2 * Copyright (c) 1980, 1991, 1993
3 * The Regents of the University of California. All rights reserved.
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
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.
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
32 static char sccsid[] = "@(#)tape.c 8.4 (Berkeley) 5/1/95";
34 static const char rcsid[] =
38 #include <sys/param.h>
39 #include <sys/socket.h>
43 #include <ufs/ufs/dinode.h>
44 #include <ufs/ffs/fs.h>
46 #include <protocols/dumprestore.h>
61 int writesize; /* size of malloc()ed buffer for tape */
62 int64_t lastspclrec = -1; /* tape block number of last written header */
63 int trecno = 0; /* next record to write in current block */
64 extern long blocksperfile; /* number of blocks per output file */
65 long blocksthisvol; /* number of blocks on current output file */
66 extern int ntrec; /* blocking factor on tape */
72 static int atomic(ssize_t (*)(), int, char *, int);
73 static void doslave(int, int);
74 static void enslave(void);
75 static void flushtape(void);
76 static void killall(void);
77 static void rollforward(void);
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.
93 #define SLAVES 3 /* 1 slave writing, 1 reading, 1 for slack */
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 */
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 */
108 char (*nextblock)[TP_BSIZE];
110 int master; /* pid of master, for sending error signals */
111 int tenths; /* length of tape used per block written */
112 static volatile sig_atomic_t caught; /* have we caught the signal to proceed? */
113 static volatile sig_atomic_t ready; /* reached the lock point without having */
114 /* received the SIGUSR2 signal from the prev slave? */
115 static jmp_buf jmpbuf; /* where to jump to if we are ready when the */
116 /* SIGUSR2 arrives from the previous slave */
121 int pgoff = getpagesize() - 1;
125 writesize = ntrec * TP_BSIZE;
126 reqsiz = (ntrec + 1) * sizeof(struct req);
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.
133 if (blocksperfile == 0 && !unlimited)
134 tenths = writesize / density +
135 (cartridge ? 16 : density == 625 ? 5 : 8);
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().
141 for (i = 0; i <= SLAVES; i++) {
143 malloc((unsigned)(reqsiz + writesize + pgoff + TP_BSIZE));
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;
154 nextblock = slp->tblock;
159 writerec(char *dp, int isspcl)
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));
167 lastspclrec = spcl.c_tapea;
175 dumpblock(ufs2_daddr_t blkno, int size)
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;
186 spcl.c_tapea += avail;
189 dblkno += avail << (tp_bshift - dev_bshift);
197 tperror(int signo __unused)
201 msg("write error on %s\n", tape);
202 quit("Cannot recover\n");
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?"))
209 msg("Closing this volume. Prepare to restart with new media;\n");
210 msg("this dump volume will be rewritten.\n");
218 sigpipe(int signo __unused)
221 quit("Broken pipe\n");
228 int64_t lastfirstrec;
230 int siz = (char *)nextblock - (char *)slp->req;
232 slp->req[trecno].count = 0; /* Sentinel */
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 */
238 lastfirstrec = slp->firstrec;
240 if (++slp >= &slaves[SLAVES])
243 /* Read results back from next slave */
245 if (atomic(read, slp->fd, (char *)&got, sizeof got)
247 perror(" DUMP: error reading command pipe in master");
252 /* Check for end of tape */
253 if (got < writesize) {
254 msg("End of tape detected\n");
257 * Drain the results, don't care what the values were.
258 * If we read them here then trewind won't...
260 for (i = 0; i < SLAVES; i++) {
261 if (slaves[i].sent) {
262 if (atomic(read, slaves[i].fd,
263 (char *)&got, sizeof got)
265 perror(" DUMP: error reading command pipe in master");
279 if (spcl.c_type != TS_END) {
280 for (i = 0; i < spcl.c_count; i++)
281 if (spcl.c_addr[i] != 0)
284 slp->count = lastspclrec + blks + 1 - spcl.c_tapea;
285 slp->tapea = spcl.c_tapea;
286 slp->firstrec = lastfirstrec + ntrec;
288 nextblock = slp->tblock;
291 blockswritten += ntrec;
292 blocksthisvol += ntrec;
293 if (!pipeout && !unlimited && (blocksperfile ?
294 (blocksthisvol >= blocksperfile) : (asize > tsize))) {
308 for (f = 0; f < SLAVES; f++) {
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.
315 * fixme: punt for now.
317 if (slaves[f].sent) {
318 if (atomic(read, slaves[f].fd, (char *)&got, sizeof got)
320 perror(" DUMP: error reading command pipe in master");
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");
330 (void) close(slaves[f].fd);
332 while (wait((int *)NULL) >= 0) /* wait for any signals from slaves */
338 msg("Closing %s\n", tape);
342 (void)pclose(popenfp);
349 while (rmtopen(tape, 0) < 0)
355 if (fstat(tapefd, &sb) == 0 && S_ISFIFO(sb.st_mode)) {
359 (void) close(tapefd);
360 while ((f = open(tape, 0)) < 0)
368 time_t tstart_changevol, tend_changevol;
373 (void)time((time_t *)&(tstart_changevol));
375 msg("Change Volumes: Mount volume #%d\n", tapeno+1);
376 broadcast("CHANGE DUMP VOLUMES!\a\a\n");
378 while (!query("Is the new volume mounted and ready to go?"))
379 if (query("Do you want to abort?")) {
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);
391 struct req *p, *q, *prev;
395 union u_spcl *ntb, *otb;
396 tslp = &slaves[SLAVES];
397 ntb = (union u_spcl *)tslp->tblock[1];
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.
405 for (i = 0; i < SLAVES; i++) {
407 otb = (union u_spcl *)slp->tblock;
410 * For each request in the current slave, copy it to tslp.
414 for (p = slp->req; p->count > 0; p += p->count) {
417 *ntb++ = *otb++; /* copy the datablock also */
422 quit("rollforward: protocol botch");
434 nextblock = tslp->tblock;
435 savedtapea = spcl.c_tapea;
436 spcl.c_tapea = slp->tapea;
438 spcl.c_tapea = savedtapea;
439 lastspclrec = savedtapea - 1;
441 size = (char *)ntb - (char *)q;
442 if (atomic(write, slp->fd, (char *)q, size) != size) {
443 perror(" DUMP: error writing command pipe");
447 if (++slp >= &slaves[SLAVES])
452 if (prev->dblk != 0) {
454 * If the last one was a disk block, make the
455 * first of this one be the last bit of that disk
458 q->dblk = prev->dblk +
459 prev->count * (TP_BSIZE / DEV_BSIZE);
460 ntb = (union u_spcl *)tslp->tblock;
463 * It wasn't a disk block. Copy the data to its
464 * new location in the buffer.
467 *((union u_spcl *)tslp->tblock) = *ntb;
468 ntb = (union u_spcl *)tslp->tblock[1];
472 nextblock = slp->tblock;
478 * Clear the first slaves' response. One hopes that it
479 * worked ok, otherwise the tape is much too short!
482 if (atomic(read, slp->fd, (char *)&got, sizeof got)
484 perror(" DUMP: error reading command pipe in master");
489 if (got != writesize) {
490 quit("EOT detected at start of the tape!\n");
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.
505 startnewtape(int top)
511 sig_t interrupt_save;
513 interrupt_save = signal(SIGINT, SIG_IGN);
514 parentpid = getpid();
517 (void)signal(SIGINT, interrupt_save);
519 * All signals are inherited...
521 setproctitle(NULL); /* Restore the proctitle. */
524 msg("Context save fork fails in parent %d\n", parentpid);
530 * save the context by waiting
531 * until the child doing all of the work returns.
532 * don't catch the interrupt
534 signal(SIGINT, SIG_IGN);
536 msg("Tape: %d; parent process: %d child process %d\n",
537 tapeno+1, parentpid, childpid);
539 if (waitpid(childpid, &status, 0) == -1)
540 msg("Waiting for child %d: %s\n", childpid,
543 msg("Child %d returns LOB status %o\n",
544 childpid, status&0xFF);
546 status = (status >> 8) & 0xFF;
550 msg("Child %d finishes X_FINOK\n", childpid);
553 msg("Child %d finishes X_ABORT\n", childpid);
556 msg("Child %d finishes X_REWRITE\n", childpid);
559 msg("Child %d finishes unknown %d\n",
570 goto restore_check_point;
572 msg("Bad return code from dump: %d\n", status);
576 } else { /* we are the child; just continue */
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());
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.
587 tapeno++; /* current tape sequence */
588 if (nexttape || strchr(tape, ',')) {
589 if (nexttape && *nexttape)
591 if ((p = strchr(tape, ',')) != NULL) {
596 msg("Dumping volume %d on %s\n", tapeno, tape);
599 tapefd = STDOUT_FILENO;
600 } else if (popenout) {
601 char volno[sizeof("2147483647")];
603 (void)sprintf(volno, "%d", spcl.c_volume + 1);
604 if (setenv("DUMP_VOLUME", volno, 1) == -1) {
605 msg("Cannot set $DUMP_VOLUME.\n");
608 popenfp = popen(popenout, "w");
609 if (popenfp == NULL) {
610 msg("Cannot open output pipeline \"%s\".\n",
614 tapefd = fileno(popenfp);
617 while ((tapefd = (host ? rmtopen(tape, 2) :
618 open(tape, O_WRONLY|O_CREAT, 0666))) < 0)
621 open(tape, O_WRONLY|O_CREAT, 0666)) < 0)
624 msg("Cannot open output \"%s\".\n", tape);
625 if (!query("Do you want to retry the open?"))
630 enslave(); /* Share open tape file descriptor with slaves */
632 close(tapefd); /* Give up our copy of it. */
633 signal(SIGINFO, infosch);
638 newtape++; /* new tape signal */
639 spcl.c_count = slp->count;
641 * measure firstrec in TP_BSIZE units since restore doesn't
642 * know the correct ntrec value...
644 spcl.c_firstrec = slp->firstrec;
646 spcl.c_type = TS_TAPE;
647 writeheader((ino_t)slp->inode);
649 msg("Volume %d begins with blocks from inode %d\n",
655 dumpabort(int signo __unused)
658 if (master != 0 && master != getpid())
659 /* Signals master to call dumpabort */
660 (void) kill(master, SIGTERM);
663 msg("The ENTIRE dump is aborted.\n");
677 msg("pid = %d exits with status %d\n", getpid(), status);
683 * proceed - handler for SIGUSR2, used to synchronize IO between the slaves.
686 proceed(int signo __unused)
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 */
707 for (i = 0; i < SLAVES; i++) {
708 if (i == slp - &slaves[0]) {
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",
719 slaves[i].fd = cmd[1];
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 */
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);
743 for (i = 0; i < SLAVES; i++)
744 if (slaves[i].pid > 0) {
745 (void) kill(slaves[i].pid, SIGKILL);
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.
758 doslave(int cmd, int slave_number)
761 int nextslave, size, wrote, eot_count;
764 * Need our own seek pointer.
766 (void) close(diskfd);
767 if ((diskfd = open(disk, O_RDONLY)) < 0)
768 quit("slave couldn't reopen disk: %s\n", strerror(errno));
771 * Need the pid of the next slave in the loop...
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");
779 * Get list of blocks to dump, read the blocks into tape buffer
781 while ((nread = atomic(read, cmd, (char *)slp->req, reqsiz)) == reqsiz) {
782 struct req *p = slp->req;
784 for (trecno = 0; trecno < ntrec;
785 trecno += p->count, p += p->count) {
787 bread(p->dblk, slp->tblock[trecno],
788 p->count * TP_BSIZE);
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");
796 if (setjmp(jmpbuf) == 0) {
804 /* Try to write the data... */
809 while (eot_count < 10 && size < writesize) {
812 wrote = rmtwrite(slp->tblock[0]+size,
816 wrote = write(tapefd, slp->tblock[0]+size,
819 printf("slave %d wrote %d\n", slave_number, wrote);
829 if (size != writesize)
830 printf("slave %d only wrote %d out of %d bytes and gave up.\n",
831 slave_number, size, writesize);
835 * Handle ENOSPC as an EOT condition.
837 if (wrote < 0 && errno == ENOSPC) {
846 (void) kill(master, SIGUSR1);
851 * pass size of write back to master
854 (void) atomic(write, cmd, (char *)&size, sizeof size);
858 * If partial write, don't want next slave to go.
859 * Also jolts him awake.
861 (void) kill(nextslave, SIGUSR2);
864 quit("error reading command pipe: %s\n", strerror(errno));
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).
873 atomic(ssize_t (*func)(), int fd, char *buf, int count)
875 int got, need = count;
877 while ((got = (*func)(fd, buf, need)) > 0 && (need -= got) > 0)
879 return (got < 0 ? got : count - need);