another missing splx; from shannon@sun.UUCP
[unix-history] / usr / src / sys / kern / vfs_cluster.c
index 07615dd..0184865 100644 (file)
@@ -1,4 +1,6 @@
-/*     vfs_cluster.c   4.1     %G%     */
+/*     vfs_cluster.c   4.44    83/05/02        */
+
+#include "../machine/pte.h"
 
 #include "../h/param.h"
 #include "../h/systm.h"
 
 #include "../h/param.h"
 #include "../h/systm.h"
 #include "../h/conf.h"
 #include "../h/proc.h"
 #include "../h/seg.h"
 #include "../h/conf.h"
 #include "../h/proc.h"
 #include "../h/seg.h"
-#include "../h/pte.h"
 #include "../h/vm.h"
 #include "../h/vm.h"
-
-/*
- * The following several routines allocate and free
- * buffers with various side effects.  In general the
- * arguments to an allocate routine are a device and
- * a block number, and the value is a pointer to
- * to the buffer header; the buffer is marked "busy"
- * so that no one else can touch it.  If the block was
- * already in core, no I/O need be done; if it is
- * already busy, the process waits until it becomes free.
- * The following routines allocate a buffer:
- *     getblk
- *     bread
- *     breada
- *     baddr   (if it is incore)
- * Eventually the buffer must be released, possibly with the
- * side effect of writing it out, by using one of
- *     bwrite
- *     bdwrite
- *     bawrite
- *     brelse
- */
-
-#define        BUFHSZ  63
-#define        BUFHASH(blkno)  (blkno % BUFHSZ)
-short  bufhash[BUFHSZ];
-
-/*
- * Initialize hash links for buffers.
- */
-bhinit()
-{
-       register int i;
-
-       for (i = 0; i < BUFHSZ; i++)
-               bufhash[i] = -1;
-}
-
-/* #define     DISKMON 1 */
-
-#ifdef DISKMON
-struct {
-       int     nbuf;
-       long    nread;
-       long    nreada;
-       long    ncache;
-       long    nwrite;
-       long    bufcount[NBUF];
-} io_info;
-#endif
-
-/*
- * Swap IO headers -
- * They contain the necessary information for the swap I/O.
- * At any given time, a swap header can be in three
- * different lists. When free it is in the free list, 
- * when allocated and the I/O queued, it is on the swap 
- * device list, and finally, if the operation was a dirty 
- * page push, when the I/O completes, it is inserted 
- * in a list of cleaned pages to be processed by the pageout daemon.
- */
-struct buf swbuf[NSWBUF];
-short  swsize[NSWBUF];         /* CAN WE JUST USE B_BCOUNT? */
-int    swpf[NSWBUF];
-
-
-#ifdef FASTVAX
-#define        notavail(bp) \
-{ \
-       int s = spl6(); \
-       (bp)->av_back->av_forw = (bp)->av_forw; \
-       (bp)->av_forw->av_back = (bp)->av_back; \
-       (bp)->b_flags |= B_BUSY; \
-       splx(s); \
-}
-#endif
+#include "../h/trace.h"
 
 /*
  * Read in (if necessary) the block and return a buffer pointer.
  */
 struct buf *
 
 /*
  * Read in (if necessary) the block and return a buffer pointer.
  */
 struct buf *
-bread(dev, blkno)
-dev_t dev;
-daddr_t blkno;
+bread(dev, blkno, size)
+       dev_t dev;
+       daddr_t blkno;
+       int size;
 {
        register struct buf *bp;
 
 {
        register struct buf *bp;
 
-       bp = getblk(dev, blkno);
+       if (size == 0)
+               panic("bread: size 0");
+       bp = getblk(dev, blkno, size);
        if (bp->b_flags&B_DONE) {
        if (bp->b_flags&B_DONE) {
-#ifdef DISKMON
-               io_info.ncache++;
-#endif
+               trace(TR_BREADHIT, dev, blkno);
                return(bp);
        }
        bp->b_flags |= B_READ;
                return(bp);
        }
        bp->b_flags |= B_READ;
-       bp->b_bcount = BSIZE;
+       if (bp->b_bcount > bp->b_bufsize)
+               panic("bread");
        (*bdevsw[major(dev)].d_strategy)(bp);
        (*bdevsw[major(dev)].d_strategy)(bp);
-#ifdef DISKMON
-       io_info.nread++;
-#endif
-       u.u_vm.vm_inblk++;              /* pay for read */
-       iowait(bp);
+       trace(TR_BREADMISS, dev, blkno);
+       u.u_ru.ru_inblock++;            /* pay for read */
+       biowait(bp);
        return(bp);
 }
 
        return(bp);
 }
 
@@ -120,43 +46,60 @@ daddr_t blkno;
  * read-ahead block (which is not allocated to the caller)
  */
 struct buf *
  * read-ahead block (which is not allocated to the caller)
  */
 struct buf *
-breada(dev, blkno, rablkno)
-dev_t dev;
-daddr_t blkno, rablkno;
+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;
 {
        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)) {
        if (!incore(dev, blkno)) {
-               bp = getblk(dev, blkno);
+               bp = getblk(dev, blkno, size);
                if ((bp->b_flags&B_DONE) == 0) {
                        bp->b_flags |= B_READ;
                if ((bp->b_flags&B_DONE) == 0) {
                        bp->b_flags |= B_READ;
-                       bp->b_bcount = BSIZE;
+                       if (bp->b_bcount > bp->b_bufsize)
+                               panic("breada");
                        (*bdevsw[major(dev)].d_strategy)(bp);
                        (*bdevsw[major(dev)].d_strategy)(bp);
-#ifdef DISKMON
-                       io_info.nread++;
-#endif
-                       u.u_vm.vm_inblk++;              /* pay for read */
-               }
+                       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)) {
        if (rablkno && !incore(dev, rablkno)) {
-               rabp = getblk(dev, rablkno);
-               if (rabp->b_flags & B_DONE)
+               rabp = getblk(dev, rablkno, rabsize);
+               if (rabp->b_flags & B_DONE) {
                        brelse(rabp);
                        brelse(rabp);
-               else {
+                       trace(TR_BREADHITRA, dev, blkno);
+               } else {
                        rabp->b_flags |= B_READ|B_ASYNC;
                        rabp->b_flags |= B_READ|B_ASYNC;
-                       rabp->b_bcount = BSIZE;
+                       if (rabp->b_bcount > rabp->b_bufsize)
+                               panic("breadrabp");
                        (*bdevsw[major(dev)].d_strategy)(rabp);
                        (*bdevsw[major(dev)].d_strategy)(rabp);
-#ifdef DISKMON
-                       io_info.nreada++;
-#endif
-                       u.u_vm.vm_inblk++;              /* pay in advance */
+                       trace(TR_BREADMISSRA, dev, rablock);
+                       u.u_ru.ru_inblock++;            /* pay in advance */
                }
        }
                }
        }
-       if(bp == NULL)
-               return(bread(dev, blkno));
-       iowait(bp);
-       return(bp);
+
+       /*
+        * 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);
 }
 
 /*
 }
 
 /*
@@ -164,26 +107,29 @@ daddr_t blkno, rablkno;
  * Then release the buffer.
  */
 bwrite(bp)
  * Then release the buffer.
  */
 bwrite(bp)
-register struct buf *bp;
+       register struct buf *bp;
 {
        register flag;
 
        flag = bp->b_flags;
 {
        register flag;
 
        flag = bp->b_flags;
-       bp->b_flags &= ~(B_READ | B_DONE | B_ERROR | B_DELWRI | B_AGE);
-       bp->b_bcount = BSIZE;
-#ifdef DISKMON
-       io_info.nwrite++;
-#endif
+       bp->b_flags &= ~(B_READ | B_DONE | B_ERROR | B_DELWRI);
        if ((flag&B_DELWRI) == 0)
        if ((flag&B_DELWRI) == 0)
-               u.u_vm.vm_oublk++;              /* noone paid yet */
+               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);
        (*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) {
        if ((flag&B_ASYNC) == 0) {
-               iowait(bp);
+               biowait(bp);
                brelse(bp);
        } else if (flag & B_DELWRI)
                bp->b_flags |= B_AGE;
                brelse(bp);
        } else if (flag & B_DELWRI)
                bp->b_flags |= B_AGE;
-       else
-               geterror(bp);
 }
 
 /*
 }
 
 /*
@@ -195,14 +141,14 @@ register struct buf *bp;
  * in the same order as requested.
  */
 bdwrite(bp)
  * in the same order as requested.
  */
 bdwrite(bp)
-register struct buf *bp;
+       register struct buf *bp;
 {
 {
-       register struct buf *dp;
+       register int flags;
 
        if ((bp->b_flags&B_DELWRI) == 0)
 
        if ((bp->b_flags&B_DELWRI) == 0)
-               u.u_vm.vm_oublk++;              /* noone paid yet */
-       dp = bdevsw[major(bp->b_dev)].d_tab;
-       if(dp->b_flags & B_TAPE)
+               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;
                bawrite(bp);
        else {
                bp->b_flags |= B_DELWRI | B_DONE;
@@ -214,7 +160,7 @@ register struct buf *bp;
  * Release the buffer, start I/O on it, but don't wait for completion.
  */
 bawrite(bp)
  * Release the buffer, start I/O on it, but don't wait for completion.
  */
 bawrite(bp)
-register struct buf *bp;
+       register struct buf *bp;
 {
 
        bp->b_flags |= B_ASYNC;
 {
 
        bp->b_flags |= B_ASYNC;
@@ -222,37 +168,50 @@ register struct buf *bp;
 }
 
 /*
 }
 
 /*
- * release the buffer, with no I/O implied.
+ * Release the buffer, with no I/O implied.
  */
 brelse(bp)
  */
 brelse(bp)
-register struct buf *bp;
+       register struct buf *bp;
 {
 {
-       register struct buf **backp;
+       register struct buf *flist;
        register s;
 
        register s;
 
+       /*
+        * 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 (bp->b_flags&B_WANTED)
                wakeup((caddr_t)bp);
-       if (bfreelist.b_flags&B_WANTED) {
-               bfreelist.b_flags &= ~B_WANTED;
-               wakeup((caddr_t)&bfreelist);
-       }
-       if ((bp->b_flags&B_ERROR) && bp->b_dev != NODEV) {
-               bunhash(bp);
-               bp->b_dev = NODEV;  /* no assoc. on error */
+       if (bfreelist[0].b_flags&B_WANTED) {
+               bfreelist[0].b_flags &= ~B_WANTED;
+               wakeup((caddr_t)bfreelist);
        }
        }
+       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.
+        */
        s = spl6();
        s = spl6();
-       if(bp->b_flags & (B_AGE|B_ERROR)) {
-               backp = &bfreelist.av_forw;
-               (*backp)->av_back = bp;
-               bp->av_forw = *backp;
-               *backp = bp;
-               bp->av_back = &bfreelist;
+       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 {
        } else {
-               backp = &bfreelist.av_back;
-               (*backp)->av_forw = bp;
-               bp->av_back = *backp;
-               *backp = bp;
-               bp->av_forw = &bfreelist;
+               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);
        }
        bp->b_flags &= ~(B_WANTED|B_BUSY|B_ASYNC|B_AGE);
        splx(s);
        }
        bp->b_flags &= ~(B_WANTED|B_BUSY|B_ASYNC|B_AGE);
        splx(s);
@@ -263,27 +222,29 @@ register struct buf *bp;
  * (mainly to avoid getting hung up on a wait in breada)
  */
 incore(dev, blkno)
  * (mainly to avoid getting hung up on a wait in breada)
  */
 incore(dev, blkno)
-dev_t dev;
-daddr_t blkno;
+       dev_t dev;
+       daddr_t blkno;
 {
        register struct buf *bp;
 {
        register struct buf *bp;
-       register int dblkno = fsbtodb(blkno);
+       register struct buf *dp;
 
 
-       for (bp = &buf[bufhash[BUFHASH(blkno)]]; bp != &buf[-1];
-           bp = &buf[bp->b_hlink])
-               if (bp->b_blkno == dblkno && bp->b_dev == dev)
+       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 *
                        return (1);
        return (0);
 }
 
 struct buf *
-baddr(dev, blkno)
-dev_t dev;
-daddr_t blkno;
+baddr(dev, blkno, size)
+       dev_t dev;
+       daddr_t blkno;
+       int size;
 {
 
        if (incore(dev, blkno))
 {
 
        if (incore(dev, blkno))
-               return (bread(dev, blkno));
+               return (bread(dev, blkno, size));
        return (0);
 }
 
        return (0);
 }
 
@@ -291,193 +252,385 @@ daddr_t blkno;
  * 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.
  * 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.
  */
 struct buf *
  */
 struct buf *
-getblk(dev, blkno)
-dev_t dev;
-daddr_t blkno;
+getblk(dev, blkno, size)
+       dev_t dev;
+       daddr_t blkno;
+       int size;
 {
 {
-       register struct buf *bp, *dp, *ep;
-       register int i, x;
-       register int dblkno = fsbtodb(blkno);
-
-    loop:
-       (void) spl0();
-       for (bp = &buf[bufhash[BUFHASH(blkno)]]; bp != &buf[-1];
-           bp = &buf[bp->b_hlink]) {
-               if (bp->b_blkno != dblkno || bp->b_dev != dev)
+       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;
                        continue;
-               (void) spl6();
+               s = spl6();
                if (bp->b_flags&B_BUSY) {
                        bp->b_flags |= B_WANTED;
                        sleep((caddr_t)bp, PRIBIO+1);
                if (bp->b_flags&B_BUSY) {
                        bp->b_flags |= B_WANTED;
                        sleep((caddr_t)bp, PRIBIO+1);
+                       splx(s);
                        goto loop;
                }
                        goto loop;
                }
-               (void) spl0();
-#ifdef DISKMON
-               i = 0;
-               dp = bp->av_forw;
-               while (dp != &bfreelist) {
-                       i++;
-                       dp = dp->av_forw;
-               }
-               if (i<NBUF)
-                       io_info.bufcount[i]++;
-#endif
+               splx(s);
                notavail(bp);
                notavail(bp);
+               if (brealloc(bp, size) == 0)
+                       goto loop;
                bp->b_flags |= B_CACHE;
                return(bp);
        }
        if (major(dev) >= nblkdev)
                panic("blkdev");
                bp->b_flags |= B_CACHE;
                return(bp);
        }
        if (major(dev) >= nblkdev)
                panic("blkdev");
-       dp = bdevsw[major(dev)].d_tab;
-       if (dp == NULL)
-               panic("devtab");
-       (void) spl6();
-       if (bfreelist.av_forw == &bfreelist) {
-               bfreelist.b_flags |= B_WANTED;
-               sleep((caddr_t)&bfreelist, PRIBIO+1);
+       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;
                goto loop;
-       }
-       spl0();
-       bp = bfreelist.av_forw;
-       notavail(bp);
-       if (bp->b_flags & B_DELWRI) {
-               bp->b_flags |= B_ASYNC;
-               bwrite(bp);
+       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;
                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)
        if (bp->b_dev == NODEV)
-               goto done;
-       /* INLINE EXPANSION OF bunhash(bp) */
-       (void) spl6();
-       i = BUFHASH(dbtofsb(bp->b_blkno));
-       x = bp - buf;
-       if (bufhash[i] == x) {
-               bufhash[i] = bp->b_hlink;
-       } else {
-               for (ep = &buf[bufhash[i]]; ep != &buf[-1];
-                   ep = &buf[ep->b_hlink])
-                       if (ep->b_hlink == x) {
-                               ep->b_hlink = bp->b_hlink;
-                               goto done;
+               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;
+               }
+               splx(s);
+               notavail(ep);
+               if (ep->b_flags & B_DELWRI) {
+                       bwrite(ep);
+                       goto loop;
+               }
+               ep->b_flags |= B_INVAL;
+               brelse(ep);
+       }
+       return (allocbuf(bp, size));
+}
+
+/*
+ * Expand or contract the actual memory allocated to a buffer.
+ * If no memory is available, release buffer and take error exit
+ */
+allocbuf(tp, size)
+       register struct buf *tp;
+       int size;
+{
+       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;
+               }
+               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;
+                               }
                        }
                        }
-               panic("getblk");
+               } 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);
        }
        }
-done:
-       (void) spl0();
-       /* END INLINE EXPANSION */
-       bp->b_flags = B_BUSY;
-       bp->b_back->b_forw = bp->b_forw;
-       bp->b_forw->b_back = bp->b_back;
-       bp->b_forw = dp->b_forw;
-       bp->b_back = dp;
-       dp->b_forw->b_back = bp;
-       dp->b_forw = bp;
-       bp->b_dev = dev;
-       bp->b_blkno = dblkno;
-       i = BUFHASH(blkno);
-       bp->b_hlink = bufhash[i];
-       bufhash[i] = bp - buf;
-       return(bp);
+       tp->b_bufsize = sizealloc;
+#endif
+out:
+       tp->b_bcount = size;
+       return (1);
 }
 
 /*
 }
 
 /*
- * get an empty block,
- * not assigned to any particular device
+ * Release space associated with a buffer.
  */
  */
-struct buf *
-geteblk()
+bfree(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;
+       }
+#endif
+       bp->b_bcount = 0;
+}
+
+#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.
+ */
+bfreemem()
 {
        register struct buf *bp, *dp;
 {
        register struct buf *bp, *dp;
+       int s;
 
 loop:
 
 loop:
-       (void) spl6();
-       while (bfreelist.av_forw == &bfreelist) {
-               bfreelist.b_flags |= B_WANTED;
-               sleep((caddr_t)&bfreelist, PRIBIO+1);
+       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);
        }
        }
-       (void) spl0();
-       dp = &bfreelist;
-       bp = bfreelist.av_forw;
+       bp = dp->av_forw;
        notavail(bp);
        if (bp->b_flags & B_DELWRI) {
                bp->b_flags |= B_ASYNC;
                bwrite(bp);
                goto loop;
        }
        notavail(bp);
        if (bp->b_flags & B_DELWRI) {
                bp->b_flags |= B_ASYNC;
                bwrite(bp);
                goto loop;
        }
-       if (bp->b_dev != NODEV)
-               bunhash(bp);
-       bp->b_flags = B_BUSY;
-       bp->b_back->b_forw = bp->b_forw;
-       bp->b_forw->b_back = bp->b_back;
-       bp->b_forw = dp->b_forw;
-       bp->b_back = dp;
-       dp->b_forw->b_back = bp;
-       dp->b_forw = bp;
+       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_dev = (dev_t)NODEV;
-       bp->b_hlink = -1;
-       return(bp);
+       bp->b_error = 0;
+       brelse(bp);
+       return (1);
 }
 }
+#endif
 
 
-bunhash(bp)
-       register struct buf *bp;
+/*
+ * Find a buffer which is available for use.
+ * Select something from a free list.
+ * Preference is to AGE list, then LRU list.
+ */
+struct buf *
+getnewbuf()
 {
 {
-       register struct buf *ep;
-       register int i, x, s;
+       register struct buf *bp, *dp;
+       int s;
 
 
-       if (bp->b_dev == NODEV)
-               return;
+loop:
        s = spl6();
        s = spl6();
-       i = BUFHASH(dbtofsb(bp->b_blkno));
-       x = bp - buf;
-       if (bufhash[i] == x) {
-               bufhash[i] = bp->b_hlink;
-               goto ret;
+#ifndef sun
+       for (dp = &bfreelist[BQ_AGE]; dp > bfreelist; dp--)
+#else
+       for (dp = &bfreelist[BQ_EMPTY]; dp > bfreelist; dp--)
+#endif
+               if (dp->av_forw != dp)
+                       break;
+       if (dp == bfreelist) {          /* no free blocks */
+               dp->b_flags |= B_WANTED;
+               sleep((caddr_t)dp, PRIBIO+1);
+               splx(s);
+               goto loop;
        }
        }
-       for (ep = &buf[bufhash[i]]; ep != &buf[-1];
-           ep = &buf[ep->b_hlink])
-               if (ep->b_hlink == x) {
-                       ep->b_hlink = bp->b_hlink;
-                       goto ret;
-               }
-       panic("bunhash");
-ret:
        splx(s);
        splx(s);
+       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;
+       return (bp);
 }
 
 /*
  * Wait for I/O completion on the buffer; return errors
  * to the user.
  */
 }
 
 /*
  * Wait for I/O completion on the buffer; return errors
  * to the user.
  */
-iowait(bp)
-register struct buf *bp;
+biowait(bp)
+       register struct buf *bp;
 {
 {
+       int s;
 
 
-       (void) spl6();
+       s = spl6();
        while ((bp->b_flags&B_DONE)==0)
                sleep((caddr_t)bp, PRIBIO);
        while ((bp->b_flags&B_DONE)==0)
                sleep((caddr_t)bp, PRIBIO);
-       (void) spl0();
-       geterror(bp);
-}
-
-#ifndef FASTVAX
-/*
- * Unlink a buffer from the available list and mark it busy.
- * (internal interface)
- */
-notavail(bp)
-register struct buf *bp;
-{
-       register s;
-
-       s = spl6();
-       bp->av_back->av_forw = bp->av_forw;
-       bp->av_forw->av_back = bp->av_back;
-       bp->b_flags |= B_BUSY;
        splx(s);
        splx(s);
+       if (u.u_error == 0)                     /* XXX */
+               u.u_error = geterror(bp);
 }
 }
-#endif
 
 /*
  * Mark I/O complete on a buffer. If the header
 
 /*
  * Mark I/O complete on a buffer. If the header
@@ -487,28 +640,34 @@ register struct buf *bp;
  * release it if I/O is asynchronous, and wake 
  * up anyone waiting for it.
  */
  * release it if I/O is asynchronous, and wake 
  * up anyone waiting for it.
  */
-iodone(bp)
-register struct buf *bp;
+biodone(bp)
+       register struct buf *bp;
 {
        register int s;
 
        if (bp->b_flags & B_DONE)
 {
        register int s;
 
        if (bp->b_flags & B_DONE)
-               panic("dup iodone");
+               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->b_flags |= B_DONE;
        if (bp->b_flags & B_DIRTY) {
                if (bp->b_flags & B_ERROR)
                        panic("IO err in push");
                s = spl6();
-               cnt.v_pgout++;
                bp->av_forw = bclnlist;
                bp->b_bcount = swsize[bp - swbuf];
                bp->b_pfcent = swpf[bp - swbuf];
                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;
        }
                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);
+               return;
+       }
        if (bp->b_flags&B_ASYNC)
                brelse(bp);
        else {
        if (bp->b_flags&B_ASYNC)
                brelse(bp);
        else {
@@ -518,139 +677,44 @@ register struct buf *bp;
 }
 
 /*
 }
 
 /*
- * Zero the core associated with a buffer.
- */
-clrbuf(bp)
-struct buf *bp;
-{
-       register *p;
-       register c;
-
-       p = bp->b_un.b_words;
-       c = BSIZE/sizeof(int);
-       do
-               *p++ = 0;
-       while (--c);
-       bp->b_resid = 0;
-}
-
-/*
- * swap I/O -
- *
- * If the flag indicates a dirty page push initiated
- * by the pageout daemon, we map the page into the i th
- * virtual page of process 2 (the daemon itself) where i is
- * the index of the swap header that has been allocated.
- * We simply initialize the header and queue the I/O but
- * do not wait for completion. When the I/O completes,
- * iodone() will link the header to a list of cleaned
- * pages to be processed by the pageout daemon.
+ * Insure that no part of a specified block is in an incore buffer.
  */
  */
-swap(p, dblkno, addr, nbytes, rdflg, flag, dev, pfcent)
-       struct proc *p;
-       swblk_t dblkno;
-       caddr_t addr;
-       int flag, nbytes;
+blkflush(dev, blkno, size)
        dev_t dev;
        dev_t dev;
-       unsigned pfcent;
+       daddr_t blkno;
+       long size;
 {
 {
-       register struct buf *bp;
-       register int c;
-       int p2dp;
-       register struct pte *dpte, *vpte;
-
-       (void) spl6();
-       while (bswlist.av_forw == NULL) {
-               bswlist.b_flags |= B_WANTED;
-               sleep((caddr_t)&bswlist, PSWP+1);
-       }
-       bp = bswlist.av_forw;
-       bswlist.av_forw = bp->av_forw;
-       (void) spl0();
-
-       bp->b_flags = B_BUSY | B_PHYS | rdflg | flag;
-       if ((bp->b_flags & (B_DIRTY|B_PGIN)) == 0)
-               if (rdflg == B_READ)
-                       sum.v_pswpin += btoc(nbytes);
-               else
-                       sum.v_pswpout += btoc(nbytes);
-       bp->b_proc = p;
-       if (flag & B_DIRTY) {
-               p2dp = ((bp - swbuf) * CLSIZE) * KLMAX;
-               dpte = dptopte(&proc[2], p2dp);
-               vpte = vtopte(p, btop(addr));
-               for (c = 0; c < nbytes; c += NBPG) {
-                       if (vpte->pg_pfnum == 0 || vpte->pg_fod)
-                               panic("swap bad pte");
-                       *dpte++ = *vpte++;
-               }
-               bp->b_un.b_addr = (caddr_t)ctob(p2dp);
-       } else
-               bp->b_un.b_addr = addr;
-       while (nbytes > 0) {
-               c = imin(ctob(120), nbytes);
-               bp->b_bcount = c;
-               bp->b_blkno = dblkno;
-               bp->b_dev = dev;
-               if (flag & B_DIRTY) {
-                       swpf[bp - swbuf] = pfcent;
-                       swsize[bp - swbuf] = nbytes;
-               }
-               (*bdevsw[major(dev)].d_strategy)(bp);
-               if (flag & B_DIRTY) {
-                       if (c < nbytes)
-                               panic("big push");
-                       return;
+       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;
                }
                }
-               (void) spl6();
-               while((bp->b_flags&B_DONE)==0)
-                       sleep((caddr_t)bp, PSWP);
-               (void) spl0();
-               bp->b_un.b_addr += c;
-               bp->b_flags &= ~B_DONE;
-               if (bp->b_flags & B_ERROR) {
-                       if ((flag & (B_UAREA|B_PAGET)) || rdflg == B_WRITE)
-                               panic("hard IO err in swap");
-                       swkill(p, (char *)0);
+               if (ep->b_flags & B_DELWRI) {
+                       splx(s);
+                       notavail(ep);
+                       bwrite(ep);
+                       goto loop;
                }
                }
-               nbytes -= c;
-               dblkno += btoc(c);
-       }
-       (void) spl6();
-       bp->b_flags &= ~(B_BUSY|B_WANTED|B_PHYS|B_PAGET|B_UAREA|B_DIRTY);
-       bp->av_forw = bswlist.av_forw;
-       bswlist.av_forw = bp;
-       if (bswlist.b_flags & B_WANTED) {
-               bswlist.b_flags &= ~B_WANTED;
-               wakeup((caddr_t)&bswlist);
-               wakeup((caddr_t)&proc[2]);
+               splx(s);
        }
        }
-       (void) spl0();
-}
-
-/*
- * If rout == 0 then killed on swap error, else
- * rout is the name of the routine where we ran out of
- * swap space.
- */
-swkill(p, rout)
-       struct proc *p;
-       char *rout;
-{
-
-       printf("%d: ", p->p_pid);
-       if (rout)
-               printf("out of swap space in %s\n", rout);
-       else
-               printf("killed on swap error\n");
-       /*
-        * To be sure no looping (e.g. in vmsched trying to
-        * swap out) mark process locked in core (as though
-        * done by user) after killing it so noone will try
-        * to swap it out.
-        */
-       psignal(p, SIGKILL);
-       p->p_flag |= SULOCK;
 }
 
 /*
 }
 
 /*
@@ -658,94 +722,30 @@ swkill(p, rout)
  * on dev (or NODEV for all)
  * are flushed out.
  * (from umount and update)
  * on dev (or NODEV for all)
  * are flushed out.
  * (from umount and update)
+ * (and temporarily pagein)
  */
 bflush(dev)
  */
 bflush(dev)
-dev_t dev;
+       dev_t dev;
 {
        register struct buf *bp;
 {
        register struct buf *bp;
+       register struct buf *flist;
+       int s;
 
 loop:
 
 loop:
-       (void) spl6();
-       for (bp = bfreelist.av_forw; bp != &bfreelist; bp = bp->av_forw) {
-               if (bp->b_flags&B_DELWRI && (dev == NODEV||dev==bp->b_dev)) {
+       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);
                        bp->b_flags |= B_ASYNC;
                        notavail(bp);
                        bwrite(bp);
+                       splx(s);
                        goto loop;
                }
        }
                        goto loop;
                }
        }
-       (void) spl0();
-}
-
-/*
- * Raw I/O. The arguments are
- *     The strategy routine for the device
- *     A buffer, which will always be a special buffer
- *       header owned exclusively by the device for this purpose
- *     The device number
- *     Read/write flag
- * Essentially all the work is computing physical addresses and
- * validating them.
- * If the user has the proper access privilidges, the process is
- * marked 'delayed unlock' and the pages involved in the I/O are
- * faulted and locked. After the completion of the I/O, the above pages
- * are unlocked.
- */
-physio(strat, bp, dev, rw, mincnt)
-int (*strat)(); 
-register struct buf *bp;
-unsigned (*mincnt)();
-{
-       register int c;
-       char *a;
-
-       if (useracc(u.u_base,u.u_count,rw==B_READ?B_WRITE:B_READ) == NULL) {
-               u.u_error = EFAULT;
-               return;
-       }
-       (void) spl6();
-       while (bp->b_flags&B_BUSY) {
-               bp->b_flags |= B_WANTED;
-               sleep((caddr_t)bp, PRIBIO+1);
-       }
-       bp->b_error = 0;
-       bp->b_proc = u.u_procp;
-       bp->b_un.b_addr = u.u_base;
-       while (u.u_count != 0 && bp->b_error==0) {
-               bp->b_flags = B_BUSY | B_PHYS | rw;
-               bp->b_dev = dev;
-               bp->b_blkno = u.u_offset >> PGSHIFT;
-               bp->b_bcount = u.u_count;
-               (*mincnt)(bp);
-               c = bp->b_bcount;
-               u.u_procp->p_flag |= SPHYSIO;
-               vslock(a = bp->b_un.b_addr, c);
-               (*strat)(bp);
-               (void) spl6();
-               while ((bp->b_flags&B_DONE) == 0)
-                       sleep((caddr_t)bp, PRIBIO);
-               vsunlock(a, c, rw);
-               u.u_procp->p_flag &= ~SPHYSIO;
-               if (bp->b_flags&B_WANTED)
-                       wakeup((caddr_t)bp);
-               (void) spl0();
-               bp->b_un.b_addr += c;
-               u.u_count -= c;
-               u.u_offset += c;
-       }
-       bp->b_flags &= ~(B_BUSY|B_WANTED|B_PHYS);
-       u.u_count = bp->b_resid;
-       geterror(bp);
-}
-
-/*ARGSUSED*/
-unsigned
-minphys(bp)
-struct buf *bp;
-{
-
-       if (bp->b_bcount > 60 * 1024)
-               bp->b_bcount = 60 * 1024;
+       splx(s);
 }
 
 /*
 }
 
 /*
@@ -755,10 +755,36 @@ struct buf *bp;
  * don't yet return specific errors.
  */
 geterror(bp)
  * don't yet return specific errors.
  */
 geterror(bp)
-register struct buf *bp;
+       register struct buf *bp;
 {
 {
+       int error = 0;
 
        if (bp->b_flags&B_ERROR)
 
        if (bp->b_flags&B_ERROR)
-               if ((u.u_error = bp->b_error)==0)
-                       u.u_error = EIO;
+               if ((error = bp->b_error)==0)
+                       return (EIO);
+       return (error);
+}
+
+/*
+ * 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;
 }
 }