112 строки
3.9 KiB
C
112 строки
3.9 KiB
C
/* -*- Mode: C; c-basic-offset:4 ; -*- */
|
|
|
|
/*
|
|
* Copyright (c) 2004-2005 The Trustees of Indiana University.
|
|
* All rights reserved.
|
|
* Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
|
|
* All rights reserved.
|
|
* Copyright (c) 2004 The Ohio State University.
|
|
* All rights reserved.
|
|
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
|
|
* University of Stuttgart. All rights reserved.
|
|
* Copyright (c) 2004-2005 The Regents of the University of California.
|
|
* All rights reserved.
|
|
* $COPYRIGHT$
|
|
*
|
|
* Additional copyrights may follow
|
|
*
|
|
* $HEADER$
|
|
*/
|
|
#include "ompi_config.h"
|
|
#include "include/types.h"
|
|
#include "datatype/datatype.h"
|
|
#include "ptl_gm.h"
|
|
#include "ptl_gm_sendfrag.h"
|
|
#include "ptl_gm_priv.h"
|
|
|
|
/*
|
|
* send fragment constructor/destructors.
|
|
*/
|
|
|
|
OBJ_CLASS_INSTANCE(mca_ptl_gm_send_frag_t,
|
|
mca_ptl_base_send_frag_t,
|
|
NULL, NULL);
|
|
|
|
/* It's not yet clear for me what's the best solution here. Block
|
|
* until we get a free request or allocate a new one. The fist case
|
|
* allow us to never take care of the gm allocated DMA buffer as all
|
|
* send fragments already have one attached, but it can stop the
|
|
* application progression. The second case require special cases: we
|
|
* should set the data in the header inside the fragment and later
|
|
* when we get some free fragments with DMA memory attached we should
|
|
* put the header back there, and send it.
|
|
*
|
|
* I will implement the first case and add the second one in my TODO
|
|
* list.
|
|
*/
|
|
mca_ptl_gm_send_frag_t*
|
|
mca_ptl_gm_alloc_send_frag( struct mca_ptl_gm_module_t* ptl,
|
|
struct mca_ptl_base_send_request_t* sendreq )
|
|
{
|
|
opal_list_item_t* item;
|
|
mca_ptl_gm_send_frag_t* frag;
|
|
int32_t rc;
|
|
|
|
/* first get a gm_send_frag */
|
|
OMPI_FREE_LIST_GET( &(ptl->gm_send_frags), item, rc );
|
|
frag = (mca_ptl_gm_send_frag_t*)item;
|
|
/* And then get some DMA memory to put the data */
|
|
OMPI_FREE_LIST_WAIT( &(ptl->gm_send_dma_frags), item, rc );
|
|
opal_atomic_sub( &(ptl->num_send_tokens), 1 );
|
|
assert( ptl->num_send_tokens >= 0 );
|
|
frag->send_buf = (void*)item;
|
|
|
|
frag->frag_send.frag_request = sendreq;
|
|
frag->frag_send.frag_base.frag_owner = (struct mca_ptl_base_module_t*)ptl;
|
|
frag->frag_send.frag_base.frag_addr = sendreq->req_send.req_addr;
|
|
frag->frag_bytes_processed = 0;
|
|
frag->frag_bytes_validated = 0;
|
|
frag->status = -1;
|
|
frag->type = PUT;
|
|
ompi_ptl_gm_init_pipeline( &(frag->pipeline) );
|
|
|
|
return frag;
|
|
}
|
|
|
|
int mca_ptl_gm_put_frag_init( struct mca_ptl_gm_send_frag_t** putfrag,
|
|
struct mca_ptl_gm_peer_t* ptl_peer,
|
|
struct mca_ptl_gm_module_t* gm_ptl,
|
|
struct mca_ptl_base_send_request_t* sendreq,
|
|
size_t offset, size_t* size, int flags )
|
|
{
|
|
ompi_convertor_t* convertor;
|
|
mca_ptl_gm_send_frag_t* frag;
|
|
|
|
frag = mca_ptl_gm_alloc_send_frag( gm_ptl, sendreq ); /*alloc_put_frag */
|
|
|
|
frag->frag_send.frag_base.frag_peer = (struct mca_ptl_base_peer_t*)ptl_peer;
|
|
frag->frag_send.frag_base.frag_size = *size;
|
|
frag->frag_offset = offset;
|
|
|
|
if( (*size) > 0 ) {
|
|
convertor = &(frag->frag_send.frag_base.frag_convertor);
|
|
/* GM use the default parameters for the convertor without any special memory
|
|
* allocation function. We have to call the prepare_for_send in order to
|
|
* initialize the missing parameters of the convertor.
|
|
*/
|
|
ompi_convertor_clone_with_position( &(sendreq->req_send.req_convertor), convertor, 1,
|
|
&offset );
|
|
}
|
|
*putfrag = frag;
|
|
|
|
return OMPI_SUCCESS;
|
|
}
|
|
|
|
/*
|
|
* recv fragment constructor/destructors.
|
|
*/
|
|
|
|
OBJ_CLASS_INSTANCE(mca_ptl_gm_recv_frag_t,
|
|
mca_ptl_base_recv_frag_t,
|
|
NULL, NULL);
|