date and time created 92/07/05 14:15:13 by bostic
[unix-history] / usr / src / sbin / dump / tape.c
index 1d5bdbe..4237024 100644 (file)
+/*-
+ * Copyright (c) 1980, 1991 The Regents of the University of California.
+ * All rights reserved.
+ *
+ * %sccs.include.redist.c%
+ */
+
 #ifndef lint
 #ifndef lint
-static char *sccsid = "@(#)tape.c      1.10 (Berkeley) %G%";
+static char sccsid[] = "@(#)tape.c     5.24 (Berkeley) %G%";
+#endif /* not lint */
+
+#ifdef sunos
+#include <sys/param.h>
+#include <stdio.h>
+#include <ctype.h>
+#include <sys/stat.h>
+#include <ufs/fs.h>
+#else
+#include <sys/param.h>
+#include <sys/wait.h>
+#include <ufs/ffs/fs.h>
 #endif
 #endif
-
-#include "dump.h"
+#include <sys/time.h>
+#include <ufs/ufs/dinode.h>
 #include <signal.h>
 #include <signal.h>
+#include <fcntl.h>
+#include <protocols/dumprestore.h>
+#include <errno.h>
+#include <setjmp.h>
+#ifdef __STDC__
+#include <unistd.h>
+#include <stdlib.h>
+#include <string.h>
+#endif
+#include <sys/socket.h>
+#include "dump.h"
+#include "pathnames.h"
 
 
-char   (*tblock)[TP_BSIZE];    /* Pointer to malloc()ed buffer for tape */
-int    writesize;              /* Size of malloc()ed buffer for tape */
-int    trecno = 0;
+int    writesize;              /* size of malloc()ed buffer for tape */
+long   lastspclrec = -1;       /* tape block number of last written header */
+int    trecno = 0;             /* next record to write in current block */
+extern long blocksperfile;     /* number of blocks per output file */
+long   blocksthisvol;          /* number of blocks on current output file */
 extern int ntrec;              /* blocking factor on tape */
 extern int ntrec;              /* blocking factor on tape */
+extern int cartridge;
+extern char *host;
+char   *nexttape;
+#ifdef RDUMP
+int    rmtopen(), rmtwrite();
+void   rmtclose();
+#endif RDUMP
+void   rollforward();
+int    atomic();
+void   doslave(), enslave(), flushtape(), killall();
 
 /*
 
 /*
- * Streaming dump mods (Caltech) - disk block reading and tape writing
+ * Concurrent dump mods (Caltech) - disk block reading and tape writing
  * are exported to several slave processes.  While one slave writes the
  * tape, the others read disk blocks; they pass control of the tape in
  * are exported to several slave processes.  While one slave writes the
  * tape, the others read disk blocks; they pass control of the tape in
- * a ring via pipes.  The parent process traverses the filesystem and
- * sends daddr's, inode records, etc, through pipes to each slave.
- * Speed from Eagle to TU77 on VAX/780 is about 140 Kbytes/second.
- * #ifdef RDUMP version is CPU-limited to about 40 Kbytes/second.
+ * a ring via signals. The parent process traverses the filesystem and
+ * sends writeheader()'s and lists of daddr's to the slaves via pipes.
+ * The following structure defines the instruction packets sent to slaves.
  */
  */
-struct req {                   /* instruction packets sent to slaves */
+struct req {
        daddr_t dblk;
        int count;
        daddr_t dblk;
        int count;
-} *req;
+};
 int reqsiz;
 
 int reqsiz;
 
-#define SLAVES 3               /* 2 slaves read disk while 3rd writes tape */
-#define LAG 2                  /* Write behind by LAG tape blocks (rdump) */
-int slavefd[SLAVES];           /* Pipes from master to each slave */
-int rotor;                     /* Current slave number */
-int master;                    /* Pid of master, for sending error signals */
-int trace = 0;                 /* Protocol trace; easily patchable with adb */
-#define  tmsg  if (trace) msg
-
-#ifdef RDUMP
-extern int rmtape;
-#endif
-
-/*
- * Allocate tape buffer contiguous with the array of instruction packets,
- * so they can be written with a single write call in flusht().
- */
+#define SLAVES 3               /* 1 slave writing, 1 reading, 1 for slack */
+struct slave {
+       int tapea;              /* header number at start of this chunk */
+       int count;              /* count to next header (used for TS_TAPE */
+                               /* after EOT) */
+       int inode;              /* inode that we are currently dealing with */
+       int fd;                 /* FD for this slave */
+       int pid;                /* PID for this slave */
+       int sent;               /* 1 == we've sent this slave requests */
+       int firstrec;           /* record number of this block */
+       char (*tblock)[TP_BSIZE]; /* buffer for data blocks */
+       struct req *req;        /* buffer for requests */
+} slaves[SLAVES+1];
+struct slave *slp;
+
+char   (*nextblock)[TP_BSIZE];
+
+int master;            /* pid of master, for sending error signals */
+int tenths;            /* length of tape used per block written */
+static int caught;     /* have we caught the signal to proceed? */
+static int ready;      /* have we reached the lock point without having */
+                       /* received the SIGUSR2 signal from the prev slave? */
+static jmp_buf jmpbuf; /* where to jump to if we are ready when the */
+                       /* SIGUSR2 arrives from the previous slave */
+
+int
 alloctape()
 {
 alloctape()
 {
+       int pgoff = getpagesize() - 1;
+       char *buf;
+       int i;
 
        writesize = ntrec * TP_BSIZE;
 
        writesize = ntrec * TP_BSIZE;
-       reqsiz = ntrec * sizeof(struct req);
-       req = (struct req *)malloc(reqsiz+writesize);   /* array of packets */
-       tblock = (char (*)[TP_BSIZE]) &req[ntrec];      /* Tape buffer */
-       return (req != NULL);
+       reqsiz = (ntrec + 1) * sizeof(struct req);
+       /*
+        * CDC 92181's and 92185's make 0.8" gaps in 1600-bpi start/stop mode
+        * (see DEC TU80 User's Guide).  The shorter gaps of 6250-bpi require
+        * repositioning after stopping, i.e, streaming mode, where the gap is
+        * variable, 0.30" to 0.45".  The gap is maximal when the tape stops.
+        */
+       if (blocksperfile == 0)
+               tenths = writesize / density +
+                   (cartridge ? 16 : density == 625 ? 5 : 8);
+       /*
+        * Allocate tape buffer contiguous with the array of instruction
+        * packets, so flushtape() can write them together with one write().
+        * Align tape buffer on page boundary to speed up tape write().
+        */
+       for (i = 0; i <= SLAVES; i++) {
+               buf = (char *)
+                   malloc((unsigned)(reqsiz + writesize + pgoff + TP_BSIZE));
+               if (buf == NULL)
+                       return(0);
+               slaves[i].tblock = (char (*)[TP_BSIZE])
+                   (((long)&buf[ntrec + 1] + pgoff) &~ pgoff);
+               slaves[i].req = (struct req *)slaves[i].tblock - ntrec - 1;
+       }
+       slp = &slaves[0];
+       slp->count = 1;
+       slp->tapea = 0;
+       slp->firstrec = 0;
+       nextblock = slp->tblock;
+       return(1);
 }
 
 }
 
-/*
- * Send special record to be put on tape
- */
-taprec(dp)
+void
+writerec(dp, isspcl)
        char *dp;
        char *dp;
+       int isspcl;
 {
 
 {
 
-       tmsg("taprec %d\n", trecno);
-       req[trecno].dblk = (daddr_t)0;
-       req[trecno].count = 1;
-       *(union u_spcl *)(*tblock++) = *(union u_spcl *)dp;
+       slp->req[trecno].dblk = (daddr_t)0;
+       slp->req[trecno].count = 1;
+       *(union u_spcl *)(*(nextblock)++) = *(union u_spcl *)dp;
+       if (isspcl)
+               lastspclrec = spcl.c_tapea;
+       trecno++;
        spcl.c_tapea++;
        spcl.c_tapea++;
-       if (++trecno >= ntrec)
-               flusht();
+       if (trecno >= ntrec)
+               flushtape();
 }
 
 }
 
-dmpblk(blkno, size)
+void
+dumpblock(blkno, size)
        daddr_t blkno;
        int size;
 {
        daddr_t blkno;
        int size;
 {
-       int tpblks, dblkno;
-       register int avail;
+       int avail, tpblks, dblkno;
 
 
-       if (size % TP_BSIZE != 0)
-               msg("bad size to dmpblk: %d\n", size);
        dblkno = fsbtodb(sblock, blkno);
        dblkno = fsbtodb(sblock, blkno);
-       tpblks = size / TP_BSIZE;
+       tpblks = size >> tp_bshift;
        while ((avail = MIN(tpblks, ntrec - trecno)) > 0) {
        while ((avail = MIN(tpblks, ntrec - trecno)) > 0) {
-               tmsg("dmpblk %d\n", avail);
-               req[trecno].dblk = dblkno;
-               req[trecno].count = avail;
+               slp->req[trecno].dblk = dblkno;
+               slp->req[trecno].count = avail;
                trecno += avail;
                spcl.c_tapea += avail;
                if (trecno >= ntrec)
                trecno += avail;
                spcl.c_tapea += avail;
                if (trecno >= ntrec)
-                       flusht();
-               dblkno += avail * (TP_BSIZE / DEV_BSIZE);
+                       flushtape();
+               dblkno += avail << (tp_bshift - dev_bshift);
                tpblks -= avail;
        }
 }
 
 int    nogripe = 0;
 
                tpblks -= avail;
        }
 }
 
 int    nogripe = 0;
 
-tperror() {
+void
+tperror()
+{
+
        if (pipeout) {
        if (pipeout) {
-               msg("Tape write error on %s\n", tape);
-               msg("Cannot recover\n");
-               dumpabort();
+               msg("write error on %s\n", tape);
+               quit("Cannot recover\n");
                /* NOTREACHED */
        }
                /* NOTREACHED */
        }
-       msg("Tape write error on tape %d\n", tapeno);
-       broadcast("TAPE ERROR!\n");
+       msg("write error %d blocks into volume %d\n", blocksthisvol, tapeno);
+       broadcast("DUMP WRITE ERROR!\n");
        if (!query("Do you want to restart?"))
                dumpabort();
        if (!query("Do you want to restart?"))
                dumpabort();
-       msg("This tape will rewind.  After it is rewound,\n");
-       msg("replace the faulty tape with a new one;\n");
+       msg("Closing this volume.  Prepare to restart with new media;\n");
        msg("this dump volume will be rewritten.\n");
        msg("this dump volume will be rewritten.\n");
+       killall();
        nogripe = 1;
        close_rewind();
        Exit(X_REWRITE);
 }
 
        nogripe = 1;
        close_rewind();
        Exit(X_REWRITE);
 }
 
-senderr()
+void
+sigpipe()
 {
 
 {
 
-       perror("dump: pipe error in command to slave");
-       dumpabort();
+       quit("Broken pipe\n");
 }
 
 }
 
-#ifdef RDUMP
-tflush(cnt)
-       int cnt;
+void
+flushtape()
 {
 {
-       int i;
+       int i, blks, got;
+       long lastfirstrec;
+#ifndef __STDC__
+       int write(), read();
+#endif
 
 
-       for (i = 0; i < ntrec; i++)
-               spclrec();
-}
-#endif RDUMP
+       int siz = (char *)nextblock - (char *)slp->req;
 
 
-flusht()
-{
-       int sig, siz = (char *)tblock - (char *)req;
-
-       tmsg("flusht %d\n", siz);
-       sig = sigblock(1<<SIGINT-1 | 1<<SIGIOT-1);  /* Don't interrupt write */
-       if (write(slavefd[rotor], req, siz) != siz)
-               senderr();
-       sigsetmask(sig);
-       if (++rotor >= SLAVES) rotor = 0;
-       tblock = (char (*)[TP_BSIZE]) &req[ntrec];
+       slp->req[trecno].count = 0;                     /* Sentinel */
+
+       if (atomic(write, slp->fd, (char *)slp->req, siz) != siz)
+               quit("error writing command pipe: %s\n", strerror(errno));
+       slp->sent = 1; /* we sent a request, read the response later */
+
+       lastfirstrec = slp->firstrec;
+
+       if (++slp >= &slaves[SLAVES])
+               slp = &slaves[0];
+
+       /* Read results back from next slave */
+       if (slp->sent) {
+               if (atomic(read, slp->fd, (char *)&got, sizeof got)
+                   != sizeof got) {
+                       perror("  DUMP: error reading command pipe in master");
+                       dumpabort();
+               }
+               slp->sent = 0;
+
+               /* Check for end of tape */
+               if (got < writesize) {
+                       msg("End of tape detected\n");
+
+                       /*
+                        * Drain the results, don't care what the values were.
+                        * If we read them here then trewind won't...
+                        */
+                       for (i = 0; i < SLAVES; i++) {
+                               if (slaves[i].sent) {
+                                       if (atomic(read, slaves[i].fd,
+                                           (char *)&got, sizeof got)
+                                           != sizeof got) {
+                                               perror("  DUMP: error reading command pipe in master");
+                                               dumpabort();
+                                       }
+                                       slaves[i].sent = 0;
+                               }
+                       }
+
+                       close_rewind();
+                       rollforward();
+                       return;
+               }
+       }
+
+       blks = 0;
+       if (spcl.c_type != TS_END) {
+               for (i = 0; i < spcl.c_count; i++)
+                       if (spcl.c_addr[i] != 0)
+                               blks++;
+       }
+       slp->count = lastspclrec + blks + 1 - spcl.c_tapea;
+       slp->tapea = spcl.c_tapea;
+       slp->firstrec = lastfirstrec + ntrec;
+       slp->inode = curino;
+       nextblock = slp->tblock;
        trecno = 0;
        trecno = 0;
-       asize += writesize/density;
-       asize += 7;                     /* inter-record gap (why fixed?) */
+       asize += tenths;
        blockswritten += ntrec;
        blockswritten += ntrec;
-       if (!pipeout && asize > tsize) {
+       blocksthisvol += ntrec;
+       if (!pipeout && (blocksperfile ?
+           (blocksthisvol >= blocksperfile) : (asize > tsize))) {
                close_rewind();
                close_rewind();
-               otape();
+               startnewtape(0);
        }
        timeest();
 }
 
        }
        timeest();
 }
 
-rewind()
+void
+trewind()
 {
 {
-       register int f;
+       int f;
+       int got;
+
+       for (f = 0; f < SLAVES; f++) {
+               /*
+                * Drain the results, but unlike EOT we DO (or should) care 
+                * what the return values were, since if we detect EOT after 
+                * we think we've written the last blocks to the tape anyway, 
+                * we have to replay those blocks with rollforward.
+                *
+                * fixme: punt for now.  
+                */
+               if (slaves[f].sent) {
+                       if (atomic(read, slaves[f].fd, (char *)&got, sizeof got)
+                           != sizeof got) {
+                               perror("  DUMP: error reading command pipe in master");
+                               dumpabort();
+                       }
+                       slaves[f].sent = 0;
+                       if (got != writesize) {
+                               msg("EOT detected in last 2 tape records!\n");
+                               msg("Use a longer tape, decrease the size estimate\n");
+                               quit("or use no size estimate at all.\n");
+                       }
+               }
+               (void) close(slaves[f].fd);
+       }
+       while (wait((int *)NULL) >= 0)  /* wait for any signals from slaves */
+               /* void */;
 
        if (pipeout)
                return;
 
        if (pipeout)
                return;
-       for (f = 0; f < SLAVES; f++)
-               close(slavefd[f]);
-       while (wait(NULL) >= 0)    ;    /* wait for any signals from slaves */
-       msg("Tape rewinding\n");
+
+       msg("Closing %s\n", tape);
+
 #ifdef RDUMP
 #ifdef RDUMP
-       rmtclose();
-       while (rmtopen(tape, 0) < 0)
-               sleep(10);
-       rmtclose();
-#else
-       close(to);
+       if (host) {
+               rmtclose();
+               while (rmtopen(tape, 0) < 0)
+                       sleep(10);
+               rmtclose();
+               return;
+       }
+#endif
+       (void) close(tapefd);
        while ((f = open(tape, 0)) < 0)
                sleep (10);
        while ((f = open(tape, 0)) < 0)
                sleep (10);
-       close(f);
-#endif
+       (void) close(f);
 }
 
 }
 
+void
 close_rewind()
 {
 close_rewind()
 {
-       rewind();
+       trewind();
+       if (nexttape)
+               return;
        if (!nogripe) {
        if (!nogripe) {
-               msg("Change Tapes: Mount tape #%d\n", tapeno+1);
-               broadcast("CHANGE TAPES!\7\7\n");
+               msg("Change Volumes: Mount volume #%d\n", tapeno+1);
+               broadcast("CHANGE DUMP VOLUMES!\7\7\n");
        }
        }
-       while (!query("Is the new tape mounted and ready to go?"))
-               if (query("Do you want to abort?"))
+       while (!query("Is the new volume mounted and ready to go?"))
+               if (query("Do you want to abort?")) {
                        dumpabort();
                        dumpabort();
+                       /*NOTREACHED*/
+               }
+}
+
+#ifdef ROLLDEBUG
+int do_sum(block)
+     union u_spcl *block;
+
+{
+       char sum = 0;
+       int i;
+
+       for (i = 0; i < TP_BSIZE; i++) {
+               sum = sum ^ block->dummy[i];
+       }
+       return(sum);
+}
+#endif
+
+void
+rollforward()
+{
+       register struct req *p, *q, *prev;
+       register struct slave *tslp;
+       int i, size, savedtapea, got;
+       union u_spcl *ntb, *otb;
+#ifdef ROLLDEBUG
+       int j; 
+#endif
+       tslp = &slaves[SLAVES];
+       ntb = (union u_spcl *)tslp->tblock[1];
+
+       /*
+        * Each of the N slaves should have requests that need to 
+        * be replayed on the next tape.  Use the extra slave buffers 
+        * (slaves[SLAVES]) to construct request lists to be sent to 
+        * each slave in turn.
+        */
+       for (i = 0; i < SLAVES; i++) {
+               q = &tslp->req[1];
+               otb = (union u_spcl *)slp->tblock;
+
+               /*
+                * For each request in the current slave, copy it to tslp. 
+                */
+#ifdef ROLLDEBUG
+               printf("replaying reqs to slave %d (%d)\n", slp - &slaves[0],
+                   slp->pid);
+               j = 0;
+#endif
+
+               for (p = slp->req; p->count > 0; p += p->count) {
+#ifdef ROLLDEBUG
+                       printf("    req %d count %d dblk %d\n",
+                              j++, p->count, p->dblk);
+                       if (p->dblk == 0)
+                               printf("\tsum %x\n", do_sum(otb));
+#endif
+                       *q = *p;
+                       if (p->dblk == 0)
+                               *ntb++ = *otb++; /* copy the datablock also */
+                       prev = q;
+                       q += q->count;
+               }
+               if (prev->dblk != 0)
+                       prev->count -= 1;
+               else
+                       ntb--;
+               q -= 1;
+               q->count = 0;
+               q = &tslp->req[0];
+               if (i == 0) {
+                       q->dblk = 0;
+                       q->count = 1;
+                       trecno = 0;
+                       nextblock = tslp->tblock;
+                       savedtapea = spcl.c_tapea;
+                       spcl.c_tapea = slp->tapea;
+                       startnewtape(0);
+                       spcl.c_tapea = savedtapea;
+                       lastspclrec = savedtapea - 1;
+               }
+               size = (char *)ntb - (char *)q;
+               if (atomic(write, slp->fd, (char *)q, size) != size) {
+                       perror("  DUMP: error writing command pipe");
+                       dumpabort();
+               }
+               slp->sent = 1;
+#ifdef ROLLDEBUG
+               printf("after the shift:\n");
+               j = 0;
+               for (p = tslp->req; p->count > 0; p += p->count) {
+                       printf("    req %d count %d dblk %d\n",
+                              j++, p->count, p->dblk);
+                       if (p->dblk == 0) {
+                               /* dump block also */
+                       }
+               }
+#endif
+               if (++slp >= &slaves[SLAVES])
+                       slp = &slaves[0];
+
+               q->count = 1;
+
+               if (prev->dblk != 0) {
+                       /*
+                        * If the last one was a disk block, make the 
+                        * first of this one be the last bit of that disk 
+                        * block...
+                        */
+                       q->dblk = prev->dblk +
+                               prev->count * (TP_BSIZE / DEV_BSIZE);
+                       ntb = (union u_spcl *)tslp->tblock;
+               } else {
+                       /*
+                        * It wasn't a disk block.  Copy the data to its 
+                        * new location in the buffer.
+                        */
+                       q->dblk = 0;
+                       *((union u_spcl *)tslp->tblock) = *ntb;
+                       ntb = (union u_spcl *)tslp->tblock[1];
+               }
+       }
+       slp->req[0] = *q;
+       nextblock = slp->tblock;
+       if (q->dblk == 0)
+               nextblock++;
+       trecno = 1;
+
+       /*
+        * Clear the first slaves' response.  One hopes that it
+        * worked ok, otherwise the tape is much too short!
+        */
+       if (slp->sent) {
+               if (atomic(read, slp->fd, (char *)&got, sizeof got)
+                   != sizeof got) {
+                       perror("  DUMP: error reading command pipe in master");
+                       dumpabort();
+               }
+               slp->sent = 0;
+
+               if (got != writesize) {
+                       quit("EOT detected at start of the tape!\n");
+               }
+       }
 }
 
 /*
 }
 
 /*
- *     We implement taking and restoring checkpoints on the tape level.
- *     When each tape is opened, a new process is created by forking; this
- *     saves all of the necessary context in the parent.  The child
- *     continues the dump; the parent waits around, saving the context.
- *     If the child returns X_REWRITE, then it had problems writing that tape;
- *     this causes the parent to fork again, duplicating the context, and
- *     everything continues as if nothing had happened.
+ * We implement taking and restoring checkpoints on the tape level.
+ * When each tape is opened, a new process is created by forking; this
+ * saves all of the necessary context in the parent.  The child
+ * continues the dump; the parent waits around, saving the context.
+ * If the child returns X_REWRITE, then it had problems writing that tape;
+ * this causes the parent to fork again, duplicating the context, and
+ * everything continues as if nothing had happened.
  */
  */
-
-otape()
+void
+startnewtape(top)
+       int top;
 {
        int     parentpid;
        int     childpid;
        int     status;
        int     waitpid;
 {
        int     parentpid;
        int     childpid;
        int     status;
        int     waitpid;
-       int     interrupt();
+       char    *p;
+#ifdef sunos
+       void    (*interrupt_save)();
+       char    *index();
+#else
+       sig_t   interrupt_save;
+#endif
 
 
+       interrupt_save = signal(SIGINT, SIG_IGN);
        parentpid = getpid();
 
     restore_check_point:
        parentpid = getpid();
 
     restore_check_point:
-       signal(SIGINT, interrupt);
+       (void)signal(SIGINT, interrupt_save);
        /*
         *      All signals are inherited...
         */
        /*
         *      All signals are inherited...
         */
@@ -242,7 +556,7 @@ otape()
                        case X_FINOK:
                                msg("Child %d finishes X_FINOK\n", childpid);
                                break;
                        case X_FINOK:
                                msg("Child %d finishes X_FINOK\n", childpid);
                                break;
-                       case X_ABORT:
+                       case X_ABORT:   
                                msg("Child %d finishes X_ABORT\n", childpid);
                                break;
                        case X_REWRITE:
                                msg("Child %d finishes X_ABORT\n", childpid);
                                break;
                        case X_REWRITE:
@@ -250,7 +564,7 @@ otape()
                                break;
                        default:
                                msg("Child %d finishes unknown %d\n",
                                break;
                        default:
                                msg("Child %d finishes unknown %d\n",
-                                   childpid, status);
+                                       childpid, status);
                                break;
                }
 #endif TDEBUG
                                break;
                }
 #endif TDEBUG
@@ -271,177 +585,310 @@ otape()
                sleep(4);       /* allow time for parent's message to get out */
                msg("Child on Tape %d has parent %d, my pid = %d\n",
                        tapeno+1, parentpid, getpid());
                sleep(4);       /* allow time for parent's message to get out */
                msg("Child on Tape %d has parent %d, my pid = %d\n",
                        tapeno+1, parentpid, getpid());
-#endif
+#endif TDEBUG
+               /*
+                * If we have a name like "/dev/rmt0,/dev/rmt1",
+                * use the name before the comma first, and save
+                * the remaining names for subsequent volumes.
+                */
+               tapeno++;               /* current tape sequence */
+               if (nexttape || index(tape, ',')) {
+                       if (nexttape && *nexttape)
+                               tape = nexttape;
+                       if (p = index(tape, ',')) {
+                               *p = '\0';
+                               nexttape = p + 1;
+                       } else
+                               nexttape = NULL;
+                       msg("Dumping volume %d on %s\n", tapeno, tape);
+               }
 #ifdef RDUMP
 #ifdef RDUMP
-               while ((to = rmtopen(tape, 2)) < 0)
+               while ((tapefd = (host ? rmtopen(tape, 2) :
+                       pipeout ? 1 : open(tape, O_WRONLY|O_CREAT, 0666))) < 0)
 #else
 #else
-               while ((to = pipeout ? 1 : creat(tape, 0666)) < 0)
+               while ((tapefd = (pipeout ? 1 : 
+                                 open(tape, O_WRONLY|O_CREAT, 0666))) < 0)
 #endif
 #endif
-                       if (!query("Cannot open tape.  Do you want to retry the open?"))
+                   {
+                       msg("Cannot open output \"%s\".\n", tape);
+                       if (!query("Do you want to retry the open?"))
                                dumpabort();
                                dumpabort();
+               }
 
                enslave();  /* Share open tape file descriptor with slaves */
 
                asize = 0;
 
                enslave();  /* Share open tape file descriptor with slaves */
 
                asize = 0;
-               tapeno++;               /* current tape sequence */
-               newtape++;              /* new tape signal */
+               blocksthisvol = 0;
+               if (top)
+                       newtape++;              /* new tape signal */
+               spcl.c_count = slp->count; 
+               /*
+                * measure firstrec in TP_BSIZE units since restore doesn't
+                * know the correct ntrec value...
+                */
+               spcl.c_firstrec = slp->firstrec;
                spcl.c_volume++;
                spcl.c_type = TS_TAPE;
                spcl.c_volume++;
                spcl.c_type = TS_TAPE;
-               spclrec();
+               spcl.c_flags |= DR_NEWHEADER;
+               writeheader((ino_t)slp->inode);
+               spcl.c_flags &=~ DR_NEWHEADER;
                if (tapeno > 1)
                if (tapeno > 1)
-                       msg("Tape %d begins with blocks from ino %d\n",
-                               tapeno, ino);
+                       msg("Volume %d begins with blocks from inode %d\n",
+                               tapeno, slp->inode);
        }
 }
 
        }
 }
 
+void
 dumpabort()
 {
 dumpabort()
 {
+
        if (master != 0 && master != getpid())
        if (master != 0 && master != getpid())
-               kill(master, SIGIOT);
-       msg("The ENTIRE dump is aborted.\n");
+               /* Signals master to call dumpabort */
+               (void) kill(master, SIGTERM);
+       else {
+               killall();
+               msg("The ENTIRE dump is aborted.\n");
+       }
        Exit(X_ABORT);
 }
 
        Exit(X_ABORT);
 }
 
+void
 Exit(status)
 Exit(status)
+       int status;
 {
 {
+
 #ifdef TDEBUG
        msg("pid = %d exits with status %d\n", getpid(), status);
 #endif TDEBUG
 #ifdef TDEBUG
        msg("pid = %d exits with status %d\n", getpid(), status);
 #endif TDEBUG
-       exit(status);
+       (void) exit(status);
 }
 
 }
 
-#define OK 020
-char tok = OK;
+/*
+ * proceed - handler for SIGUSR2, used to synchronize IO between the slaves.
+ */
+void
+proceed()
+{
+
+       if (ready)
+               longjmp(jmpbuf, 1);
+       caught++;
+}
 
 
+void
 enslave()
 {
 enslave()
 {
-       int prev[2], next[2], cmd[2];   /* file descriptors for pipes */
-       int i, j, slavepid;
+       int cmd[2];
+       register int i, j;
 
        master = getpid();
 
        master = getpid();
-       signal(SIGPIPE, dumpabort);
-       signal(SIGIOT, tperror); /* SIGIOT asks for restart from checkpoint */
-       pipe(prev);
-       for (i = rotor = 0; i < SLAVES; ++i) {
-               if ((i < SLAVES - 1 && pipe(next) < 0) || pipe(cmd) < 0
-                               || (slavepid = fork()) < 0) {
-                       perror("  DUMP: too many slaves");
-                       dumpabort();
+
+       signal(SIGTERM, dumpabort);  /* Slave sends SIGTERM on dumpabort() */
+       signal(SIGPIPE, sigpipe);
+       signal(SIGUSR1, tperror);    /* Slave sends SIGUSR1 on tape errors */
+       signal(SIGUSR2, proceed);    /* Slave sends SIGUSR2 to next slave */
+
+       for (i = 0; i < SLAVES; i++) {
+               if (i == slp - &slaves[0]) {
+                       caught = 1;
+               } else {
+                       caught = 0;
                }
                }
-               if (i >= SLAVES - 1)
-                       next[1] = prev[1];          /* Last slave loops back */
-               slavefd[i] = cmd[1];
-               if (slavepid == 0) {                /* Slave starts up here */
+
+               if (socketpair(AF_UNIX, SOCK_STREAM, 0, cmd) < 0 ||
+                   (slaves[i].pid = fork()) < 0)
+                       quit("too many slaves, %d (recompile smaller): %s\n",
+                           i, strerror(errno));
+
+               slaves[i].fd = cmd[1];
+               slaves[i].sent = 0;
+               if (slaves[i].pid == 0) {           /* Slave starts up here */
                        for (j = 0; j <= i; j++)
                        for (j = 0; j <= i; j++)
-                               close(slavefd[j]);
-                       if (i < SLAVES - 1) {
-                               close(prev[1]);
-                               close(next[0]);
-                       } else {                    /* Insert initial token */
-                               if (write(next[1], &tok, 1) != 1)
-                                       ringerr();
-                       }
-                       doslave(i, cmd[0], prev[0], next[1]);
-                       close(next[1]);
-                       j = read(prev[0], &tok, 1);   /* Eat the final token */
-#ifdef RDUMP                               /* Read remaining acknowledges */
-                       for (; j > 0 && (tok &~ OK) > 0; tok--) {
-                               if (rmtwrite2() != writesize && (tok & OK)) {
-                                       kill(master, SIGIOT);
-                                       tok &= ~OK;
-                               }
-                       }
-#endif
+                               (void) close(slaves[j].fd);
+                       signal(SIGINT, SIG_IGN);    /* Master handles this */
+                       doslave(cmd[0], i);
                        Exit(X_FINOK);
                }
                        Exit(X_FINOK);
                }
-               close(cmd[0]);
-               close(next[1]);
-               close(prev[0]);
-               prev[0] = next[0];
        }
        }
-       master = 0;
+       
+       for (i = 0; i < SLAVES; i++)
+               (void) atomic(write, slaves[i].fd, 
+                             (char *) &slaves[(i + 1) % SLAVES].pid, 
+                             sizeof slaves[0].pid);
+               
+       master = 0; 
 }
 
 }
 
-/*
- * Somebody must have died, should never happen
- */
-ringerr()
+void
+killall()
 {
 {
-       perror("  DUMP: token passing error");
-       kill(master, SIGPIPE);
-       Exit(X_ABORT);
+       register int i;
+
+       for (i = 0; i < SLAVES; i++)
+               if (slaves[i].pid > 0)
+                       (void) kill(slaves[i].pid, SIGKILL);
 }
 
 }
 
-doslave(num, cmd, prev, next)
-       int num, cmd, prev, next;
+/*
+ * Synchronization - each process has a lockfile, and shares file
+ * descriptors to the following process's lockfile.  When our write
+ * completes, we release our lock on the following process's lock-
+ * file, allowing the following process to lock it and proceed. We
+ * get the lock back for the next cycle by swapping descriptors.
+ */
+void
+doslave(cmd, slave_number)
+       register int cmd;
+        int slave_number;
 {
 {
-       tmsg("slave %d\n", num);
-       signal(SIGINT, SIG_IGN);                /* Master handles it */
-       signal(SIGTERM, SIG_IGN);
-       signal(SIGPIPE, ringerr);
-       close(fi);
-       if ((fi = open(disk, 0)) < 0) {         /* Need our own seek pointer */
-               perror("  DUMP: can't reopen disk");
-               kill(master, SIGPIPE);
-               Exit(X_ABORT);
+       register int nread;
+       int nextslave, size, wrote, eot_count;
+#ifndef __STDC__
+       int read();
+#endif
+#ifdef ROLLDEBUG
+       int dodump = 2;
+       FILE *out;
+       char name[64];
+#endif
+
+       /*
+        * Need our own seek pointer.
+        */
+       (void) close(diskfd);
+       if ((diskfd = open(disk, O_RDONLY)) < 0)
+               quit("slave couldn't reopen disk: %s\n", strerror(errno));
+
+       /*
+        * Need the pid of the next slave in the loop...
+        */
+       if ((nread = atomic(read, cmd, (char *)&nextslave, sizeof nextslave))
+           != sizeof nextslave) {
+               quit("master/slave protocol botched - didn't get pid of next slave.\n");
        }
        }
-       while (readpipe(cmd, req, reqsiz) == reqsiz) {
-               register struct req *p = req;
-               for (trecno = 0; trecno < ntrec; trecno += p->count, p += p->count) {
+
+#ifdef ROLLDEBUG
+       sprintf(name, "slave.%d", slave_number);
+       out = fopen(name, "w");
+#endif
+       /*
+        * Get list of blocks to dump, read the blocks into tape buffer
+        */
+       while ((nread = atomic(read, cmd, (char *)slp->req, reqsiz)) == reqsiz) {
+               register struct req *p = slp->req;
+#ifdef ROLLDEBUG
+               int req_count = 0;
+#endif
+
+               for (trecno = 0; trecno < ntrec;
+                    trecno += p->count, p += p->count) {
                        if (p->dblk) {
                        if (p->dblk) {
-                               tmsg("%d READS %d\n", num, p->count);
-                               bread(p->dblk, tblock[trecno],
-                                   p->count * TP_BSIZE);
+                               bread(p->dblk, slp->tblock[trecno],
+                                       p->count * TP_BSIZE);
                        } else {
                        } else {
-                               tmsg("%d PIPEIN %d\n", num, p->count);
-                               if (p->count != 1)
-                                       ringerr();
-                               if (readpipe(cmd, tblock[trecno], TP_BSIZE) != TP_BSIZE)
-                                       senderr();
+                               if (p->count != 1 || atomic(read, cmd,
+                                   (char *)slp->tblock[trecno], 
+                                   TP_BSIZE) != TP_BSIZE)
+                                      quit("master/slave protocol botched.\n");
                        }
                        }
+#ifdef ROLLDEBUG
+                       if (dodump) {
+                               (void) fprintf(out, "    req %d count %d dblk %d\n",
+                                       req_count++, p->count, p->dblk);
+                               if (p->dblk == 0) {
+                                       (void) fprintf(out, "\tsum %x\n",
+                                               do_sum(slp->tblock[trecno]));
+                               }
+                       }
+#endif
                }
                }
-               if (read(prev, &tok, 1) != 1)
-                       ringerr();      /* Wait your turn */
-               tmsg("%d WRITE\n", num);
-#ifdef RDUMP
-               if (tok & OK) {
-                       rmtwrite0(writesize);
-                       rmtwrite1(tblock[0], writesize);
-                       tok++;          /* Number of writes in progress */
+#ifdef ROLLDEBUG
+               if (dodump) {
+                       (void) fprintf(out, "\n");
+               }
+               if (--dodump == 0) {
+                       (void) fclose(out);
                }
                }
-               if (tok > (LAG|OK) && (--tok, rmtwrite2() != writesize)) {
-#else
-               if ((tok & OK) &&
-                   write(to, tblock[0], writesize) != writesize) {
-                       perror(tape);
 #endif
 #endif
-                       kill(master, SIGIOT);   /* restart from checkpoint */
-                       tok &= ~OK;
+               if (setjmp(jmpbuf) == 0) {
+                       ready = 1;
+                       if (!caught)
+                               (void) pause();
                }
                }
-               if (write(next, &tok, 1) != 1)
-                       ringerr(); /* Next slave's turn */
+               ready = 0;
+               caught = 0;
+
+               /* Try to write the data... */
+               eot_count = 0;
+               size = 0;
+
+               while (eot_count < 10 && size < writesize) {
+#ifdef RDUMP
+                       if (host)
+                               wrote = rmtwrite(slp->tblock[0]+size,
+                                   writesize-size);
+                       else
+#endif
+                               wrote = write(tapefd, slp->tblock[0]+size,
+                                   writesize-size);
+#ifdef WRITEDEBUG
+                       printf("slave %d wrote %d\n", slave_number, wrote);
+#endif
+                       if (wrote < 0) 
+                               break;
+                       if (wrote == 0)
+                               eot_count++;
+                       size += wrote;
+               }
+
+#ifdef WRITEDEBUG
+               if (size != writesize) 
+                printf("slave %d only wrote %d out of %d bytes and gave up.\n",
+                    slave_number, size, writesize);
+#endif
+
+               if (eot_count > 0)
+                       size = 0;
+
+               /*
+                * fixme: Pyramids running OSx return ENOSPC
+                * at EOT on 1/2 inch drives.
+                */
+               if (size < 0) {
+                       (void) kill(master, SIGUSR1);
+                       for (;;)
+                               (void) sigpause(0);
+               } else {
+                       /*
+                        * pass size of write back to master
+                        * (for EOT handling)
+                        */
+                       (void) atomic(write, cmd, (char *)&size, sizeof size);
+               } 
+
+               /*
+                * If partial write, don't want next slave to go.
+                * Also jolts him awake.
+                */
+               (void) kill(nextslave, SIGUSR2);
        }
        }
-       tmsg("%d CLOSE\n", num);
+       if (nread != 0)
+               quit("error reading command pipe: %s\n", strerror(errno));
 }
 
 /*
 }
 
 /*
- * Since a read from a pipe may not return all we asked for
- * we must loop until we get all we need
+ * Since a read from a pipe may not return all we asked for,
+ * or a write may not write all we ask if we get a signal,
+ * loop until the count is satisfied (or error).
  */
  */
-readpipe(fd, buf, cnt)
-       int fd;
+int
+atomic(func, fd, buf, count)
+       int (*func)(), fd, count;
        char *buf;
        char *buf;
-       int cnt;
 {
 {
-       int rd, got;
-
-       for (rd = cnt; rd > 0; rd -= got) {
-               got = read(fd, buf, rd);
-               if (got < 0)
-                       return (got);
-               if (got == 0)
-                       return (cnt - rd);
+       int got, need = count;
+
+       while ((got = (*func)(fd, buf, need)) > 0 && (need -= got) > 0)
                buf += got;
                buf += got;
-       }
-       return (cnt);
+       return (got < 0 ? got : count - need);
 }
 }