windows-buf-scache-interlock-20080222
[openafs.git] / src / WINNT / afsd / cm_buf.c
index 9b28eb4..ccc5098 100644 (file)
 #include <afs/param.h>
 #include <afs/stds.h>
 
-#ifndef DJGPP
 #include <windows.h>
-#endif
 #include <osi.h>
 #include <stdio.h>
 #include <assert.h>
 #include <strsafe.h>
+#include <math.h>
 
 #include "afsd.h"
 #include "cm_memmap.h"
@@ -88,64 +87,156 @@ extern int cm_diskCacheEnabled;
 /* set this to 1 when we are terminating to prevent access attempts */
 static int buf_ShutdownFlag = 0;
 
+void buf_HoldLocked(cm_buf_t *bp)
+{
+    osi_assertx(bp->magic == CM_BUF_MAGIC,"incorrect cm_buf_t magic");
+    InterlockedIncrement(&bp->refCount);
+}
+
 /* hold a reference to an already held buffer */
 void buf_Hold(cm_buf_t *bp)
 {
-    osi_assert(bp->magic == CM_BUF_MAGIC);
-    lock_ObtainWrite(&buf_globalLock);
-    bp->refCount++;
-    lock_ReleaseWrite(&buf_globalLock);
+    lock_ObtainRead(&buf_globalLock);
+    buf_HoldLocked(bp);
+    lock_ReleaseRead(&buf_globalLock);
 }
 
-/* incremental sync daemon.  Writes 1/10th of all the buffers every 5000 ms */
+/* code to drop reference count while holding buf_globalLock */
+void buf_ReleaseLocked(cm_buf_t *bp, afs_uint32 writeLocked)
+{
+    afs_int32 refCount;
+
+    if (writeLocked)
+        lock_AssertWrite(&buf_globalLock);
+    else
+        lock_AssertRead(&buf_globalLock);
+
+    /* ensure that we're in the LRU queue if our ref count is 0 */
+    osi_assertx(bp->magic == CM_BUF_MAGIC,"incorrect cm_buf_t magic");
+
+    refCount = InterlockedDecrement(&bp->refCount);
+#ifdef DEBUG
+    if (refCount < 0)
+       osi_panic("buf refcount 0",__FILE__,__LINE__);;
+#else
+    osi_assertx(refCount >= 0, "cm_buf_t refCount == 0");
+#endif
+    if (refCount == 0) {
+        /* 
+         * If we are read locked there could be a race condition
+         * with buf_Find() so we must obtain a write lock and
+         * double check that the refCount is actually zero
+         * before we remove the buffer from the LRU queue.
+         */
+        if (!writeLocked) {
+            lock_ReleaseRead(&buf_globalLock);
+            lock_ObtainWrite(&buf_globalLock);
+        }
+
+        if (bp->refCount == 0 &&
+            !(bp->flags & CM_BUF_INLRU)) {
+            osi_QAdd((osi_queue_t **) &cm_data.buf_freeListp, &bp->q);
+
+            /* watch for transition from empty to one element */
+            if (!cm_data.buf_freeListEndp)
+                cm_data.buf_freeListEndp = cm_data.buf_freeListp;
+            bp->flags |= CM_BUF_INLRU;
+        }
+
+        if (!writeLocked) {
+            lock_ReleaseWrite(&buf_globalLock);
+            lock_ObtainRead(&buf_globalLock);
+        }
+    }
+}       
+
+/* release a buffer.  Buffer must be referenced, but unlocked. */
+void buf_Release(cm_buf_t *bp)
+{
+    afs_int32 refCount;
+
+    /* ensure that we're in the LRU queue if our ref count is 0 */
+    osi_assertx(bp->magic == CM_BUF_MAGIC,"incorrect cm_buf_t magic");
+
+    refCount = InterlockedDecrement(&bp->refCount);
+#ifdef DEBUG
+    if (refCount < 0)
+       osi_panic("buf refcount 0",__FILE__,__LINE__);;
+#else
+    osi_assertx(refCount >= 0, "cm_buf_t refCount == 0");
+#endif
+    if (refCount == 0) {
+        lock_ObtainWrite(&buf_globalLock);
+        if (bp->refCount == 0 && 
+            !(bp->flags & CM_BUF_INLRU)) {
+            osi_QAdd((osi_queue_t **) &cm_data.buf_freeListp, &bp->q);
+
+            /* watch for transition from empty to one element */
+            if (!cm_data.buf_freeListEndp)
+                cm_data.buf_freeListEndp = cm_data.buf_freeListp;
+            bp->flags |= CM_BUF_INLRU;
+        }
+        lock_ReleaseWrite(&buf_globalLock);
+    }
+}
+
+/* incremental sync daemon.  Writes all dirty buffers every 5000 ms */
 void buf_IncrSyncer(long parm)
 {
-    cm_buf_t *bp;                      /* buffer we're hacking on; held */
+    cm_buf_t **bpp, *bp;
     long i;                            /* counter */
-    long nAtOnce;                      /* how many to do at once */
+    long wasDirty = 0;
     cm_req_t req;
 
-    lock_ObtainWrite(&buf_globalLock);
-    bp = cm_data.buf_allp;
-    bp->refCount++;
-    lock_ReleaseWrite(&buf_globalLock);
-    nAtOnce = cm_data.buf_nbuffers / 10;
     while (buf_ShutdownFlag == 0) {
-#ifndef DJGPP
-        i = SleepEx(5000, 1);
-        if (i != 0) continue;
-#else
-        thrd_Sleep(5000);
-#endif /* DJGPP */
-            
-        if (buf_ShutdownFlag == 1)
-            return;
+        if (!wasDirty) {
+            i = SleepEx(5000, 1);
+            if (i != 0) continue;
+       }
 
-        /* now go through our percentage of the buffers */
-        for (i=0; i<nAtOnce; i++) {
-            /* don't want its identity changing while we're
-             * messing with it, so must do all of this with
-             * bp held.
-             */
-
-            /* start cleaning the buffer; don't touch log pages since
-             * the log code counts on knowing exactly who is writing
-             * a log page at any given instant.
-             */
-            cm_InitReq(&req);
-            req.flags |= CM_REQ_NORETRY;
-            buf_CleanAsync(bp, &req);
+       wasDirty = 0;
 
-            /* now advance to the next buffer; the allp chain never changes,
-             * and so can be followed even when holding no locks.
-             */
-            lock_ObtainWrite(&buf_globalLock);
-            buf_LockedRelease(bp);
-            bp = bp->allp;
-            if (!bp) 
-                bp = cm_data.buf_allp;
-            bp->refCount++;
-            lock_ReleaseWrite(&buf_globalLock);
+        /* now go through our percentage of the buffers */
+        for (bpp = &cm_data.buf_dirtyListp; bp = *bpp; ) {
+
+           /* all dirty buffers are held when they are added to the
+            * dirty list.  No need for an additional hold.
+            */
+
+           if (bp->flags & CM_BUF_DIRTY) {
+               /* start cleaning the buffer; don't touch log pages since
+                * the log code counts on knowing exactly who is writing
+                * a log page at any given instant.
+                */
+               cm_InitReq(&req);
+               req.flags |= CM_REQ_NORETRY;
+               wasDirty |= buf_CleanAsync(bp, &req);
+           }
+
+           /* the buffer may or may not have been dirty
+            * and if dirty may or may not have been cleaned
+            * successfully.  check the dirty flag again.  
+            */
+           if (!(bp->flags & CM_BUF_DIRTY)) {
+               lock_ObtainMutex(&bp->mx);
+               if (!(bp->flags & CM_BUF_DIRTY)) {
+                   /* remove the buffer from the dirty list */
+                   lock_ObtainWrite(&buf_globalLock);
+                   *bpp = bp->dirtyp;
+                   bp->dirtyp = NULL;
+                   if (cm_data.buf_dirtyListp == NULL)
+                       cm_data.buf_dirtyListEndp = NULL;
+                   buf_ReleaseLocked(bp, TRUE);
+                   lock_ReleaseWrite(&buf_globalLock);
+               } else {
+                   /* advance the pointer so we don't loop forever */
+                   bpp = &bp->dirtyp;
+               }
+               lock_ReleaseMutex(&bp->mx);
+           } else {
+               /* advance the pointer so we don't loop forever */
+               bpp = &bp->dirtyp;
+           }
         }      /* for loop over a bunch of buffers */
     }          /* whole daemon's while loop */
 }
@@ -154,7 +245,14 @@ long
 buf_ValidateBuffers(void)
 {
     cm_buf_t * bp, *bpf, *bpa, *bpb;
-    afs_uint32 countb = 0, countf = 0, counta = 0;
+    afs_uint64 countb = 0, countf = 0, counta = 0;
+
+    if (cm_data.buf_freeListp == NULL && cm_data.buf_freeListEndp != NULL ||
+         cm_data.buf_freeListp != NULL && cm_data.buf_freeListEndp == NULL) {
+        afsi_log("cm_ValidateBuffers failure: inconsistent free list pointers");
+        fprintf(stderr, "cm_ValidateBuffers failure: inconsistent free list pointers\n");
+        return -9;                  
+    }
 
     for (bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) { 
         if (bp->magic != CM_BUF_MAGIC) {
@@ -227,7 +325,7 @@ void buf_Shutdown(void)
 /* initialize the buffer package; called with no locks
  * held during the initialization phase.
  */
-long buf_Init(int newFile, cm_buf_ops_t *opsp, long nbuffers)
+long buf_Init(int newFile, cm_buf_ops_t *opsp, afs_uint64 nbuffers)
 {
     static osi_once_t once;
     cm_buf_t *bp;
@@ -257,10 +355,10 @@ long buf_Init(int newFile, cm_buf_ops_t *opsp, long nbuffers)
             cm_data.buf_nOrigBuffers = cm_data.buf_nbuffers;
  
             /* lower hash size to a prime number */
-            cm_data.buf_hashSize = osi_PrimeLessThan(CM_BUF_HASHSIZE);
+           cm_data.buf_hashSize = osi_PrimeLessThan((afs_uint32)(cm_data.buf_nbuffers/7 + 1));
  
             /* create hash table */
-            memset((void *)cm_data.buf_hashTablepp, 0, cm_data.buf_hashSize * sizeof(cm_buf_t *));
+            memset((void *)cm_data.buf_scacheHashTablepp, 0, cm_data.buf_hashSize * sizeof(cm_buf_t *));
             
             /* another hash table */
             memset((void *)cm_data.buf_fileHashTablepp, 0, cm_data.buf_hashSize * sizeof(cm_buf_t *));
@@ -271,8 +369,10 @@ long buf_Init(int newFile, cm_buf_ops_t *opsp, long nbuffers)
             cm_data.buf_allp = NULL;
             
             for (i=0; i<cm_data.buf_nbuffers; i++) {
-                osi_assert(bp >= cm_data.bufHeaderBaseAddress && bp < (cm_buf_t *)cm_data.bufDataBaseAddress);
-                osi_assert(data >= cm_data.bufDataBaseAddress && data < cm_data.bufEndOfData);
+                osi_assertx(bp >= cm_data.bufHeaderBaseAddress && bp < (cm_buf_t *)cm_data.bufDataBaseAddress, 
+                            "invalid cm_buf_t address");
+                osi_assertx(data >= cm_data.bufDataBaseAddress && data < cm_data.bufEndOfData,
+                            "invalid cm_buf_t data address");
                 
                 /* allocate and zero some storage */
                 memset(bp, 0, sizeof(cm_buf_t));
@@ -334,9 +434,7 @@ long buf_Init(int newFile, cm_buf_ops_t *opsp, long nbuffers)
                                "buf_IncrSyncer");
 
         osi_assertx(phandle != NULL, "buf: can't create incremental sync proc");
-#ifndef DJGPP
         CloseHandle(phandle);
-#endif /* !DJGPP */
     }
 
 #ifdef TESTING
@@ -348,9 +446,8 @@ long buf_Init(int newFile, cm_buf_ops_t *opsp, long nbuffers)
 /* add nbuffers to the buffer pool, if possible.
  * Called with no locks held.
  */
-long buf_AddBuffers(long nbuffers)
+long buf_AddBuffers(afs_uint64 nbuffers)
 {
-#ifndef DJGPP
     /* The size of a virtual cache cannot be changed after it has
      * been created.  Subsequent calls to MapViewofFile() with
      * an existing mapping object name would not allow the 
@@ -364,49 +461,12 @@ long buf_AddBuffers(long nbuffers)
               nbuffers, cm_data.buf_nbuffers);
 
     return CM_ERROR_INVAL;
-#else
-    cm_buf_t *bp;
-    int i;
-    char *data;
-
-    data = malloc(buf_nbuffers * cm_data.buf_blockSize);
-
-    /* Create buffer headers and put in free list */
-    bp = malloc(nbuffers * sizeof(*bp));
-
-    for (i=0; i<nbuffers; i++) {
-        memset(bp, 0, sizeof(*bp));
-        
-        lock_InitializeMutex(&bp->mx, "cm_buf_t");
-
-        /* grab appropriate number of bytes from aligned zone */
-        bp->datap = data;
-
-        bp->flags |= CM_BUF_INLRU;
-
-        lock_ObtainWrite(&buf_globalLock);
-        /* note that buf_allp chain is covered by buf_globalLock now */
-        bp->allp = cm_data.buf_allp;
-        cm_data.buf_allp = bp;
-        osi_QAdd((osi_queue_t **) &cm_data.buf_freeListp, &bp->q);
-        if (!cm_data.buf_freeListEndp) 
-            cm_data.buf_freeListEndp = bp;
-        cm_data.buf_nbuffers++;
-        lock_ReleaseWrite(&buf_globalLock);
-
-        bp++;
-        data += cm_data.buf_blockSize;
-       
-    }   /* for loop over all buffers */
-
-    return 0;
-#endif /* DJGPP */
 }       
 
 /* interface to set the number of buffers to an exact figure.
  * Called with no locks held.
  */
-long buf_SetNBuffers(long nbuffers)
+long buf_SetNBuffers(afs_uint64 nbuffers)
 {
     if (nbuffers < 10) 
         return CM_ERROR_INVAL;
@@ -418,22 +478,16 @@ long buf_SetNBuffers(long nbuffers)
         return CM_ERROR_INVAL;
 }
 
-/* release a buffer.  Buffer must be referenced, but unlocked. */
-void buf_Release(cm_buf_t *bp)
-{
-    lock_ObtainWrite(&buf_globalLock);
-    buf_LockedRelease(bp);
-    lock_ReleaseWrite(&buf_globalLock);
-}
-
 /* wait for reading or writing to clear; called with write-locked
- * buffer, and returns with locked buffer.
+ * buffer and unlocked scp and returns with locked buffer.
  */
 void buf_WaitIO(cm_scache_t * scp, cm_buf_t *bp)
 {
+    int release = 0;
+
     if (scp)
-        osi_assert(scp->magic == CM_SCACHE_MAGIC);
-    osi_assert(bp->magic == CM_BUF_MAGIC);
+        osi_assertx(scp->magic == CM_SCACHE_MAGIC, "invalid cm_scache_t magic");
+    osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
 
     while (1) {
         /* if no IO is happening, we're done */
@@ -447,29 +501,36 @@ void buf_WaitIO(cm_scache_t * scp, cm_buf_t *bp)
         if ( bp->flags & CM_BUF_WAITING ) {
             bp->waitCount++;
             bp->waitRequests++;
-            osi_Log1(buf_logp, "buf_WaitIO CM_BUF_WAITING already set for 0x%x", bp);
+            osi_Log1(buf_logp, "buf_WaitIO CM_BUF_WAITING already set for 0x%p", bp);
         } else {
-            osi_Log1(buf_logp, "buf_WaitIO CM_BUF_WAITING set for 0x%x", bp);
+            osi_Log1(buf_logp, "buf_WaitIO CM_BUF_WAITING set for 0x%p", bp);
             bp->flags |= CM_BUF_WAITING;
             bp->waitCount = bp->waitRequests = 1;
         }
-        osi_SleepM((long) bp, &bp->mx);
+        osi_SleepM((LONG_PTR)bp, &bp->mx);
+
+       smb_UpdateServerPriority();
+
         lock_ObtainMutex(&bp->mx);
-        osi_Log1(buf_logp, "buf_WaitIO conflict wait done for 0x%x", bp);
+        osi_Log1(buf_logp, "buf_WaitIO conflict wait done for 0x%p", bp);
         bp->waitCount--;
         if (bp->waitCount == 0) {
-            osi_Log1(afsd_logp, "buf_WaitIO CM_BUF_WAITING reset for 0x%x", bp);
+            osi_Log1(buf_logp, "buf_WaitIO CM_BUF_WAITING reset for 0x%p", bp);
             bp->flags &= ~CM_BUF_WAITING;
             bp->waitRequests = 0;
         }
 
+        if ( !scp ) {
+            if (scp = cm_FindSCache(&bp->fid))
+                release = 1;
+        }
         if ( scp ) {
             lock_ObtainMutex(&scp->mx);
             if (scp->flags & CM_SCACHEFLAG_WAITING) {
-                osi_Log1(buf_logp, "buf_WaitIO waking scp 0x%x", scp);
-                osi_Wakeup(&scp->flags);
-                lock_ReleaseMutex(&scp->mx);
+                osi_Log1(buf_logp, "buf_WaitIO waking scp 0x%p", scp);
+                osi_Wakeup((LONG_PTR)&scp->flags);
             }
+           lock_ReleaseMutex(&scp->mx);
         }
     }
         
@@ -477,43 +538,29 @@ void buf_WaitIO(cm_scache_t * scp, cm_buf_t *bp)
      * the I/O to complete.  Do so.
      */
     if (bp->flags & CM_BUF_WAITING) {
-        osi_Log1(buf_logp, "buf_WaitIO Waking bp 0x%x", bp);
-        osi_Wakeup((long) bp);
+        osi_Log1(buf_logp, "buf_WaitIO Waking bp 0x%p", bp);
+        osi_Wakeup((LONG_PTR) bp);
     }
-    osi_Log1(buf_logp, "WaitIO finished wait for bp 0x%x", (long) bp);
-}
-
-/* code to drop reference count while holding buf_globalLock */
-void buf_LockedRelease(cm_buf_t *bp)
-{
-    /* ensure that we're in the LRU queue if our ref count is 0 */
-    osi_assert(bp->refCount > 0);
-    if (--bp->refCount == 0) {
-        if (!(bp->flags & CM_BUF_INLRU)) {
-            osi_QAdd((osi_queue_t **) &cm_data.buf_freeListp, &bp->q);
+    osi_Log1(buf_logp, "WaitIO finished wait for bp 0x%p", bp);
 
-            /* watch for transition from empty to one element */
-            if (!cm_data.buf_freeListEndp)
-                cm_data.buf_freeListEndp = cm_data.buf_freeListp;
-            bp->flags |= CM_BUF_INLRU;
-        }
-    }
-}       
+    if (scp && release)
+       cm_ReleaseSCache(scp);
+}
 
 /* find a buffer, if any, for a particular file ID and offset.  Assumes
  * that buf_globalLock is write locked when called.
  */
-cm_buf_t *buf_LockedFind(struct cm_scache *scp, osi_hyper_t *offsetp)
+cm_buf_t *buf_FindLocked(struct cm_scache *scp, osi_hyper_t *offsetp)
 {
-    long i;
+    afs_uint32 i;
     cm_buf_t *bp;
 
     i = BUF_HASH(&scp->fid, offsetp);
-    for(bp = cm_data.buf_hashTablepp[i]; bp; bp=bp->hashp) {
+    for(bp = cm_data.buf_scacheHashTablepp[i]; bp; bp=bp->hashp) {
         if (cm_FidCmp(&scp->fid, &bp->fid) == 0
              && offsetp->LowPart == bp->offset.LowPart
              && offsetp->HighPart == bp->offset.HighPart) {
-            bp->refCount++;
+            buf_HoldLocked(bp);
             break;
         }
     }
@@ -529,9 +576,9 @@ cm_buf_t *buf_Find(struct cm_scache *scp, osi_hyper_t *offsetp)
 {
     cm_buf_t *bp;
 
-    lock_ObtainWrite(&buf_globalLock);
-    bp = buf_LockedFind(scp, offsetp);
-    lock_ReleaseWrite(&buf_globalLock);
+    lock_ObtainRead(&buf_globalLock);
+    bp = buf_FindLocked(scp, offsetp);
+    lock_ReleaseRead(&buf_globalLock);
 
     return bp;
 }       
@@ -542,42 +589,83 @@ cm_buf_t *buf_Find(struct cm_scache *scp, osi_hyper_t *offsetp)
  * Makes sure that there's only one person writing this block
  * at any given time, and also ensures that the log is forced sufficiently far,
  * if this buffer contains logged data.
+ *
+ * Returns non-zero if the buffer was dirty.
  */
-void buf_LockedCleanAsync(cm_buf_t *bp, cm_req_t *reqp)
+long buf_CleanAsyncLocked(cm_buf_t *bp, cm_req_t *reqp)
 {
     long code = 0;
+    long isdirty = 0;
+    cm_scache_t * scp = NULL;
+    osi_hyper_t offset;
 
-    osi_assert(bp->magic == CM_BUF_MAGIC);
+    osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
 
     while ((bp->flags & CM_BUF_DIRTY) == CM_BUF_DIRTY) {
+       isdirty = 1;
         lock_ReleaseMutex(&bp->mx);
 
-        code = (*cm_buf_opsp->Writep)(&bp->fid, &bp->offset,
-                                       cm_data.buf_blockSize, 0, bp->userp,
-                                       reqp);
-                
-        lock_ObtainMutex(&bp->mx);
-        if (code) 
-            break;
+       scp = cm_FindSCache(&bp->fid);
+       if (scp) {
+           osi_Log2(buf_logp, "buf_CleanAsyncLocked starts I/O on scp 0x%p buf 0x%p", scp, bp);
+
+            offset = bp->offset;
+            LargeIntegerAdd(offset, ConvertLongToLargeInteger(bp->dirty_offset));
+           code = (*cm_buf_opsp->Writep)(scp, &offset, bp->dirty_length, 0, bp->userp, reqp);
+           osi_Log3(buf_logp, "buf_CleanAsyncLocked I/O on scp 0x%p buf 0x%p, done=%d", scp, bp, code);
+
+           cm_ReleaseSCache(scp);
+           scp = NULL;
+       } else {
+           osi_Log1(buf_logp, "buf_CleanAsyncLocked unable to start I/O - scp not found buf 0x%p", bp);
+           code = CM_ERROR_NOSUCHFILE;
+       }    
+        
+       lock_ObtainMutex(&bp->mx);
+       /* if the Write routine returns No Such File, clear the dirty flag
+        * because we aren't going to be able to write this data to the file
+        * server.
+        */
+       if (code == CM_ERROR_NOSUCHFILE || code == CM_ERROR_BADFD){
+           bp->flags &= ~CM_BUF_DIRTY;
+           bp->flags |= CM_BUF_ERROR;
+            bp->dirty_offset = 0;
+            bp->dirty_length = 0;
+           bp->error = code;
+           bp->dataVersion = -1; /* bad */
+           bp->dirtyCounter++;
+       }
 
 #ifdef DISKCACHE95
         /* Disk cache support */
         /* write buffer to disk cache (synchronous for now) */
         diskcache_Update(bp->dcp, bp->datap, cm_data.buf_blockSize, bp->dataVersion);
 #endif /* DISKCACHE95 */
+
+       /* if we get here and retries are not permitted 
+        * then we need to exit this loop regardless of 
+        * whether or not we were able to clear the dirty bit
+        */
+       if (reqp->flags & CM_REQ_NORETRY)
+           break;
     };
 
+    if (!(bp->flags & CM_BUF_DIRTY)) {
+       /* remove buffer from dirty buffer queue */
+
+    }
+
     /* do logging after call to GetLastError, or else */
-    osi_Log2(buf_logp, "buf_CleanAsync starts I/O on 0x%x, done=%d", bp, code);
         
     /* if someone was waiting for the I/O that just completed or failed,
      * wake them up.
      */
     if (bp->flags & CM_BUF_WAITING) {
         /* turn off flags and wakeup users */
-        osi_Log1(buf_logp, "buf_WaitIO Waking bp 0x%x", bp);
-        osi_Wakeup((long) bp);
+        osi_Log1(buf_logp, "buf_WaitIO Waking bp 0x%p", bp);
+        osi_Wakeup((LONG_PTR) bp);
     }
+    return isdirty;
 }
 
 /* Called with a zero-ref count buffer and with the buf_globalLock write locked.
@@ -586,12 +674,12 @@ void buf_LockedCleanAsync(cm_buf_t *bp, cm_req_t *reqp)
  */
 void buf_Recycle(cm_buf_t *bp)
 {
-    int i;
+    afs_uint32 i;
     cm_buf_t **lbpp;
     cm_buf_t *tbp;
     cm_buf_t *prevBp, *nextBp;
 
-    osi_assert(bp->magic == CM_BUF_MAGIC);
+    osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
 
     /* if we get here, we know that the buffer still has a 0 ref count,
      * and that it is clean and has no currently pending I/O.  This is
@@ -600,32 +688,37 @@ void buf_Recycle(cm_buf_t *bp)
      * have any lock conflicts, so we can grab the buffer lock out of
      * order in the locking hierarchy.
      */
-    osi_Log2( buf_logp, "buf_Recycle recycles 0x%x, off 0x%x",
-              bp, bp->offset.LowPart);
+    osi_Log3( buf_logp, "buf_Recycle recycles 0x%p, off 0x%x:%08x",
+              bp, bp->offset.HighPart, bp->offset.LowPart);
 
-    osi_assert(bp->refCount == 0);
-    osi_assert(!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING | CM_BUF_DIRTY)));
+    osi_assertx(bp->refCount == 0, "cm_buf_t refcount != 0");
+    osi_assertx(!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING | CM_BUF_DIRTY)),
+                "incorrect cm_buf_t flags");
     lock_AssertWrite(&buf_globalLock);
 
     if (bp->flags & CM_BUF_INHASH) {
         /* Remove from hash */
 
         i = BUF_HASH(&bp->fid, &bp->offset);
-        lbpp = &(cm_data.buf_hashTablepp[i]);
+        lbpp = &(cm_data.buf_scacheHashTablepp[i]);
         for(tbp = *lbpp; tbp; lbpp = &tbp->hashp, tbp = *lbpp) {
-            if (tbp == bp) break;
+            if (tbp == bp) 
+                break;
         }
 
         /* we better find it */
         osi_assertx(tbp != NULL, "buf_Recycle: hash table screwup");
 
         *lbpp = bp->hashp;     /* hash out */
+        bp->hashp = NULL;
 
         /* Remove from file hash */
 
         i = BUF_FILEHASH(&bp->fid);
         prevBp = bp->fileHashBackp;
+        bp->fileHashBackp = NULL;
         nextBp = bp->fileHashp;
+        bp->fileHashp = NULL;
         if (prevBp)
             prevBp->fileHashp = nextBp;
         else
@@ -659,9 +752,9 @@ void buf_Recycle(cm_buf_t *bp)
  */
 long buf_GetNewLocked(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bufpp)
 {
-    cm_buf_t *bp;              /* buffer we're dealing with */
+    cm_buf_t *bp;      /* buffer we're dealing with */
     cm_buf_t *nextBp;  /* next buffer in file hash chain */
-    long i;                    /* temp */
+    afs_uint32 i;      /* temp */
     cm_req_t req;
 
     cm_InitReq(&req);  /* just in case */
@@ -672,16 +765,31 @@ long buf_GetNewLocked(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bu
 
     while(1) {
       retry:
+        lock_ObtainRead(&scp->bufCreateLock);
         lock_ObtainWrite(&buf_globalLock);
         /* check to see if we lost the race */
         if (scp) {
-            if (bp = buf_LockedFind(scp, offsetp)) {
+            if (bp = buf_FindLocked(scp, offsetp)) {
+               /* Do not call buf_ReleaseLocked() because we 
+                * do not want to allow the buffer to be added
+                * to the free list.
+                */
                 bp->refCount--;
                 lock_ReleaseWrite(&buf_globalLock);
                 return CM_BUF_EXISTS;
             }
         }
 
+       /* does this fix the problem below?  it's a simple solution. */
+       if (!cm_data.buf_freeListEndp)
+       {
+           lock_ReleaseWrite(&buf_globalLock);
+            lock_ReleaseRead(&scp->bufCreateLock);
+           osi_Log0(afsd_logp, "buf_GetNewLocked: Free Buffer List is empty - sleeping 200ms");
+           Sleep(200);
+           goto retry;
+       }
+
         /* for debugging, assert free list isn't empty, although we
          * really should try waiting for a running tranasction to finish
          * instead of this; or better, we should have a transaction
@@ -731,7 +839,7 @@ long buf_GetNewLocked(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bu
                  * just the lock required to minimize contention
                  * on the big lock.
                  */
-                bp->refCount++;
+                buf_HoldLocked(bp);
                 lock_ReleaseWrite(&buf_globalLock);
 
                 /* grab required lock and clean; this only
@@ -766,10 +874,13 @@ long buf_GetNewLocked(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bu
             if (scp) {
                 bp->flags |= CM_BUF_INHASH;
                 bp->fid = scp->fid;
+#ifdef DEBUG
+               bp->scp = scp;
+#endif
                 bp->offset = *offsetp;
                 i = BUF_HASH(&scp->fid, offsetp);
-                bp->hashp = cm_data.buf_hashTablepp[i];
-                cm_data.buf_hashTablepp[i] = bp;
+                bp->hashp = cm_data.buf_scacheHashTablepp[i];
+                cm_data.buf_scacheHashTablepp[i] = bp;
                 i = BUF_FILEHASH(&scp->fid);
                 nextBp = cm_data.buf_fileHashTablepp[i];
                 bp->fileHashp = nextBp;
@@ -779,17 +890,12 @@ long buf_GetNewLocked(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bu
                 cm_data.buf_fileHashTablepp[i] = bp;
             }
 
-            /* prepare to return it.  Start by giving it a good
-             * refcount */
-            bp->refCount = 1;
-                        
-            /* and since it has a non-zero ref count, we should move
-             * it from the lru queue.  It better be still there,
-             * since we've held the global (big) lock since we found
-             * it there.
+            /* we should move it from the lru queue.  It better still be there,
+             * since we've held the global (big) lock since we found it there.
              */
             osi_assertx(bp->flags & CM_BUF_INLRU,
                          "buf_GetNewLocked: LRU screwup");
+
             if (cm_data.buf_freeListEndp == bp) {
                 /* we're the last guy in this queue, so maintain it */
                 cm_data.buf_freeListEndp = (cm_buf_t *) osi_QPrev(&bp->q);
@@ -797,14 +903,21 @@ long buf_GetNewLocked(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bu
             osi_QRemove((osi_queue_t **) &cm_data.buf_freeListp, &bp->q);
             bp->flags &= ~CM_BUF_INLRU;
 
-            /* finally, grab the mutex so that people don't use it
+            /* grab the mutex so that people don't use it
              * before the caller fills it with data.  Again, no one    
              * should have been able to get to this dude to lock it.
              */
-            osi_assertx(lock_TryMutex(&bp->mx),
-                         "buf_GetNewLocked: TryMutex failed");
+           if (!lock_TryMutex(&bp->mx)) {
+               osi_Log2(afsd_logp, "buf_GetNewLocked bp 0x%p cannot be mutex locked.  refCount %d should be 0",
+                        bp, bp->refCount);
+               osi_panic("buf_GetNewLocked: TryMutex failed",__FILE__,__LINE__);
+           }
 
+           /* prepare to return it.  Give it a refcount */
+            bp->refCount = 1;
+                        
             lock_ReleaseWrite(&buf_globalLock);
+            lock_ReleaseRead(&scp->bufCreateLock);
             *bufpp = bp;
 
 #ifdef TESTING
@@ -813,6 +926,9 @@ long buf_GetNewLocked(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bu
             return 0;
         } /* for all buffers in lru queue */
         lock_ReleaseWrite(&buf_globalLock);
+        lock_ReleaseRead(&scp->bufCreateLock);
+       osi_Log0(afsd_logp, "buf_GetNewLocked: Free Buffer List has no buffers with a zero refcount - sleeping 100ms");
+       Sleep(100);             /* give some time for a buffer to be freed */
     }  /* while loop over everything */
     /* not reached */
 } /* the proc */
@@ -865,8 +981,8 @@ long buf_GetNew(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bufpp)
      */
     lock_ReleaseMutex(&bp->mx);
     *bufpp = bp;
-    osi_Log3(buf_logp, "buf_GetNew returning bp 0x%x for file 0x%x, offset 0x%x",
-              bp, (long) scp, offsetp->LowPart);
+    osi_Log4(buf_logp, "buf_GetNew returning bp 0x%p for scp 0x%p, offset 0x%x:%08x",
+              bp, scp, offsetp->HighPart, offsetp->LowPart);
     return 0;
 }
 
@@ -897,16 +1013,17 @@ long buf_Get(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bufpp)
         if (bp) {
             /* lock it and break out */
             lock_ObtainMutex(&bp->mx);
-            break;
 
 #ifdef DISKCACHE95
             /* touch disk chunk to update LRU info */
             diskcache_Touch(bp->dcp);
 #endif /* DISKCACHE95 */
+            break;
         }
 
         /* otherwise, we have to create a page */
         code = buf_GetNewLocked(scp, &pageOffset, &bp);
+       /* bp->mx is now held */
 
         /* check if the buffer was created in a race condition branch.
          * If so, go around so we can hold a reference to it. 
@@ -932,13 +1049,7 @@ long buf_Get(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bufpp)
      */
     if (created) {
         /* load the page; freshly created pages should be idle */
-        osi_assert(!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING)));
-
-        /* setup offset, event */
-#ifndef DJGPP  /* doesn't seem to be used */
-        bp->over.Offset = bp->offset.LowPart;
-        bp->over.OffsetHigh = bp->offset.HighPart;
-#endif /* !DJGPP */
+        osi_assertx(!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING)), "incorrect cm_buf_t flags");
 
         /* start the I/O; may drop lock */
         bp->flags |= CM_BUF_READING;
@@ -951,15 +1062,13 @@ long buf_Get(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bufpp)
 
         if (code != 0) {
             /* failure or queued */
-#ifndef DJGPP   /* cm_bufRead always returns 0 */
             if (code != ERROR_IO_PENDING) {
-#endif
                 bp->error = code;
                 bp->flags |= CM_BUF_ERROR;
                 bp->flags &= ~CM_BUF_READING;
                 if (bp->flags & CM_BUF_WAITING) {
-                    osi_Log1(buf_logp, "buf_Get Waking bp 0x%x", bp);
-                    osi_Wakeup((long) bp);
+                    osi_Log1(buf_logp, "buf_Get Waking bp 0x%p", bp);
+                    osi_Wakeup((LONG_PTR) bp);
                 }
                 lock_ReleaseMutex(&bp->mx);
                 buf_Release(bp);
@@ -967,9 +1076,7 @@ long buf_Get(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bufpp)
                 buf_ValidateBufQueues();
 #endif /* TESTING */
                 return code;
-#ifndef DJGPP
             }
-#endif
         } else {
             /* otherwise, I/O completed instantly and we're done, except
              * for padding the xfr out with 0s and checking for EOF
@@ -981,8 +1088,8 @@ long buf_Get(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bufpp)
             }
             bp->flags &= ~CM_BUF_READING;
             if (bp->flags & CM_BUF_WAITING) {
-                osi_Log1(buf_logp, "buf_Get Waking bp 0x%x", bp);
-                osi_Wakeup((long) bp);
+                osi_Log1(buf_logp, "buf_Get Waking bp 0x%p", bp);
+                osi_Wakeup((LONG_PTR) bp);
             }
         }
 
@@ -1012,8 +1119,8 @@ long buf_Get(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bufpp)
     }
     lock_ReleaseWrite(&buf_globalLock);
 
-    osi_Log3(buf_logp, "buf_Get returning bp 0x%x for file 0x%x, offset 0x%x",
-              bp, (long) scp, offsetp->LowPart);
+    osi_Log4(buf_logp, "buf_Get returning bp 0x%p for scp 0x%p, offset 0x%x:%08x",
+              bp, scp, offsetp->HighPart, offsetp->LowPart);
 #ifdef TESTING
     buf_ValidateBufQueues();
 #endif /* TESTING */
@@ -1045,19 +1152,22 @@ long buf_CountFreeList(void)
 }
 
 /* clean a buffer synchronously */
-void buf_CleanAsync(cm_buf_t *bp, cm_req_t *reqp)
+long buf_CleanAsync(cm_buf_t *bp, cm_req_t *reqp)
 {
-    osi_assert(bp->magic == CM_BUF_MAGIC);
+    long code;
+    osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
 
     lock_ObtainMutex(&bp->mx);
-    buf_LockedCleanAsync(bp, reqp);
+    code = buf_CleanAsyncLocked(bp, reqp);
     lock_ReleaseMutex(&bp->mx);
+
+    return code;
 }       
 
 /* wait for a buffer's cleaning to finish */
 void buf_CleanWait(cm_scache_t * scp, cm_buf_t *bp)
 {
-    osi_assert(bp->magic == CM_BUF_MAGIC);
+    osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
 
     lock_ObtainMutex(&bp->mx);
     if (bp->flags & CM_BUF_WRITING) {
@@ -1071,18 +1181,64 @@ void buf_CleanWait(cm_scache_t * scp, cm_buf_t *bp)
  *
  * The buffer must be locked before calling this routine.
  */
-void buf_SetDirty(cm_buf_t *bp)
+void buf_SetDirty(cm_buf_t *bp, afs_uint32 offset, afs_uint32 length)
 {
-    osi_assert(bp->magic == CM_BUF_MAGIC);
-    osi_assert(bp->refCount > 0);
-       
-    osi_Log1(buf_logp, "buf_SetDirty 0x%x", bp);
+    osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
+    osi_assertx(bp->refCount > 0, "cm_buf_t refcount 0");
+
+    if (bp->flags & CM_BUF_DIRTY) {
 
-    /* set dirty bit */
-    bp->flags |= CM_BUF_DIRTY;
+       osi_Log1(buf_logp, "buf_SetDirty 0x%p already dirty", bp);
 
-    /* and turn off EOF flag, since it has associated data now */
-    bp->flags &= ~CM_BUF_EOF;
+        if (bp->dirty_offset <= offset) {
+            if (bp->dirty_offset + bp->dirty_length >= offset + length) {
+                /* dirty_length remains the same */
+            } else {
+                bp->dirty_length = offset + length - bp->dirty_offset;
+            }
+        } else /* bp->dirty_offset > offset */ {
+            if (bp->dirty_offset + bp->dirty_length >= offset + length) {
+                bp->dirty_length = bp->dirty_offset + bp->dirty_length - offset;
+            } else {
+                bp->dirty_length = length;
+            }
+            bp->dirty_offset = offset;
+        }
+    } else {
+       osi_Log1(buf_logp, "buf_SetDirty 0x%p", bp);
+
+        /* set dirty bit */
+        bp->flags |= CM_BUF_DIRTY;
+
+        /* and turn off EOF flag, since it has associated data now */
+        bp->flags &= ~CM_BUF_EOF;
+
+        bp->dirty_offset = offset;
+        bp->dirty_length = length;
+
+        /* and add to the dirty list.  
+         * we obtain a hold on the buffer for as long as it remains 
+         * in the list.  buffers are only removed from the list by 
+         * the buf_IncrSyncer function regardless of when else the
+         * dirty flag might be cleared.
+         *
+         * This should never happen but just in case there is a bug
+         * elsewhere, never add to the dirty list if the buffer is 
+         * already there.
+         */
+        lock_ObtainWrite(&buf_globalLock);
+        if (bp->dirtyp == NULL && cm_data.buf_dirtyListEndp != bp) {
+            buf_HoldLocked(bp);
+            if (!cm_data.buf_dirtyListp) {
+                cm_data.buf_dirtyListp = cm_data.buf_dirtyListEndp = bp;
+            } else {
+                cm_data.buf_dirtyListEndp->dirtyp = bp;
+                cm_data.buf_dirtyListEndp = bp;
+            }
+            bp->dirtyp = NULL;
+        }
+        lock_ReleaseWrite(&buf_globalLock);
+    }
 }
 
 /* clean all buffers, reset log pointers and invalidate all buffers.
@@ -1107,31 +1263,33 @@ void buf_SetDirty(cm_buf_t *bp)
  */
 long buf_CleanAndReset(void)
 {
-    long i;
+    afs_uint32 i;
     cm_buf_t *bp;
     cm_req_t req;
 
-    lock_ObtainWrite(&buf_globalLock);
+    lock_ObtainRead(&buf_globalLock);
     for(i=0; i<cm_data.buf_hashSize; i++) {
-        for(bp = cm_data.buf_hashTablepp[i]; bp; bp = bp->hashp) {
+        for(bp = cm_data.buf_scacheHashTablepp[i]; bp; bp = bp->hashp) {
             if ((bp->flags & CM_BUF_DIRTY) == CM_BUF_DIRTY) {
-                bp->refCount++;
-                lock_ReleaseWrite(&buf_globalLock);
+                buf_HoldLocked(bp);
+                lock_ReleaseRead(&buf_globalLock);
 
                 /* now no locks are held; clean buffer and go on */
                 cm_InitReq(&req);
-                buf_CleanAsync(bp, &req);
-                buf_CleanWait(NULL, bp);
+               req.flags |= CM_REQ_NORETRY;
+
+               buf_CleanAsync(bp, &req);
+               buf_CleanWait(NULL, bp);
 
                 /* relock and release buffer */
-                lock_ObtainWrite(&buf_globalLock);
-                buf_LockedRelease(bp);
+                lock_ObtainRead(&buf_globalLock);
+                buf_ReleaseLocked(bp, FALSE);
             } /* dirty */
         } /* over one bucket */
     }  /* for loop over all hash buckets */
 
     /* release locks */
-    lock_ReleaseWrite(&buf_globalLock);
+    lock_ReleaseRead(&buf_globalLock);
 
 #ifdef TESTING
     buf_ValidateBufQueues();
@@ -1144,14 +1302,14 @@ long buf_CleanAndReset(void)
 /* called without global lock being held, reserves buffers for callers
  * that need more than one held (not locked) at once.
  */
-void buf_ReserveBuffers(long nbuffers)
+void buf_ReserveBuffers(afs_uint64 nbuffers)
 {
     lock_ObtainWrite(&buf_globalLock);
     while (1) {
         if (cm_data.buf_reservedBufs + nbuffers > cm_data.buf_maxReservedBufs) {
             cm_data.buf_reserveWaiting = 1;
             osi_Log1(buf_logp, "buf_ReserveBuffers waiting for %d bufs", nbuffers);
-            osi_SleepW((long) &cm_data.buf_reservedBufs, &buf_globalLock);
+            osi_SleepW((LONG_PTR) &cm_data.buf_reservedBufs, &buf_globalLock);
             lock_ObtainWrite(&buf_globalLock);
         }
         else {
@@ -1162,7 +1320,7 @@ void buf_ReserveBuffers(long nbuffers)
     lock_ReleaseWrite(&buf_globalLock);
 }
 
-int buf_TryReserveBuffers(long nbuffers)
+int buf_TryReserveBuffers(afs_uint64 nbuffers)
 {
     int code;
 
@@ -1181,13 +1339,13 @@ int buf_TryReserveBuffers(long nbuffers)
 /* called without global lock held, releases reservation held by
  * buf_ReserveBuffers.
  */
-void buf_UnreserveBuffers(long nbuffers)
+void buf_UnreserveBuffers(afs_uint64 nbuffers)
 {
     lock_ObtainWrite(&buf_globalLock);
     cm_data.buf_reservedBufs -= nbuffers;
     if (cm_data.buf_reserveWaiting) {
         cm_data.buf_reserveWaiting = 0;
-        osi_Wakeup((long) &cm_data.buf_reservedBufs);
+        osi_Wakeup((LONG_PTR) &cm_data.buf_reservedBufs);
     }
     lock_ReleaseWrite(&buf_globalLock);
 }       
@@ -1205,25 +1363,23 @@ long buf_Truncate(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp,
     osi_hyper_t bufEnd;
     long code;
     long bufferPos;
-    int didRelease;
-    long i;
+    afs_uint32 i;
 
     /* assert that cm_bufCreateLock is held in write mode */
     lock_AssertWrite(&scp->bufCreateLock);
 
     i = BUF_FILEHASH(&scp->fid);
 
-    lock_ObtainWrite(&buf_globalLock);
+    lock_ObtainRead(&buf_globalLock);
     bufp = cm_data.buf_fileHashTablepp[i];
     if (bufp == NULL) {
-        lock_ReleaseWrite(&buf_globalLock);
+        lock_ReleaseRead(&buf_globalLock);
         return 0;
     }
 
-    bufp->refCount++;
-    lock_ReleaseWrite(&buf_globalLock);
-    for(; bufp; bufp = nbufp) {
-        didRelease = 0;
+    buf_HoldLocked(bufp);
+    lock_ReleaseRead(&buf_globalLock);
+    while (bufp) {
         lock_ObtainMutex(&bufp->mx);
 
         bufEnd.HighPart = 0;
@@ -1244,14 +1400,15 @@ long buf_Truncate(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp,
                           | CM_SCACHESYNC_GETSTATUS
                           | CM_SCACHESYNC_SETSIZE
                           | CM_SCACHESYNC_BUFLOCKED);
-        /* if we succeeded in our locking, and this applies to the right
+
+       
+       /* if we succeeded in our locking, and this applies to the right
          * file, and the truncate request overlaps the buffer either
          * totally or partially, then do something.
          */
         if (code == 0 && cm_FidCmp(&bufp->fid, &scp->fid) == 0
              && LargeIntegerLessThan(*sizep, bufEnd)) {
 
-            lock_ObtainWrite(&buf_globalLock);
 
             /* destroy the buffer, turning off its dirty bit, if
              * we're truncating the whole buffer.  Otherwise, set
@@ -1261,6 +1418,8 @@ long buf_Truncate(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp,
             if (LargeIntegerLessThanOrEqualTo(*sizep, bufp->offset)) {
                 /* truncating the entire page */
                 bufp->flags &= ~CM_BUF_DIRTY;
+                bufp->dirty_offset = 0;
+                bufp->dirty_length = 0;
                 bufp->dataVersion = -1;        /* known bad */
                 bufp->dirtyCounter++;
             }
@@ -1274,46 +1433,38 @@ long buf_Truncate(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp,
                  * visible again.
                  */
                 bufferPos = sizep->LowPart & (cm_data.buf_blockSize - 1);
-                osi_assert(bufferPos != 0);
+                osi_assertx(bufferPos != 0, "non-zero bufferPos");
                 memset(bufp->datap + bufferPos, 0,
                         cm_data.buf_blockSize - bufferPos);
             }
-
-            lock_ReleaseWrite(&buf_globalLock);
         }
                
+       cm_SyncOpDone( scp, bufp, 
+                      CM_SCACHESYNC_NEEDCALLBACK | CM_SCACHESYNC_GETSTATUS
+                      | CM_SCACHESYNC_SETSIZE | CM_SCACHESYNC_BUFLOCKED);
+
         lock_ReleaseMutex(&scp->mx);
         lock_ReleaseMutex(&bufp->mx);
-        if (!didRelease) {
-            lock_ObtainWrite(&buf_globalLock);
-            nbufp = bufp->fileHashp;
-            if (nbufp) nbufp->refCount++;
-            buf_LockedRelease(bufp);
-            lock_ReleaseWrite(&buf_globalLock);
-        }
-
-        /* bail out early if we fail */
-        if (code) {
-            /* at this point, nbufp is held; bufp has already been
-             * released.
-             */
-            if (nbufp) 
-                buf_Release(nbufp);
-
-#ifdef TESTING
-            buf_ValidateBufQueues();
-#endif /* TESTING */
-
-            return code;
-        }
+    
+       if (!code) {
+           nbufp = bufp->fileHashp;
+           if (nbufp) 
+               buf_Hold(nbufp);
+       } else {
+           /* This forces the loop to end and the error code
+            * to be returned. */
+           nbufp = NULL;
+       }
+       buf_Release(bufp);
+       bufp = nbufp;
     }
 
 #ifdef TESTING
     buf_ValidateBufQueues();
 #endif /* TESTING */
 
-    /* success */
-    return 0;
+    /* done */
+    return code;
 }
 
 long buf_FlushCleanPages(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp)
@@ -1322,16 +1473,17 @@ long buf_FlushCleanPages(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp)
     cm_buf_t *bp;              /* buffer we're hacking on */
     cm_buf_t *nbp;
     int didRelease;
-    long i;
+    afs_uint32 i;
 
     i = BUF_FILEHASH(&scp->fid);
 
     code = 0;
-    lock_ObtainWrite(&buf_globalLock);
+    lock_ObtainRead(&buf_globalLock);
     bp = cm_data.buf_fileHashTablepp[i];
     if (bp) 
-        bp->refCount++;
-    lock_ReleaseWrite(&buf_globalLock);
+        buf_HoldLocked(bp);
+    lock_ReleaseRead(&buf_globalLock);
+    
     for (; bp; bp = nbp) {
         didRelease = 0;        /* haven't released this buffer yet */
 
@@ -1340,66 +1492,111 @@ long buf_FlushCleanPages(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp)
             lock_ObtainMutex(&bp->mx);
 
             /* start cleaning the buffer, and wait for it to finish */
-            buf_LockedCleanAsync(bp, reqp);
+            buf_CleanAsyncLocked(bp, reqp);
             buf_WaitIO(scp, bp);
             lock_ReleaseMutex(&bp->mx);
 
             code = (*cm_buf_opsp->Stabilizep)(scp, userp, reqp);
-            if (code) 
+            if (code && code != CM_ERROR_BADFD) 
                 goto skip;
 
-            lock_ObtainWrite(&buf_globalLock);
+           /* if the scp's FID is bad its because we received VNOVNODE 
+            * when attempting to FetchStatus before the write.  This
+            * page therefore contains data that can no longer be stored.
+            */
+           lock_ObtainMutex(&bp->mx);
+           bp->flags &= ~CM_BUF_DIRTY;
+           bp->flags |= CM_BUF_ERROR;
+           bp->error = code;
+            bp->dirty_offset = 0;
+            bp->dirty_length = 0;
+            bp->dataVersion = -1;      /* known bad */
+            bp->dirtyCounter++;
+           lock_ReleaseMutex(&bp->mx);
+
             /* actually, we only know that buffer is clean if ref
              * count is 1, since we don't have buffer itself locked.
              */
             if (!(bp->flags & CM_BUF_DIRTY)) {
+                lock_ObtainWrite(&buf_globalLock);
                 if (bp->refCount == 1) {       /* bp is held above */
-                    buf_LockedRelease(bp);
                     nbp = bp->fileHashp;
                     if (nbp) 
-                        nbp->refCount++;
+                        buf_HoldLocked(nbp);
+                    buf_ReleaseLocked(bp, TRUE);
                     didRelease = 1;
                     buf_Recycle(bp);
                 }
+                lock_ReleaseWrite(&buf_globalLock);
             }
-            lock_ReleaseWrite(&buf_globalLock);
 
-            (*cm_buf_opsp->Unstabilizep)(scp, userp);
+           if (code != CM_ERROR_BADFD)
+               (*cm_buf_opsp->Unstabilizep)(scp, userp);
         }
 
       skip:
         if (!didRelease) {
-            lock_ObtainWrite(&buf_globalLock);
-            if (nbp = bp->fileHashp) 
-                nbp->refCount++;
-            buf_LockedRelease(bp);
-            lock_ReleaseWrite(&buf_globalLock);
+            lock_ObtainRead(&buf_globalLock);
+            nbp = bp->fileHashp;
+           if (nbp)
+                buf_HoldLocked(nbp);
+            buf_ReleaseLocked(bp, FALSE);
+            lock_ReleaseRead(&buf_globalLock);
         }
     }  /* for loop over a bunch of buffers */
 
 #ifdef TESTING
-            buf_ValidateBufQueues();
+    buf_ValidateBufQueues();
 #endif /* TESTING */
 
     /* done */
     return code;
 }       
 
+/* Must be called with scp->mx held */
+long buf_ForceDataVersion(cm_scache_t * scp, afs_uint64 fromVersion, afs_uint64 toVersion)
+{
+    cm_buf_t * bp;
+    afs_uint32 i;
+    int found = 0;
+
+    lock_AssertMutex(&scp->mx);
+
+    i = BUF_FILEHASH(&scp->fid);
+
+    lock_ObtainRead(&buf_globalLock);
+
+    for (bp = cm_data.buf_fileHashTablepp[i]; bp; bp = bp->fileHashp) {
+        if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
+            if (bp->dataVersion == fromVersion) {
+                bp->dataVersion = toVersion;
+                found = 1;
+            }
+        }
+    }
+    lock_ReleaseRead(&buf_globalLock);
+
+    if (found)
+        return 0;
+    else
+        return ENOENT;
+}
+
 long buf_CleanVnode(struct cm_scache *scp, cm_user_t *userp, cm_req_t *reqp)
 {
-    long code;
+    long code = 0;
+    long wasDirty = 0;
     cm_buf_t *bp;              /* buffer we're hacking on */
     cm_buf_t *nbp;             /* next one */
-    long i;
+    afs_uint32 i;
 
     i = BUF_FILEHASH(&scp->fid);
 
-    code = 0;
-    lock_ObtainWrite(&buf_globalLock);
+    lock_ObtainRead(&buf_globalLock);
     bp = cm_data.buf_fileHashTablepp[i];
     if (bp) 
-        bp->refCount++;
-    lock_ReleaseWrite(&buf_globalLock);
+        buf_HoldLocked(bp);
+    lock_ReleaseRead(&buf_globalLock);
     for (; bp; bp = nbp) {
         /* clean buffer synchronously */
         if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
@@ -1411,24 +1608,23 @@ long buf_CleanVnode(struct cm_scache *scp, cm_user_t *userp, cm_req_t *reqp)
                 bp->userp = userp;
                 lock_ReleaseMutex(&bp->mx);
             }   
-            buf_CleanAsync(bp, reqp);
-            buf_CleanWait(scp, bp);
+            wasDirty = buf_CleanAsync(bp, reqp);
+           buf_CleanWait(scp, bp);
             lock_ObtainMutex(&bp->mx);
             if (bp->flags & CM_BUF_ERROR) {
-                if (code == 0 || code == -1) 
-                    code = bp->error;
+               code = bp->error;
                 if (code == 0) 
                     code = -1;
             }
             lock_ReleaseMutex(&bp->mx);
         }
 
-        lock_ObtainWrite(&buf_globalLock);
-        buf_LockedRelease(bp);
+        lock_ObtainRead(&buf_globalLock);
         nbp = bp->fileHashp;
         if (nbp) 
-            nbp->refCount++;
-        lock_ReleaseWrite(&buf_globalLock);
+            buf_HoldLocked(nbp);
+        buf_ReleaseLocked(bp, FALSE);
+        lock_ReleaseRead(&buf_globalLock);
     }  /* for loop over a bunch of buffers */
 
 #ifdef TESTING
@@ -1449,68 +1645,102 @@ buf_ValidateBufQueues(void)
     lock_ObtainRead(&buf_globalLock);
     for (bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
         if (bp->magic != CM_BUF_MAGIC)
-            DebugBreak();
+            osi_panic("buf magic error",__FILE__,__LINE__);
         countb++;
         bpb = bp;
     }
 
     for (bp = cm_data.buf_freeListp; bp; bp=(cm_buf_t *) osi_QNext(&bp->q)) {
         if (bp->magic != CM_BUF_MAGIC)
-            DebugBreak();
+            osi_panic("buf magic error",__FILE__,__LINE__);
         countf++;
         bpf = bp;
     }
 
     for (bp = cm_data.buf_allp; bp; bp=bp->allp) {
         if (bp->magic != CM_BUF_MAGIC)
-            DebugBreak();
+            osi_panic("buf magic error",__FILE__,__LINE__);
         counta++;
         bpa = bp;
     }
     lock_ReleaseRead(&buf_globalLock);
 
     if (countb != countf)
-        DebugBreak();
+       osi_panic("buf magic error",__FILE__,__LINE__);
 
     if (counta != cm_data.buf_nbuffers)
-        DebugBreak();  
+       osi_panic("buf magic error",__FILE__,__LINE__);
 }
 #endif /* TESTING */
 
-/* dump the contents of the buf_hashTablepp. */
+/* dump the contents of the buf_scacheHashTablepp. */
 int cm_DumpBufHashTable(FILE *outputFile, char *cookie, int lock)
 {
     int zilch;
     cm_buf_t *bp;
     char output[1024];
-    int i;
+    afs_uint32 i;
   
-    if (cm_data.buf_hashTablepp == NULL)
+    if (cm_data.buf_scacheHashTablepp == NULL)
         return -1;
 
     if (lock)
         lock_ObtainRead(&buf_globalLock);
   
-    StringCbPrintfA(output, sizeof(output), "%s - dumping buf_HashTable - buf_hashSize=%d\n", 
+    StringCbPrintfA(output, sizeof(output), "%s - dumping buf_HashTable - buf_hashSize=%d\r\n", 
                     cookie, cm_data.buf_hashSize);
-    WriteFile(outputFile, output, strlen(output), &zilch, NULL);
+    WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
   
     for (i = 0; i < cm_data.buf_hashSize; i++)
     {
-        for (bp = cm_data.buf_hashTablepp[i]; bp; bp=bp->hashp) 
+        for (bp = cm_data.buf_scacheHashTablepp[i]; bp; bp=bp->hashp) 
         {
-            if (bp->refCount)
-            {
-                StringCbPrintfA(output, sizeof(output), "vnode=%d, unique=%d), size=%d refCount=%d\n", 
-                        cookie, (void *)bp, i, bp->fid.cell, bp->fid.volume, 
-                        bp->fid.vnode, bp->fid.unique, bp->size, bp->refCount);
-                WriteFile(outputFile, output, strlen(output), &zilch, NULL);
-            }
+           StringCbPrintfA(output, sizeof(output), 
+                           "%s bp=0x%08X, hash=%d, fid (cell=%d, volume=%d, "
+                           "vnode=%d, unique=%d), offset=%x:%08x, dv=%I64d, "
+                           "flags=0x%x, cmFlags=0x%x, refCount=%d\r\n",
+                            cookie, (void *)bp, i, bp->fid.cell, bp->fid.volume, 
+                            bp->fid.vnode, bp->fid.unique, bp->offset.HighPart, 
+                            bp->offset.LowPart, bp->dataVersion, bp->flags, 
+                            bp->cmFlags, bp->refCount);
+           WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
         }
     }
   
-    StringCbPrintfA(output, sizeof(output), "%s - Done dumping buf_HashTable.\n", cookie);
-    WriteFile(outputFile, output, strlen(output), &zilch, NULL);
+    StringCbPrintfA(output, sizeof(output), "%s - Done dumping buf_HashTable.\r\n", cookie);
+    WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
+
+    StringCbPrintfA(output, sizeof(output), "%s - dumping buf_freeListEndp\r\n", cookie);
+    WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
+    for(bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
+       StringCbPrintfA(output, sizeof(output), 
+                        "%s bp=0x%08X, fid (cell=%d, volume=%d, "
+                        "vnode=%d, unique=%d), offset=%x:%08x, dv=%I64d, "
+                        "flags=0x%x, cmFlags=0x%x, refCount=%d\r\n",
+                        cookie, (void *)bp, bp->fid.cell, bp->fid.volume, 
+                        bp->fid.vnode, bp->fid.unique, bp->offset.HighPart, 
+                        bp->offset.LowPart, bp->dataVersion, bp->flags, 
+                        bp->cmFlags, bp->refCount);
+       WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
+    }
+    StringCbPrintfA(output, sizeof(output), "%s - Done dumping buf_FreeListEndp.\r\n", cookie);
+    WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
+
+    StringCbPrintfA(output, sizeof(output), "%s - dumping buf_dirtyListEndp\r\n", cookie);
+    WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
+    for(bp = cm_data.buf_dirtyListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
+       StringCbPrintfA(output, sizeof(output), 
+                        "%s bp=0x%08X, fid (cell=%d, volume=%d, "
+                        "vnode=%d, unique=%d), offset=%x:%08x, dv=%I64d, "
+                        "flags=0x%x, cmFlags=0x%x, refCount=%d\r\n",
+                        cookie, (void *)bp, bp->fid.cell, bp->fid.volume, 
+                        bp->fid.vnode, bp->fid.unique, bp->offset.HighPart, 
+                        bp->offset.LowPart, bp->dataVersion, bp->flags, 
+                        bp->cmFlags, bp->refCount);
+       WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
+    }
+    StringCbPrintfA(output, sizeof(output), "%s - Done dumping buf_dirtyListEndp.\r\n", cookie);
+    WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
 
     if (lock)
         lock_ReleaseRead(&buf_globalLock);
@@ -1538,3 +1768,49 @@ void buf_ForceTrace(BOOL flush)
         FlushFileBuffers(handle);
     CloseHandle(handle);
 }
+
+long buf_DirtyBuffersExist(cm_fid_t *fidp)
+{
+    cm_buf_t *bp;
+    afs_uint32 bcount = 0;
+    afs_uint32 i;
+
+    i = BUF_FILEHASH(fidp);
+
+    for (bp = cm_data.buf_fileHashTablepp[i]; bp; bp=bp->allp, bcount++) {
+       if (!cm_FidCmp(fidp, &bp->fid) && (bp->flags & CM_BUF_DIRTY))
+           return 1;
+    }
+    return 0;
+}
+
+#if 0
+long buf_CleanDirtyBuffers(cm_scache_t *scp)
+{
+    cm_buf_t *bp;
+    afs_uint32 bcount = 0;
+    cm_fid_t * fidp = &scp->fid;
+
+    for (bp = cm_data.buf_allp; bp; bp=bp->allp, bcount++) {
+       if (!cm_FidCmp(fidp, &bp->fid) && (bp->flags & CM_BUF_DIRTY)) {
+            buf_Hold(bp);
+           lock_ObtainMutex(&bp->mx);
+           bp->cmFlags &= ~CM_BUF_CMSTORING;
+           bp->flags &= ~CM_BUF_DIRTY;
+            bp->dirty_offset = 0;
+            bp->dirty_length = 0;
+           bp->flags |= CM_BUF_ERROR;
+           bp->error = VNOVNODE;
+           bp->dataVersion = -1; /* bad */
+           bp->dirtyCounter++;
+           if (bp->flags & CM_BUF_WAITING) {
+               osi_Log2(buf_logp, "BUF CleanDirtyBuffers Waking [scp 0x%x] bp 0x%x", scp, bp);
+               osi_Wakeup((long) &bp);
+           }
+           lock_ReleaseMutex(&bp->mx);
+           buf_Release(bp);
+       }
+    }
+    return 0;
+}
+#endif