1
1
openmpi/ompi/mca/pml/dr/pml_dr_recvreq.h
George Bosilca e361bcb64c Send optimizations.
1. The send path get shorter. The BTL is allowed to return > 0 to specify that the
   descriptor was pushed to the networks, and that the memory attached to it is 
   available again for the upper layer. The MCA_BTL_DES_SEND_ALWAYS_CALLBACK flag
   can be used by the PML to force the BTL to always trigger the callback.
   Unmodified BTL will continue to work as expected, as they will return OMPI_SUCCESS
   which force the PML to have exactly the same behavior as before. Some BTLs have
   been modified: self, sm, tcp, mx.
2. Add send immediate interface to BTL.
   The idea is to have a mechanism of allowing the BTL to take advantage of
   send optimizations such as the ability to deliver data "inline". Some
   network APIs such as Portals allow data to be sent using a "thin" event
   without packing data into a memory descriptor. This interface change
   allows the BTL to use such capabilities and allows for other optimizations
   in the future. All existing BTLs except for Portals and sm have this interface
   set to NULL.

This commit was SVN r18551.
2008-05-30 03:58:39 +00:00

408 строки
20 KiB
C

/*
* Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
* University Research and Technology
* Corporation. All rights reserved.
* Copyright (c) 2004-2007 The University of Tennessee and The University
* of Tennessee Research Foundation. All rights
* reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2006 The Regents of the University of California.
* All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
/**
* @file
*/
#ifndef OMPI_PML_DR_RECV_REQUEST_H
#define OMPI_PML_DR_RECV_REQUEST_H
#include "ompi_config.h"
#include "ompi/mca/mpool/base/base.h"
#include "ompi/mca/pml/base/pml_base_recvreq.h"
#include "pml_dr.h"
#include "pml_dr_hdr.h"
#include "pml_dr_vfrag.h"
#include "pml_dr_comm.h"
#if defined(c_plusplus) || defined(__cplusplus)
extern "C" {
#endif
struct mca_pml_dr_recv_request_t {
mca_pml_base_recv_request_t req_recv;
size_t req_bytes_received;
size_t req_bytes_delivered;
bool req_acked;
/* filled in after match */
struct mca_pml_dr_comm_proc_t* req_proc;
struct mca_pml_dr_endpoint_t* req_endpoint;
opal_mutex_t* req_mutex;
/* vfrag state */
mca_pml_dr_vfrag_t *req_vfrag;
mca_pml_dr_vfrag_t req_vfrag0;
opal_list_t req_vfrags;
};
typedef struct mca_pml_dr_recv_request_t mca_pml_dr_recv_request_t;
OBJ_CLASS_DECLARATION(mca_pml_dr_recv_request_t);
/**
* Allocate a recv request from the modules free list.
*
* @param rc (OUT) OMPI_SUCCESS or error status on failure.
* @return Receive request.
*/
#define MCA_PML_DR_RECV_REQUEST_ALLOC(recvreq, rc) \
do { \
ompi_free_list_item_t* item; \
rc = OMPI_SUCCESS; \
OMPI_FREE_LIST_GET(&mca_pml_base_recv_requests, item, rc); \
recvreq = (mca_pml_dr_recv_request_t*)item; \
} while(0)
/**
* Initialize a receive request with call parameters.
*
* @param request (IN) Receive request.
* @param addr (IN) User buffer.
* @param count (IN) Number of elements of indicated datatype.
* @param datatype (IN) User defined datatype.
* @param src (IN) Source rank w/in the communicator.
* @param tag (IN) User defined tag.
* @param comm (IN) Communicator.
* @param persistent (IN) Is this a ersistent request.
*/
#define MCA_PML_DR_RECV_REQUEST_INIT( \
request, \
addr, \
count, \
datatype, \
src, \
tag, \
comm, \
persistent) \
do { \
MCA_PML_BASE_RECV_REQUEST_INIT( \
&(request)->req_recv, \
addr, \
count, \
datatype, \
src, \
tag, \
comm, \
persistent); \
} while(0)
/**
* Mark a recv request complete.
*
* @param request (IN) Receive request.
*/
#define MCA_PML_DR_RECV_REQUEST_PML_COMPLETE(recvreq) \
do { \
ompi_free_list_item_t* item; \
assert( false == recvreq->req_recv.req_base.req_pml_complete ); \
OPAL_THREAD_LOCK((recvreq)->req_mutex); \
while(NULL != (item = (ompi_free_list_item_t*) \
opal_list_remove_first(&(recvreq)->req_vfrags))) { \
OMPI_FREE_LIST_RETURN(&mca_pml_dr.vfrags, item); \
} \
OPAL_THREAD_UNLOCK((recvreq)->req_mutex); \
\
opal_list_remove_item(&(recvreq)->req_proc->matched_receives, \
(opal_list_item_t*)(recvreq)); \
\
/* initialize request status */ \
recvreq->req_recv.req_base.req_pml_complete = true; \
if (recvreq->req_bytes_received > recvreq->req_bytes_delivered) { \
recvreq->req_recv.req_base.req_ompi.req_status._count = \
recvreq->req_bytes_delivered; \
recvreq->req_recv.req_base.req_ompi.req_status.MPI_ERROR = \
MPI_ERR_TRUNCATE; \
} else { \
recvreq->req_recv.req_base.req_ompi.req_status._count = \
recvreq->req_bytes_received; \
} \
ompi_request_complete( &(recvreq->req_recv.req_base.req_ompi), true ); \
\
if( true == recvreq->req_recv.req_base.req_free_called ) { \
MCA_PML_DR_RECV_REQUEST_RETURN( recvreq ); \
} \
} while(0)
/**
* Return a recv request to the modules free list.
*
* @param request (IN) Receive request.
*/
#define MCA_PML_DR_RECV_REQUEST_RETURN(recvreq) \
do { \
/* decrement reference counts */ \
MCA_PML_BASE_RECV_REQUEST_FINI(&(recvreq)->req_recv); \
OMPI_FREE_LIST_RETURN(&mca_pml_base_recv_requests, \
(ompi_free_list_item_t*)(recvreq)); \
} while(0)
/**
* Attempt to match the request against the unexpected fragment list
* for all source ranks w/in the communicator.
*
* @param request (IN) Request to match.
*/
void mca_pml_dr_recv_request_match_wild(mca_pml_dr_recv_request_t* request);
/**
* Attempt to match the request against the unexpected fragment list
* for a specific source rank.
*
* @param request (IN) Request to match.
*/
void mca_pml_dr_recv_request_match_specific(mca_pml_dr_recv_request_t* request);
/**
* Ack a matched request.
*/
void mca_pml_dr_recv_request_ack(
mca_btl_base_module_t* blt,
mca_pml_dr_recv_request_t* recvreq,
mca_pml_dr_common_hdr_t* hdr,
ompi_ptr_t src_ptr,
size_t vlen,
uint64_t vmask);
/**
* Start an initialized request.
*
* @param request Receive request.
* @return OMPI_SUCESS or error status on failure.
*/
#define MCA_PML_DR_RECV_REQUEST_START(request) \
do { \
/* init/re-init the request */ \
(request)->req_bytes_received = 0; \
(request)->req_bytes_delivered = 0; \
(request)->req_acked = false; \
(request)->req_recv.req_base.req_pml_complete = false; \
(request)->req_recv.req_base.req_ompi.req_complete = false; \
(request)->req_recv.req_base.req_ompi.req_state = OMPI_REQUEST_ACTIVE; \
(request)->req_vfrag = &(request)->req_vfrag0; \
(request)->req_proc = NULL; \
(request)->req_endpoint = NULL; \
\
/* always set the req_status.MPI_TAG to ANY_TAG before starting the \
* request. This field is used if cancelled to find out if the request \
* has been matched or not. \
*/ \
(request)->req_recv.req_base.req_ompi.req_status.MPI_TAG = OMPI_ANY_TAG; \
(request)->req_recv.req_base.req_ompi.req_status.MPI_ERROR = OMPI_SUCCESS; \
(request)->req_recv.req_base.req_ompi.req_status._cancelled = 0; \
\
/* attempt to match unexpected recv */ \
if((request)->req_recv.req_base.req_peer == OMPI_ANY_SOURCE) { \
mca_pml_dr_recv_request_match_wild(request); \
} else { \
mca_pml_dr_recv_request_match_specific(request); \
} \
} while (0)
/**
* Initialize request when match is made
*/
#define MCA_PML_DR_RECV_REQUEST_MATCHED(request,comm,proc,hdr) \
do { \
(request)->req_mutex = &comm->matching_lock; \
(request)->req_proc = proc; \
(request)->req_endpoint = (mca_pml_dr_endpoint_t*)proc->ompi_proc->proc_pml; \
(request)->req_recv.req_base.req_ompi.req_status.MPI_TAG = (hdr)->hdr_tag; \
(request)->req_recv.req_base.req_ompi.req_status.MPI_SOURCE = proc->comm_rank; \
(request)->req_vfrag0.vf_id = (hdr)->hdr_common.hdr_vid; \
opal_list_append(&proc->matched_receives, (opal_list_item_t*)request); \
ompi_seq_tracker_insert(&request->req_endpoint->seq_recvs_matched, \
(hdr)->hdr_common.hdr_vid); \
} while(0)
/**
* Setup convertor if message length is non-zero
*/
#define MCA_PML_DR_RECV_REQUEST_BYTES_PACKED(request, bytes_packed) \
do { \
bool do_csum = mca_pml_dr.enable_csum && \
(request->req_endpoint->bml_endpoint->btl_flags_or & MCA_BTL_FLAGS_NEED_CSUM); \
(request)->req_recv.req_bytes_packed = bytes_packed; \
if((request)->req_recv.req_bytes_packed != 0) { \
ompi_proc_t *proc = (request)->req_proc->ompi_proc; \
ompi_convertor_copy_and_prepare_for_recv( proc->proc_convertor, \
(request)->req_recv.req_base.req_datatype, \
(request)->req_recv.req_base.req_count, \
(request)->req_recv.req_base.req_addr, \
(do_csum ? CONVERTOR_WITH_CHECKSUM: 0), \
&(request)->req_recv.req_base.req_convertor ); \
} \
} while (0)
/**
*
*/
#define MCA_PML_DR_RECV_REQUEST_UNPACK( \
request, \
segments, \
num_segments, \
seg_offset, \
data_offset, \
bytes_received, \
bytes_delivered, \
csum) \
do { \
if(request->req_recv.req_bytes_packed > 0) { \
struct iovec iov[MCA_BTL_DES_MAX_SEGMENTS]; \
uint32_t iov_count = 0; \
size_t max_data = bytes_received; \
size_t n, offset = seg_offset; \
bool do_csum = mca_pml_dr.enable_csum && \
(request->req_endpoint->bml_endpoint->btl_flags_or & MCA_BTL_FLAGS_NEED_CSUM); \
\
for(n=0; n<num_segments; n++) { \
mca_btl_base_segment_t* segment = segments+n; \
if(offset >= segment->seg_len) { \
offset -= segment->seg_len; \
} else { \
iov[iov_count].iov_len = segment->seg_len - offset; \
iov[iov_count].iov_base = (IOVBASE_TYPE*)((unsigned char*)segment->seg_addr.pval + offset); \
offset = 0; \
iov_count++; \
} \
} \
ompi_convertor_set_position( &(request->req_recv.req_base.req_convertor), \
&data_offset); \
assert((request->req_recv.req_base.req_convertor.flags & CONVERTOR_COMPLETED) == 0); \
ompi_convertor_unpack( &(request)->req_recv.req_base.req_convertor, \
iov, \
&iov_count, \
&max_data); \
bytes_delivered = max_data; \
if(bytes_received && !bytes_delivered) assert(0); \
csum = (do_csum ? \
request->req_recv.req_base.req_convertor.checksum : OPAL_CSUM_ZERO); \
} else { \
bytes_delivered = 0; \
csum = OPAL_CSUM_ZERO; \
} \
} while (0)
/**
*
*/
void mca_pml_dr_recv_request_progress(
mca_pml_dr_recv_request_t* req,
struct mca_btl_base_module_t* btl,
mca_btl_base_segment_t* segments,
size_t num_segments);
/**
*
*/
void mca_pml_dr_recv_request_matched_probe(
mca_pml_dr_recv_request_t* req,
struct mca_btl_base_module_t* btl,
mca_btl_base_segment_t* segments,
size_t num_segments);
/**
*
*/
void mca_pml_dr_recv_request_schedule(
mca_pml_dr_recv_request_t* req);
/**
* Look for matched receive.
* Must be called w/ matching lock held.
*/
static inline struct mca_pml_dr_recv_request_t* mca_pml_dr_comm_proc_check_matched(
mca_pml_dr_comm_proc_t* dr_proc,
uint32_t vfrag_id)
{
opal_list_item_t* item;
for(item = opal_list_get_first(&dr_proc->matched_receives);
item != opal_list_get_end(&dr_proc->matched_receives);
item = opal_list_get_next(item)) {
struct mca_pml_dr_recv_request_t* recvreq = (struct mca_pml_dr_recv_request_t*)item;
if(recvreq->req_vfrag0.vf_id == vfrag_id)
return recvreq;
}
return NULL;
}
/*
*
*/
#define MCA_PML_DR_RECV_REQUEST_VFRAG_LOOKUP(recvreq,hdr,vfrag) \
do { \
if((recvreq)->req_vfrag->vf_id == (hdr)->hdr_common.hdr_vid) { \
vfrag = (recvreq)->req_vfrag; \
} else { \
opal_list_item_t* item; \
int rc; \
\
vfrag = NULL; \
OPAL_THREAD_LOCK(recvreq->req_mutex); \
for(item = opal_list_get_first(&(recvreq)->req_vfrags); \
item != opal_list_get_end(&(recvreq)->req_vfrags); \
item = opal_list_get_next(item)) { \
mca_pml_dr_vfrag_t* vf = (mca_pml_dr_vfrag_t*)item; \
if(vf->vf_id == (hdr)->hdr_common.hdr_vid) { \
vfrag = vf; \
break; \
} \
} \
if(NULL == vfrag) { \
MCA_PML_DR_VFRAG_ALLOC(vfrag,rc); \
if(NULL != vfrag) { \
MCA_PML_DR_VFRAG_INIT(vfrag); \
(vfrag)->vf_id = (hdr)->hdr_common.hdr_vid; \
(vfrag)->vf_len = (hdr)->hdr_vlen; \
if((hdr)->hdr_vlen == 64) { \
(vfrag)->vf_mask = ~(uint64_t)0; \
} else { \
(vfrag)->vf_mask = (((uint64_t)1 << (hdr)->hdr_vlen)-1); \
} \
opal_list_append(&(recvreq)->req_vfrags, (opal_list_item_t*)vfrag); \
(recvreq)->req_vfrag = vfrag; \
} \
} \
OPAL_THREAD_UNLOCK(recvreq->req_mutex); \
} \
} while(0)
#if defined(c_plusplus) || defined(__cplusplus)
}
#endif
#endif