[svn-r6387] Purpose:

    Bug Fix

Description:
    Metadata cache in parallel I/O can cause hangs in applications which
    perform independent I/O on chunked datasets, because the metadata cache
    can attempt to flush out dirty metadata from only a single process, instead
    of collectively from all processes.

Solution:
    Pass a dataset transfer property list down from every API function which
    could possibly trigger metadata I/O.

    Then, split the metadata cache into two sets of entries to allow dirty
    metadata to be set aside when a hash table collision occurs during
    independent I/O.

Platforms tested:
    Tested h5committest {arabica (fortran), eirene (fortran, C++)
        modi4 (parallel, fortran)}

    FreeBSD 4.7 (sleipnir) serial & parallel

Misc. update:
    Updated release_docs/RELEASE

1 files changed, 37 insertions, 25 deletions

diff --git a/src/H5ACprivate.h b/src/H5ACprivate.h
index d587df6..96cf05c 100644
--- a/src/H5ACprivate.h
+++ b/src/H5ACprivate.h
@@ -34,6 +34,16 @@
 #  undef H5AC_DEBUG
 #endif
 
+/* Dataset transfer property list for flush calls */
+/* (Collective set, "block before metadata write" set and "library internal" set) */
+/* (Global variable declaration, definition is in H5AC.c) */
+extern hid_t H5AC_dxpl_id;
+
+/* Dataset transfer property list for independent metadata I/O calls */
+/* (just "library internal" set - i.e. independent transfer mode) */
+/* (Global variable declaration, definition is in H5AC.c) */
+extern hid_t H5AC_ind_dxpl_id;
+
 /*
  * Class methods pertaining to caching.	 Each type of cached object will
  * have a constant variable with permanent life-span that describes how
@@ -58,13 +68,15 @@ typedef enum H5AC_subid_t {
     H5AC_NTYPES		= 5	/*THIS MUST BE LAST!			     */
 } H5AC_subid_t;
 
-typedef void *(*H5AC_load_func_t)(H5F_t*, hid_t dxpl_id, haddr_t addr, const void *udata1, void *udata2);
-typedef herr_t (*H5AC_flush_func_t)(H5F_t*, hid_t dxpl_id, hbool_t dest, haddr_t addr, void *thing);
+typedef void *(*H5AC_load_func_t)(H5F_t *f, hid_t dxpl_id, haddr_t addr, const void *udata1, void *udata2);
+typedef herr_t (*H5AC_flush_func_t)(H5F_t *f, hid_t dxpl_id, hbool_t dest, haddr_t addr, void *thing);
+typedef herr_t (*H5AC_dest_func_t)(H5F_t *f, void *thing);
 
 typedef struct H5AC_class_t {
     H5AC_subid_t	id;
     H5AC_load_func_t     load;
     H5AC_flush_func_t    flush;
+    H5AC_dest_func_t    dest;
 } H5AC_class_t;
 
 /*
@@ -81,6 +93,7 @@ typedef struct H5AC_class_t {
 typedef struct H5AC_info_t {
     const H5AC_class_t	*type;		/*type of object stored here	     */
     haddr_t		addr;		/*file address for object	     */
+    hbool_t             dirty;          /* 'Dirty' flag for cached object */
 } H5AC_info_t;
 typedef H5AC_info_t *H5AC_info_ptr_t;   /* Typedef for free lists */
 
@@ -95,6 +108,7 @@ typedef struct H5AC_prot_t {
 typedef struct H5AC_t {
     unsigned	nslots;			/*number of cache slots		     */
     H5AC_info_t **slot;		/*the cache slots, an array of pointers to the cached objects */
+    H5AC_info_t **dslot;	/*"held object" cache slots, an array of pointers to dirty cached objects */
 #ifdef H5AC_DEBUG
     H5AC_prot_t *prot;		/*the protected slots		     */
 #endif /* H5AC_DEBUG */
@@ -105,6 +119,11 @@ typedef struct H5AC_t {
 	unsigned	nmisses;		/*number of cache misses	     */
 	unsigned	ninits;			/*number of cache inits		     */
 	unsigned	nflushes;		/*number of flushes to disk	     */
+#ifdef H5_HAVE_PARALLEL
+	unsigned	ndestroys;		/*number of cache destroys	     */
+	unsigned	nholds;			/*number of cache holds	     */
+	unsigned	nrestores;		/*number of cache restores	     */
+#endif /* H5_HAVE_PARALLEL */
     } diagnostics[H5AC_NTYPES];		/*diagnostics for each type of object*/
 #endif /* H5AC_DEBUG */
 } H5AC_t;
@@ -114,39 +133,32 @@ typedef struct H5AC_t {
 #define H5AC_BLOCK_BEFORE_META_WRITE_NAME       "H5AC_block_before_meta_write"
 #define H5AC_BLOCK_BEFORE_META_WRITE_SIZE       sizeof(unsigned)
 #define H5AC_BLOCK_BEFORE_META_WRITE_DEF        0
+
+/* Definitions for "library internal" property */
+#define H5AC_LIBRARY_INTERNAL_NAME       "H5AC_library_internal"
+#define H5AC_LIBRARY_INTERNAL_SIZE       sizeof(unsigned)
+#define H5AC_LIBRARY_INTERNAL_DEF        0
 #endif /* H5_HAVE_PARALLEL */
 
 /*
  * Library prototypes.
  */
-H5_DLL herr_t H5AC_dest(H5F_t *f);
-H5_DLL void *H5AC_find_f(H5F_t *f, const H5AC_class_t *type, haddr_t addr,
-			  const void *udata1, void *udata2);
-H5_DLL void *H5AC_protect(H5F_t *f, const H5AC_class_t *type, haddr_t addr,
+H5_DLL herr_t H5AC_init(void);
+H5_DLL herr_t H5AC_create(H5F_t *f, int size_hint);
+H5_DLL herr_t H5AC_set(H5F_t *f, hid_t dxpl_id, const H5AC_class_t *type, haddr_t addr,
+			void *thing);
+H5_DLL void *H5AC_protect(H5F_t *f, hid_t dxpl_id, const H5AC_class_t *type, haddr_t addr,
 			   const void *udata1, void *udata2);
-H5_DLL herr_t H5AC_unprotect(H5F_t *f, const H5AC_class_t *type, haddr_t addr,
+H5_DLL herr_t H5AC_unprotect(H5F_t *f, hid_t dxpl_id, const H5AC_class_t *type, haddr_t addr,
 			      void *thing);
-H5_DLL herr_t H5AC_flush(H5F_t *f, const H5AC_class_t *type, haddr_t addr,
+H5_DLL void *H5AC_find(H5F_t *f, hid_t dxpl_id, const H5AC_class_t *type,
+        haddr_t addr, const void *udata1, void *udata2);
+H5_DLL herr_t H5AC_flush(H5F_t *f, hid_t dxpl_id, const H5AC_class_t *type, haddr_t addr,
 			  hbool_t destroy);
-H5_DLL herr_t H5AC_create(H5F_t *f, int size_hint);
-H5_DLL herr_t H5AC_rename(H5F_t *f, const H5AC_class_t *type,
+H5_DLL herr_t H5AC_rename(H5F_t *f, hid_t dxpl_id, const H5AC_class_t *type,
 			   haddr_t old_addr, haddr_t new_addr);
-H5_DLL herr_t H5AC_set(H5F_t *f, const H5AC_class_t *type, haddr_t addr,
-			void *thing);
+H5_DLL herr_t H5AC_dest(H5F_t *f, hid_t dxpl_id);
 H5_DLL herr_t H5AC_debug(H5F_t *f);
 
-/* This seems to be slower than just calling H5AC_find_f, which performs the
- * cache lookup also. - QAK
- */
-#ifdef OLD_WAY
-#define H5AC_find(F,TYPE,ADDR,UDATA1,UDATA2)				      \
-   ((F)->shared->cache->slot[H5AC_HASH(F,ADDR)]!=NULL &&	      \
-    ((F)->shared->cache->slot[H5AC_HASH(F,ADDR)]->type==(TYPE) &&	      \
-     (F)->shared->cache->slot[H5AC_HASH(F,ADDR)]->addr==ADDR) ?		      \
-    ((F)->shared->cache->diagnostics[(TYPE)->id].nhits++,		      \
-     (F)->shared->cache->slot[H5AC_HASH(F,ADDR)]) :		      \
-    H5AC_find_f(F, TYPE, ADDR, UDATA1, UDATA2))
-#endif /* OLD_WAY */
-     
 #endif /* !_H5ACprivate_H */