Fix some collective metadata read issues

6 files changed, 29 insertions, 64 deletions

diff --git a/src/H5C.c b/src/H5C.c
index 3e4f1ad..288f3db 100644
--- a/src/H5C.c
+++ b/src/H5C.c
@@ -1483,31 +1483,17 @@ H5C_insert_entry(H5F_t *             f,
     H5C__UPDATE_STATS_FOR_INSERTION(cache_ptr, entry_ptr)
 
 #ifdef H5_HAVE_PARALLEL
-    if(H5F_HAS_FEATURE(f, H5FD_FEAT_HAS_MPI)) {
-        coll_access = (H5P_USER_TRUE == f->coll_md_read ? TRUE : FALSE);
-
-        /* If not explicitly disabled, get the cmdr setting from the API context */
-        if(!coll_access && H5P_FORCE_FALSE != f->coll_md_read)
-            coll_access = H5CX_get_coll_metadata_read();
-    } /* end if */
+    if(H5F_HAS_FEATURE(f, H5FD_FEAT_HAS_MPI))
+        coll_access = H5CX_get_coll_metadata_read();
 
     entry_ptr->coll_access = coll_access;
     if(coll_access) {
         H5C__INSERT_IN_COLL_LIST(cache_ptr, entry_ptr, FAIL)
 
         /* Make sure the size of the collective entries in the cache remain in check */
-        if(H5P_USER_TRUE == f->coll_md_read) {
-            if(cache_ptr->max_cache_size * 80 < cache_ptr->coll_list_size * 100) {
-                if(H5C_clear_coll_entries(cache_ptr, TRUE) < 0)
-                    HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, "can't clear collective metadata entries")
-            } /* end if */
-        } /* end if */
-        else {
-            if(cache_ptr->max_cache_size * 40 < cache_ptr->coll_list_size * 100) {
-                if(H5C_clear_coll_entries(cache_ptr, TRUE) < 0)
-                    HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, "can't clear collective metadata entries")
-            } /* end if */
-        } /* end else */
+        if(cache_ptr->max_cache_size * 80 < cache_ptr->coll_list_size * 100)
+            if(H5C_clear_coll_entries(cache_ptr, TRUE) < 0)
+                HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, FAIL, "can't clear collective metadata entries")
     } /* end if */
 #endif
 
@@ -2243,13 +2229,8 @@ H5C_protect(H5F_t *		f,
     ring = H5CX_get_ring();
 
 #ifdef H5_HAVE_PARALLEL
-    if(H5F_HAS_FEATURE(f, H5FD_FEAT_HAS_MPI)) {
-        coll_access = (H5P_USER_TRUE == f->coll_md_read ? TRUE : FALSE);
-
-        /* If not explicitly disabled, get the cmdr setting from the API context */
-        if(!coll_access && H5P_FORCE_FALSE != f->coll_md_read)
-            coll_access = H5CX_get_coll_metadata_read();
-    } /* end if */
+    if(H5F_HAS_FEATURE(f, H5FD_FEAT_HAS_MPI))
+        coll_access = H5CX_get_coll_metadata_read();
 #endif /* H5_HAVE_PARALLEL */
 
     /* first check to see if the target is in cache */
@@ -2286,7 +2267,7 @@ H5C_protect(H5F_t *		f,
            the entry in their cache still have to participate in the
            bcast. */
 #ifdef H5_HAVE_PARALLEL
-        if(H5F_HAS_FEATURE(f, H5FD_FEAT_HAS_MPI) && coll_access) {
+        if(coll_access) {
             if(!(entry_ptr->is_dirty) && !(entry_ptr->coll_access)) {
                 MPI_Comm  comm;           /* File MPI Communicator */
                 int       mpi_code;       /* MPI error code */
@@ -2601,21 +2582,11 @@ H5C_protect(H5F_t *		f,
     } /* end if */
 
 #ifdef H5_HAVE_PARALLEL
-    if(H5F_HAS_FEATURE(f, H5FD_FEAT_HAS_MPI)) {
-        /* Make sure the size of the collective entries in the cache remain in check */
-        if(coll_access) {
-            if(H5P_USER_TRUE == f->coll_md_read) {
-                if(cache_ptr->max_cache_size * 80 < cache_ptr->coll_list_size * 100)
-                    if(H5C_clear_coll_entries(cache_ptr, TRUE) < 0)
-                        HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, NULL, "can't clear collective metadata entries")
-            } /* end if */
-            else {
-                if(cache_ptr->max_cache_size * 40 < cache_ptr->coll_list_size * 100)
-                    if(H5C_clear_coll_entries(cache_ptr, TRUE) < 0)
-                        HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, NULL, "can't clear collective metadata entries")
-            } /* end else */
-        } /* end if */
-    } /* end if */
+    /* Make sure the size of the collective entries in the cache remain in check */
+    if(coll_access)
+        if(cache_ptr->max_cache_size * 80 < cache_ptr->coll_list_size * 100)
+            if(H5C_clear_coll_entries(cache_ptr, TRUE) < 0)
+                HGOTO_ERROR(H5E_CACHE, H5E_CANTFLUSH, NULL, "can't clear collective metadata entries")
 #endif /* H5_HAVE_PARALLEL */
 
 done:
diff --git a/src/H5CX.c b/src/H5CX.c
index 9ef2886..6b5e510 100644
--- a/src/H5CX.c
+++ b/src/H5CX.c
@@ -807,8 +807,8 @@ H5CX_set_apl(hid_t *acspl_id, const H5P_libclass_t *libclass,
 
 #ifdef H5_HAVE_PARALLEL
         /* If this routine is not guaranteed to be collective (i.e. it doesn't
-         * modify the structural metadata in a file), check if we should use
-         * a collective metadata read.
+         * modify the structural metadata in a file), check if the application
+         * specified a collective metadata read for just this operation.
          */
         if(!is_collective) {
             H5P_genplist_t *plist;                  /* Property list pointer */
diff --git a/src/H5Cimage.c b/src/H5Cimage.c
index 26b6506..298a2dd 100644
--- a/src/H5Cimage.c
+++ b/src/H5Cimage.c
@@ -1065,9 +1065,9 @@ H5C__read_cache_image(H5F_t *f, H5C_t *cache_ptr)
         H5C__UPDATE_STATS_FOR_CACHE_IMAGE_READ(cache_ptr)
 
 #ifdef H5_HAVE_PARALLEL
-	if ( aux_ptr ) {
+        if ( aux_ptr ) {
 
-	    /* Broadcast cache image */
+	        /* Broadcast cache image */
             if ( MPI_SUCCESS != 
                  (mpi_result = MPI_Bcast(cache_ptr->image_buffer, 
                                          (int)cache_ptr->image_len, MPI_BYTE, 
diff --git a/src/H5Dchunk.c b/src/H5Dchunk.c
index 3fa4a23..13aced6 100644
--- a/src/H5Dchunk.c
+++ b/src/H5Dchunk.c
@@ -2929,9 +2929,6 @@ H5D__chunk_lookup(const H5D_t *dset, const hsize_t *scaled,
         /* Check for cached information */
         if(!H5D__chunk_cinfo_cache_found(&dset->shared->cache.chunk.last, udata)) {
             H5D_chk_idx_info_t idx_info;        /* Chunked index info */
-#ifdef H5_HAVE_PARALLEL
-            H5P_coll_md_read_flag_t temp_cmr;   /* Temp value to hold the coll metadata read setting */
-#endif /* H5_HAVE_PARALLEL */
 
             /* Compose chunked index info struct */
             idx_info.f = dset->oloc.file;
@@ -2940,25 +2937,18 @@ H5D__chunk_lookup(const H5D_t *dset, const hsize_t *scaled,
             idx_info.storage = &dset->shared->layout.storage.u.chunk;
 
 #ifdef H5_HAVE_PARALLEL
-            if(H5F_HAS_FEATURE(idx_info.f, H5FD_FEAT_HAS_MPI)) {
-                /* disable collective metadata read for chunk indexes
-                   as it is highly unlikely that users would read the
-                   same chunks from all processes. MSC - might turn on
-                   for root node? */
-                temp_cmr = H5F_COLL_MD_READ(idx_info.f);
-                H5F_set_coll_md_read(idx_info.f, H5P_FORCE_FALSE);
-            } /* end if */
+            /* Disable collective metadata read for chunk indexes as it is
+             * highly unlikely that users would read the same chunks from all
+             * processes.
+             */
+            if(H5F_HAS_FEATURE(idx_info.f, H5FD_FEAT_HAS_MPI))
+                H5CX_set_coll_metadata_read(FALSE);
 #endif /* H5_HAVE_PARALLEL */
 
             /* Go get the chunk information */
             if((dset->shared->layout.storage.u.chunk.ops->get_addr)(&idx_info, udata) < 0)
                 HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "can't query chunk address")
 
-#ifdef H5_HAVE_PARALLEL
-            if(H5F_HAS_FEATURE(idx_info.f, H5FD_FEAT_HAS_MPI))
-                H5F_set_coll_md_read(idx_info.f, temp_cmr);
-#endif /* H5_HAVE_PARALLEL */
-
             /*
              * Cache the information retrieved.
              *
diff --git a/src/H5Dmpio.c b/src/H5Dmpio.c
index 7352375..6803b60 100644
--- a/src/H5Dmpio.c
+++ b/src/H5Dmpio.c
@@ -800,6 +800,10 @@ H5D__chunk_collective_io(H5D_io_info_t *io_info, const H5D_type_info_t *type_inf
     HDassert(type_info);
     HDassert(fm);
 
+    /* Disable collective metadata reads for chunked dataset I/O operations
+     * in order to prevent potential hangs */
+    H5CX_set_coll_metadata_read(FALSE);
+
     /* Check the optional property list for the collective chunk IO optimization option */
     if(H5CX_get_mpio_chunk_opt_mode(&chunk_opt_mode) < 0)
         HGOTO_ERROR(H5E_DATASET, H5E_CANTGET, FAIL, "couldn't get chunk optimization option")
@@ -2313,7 +2317,7 @@ if(H5DEBUG(D))
 
         /* Broadcasting the MPI_IO option info. and chunk address info. */
         if(MPI_SUCCESS != (mpi_code = MPI_Bcast(total_chunk_addr_array, (int)(sizeof(haddr_t) * fm->layout->u.chunk.nchunks), MPI_BYTE, (int)0, io_info->comm)))
-           HMPI_GOTO_ERROR(FAIL, "MPI_BCast failed", mpi_code)
+            HMPI_GOTO_ERROR(FAIL, "MPI_BCast failed", mpi_code)
     } /* end if */
 
     /* Start at first node in chunk skip list */
diff --git a/src/H5Z.c b/src/H5Z.c
index 14b0400..baea0d1 100644
--- a/src/H5Z.c
+++ b/src/H5Z.c
@@ -603,7 +603,7 @@ H5Z__flush_file_cb(void *obj_ptr, hid_t H5_ATTR_UNUSED obj_id, void *key)
             /* Sanity check for collectively calling H5Zunregister, if requested */
             /* (Sanity check assumes that a barrier on one file's comm
              *  is sufficient (i.e. that there aren't different comms for
-             *  different files).  -QAK, 2018/02/14
+             *  different files).  -QAK, 2018/02/14)
              */
             if(H5_coll_api_sanity_check_g && !object->sanity_checked) {
                 MPI_Comm mpi_comm;      /* File's communicator */