2 * Ported to Linux's Second Extended File System as part of the
3 * dump and restore backup suit
4 * Remy Card <card@Linux.EU.Org>, 1994-1997
5 * Stelian Pop <pop@cybercable.fr>, 1999-2000
9 * Copyright (c) 1980, 1991, 1993
10 * The Regents of the University of California. All rights reserved.
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
20 * 3. All advertising materials mentioning features or use of this software
21 * must display the following acknowledgement:
22 * This product includes software developed by the University of
23 * California, Berkeley and its contributors.
24 * 4. Neither the name of the University nor the names of its contributors
25 * may be used to endorse or promote products derived from this software
26 * without specific prior written permission.
28 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
29 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
32 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
42 static const char rcsid[] =
43 "$Id: tape.c,v 1.14 2000/03/01 10:16:05 stelian Exp $";
47 #include <sys/types.h>
48 #include <linux/types.h>
50 #include <sys/param.h>
51 #include <sys/socket.h>
55 #include <linux/ext2_fs.h>
56 #include <bsdcompat.h>
59 #include <sys/vnode.h>
62 #include <ufs/inode.h>
64 #include <ufs/ufs/dinode.h>
65 #include <ufs/ffs/fs.h>
67 #endif /* __linux__ */
69 #include <protocols/dumprestore.h>
85 #include <ext2fs/ext2fs.h>
90 int writesize; /* size of malloc()ed buffer for tape */
91 long lastspclrec = -1; /* tape block number of last written header */
92 int trecno = 0; /* next record to write in current block */
93 extern long blocksperfile; /* number of blocks per output file */
94 long blocksthisvol; /* number of blocks on current output file */
95 extern int ntrec; /* blocking factor on tape */
101 static ssize_t atomic_read __P((int, void *, size_t));
102 static ssize_t atomic_write __P((int, const void *, size_t));
103 static void doslave __P((int, int));
104 static void enslave __P((void));
105 static void flushtape __P((void));
106 static void killall __P((void));
107 static void rollforward __P((void));
110 * Concurrent dump mods (Caltech) - disk block reading and tape writing
111 * are exported to several slave processes. While one slave writes the
112 * tape, the others read disk blocks; they pass control of the tape in
113 * a ring via signals. The parent process traverses the filesystem and
114 * sends writeheader()'s and lists of daddr's to the slaves via pipes.
115 * The following structure defines the instruction packets sent to slaves.
123 #define SLAVES 3 /* 1 slave writing, 1 reading, 1 for slack */
125 int tapea; /* header number at start of this chunk */
126 int count; /* count to next header (used for TS_TAPE */
128 int inode; /* inode that we are currently dealing with */
129 int fd; /* FD for this slave */
130 int pid; /* PID for this slave */
131 int sent; /* 1 == we've sent this slave requests */
132 int firstrec; /* record number of this block */
133 char (*tblock)[TP_BSIZE]; /* buffer for data blocks */
134 struct req *req; /* buffer for requests */
138 char (*nextblock)[TP_BSIZE];
140 static time_t tstart_volume; /* time of volume start */
141 static int tapea_volume; /* value of spcl.c_tapea at volume start */
143 int master; /* pid of master, for sending error signals */
144 int tenths; /* length of tape used per block written */
145 static int caught; /* have we caught the signal to proceed? */
146 static int ready; /* have we reached the lock point without having */
147 /* received the SIGUSR2 signal from the prev slave? */
148 static sigjmp_buf jmpbuf; /* where to jump to if we are ready when the */
149 /* SIGUSR2 arrives from the previous slave */
154 int pgoff = getpagesize() - 1;
158 writesize = ntrec * TP_BSIZE;
159 reqsiz = (ntrec + 1) * sizeof(struct req);
161 * CDC 92181's and 92185's make 0.8" gaps in 1600-bpi start/stop mode
162 * (see DEC TU80 User's Guide). The shorter gaps of 6250-bpi require
163 * repositioning after stopping, i.e, streaming mode, where the gap is
164 * variable, 0.30" to 0.45". The gap is maximal when the tape stops.
166 if (blocksperfile == 0 && !unlimited)
167 tenths = writesize / density +
168 (cartridge ? 16 : density == 625 ? 5 : 8);
170 * Allocate tape buffer contiguous with the array of instruction
171 * packets, so flushtape() can write them together with one write().
172 * Align tape buffer on page boundary to speed up tape write().
174 for (i = 0; i <= SLAVES; i++) {
176 malloc((unsigned)(reqsiz + writesize + pgoff + TP_BSIZE));
179 slaves[i].tblock = (char (*)[TP_BSIZE])
181 (((long)&buf[reqsiz] + pgoff) &~ pgoff);
183 (((long)&buf[ntrec + 1] + pgoff) &~ pgoff);
185 slaves[i].req = (struct req *)slaves[i].tblock - ntrec - 1;
191 nextblock = slp->tblock;
196 writerec(const void *dp, int isspcl)
199 slp->req[trecno].dblk = (daddr_t)0;
200 slp->req[trecno].count = 1;
201 /* XXX post increment triggers an egcs-1.1.2-12 bug on alpha/sparc */
202 *(union u_spcl *)(*(nextblock)) = *(union u_spcl *)dp;
205 lastspclrec = spcl.c_tapea;
213 dumpblock(daddr_t blkno, int size)
215 int avail, tpblks, dblkno;
217 dblkno = fsbtodb(sblock, blkno);
218 tpblks = size >> tp_bshift;
219 while ((avail = MIN(tpblks, ntrec - trecno)) > 0) {
220 slp->req[trecno].dblk = dblkno;
221 slp->req[trecno].count = avail;
223 spcl.c_tapea += avail;
226 dblkno += avail << (tp_bshift - dev_bshift);
238 msg("write error on %s\n", tape);
239 quit("Cannot recover\n");
242 msg("write error %d blocks into volume %d\n", blocksthisvol, tapeno);
243 broadcast("DUMP WRITE ERROR!\n");
244 if (!query("Do you want to restart?"))
246 msg("Closing this volume. Prepare to restart with new media;\n");
247 msg("this dump volume will be rewritten.\n");
258 quit("Broken pipe\n");
263 * Update xferrate stats
276 ttaken = tnow - tstart_volume;
277 blocks = spcl.c_tapea - tapea_volume;
278 msg("Volume %d completed at: %s", tapeno,
285 msg("Volume %d took %d:%02d:%02d\n", tapeno,
286 ttaken / 3600, (ttaken % 3600) / 60, ttaken % 60);
287 msg("Volume %d transfer rate: %ld KB/s\n", tapeno,
289 xferrate += blocks / ttaken;
297 * information message upon receipt of SIGINFO
298 * (derived from optr.c::timeest())
301 statussig(int notused)
305 int save_errno = errno;
307 if (blockswritten < 500)
312 (void) time((time_t *) &tnow);
314 if (blockswritten > tapesize)
315 tapesize = blockswritten;
316 deltat = tstart_writing - tnow + (1.0 * (tnow - tstart_writing))
317 / blockswritten * tapesize;
318 (void)snprintf(msgbuf, sizeof(msgbuf),
319 "%3.2f%% done at %ld KB/s, finished in %d:%02d\n",
320 (blockswritten * 100.0) / tapesize,
321 (spcl.c_tapea - tapea_volume) / (tnow - tstart_volume),
322 (int)(deltat / 3600), (int)((deltat % 3600) / 60));
323 write(STDERR_FILENO, msgbuf, strlen(msgbuf));
334 int siz = (char *)nextblock - (char *)slp->req;
336 slp->req[trecno].count = 0; /* Sentinel */
338 if (atomic_write( slp->fd, (char *)slp->req, siz) != siz)
339 quit("error writing command pipe: %s\n", strerror(errno));
340 slp->sent = 1; /* we sent a request, read the response later */
342 lastfirstrec = slp->firstrec;
344 if (++slp >= &slaves[SLAVES])
347 /* Read results back from next slave */
349 if (atomic_read( slp->fd, (char *)&got, sizeof got)
351 perror(" DUMP: error reading command pipe in master");
356 /* Check for end of tape */
357 if (got < writesize) {
358 msg("End of tape detected\n");
361 * Drain the results, don't care what the values were.
362 * If we read them here then trewind won't...
364 for (i = 0; i < SLAVES; i++) {
365 if (slaves[i].sent) {
366 if (atomic_read( slaves[i].fd,
367 (char *)&got, sizeof got)
369 perror(" DUMP: error reading command pipe in master");
383 if (spcl.c_type != TS_END) {
384 for (i = 0; i < spcl.c_count; i++)
385 if (spcl.c_addr[i] != 0)
388 slp->count = lastspclrec + blks + 1 - spcl.c_tapea;
389 slp->tapea = spcl.c_tapea;
390 slp->firstrec = lastfirstrec + ntrec;
392 nextblock = slp->tblock;
395 blockswritten += ntrec;
396 blocksthisvol += ntrec;
397 if (!pipeout && !unlimited && (blocksperfile ?
398 (blocksthisvol >= blocksperfile) : (asize > tsize))) {
411 for (f = 0; f < SLAVES; f++) {
413 * Drain the results, but unlike EOT we DO (or should) care
414 * what the return values were, since if we detect EOT after
415 * we think we've written the last blocks to the tape anyway,
416 * we have to replay those blocks with rollforward.
418 * fixme: punt for now.
420 if (slaves[f].sent) {
421 if (atomic_read( slaves[f].fd, (char *)&got, sizeof got)
423 perror(" DUMP: error reading command pipe in master");
427 if (got != writesize) {
428 msg("EOT detected in last 2 tape records!\n");
429 msg("Use a longer tape, decrease the size estimate\n");
430 quit("or use no size estimate at all.\n");
433 (void) close(slaves[f].fd);
435 while (wait((int *)NULL) >= 0) /* wait for any signals from slaves */
441 msg("Closing %s\n", tape);
446 while (rmtopen(tape, 0) < 0)
452 (void) close(tapefd);
453 while ((f = open(tape, 0)) < 0)
463 if (nexttape || Mflag)
466 msg("Change Volumes: Mount volume #%d\n", tapeno+1);
467 broadcast("CHANGE DUMP VOLUMES!\7\7\n");
469 while (!query("Is the new volume mounted and ready to go?"))
470 if (query("Do you want to abort?")) {
479 register struct req *p, *q, *prev;
480 register struct slave *tslp;
481 int i, size, savedtapea, got;
482 union u_spcl *ntb, *otb;
487 tslp = &slaves[SLAVES];
488 ntb = (union u_spcl *)tslp->tblock[1];
491 * Each of the N slaves should have requests that need to
492 * be replayed on the next tape. Use the extra slave buffers
493 * (slaves[SLAVES]) to construct request lists to be sent to
494 * each slave in turn.
496 for (i = 0; i < SLAVES; i++) {
498 otb = (union u_spcl *)slp->tblock;
501 * For each request in the current slave, copy it to tslp.
505 for (p = slp->req; p->count > 0; p += p->count) {
508 *ntb++ = *otb++; /* copy the datablock also */
513 quit("rollforward: protocol botch");
525 nextblock = tslp->tblock;
526 savedtapea = spcl.c_tapea;
527 spcl.c_tapea = slp->tapea;
529 spcl.c_tapea = savedtapea;
530 lastspclrec = savedtapea - 1;
532 size = (char *)ntb - (char *)q;
533 if (atomic_write( slp->fd, (char *)q, size) != size) {
534 perror(" DUMP: error writing command pipe");
539 lastfirstrec = slp->firstrec;
541 if (++slp >= &slaves[SLAVES])
546 if (prev->dblk != 0) {
548 * If the last one was a disk block, make the
549 * first of this one be the last bit of that disk
552 q->dblk = prev->dblk +
553 prev->count * (TP_BSIZE / DEV_BSIZE);
554 ntb = (union u_spcl *)tslp->tblock;
557 * It wasn't a disk block. Copy the data to its
558 * new location in the buffer.
561 *((union u_spcl *)tslp->tblock) = *ntb;
562 ntb = (union u_spcl *)tslp->tblock[1];
566 nextblock = slp->tblock;
569 *(union u_spcl *)(*(nextblock)++) = *(union u_spcl *)tslp->tblock;
576 * Clear the first slaves' response. One hopes that it
577 * worked ok, otherwise the tape is much too short!
580 if (atomic_read( slp->fd, (char *)&got, sizeof got)
582 perror(" DUMP: error reading command pipe in master");
587 if (got != writesize) {
588 quit("EOT detected at start of the tape!\n");
594 if (spcl.c_type != TS_END) {
595 for (i = 0; i < spcl.c_count; i++)
596 if (spcl.c_addr[i] != 0)
600 slp->firstrec = lastfirstrec + ntrec;
601 slp->count = lastspclrec + blks + 1 - spcl.c_tapea;
604 blockswritten += ntrec;
605 blocksthisvol += ntrec;
610 * We implement taking and restoring checkpoints on the tape level.
611 * When each tape is opened, a new process is created by forking; this
612 * saves all of the necessary context in the parent. The child
613 * continues the dump; the parent waits around, saving the context.
614 * If the child returns X_REWRITE, then it had problems writing that tape;
615 * this causes the parent to fork again, duplicating the context, and
616 * everything continues as if nothing had happened.
619 startnewtape(int top)
627 void (*interrupt_save) __P((int signo));
628 #else /* __linux__ */
630 void (*interrupt_save)();
632 sig_t interrupt_save;
634 #endif /* __linux__ */
636 interrupt_save = signal(SIGINT, SIG_IGN);
637 parentpid = getpid();
638 tapea_volume = spcl.c_tapea;
640 (void)time4(&tstart_volume);
642 (void)time((&tstart_volume);
646 (void)signal(SIGINT, interrupt_save);
648 * All signals are inherited...
652 msg("Context save fork fails in parent %d\n", parentpid);
658 * save the context by waiting
659 * until the child doing all of the work returns.
660 * don't catch the interrupt
662 signal(SIGINT, SIG_IGN);
664 msg("Tape: %d; parent process: %d child process %d\n",
665 tapeno+1, parentpid, childpid);
667 while ((waitpid = wait(&status)) != childpid)
668 if (waitpid != rshpid)
669 msg("Parent %d waiting for child %d has another child %d return\n",
670 parentpid, childpid, waitpid);
672 msg("Child %d returns LOB status %o\n",
673 childpid, status&0xFF);
675 status = (status >> 8) & 0xFF;
679 msg("Child %d finishes X_FINOK\n", childpid);
682 msg("Child %d finishes X_ABORT\n", childpid);
685 msg("Child %d finishes X_REWRITE\n", childpid);
688 msg("Child %d finishes unknown %d\n",
699 goto restore_check_point;
701 msg("Bad return code from dump: %d\n", status);
705 } else { /* we are the child; just continue */
707 sleep(4); /* allow time for parent's message to get out */
708 msg("Child on Tape %d has parent %d, my pid = %d\n",
709 tapeno+1, parentpid, getpid());
712 * If we have a name like "/dev/rmt0,/dev/rmt1",
713 * use the name before the comma first, and save
714 * the remaining names for subsequent volumes.
716 tapeno++; /* current tape sequence */
718 snprintf(tape, MAXPATHLEN, "%s%03d", tapeprefix, tapeno);
719 tape[MAXPATHLEN - 1] = '\0';
720 msg("Dumping volume %d on %s\n", tapeno, tape);
722 else if (nexttape || strchr(tapeprefix, ',')) {
723 if (nexttape && *nexttape)
724 tapeprefix = nexttape;
725 if ((p = strchr(tapeprefix, ',')) != NULL) {
730 strncpy(tape, tapeprefix, MAXPATHLEN);
731 tape[MAXPATHLEN - 1] = '\0';
732 msg("Dumping volume %d on %s\n", tapeno, tape);
735 while ((tapefd = (host ? rmtopen(tape, 2) : pipeout ?
737 open(tape, O_WRONLY|O_CREAT, 0666))) < 0)
739 while ((tapefd = (pipeout ? fileno(stdout) :
740 open(tape, O_WRONLY|O_CREAT, 0666))) < 0)
743 msg("Cannot open output \"%s\".\n", tape);
744 if (!query("Do you want to retry the open?"))
748 enslave(); /* Share open tape file descriptor with slaves */
753 newtape++; /* new tape signal */
754 spcl.c_count = slp->count;
756 * measure firstrec in TP_BSIZE units since restore doesn't
757 * know the correct ntrec value...
759 spcl.c_firstrec = slp->firstrec;
761 spcl.c_type = TS_TAPE;
762 spcl.c_flags |= DR_NEWHEADER;
763 writeheader((ino_t)slp->inode);
764 spcl.c_flags &=~ DR_NEWHEADER;
765 msg("Volume %d started at: %s", tapeno,
767 ctime4(&tstart_volume));
769 ctime(&tstart_volume));
772 msg("Volume %d begins with blocks from inode %d\n",
781 if (master != 0 && master != getpid())
782 /* Signals master to call dumpabort */
783 (void) kill(master, SIGTERM);
786 msg("The ENTIRE dump is aborted.\n");
799 msg("pid = %d exits with status %d\n", getpid(), status);
805 * proceed - handler for SIGUSR2, used to synchronize IO between the slaves.
812 siglongjmp(jmpbuf, 1);
820 #ifdef LINUX_FORK_BUG
828 signal(SIGTERM, dumpabort); /* Slave sends SIGTERM on dumpabort() */
829 signal(SIGPIPE, sigpipe);
830 signal(SIGUSR1, tperror); /* Slave sends SIGUSR1 on tape errors */
831 signal(SIGUSR2, proceed); /* Slave sends SIGUSR2 to next slave */
833 for (i = 0; i < SLAVES; i++) {
834 if (i == slp - &slaves[0]) {
840 if (socketpair(AF_UNIX, SOCK_STREAM, 0, cmd) < 0 ||
841 (slaves[i].pid = fork()) < 0)
842 quit("too many slaves, %d (recompile smaller): %s\n",
845 slaves[i].fd = cmd[1];
847 if (slaves[i].pid == 0) { /* Slave starts up here */
848 for (j = 0; j <= i; j++)
849 (void) close(slaves[j].fd);
850 signal(SIGINT, SIG_IGN); /* Master handles this */
852 signal(SIGINFO, SIG_IGN);
855 #ifdef LINUX_FORK_BUG
856 if (atomic_write( cmd[0], (char *) &i, sizeof i)
858 quit("master/slave protocol botched 3\n");
865 #ifdef LINUX_FORK_BUG
867 * Wait for all slaves to _actually_ start to circumvent a bug in
868 * Linux kernels >= 2.1.3 where a signal sent to a child that hasn't
869 * returned from fork() causes a SEGV in the child process
871 for (i = 0; i < SLAVES; i++)
872 if (atomic_read( slaves[i].fd, (char *) &j, sizeof j) != sizeof j)
873 quit("master/slave protocol botched 4\n");
876 for (i = 0; i < SLAVES; i++)
877 (void) atomic_write( slaves[i].fd,
878 (char *) &slaves[(i + 1) % SLAVES].pid,
879 sizeof slaves[0].pid);
889 for (i = 0; i < SLAVES; i++)
890 if (slaves[i].pid > 0) {
891 (void) kill(slaves[i].pid, SIGKILL);
897 * Synchronization - each process has a lockfile, and shares file
898 * descriptors to the following process's lockfile. When our write
899 * completes, we release our lock on the following process's lock-
900 * file, allowing the following process to lock it and proceed. We
901 * get the lock back for the next cycle by swapping descriptors.
904 doslave(int cmd, int slave_number)
907 int nextslave, size, eot_count;
908 volatile int wrote = 0;
915 * Need our own seek pointer.
917 (void) close(diskfd);
918 if ((diskfd = open(disk, O_RDONLY)) < 0)
919 quit("slave couldn't reopen disk: %s\n", strerror(errno));
922 retval = dump_fs_open(disk, &fs);
924 quit("slave couldn't reopen disk: %s\n", error_message(retval));
925 #endif /* __linux__ */
928 * Need the pid of the next slave in the loop...
930 if ((nread = atomic_read( cmd, (char *)&nextslave, sizeof nextslave))
931 != sizeof nextslave) {
932 quit("master/slave protocol botched - didn't get pid of next slave.\n");
936 * Get list of blocks to dump, read the blocks into tape buffer
938 while ((nread = atomic_read( cmd, (char *)slp->req, reqsiz)) == reqsiz) {
939 register struct req *p = slp->req;
941 for (trecno = 0; trecno < ntrec;
942 trecno += p->count, p += p->count) {
944 bread(p->dblk, slp->tblock[trecno],
945 p->count * TP_BSIZE);
947 if (p->count != 1 || atomic_read( cmd,
948 (char *)slp->tblock[trecno],
949 TP_BSIZE) != TP_BSIZE)
950 quit("master/slave protocol botched.\n");
953 if (setjmp(jmpbuf) == 0) {
961 /* Try to write the data... */
966 while (eot_count < 10 && size < writesize) {
969 wrote = rmtwrite(slp->tblock[0]+size,
973 wrote = write(tapefd, slp->tblock[0]+size,
976 printf("slave %d wrote %d\n", slave_number, wrote);
986 if (size != writesize)
987 printf("slave %d only wrote %d out of %d bytes and gave up.\n",
988 slave_number, size, writesize);
992 * Handle ENOSPC as an EOT condition.
994 if (wrote < 0 && errno == ENOSPC) {
1003 (void) kill(master, SIGUSR1);
1004 sigemptyset(&sigset);
1006 sigsuspend(&sigset);
1009 * pass size of write back to master
1010 * (for EOT handling)
1012 (void) atomic_write( cmd, (char *)&size, sizeof size);
1016 * If partial write, don't want next slave to go.
1017 * Also jolts him awake.
1019 (void) kill(nextslave, SIGUSR2);
1022 quit("error reading command pipe: %s\n", strerror(errno));
1026 * Since a read from a pipe may not return all we asked for,
1027 * or a write may not write all we ask if we get a signal,
1028 * loop until the count is satisfied (or error).
1031 atomic_read(int fd, void *buf, size_t count)
1033 int got, need = count;
1036 while ((got = read(fd, buf, need)) > 0 && (need -= got) > 0)
1038 } while (got == -1 && errno == EINTR);
1039 return (got < 0 ? got : count - need);
1043 * Since a read from a pipe may not return all we asked for,
1044 * or a write may not write all we ask if we get a signal,
1045 * loop until the count is satisfied (or error).
1048 atomic_write(int fd, const void *buf, size_t count)
1050 int got, need = count;
1053 while ((got = write(fd, buf, need)) > 0 && (need -= got) > 0)
1055 } while (got == -1 && errno == EINTR);
1056 return (got < 0 ? got : count - need);