update syscalls.master

[unix-history] / usr / src / sys / kern / vfs_cluster.c

diff --git a/usr/src/sys/kern/vfs_cluster.c b/usr/src/sys/kern/vfs_cluster.c
index c257aac..410b555 100644 (file)
--- a/usr/src/sys/kern/vfs_cluster.c
+++ b/usr/src/sys/kern/vfs_cluster.c
@@ -1,788 +1,540 @@
-/*     vfs_cluster.c   4.42    83/04/04        */
-
-#include "../machine/pte.h"
-
-#include "../h/param.h"
-#include "../h/systm.h"
-#include "../h/dir.h"
-#include "../h/user.h"
-#include "../h/buf.h"
-#include "../h/conf.h"
-#include "../h/proc.h"
-#include "../h/seg.h"
-#include "../h/vm.h"
-#include "../h/trace.h"
-
-/*
- * Read in (if necessary) the block and return a buffer pointer.
- */
-struct buf *
-bread(dev, blkno, size)
-       dev_t dev;
-       daddr_t blkno;
-       int size;
-{
-       register struct buf *bp;
-
-       if (size == 0)
-               panic("bread: size 0");
-       bp = getblk(dev, blkno, size);
-       if (bp->b_flags&B_DONE) {
-               trace(TR_BREADHIT, dev, blkno);
-               return(bp);
-       }
-       bp->b_flags |= B_READ;
-       if (bp->b_bcount > bp->b_bufsize)
-               panic("bread");
-       (*bdevsw[major(dev)].d_strategy)(bp);
-       trace(TR_BREADMISS, dev, blkno);
-       u.u_ru.ru_inblock++;            /* pay for read */
-       biowait(bp);
-       return(bp);
-}
-
-/*
- * Read in the block, like bread, but also start I/O on the
- * read-ahead block (which is not allocated to the caller)
- */
-struct buf *
-breada(dev, blkno, size, rablkno, rabsize)
-       dev_t dev;
-       daddr_t blkno; int size;
-       daddr_t rablkno; int rabsize;
-{
-       register struct buf *bp, *rabp;
-
-       bp = NULL;
-       /*
-        * If the block isn't in core, then allocate
-        * a buffer and initiate i/o (getblk checks
-        * for a cache hit).
-        */
-       if (!incore(dev, blkno)) {
-               bp = getblk(dev, blkno, size);
-               if ((bp->b_flags&B_DONE) == 0) {
-                       bp->b_flags |= B_READ;
-                       if (bp->b_bcount > bp->b_bufsize)
-                               panic("breada");
-                       (*bdevsw[major(dev)].d_strategy)(bp);
-                       trace(TR_BREADMISS, dev, blkno);
-                       u.u_ru.ru_inblock++;            /* pay for read */
-               } else
-                       trace(TR_BREADHIT, dev, blkno);
-       }
-
-       /*
-        * If there's a read-ahead block, start i/o
-        * on it also (as above).
-        */
-       if (rablkno && !incore(dev, rablkno)) {
-               rabp = getblk(dev, rablkno, rabsize);
-               if (rabp->b_flags & B_DONE) {
-                       brelse(rabp);
-                       trace(TR_BREADHITRA, dev, blkno);
-               } else {
-                       rabp->b_flags |= B_READ|B_ASYNC;
-                       if (rabp->b_bcount > rabp->b_bufsize)
-                               panic("breadrabp");
-                       (*bdevsw[major(dev)].d_strategy)(rabp);
-                       trace(TR_BREADMISSRA, dev, rablock);
-                       u.u_ru.ru_inblock++;            /* pay in advance */
-               }
-       }
-
-       /*
-        * If block was in core, let bread get it.
-        * If block wasn't in core, then the read was started
-        * above, and just wait for it.
-        */
-       if (bp == NULL)
-               return (bread(dev, blkno, size));
-       biowait(bp);
-       return (bp);
-}
-
-/*
- * Write the buffer, waiting for completion.
- * Then release the buffer.
+/*-
+ * Copyright (c) 1993 The Regents of the University of California.
+ * All rights reserved.
+ *
+ * %sccs.include.redist.c%
+ *
+ *     @(#)vfs_cluster.c       7.60 (Berkeley) %G%

  */
  */

-bwrite(bp)
-       register struct buf *bp;
-{
-       register flag;

 
 

-       flag = bp->b_flags;
-       bp->b_flags &= ~(B_READ | B_DONE | B_ERROR | B_DELWRI);
-       if ((flag&B_DELWRI) == 0)
-               u.u_ru.ru_oublock++;            /* noone paid yet */
-       trace(TR_BWRITE, bp->b_dev, bp->b_blkno);
-       if (bp->b_bcount > bp->b_bufsize)
-               panic("bwrite");
-       (*bdevsw[major(bp->b_dev)].d_strategy)(bp);
-
-       /*
-        * If the write was synchronous, then await i/o completion.
-        * If the write was "delayed", then we put the buffer on
-        * the q of blocks awaiting i/o completion status.
-        */
-       if ((flag&B_ASYNC) == 0) {
-               biowait(bp);
-               brelse(bp);
-       } else if (flag & B_DELWRI)
-               bp->b_flags |= B_AGE;
-}
+#include <sys/param.h>
+#include <sys/proc.h>
+#include <sys/buf.h>
+#include <sys/vnode.h>
+#include <sys/mount.h>
+#include <sys/trace.h>
+#include <sys/malloc.h>
+#include <sys/resourcevar.h>
+#include <libkern/libkern.h>
+#include <ufs/ufs/quota.h>
+#include <ufs/ufs/inode.h>

 
 /*
 
 /*

- * Release the buffer, marking it so that if it is grabbed
- * for another purpose it will be written out before being
- * given up (e.g. when writing a partial block where it is
- * assumed that another write for the same block will soon follow).
- * This can't be done for magtape, since writes must be done
- * in the same order as requested.
+ * Local declarations

  */
  */

-bdwrite(bp)
-       register struct buf *bp;
-{
-       register int flags;
-
-       if ((bp->b_flags&B_DELWRI) == 0)
-               u.u_ru.ru_oublock++;            /* noone paid yet */
-       flags = bdevsw[major(bp->b_dev)].d_flags;
-       if(flags & B_TAPE)
-               bawrite(bp);
-       else {
-               bp->b_flags |= B_DELWRI | B_DONE;
-               brelse(bp);
-       }
-}
+struct buf *cluster_newbuf __P((struct vnode *, struct buf *, long, daddr_t,
+           daddr_t, long, int));
+struct buf *cluster_rbuild __P((struct vnode *, u_quad_t, struct buf *,
+           daddr_t, daddr_t, long, int, long));
+void       cluster_wbuild __P((struct vnode *, struct buf *, long size,
+           daddr_t start_lbn, int len, daddr_t lbn));

 
 /*
 
 /*

- * Release the buffer, start I/O on it, but don't wait for completion.
+ * We could optimize this by keeping track of where the last read-ahead
+ * was, but it would involve adding fields to the vnode.  For now, let's
+ * just get it working.
+ *
+ * This replaces bread.  If this is a bread at the beginning of a file and
+ * lastr is 0, we assume this is the first read and we'll read up to two
+ * blocks if they are sequential.  After that, we'll do regular read ahead
+ * in clustered chunks.
+ *
+ * There are 4 or 5 cases depending on how you count:
+ *     Desired block is in the cache:
+ *         1 Not sequential access (0 I/Os).
+ *         2 Access is sequential, do read-ahead (1 ASYNC).
+ *     Desired block is not in cache:
+ *         3 Not sequential access (1 SYNC).
+ *         4 Sequential access, next block is contiguous (1 SYNC).
+ *         5 Sequential access, next block is not contiguous (1 SYNC, 1 ASYNC)
+ *
+ * There are potentially two buffers that require I/O.
+ *     bp is the block requested.
+ *     rbp is the read-ahead block.
+ *     If either is NULL, then you don't have to do the I/O.

  */
  */

-bawrite(bp)
-       register struct buf *bp;
+cluster_read(vp, filesize, lblkno, size, cred, bpp)
+       struct vnode *vp;
+       u_quad_t filesize;
+       daddr_t lblkno;
+       long size;
+       struct ucred *cred;
+       struct buf **bpp;

 {
 {

+       struct buf *bp, *rbp;
+       daddr_t blkno, ioblkno;
+       long flags;
+       int error, num_ra, alreadyincore;

 
 

-       bp->b_flags |= B_ASYNC;
-       bwrite(bp);
-}
-
-/*
- * Release the buffer, with no I/O implied.
- */
-brelse(bp)
-       register struct buf *bp;
-{
-       register struct buf *flist;
-       register s;
+#ifdef DIAGNOSTIC
+       if (size == 0)
+               panic("cluster_read: size = 0");
+#endif

 
 

-       /*
-        * If someone's waiting for the buffer, or
-        * is waiting for a buffer wake 'em up.
-        */
-       if (bp->b_flags&B_WANTED)
-               wakeup((caddr_t)bp);
-       if (bfreelist[0].b_flags&B_WANTED) {
-               bfreelist[0].b_flags &= ~B_WANTED;
-               wakeup((caddr_t)bfreelist);
+       error = 0;
+       flags = B_READ;
+       *bpp = bp = getblk(vp, lblkno, size, 0, 0);
+       if (bp->b_flags & (B_CACHE | B_DONE | B_DELWRI)) {
+               /*
+                * Desired block is in cache; do any readahead ASYNC.
+                * Case 1, 2.
+                */
+               trace(TR_BREADHIT, pack(vp, size), lblkno);
+               flags |= B_ASYNC;
+               ioblkno = lblkno +
+                   (lblkno < vp->v_ralen ? vp->v_ralen >> 1 : vp->v_ralen);
+               alreadyincore = (int)incore(vp, ioblkno);
+               bp = NULL;
+       } else {
+               /* Block wasn't in cache, case 3, 4, 5. */
+               trace(TR_BREADMISS, pack(vp, size), lblkno);
+               ioblkno = lblkno;
+               bp->b_flags |= flags;
+               alreadyincore = 0;
+               curproc->p_stats->p_ru.ru_inblock++;            /* XXX */

        }
        }

-       if (bp->b_flags&B_ERROR)
-               if (bp->b_flags & B_LOCKED)
-                       bp->b_flags &= ~B_ERROR;        /* try again later */
-               else
-                       bp->b_dev = NODEV;              /* no assoc */
-

        /*
        /*

-        * Stick the buffer back on a free list.
+        * XXX
+        * Replace 1 with a window size based on some permutation of
+        * maxcontig and rot_delay.  This will let you figure out how
+        * many blocks you should read-ahead (case 2, 4, 5).
+        *
+        * If the access isn't sequential, cut the window size in half.

         */
         */

-       s = spl6();
-       if (bp->b_bufsize <= 0) {
-               /* block has no buffer ... put at front of unused buffer list */
-               flist = &bfreelist[BQ_EMPTY];
-               binsheadfree(bp, flist);
-       } else if (bp->b_flags & (B_ERROR|B_INVAL)) {
-               /* block has no info ... put at front of most free list */
-               flist = &bfreelist[BQ_AGE];
-               binsheadfree(bp, flist);
-       } else {
-               if (bp->b_flags & B_LOCKED)
-                       flist = &bfreelist[BQ_LOCKED];
-               else if (bp->b_flags & B_AGE)
-                       flist = &bfreelist[BQ_AGE];
-               else
-                       flist = &bfreelist[BQ_LRU];
-               binstailfree(bp, flist);
+       rbp = NULL;
+       if (lblkno != vp->v_lastr + 1 && lblkno != 0)
+               vp->v_ralen = max(vp->v_ralen >> 1, 1);
+       else if ((ioblkno + 1) * size < filesize && !alreadyincore &&
+           !(error = VOP_BMAP(vp, ioblkno, NULL, &blkno, &num_ra))) {
+               /*
+                * Reading sequentially, and the next block is not in the
+                * cache.  We are going to try reading ahead. If this is
+                * the first read of a file, then limit read-ahead to a
+                * single block, else read as much as we're allowed.
+                */
+               if (num_ra > vp->v_ralen) {
+                       num_ra = vp->v_ralen;
+                       vp->v_ralen = min(MAXPHYS / size, vp->v_ralen << 1);
+               } else 
+                       vp->v_ralen = num_ra + 1;
+
+
+               if (num_ra)                             /* case 2, 4 */
+                       rbp = cluster_rbuild(vp, filesize,
+                           bp, ioblkno, blkno, size, num_ra, flags);
+               else if (lblkno != 0 && ioblkno == lblkno) {
+                       /* Case 5: check how many blocks to read ahead */
+                       ++ioblkno;
+                       if ((ioblkno + 1) * size > filesize ||
+                           (error = VOP_BMAP(vp,
+                           ioblkno, NULL, &blkno, &num_ra)))
+                               goto skip_readahead;
+                       flags |= B_ASYNC;
+                       if (num_ra)
+                               rbp = cluster_rbuild(vp, filesize,
+                                   NULL, ioblkno, blkno, size, num_ra, flags);
+                       else {
+                               rbp = getblk(vp, ioblkno, size, 0, 0);
+                               rbp->b_flags |= flags;
+                               rbp->b_blkno = blkno;
+                       }
+               } else if (lblkno != 0) {
+                       /* case 2; read ahead single block */
+                       rbp = getblk(vp, ioblkno, size, 0, 0);
+                       rbp->b_flags |= flags;
+                       rbp->b_blkno = blkno;
+               } else if (bp)                          /* case 1, 3, block 0 */
+                       bp->b_blkno = blkno;
+               /* Case 1 on block 0; not really doing sequential I/O */
+
+               if (rbp == bp)          /* case 4 */
+                       rbp = NULL;
+               else if (rbp) {                 /* case 2, 5 */
+                       trace(TR_BREADMISSRA,
+                           pack(vp, (num_ra + 1) * size), ioblkno);
+                       curproc->p_stats->p_ru.ru_inblock++;    /* XXX */
+               }

        }
        }

-       bp->b_flags &= ~(B_WANTED|B_BUSY|B_ASYNC|B_AGE);
-       splx(s);
-}

 
 

-/*
- * See if the block is associated with some buffer
- * (mainly to avoid getting hung up on a wait in breada)
- */
-incore(dev, blkno)
-       dev_t dev;
-       daddr_t blkno;
-{
-       register struct buf *bp;
-       register struct buf *dp;
-
-       dp = BUFHASH(dev, blkno);
-       for (bp = dp->b_forw; bp != dp; bp = bp->b_forw)
-               if (bp->b_blkno == blkno && bp->b_dev == dev &&
-                   (bp->b_flags & B_INVAL) == 0)
-                       return (1);
-       return (0);
-}
-
-struct buf *
-baddr(dev, blkno, size)
-       dev_t dev;
-       daddr_t blkno;
-       int size;
-{
+       /* XXX Kirk, do we need to make sure the bp has creds? */
+skip_readahead:
+       if (bp)
+               if (bp->b_flags & (B_DONE | B_DELWRI))
+                       panic("cluster_read: DONE bp");
+               else 
+                       error = VOP_STRATEGY(bp);
+
+       if (rbp)
+               if (error || rbp->b_flags & (B_DONE | B_DELWRI)) {
+                       rbp->b_flags &= ~(B_ASYNC | B_READ);
+                       brelse(rbp);
+               } else
+                       (void) VOP_STRATEGY(rbp);

 
 

-       if (incore(dev, blkno))
-               return (bread(dev, blkno, size));
-       return (0);
+       if (bp)
+               return(biowait(bp));
+       return(error);

 }
 
 /*
 }
 
 /*

- * Assign a buffer for the given block.  If the appropriate
- * block is already associated, return it; otherwise search
- * for the oldest non-busy buffer and reassign it.
- *
- * We use splx here because this routine may be called
- * on the interrupt stack during a dump, and we don't
- * want to lower the ipl back to 0.
+ * If blocks are contiguous on disk, use this to provide clustered
+ * read ahead.  We will read as many blocks as possible sequentially
+ * and then parcel them up into logical blocks in the buffer hash table.

  */
 struct buf *
  */
 struct buf *

-getblk(dev, blkno, size)
-       dev_t dev;
+cluster_rbuild(vp, filesize, bp, lbn, blkno, size, run, flags)
+       struct vnode *vp;
+       u_quad_t filesize;
+       struct buf *bp;
+       daddr_t lbn;

        daddr_t blkno;
        daddr_t blkno;

-       int size;
+       long size;
+       int run;
+       long flags;

 {
 {

-       register struct buf *bp, *dp;
-       int s;
-
-       if ((unsigned)blkno >= 1 << (sizeof(int)*NBBY-PGSHIFT)) /* XXX */
-               blkno = 1 << ((sizeof(int)*NBBY-PGSHIFT) + 1);
-       /*
-        * Search the cache for the block.  If we hit, but
-        * the buffer is in use for i/o, then we wait until
-        * the i/o has completed.
-        */
-       dp = BUFHASH(dev, blkno);
-loop:
-       for (bp = dp->b_forw; bp != dp; bp = bp->b_forw) {
-               if (bp->b_blkno != blkno || bp->b_dev != dev ||
-                   bp->b_flags&B_INVAL)
-                       continue;
-               s = spl6();
-               if (bp->b_flags&B_BUSY) {
-                       bp->b_flags |= B_WANTED;
-                       sleep((caddr_t)bp, PRIBIO+1);
-                       splx(s);
-                       goto loop;
+       struct cluster_save *b_save;
+       struct buf *tbp;
+       daddr_t bn;
+       int i, inc;
+
+#ifdef DIAGNOSTIC
+       if (size != vp->v_mount->mnt_stat.f_iosize)
+               panic("cluster_rbuild: size %d != filesize %d\n",
+                       size, vp->v_mount->mnt_stat.f_iosize);
+#endif
+       if (size * (lbn + run + 1) > filesize)
+               --run;
+       if (run == 0) {
+               if (!bp) {
+                       bp = getblk(vp, lbn, size, 0, 0);
+                       bp->b_blkno = blkno;
+                       bp->b_flags |= flags;

                }
                }

-               splx(s);
-               notavail(bp);
-               if (brealloc(bp, size) == 0)
-                       goto loop;
-               bp->b_flags |= B_CACHE;

                return(bp);
        }
                return(bp);
        }

-       if (major(dev) >= nblkdev)
-               panic("blkdev");
-       bp = getnewbuf();
-       bfree(bp);
-       bremhash(bp);
-       binshash(bp, dp);
-       bp->b_dev = dev;
-       bp->b_blkno = blkno;
-       bp->b_error = 0;
-       if (brealloc(bp, size) == 0)
-               goto loop;
-       return(bp);
-}
-
-/*
- * get an empty block,
- * not assigned to any particular device
- */
-struct buf *
-geteblk(size)
-       int size;
-{
-       register struct buf *bp, *flist;
-
-loop:
-       bp = getnewbuf();
-       bp->b_flags |= B_INVAL;
-       bfree(bp);
-       bremhash(bp);
-       flist = &bfreelist[BQ_AGE];
-       binshash(bp, flist);
-       bp->b_dev = (dev_t)NODEV;
-       bp->b_error = 0;
-       if (brealloc(bp, size) == 0)
-               goto loop;
-       return(bp);
-}
-
-/*
- * Allocate space associated with a buffer.
- * If can't get space, buffer is released
- */
-brealloc(bp, size)
-       register struct buf *bp;
-       int size;
-{
-       daddr_t start, last;
-       register struct buf *ep;
-       struct buf *dp;
-       int s;
-
-       /*
-        * First need to make sure that all overlaping previous I/O
-        * is dispatched with.
-        */
-       if (size == bp->b_bcount)
-               return (1);
-       if (size < bp->b_bcount) { 
-               if (bp->b_flags & B_DELWRI) {
-                       bwrite(bp);
-                       return (0);
-               }
-               if (bp->b_flags & B_LOCKED)
-                       panic("brealloc");
-               return (allocbuf(bp, size));
-       }
-       bp->b_flags &= ~B_DONE;
-       if (bp->b_dev == NODEV)
-               return (allocbuf(bp, size));

 
 

-       /*
-        * Search cache for any buffers that overlap the one that we
-        * are trying to allocate. Overlapping buffers must be marked
-        * invalid, after being written out if they are dirty. (indicated
-        * by B_DELWRI) A disk block must be mapped by at most one buffer
-        * at any point in time. Care must be taken to avoid deadlocking
-        * when two buffer are trying to get the same set of disk blocks.
-        */
-       start = bp->b_blkno;
-       last = start + (size / DEV_BSIZE) - 1;
-       dp = BUFHASH(bp->b_dev, bp->b_blkno);
-loop:
-       for (ep = dp->b_forw; ep != dp; ep = ep->b_forw) {
-               if (ep == bp || ep->b_dev != bp->b_dev || (ep->b_flags&B_INVAL))
-                       continue;
-               /* look for overlap */
-               if (ep->b_bcount == 0 || ep->b_blkno > last ||
-                   ep->b_blkno + (ep->b_bcount / DEV_BSIZE) <= start)
-                       continue;
-               s = spl6();
-               if (ep->b_flags&B_BUSY) {
-                       ep->b_flags |= B_WANTED;
-                       sleep((caddr_t)ep, PRIBIO+1);
-                       splx(s);
-                       goto loop;
+       bp = cluster_newbuf(vp, bp, flags, blkno, lbn, size, run + 1);
+       if (bp->b_flags & (B_DONE | B_DELWRI))
+               return (bp);
+
+       b_save = malloc(sizeof(struct buf *) * run + sizeof(struct cluster_save),
+           M_SEGMENT, M_WAITOK);
+       b_save->bs_bufsize = b_save->bs_bcount = size;
+       b_save->bs_nchildren = 0;
+       b_save->bs_children = (struct buf **)(b_save + 1);
+       b_save->bs_saveaddr = bp->b_saveaddr;
+       bp->b_saveaddr = (caddr_t) b_save;
+
+       inc = size / DEV_BSIZE;
+       for (bn = blkno + inc, i = 1; i <= run; ++i, bn += inc) {
+               if (incore(vp, lbn + i)) {
+                       if (i == 1) {
+                               bp->b_saveaddr = b_save->bs_saveaddr;
+                               bp->b_flags &= ~B_CALL;
+                               bp->b_iodone = NULL;
+                               allocbuf(bp, size);
+                               free(b_save, M_SEGMENT);
+                       } else
+                               allocbuf(bp, size * i);
+                       break;

                }
                }

-               splx(s);
-               notavail(ep);
-               if (ep->b_flags & B_DELWRI) {
-                       bwrite(ep);
-                       goto loop;
+               tbp = getblk(vp, lbn + i, 0, 0, 0);
+               tbp->b_bcount = tbp->b_bufsize = size;
+               tbp->b_blkno = bn;
+               {
+                       daddr_t temp;
+                       VOP_BMAP(tbp->b_vp, tbp->b_lblkno, NULL, &temp, NULL);
+                       if (temp != bn) {
+                               printf("Block: %d Assigned address: %x Bmap address: %x\n",
+                                           tbp->b_lblkno, tbp->b_blkno, temp);
+                               panic("cluster_rbuild: wrong disk address");
+                       }

                }
                }

-               ep->b_flags |= B_INVAL;
-               brelse(ep);
+               tbp->b_flags |= flags | B_READ | B_ASYNC;
+               ++b_save->bs_nchildren;
+               b_save->bs_children[i - 1] = tbp;

        }
        }

-       return (allocbuf(bp, size));
+       if (!(bp->b_flags & B_ASYNC))
+               vp->v_ralen = max(vp->v_ralen - 1, 1);
+       return(bp);

 }
 
 /*
 }
 
 /*

- * Expand or contract the actual memory allocated to a buffer.
- * If no memory is available, release buffer and take error exit
+ * Either get a new buffer or grow the existing one.

  */
  */

-allocbuf(tp, size)
-       register struct buf *tp;
-       int size;
+struct buf *
+cluster_newbuf(vp, bp, flags, blkno, lblkno, size, run)
+       struct vnode *vp;
+       struct buf *bp;
+       long flags;
+       daddr_t blkno;
+       daddr_t lblkno;
+       long size;
+       int run;

 {
 {

-       register struct buf *bp, *ep;
-       int sizealloc, take;
-#ifdef sun
-       register char *a;
-       int osize;
-#endif
-
-#ifndef sun
-       sizealloc = roundup(size, CLBYTES);
-#else
-       sizealloc = roundup(size, BUFALLOCSIZE);
-#endif
-       /*
-        * Buffer size does not change
-        */
-       if (sizealloc == tp->b_bufsize)
-               goto out;
-#ifndef sun
-       /*
-        * Buffer size is shrinking.
-        * Place excess space in a buffer header taken from the
-        * BQ_EMPTY buffer list and placed on the "most free" list.
-        * If no extra buffer headers are available, leave the
-        * extra space in the present buffer.
-        */
-       if (sizealloc < tp->b_bufsize) {
-               ep = bfreelist[BQ_EMPTY].av_forw;
-               if (ep == &bfreelist[BQ_EMPTY])
-                       goto out;
-               notavail(ep);
-               pagemove(tp->b_un.b_addr + sizealloc, ep->b_un.b_addr,
-                   (int)tp->b_bufsize - sizealloc);
-               ep->b_bufsize = tp->b_bufsize - sizealloc;
-               tp->b_bufsize = sizealloc;
-               ep->b_flags |= B_INVAL;
-               ep->b_bcount = 0;
-               brelse(ep);
-               goto out;
-       }
-       /*
-        * More buffer space is needed. Get it out of buffers on
-        * the "most free" list, placing the empty headers on the
-        * BQ_EMPTY buffer header list.
-        */
-       while (tp->b_bufsize < sizealloc) {
-               take = sizealloc - tp->b_bufsize;
-               bp = getnewbuf();
-               if (take >= bp->b_bufsize)
-                       take = bp->b_bufsize;
-               pagemove(&bp->b_un.b_addr[bp->b_bufsize - take],
-                   &tp->b_un.b_addr[tp->b_bufsize], take);
-               tp->b_bufsize += take;
-               bp->b_bufsize = bp->b_bufsize - take;
-               if (bp->b_bcount > bp->b_bufsize)
-                       bp->b_bcount = bp->b_bufsize;
-               if (bp->b_bufsize <= 0) {
-                       bremhash(bp);
-                       binshash(bp, &bfreelist[BQ_EMPTY]);
-                       bp->b_dev = (dev_t)NODEV;
-                       bp->b_error = 0;
-                       bp->b_flags |= B_INVAL;
+       if (!bp) {
+               bp = getblk(vp, lblkno, size, 0, 0);
+               if (bp->b_flags & (B_DONE | B_DELWRI)) {
+                       bp->b_blkno = blkno;
+                       return(bp);

                }
                }

-               brelse(bp);
-       }
-#else
-       /*
-        * Buffer size is shrinking
-        * Just put the tail end back in the map
-        */
-       if (sizealloc < tp->b_bufsize) {
-               rmfree(buffermap, (long)(tp->b_bufsize - sizealloc),
-                       (long)(tp->b_un.b_addr + sizealloc));
-               tp->b_bufsize = sizealloc;
-               goto out;

        }
        }

-       /*
-        * Buffer is being expanded or created
-        * If being expanded, attempt to get contiguous
-        * section, otherwise get a new chunk and copy.
-        * If no space, free up a buffer on the AGE list
-        * and try again.
-        */
-       do {
-               if ((osize = tp->b_bufsize)) {
-                       a = (char *)rmget(buffermap, (long)(sizealloc-osize),
-                               (long)(tp->b_un.b_addr + osize));
-                       if (a == 0) {
-                               a = (char *)rmalloc(buffermap, (long)sizealloc);
-                               if (a != 0) {
-                                       bcopy(tp->b_un.b_addr, a, osize);
-                                       rmfree(buffermap, (long)osize,
-                                               (long)tp->b_un.b_addr);
-                                       tp->b_un.b_addr = a;
-                               }
-                       }
-               } else {
-                       a = (char *)rmalloc(buffermap, (long)sizealloc);
-                       if (a != 0)
-                               tp->b_un.b_addr = a;
-               }
-       } while (a == 0 && bfreemem());
-       if (a == 0) {
-               brelse(tp);
-               return (0);
-       }
-       tp->b_bufsize = sizealloc;
-#endif
-out:
-       tp->b_bcount = size;
-       return (1);
+       allocbuf(bp, run * size);
+       bp->b_blkno = blkno;
+       bp->b_iodone = cluster_callback;
+       bp->b_flags |= flags | B_CALL;
+       return(bp);

 }
 
 /*
 }
 
 /*

- * Release space associated with a buffer.
+ * Cleanup after a clustered read or write.

  */
  */

-bfree(bp)
+void
+cluster_callback(bp)

        struct buf *bp;
 {
        struct buf *bp;
 {

-#ifdef sun
-       if (bp->b_bufsize) {
-               rmfree(buffermap, (long)bp->b_bufsize, (long)bp->b_un.b_addr);
-               bp->b_bufsize = 0;
+       struct cluster_save *b_save;
+       struct buf **tbp;
+       long bsize;
+       caddr_t cp;
+       daddr_t daddr;
+       b_save = (struct cluster_save *)(bp->b_saveaddr);
+       bp->b_saveaddr = b_save->bs_saveaddr;
+
+       cp = bp->b_un.b_addr + b_save->bs_bufsize;
+       daddr = bp->b_blkno + b_save->bs_bufsize / DEV_BSIZE;
+       for (tbp = b_save->bs_children; b_save->bs_nchildren--; ++tbp) {
+               pagemove(cp, (*tbp)->b_un.b_addr, (*tbp)->b_bufsize);
+               cp += (*tbp)->b_bufsize;
+               bp->b_bufsize -= (*tbp)->b_bufsize;
+               if ((*tbp)->b_blkno != daddr) {
+                       struct inode *ip;
+                       printf("cluster_callback: bad disk address:\n");
+                       printf("Clustered Block: %d DiskAddr: %x bytes left: %d\n",
+                           bp->b_lblkno, bp->b_blkno, bp->b_bufsize);
+                       printf("\toriginal size: %d flags: %x\n", bp->b_bcount,
+                           bp->b_flags);
+                       printf("Child Block: %d DiskAddr: %x bytes: %d\n",
+                           (*tbp)->b_lblkno, (*tbp)->b_blkno,
+                           (*tbp)->b_bufsize);
+                       ip = VTOI((*tbp)->b_vp);
+                       printf("daddr: %x i_size %qd\n", daddr, ip->i_size);
+                       if ((*tbp)->b_lblkno < NDADDR)
+                               printf("Child block pointer from inode: %x\n",
+                                   ip->i_din.di_db[(*tbp)->b_lblkno]);
+                       spl0();
+                       panic ("cluster_callback: bad disk address");
+               }
+               daddr += (*tbp)->b_bufsize / DEV_BSIZE;
+               biodone(*tbp);

        }
        }

+#ifdef DIAGNOSTIC
+       if (bp->b_bufsize != b_save->bs_bufsize)
+               panic ("cluster_callback: more space to reclaim");

 #endif
 #endif

-       bp->b_bcount = 0;
+       bp->b_bcount = bp->b_bufsize;
+       bp->b_iodone = NULL;
+       free(b_save, M_SEGMENT);
+       if (bp->b_flags & B_ASYNC)
+               brelse(bp);
+       else
+               wakeup((caddr_t)bp);

 }
 
 }
 

-#ifdef sun

 /*
 /*

- * Attempt to free up buffer space by flushing
- * something in the free list.
- * Don't wait for something, that could cause deadlocks
- * We start with BQ_AGE because we know BQ_EMPTY take no memory.
+ * Do clustered write for FFS.
+ *
+ * Three cases:
+ *     1. Write is not sequential (write asynchronously)
+ *     Write is sequential:
+ *     2.      beginning of cluster - begin cluster
+ *     3.      middle of a cluster - add to cluster
+ *     4.      end of a cluster - asynchronously write cluster

  */
  */

-bfreemem()
+void
+cluster_write(bp, filesize)
+        struct buf *bp;
+       u_quad_t filesize;

 {
 {

-       register struct buf *bp, *dp;
-       int s;
-
-loop:
-       s = spl6();
-       for (dp = &bfreelist[BQ_AGE]; dp > bfreelist; dp--)
-               if (dp->av_forw != dp)
-                       break;
-       splx(s);
-       if (dp == bfreelist) {          /* no free blocks */
-               return (0);
-       }
-       bp = dp->av_forw;
-       notavail(bp);
-       if (bp->b_flags & B_DELWRI) {
-               bp->b_flags |= B_ASYNC;
-               bwrite(bp);
-               goto loop;
-       }
-       trace(TR_BRELSE, bp->b_dev, bp->b_blkno);
-       bp->b_flags = B_BUSY | B_INVAL;
-       bfree(bp);
-       bremhash(bp);
-       binshash(bp, &bfreelist[BQ_EMPTY]);
-       bp->b_dev = (dev_t)NODEV;
-       bp->b_error = 0;
-       brelse(bp);
-       return (1);
+        struct vnode *vp;
+        daddr_t lbn;
+        int clen;
+
+        vp = bp->b_vp;
+        lbn = bp->b_lblkno;
+
+       /* Initialize vnode to beginning of file. */
+       if (lbn == 0)
+               vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
+
+        if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 ||
+           (bp->b_blkno != vp->v_lasta + bp->b_bcount / DEV_BSIZE)) {
+               if (vp->v_clen != 0)
+                       /*
+                        * Write is not sequential.
+                        */
+                       cluster_wbuild(vp, NULL, bp->b_bcount, vp->v_cstart,
+                           vp->v_lastw - vp->v_cstart + 1, lbn);
+               /*
+                * Consider beginning a cluster.
+                */
+               if ((lbn + 1) * bp->b_bcount == filesize)
+                       /* End of file, make cluster as large as possible */
+                       clen = MAXBSIZE / vp->v_mount->mnt_stat.f_iosize - 1;
+               else if (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &clen)) {
+                       bawrite(bp);
+                       vp->v_clen = 0;
+                       vp->v_lasta = bp->b_blkno;
+                       vp->v_cstart = lbn + 1;
+                       vp->v_lastw = lbn;
+                       return;
+               } else
+                       clen = 0;
+                vp->v_clen = clen;
+                if (clen == 0) {               /* I/O not contiguous */
+                       vp->v_cstart = lbn + 1;
+                        bawrite(bp);
+                } else {                       /* Wait for rest of cluster */
+                       vp->v_cstart = lbn;
+                        bdwrite(bp);
+               }
+        } else if (lbn == vp->v_cstart + vp->v_clen) {
+               /*
+                * At end of cluster, write it out.
+                */
+               cluster_wbuild(vp, bp, bp->b_bcount, vp->v_cstart,
+                   vp->v_clen + 1, lbn);
+               vp->v_clen = 0;
+               vp->v_cstart = lbn + 1;
+        } else
+               /*
+                * In the middle of a cluster, so just delay the
+                * I/O for now.
+                */
+                bdwrite(bp);
+        vp->v_lastw = lbn;
+       vp->v_lasta = bp->b_blkno;

 }
 }

-#endif
+

 
 /*
 
 /*

- * Find a buffer which is available for use.
- * Select something from a free list.
- * Preference is to AGE list, then LRU list.
+ * This is an awful lot like cluster_rbuild...wish they could be combined.
+ * The last lbn argument is the current block on which I/O is being
+ * performed.  Check to see that it doesn't fall in the middle of
+ * the current block.

  */
  */

-struct buf *
-getnewbuf()
+void
+cluster_wbuild(vp, last_bp, size, start_lbn, len, lbn)
+       struct vnode *vp;
+       struct buf *last_bp;
+       long size;
+       daddr_t start_lbn;
+       int len;
+       daddr_t lbn;

 {
 {

-       register struct buf *bp, *dp;
-       int s;
-
-loop:
-       s = spl6();
-#ifndef sun
-       for (dp = &bfreelist[BQ_AGE]; dp > bfreelist; dp--)
-#else
-       for (dp = &bfreelist[BQ_EMPTY]; dp > bfreelist; dp--)
+       struct cluster_save *b_save;
+       struct buf *bp, *tbp;
+       caddr_t cp;
+       int i, s;
+
+#ifdef DIAGNOSTIC
+       if (size != vp->v_mount->mnt_stat.f_iosize)
+               panic("cluster_wbuild: size %d != filesize %d\n",
+                       size, vp->v_mount->mnt_stat.f_iosize);

 #endif
 #endif

-               if (dp->av_forw != dp)
-                       break;
-       if (dp == bfreelist) {          /* no free blocks */
-               dp->b_flags |= B_WANTED;
-               sleep((caddr_t)dp, PRIBIO+1);
-               goto loop;
-       }
-       splx(s);
-       bp = dp->av_forw;
-       notavail(bp);
-       if (bp->b_flags & B_DELWRI) {
-               bp->b_flags |= B_ASYNC;
-               bwrite(bp);
-               goto loop;
+redo:
+       while ((!incore(vp, start_lbn) || start_lbn == lbn) && len) {
+               ++start_lbn;
+               --len;

        }
        }

-       trace(TR_BRELSE, bp->b_dev, bp->b_blkno);
-       bp->b_flags = B_BUSY;
-       return (bp);
-}

 
 

-/*
- * Wait for I/O completion on the buffer; return errors
- * to the user.
- */
-biowait(bp)
-       register struct buf *bp;
-{
-       int s;
-
-       s = spl6();
-       while ((bp->b_flags&B_DONE)==0)
-               sleep((caddr_t)bp, PRIBIO);
-       splx(s);
-       if (u.u_error == 0)                     /* XXX */
-               u.u_error = geterror(bp);
-}
-
-/*
- * Mark I/O complete on a buffer. If the header
- * indicates a dirty page push completion, the
- * header is inserted into the ``cleaned'' list
- * to be processed by the pageout daemon. Otherwise
- * release it if I/O is asynchronous, and wake 
- * up anyone waiting for it.
- */
-biodone(bp)
-       register struct buf *bp;
-{
-       register int s;
-
-       if (bp->b_flags & B_DONE)
-               panic("dup biodone");
-       bp->b_flags |= B_DONE;
-       if (bp->b_flags & B_DIRTY) {
-               if (bp->b_flags & B_ERROR)
-                       panic("IO err in push");
-               s = spl6();
-               bp->av_forw = bclnlist;
-               bp->b_bcount = swsize[bp - swbuf];
-               bp->b_pfcent = swpf[bp - swbuf];
-               cnt.v_pgout++;
-               cnt.v_pgpgout += bp->b_bcount / NBPG;
-               bclnlist = bp;
-               if (bswlist.b_flags & B_WANTED)
-                       wakeup((caddr_t)&proc[2]);
-               splx(s);
-               return;
-       }
-       if (bp->b_flags & B_CALL) {
-               bp->b_flags &= ~B_CALL;
-               (*bp->b_iodone)(bp);
+       /* Get more memory for current buffer */
+       if (len <= 1) {
+               if (last_bp) {
+                       bawrite(last_bp);
+               } else if (len) {
+                       bp = getblk(vp, start_lbn, size, 0, 0);
+                       bawrite(bp);
+               }

                return;
        }
                return;
        }

-       if (bp->b_flags&B_ASYNC)
+
+       bp = getblk(vp, start_lbn, size, 0, 0);
+       if (!(bp->b_flags & B_DELWRI)) {
+               ++start_lbn;
+               --len;

                brelse(bp);
                brelse(bp);

-       else {
-               bp->b_flags &= ~B_WANTED;
-               wakeup((caddr_t)bp);
+               goto redo;

        }
        }

-}

 
 

-/*
- * Insure that no part of a specified block is in an incore buffer.
- */
-blkflush(dev, blkno, size)
-       dev_t dev;
-       daddr_t blkno;
-       long size;
-{
-       register struct buf *ep;
-       struct buf *dp;
-       daddr_t start, last;
-       int s;
-
-       start = blkno;
-       last = start + (size / DEV_BSIZE) - 1;
-       dp = BUFHASH(dev, blkno);
-loop:
-       for (ep = dp->b_forw; ep != dp; ep = ep->b_forw) {
-               if (ep->b_dev != dev || (ep->b_flags&B_INVAL))
-                       continue;
-               /* look for overlap */
-               if (ep->b_bcount == 0 || ep->b_blkno > last ||
-                   ep->b_blkno + (ep->b_bcount / DEV_BSIZE) <= start)
-                       continue;
-               s = spl6();
-               if (ep->b_flags&B_BUSY) {
-                       ep->b_flags |= B_WANTED;
-                       sleep((caddr_t)ep, PRIBIO+1);
-                       splx(s);
-                       goto loop;
-               }
-               if (ep->b_flags & B_DELWRI) {
-                       splx(s);
-                       notavail(ep);
-                       bwrite(ep);
-                       goto loop;
-               }
-               splx(s);
-       }
-}
+       --len;
+       b_save = malloc(sizeof(struct buf *) * len + sizeof(struct cluster_save),
+           M_SEGMENT, M_WAITOK);
+       b_save->bs_bcount = bp->b_bcount;
+       b_save->bs_bufsize = bp->b_bufsize;
+       b_save->bs_nchildren = 0;
+       b_save->bs_children = (struct buf **)(b_save + 1);
+       b_save->bs_saveaddr = bp->b_saveaddr;
+       bp->b_saveaddr = (caddr_t) b_save;
+
+
+       bp->b_flags |= B_CALL;
+       bp->b_iodone = cluster_callback;
+       cp = bp->b_un.b_addr + bp->b_bufsize;
+       for (++start_lbn, i = 0; i < len; ++i, ++start_lbn) {
+               if (!incore(vp, start_lbn) || start_lbn == lbn)
+                       break;

 
 

-/*
- * make sure all write-behind blocks
- * on dev (or NODEV for all)
- * are flushed out.
- * (from umount and update)
- * (and temporarily pagein)
- */
-bflush(dev)
-       dev_t dev;
-{
-       register struct buf *bp;
-       register struct buf *flist;
-       int s;
-
-loop:
-       s = spl6();
-       for (flist = bfreelist; flist < &bfreelist[BQ_EMPTY]; flist++)
-       for (bp = flist->av_forw; bp != flist; bp = bp->av_forw) {
-               if ((bp->b_flags & B_DELWRI) == 0)
-                       continue;
-               if (dev == NODEV || dev == bp->b_dev) {
-                       bp->b_flags |= B_ASYNC;
-                       notavail(bp);
-                       bwrite(bp);
-                       goto loop;
+               if (last_bp == NULL || start_lbn != last_bp->b_lblkno) {
+                       tbp = getblk(vp, start_lbn, size, 0, 0);
+#ifdef DIAGNOSTIC
+                       if (tbp->b_bcount != tbp->b_bufsize)
+                               panic("cluster_wbuild: Buffer too big");
+#endif
+                       if (!(tbp->b_flags & B_DELWRI)) {
+                               brelse(tbp);
+                               break;
+                       }
+               } else
+                       tbp = last_bp;
+
+               ++b_save->bs_nchildren;
+
+               /* Move memory from children to parent */
+               if (tbp->b_blkno != (bp->b_blkno + bp->b_bufsize / DEV_BSIZE)) {
+                       printf("Clustered Block: %d addr %x bufsize: %d\n",
+                           bp->b_lblkno, bp->b_blkno, bp->b_bufsize);
+                       printf("Child Block: %d addr: %x\n", tbp->b_lblkno,
+                           tbp->b_blkno);
+                       panic("Clustered write to wrong blocks");

                }
                }

-       }
-       splx(s);
-}

 
 

-/*
- * Pick up the device's error number and pass it to the user;
- * if there is an error but the number is 0 set a generalized
- * code.  Actually the latter is always true because devices
- * don't yet return specific errors.
- */
-geterror(bp)
-       register struct buf *bp;
-{
-       int error = 0;
+               pagemove(tbp->b_un.b_daddr, cp, size);
+               bp->b_bcount += size;
+               bp->b_bufsize += size;

 
 

-       if (bp->b_flags&B_ERROR)
-               if ((error = bp->b_error)==0)
-                       return (EIO);
-       return (error);
-}
+               tbp->b_flags &= ~(B_READ | B_DONE | B_ERROR | B_DELWRI);
+               tbp->b_flags |= B_ASYNC;
+               s = splbio();
+               reassignbuf(tbp, tbp->b_vp);            /* put on clean list */
+               ++tbp->b_vp->v_numoutput;
+               splx(s);
+               b_save->bs_children[i] = tbp;

 
 

-/*
- * Invalidate in core blocks belonging to closed or umounted filesystem
- *
- * This is not nicely done at all - the buffer ought to be removed from the
- * hash chains & have its dev/blkno fields clobbered, but unfortunately we
- * can't do that here, as it is quite possible that the block is still
- * being used for i/o. Eventually, all disc drivers should be forced to
- * have a close routine, which ought ensure that the queue is empty, then
- * properly flush the queues. Until that happy day, this suffices for
- * correctness.                                                ... kre
- */
-binval(dev)
-       dev_t dev;
-{
-       register struct buf *bp;
-       register struct bufhd *hp;
-#define dp ((struct buf *)hp)
-
-       for (hp = bufhash; hp < &bufhash[BUFHSZ]; hp++)
-               for (bp = dp->b_forw; bp != dp; bp = bp->b_forw)
-                       if (bp->b_dev == dev)
-                               bp->b_flags |= B_INVAL;
+               cp += tbp->b_bufsize;
+       }
+
+       if (i == 0) {
+               /* None to cluster */
+               bp->b_saveaddr = b_save->bs_saveaddr;
+               bp->b_flags &= ~B_CALL;
+               bp->b_iodone = NULL;
+               free(b_save, M_SEGMENT);
+       }
+       bawrite(bp);
+       if (i < len) {
+               len -= i + 1;
+               start_lbn += 1;
+               goto redo;
+       }

 }
 }