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 1.18 2000/03/11 15:29:01 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>
76 #include <compaterr.h>
86 #include <ext2fs/ext2fs.h>
91 int writesize; /* size of malloc()ed buffer for tape */
92 long lastspclrec = -1; /* tape block number of last written header */
93 int trecno = 0; /* next record to write in current block */
94 extern long blocksperfile; /* number of blocks per output file */
95 long blocksthisvol; /* number of blocks on current output file */
96 extern int ntrec; /* blocking factor on tape */
103 static ssize_t atomic_read __P((int, void *, size_t));
104 static ssize_t atomic_write __P((int, const void *, size_t));
105 static void doslave __P((int, int));
106 static void enslave __P((void));
107 static void flushtape __P((void));
108 static void killall __P((void));
109 static void rollforward __P((void));
110 static int system_command __P((const char *, const char *, int));
113 * Concurrent dump mods (Caltech) - disk block reading and tape writing
114 * are exported to several slave processes. While one slave writes the
115 * tape, the others read disk blocks; they pass control of the tape in
116 * a ring via signals. The parent process traverses the filesystem and
117 * sends writeheader()'s and lists of daddr's to the slaves via pipes.
118 * The following structure defines the instruction packets sent to slaves.
126 #define SLAVES 3 /* 1 slave writing, 1 reading, 1 for slack */
128 int tapea; /* header number at start of this chunk */
129 int count; /* count to next header (used for TS_TAPE */
131 int inode; /* inode that we are currently dealing with */
132 int fd; /* FD for this slave */
133 int pid; /* PID for this slave */
134 int sent; /* 1 == we've sent this slave requests */
135 int firstrec; /* record number of this block */
136 char (*tblock)[TP_BSIZE]; /* buffer for data blocks */
137 struct req *req; /* buffer for requests */
141 char (*nextblock)[TP_BSIZE];
143 static time_t tstart_volume; /* time of volume start */
144 static int tapea_volume; /* value of spcl.c_tapea at volume start */
146 int master; /* pid of master, for sending error signals */
147 int tenths; /* length of tape used per block written */
148 static int caught; /* have we caught the signal to proceed? */
149 static int ready; /* have we reached the lock point without having */
150 /* received the SIGUSR2 signal from the prev slave? */
151 static sigjmp_buf jmpbuf; /* where to jump to if we are ready when the */
152 /* SIGUSR2 arrives from the previous slave */
157 int pgoff = getpagesize() - 1;
161 writesize = ntrec * TP_BSIZE;
162 reqsiz = (ntrec + 1) * sizeof(struct req);
164 * CDC 92181's and 92185's make 0.8" gaps in 1600-bpi start/stop mode
165 * (see DEC TU80 User's Guide). The shorter gaps of 6250-bpi require
166 * repositioning after stopping, i.e, streaming mode, where the gap is
167 * variable, 0.30" to 0.45". The gap is maximal when the tape stops.
169 if (blocksperfile == 0 && !unlimited)
170 tenths = writesize / density +
171 (cartridge ? 16 : density == 625 ? 5 : 8);
173 * Allocate tape buffer contiguous with the array of instruction
174 * packets, so flushtape() can write them together with one write().
175 * Align tape buffer on page boundary to speed up tape write().
177 for (i = 0; i <= SLAVES; i++) {
179 malloc((unsigned)(reqsiz + writesize + pgoff + TP_BSIZE));
182 slaves[i].tblock = (char (*)[TP_BSIZE])
184 (((long)&buf[reqsiz] + pgoff) &~ pgoff);
186 (((long)&buf[ntrec + 1] + pgoff) &~ pgoff);
188 slaves[i].req = (struct req *)slaves[i].tblock - ntrec - 1;
194 nextblock = slp->tblock;
199 writerec(const void *dp, int isspcl)
202 slp->req[trecno].dblk = (daddr_t)0;
203 slp->req[trecno].count = 1;
204 /* XXX post increment triggers an egcs-1.1.2-12 bug on alpha/sparc */
205 *(union u_spcl *)(*(nextblock)) = *(union u_spcl *)dp;
208 lastspclrec = spcl.c_tapea;
216 dumpblock(daddr_t blkno, int size)
218 int avail, tpblks, dblkno;
220 dblkno = fsbtodb(sblock, blkno);
221 tpblks = size >> tp_bshift;
222 while ((avail = MIN(tpblks, ntrec - trecno)) > 0) {
223 slp->req[trecno].dblk = dblkno;
224 slp->req[trecno].count = avail;
226 spcl.c_tapea += avail;
229 dblkno += avail << (tp_bshift - dev_bshift);
241 msg("write error on %s\n", tape);
242 quit("Cannot recover\n");
245 msg("write error %d blocks into volume %d\n", blocksthisvol, tapeno);
246 broadcast("DUMP WRITE ERROR!\n");
247 if (!query("Do you want to restart?"))
249 msg("Closing this volume. Prepare to restart with new media;\n");
250 msg("this dump volume will be rewritten.\n");
261 quit("Broken pipe\n");
266 * Update xferrate stats
279 ttaken = tnow - tstart_volume;
280 blocks = spcl.c_tapea - tapea_volume;
281 msg("Volume %d completed at: %s", tapeno,
288 msg("Volume %d took %d:%02d:%02d\n", tapeno,
289 ttaken / 3600, (ttaken % 3600) / 60, ttaken % 60);
290 msg("Volume %d transfer rate: %ld KB/s\n", tapeno,
292 xferrate += blocks / ttaken;
300 * information message upon receipt of SIGINFO
301 * (derived from optr.c::timeest())
304 statussig(int notused)
308 int save_errno = errno;
310 if (blockswritten < 500)
315 (void) time((time_t *) &tnow);
317 if (blockswritten > tapesize)
318 tapesize = blockswritten;
319 deltat = tstart_writing - tnow + (1.0 * (tnow - tstart_writing))
320 / blockswritten * tapesize;
321 (void)snprintf(msgbuf, sizeof(msgbuf),
322 "%3.2f%% done at %ld KB/s, finished in %d:%02d\n",
323 (blockswritten * 100.0) / tapesize,
324 (spcl.c_tapea - tapea_volume) / (tnow - tstart_volume),
325 (int)(deltat / 3600), (int)((deltat % 3600) / 60));
326 write(STDERR_FILENO, msgbuf, strlen(msgbuf));
337 int siz = (char *)nextblock - (char *)slp->req;
339 slp->req[trecno].count = 0; /* Sentinel */
341 if (atomic_write( slp->fd, (char *)slp->req, siz) != siz)
342 quit("error writing command pipe: %s\n", strerror(errno));
343 slp->sent = 1; /* we sent a request, read the response later */
345 lastfirstrec = slp->firstrec;
347 if (++slp >= &slaves[SLAVES])
350 /* Read results back from next slave */
352 if (atomic_read( slp->fd, (char *)&got, sizeof got)
354 perror(" DUMP: error reading command pipe in master");
359 /* Check for end of tape */
360 if (got < writesize) {
361 msg("End of tape detected\n");
364 * Drain the results, don't care what the values were.
365 * If we read them here then trewind won't...
367 for (i = 0; i < SLAVES; i++) {
368 if (slaves[i].sent) {
369 if (atomic_read( slaves[i].fd,
370 (char *)&got, sizeof got)
372 perror(" DUMP: error reading command pipe in master");
386 if (spcl.c_type != TS_END) {
387 for (i = 0; i < spcl.c_count; i++)
388 if (spcl.c_addr[i] != 0)
391 slp->count = lastspclrec + blks + 1 - spcl.c_tapea;
392 slp->tapea = spcl.c_tapea;
393 slp->firstrec = lastfirstrec + ntrec;
395 nextblock = slp->tblock;
398 blockswritten += ntrec;
399 blocksthisvol += ntrec;
400 if (!pipeout && !unlimited && (blocksperfile ?
401 (blocksthisvol >= blocksperfile) : (asize > tsize))) {
409 * Executes the command in a shell.
410 * Returns -1 if an error occured, the exit status of
411 * the command on success.
413 int system_command(const char *command, const char *device, int volnum) {
415 char commandstr[4096];
419 perror(" DUMP: unable to fork");
425 snprintf(commandstr, sizeof(commandstr), "%s %s %d", command, device, volnum);
426 commandstr[sizeof(commandstr) - 1] = '\0';
427 execl("/bin/sh", "sh", "-c", commandstr, NULL);
428 perror(" DUMP: unable to execute shell");
432 if (waitpid(pid, &status, 0) == -1) {
433 if (errno != EINTR) {
434 perror(" DUMP: waitpid error");
438 if (WIFEXITED(status))
439 return WEXITSTATUS(status);
452 for (f = 0; f < SLAVES; f++) {
454 * Drain the results, but unlike EOT we DO (or should) care
455 * what the return values were, since if we detect EOT after
456 * we think we've written the last blocks to the tape anyway,
457 * we have to replay those blocks with rollforward.
459 * fixme: punt for now.
461 if (slaves[f].sent) {
462 if (atomic_read( slaves[f].fd, (char *)&got, sizeof got)
464 perror(" DUMP: error reading command pipe in master");
468 if (got != writesize) {
469 msg("EOT detected in last 2 tape records!\n");
470 msg("Use a longer tape, decrease the size estimate\n");
471 quit("or use no size estimate at all.\n");
474 (void) close(slaves[f].fd);
476 while (wait((int *)NULL) >= 0) /* wait for any signals from slaves */
481 msg("Closing %s\n", tape);
486 while (rmtopen(tape, 0) < 0)
493 (void) close(tapefd);
494 while ((f = open(tape, 0)) < 0)
500 msg("Launching %s\n", eot_script);
501 eot_code = system_command(eot_script, tape, tapeno);
503 if (eot_code != 0 && eot_code != 1) {
504 msg("Dump aborted by the end of tape script\n");
516 if (nexttape || Mflag || (eot_code == 0) )
519 msg("Change Volumes: Mount volume #%d\n", tapeno+1);
520 broadcast("CHANGE DUMP VOLUMES!\7\7\n");
522 while (!query("Is the new volume mounted and ready to go?"))
523 if (query("Do you want to abort?")) {
532 register struct req *p, *q, *prev;
533 register struct slave *tslp;
534 int i, size, savedtapea, got;
535 union u_spcl *ntb, *otb;
540 tslp = &slaves[SLAVES];
541 ntb = (union u_spcl *)tslp->tblock[1];
544 * Each of the N slaves should have requests that need to
545 * be replayed on the next tape. Use the extra slave buffers
546 * (slaves[SLAVES]) to construct request lists to be sent to
547 * each slave in turn.
549 for (i = 0; i < SLAVES; i++) {
551 otb = (union u_spcl *)slp->tblock;
554 * For each request in the current slave, copy it to tslp.
558 for (p = slp->req; p->count > 0; p += p->count) {
561 *ntb++ = *otb++; /* copy the datablock also */
566 quit("rollforward: protocol botch");
578 nextblock = tslp->tblock;
579 savedtapea = spcl.c_tapea;
580 spcl.c_tapea = slp->tapea;
582 spcl.c_tapea = savedtapea;
583 lastspclrec = savedtapea - 1;
585 size = (char *)ntb - (char *)q;
586 if (atomic_write( slp->fd, (char *)q, size) != size) {
587 perror(" DUMP: error writing command pipe");
592 lastfirstrec = slp->firstrec;
594 if (++slp >= &slaves[SLAVES])
599 if (prev->dblk != 0) {
601 * If the last one was a disk block, make the
602 * first of this one be the last bit of that disk
605 q->dblk = prev->dblk +
606 prev->count * (TP_BSIZE / DEV_BSIZE);
607 ntb = (union u_spcl *)tslp->tblock;
610 * It wasn't a disk block. Copy the data to its
611 * new location in the buffer.
614 *((union u_spcl *)tslp->tblock) = *ntb;
615 ntb = (union u_spcl *)tslp->tblock[1];
619 nextblock = slp->tblock;
622 *(union u_spcl *)(*(nextblock)++) = *(union u_spcl *)tslp->tblock;
629 * Clear the first slaves' response. One hopes that it
630 * worked ok, otherwise the tape is much too short!
633 if (atomic_read( slp->fd, (char *)&got, sizeof got)
635 perror(" DUMP: error reading command pipe in master");
640 if (got != writesize) {
641 quit("EOT detected at start of the tape!\n");
647 if (spcl.c_type != TS_END) {
648 for (i = 0; i < spcl.c_count; i++)
649 if (spcl.c_addr[i] != 0)
653 slp->firstrec = lastfirstrec + ntrec;
654 slp->count = lastspclrec + blks + 1 - spcl.c_tapea;
657 blockswritten += ntrec;
658 blocksthisvol += ntrec;
663 * We implement taking and restoring checkpoints on the tape level.
664 * When each tape is opened, a new process is created by forking; this
665 * saves all of the necessary context in the parent. The child
666 * continues the dump; the parent waits around, saving the context.
667 * If the child returns X_REWRITE, then it had problems writing that tape;
668 * this causes the parent to fork again, duplicating the context, and
669 * everything continues as if nothing had happened.
672 startnewtape(int top)
680 void (*interrupt_save) __P((int signo));
681 #else /* __linux__ */
683 void (*interrupt_save)();
685 sig_t interrupt_save;
687 #endif /* __linux__ */
689 interrupt_save = signal(SIGINT, SIG_IGN);
690 parentpid = getpid();
691 tapea_volume = spcl.c_tapea;
693 (void)time4(&tstart_volume);
695 (void)time((&tstart_volume);
699 (void)signal(SIGINT, interrupt_save);
701 * All signals are inherited...
705 msg("Context save fork fails in parent %d\n", parentpid);
711 * save the context by waiting
712 * until the child doing all of the work returns.
713 * don't catch the interrupt
715 signal(SIGINT, SIG_IGN);
717 msg("Tape: %d; parent process: %d child process %d\n",
718 tapeno+1, parentpid, childpid);
720 while ((waitpid = wait(&status)) != childpid)
721 if (waitpid != rshpid)
722 msg("Parent %d waiting for child %d has another child %d return\n",
723 parentpid, childpid, waitpid);
725 msg("Child %d returns LOB status %o\n",
726 childpid, status&0xFF);
728 status = (status >> 8) & 0xFF;
732 msg("Child %d finishes X_FINOK\n", childpid);
735 msg("Child %d finishes X_ABORT\n", childpid);
738 msg("Child %d finishes X_REWRITE\n", childpid);
741 msg("Child %d finishes unknown %d\n",
752 goto restore_check_point;
754 msg("Bad return code from dump: %d\n", status);
758 } else { /* we are the child; just continue */
760 sleep(4); /* allow time for parent's message to get out */
761 msg("Child on Tape %d has parent %d, my pid = %d\n",
762 tapeno+1, parentpid, getpid());
765 * If we have a name like "/dev/rmt0,/dev/rmt1",
766 * use the name before the comma first, and save
767 * the remaining names for subsequent volumes.
769 tapeno++; /* current tape sequence */
771 snprintf(tape, MAXPATHLEN, "%s%03d", tapeprefix, tapeno);
772 tape[MAXPATHLEN - 1] = '\0';
773 msg("Dumping volume %d on %s\n", tapeno, tape);
775 else if (nexttape || strchr(tapeprefix, ',')) {
776 if (nexttape && *nexttape)
777 tapeprefix = nexttape;
778 if ((p = strchr(tapeprefix, ',')) != NULL) {
783 strncpy(tape, tapeprefix, MAXPATHLEN);
784 tape[MAXPATHLEN - 1] = '\0';
785 msg("Dumping volume %d on %s\n", tapeno, tape);
788 while ((tapefd = (host ? rmtopen(tape, 2) : pipeout ?
790 open(tape, O_WRONLY|O_CREAT, 0666))) < 0)
792 while ((tapefd = (pipeout ? fileno(stdout) :
793 open(tape, O_WRONLY|O_CREAT, 0666))) < 0)
796 msg("Cannot open output \"%s\".\n", tape);
797 if (!query("Do you want to retry the open?"))
801 enslave(); /* Share open tape file descriptor with slaves */
806 newtape++; /* new tape signal */
807 spcl.c_count = slp->count;
809 * measure firstrec in TP_BSIZE units since restore doesn't
810 * know the correct ntrec value...
812 spcl.c_firstrec = slp->firstrec;
814 spcl.c_type = TS_TAPE;
815 spcl.c_flags |= DR_NEWHEADER;
816 writeheader((ino_t)slp->inode);
817 spcl.c_flags &=~ DR_NEWHEADER;
818 msg("Volume %d started at: %s", tapeno,
820 ctime4(&tstart_volume));
822 ctime(&tstart_volume));
825 msg("Volume %d begins with blocks from inode %d\n",
834 if (master != 0 && master != getpid())
835 /* Signals master to call dumpabort */
836 (void) kill(master, SIGTERM);
839 msg("The ENTIRE dump is aborted.\n");
852 msg("pid = %d exits with status %d\n", getpid(), status);
858 * proceed - handler for SIGUSR2, used to synchronize IO between the slaves.
865 siglongjmp(jmpbuf, 1);
873 #ifdef LINUX_FORK_BUG
881 signal(SIGTERM, dumpabort); /* Slave sends SIGTERM on dumpabort() */
882 signal(SIGPIPE, sigpipe);
883 signal(SIGUSR1, tperror); /* Slave sends SIGUSR1 on tape errors */
884 signal(SIGUSR2, proceed); /* Slave sends SIGUSR2 to next slave */
886 for (i = 0; i < SLAVES; i++) {
887 if (i == slp - &slaves[0]) {
893 if (socketpair(AF_UNIX, SOCK_STREAM, 0, cmd) < 0 ||
894 (slaves[i].pid = fork()) < 0)
895 quit("too many slaves, %d (recompile smaller): %s\n",
898 slaves[i].fd = cmd[1];
900 if (slaves[i].pid == 0) { /* Slave starts up here */
901 for (j = 0; j <= i; j++)
902 (void) close(slaves[j].fd);
903 signal(SIGINT, SIG_IGN); /* Master handles this */
905 signal(SIGINFO, SIG_IGN);
908 #ifdef LINUX_FORK_BUG
909 if (atomic_write( cmd[0], (char *) &i, sizeof i)
911 quit("master/slave protocol botched 3\n");
918 #ifdef LINUX_FORK_BUG
920 * Wait for all slaves to _actually_ start to circumvent a bug in
921 * Linux kernels >= 2.1.3 where a signal sent to a child that hasn't
922 * returned from fork() causes a SEGV in the child process
924 for (i = 0; i < SLAVES; i++)
925 if (atomic_read( slaves[i].fd, (char *) &j, sizeof j) != sizeof j)
926 quit("master/slave protocol botched 4\n");
929 for (i = 0; i < SLAVES; i++)
930 (void) atomic_write( slaves[i].fd,
931 (char *) &slaves[(i + 1) % SLAVES].pid,
932 sizeof slaves[0].pid);
942 for (i = 0; i < SLAVES; i++)
943 if (slaves[i].pid > 0) {
944 (void) kill(slaves[i].pid, SIGKILL);
950 * Synchronization - each process has a lockfile, and shares file
951 * descriptors to the following process's lockfile. When our write
952 * completes, we release our lock on the following process's lock-
953 * file, allowing the following process to lock it and proceed. We
954 * get the lock back for the next cycle by swapping descriptors.
957 doslave(int cmd, int slave_number)
960 int nextslave, size, eot_count;
961 volatile int wrote = 0;
968 * Need our own seek pointer.
970 (void) close(diskfd);
971 if ((diskfd = open(disk, O_RDONLY)) < 0)
972 quit("slave couldn't reopen disk: %s\n", strerror(errno));
975 retval = dump_fs_open(disk, &fs);
977 quit("slave couldn't reopen disk: %s\n", error_message(retval));
978 #endif /* __linux__ */
981 * Need the pid of the next slave in the loop...
983 if ((nread = atomic_read( cmd, (char *)&nextslave, sizeof nextslave))
984 != sizeof nextslave) {
985 quit("master/slave protocol botched - didn't get pid of next slave.\n");
989 * Get list of blocks to dump, read the blocks into tape buffer
991 while ((nread = atomic_read( cmd, (char *)slp->req, reqsiz)) == reqsiz) {
992 register struct req *p = slp->req;
994 for (trecno = 0; trecno < ntrec;
995 trecno += p->count, p += p->count) {
997 bread(p->dblk, slp->tblock[trecno],
998 p->count * TP_BSIZE);
1000 if (p->count != 1 || atomic_read( cmd,
1001 (char *)slp->tblock[trecno],
1002 TP_BSIZE) != TP_BSIZE)
1003 quit("master/slave protocol botched.\n");
1006 if (setjmp(jmpbuf) == 0) {
1014 /* Try to write the data... */
1019 while (eot_count < 10 && size < writesize) {
1022 wrote = rmtwrite(slp->tblock[0]+size,
1026 wrote = write(tapefd, slp->tblock[0]+size,
1029 printf("slave %d wrote %d\n", slave_number, wrote);
1039 if (size != writesize)
1040 printf("slave %d only wrote %d out of %d bytes and gave up.\n",
1041 slave_number, size, writesize);
1045 * Handle ENOSPC as an EOT condition.
1047 if (wrote < 0 && errno == ENOSPC) {
1056 (void) kill(master, SIGUSR1);
1057 sigemptyset(&sigset);
1059 sigsuspend(&sigset);
1062 * pass size of write back to master
1063 * (for EOT handling)
1065 (void) atomic_write( cmd, (char *)&size, sizeof size);
1069 * If partial write, don't want next slave to go.
1070 * Also jolts him awake.
1072 (void) kill(nextslave, SIGUSR2);
1075 quit("error reading command pipe: %s\n", strerror(errno));
1079 * Since a read from a pipe may not return all we asked for,
1080 * or a write may not write all we ask if we get a signal,
1081 * loop until the count is satisfied (or error).
1084 atomic_read(int fd, void *buf, size_t count)
1086 int got, need = count;
1089 while ((got = read(fd, buf, need)) > 0 && (need -= got) > 0)
1091 } while (got == -1 && errno == EINTR);
1092 return (got < 0 ? got : count - need);
1096 * Since a read from a pipe may not return all we asked for,
1097 * or a write may not write all we ask if we get a signal,
1098 * loop until the count is satisfied (or error).
1101 atomic_write(int fd, const void *buf, size_t count)
1103 int got, need = count;
1106 while ((got = write(fd, buf, need)) > 0 && (need -= got) > 0)
1108 } while (got == -1 && errno == EINTR);
1109 return (got < 0 ? got : count - need);