int mca_pml_ob1_irecv(void *addr,
size_t count,
ompi_datatype_t * datatype,
int src,
int tag,
struct ompi_communicator_t *comm,
struct ompi_request_t **request)
{
mca_pml_ob1_recv_request_t *recvreq;
MCA_PML_OB1_RECV_REQUEST_ALLOC(recvreq);
if (NULL == recvreq)
return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
recvreq->req_recv.req_base.req_type = MCA_PML_REQUEST_RECV;
MCA_PML_OB1_RECV_REQUEST_INIT(recvreq,
addr,
count, datatype, src, tag, comm, false);
PERUSE_TRACE_COMM_EVENT (PERUSE_COMM_REQ_ACTIVATE,
&((recvreq)->req_recv.req_base),
PERUSE_RECV);
MCA_PML_OB1_RECV_REQUEST_START(recvreq);
*request = (ompi_request_t *) recvreq;
return OMPI_SUCCESS;
}
int mca_pml_ob1_recv(void *addr,
size_t count,
ompi_datatype_t * datatype,
int src,
int tag,
struct ompi_communicator_t *comm,
ompi_status_public_t * status)
{
mca_pml_ob1_recv_request_t recvreq;
int rc;
OBJ_CONSTRUCT(&recvreq, mca_pml_ob1_recv_request_t);
MCA_PML_OB1_RECV_REQUEST_INIT(&recvreq, addr, count, datatype,
src, tag, comm, false);
PERUSE_TRACE_COMM_EVENT (PERUSE_COMM_REQ_ACTIVATE,
&(recvreq.req_recv.req_base),
PERUSE_RECV);
MCA_PML_OB1_RECV_REQUEST_START(&recvreq);
ompi_request_wait_completion(&recvreq.req_recv.req_base.req_ompi);
if (NULL != status) { /* return status */
*status = recvreq.req_recv.req_base.req_ompi.req_status;
}
rc = recvreq.req_recv.req_base.req_ompi.req_status.MPI_ERROR;
MCA_PML_BASE_RECV_REQUEST_FINI(&recvreq.req_recv);
OBJ_DESTRUCT(&recvreq);
return rc;
}
int mca_pml_ob1_isend_init(const void *buf,
size_t count,
ompi_datatype_t * datatype,
int dst,
int tag,
mca_pml_base_send_mode_t sendmode,
ompi_communicator_t * comm,
ompi_request_t ** request)
{
mca_pml_ob1_send_request_t *sendreq = NULL;
MCA_PML_OB1_SEND_REQUEST_ALLOC(comm, dst, sendreq);
if (NULL == sendreq)
return OMPI_ERR_OUT_OF_RESOURCE;
MCA_PML_OB1_SEND_REQUEST_INIT(sendreq,
buf,
count,
datatype,
dst, tag,
comm, sendmode, true);
PERUSE_TRACE_COMM_EVENT (PERUSE_COMM_REQ_ACTIVATE,
&(sendreq)->req_send.req_base,
PERUSE_SEND);
*request = (ompi_request_t *) sendreq;
return OMPI_SUCCESS;
}
int mca_pml_ob1_isend_init(const void *buf,
size_t count,
ompi_datatype_t * datatype,
int dst,
int tag,
mca_pml_base_send_mode_t sendmode,
ompi_communicator_t * comm,
ompi_request_t ** request)
{
mca_pml_ob1_send_request_t *sendreq = NULL;
MCA_PML_OB1_SEND_REQUEST_ALLOC(comm, dst, sendreq);
if (NULL == sendreq)
return OMPI_ERR_OUT_OF_RESOURCE;
MCA_PML_OB1_SEND_REQUEST_INIT(sendreq,
buf,
count,
datatype,
dst, tag,
comm, sendmode, true);
PERUSE_TRACE_COMM_EVENT (PERUSE_COMM_REQ_ACTIVATE,
&(sendreq)->req_send.req_base,
PERUSE_SEND);
/* Work around a leak in start by marking this request as complete. The
* problem occured because we do not have a way to differentiate an
* inital request and an incomplete pml request in start. This line
* allows us to detect this state. */
sendreq->req_send.req_base.req_pml_complete = true;
*request = (ompi_request_t *) sendreq;
return OMPI_SUCCESS;
}
int mca_pml_csum_irecv(void *addr,
size_t count,
ompi_datatype_t * datatype,
int src,
int tag,
struct ompi_communicator_t *comm,
struct ompi_request_t **request)
{
int rc;
mca_pml_csum_recv_request_t *recvreq;
MCA_PML_CSUM_RECV_REQUEST_ALLOC(recvreq, rc);
if (NULL == recvreq)
return rc;
MCA_PML_CSUM_RECV_REQUEST_INIT(recvreq,
addr,
count, datatype, src, tag, comm, false);
PERUSE_TRACE_COMM_EVENT (PERUSE_COMM_REQ_ACTIVATE,
&((recvreq)->req_recv.req_base),
PERUSE_RECV);
MCA_PML_CSUM_RECV_REQUEST_START(recvreq);
*request = (ompi_request_t *) recvreq;
return OMPI_SUCCESS;
}
int mca_pml_ob1_irecv_init(void *addr,
size_t count,
ompi_datatype_t * datatype,
int src,
int tag,
struct ompi_communicator_t *comm,
struct ompi_request_t **request)
{
mca_pml_ob1_recv_request_t *recvreq;
MCA_PML_OB1_RECV_REQUEST_ALLOC(recvreq);
if (NULL == recvreq)
return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
recvreq->req_recv.req_base.req_type = MCA_PML_REQUEST_RECV;
MCA_PML_OB1_RECV_REQUEST_INIT(recvreq,
addr,
count, datatype, src, tag, comm, true);
PERUSE_TRACE_COMM_EVENT (PERUSE_COMM_REQ_ACTIVATE,
&((recvreq)->req_recv.req_base),
PERUSE_RECV);
/* Work around a leak in start by marking this request as complete. The
* problem occured because we do not have a way to differentiate an
* inital request and an incomplete pml request in start. This line
* allows us to detect this state. */
recvreq->req_recv.req_base.req_pml_complete = true;
*request = (ompi_request_t *) recvreq;
return OMPI_SUCCESS;
}
int mca_pml_csum_recv(void *addr,
size_t count,
ompi_datatype_t * datatype,
int src,
int tag,
struct ompi_communicator_t *comm,
ompi_status_public_t * status)
{
int rc;
mca_pml_csum_recv_request_t *recvreq;
MCA_PML_CSUM_RECV_REQUEST_ALLOC(recvreq, rc);
if (NULL == recvreq)
return rc;
MCA_PML_CSUM_RECV_REQUEST_INIT(recvreq,
addr,
count, datatype, src, tag, comm, false);
PERUSE_TRACE_COMM_EVENT (PERUSE_COMM_REQ_ACTIVATE,
&((recvreq)->req_recv.req_base),
PERUSE_RECV);
MCA_PML_CSUM_RECV_REQUEST_START(recvreq);
ompi_request_wait_completion(&recvreq->req_recv.req_base.req_ompi);
if (NULL != status) { /* return status */
OMPI_STATUS_SET(status, &recvreq->req_recv.req_base.req_ompi.req_status);
}
rc = recvreq->req_recv.req_base.req_ompi.req_status.MPI_ERROR;
ompi_request_free( (ompi_request_t**)&recvreq );
return rc;
}
int mca_pml_ob1_isend(const void *buf,
size_t count,
ompi_datatype_t * datatype,
int dst,
int tag,
mca_pml_base_send_mode_t sendmode,
ompi_communicator_t * comm,
ompi_request_t ** request)
{
mca_pml_ob1_comm_proc_t *ob1_proc = mca_pml_ob1_peer_lookup (comm, dst);
mca_pml_ob1_send_request_t *sendreq = NULL;
ompi_proc_t *dst_proc = ob1_proc->ompi_proc;
mca_bml_base_endpoint_t* endpoint = mca_bml_base_get_endpoint (dst_proc);
int16_t seqn;
int rc;
if (OPAL_UNLIKELY(NULL == endpoint)) {
return OMPI_ERR_UNREACH;
}
seqn = (uint16_t) OPAL_THREAD_ADD32(&ob1_proc->send_sequence, 1);
if (MCA_PML_BASE_SEND_SYNCHRONOUS != sendmode) {
rc = mca_pml_ob1_send_inline (buf, count, datatype, dst, tag, seqn, dst_proc,
endpoint, comm);
if (OPAL_LIKELY(0 <= rc)) {
/* NTH: it is legal to return ompi_request_empty since the only valid
* field in a send completion status is whether or not the send was
* cancelled (which it can't be at this point anyway). */
*request = &ompi_request_empty;
return OMPI_SUCCESS;
}
}
MCA_PML_OB1_SEND_REQUEST_ALLOC(comm, dst, sendreq);
if (NULL == sendreq)
return OMPI_ERR_OUT_OF_RESOURCE;
MCA_PML_OB1_SEND_REQUEST_INIT(sendreq,
buf,
count,
datatype,
dst, tag,
comm, sendmode, false);
PERUSE_TRACE_COMM_EVENT (PERUSE_COMM_REQ_ACTIVATE,
&(sendreq)->req_send.req_base,
PERUSE_SEND);
MCA_PML_OB1_SEND_REQUEST_START_W_SEQ(sendreq, endpoint, seqn, rc);
*request = (ompi_request_t *) sendreq;
return rc;
}
int mca_pml_ob1_recv(void *addr,
size_t count,
ompi_datatype_t * datatype,
int src,
int tag,
struct ompi_communicator_t *comm,
ompi_status_public_t * status)
{
mca_pml_ob1_recv_request_t *recvreq = NULL;
int rc;
#if !OMPI_ENABLE_THREAD_MULTIPLE
recvreq = mca_pml_ob1_recvreq;
mca_pml_ob1_recvreq = NULL;
if( OPAL_UNLIKELY(NULL == recvreq) )
#endif /* !OMPI_ENABLE_THREAD_MULTIPLE */
{
MCA_PML_OB1_RECV_REQUEST_ALLOC(recvreq);
if (NULL == recvreq)
return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
}
MCA_PML_OB1_RECV_REQUEST_INIT(recvreq, addr, count, datatype,
src, tag, comm, false);
PERUSE_TRACE_COMM_EVENT (PERUSE_COMM_REQ_ACTIVATE,
&(recvreq->req_recv.req_base),
PERUSE_RECV);
MCA_PML_OB1_RECV_REQUEST_START(recvreq);
ompi_request_wait_completion(&recvreq->req_recv.req_base.req_ompi);
if (NULL != status) { /* return status */
*status = recvreq->req_recv.req_base.req_ompi.req_status;
}
rc = recvreq->req_recv.req_base.req_ompi.req_status.MPI_ERROR;
#if OMPI_ENABLE_THREAD_MULTIPLE
MCA_PML_OB1_RECV_REQUEST_RETURN(recvreq);
#else
if( NULL != mca_pml_ob1_recvreq ) {
MCA_PML_OB1_RECV_REQUEST_RETURN(recvreq);
} else {
mca_pml_ob1_recv_request_fini (recvreq);
mca_pml_ob1_recvreq = recvreq;
}
#endif
return rc;
}
int mca_pml_ob1_recv(void *addr,
size_t count,
ompi_datatype_t * datatype,
int src,
int tag,
struct ompi_communicator_t *comm,
ompi_status_public_t * status)
{
mca_pml_ob1_recv_request_t *recvreq = NULL;
int rc;
if (OPAL_LIKELY(!ompi_mpi_thread_multiple)) {
recvreq = mca_pml_ob1_recvreq;
mca_pml_ob1_recvreq = NULL;
}
if( OPAL_UNLIKELY(NULL == recvreq) ) {
MCA_PML_OB1_RECV_REQUEST_ALLOC(recvreq);
if (NULL == recvreq)
return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
}
recvreq->req_recv.req_base.req_type = MCA_PML_REQUEST_RECV;
MCA_PML_OB1_RECV_REQUEST_INIT(recvreq, addr, count, datatype,
src, tag, comm, false);
PERUSE_TRACE_COMM_EVENT (PERUSE_COMM_REQ_ACTIVATE,
&(recvreq->req_recv.req_base),
PERUSE_RECV);
MCA_PML_OB1_RECV_REQUEST_START(recvreq);
ompi_request_wait_completion(&recvreq->req_recv.req_base.req_ompi);
if (NULL != status) { /* return status */
*status = recvreq->req_recv.req_base.req_ompi.req_status;
}
rc = recvreq->req_recv.req_base.req_ompi.req_status.MPI_ERROR;
if (OPAL_UNLIKELY(ompi_mpi_thread_multiple || NULL != mca_pml_ob1_recvreq)) {
MCA_PML_OB1_RECV_REQUEST_RETURN(recvreq);
} else {
mca_pml_ob1_recv_request_fini (recvreq);
mca_pml_ob1_recvreq = recvreq;
}
return rc;
}
int mca_pml_ob1_recv(void *addr,
size_t count,
ompi_datatype_t * datatype,
int src,
int tag,
struct ompi_communicator_t *comm,
ompi_status_public_t * status)
{
mca_pml_ob1_recv_request_t *recvreq = NULL;
int rc;
if (OPAL_LIKELY(!ompi_mpi_thread_multiple)) {
recvreq = mca_pml_ob1_recvreq;
mca_pml_ob1_recvreq = NULL;
}
if( OPAL_UNLIKELY(NULL == recvreq) ) {
MCA_PML_OB1_RECV_REQUEST_ALLOC(recvreq);
if (NULL == recvreq)
return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
}
recvreq->req_recv.req_base.req_type = MCA_PML_REQUEST_RECV;
MCA_PML_OB1_RECV_REQUEST_INIT(recvreq, addr, count, datatype,
src, tag, comm, false);
PERUSE_TRACE_COMM_EVENT (PERUSE_COMM_REQ_ACTIVATE,
&(recvreq->req_recv.req_base),
PERUSE_RECV);
MCA_PML_OB1_RECV_REQUEST_START(recvreq);
ompi_request_wait_completion(&recvreq->req_recv.req_base.req_ompi);
if (NULL != status) { /* return status */
*status = recvreq->req_recv.req_base.req_ompi.req_status;
}
rc = recvreq->req_recv.req_base.req_ompi.req_status.MPI_ERROR;
if (recvreq->req_recv.req_base.req_pml_complete) {
/* make buffer defined when the request is compeleted,
and before releasing the objects. */
MEMCHECKER(
memchecker_call(&opal_memchecker_base_mem_defined,
recvreq->req_recv.req_base.req_addr,
recvreq->req_recv.req_base.req_count,
recvreq->req_recv.req_base.req_datatype);
);
/*
* The free call mark the final stage in a request life-cycle. Starting from this
* point the request is completed at both PML and user level, and can be used
* for others p2p communications. Therefore, in the case of the OB1 PML it should
* be added to the free request list.
*/
static int mca_pml_ob1_send_request_free(struct ompi_request_t** request)
{
mca_pml_ob1_send_request_t* sendreq = *(mca_pml_ob1_send_request_t**)request;
if(false == sendreq->req_send.req_base.req_free_called) {
sendreq->req_send.req_base.req_free_called = true;
PERUSE_TRACE_COMM_EVENT( PERUSE_COMM_REQ_NOTIFY,
&(sendreq->req_send.req_base), PERUSE_SEND );
if( true == sendreq->req_send.req_base.req_pml_complete ) {
/* make buffer defined when the request is compeleted,
and before releasing the objects. */
MEMCHECKER(
memchecker_call(&opal_memchecker_base_mem_defined,
sendreq->req_send.req_base.req_addr,
sendreq->req_send.req_base.req_count,
sendreq->req_send.req_base.req_datatype);
);
static int mca_pml_ob1_recv_request_free(struct ompi_request_t** request)
{
mca_pml_ob1_recv_request_t* recvreq = *(mca_pml_ob1_recv_request_t**)request;
assert( false == recvreq->req_recv.req_base.req_free_called );
OPAL_THREAD_LOCK(&ompi_request_lock);
recvreq->req_recv.req_base.req_free_called = true;
PERUSE_TRACE_COMM_EVENT( PERUSE_COMM_REQ_NOTIFY,
&(recvreq->req_recv.req_base), PERUSE_RECV );
if( true == recvreq->req_recv.req_base.req_pml_complete ) {
/* make buffer defined when the request is compeleted,
and before releasing the objects. */
MEMCHECKER(
memchecker_call(&opal_memchecker_base_mem_defined,
recvreq->req_recv.req_base.req_addr,
recvreq->req_recv.req_base.req_count,
recvreq->req_recv.req_base.req_datatype);
);
int mca_pml_ob1_send(void *buf,
size_t count,
ompi_datatype_t * datatype,
int dst,
int tag,
mca_pml_base_send_mode_t sendmode,
ompi_communicator_t * comm)
{
mca_pml_ob1_comm_t* ob1_comm = comm->c_pml_comm;
ompi_proc_t *dst_proc = ompi_comm_peer_lookup (comm, dst);
mca_bml_base_endpoint_t* endpoint = (mca_bml_base_endpoint_t*)
dst_proc->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_BML];
mca_pml_ob1_send_request_t *sendreq =
alloca(mca_pml_base_send_requests.fl_frag_size);
int16_t seqn;
int rc;
if (OPAL_UNLIKELY(MCA_PML_BASE_SEND_BUFFERED == sendmode)) {
/* large buffered sends *need* a real request so use isend instead */
ompi_request_t *brequest;
rc = mca_pml_ob1_isend (buf, count, datatype, dst, tag, sendmode, comm, &brequest);
if (OPAL_UNLIKELY(OMPI_SUCCESS != rc)) {
return rc;
}
/* free the request and return. don't care if it completes now */
ompi_request_free (&brequest);
return OMPI_SUCCESS;
}
if (OPAL_UNLIKELY(NULL == endpoint)) {
return OMPI_ERR_UNREACH;
}
seqn = (uint16_t) OPAL_THREAD_ADD32(&ob1_comm->procs[dst].send_sequence, 1);
if (MCA_PML_BASE_SEND_SYNCHRONOUS != sendmode) {
rc = mca_pml_ob1_send_inline (buf, count, datatype, dst, tag, seqn, dst_proc,
endpoint, comm);
if (OPAL_LIKELY(0 <= rc)) {
return OMPI_SUCCESS;
}
}
OBJ_CONSTRUCT(sendreq, mca_pml_ob1_send_request_t);
sendreq->req_send.req_base.req_proc = dst_proc;
sendreq->src_des = NULL;
MCA_PML_OB1_SEND_REQUEST_INIT(sendreq,
buf,
count,
datatype,
dst, tag,
comm, sendmode, false);
PERUSE_TRACE_COMM_EVENT (PERUSE_COMM_REQ_ACTIVATE,
&sendreq->req_send.req_base,
PERUSE_SEND);
MCA_PML_OB1_SEND_REQUEST_START_W_SEQ(sendreq, endpoint, seqn, rc);
if (rc != OMPI_SUCCESS) {
return rc;
}
ompi_request_wait_completion(&sendreq->req_send.req_base.req_ompi);
rc = sendreq->req_send.req_base.req_ompi.req_status.MPI_ERROR;
MCA_PML_BASE_SEND_REQUEST_FINI(&sendreq->req_send);
OBJ_DESTRUCT(sendreq);
return rc;
}
int mca_pml_ob1_send(const void *buf,
size_t count,
ompi_datatype_t * datatype,
int dst,
int tag,
mca_pml_base_send_mode_t sendmode,
ompi_communicator_t * comm)
{
mca_pml_ob1_comm_proc_t *ob1_proc = mca_pml_ob1_peer_lookup (comm, dst);
ompi_proc_t *dst_proc = ob1_proc->ompi_proc;
mca_bml_base_endpoint_t* endpoint = mca_bml_base_get_endpoint (dst_proc);
mca_pml_ob1_send_request_t *sendreq = NULL;
int16_t seqn;
int rc;
if (OPAL_UNLIKELY(NULL == endpoint)) {
return OMPI_ERR_UNREACH;
}
if (OPAL_UNLIKELY(MCA_PML_BASE_SEND_BUFFERED == sendmode)) {
/* large buffered sends *need* a real request so use isend instead */
ompi_request_t *brequest;
rc = mca_pml_ob1_isend (buf, count, datatype, dst, tag, sendmode, comm, &brequest);
if (OPAL_UNLIKELY(OMPI_SUCCESS != rc)) {
return rc;
}
ompi_request_wait_completion (brequest);
ompi_request_free (&brequest);
return OMPI_SUCCESS;
}
seqn = (uint16_t) OPAL_THREAD_ADD32(&ob1_proc->send_sequence, 1);
/**
* The immediate send will not have a request, so they are
* intracable from the point of view of any debugger attached to
* the parallel application.
*/
if (MCA_PML_BASE_SEND_SYNCHRONOUS != sendmode) {
rc = mca_pml_ob1_send_inline (buf, count, datatype, dst, tag, seqn, dst_proc,
endpoint, comm);
if (OPAL_LIKELY(0 <= rc)) {
return OMPI_SUCCESS;
}
}
if (OPAL_LIKELY(!ompi_mpi_thread_multiple)) {
sendreq = mca_pml_ob1_sendreq;
mca_pml_ob1_sendreq = NULL;
}
if( OPAL_UNLIKELY(NULL == sendreq) ) {
MCA_PML_OB1_SEND_REQUEST_ALLOC(comm, dst, sendreq);
if (NULL == sendreq)
return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
}
sendreq->req_send.req_base.req_proc = dst_proc;
sendreq->rdma_frag = NULL;
MCA_PML_OB1_SEND_REQUEST_INIT(sendreq,
buf,
count,
datatype,
dst, tag,
comm, sendmode, false);
PERUSE_TRACE_COMM_EVENT (PERUSE_COMM_REQ_ACTIVATE,
&sendreq->req_send.req_base,
PERUSE_SEND);
MCA_PML_OB1_SEND_REQUEST_START_W_SEQ(sendreq, endpoint, seqn, rc);
if (OPAL_LIKELY(rc == OMPI_SUCCESS)) {
ompi_request_wait_completion(&sendreq->req_send.req_base.req_ompi);
rc = sendreq->req_send.req_base.req_ompi.req_status.MPI_ERROR;
}
if (OPAL_UNLIKELY(ompi_mpi_thread_multiple || NULL != mca_pml_ob1_sendreq)) {
MCA_PML_OB1_SEND_REQUEST_RETURN(sendreq);
} else {
mca_pml_ob1_send_request_fini (sendreq);
mca_pml_ob1_sendreq = sendreq;
}
return rc;
}
int
mca_pml_bfo_mrecv( void *buf,
size_t count,
ompi_datatype_t *datatype,
struct ompi_message_t **message,
ompi_status_public_t* status )
{
mca_pml_bfo_recv_frag_t* frag;
mca_pml_bfo_recv_request_t *recvreq;
mca_pml_bfo_hdr_t *hdr;
int src, tag, rc;
ompi_communicator_t *comm;
mca_pml_bfo_comm_proc_t* proc;
mca_pml_bfo_comm_t* bfo_comm;
uint64_t seq;
/* get the request from the message and the frag from the request
before we overwrite everything */
comm = (*message)->comm;
recvreq = (mca_pml_bfo_recv_request_t*) (*message)->req_ptr;
frag = (mca_pml_bfo_recv_frag_t*) recvreq->req_recv.req_base.req_addr;
src = recvreq->req_recv.req_base.req_ompi.req_status.MPI_SOURCE;
tag = recvreq->req_recv.req_base.req_ompi.req_status.MPI_TAG;
seq = recvreq->req_recv.req_base.req_sequence;
bfo_comm = recvreq->req_recv.req_base.req_comm->c_pml_comm;
/* make the request a recv request again */
/* The old request kept pointers to comm and the char datatype.
We're about to release those, but need to make sure comm
doesn't go out of scope (we don't care about the char datatype
anymore). So retain comm, then release the frag, then reinit
the frag (which will retain comm), then release comm (but the
frag still has it's ref, so it'll stay in scope). Make
sense? */
OBJ_RETAIN(comm);
MCA_PML_BASE_RECV_REQUEST_FINI(&recvreq->req_recv);
recvreq->req_recv.req_base.req_type = MCA_PML_REQUEST_RECV;
MCA_PML_BFO_RECV_REQUEST_INIT(recvreq,
buf,
count, datatype,
src, tag, comm, false);
OBJ_RELEASE(comm);
PERUSE_TRACE_COMM_EVENT (PERUSE_COMM_REQ_ACTIVATE,
&((recvreq)->req_recv.req_base),
PERUSE_RECV);
/* init/re-init the request */
recvreq->req_lock = 0;
recvreq->req_pipeline_depth = 0;
recvreq->req_bytes_received = 0;
recvreq->req_rdma_cnt = 0;
recvreq->req_rdma_idx = 0;
recvreq->req_pending = false;
MCA_PML_BASE_RECV_START(&recvreq->req_recv.req_base);
/* Note - sequence number already assigned */
recvreq->req_recv.req_base.req_sequence = seq;
proc = &bfo_comm->procs[recvreq->req_recv.req_base.req_peer];
recvreq->req_recv.req_base.req_proc = proc->ompi_proc;
prepare_recv_req_converter(recvreq);
/* we can't go through the match, since we already have the match.
Cheat and do what REQUEST_START does, but without the frag
search */
hdr = (mca_pml_bfo_hdr_t*)frag->segments->seg_addr.pval;
switch(hdr->hdr_common.hdr_type) {
case MCA_PML_BFO_HDR_TYPE_MATCH:
mca_pml_bfo_recv_request_progress_match(recvreq, frag->btl, frag->segments,
frag->num_segments);
break;
case MCA_PML_BFO_HDR_TYPE_RNDV:
mca_pml_bfo_recv_request_progress_rndv(recvreq, frag->btl, frag->segments,
frag->num_segments);
break;
case MCA_PML_BFO_HDR_TYPE_RGET:
mca_pml_bfo_recv_request_progress_rget(recvreq, frag->btl, frag->segments,
frag->num_segments);
break;
default:
assert(0);
}
ompi_message_return(*message);
*message = MPI_MESSAGE_NULL;
ompi_request_wait_completion(&(recvreq->req_recv.req_base.req_ompi));
MCA_PML_BFO_RECV_FRAG_RETURN(frag);
if (NULL != status) { /* return status */
*status = recvreq->req_recv.req_base.req_ompi.req_status;
}
rc = recvreq->req_recv.req_base.req_ompi.req_status.MPI_ERROR;
ompi_request_free( (ompi_request_t**)&recvreq );
return rc;
}
int
mca_pml_ob1_imrecv( void *buf,
size_t count,
ompi_datatype_t *datatype,
struct ompi_message_t **message,
struct ompi_request_t **request )
{
mca_pml_ob1_recv_frag_t* frag;
mca_pml_ob1_recv_request_t *recvreq;
mca_pml_ob1_hdr_t *hdr;
int src, tag;
ompi_communicator_t *comm;
mca_pml_ob1_comm_proc_t* proc;
mca_pml_ob1_comm_t* ob1_comm;
uint64_t seq;
/* get the request from the message and the frag from the request
before we overwrite everything */
recvreq = (mca_pml_ob1_recv_request_t*) (*message)->req_ptr;
frag = (mca_pml_ob1_recv_frag_t*) recvreq->req_recv.req_base.req_addr;
src = recvreq->req_recv.req_base.req_ompi.req_status.MPI_SOURCE;
tag = recvreq->req_recv.req_base.req_ompi.req_status.MPI_TAG;
comm = (*message)->comm;
ob1_comm = recvreq->req_recv.req_base.req_comm->c_pml_comm;
seq = recvreq->req_recv.req_base.req_sequence;
/* make the request a recv request again */
recvreq->req_recv.req_base.req_type = MCA_PML_REQUEST_RECV;
MCA_PML_OB1_RECV_REQUEST_INIT(recvreq,
buf,
count, datatype,
src, tag, comm, false);
PERUSE_TRACE_COMM_EVENT (PERUSE_COMM_REQ_ACTIVATE,
&((recvreq)->req_recv.req_base),
PERUSE_RECV);
/* init/re-init the request */
recvreq->req_lock = 0;
recvreq->req_pipeline_depth = 0;
recvreq->req_bytes_received = 0;
/* What about req_rdma_cnt ? */
recvreq->req_rdma_idx = 0;
recvreq->req_pending = false;
recvreq->req_ack_sent = false;
MCA_PML_BASE_RECV_START(&recvreq->req_recv.req_base);
/* Note - sequence number already assigned */
recvreq->req_recv.req_base.req_sequence = seq;
proc = &ob1_comm->procs[recvreq->req_recv.req_base.req_peer];
recvreq->req_recv.req_base.req_proc = proc->ompi_proc;
prepare_recv_req_converter(recvreq);
/* we can't go throught he match, since we already have the match.
Cheat and do what REQUEST_START does, but without the frag
search */
hdr = (mca_pml_ob1_hdr_t*)frag->segments->seg_addr.pval;
switch(hdr->hdr_common.hdr_type) {
case MCA_PML_OB1_HDR_TYPE_MATCH:
mca_pml_ob1_recv_request_progress_match(recvreq, frag->btl, frag->segments,
frag->num_segments);
break;
case MCA_PML_OB1_HDR_TYPE_RNDV:
mca_pml_ob1_recv_request_progress_rndv(recvreq, frag->btl, frag->segments,
frag->num_segments);
break;
case MCA_PML_OB1_HDR_TYPE_RGET:
mca_pml_ob1_recv_request_progress_rget(recvreq, frag->btl, frag->segments,
frag->num_segments);
break;
default:
assert(0);
}
MCA_PML_OB1_RECV_FRAG_RETURN(frag);
ompi_message_return(*message);
*message = MPI_MESSAGE_NULL;
*request = (ompi_request_t *) recvreq;
return OMPI_SUCCESS;
}
int mca_pml_ob1_send(void *buf,
size_t count,
ompi_datatype_t * datatype,
int dst,
int tag,
mca_pml_base_send_mode_t sendmode,
ompi_communicator_t * comm)
{
mca_pml_ob1_comm_t* ob1_comm = comm->c_pml_comm;
ompi_proc_t *dst_proc = ompi_comm_peer_lookup (comm, dst);
mca_bml_base_endpoint_t* endpoint = (mca_bml_base_endpoint_t*)
dst_proc->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_BML];
mca_pml_ob1_send_request_t *sendreq = NULL;
int16_t seqn;
int rc;
if (OPAL_UNLIKELY(MCA_PML_BASE_SEND_BUFFERED == sendmode)) {
/* large buffered sends *need* a real request so use isend instead */
ompi_request_t *brequest;
rc = mca_pml_ob1_isend (buf, count, datatype, dst, tag, sendmode, comm, &brequest);
if (OPAL_UNLIKELY(OMPI_SUCCESS != rc)) {
return rc;
}
/* free the request and return. don't care if it completes now */
ompi_request_free (&brequest);
return OMPI_SUCCESS;
}
if (OPAL_UNLIKELY(NULL == endpoint)) {
return OMPI_ERR_UNREACH;
}
seqn = (uint16_t) OPAL_THREAD_ADD32(&ob1_comm->procs[dst].send_sequence, 1);
/**
* The immediate send will not have a request, so they are
* intracable from the point of view of any debugger attached to
* the parallel application.
*/
if (MCA_PML_BASE_SEND_SYNCHRONOUS != sendmode) {
rc = mca_pml_ob1_send_inline (buf, count, datatype, dst, tag, seqn, dst_proc,
endpoint, comm);
if (OPAL_LIKELY(0 <= rc)) {
return OMPI_SUCCESS;
}
}
#if !OMPI_ENABLE_THREAD_MULTIPLE
sendreq = mca_pml_ob1_sendreq;
if( OPAL_UNLIKELY(NULL == sendreq) )
#endif /* !OMPI_ENABLE_THREAD_MULTIPLE */
{
MCA_PML_OB1_SEND_REQUEST_ALLOC(comm, dst, sendreq);
if (NULL == sendreq)
return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
#if !OMPI_ENABLE_THREAD_MULTIPLE
mca_pml_ob1_sendreq = sendreq;
#endif /* !OMPI_ENABLE_THREAD_MULTIPLE */
}
sendreq->req_send.req_base.req_proc = dst_proc;
sendreq->rdma_frag = NULL;
MCA_PML_OB1_SEND_REQUEST_INIT(sendreq,
buf,
count,
datatype,
dst, tag,
comm, sendmode, false);
PERUSE_TRACE_COMM_EVENT (PERUSE_COMM_REQ_ACTIVATE,
&sendreq->req_send.req_base,
PERUSE_SEND);
MCA_PML_OB1_SEND_REQUEST_START_W_SEQ(sendreq, endpoint, seqn, rc);
if (OPAL_LIKELY(rc == OMPI_SUCCESS)) {
ompi_request_wait_completion(&sendreq->req_send.req_base.req_ompi);
rc = sendreq->req_send.req_base.req_ompi.req_status.MPI_ERROR;
}
#if OMPI_ENABLE_THREAD_MULTIPLE
MCA_PML_OB1_SEND_REQUEST_RETURN(sendreq);
#else
mca_pml_ob1_send_request_fini (sendreq);
#endif
return rc;
}
