[svn-r5814] Purpose:

    Bug Fix

Description:
    It was possible to create corrupted metadata information (either in memory
    or in the file or both) with a parallel I/O program because of the way
    metadata writes were being handled for writes out of the metadata cache.

Solution:
    Added a dataset transfer property called "block before metadata write"
    which is used by the MPI-I/O and MPI-posix drivers to sync up all the
    processes before attempting a metadata write.  This property is currently
    only for metadata writes from the metadata cache.

Platforms tested:
    IRIX64 6.5 (modi4) w/parallel

1 files changed, 37 insertions, 19 deletions

diff --git a/src/H5FDmpiposix.c b/src/H5FDmpiposix.c
index 82533d0..32e5535 100644
--- a/src/H5FDmpiposix.c
+++ b/src/H5FDmpiposix.c
@@ -32,6 +32,7 @@
  *
  */
 #include "H5private.h"		/*library functions			*/
+#include "H5ACprivate.h"        /* Metadata cache */
 #include "H5Eprivate.h"		/*error handling			*/
 #include "H5Fprivate.h"		/*files					*/
 #include "H5FDprivate.h"	/*file driver			        */
@@ -965,21 +966,31 @@ done:
  *
  * Modifications:
  *
+ *              Quincey Koziol - 2002/07/18
+ *              Added "block_before_meta_write" dataset transfer flag, which
+ *              is set during writes from a metadata cache flush and indicates
+ *              that all the processes must sync up before (one of them)
+ *              writing metadata.
+ *
  *-------------------------------------------------------------------------
  */
 static herr_t
-H5FD_mpiposix_write(H5FD_t *_file, H5FD_mem_t type, hid_t UNUSED dxpl_id, haddr_t addr,
+H5FD_mpiposix_write(H5FD_t *_file, H5FD_mem_t type, hid_t dxpl_id, haddr_t addr,
 		size_t size, const void *buf)
 {
     H5FD_mpiposix_t	*file = (H5FD_mpiposix_t*)_file;
     int			mpi_code;	/* MPI return code */
     ssize_t	        nbytes;         /* Number of bytes written each I/O call */
+    H5P_genplist_t      *plist;         /* Property list pointer */
+    unsigned		block_before_meta_write=0;      /* Whether to block before a metadata write */
     herr_t             	ret_value=SUCCEED;      /* Return value */
 
     FUNC_ENTER_NOAPI(H5FD_mpiposix_write, FAIL);
 
     assert(file);
     assert(H5FD_MPIPOSIX==file->pub.driver_id);
+    assert(H5I_GENPROP_LST==H5I_get_type(dxpl_id));
+    assert(TRUE==H5P_isa_class(dxpl_id,H5P_DATASET_XFER));
     assert(buf);
 
     /* Check for overflow conditions */
@@ -990,17 +1001,36 @@ H5FD_mpiposix_write(H5FD_t *_file, H5FD_mem_t type, hid_t UNUSED dxpl_id, haddr_
     if (addr+size>file->eoa)
         HGOTO_ERROR(H5E_ARGS, H5E_OVERFLOW, FAIL, "addr overflow");
     
-    /* Only p<round> will do the actual write if all procs in comm write same data */
-    if ((type!=H5FD_MEM_DRAW) && H5_mpiposix_1_metawrite_g) {
-        if (file->mpi_rank != file->mpi_round)
-            HGOTO_DONE(SUCCEED) /* skip the actual write */
+    /* Obtain the data transfer properties */
+    if(NULL == (plist = H5I_object(dxpl_id)))
+        HGOTO_ERROR(H5E_ARGS, H5E_BADTYPE, FAIL, "not a file access property list");
+
+    /* Metadata specific actions */
+    if(type!=H5FD_MEM_DRAW) {
+        /* Check if we need to syncronize all processes before attempting metadata write
+         * (Prevents race condition where the process writing the metadata goes ahead
+         * and writes the metadata to the file before all the processes have
+         * read the data, "transmitting" data from the "future" to the reading
+         * process. -QAK )
+         */
+        if(H5P_exist_plist(plist,H5AC_BLOCK_BEFORE_META_WRITE_NAME)>0)
+            if(H5P_get(plist,H5AC_BLOCK_BEFORE_META_WRITE_NAME,&block_before_meta_write)<0)
+                HGOTO_ERROR(H5E_PLIST, H5E_CANTGET, FAIL, "can't get H5AC property");
+
+        if(block_before_meta_write)
+            if (MPI_SUCCESS!= (mpi_code=MPI_Barrier(file->comm)))
+                HMPI_GOTO_ERROR(FAIL, "MPI_Barrier failed", mpi_code);
+
+        /* Only p<round> will do the actual write if all procs in comm write same metadata */
+        if (H5_mpiposix_1_metawrite_g)
+            if (file->mpi_rank != file->mpi_round)
+                HGOTO_DONE(SUCCEED) /* skip the actual write */
     } /* end if */
 
     /* Seek to the correct location */
     if ((addr!=file->pos || OP_WRITE!=file->op) &&
             file_seek(file->fd, (file_offset_t)addr, SEEK_SET)<0)
-        HGOTO_ERROR(H5E_IO, H5E_SEEKERROR, FAIL,
-		      "unable to seek to proper position");
+        HGOTO_ERROR(H5E_IO, H5E_SEEKERROR, FAIL, "unable to seek to proper position");
 
     /*
      * Write the data, being careful of interrupted system calls and partial
@@ -1039,18 +1069,6 @@ done:
 
             /* Round-robin rotate to the next process */
             file->mpi_round = (++file->mpi_round)%file->mpi_size;
-#ifdef QAK
-    {
-        int max,min;
-
-        MPI_Allreduce(&file->mpi_round, &max, 1, MPI_INT, MPI_MAX, file->comm);
-        MPI_Allreduce(&file->mpi_round, &min, 1, MPI_INT, MPI_MIN, file->comm);
-        if(max!=file->mpi_round)
-            printf("%s: rank=%d, round=%d, max=%d\n",FUNC,file->mpi_rank,file->mpi_round,max);
-        if(min!=file->mpi_round)
-            printf("%s: rank=%d, round=%d, min=%d\n",FUNC,file->mpi_rank,file->mpi_round,min);
-    }
-#endif /* QAK */
         } /* end if */
     } /* end else */