\n");
215 	}
216 	while (!query("Is the new tape mounted and ready to go?"))
217 		if (query("Do you want to abort?")) {
218 			dumpabort();
219 			/*NOTREACHED*/
220 		}
221 }
222 
223 /*
224  *	We implement taking and restoring checkpoints on the tape level.
225  *	When each tape is opened, a new process is created by forking; this
226  *	saves all of the necessary context in the parent.  The child
227  *	continues the dump; the parent waits around, saving the context.
228  *	If the child returns X_REWRITE, then it had problems writing that tape;
229  *	this causes the parent to fork again, duplicating the context, and
230  *	everything continues as if nothing had happened.
231  */
232 
233 void
234 otape()
235 {
236 	int	parentpid;
237 	int	childpid;
238 	int	status;
239 	int	waitpid;
240 	sig_t	interrupt;
241 	int	blks, i;
242 
243 	interrupt = signal(SIGINT, SIG_IGN);
244 	parentpid = getpid();
245 
246     restore_check_point:
247 	(void)signal(SIGINT, interrupt);
248 	/*
249 	 *	All signals are inherited...
250 	 */
251 	childpid = fork();
252 	if (childpid < 0) {
253 		msg("Context save fork fails in parent %d\n", parentpid);
254 		Exit(X_ABORT);
255 	}
256 	if (childpid != 0) {
257 		/*
258 		 *	PARENT:
259 		 *	save the context by waiting
260 		 *	until the child doing all of the work returns.
261 		 *	don't catch the interrupt
262 		 */
263 		signal(SIGINT, SIG_IGN);
264 #ifdef TDEBUG
265 		msg("Tape: %d; parent process: %d child process %d\n",
266 			tapeno+1, parentpid, childpid);
267 #endif TDEBUG
268 		while ((waitpid = wait(&status)) != childpid)
269 			msg("Parent %d waiting for child %d has another child %d return\n",
270 				parentpid, childpid, waitpid);
271 		if (status & 0xFF) {
272 			msg("Child %d returns LOB status %o\n",
273 				childpid, status&0xFF);
274 		}
275 		status = (status >> 8) & 0xFF;
276 #ifdef TDEBUG
277 		switch(status) {
278 			case X_FINOK:
279 				msg("Child %d finishes X_FINOK\n", childpid);
280 				break;
281 			case X_ABORT:
282 				msg("Child %d finishes X_ABORT\n", childpid);
283 				break;
284 			case X_REWRITE:
285 				msg("Child %d finishes X_REWRITE\n", childpid);
286 				break;
287 			default:
288 				msg("Child %d finishes unknown %d\n",
289 					childpid, status);
290 				break;
291 		}
292 #endif TDEBUG
293 		switch(status) {
294 			case X_FINOK:
295 				Exit(X_FINOK);
296 			case X_ABORT:
297 				Exit(X_ABORT);
298 			case X_REWRITE:
299 				goto restore_check_point;
300 			default:
301 				msg("Bad return code from dump: %d\n", status);
302 				Exit(X_ABORT);
303 		}
304 		/*NOTREACHED*/
305 	} else {	/* we are the child; just continue */
306 #ifdef TDEBUG
307 		sleep(4);	/* allow time for parent's message to get out */
308 		msg("Child on Tape %d has parent %d, my pid = %d\n",
309 			tapeno+1, parentpid, getpid());
310 #endif TDEBUG
311 #ifdef RDUMP
312 		while ((to = (host ? rmtopen(tape, 2) :
313 			pipeout ? 1 : creat(tape, 0666))) < 0)
314 #else RDUMP
315 		while ((to = pipeout ? 1 : creat(tape, 0666)) < 0)
316 #endif RDUMP
317 		    {
318 			msg("Cannot open tape \"%s\".\n", tape);
319 			if (!query("Do you want to retry the open?"))
320 				dumpabort();
321 		}
322 
323 		enslave();  /* Share open tape file descriptor with slaves */
324 
325 		asize = 0;
326 		tapeno++;		/* current tape sequence */
327 		newtape++;		/* new tape signal */
328 		blks = 0;
329 		if (spcl.c_type != TS_END)
330 			for (i = 0; i < spcl.c_count; i++)
331 				if (spcl.c_addr[i] != 0)
332 					blks++;
333 		spcl.c_count = blks + 1 - spcl.c_tapea + lastspclrec;
334 		spcl.c_volume++;
335 		spcl.c_type = TS_TAPE;
336 		spcl.c_flags |= DR_NEWHEADER;
337 		spclrec();
338 		spcl.c_flags &=~ DR_NEWHEADER;
339 		if (tapeno > 1)
340 			msg("Tape %d begins with blocks from ino %d\n",
341 				tapeno, ino);
342 	}
343 }
344 
345 void
346 dumpabort()
347 {
348 	if (master != 0 && master != getpid())
349 		kill(master, SIGTERM);	/* Signals master to call dumpabort */
350 	else {
351 		killall();
352 		msg("The ENTIRE dump is aborted.\n");
353 	}
354 	Exit(X_ABORT);
355 }
356 
357 void
358 Exit(status)
359 	int status;
360 {
361 #ifdef TDEBUG
362 	msg("pid = %d exits with status %d\n", getpid(), status);
363 #endif TDEBUG
364 	exit(status);
365 }
366 
367 /*
368  * could use pipe() for this if flock() worked on pipes
369  */
370 void
371 lockfile(fd)
372 	int fd[2];
373 {
374 	char tmpname[20];
375 
376 	strcpy(tmpname, _PATH_LOCK);
377 	mktemp(tmpname);
378 	if ((fd[1] = creat(tmpname, 0400)) < 0)
379 		quit("cannot create lockfile %s: %s\n",
380 		    tmpname, strerror(errno));
381 	if ((fd[0] = open(tmpname, 0)) < 0)
382 		quit("cannot reopen lockfile %s: %s\n",
383 		    tmpname, strerror(errno));
384 	(void) unlink(tmpname);
385 }
386 
387 void
388 enslave()
389 {
390 	int first[2], prev[2], next[2], cmd[2];     /* file descriptors */
391 	register int i, j;
392 
393 	master = getpid();
394 	signal(SIGTERM, dumpabort); /* Slave sends SIGTERM on dumpabort() */
395 	signal(SIGPIPE, sigpipe);
396 	signal(SIGUSR1, tperror);    /* Slave sends SIGUSR1 on tape errors */
397 	lockfile(first);
398 	for (i = 0; i < SLAVES; i++) {
399 		if (i == 0) {
400 			prev[0] = first[1];
401 			prev[1] = first[0];
402 		} else {
403 			prev[0] = next[0];
404 			prev[1] = next[1];
405 			flock(prev[1], LOCK_EX);
406 		}
407 		if (i < SLAVES - 1) {
408 			lockfile(next);
409 		} else {
410 			next[0] = first[0];
411 			next[1] = first[1];	    /* Last slave loops back */
412 		}
413 		if (pipe(cmd) < 0 || (slavepid[i] = fork()) < 0)
414 			quit("too many slaves, %d (recompile smaller): %s\n",
415 			    i, strerror(errno));
416 		slavefd[i] = cmd[1];
417 		if (slavepid[i] == 0) { 	    /* Slave starts up here */
418 			for (j = 0; j <= i; j++)
419 				close(slavefd[j]);
420 			signal(SIGINT, SIG_IGN);    /* Master handles this */
421 			doslave(cmd[0], prev, next);
422 			Exit(X_FINOK);
423 		}
424 		close(cmd[0]);
425 		if (i > 0) {
426 			close(prev[0]);
427 			close(prev[1]);
428 		}
429 	}
430 	close(first[0]);
431 	close(first[1]);
432 	master = 0; rotor = 0;
433 }
434 
435 void
436 killall()
437 {
438 	register int i;
439 
440 	for (i = 0; i < SLAVES; i++)
441 		if (slavepid[i] > 0)
442 			kill(slavepid[i], SIGKILL);
443 }
444 
445 /*
446  * Synchronization - each process has a lockfile, and shares file
447  * descriptors to the following process's lockfile.  When our write
448  * completes, we release our lock on the following process's lock-
449  * file, allowing the following process to lock it and proceed. We
450  * get the lock back for the next cycle by swapping descriptors.
451  */
452 void
453 doslave(cmd, prev, next)
454 	register int cmd, prev[2], next[2];
455 {
456 	register int nread, toggle = 0;
457 
458 	close(fi);
459 	if ((fi = open(disk, 0)) < 0) 	/* Need our own seek pointer */
460 		quit("slave couldn't reopen disk: %s\n", strerror(errno));
461 	/*
462 	 * Get list of blocks to dump, read the blocks into tape buffer
463 	 */
464 	while ((nread = atomic(read, cmd, req, reqsiz)) == reqsiz) {
465 		register struct req *p = req;
466 		for (trecno = 0; trecno < ntrec; trecno += p->count, p += p->count) {
467 			if (p->dblk) {
468 				bread(p->dblk, tblock[trecno],
469 					p->count * TP_BSIZE);
470 			} else {
471 				if (p->count != 1 || atomic(read, cmd,
472 				    tblock[trecno], TP_BSIZE) != TP_BSIZE)
473 					quit("master/slave protocol botched.\n");
474 			}
475 		}
476 		flock(prev[toggle], LOCK_EX);	/* Wait our turn */
477 
478 #ifdef RDUMP
479 		if ((host ? rmtwrite(tblock[0], writesize)
480 			: write(to, tblock[0], writesize)) != writesize) {
481 #else RDUMP
482 		if (write(to, tblock[0], writesize) != writesize) {
483 #endif RDUMP
484 			kill(master, SIGUSR1);
485 			for (;;)
486 				sigpause(0);
487 		}
488 		toggle ^= 1;
489 		flock(next[toggle], LOCK_UN);	/* Next slave's turn */
490 	}					/* Also jolts him awake */
491 	if (nread != 0)
492 		quit("error reading command pipe: %s\n", strerror(errno));
493 }
494 
495 /*
496  * Since a read from a pipe may not return all we asked for,
497  * or a write may not write all we ask if we get a signal,
498  * loop until the count is satisfied (or error).
499  */
500 int
501 atomic(func, fd, buf, count)
502 	int (*func)(), fd, count;
503 	char *buf;
504 {
505 	int got, need = count;
506 
507 	while ((got = (*func)(fd, buf, need)) > 0 && (need -= got) > 0)
508 		buf += got;
509 	return (got < 0 ? got : count - need);
510 }
511