summaryrefslogtreecommitdiffstats
path: root/src/H5Flow.c
diff options
context:
space:
mode:
authorRobert Kim Yates <rkyates@llnl.gov>1998-08-17 15:20:56 (GMT)
committerRobert Kim Yates <rkyates@llnl.gov>1998-08-17 15:20:56 (GMT)
commita1e8ce1d81cbb8b6029c1bb972fcc527cd179f0d (patch)
tree88a5b2a2942c309949011084b019596f7929d0e2 /src/H5Flow.c
parent8ed18c9e73733939216251e1d154b18e7d34bad3 (diff)
downloadhdf5-a1e8ce1d81cbb8b6029c1bb972fcc527cd179f0d.zip
hdf5-a1e8ce1d81cbb8b6029c1bb972fcc527cd179f0d.tar.gz
hdf5-a1e8ce1d81cbb8b6029c1bb972fcc527cd179f0d.tar.bz2
[svn-r599] Added code to convert HDF dataspace selections to MPI datatypes.
Diffstat (limited to 'src/H5Flow.c')
-rw-r--r--src/H5Flow.c53
1 files changed, 46 insertions, 7 deletions
diff --git a/src/H5Flow.c b/src/H5Flow.c
index f929e44..5ddc299 100644
--- a/src/H5Flow.c
+++ b/src/H5Flow.c
@@ -269,6 +269,11 @@ H5F_low_read(H5F_low_t *lf, const H5F_access_t *access_parms,
* June 2, 1998 Albert Cheng
* Added xfer_mode argument
*
+ * rky 980816
+ * Accommodate fancy MPI compound datatype writes.
+ * Also, parallel writes now abort if writing beyond eof
+ * (rather than just printing an error message
+ * and continuing, as they used to).
*-------------------------------------------------------------------------
*/
herr_t
@@ -279,19 +284,40 @@ H5F_low_write(H5F_low_t *lf, const H5F_access_t *access_parms,
herr_t ret_value = FAIL;
haddr_t tmp_addr;
+#ifdef HAVE_PARALLEL
+ int use_types;
+#endif
+
FUNC_ENTER(H5F_low_write, FAIL);
assert(lf && lf->type);
assert(addr && addr_defined(addr));
assert(buf);
- /* Extend the file eof marker if we write past it */
- tmp_addr = *addr;
- H5F_addr_inc(&tmp_addr, (hsize_t)size);
- if (H5F_addr_gt(&tmp_addr, &(lf->eof))) {
- fprintf(stderr, "H5F: extending file w/o allocation\n");
- lf->eof = tmp_addr;
- }
+ /* check for writing past the end of file marker */
+#ifdef HAVE_PARALLEL
+ if (H5F_LOW_MPIO==access_parms->driver
+ && access_parms->u.mpio.use_types) {
+ /* In the case of fancy use of MPI datatypes, the addr and size
+ * parameters have a very peculiar interpretation.
+ * It is logically possible, but quite complex, to calculate
+ * the physical offset that the last byte to be written will have
+ * (assuming the write doesn't fail partway thru, which it may).
+ * Instead, we fix up the eof _after_ the write
+ * (I really don't know if that's OK or not... rky 980816) */
+ use_types = access_parms->u.mpio.use_types; /* check after write */
+ } else {
+#endif /* HAVE_PARALLEL */
+ /* writing a simple block of bytes; can check for writing beyond eof */
+ tmp_addr = *addr;
+ H5F_addr_inc(&tmp_addr, (hsize_t)size);
+ if (H5F_addr_gt(&tmp_addr, &(lf->eof))) {
+ fprintf(stderr, "H5F: extending file w/o allocation\n");
+ lf->eof = tmp_addr;
+ }
+#ifdef HAVE_PARALLEL
+ } /* end else */
+#endif /* HAVE_PARALLEL */
/* Write the data */
if (lf->type->write) {
@@ -303,6 +329,19 @@ H5F_low_write(H5F_low_t *lf, const H5F_access_t *access_parms,
HRETURN_ERROR(H5E_IO, H5E_UNSUPPORTED, FAIL, "no write method");
}
+#ifdef HAVE_PARALLEL
+ /* fix up eof for MPI use_type writes */
+ if (H5F_LOW_MPIO==access_parms->driver && use_types) {
+ /* set logical eof to current physical eof
+ * (ephemeral though it may be...) */
+ MPI_Offset size;
+ if (MPI_SUCCESS != MPI_File_get_size(lf->u.mpio.f,&size)) {
+ HRETURN_ERROR(H5E_IO, H5E_MPI, NULL, "couldn't get file size" );
+ }
+ lf->eof.offset = size;
+ }
+#endif
+
FUNC_LEAVE(ret_value);
}