1
1
openmpi/ompi/mca/pml/ob1/pml_ob1_recvreq.h
bosilca b90c83840f Refactor the request completion (#1422)
* Remodel the request.
Added the wait sync primitive and integrate it into the PML and MTL
infrastructure. The multi-threaded requests are now significantly
less heavy and less noisy (only the threads associated with completed
requests are signaled).

* Fix the condition to release the request.
2016-05-24 18:20:51 -05:00

465 строки
18 KiB
C

/* -*- 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
* Corporation. All rights reserved.
* Copyright (c) 2004-2016 The University of Tennessee and The University
* of Tennessee Research Foundation. All rights
* reserved.
* Copyright (c) 2004-2007 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 (c) 2008 UT-Battelle, LLC. All rights reserved.
* Copyright (c) 2011-2015 Los Alamos National Security, LLC. All rights
* reserved.
* Copyright (c) 2014 Research Organization for Information Science
* and Technology (RIST). All rights reserved.
*
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
/**
* @file
*/
#ifndef OMPI_PML_OB1_RECV_REQUEST_H
#define OMPI_PML_OB1_RECV_REQUEST_H
#include "pml_ob1.h"
#include "pml_ob1_rdma.h"
#include "pml_ob1_rdmafrag.h"
#include "ompi/proc/proc.h"
#include "ompi/mca/pml/ob1/pml_ob1_comm.h"
#include "opal/mca/mpool/base/base.h"
#include "ompi/mca/pml/base/pml_base_recvreq.h"
BEGIN_C_DECLS
struct mca_pml_ob1_recv_request_t {
mca_pml_base_recv_request_t req_recv;
opal_ptr_t remote_req_send;
int32_t req_lock;
size_t req_pipeline_depth;
size_t req_bytes_received; /**< amount of data transferred into the user buffer */
size_t req_bytes_expected; /**< local size of the data as suggested by the user */
size_t req_rdma_offset;
size_t req_send_offset;
uint32_t req_rdma_cnt;
uint32_t req_rdma_idx;
bool req_pending;
bool req_ack_sent; /**< whether ack was sent to the sender */
bool req_match_received; /**< Prevent request to be completed prematurely */
opal_mutex_t lock;
mca_bml_base_btl_t *rdma_bml;
mca_btl_base_registration_handle_t *local_handle;
/** The size of this array is set from mca_pml_ob1.max_rdma_per_request */
mca_pml_ob1_com_btl_t req_rdma[];
};
typedef struct mca_pml_ob1_recv_request_t mca_pml_ob1_recv_request_t;
OBJ_CLASS_DECLARATION(mca_pml_ob1_recv_request_t);
static inline bool lock_recv_request(mca_pml_ob1_recv_request_t *recvreq)
{
return OPAL_THREAD_ADD32(&recvreq->req_lock, 1) == 1;
}
static inline bool unlock_recv_request(mca_pml_ob1_recv_request_t *recvreq)
{
return OPAL_THREAD_ADD32(&recvreq->req_lock, -1) == 0;
}
/**
* 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_OB1_RECV_REQUEST_ALLOC(recvreq) \
do { \
recvreq = (mca_pml_ob1_recv_request_t *) \
opal_free_list_get (&mca_pml_base_recv_requests); \
} 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_OB1_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 the request as completed at MPI level for internal purposes.
*
* @param recvreq (IN) Receive request.
*/
#define MCA_PML_OB1_RECV_REQUEST_MPI_COMPLETE( recvreq ) \
do { \
PERUSE_TRACE_COMM_EVENT( PERUSE_COMM_REQ_COMPLETE, \
&(recvreq->req_recv.req_base), PERUSE_RECV ); \
ompi_request_complete( &(recvreq->req_recv.req_base.req_ompi), true ); \
} while (0)
static inline void mca_pml_ob1_recv_request_fini (mca_pml_ob1_recv_request_t *recvreq)
{
MCA_PML_BASE_RECV_REQUEST_FINI(&recvreq->req_recv);
if ((recvreq)->local_handle) {
mca_bml_base_deregister_mem (recvreq->rdma_bml, recvreq->local_handle);
recvreq->local_handle = NULL;
}
}
/*
* Free the PML receive request
*/
#define MCA_PML_OB1_RECV_REQUEST_RETURN(recvreq) \
{ \
mca_pml_ob1_recv_request_fini (recvreq); \
opal_free_list_return (&mca_pml_base_recv_requests, \
(opal_free_list_item_t*)(recvreq)); \
}
/**
* Complete receive request. Request structure cannot be accessed after calling
* this function any more.
*
* @param recvreq (IN) Receive request.
*/
static inline void
recv_request_pml_complete(mca_pml_ob1_recv_request_t *recvreq)
{
size_t i;
if(false == recvreq->req_recv.req_base.req_pml_complete){
if(recvreq->req_recv.req_bytes_packed > 0) {
PERUSE_TRACE_COMM_EVENT( PERUSE_COMM_REQ_XFER_END,
&recvreq->req_recv.req_base, PERUSE_RECV );
}
for(i = 0; i < recvreq->req_rdma_cnt; i++) {
struct mca_btl_base_registration_handle_t *handle = recvreq->req_rdma[i].btl_reg;
mca_bml_base_btl_t *bml_btl = recvreq->req_rdma[i].bml_btl;
if (NULL != handle) {
mca_bml_base_deregister_mem (bml_btl, handle);
}
}
recvreq->req_rdma_cnt = 0;
if(true == recvreq->req_recv.req_base.req_free_called) {
if( MPI_SUCCESS != recvreq->req_recv.req_base.req_ompi.req_status.MPI_ERROR ) {
ompi_mpi_abort(&ompi_mpi_comm_world.comm, MPI_ERR_REQUEST);
}
MCA_PML_OB1_RECV_REQUEST_RETURN(recvreq);
} else {
/* initialize request status */
recvreq->req_recv.req_base.req_pml_complete = true;
recvreq->req_recv.req_base.req_ompi.req_status._ucount =
recvreq->req_bytes_received;
if (recvreq->req_recv.req_bytes_packed > recvreq->req_bytes_expected) {
recvreq->req_recv.req_base.req_ompi.req_status._ucount =
recvreq->req_recv.req_bytes_packed;
recvreq->req_recv.req_base.req_ompi.req_status.MPI_ERROR =
MPI_ERR_TRUNCATE;
}
if (OPAL_UNLIKELY(recvreq->local_handle)) {
mca_bml_base_deregister_mem (recvreq->rdma_bml, recvreq->local_handle);
recvreq->local_handle = NULL;
}
MCA_PML_OB1_RECV_REQUEST_MPI_COMPLETE(recvreq);
}
}
}
static inline bool
recv_request_pml_complete_check(mca_pml_ob1_recv_request_t *recvreq)
{
#if OPAL_ENABLE_MULTI_THREADS
opal_atomic_rmb();
#endif
if(recvreq->req_match_received &&
recvreq->req_bytes_received >= recvreq->req_recv.req_bytes_packed &&
lock_recv_request(recvreq)) {
recv_request_pml_complete(recvreq);
return true;
}
return false;
}
extern void mca_pml_ob1_recv_req_start(mca_pml_ob1_recv_request_t *req);
#define MCA_PML_OB1_RECV_REQUEST_START(r) mca_pml_ob1_recv_req_start(r)
static inline void prepare_recv_req_converter(mca_pml_ob1_recv_request_t *req)
{
if( req->req_recv.req_base.req_datatype->super.size | req->req_recv.req_base.req_count ) {
opal_convertor_copy_and_prepare_for_recv(
req->req_recv.req_base.req_proc->super.proc_convertor,
&(req->req_recv.req_base.req_datatype->super),
req->req_recv.req_base.req_count,
req->req_recv.req_base.req_addr,
0,
&req->req_recv.req_base.req_convertor);
opal_convertor_get_unpacked_size(&req->req_recv.req_base.req_convertor,
&req->req_bytes_expected);
}
}
#define MCA_PML_OB1_RECV_REQUEST_MATCHED(request, hdr) \
recv_req_matched(request, hdr)
static inline void recv_req_matched(mca_pml_ob1_recv_request_t *req,
mca_pml_ob1_match_hdr_t *hdr)
{
req->req_recv.req_base.req_ompi.req_status.MPI_SOURCE = hdr->hdr_src;
req->req_recv.req_base.req_ompi.req_status.MPI_TAG = hdr->hdr_tag;
req->req_match_received = true;
#if OPAL_ENABLE_MULTI_THREADS
opal_atomic_wmb();
#endif
if(req->req_recv.req_bytes_packed > 0) {
#if OPAL_ENABLE_HETEROGENEOUS_SUPPORT
if(MPI_ANY_SOURCE == req->req_recv.req_base.req_peer) {
/* non wildcard prepared during post recv */
prepare_recv_req_converter(req);
}
#endif /* OPAL_ENABLE_HETEROGENEOUS_SUPPORT */
PERUSE_TRACE_COMM_EVENT(PERUSE_COMM_REQ_XFER_BEGIN,
&req->req_recv.req_base, PERUSE_RECV);
}
}
/**
*
*/
#define MCA_PML_OB1_RECV_REQUEST_UNPACK( request, \
segments, \
num_segments, \
seg_offset, \
data_offset, \
bytes_received, \
bytes_delivered) \
do { \
bytes_delivered = 0; \
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; \
mca_btl_base_segment_t* segment = segments; \
\
for( n = 0; n < num_segments; n++, segment++ ) { \
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); \
iov_count++; \
offset = 0; \
} \
} \
OPAL_THREAD_LOCK(&request->lock); \
PERUSE_TRACE_COMM_OMPI_EVENT (PERUSE_COMM_REQ_XFER_CONTINUE, \
&(request->req_recv.req_base), max_data, \
PERUSE_RECV); \
opal_convertor_set_position( &(request->req_recv.req_base.req_convertor), \
&data_offset ); \
opal_convertor_unpack( &(request)->req_recv.req_base.req_convertor, \
iov, \
&iov_count, \
&max_data ); \
bytes_delivered = max_data; \
OPAL_THREAD_UNLOCK(&request->lock); \
} \
} while (0)
/**
*
*/
void mca_pml_ob1_recv_request_progress_match(
mca_pml_ob1_recv_request_t* req,
struct mca_btl_base_module_t* btl,
mca_btl_base_segment_t* segments,
size_t num_segments);
/**
*
*/
void mca_pml_ob1_recv_request_progress_frag(
mca_pml_ob1_recv_request_t* req,
struct mca_btl_base_module_t* btl,
mca_btl_base_segment_t* segments,
size_t num_segments);
#if OPAL_CUDA_SUPPORT
void mca_pml_ob1_recv_request_frag_copy_start(
mca_pml_ob1_recv_request_t* req,
struct mca_btl_base_module_t* btl,
mca_btl_base_segment_t* segments,
size_t num_segments,
mca_btl_base_descriptor_t* des);
void mca_pml_ob1_recv_request_frag_copy_finished(struct mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* ep,
struct mca_btl_base_descriptor_t* des,
int status );
#endif /* OPAL_CUDA_SUPPORT */
/**
*
*/
void mca_pml_ob1_recv_request_progress_rndv(
mca_pml_ob1_recv_request_t* req,
struct mca_btl_base_module_t* btl,
mca_btl_base_segment_t* segments,
size_t num_segments);
/**
*
*/
void mca_pml_ob1_recv_request_progress_rget(
mca_pml_ob1_recv_request_t* req,
struct mca_btl_base_module_t* btl,
mca_btl_base_segment_t* segments,
size_t num_segments);
/**
*
*/
void mca_pml_ob1_recv_request_matched_probe(
mca_pml_ob1_recv_request_t* req,
struct mca_btl_base_module_t* btl,
mca_btl_base_segment_t* segments,
size_t num_segments);
/**
*
*/
int mca_pml_ob1_recv_request_schedule_once(
mca_pml_ob1_recv_request_t* req, mca_bml_base_btl_t* start_bml_btl);
static inline int mca_pml_ob1_recv_request_schedule_exclusive(
mca_pml_ob1_recv_request_t* req,
mca_bml_base_btl_t* start_bml_btl)
{
int rc;
do {
rc = mca_pml_ob1_recv_request_schedule_once(req, start_bml_btl);
if(rc == OMPI_ERR_OUT_OF_RESOURCE)
break;
} while(!unlock_recv_request(req));
if(OMPI_SUCCESS == rc)
recv_request_pml_complete_check(req);
return rc;
}
static inline void mca_pml_ob1_recv_request_schedule(
mca_pml_ob1_recv_request_t* req,
mca_bml_base_btl_t* start_bml_btl)
{
if(!lock_recv_request(req))
return;
(void)mca_pml_ob1_recv_request_schedule_exclusive(req, start_bml_btl);
}
#define MCA_PML_OB1_ADD_ACK_TO_PENDING(P, S, D, O, Sz) \
do { \
mca_pml_ob1_pckt_pending_t *_pckt; \
\
MCA_PML_OB1_PCKT_PENDING_ALLOC(_pckt); \
_pckt->hdr.hdr_common.hdr_type = MCA_PML_OB1_HDR_TYPE_ACK; \
_pckt->hdr.hdr_ack.hdr_src_req.lval = (S); \
_pckt->hdr.hdr_ack.hdr_dst_req.pval = (D); \
_pckt->hdr.hdr_ack.hdr_send_offset = (O); \
_pckt->hdr.hdr_ack.hdr_send_size = (Sz); \
_pckt->proc = (P); \
_pckt->bml_btl = NULL; \
OPAL_THREAD_LOCK(&mca_pml_ob1.lock); \
opal_list_append(&mca_pml_ob1.pckt_pending, \
(opal_list_item_t*)_pckt); \
OPAL_THREAD_UNLOCK(&mca_pml_ob1.lock); \
} while(0)
int mca_pml_ob1_recv_request_ack_send_btl(ompi_proc_t* proc,
mca_bml_base_btl_t* bml_btl, uint64_t hdr_src_req, void *hdr_dst_req,
uint64_t hdr_rdma_offset, uint64_t size, bool nordma);
static inline int mca_pml_ob1_recv_request_ack_send(ompi_proc_t* proc,
uint64_t hdr_src_req, void *hdr_dst_req, uint64_t hdr_send_offset,
uint64_t size, bool nordma)
{
size_t i;
mca_bml_base_btl_t* bml_btl;
mca_bml_base_endpoint_t* endpoint = mca_bml_base_get_endpoint (proc);
assert (NULL != endpoint);
for(i = 0; i < mca_bml_base_btl_array_get_size(&endpoint->btl_eager); i++) {
bml_btl = mca_bml_base_btl_array_get_next(&endpoint->btl_eager);
if(mca_pml_ob1_recv_request_ack_send_btl(proc, bml_btl, hdr_src_req,
hdr_dst_req, hdr_send_offset, size, nordma) == OMPI_SUCCESS)
return OMPI_SUCCESS;
}
MCA_PML_OB1_ADD_ACK_TO_PENDING(proc, hdr_src_req, hdr_dst_req,
hdr_send_offset, size);
return OMPI_ERR_OUT_OF_RESOURCE;
}
int mca_pml_ob1_recv_request_get_frag(mca_pml_ob1_rdma_frag_t* frag);
/* This function tries to continue recvreq that stuck due to resource
* unavailability. Recvreq is added to recv_pending list if scheduling of put
* operation cannot be accomplished for some reason. */
void mca_pml_ob1_recv_request_process_pending(void);
END_C_DECLS
#endif