What: Allocate free list payload even if a payload size is specified even if no
mpool is specified.
When: Thursday, Feb 23, 2012
Why: The current behavior is to ignore the payload size if no mpool is
specified. I see no reason why a payload buffer should't be allocated in the no
mpool case. Thoughts?
Patch is attached.
-Nathan Hjelm
HPC-3, LANL
diff --git a/ompi/class/ompi_free_list.c b/ompi/class/ompi_free_list.c
index d468a70..e3c0988 100644
--- a/ompi/class/ompi_free_list.c
+++ b/ompi/class/ompi_free_list.c
@@ -1,4 +1,4 @@
-/* -*- Mode: C; c-basic-offset:4 ; -*- */
+/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
* University Research and Technology
@@ -13,6 +13,8 @@
* Copyright (c) 2006-2007 Mellanox Technologies. All rights reserved.
* Copyright (c) 2010 Cisco Systems, Inc. All rights reserved.
* Copyright (c) 2011 NVIDIA Corporation. All rights reserved.
+ * Copyright (c) 2012 Los Alamos National Security, LLC. All rights
+ * reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
@@ -70,23 +72,19 @@ static void ompi_free_list_destruct(ompi_free_list_t* fl)
}
#endif
- if( NULL != fl->fl_mpool ) {
- while(NULL != (item = opal_list_remove_first(&(fl->fl_allocations)))) {
- fl_mem = (ompi_free_list_memory_t*)item;
+ while(NULL != (item = opal_list_remove_first(&(fl->fl_allocations)))) {
+ fl_mem = (ompi_free_list_memory_t*)item;
+ if( NULL != fl->fl_mpool ) {
fl->fl_mpool->mpool_free(fl->fl_mpool, fl_mem->ptr,
fl_mem->registration);
-
- /* destruct the item (we constructed it), then free the memory
chunk */
- OBJ_DESTRUCT(item);
- free(item);
- }
- } else {
- while(NULL != (item = opal_list_remove_first(&(fl->fl_allocations)))) {
- /* destruct the item (we constructed it), then free the memory
chunk */
- OBJ_DESTRUCT(item);
- free(item);
+ } else if (fl_mem->ptr) {
+ free (fl_mem->ptr);
}
+
+ /* destruct the item (we constructed it), then free the memory chunk */
+ OBJ_DESTRUCT(item);
+ free(item);
}
OBJ_DESTRUCT(&fl->fl_allocations);
@@ -171,7 +169,7 @@ int ompi_free_list_init_ex_new(
}
int ompi_free_list_grow(ompi_free_list_t* flist, size_t num_elements)
{
- unsigned char *ptr, *mpool_alloc_ptr = NULL;
+ unsigned char *ptr, *mpool_alloc_ptr = NULL, *payload_ptr;
ompi_free_list_memory_t *alloc_ptr;
size_t i, alloc_size, head_size, elem_size = 0;
mca_mpool_base_registration_t *reg = NULL;
@@ -201,7 +199,7 @@ int ompi_free_list_grow(ompi_free_list_t* flist, size_t
num_elements)
elem_size = OPAL_ALIGN(flist->fl_payload_buffer_size,
flist->fl_payload_buffer_alignment, size_t);
if(elem_size != 0) {
- mpool_alloc_ptr = (unsigned char *)
flist->fl_mpool->mpool_alloc(flist->fl_mpool,
+ payload_ptr = mpool_alloc_ptr = (unsigned char *)
flist->fl_mpool->mpool_alloc(flist->fl_mpool,
num_elements * elem_size,
flist->fl_payload_buffer_alignment,
MCA_MPOOL_FLAGS_CACHE_BYPASS |
MCA_MPOOL_FLAGS_CUDA_REGISTER_MEM, ®);
if(NULL == mpool_alloc_ptr) {
@@ -209,6 +207,26 @@ int ompi_free_list_grow(ompi_free_list_t* flist, size_t
num_elements)
return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
}
}
+ } else if (0 != flist->fl_payload_buffer_size) {
+ elem_size = OPAL_ALIGN(flist->fl_payload_buffer_size,
+ flist->fl_payload_buffer_alignment, size_t);
+ if(elem_size != 0) {
+#ifdef HAVE_POSIX_MEMALIGN
+ posix_memalign ((void **) &mpool_alloc_ptr,
flist->fl_payload_buffer_alignment,
+ num_elements * elem_size);
+ payload_ptr = mpool_alloc_ptr;
+#else
+ mpool_alloc_ptr = malloc (num_elements * elem_size +
+ flist->fl_payload_buffer_alignment);
+ payload_ptr = (void*)OPAL_ALIGN((uintptr_t)mpool_alloc_ptr,
+ flist->fl_payload_buffer_alignment,
+ uintptr_t);
+#endif
+ if(NULL == mpool_alloc_ptr) {
+ free(alloc_ptr);
+ return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
+ }
+ }
}
/* make the alloc_ptr a list item, save the chunk in the allocations list,
@@ -225,7 +243,7 @@ int ompi_free_list_grow(ompi_free_list_t* flist, size_t
num_elements)
for(i=0; i<num_elements; i++) {
ompi_free_list_item_t* item = (ompi_free_list_item_t*)ptr;
item->registration = reg;
- item->ptr = mpool_alloc_ptr;
+ item->ptr = payload_ptr;
OBJ_CONSTRUCT_INTERNAL(item, flist->fl_frag_class);
@@ -236,7 +254,7 @@ int ompi_free_list_grow(ompi_free_list_t* flist, size_t
num_elements)
opal_atomic_lifo_push(&(flist->super), &(item->super));
ptr += head_size;
- mpool_alloc_ptr += elem_size;
+ payload_ptr += elem_size;
}
flist->fl_num_allocated += num_elements;
diff --git a/ompi/mca/btl/ugni/btl_ugni_frag.c
b/ompi/mca/btl/ugni/btl_ugni_frag.c
index de7857b..16f3612 100644
--- a/ompi/mca/btl/ugni/btl_ugni_frag.c
+++ b/ompi/mca/btl/ugni/btl_ugni_frag.c
@@ -17,14 +17,11 @@ static inline void mca_btl_ugni_smsg_frag_constructor
(mca_btl_ugni_base_frag_t
{
/* send memory does not need to be registered so we do not need a mpool */
memset ((char *) frag + sizeof (frag->base), 0, sizeof (*frag) - sizeof
(frag->base));
- frag->buffer = frag->segments[0].seg_addr.pval = calloc (1,
mca_btl_ugni_component.smsg_max_data);
+ frag->segments[0].seg_addr.pval = frag->base.super.ptr;
}
static inline void mca_btl_ugni_frag_destructor (mca_btl_ugni_base_frag_t
*frag)
{
- if (NULL != frag->buffer) {
- free (frag->buffer);
- }
}
static inline void mca_btl_ugni_rdma_frag_constructor
(mca_btl_ugni_base_frag_t *frag)
diff --git a/ompi/mca/btl/ugni/btl_ugni_frag.h
b/ompi/mca/btl/ugni/btl_ugni_frag.h
index b115963..d01ff17 100644
--- a/ompi/mca/btl/ugni/btl_ugni_frag.h
+++ b/ompi/mca/btl/ugni/btl_ugni_frag.h
@@ -48,7 +48,6 @@ typedef struct mca_btl_ugni_base_frag_t {
mca_btl_base_endpoint_t *endpoint;
mca_btl_ugni_reg_t *registration;
ompi_free_list_t *my_list;
- void *buffer;
uint32_t msg_id;
} mca_btl_ugni_base_frag_t;
