/* Send a zero-sized message with eager synchronous. This is a temporary
routine, as we may want to replace this with a counterpart to the
Eager Short message */
int MPIDI_CH3_EagerSyncZero(MPID_Request **sreq_p, int rank, int tag,
MPID_Comm * comm, int context_offset )
{
int mpi_errno = MPI_SUCCESS;
MPIDI_CH3_Pkt_t upkt;
MPIDI_CH3_Pkt_eager_sync_send_t * const es_pkt = &upkt.eager_sync_send;
MPIDI_VC_t * vc;
MPID_Request *sreq = *sreq_p;
MPIU_DBG_MSG(CH3_OTHER,VERBOSE,"sending zero length message");
/* MT FIXME what are the two operations we are waiting for? the send and
* the sync response? */
MPID_cc_set(&sreq->cc, 2);
MPIDI_Request_set_msg_type(sreq, MPIDI_REQUEST_EAGER_MSG);
sreq->dev.OnDataAvail = 0;
MPIDI_Pkt_init(es_pkt, MPIDI_CH3_PKT_EAGER_SYNC_SEND);
es_pkt->match.parts.rank = comm->rank;
es_pkt->match.parts.tag = tag;
es_pkt->match.parts.context_id = comm->context_id + context_offset;
es_pkt->sender_req_id = sreq->handle;
es_pkt->data_sz = 0;
MPIDI_Comm_get_vc_set_active(comm, rank, &vc);
MPIDI_VC_FAI_send_seqnum(vc, seqnum);
MPIDI_Pkt_set_seqnum(es_pkt, seqnum);
MPIDI_Request_set_seqnum(sreq, seqnum);
MPIU_DBG_MSGPKT(vc,tag,es_pkt->match.parts.context_id,rank,(MPIDI_msg_sz_t)0,"EagerSync0");
MPID_THREAD_CS_ENTER(POBJ, vc->pobj_mutex);
mpi_errno = MPIDI_CH3_iSend(vc, sreq, es_pkt, sizeof(*es_pkt));
MPID_THREAD_CS_EXIT(POBJ, vc->pobj_mutex);
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno != MPI_SUCCESS)
{
MPID_Request_release(sreq);
*sreq_p = NULL;
MPIR_ERR_SETANDJUMP(mpi_errno, MPI_ERR_OTHER, "**ch3|eagermsg");
}
/* --END ERROR HANDLING-- */
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPID_Recv_init(void * buf, int count, MPI_Datatype datatype, int rank, int tag, MPID_Comm * comm, int context_offset,
MPID_Request ** request)
{
MPID_Request * rreq;
int mpi_errno = MPI_SUCCESS;
MPIDI_STATE_DECL(MPID_STATE_MPID_RECV_INIT);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_RECV_INIT);
rreq = MPID_Request_create();
if (rreq == NULL)
{
/* --BEGIN ERROR HANDLING-- */
mpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_FATAL, FCNAME, __LINE__, MPI_ERR_OTHER, "**nomemreq", 0);
/* --END ERROR HANDLING-- */
goto fn_exit;
}
MPIU_Object_set_ref(rreq, 1);
rreq->kind = MPID_PREQUEST_RECV;
rreq->comm = comm;
MPID_cc_set(&rreq->cc, 0);
MPIR_Comm_add_ref(comm);
rreq->dev.match.parts.rank = rank;
rreq->dev.match.parts.tag = tag;
rreq->dev.match.parts.context_id = comm->recvcontext_id + context_offset;
rreq->dev.user_buf = (void *) buf;
rreq->dev.user_count = count;
rreq->dev.datatype = datatype;
rreq->partner_request = NULL;
MPIDI_Request_set_type(rreq, MPIDI_REQUEST_TYPE_RECV);
if (HANDLE_GET_KIND(datatype) != HANDLE_KIND_BUILTIN)
{
MPID_Datatype_get_ptr(datatype, rreq->dev.datatype_ptr);
MPID_Datatype_add_ref(rreq->dev.datatype_ptr);
}
*request = rreq;
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_MPID_RECV_INIT);
return mpi_errno;
}
int MPID_Irecv(void * buf, int count, MPI_Datatype datatype, int rank, int tag,
MPID_Comm * comm, int context_offset,
MPID_Request ** request)
{
MPID_Request * rreq;
int found;
int mpi_errno = MPI_SUCCESS;
MPIDI_STATE_DECL(MPID_STATE_MPID_IRECV);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_IRECV);
MPIU_DBG_MSG_FMT(CH3_OTHER,VERBOSE,(MPIU_DBG_FDEST,
"rank=%d, tag=%d, context=%d",
rank, tag, comm->recvcontext_id + context_offset));
if (rank == MPI_PROC_NULL)
{
rreq = MPID_Request_create();
if (rreq != NULL)
{
/* MT FIXME should these be handled by MPID_Request_create? */
MPIU_Object_set_ref(rreq, 1);
MPID_cc_set(&rreq->cc, 0);
rreq->kind = MPID_REQUEST_RECV;
MPIR_Status_set_procnull(&rreq->status);
}
else
{
MPIU_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,"**nomemreq");
}
goto fn_exit;
}
MPIU_THREAD_CS_ENTER(MSGQUEUE,);
rreq = MPIDI_CH3U_Recvq_FDU_or_AEP(rank, tag,
comm->recvcontext_id + context_offset,
comm, buf, count, datatype, &found);
if (rreq == NULL)
{
MPIU_THREAD_CS_EXIT(MSGQUEUE,);
MPIU_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,"**nomemreq");
}
int MPIDI_CH3_iStartMsg(MPIDI_VC_t * vc, void * hdr, MPIDI_msg_sz_t hdr_sz,
MPID_Request ** sreq_ptr)
{
MPID_Request * sreq = NULL;
MPIDI_CH3I_VC *vcch = &vc->ch;
int mpi_errno = MPI_SUCCESS;
MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3_ISTARTMSG);
MPIDI_FUNC_ENTER(MPID_STATE_MPIDI_CH3_ISTARTMSG);
MPIU_Assert( hdr_sz <= sizeof(MPIDI_CH3_Pkt_t));
/* The SOCK channel uses a fixed length header, the size of which is the
maximum of all possible packet headers */
hdr_sz = sizeof(MPIDI_CH3_Pkt_t);
MPIU_DBG_STMT(CH3_CHANNEL,VERBOSE,
MPIDI_DBG_Print_packet((MPIDI_CH3_Pkt_t*)hdr));
if (vcch->state == MPIDI_CH3I_VC_STATE_CONNECTED) /* MT */
{
/* Connection already formed. If send queue is empty attempt to send
data, queuing any unsent data. */
if (MPIDI_CH3I_SendQ_empty(vcch)) /* MT */
{
MPIU_Size_t nb;
int rc;
MPIU_DBG_MSG(CH3_CHANNEL,VERBOSE,
"send queue empty, attempting to write");
MPIU_DBG_PKT(vcch->conn,hdr,"istartmsg");
/* MT: need some signalling to lock down our right to use the
channel, thus insuring that the progress engine does
not also try to write */
rc = MPIDU_Sock_write(vcch->sock, hdr, hdr_sz, &nb);
if (rc == MPI_SUCCESS)
{
MPIU_DBG_MSG_D(CH3_CHANNEL,VERBOSE,
"wrote %ld bytes", (unsigned long) nb);
if (nb == hdr_sz)
{
MPIU_DBG_MSG_D(CH3_CHANNEL,VERBOSE,
"entire write complete, " MPIDI_MSG_SZ_FMT " bytes", nb);
/* done. get us out of here as quickly as possible. */
}
else
{
MPIU_DBG_MSG_D(CH3_CHANNEL,VERBOSE,
"partial write of " MPIDI_MSG_SZ_FMT " bytes, request enqueued at head", nb);
sreq = create_request(hdr, hdr_sz, nb);
if (!sreq) {
MPIU_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,"**nomem");
}
MPIDI_CH3I_SendQ_enqueue_head(vcch, sreq);
MPIU_DBG_MSG_FMT(CH3_CHANNEL,VERBOSE,
(MPIU_DBG_FDEST,"posting write, vc=0x%p, sreq=0x%08x", vc, sreq->handle));
vcch->conn->send_active = sreq;
mpi_errno = MPIDU_Sock_post_write(vcch->conn->sock, sreq->dev.iov[0].MPID_IOV_BUF,
sreq->dev.iov[0].MPID_IOV_LEN, sreq->dev.iov[0].MPID_IOV_LEN, NULL);
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno != MPI_SUCCESS)
{
mpi_errno = MPIR_Err_create_code(mpi_errno, MPIR_ERR_FATAL, FCNAME, __LINE__, MPI_ERR_OTHER,
"**ch3|sock|postwrite", "ch3|sock|postwrite %p %p %p",
sreq, vcch->conn, vc);
goto fn_fail;
}
/* --END ERROR HANDLING-- */
}
}
/* --BEGIN ERROR HANDLING-- */
else
{
MPIU_DBG_MSG_D(CH3_CHANNEL,TYPICAL,
"ERROR - MPIDU_Sock_write failed, rc=%d", rc);
sreq = MPID_Request_create();
if (!sreq) {
MPIU_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,"**nomem");
}
sreq->kind = MPID_REQUEST_SEND;
MPID_cc_set(&(sreq->cc), 0);
sreq->status.MPI_ERROR = MPIR_Err_create_code( rc,
MPIR_ERR_RECOVERABLE, FCNAME, __LINE__,
MPI_ERR_INTERN, "**ch3|sock|writefailed",
"**ch3|sock|writefailed %d", rc );
/* Make sure that the caller sees this error */
mpi_errno = sreq->status.MPI_ERROR;
}
/* --END ERROR HANDLING-- */
}
else
{
MPIU_DBG_MSG(CH3_CHANNEL,VERBOSE,
"send in progress, request enqueued");
sreq = create_request(hdr, hdr_sz, 0);
if (!sreq) {
MPIU_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,"**nomem");
}
MPIDI_CH3I_SendQ_enqueue(vcch, sreq);
}
}
else if (vcch->state == MPIDI_CH3I_VC_STATE_CONNECTING) /* MT */
{
MPIU_DBG_VCUSE(vc,
"connecteding. enqueuing request");
/* queue the data so it can be sent after the connection is formed */
sreq = create_request(hdr, hdr_sz, 0);
if (!sreq) {
MPIU_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,"**nomem");
}
MPIDI_CH3I_SendQ_enqueue(vcch, sreq);
}
else if (vcch->state == MPIDI_CH3I_VC_STATE_UNCONNECTED) /* MT */
{
MPIU_DBG_VCUSE(vc,
"unconnected. posting connect and enqueuing request");
/* queue the data so it can be sent after the connection is formed */
sreq = create_request(hdr, hdr_sz, 0);
if (!sreq) {
MPIU_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,"**nomem");
}
MPIDI_CH3I_SendQ_enqueue(vcch, sreq);
/* Form a new connection */
MPIDI_CH3I_VC_post_connect(vc);
}
else if (vcch->state != MPIDI_CH3I_VC_STATE_FAILED)
{
/* Unable to send data at the moment, so queue it for later */
MPIU_DBG_VCUSE(vc,"forming connection, request enqueued");
sreq = create_request(hdr, hdr_sz, 0);
if (!sreq) {
MPIU_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,"**nomem");
}
MPIDI_CH3I_SendQ_enqueue(vcch, sreq);
}
/* --BEGIN ERROR HANDLING-- */
else
{
/* Connection failed, so allocate a request and return an error. */
MPIU_DBG_VCUSE(vc,"ERROR - connection failed");
sreq = MPID_Request_create();
if (!sreq) {
MPIU_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,"**nomem");
}
sreq->kind = MPID_REQUEST_SEND;
MPID_cc_set(&sreq->cc, 0);
sreq->status.MPI_ERROR = MPIR_Err_create_code( MPI_SUCCESS,
MPIR_ERR_RECOVERABLE, FCNAME, __LINE__,
MPI_ERR_INTERN, "**ch3|sock|connectionfailed",0 );
/* Make sure that the caller sees this error */
mpi_errno = sreq->status.MPI_ERROR;
}
/* --END ERROR HANDLING-- */
fn_fail:
*sreq_ptr = sreq;
MPIDI_FUNC_EXIT(MPID_STATE_MPIDI_CH3_ISTARTMSG);
return mpi_errno;
}
int MPID_Cancel_send(MPID_Request * sreq)
{
MPIDI_VC_t * vc;
int proto;
int flag;
int mpi_errno = MPI_SUCCESS;
MPIDI_STATE_DECL(MPID_STATE_MPID_CANCEL_SEND);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_CANCEL_SEND);
MPIU_Assert(sreq->kind == MPID_REQUEST_SEND);
MPIDI_Request_cancel_pending(sreq, &flag);
if (flag)
{
goto fn_exit;
}
/*
* FIXME: user requests returned by MPI_Ibsend() have a NULL comm pointer
* and no pointer to the underlying communication
* request. For now, we simply fail to cancel the request. In the future,
* we should add a new request kind to indicate that
* the request is a BSEND. Then we can properly cancel the request, much
* in the way we do persistent requests.
*/
if (sreq->comm == NULL)
{
goto fn_exit;
}
MPIDI_Comm_get_vc_set_active(sreq->comm, sreq->dev.match.parts.rank, &vc);
proto = MPIDI_Request_get_msg_type(sreq);
if (proto == MPIDI_REQUEST_SELF_MSG)
{
MPID_Request * rreq;
MPIU_DBG_MSG(CH3_OTHER,VERBOSE,
"attempting to cancel message sent to self");
MPIU_THREAD_CS_ENTER(MSGQUEUE,);
rreq = MPIDI_CH3U_Recvq_FDU(sreq->handle, &sreq->dev.match);
MPIU_THREAD_CS_EXIT(MSGQUEUE,);
if (rreq)
{
MPIU_Assert(rreq->partner_request == sreq);
MPIU_DBG_MSG_FMT(CH3_OTHER,VERBOSE,(MPIU_DBG_FDEST,
"send-to-self cancellation successful, sreq=0x%08x, rreq=0x%08x",
sreq->handle, rreq->handle));
MPIU_Object_set_ref(rreq, 0);
MPIDI_CH3_Request_destroy(rreq);
sreq->status.cancelled = TRUE;
/* no other thread should be waiting on sreq, so it is safe to
reset ref_count and cc */
MPID_cc_set(&sreq->cc, 0);
/* FIXME should be a decr and assert, not a set */
MPIU_Object_set_ref(sreq, 1);
}
else
{
sreq->status.cancelled = FALSE;
MPIU_DBG_MSG_FMT(CH3_OTHER,VERBOSE,(MPIU_DBG_FDEST,
"send-to-self cancellation failed, sreq=0x%08x, rreq=0x%08x",
sreq->handle, rreq->handle));
}
goto fn_exit;
}
MPID_Request * MPID_Request_create(void)
{
MPID_Request * req;
MPIDI_STATE_DECL(MPID_STATE_MPID_REQUEST_CREATE);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_REQUEST_CREATE);
req = MPIU_Handle_obj_alloc(&MPID_Request_mem);
if (req != NULL)
{
MPIU_DBG_MSG_P(CH3_CHANNEL,VERBOSE,
"allocated request, handle=0x%08x", req->handle);
#ifdef MPICH_DBG_OUTPUT
/*MPIU_Assert(HANDLE_GET_MPI_KIND(req->handle) == MPID_REQUEST);*/
if (HANDLE_GET_MPI_KIND(req->handle) != MPID_REQUEST)
{
int mpi_errno;
mpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_FATAL,
FCNAME, __LINE__, MPI_ERR_OTHER,
"**invalid_handle", "**invalid_handle %d", req->handle);
MPID_Abort(MPIR_Process.comm_world, mpi_errno, -1, NULL);
}
#endif
/* FIXME: This makes request creation expensive. We need to trim
this to the basics, with additional setup for special-purpose
requests (think base class and inheritance). For example, do we
*really* want to set the kind to UNDEFINED? And should the RMA
values be set only for RMA requests? */
MPIU_Object_set_ref(req, 1);
req->kind = MPID_REQUEST_UNDEFINED;
MPID_cc_set(&req->cc, 1);
req->cc_ptr = &req->cc;
/* FIXME: status fields meaningful only for receive, and even then
should not need to be set. */
req->status.MPI_SOURCE = MPI_UNDEFINED;
req->status.MPI_TAG = MPI_UNDEFINED;
req->status.MPI_ERROR = MPI_SUCCESS;
MPIR_STATUS_SET_COUNT(req->status, 0);
MPIR_STATUS_SET_CANCEL_BIT(req->status, FALSE);
req->comm = NULL;
req->greq_fns = NULL;
req->errflag = MPIR_ERR_NONE;
req->dev.datatype_ptr = NULL;
req->dev.segment_ptr = NULL;
/* Masks and flags for channel device state in an MPID_Request */
req->dev.state = 0;
req->dev.cancel_pending = FALSE;
/* FIXME: RMA ops shouldn't need to be set except when creating a
request for RMA operations */
req->dev.target_win_handle = MPI_WIN_NULL;
req->dev.source_win_handle = MPI_WIN_NULL;
req->dev.lock_queue_entry = NULL;
req->dev.dtype_info = NULL;
req->dev.dataloop = NULL;
req->dev.iov_offset = 0;
req->dev.flags = MPIDI_CH3_PKT_FLAG_NONE;
req->dev.resp_request_handle = MPI_REQUEST_NULL;
req->dev.user_buf = NULL;
req->dev.OnDataAvail = NULL;
req->dev.OnFinal = NULL;
req->dev.user_buf = NULL;
req->dev.drop_data = FALSE;
req->dev.stream_offset = 0;
#ifdef MPIDI_CH3_REQUEST_INIT
MPIDI_CH3_REQUEST_INIT(req);
#endif
}
else
{
/* FIXME: This fails to fail if debugging is turned off */
MPIU_DBG_MSG(CH3_CHANNEL,TYPICAL,"unable to allocate a request");
}
MPIDI_FUNC_EXIT(MPID_STATE_MPID_REQUEST_CREATE);
return req;
}
int MPID_Irsend(const void * buf, int count, MPI_Datatype datatype, int rank, int tag, MPID_Comm * comm, int context_offset,
MPID_Request ** request)
{
MPIDI_CH3_Pkt_t upkt;
MPIDI_CH3_Pkt_ready_send_t * const ready_pkt = &upkt.ready_send;
MPIDI_msg_sz_t data_sz;
int dt_contig;
MPI_Aint dt_true_lb;
MPID_Datatype * dt_ptr;
MPID_Request * sreq;
MPIDI_VC_t * vc;
#if defined(MPID_USE_SEQUENCE_NUMBERS)
MPID_Seqnum_t seqnum;
#endif
int mpi_errno = MPI_SUCCESS;
MPIDI_STATE_DECL(MPID_STATE_MPID_IRSEND);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_IRSEND);
MPIU_DBG_MSG_FMT(CH3_OTHER,VERBOSE,(MPIU_DBG_FDEST,
"rank=%d, tag=%d, context=%d",
rank, tag, comm->context_id + context_offset));
/* Check to make sure the communicator hasn't already been revoked */
if (comm->revoked &&
MPIR_AGREE_TAG != MPIR_TAG_MASK_ERROR_BITS(tag & ~MPIR_Process.tagged_coll_mask) &&
MPIR_SHRINK_TAG != MPIR_TAG_MASK_ERROR_BITS(tag & ~MPIR_Process.tagged_coll_mask)) {
MPIR_ERR_SETANDJUMP(mpi_errno,MPIX_ERR_REVOKED,"**revoked");
}
if (rank == comm->rank && comm->comm_kind != MPID_INTERCOMM)
{
mpi_errno = MPIDI_Isend_self(buf, count, datatype, rank, tag, comm, context_offset, MPIDI_REQUEST_TYPE_RSEND, &sreq);
goto fn_exit;
}
if (rank != MPI_PROC_NULL) {
MPIDI_Comm_get_vc_set_active(comm, rank, &vc);
#ifdef ENABLE_COMM_OVERRIDES
/* this needs to come before the sreq is created, since the override
* function is responsible for creating its own request */
if (vc->comm_ops && vc->comm_ops->irsend)
{
mpi_errno = vc->comm_ops->irsend( vc, buf, count, datatype, rank, tag, comm, context_offset, &sreq);
goto fn_exit;
}
#endif
}
MPIDI_Request_create_sreq(sreq, mpi_errno, goto fn_exit);
MPIDI_Request_set_type(sreq, MPIDI_REQUEST_TYPE_RSEND);
MPIDI_Request_set_msg_type(sreq, MPIDI_REQUEST_EAGER_MSG);
if (rank == MPI_PROC_NULL)
{
MPIU_Object_set_ref(sreq, 1);
MPID_cc_set(&sreq->cc, 0);
goto fn_exit;
}
MPIDI_Datatype_get_info(count, datatype, dt_contig, data_sz, dt_ptr, dt_true_lb);
MPIDI_Pkt_init(ready_pkt, MPIDI_CH3_PKT_READY_SEND);
ready_pkt->match.parts.rank = comm->rank;
ready_pkt->match.parts.tag = tag;
ready_pkt->match.parts.context_id = comm->context_id + context_offset;
ready_pkt->sender_req_id = MPI_REQUEST_NULL;
ready_pkt->data_sz = data_sz;
if (data_sz == 0)
{
MPIU_DBG_MSG(CH3_OTHER,VERBOSE,"sending zero length message");
sreq->dev.OnDataAvail = 0;
MPIDI_VC_FAI_send_seqnum(vc, seqnum);
MPIDI_Pkt_set_seqnum(ready_pkt, seqnum);
MPIDI_Request_set_seqnum(sreq, seqnum);
MPID_THREAD_CS_ENTER(POBJ, vc->pobj_mutex);
mpi_errno = MPIDI_CH3_iSend(vc, sreq, ready_pkt, sizeof(*ready_pkt));
MPID_THREAD_CS_EXIT(POBJ, vc->pobj_mutex);
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno != MPI_SUCCESS)
{
MPID_Request_release(sreq);
sreq = NULL;
MPIR_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**ch3|eagermsg");
goto fn_exit;
}
/* --END ERROR HANDLING-- */
goto fn_exit;
}
if (vc->ready_eager_max_msg_sz < 0 || data_sz + sizeof(MPIDI_CH3_Pkt_ready_send_t) <= vc->ready_eager_max_msg_sz) {
if (dt_contig) {
mpi_errno = MPIDI_CH3_EagerContigIsend( &sreq,
MPIDI_CH3_PKT_READY_SEND,
(char*)buf + dt_true_lb,
data_sz, rank, tag,
comm, context_offset );
}
else {
mpi_errno = MPIDI_CH3_EagerNoncontigSend( &sreq,
MPIDI_CH3_PKT_READY_SEND,
buf, count, datatype,
data_sz, rank, tag,
comm, context_offset );
/* If we're not complete, then add a reference to the datatype */
if (sreq && sreq->dev.OnDataAvail) {
sreq->dev.datatype_ptr = dt_ptr;
MPID_Datatype_add_ref(dt_ptr);
}
}
} else {
/* Do rendezvous. This will be sent as a regular send not as
a ready send, so the receiver won't know to send an error
if the receive has not been posted */
MPIDI_Request_set_msg_type( sreq, MPIDI_REQUEST_RNDV_MSG );
mpi_errno = vc->rndvSend_fn( &sreq, buf, count, datatype, dt_contig,
data_sz, dt_true_lb, rank, tag, comm,
context_offset );
if (sreq && dt_ptr != NULL) {
sreq->dev.datatype_ptr = dt_ptr;
MPID_Datatype_add_ref(dt_ptr);
}
}
fn_exit:
*request = sreq;
MPIU_DBG_STMT(CH3_OTHER,VERBOSE,{
if (sreq != NULL)
{
MPIU_DBG_MSG_P(CH3_OTHER,VERBOSE,"request allocated, handle=0x%08x", sreq->handle);
}
}
);
int MPID_Isend(const void * buf, int count, MPI_Datatype datatype, int rank,
int tag, MPID_Comm * comm, int context_offset,
MPID_Request ** request)
{
MPIDI_msg_sz_t data_sz;
int dt_contig;
MPI_Aint dt_true_lb;
MPID_Datatype * dt_ptr;
MPID_Request * sreq;
MPIDI_VC_t * vc=0;
#if defined(MPID_USE_SEQUENCE_NUMBERS)
MPID_Seqnum_t seqnum;
#endif
int mpi_errno = MPI_SUCCESS;
MPIDI_STATE_DECL(MPID_STATE_MPID_ISEND);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_ISEND);
MPIU_DBG_MSG_FMT(CH3_OTHER,VERBOSE,(MPIU_DBG_FDEST,
"rank=%d, tag=%d, context=%d",
rank, tag, comm->context_id + context_offset));
if (rank == comm->rank && comm->comm_kind != MPID_INTERCOMM)
{
#if defined (_OSU_PSM_)
goto skip_self_send; /* psm will internally do self-send, no special
handling is needed here */
#endif /* _OSU_PSM_ */
mpi_errno = MPIDI_Isend_self(buf, count, datatype, rank, tag, comm,
context_offset, MPIDI_REQUEST_TYPE_SEND, &sreq);
goto fn_exit;
}
#if defined (_OSU_PSM_)
skip_self_send:
#endif
if (rank != MPI_PROC_NULL) {
MPIDI_Comm_get_vc_set_active(comm, rank, &vc);
#ifdef ENABLE_COMM_OVERRIDES
/* this needs to come before the sreq is created, since the override
* function is responsible for creating its own request */
if (vc->comm_ops && vc->comm_ops->isend)
{
mpi_errno = vc->comm_ops->isend( vc, buf, count, datatype, rank, tag, comm, context_offset, &sreq);
goto fn_exit;
}
#endif
}
MPIDI_Request_create_sreq(sreq, mpi_errno, goto fn_exit);
MPIDI_Request_set_type(sreq, MPIDI_REQUEST_TYPE_SEND);
if (rank == MPI_PROC_NULL)
{
MPIU_Object_set_ref(sreq, 1);
MPID_cc_set(&sreq->cc, 0);
goto fn_exit;
}
MPIDI_Datatype_get_info(count, datatype, dt_contig, data_sz, dt_ptr,
dt_true_lb);
if (data_sz == 0)
{
#if defined (_OSU_PSM_)
goto eager_send;
#endif /* _OSU_PSM_ */
MPIDI_CH3_Pkt_t upkt;
MPIDI_CH3_Pkt_eager_send_t * const eager_pkt = &upkt.eager_send;
MPIDI_Request_set_msg_type(sreq, MPIDI_REQUEST_EAGER_MSG);
sreq->dev.OnDataAvail = 0;
MPIU_DBG_MSG(CH3_OTHER,VERBOSE,"sending zero length message");
MPIDI_Pkt_init(eager_pkt, MPIDI_CH3_PKT_EAGER_SEND);
eager_pkt->match.parts.rank = comm->rank;
eager_pkt->match.parts.tag = tag;
eager_pkt->match.parts.context_id = comm->context_id + context_offset;
eager_pkt->sender_req_id = sreq->handle;
eager_pkt->data_sz = 0;
MPIDI_VC_FAI_send_seqnum(vc, seqnum);
MPIDI_Pkt_set_seqnum(eager_pkt, seqnum);
MPIDI_Request_set_seqnum(sreq, seqnum);
MPIU_THREAD_CS_ENTER(CH3COMM,vc);
mpi_errno = MPIU_CALL(MPIDI_CH3,iSend(vc, sreq, eager_pkt,
sizeof(*eager_pkt)));
MPIU_THREAD_CS_EXIT(CH3COMM,vc);
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno != MPI_SUCCESS)
{
MPIU_Object_set_ref(sreq, 0);
MPIDI_CH3_Request_destroy(sreq);
sreq = NULL;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**ch3|eagermsg");
goto fn_exit;
}
/* --END ERROR HANDLING-- */
goto fn_exit;
}
#if defined (_OSU_PSM_)
if(HANDLE_GET_KIND(datatype) != HANDLE_KIND_BUILTIN) {
sreq->dev.datatype_ptr = dt_ptr;
MPID_Datatype_add_ref(dt_ptr);
sreq->psm_flags |= PSM_NEED_DTYPE_RELEASE;
}
if(vc->force_eager)
goto eager_send;
#endif /* _OSU_PSM_ */
#if defined(_OSU_MVAPICH_)
int i;
for (i = 0 ; i < rdma_num_extra_polls; i++)
{
if (rdma_global_ext_sendq_size > 1)
MPID_Progress_test();
}
#endif
/* FIXME: flow control: limit number of outstanding eager messsages
containing data and need to be buffered by the receiver */
#if defined(_OSU_MVAPICH_)
if (data_sz + sizeof(MPIDI_CH3_Pkt_eager_send_t) <= vc->eager_max_msg_sz
&& !vc->force_rndv)
#else /* defined(_OSU_MVAPICH_) */
if (data_sz + sizeof(MPIDI_CH3_Pkt_eager_send_t) <= vc->eager_max_msg_sz)
#endif /* defined(_OSU_MVAPICH_) */
{
#if defined (_OSU_PSM_)
eager_send:
#endif /* _OSU_PSM */
if (dt_contig)
{
mpi_errno = MPIDI_CH3_EagerContigIsend( &sreq,
MPIDI_CH3_PKT_EAGER_SEND,
(char*)buf + dt_true_lb,
data_sz, rank, tag,
comm, context_offset );
}
else
{
#if defined (_OSU_PSM_)
sreq->psm_flags |= PSM_NON_BLOCKING_SEND;
#endif
mpi_errno = MPIDI_CH3_EagerNoncontigSend( &sreq,
MPIDI_CH3_PKT_EAGER_SEND,
buf, count, datatype,
data_sz, rank, tag,
comm, context_offset );
#if defined (_OSU_PSM_)
goto fn_exit;
#endif
/* If we're not complete, then add a reference to the datatype */
if (sreq && sreq->dev.OnDataAvail) {
sreq->dev.datatype_ptr = dt_ptr;
MPID_Datatype_add_ref(dt_ptr);
}
}
}
else
{
/* Note that the sreq was created above */
MPIDI_Request_set_msg_type( sreq, MPIDI_REQUEST_RNDV_MSG );
mpi_errno = vc->rndvSend_fn( &sreq, buf, count, datatype, dt_contig,
data_sz, dt_true_lb, rank, tag, comm,
context_offset );
/* FIXME: fill temporary IOV or pack temporary buffer after send to
hide some latency. This requires synchronization
because the CTS packet could arrive and be processed before the
above iStartmsg completes (depending on the progress
engine, threads, etc.). */
#if defined(_OSU_MVAPICH_)
/* rndv transfers need to process CTS packet to initiate the actual RDMA transfer */
MPID_Progress_test();
#endif /* defined(_OSU_MVAPICH_) */
if (sreq && dt_ptr != NULL)
{
sreq->dev.datatype_ptr = dt_ptr;
MPID_Datatype_add_ref(dt_ptr);
}
}
fn_exit:
*request = sreq;
#if defined(_OSU_MVAPICH_)
for (i = 0 ; i < rdma_num_extra_polls; i++)
{
if (rdma_global_ext_sendq_size > 1)
MPID_Progress_test();
}
#endif
MPIU_DBG_STMT(CH3_OTHER,VERBOSE,
{
if (sreq != NULL)
{
MPIU_DBG_MSG_P(CH3_OTHER,VERBOSE,"request allocated, handle=0x%08x", sreq->handle);
}
}
);
int MPID_Isend(const void * buf, int count, MPI_Datatype datatype, int rank,
int tag, MPID_Comm * comm, int context_offset,
MPID_Request ** request)
{
MPIDI_msg_sz_t data_sz;
int dt_contig;
MPI_Aint dt_true_lb;
MPID_Datatype * dt_ptr;
MPID_Request * sreq;
MPIDI_VC_t * vc=0;
#if defined(MPID_USE_SEQUENCE_NUMBERS)
MPID_Seqnum_t seqnum;
#endif
int eager_threshold = -1;
int mpi_errno = MPI_SUCCESS;
MPIDI_STATE_DECL(MPID_STATE_MPID_ISEND);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_ISEND);
MPIU_DBG_MSG_FMT(CH3_OTHER,VERBOSE,(MPIU_DBG_FDEST,
"rank=%d, tag=%d, context=%d",
rank, tag, comm->context_id + context_offset));
if (rank == comm->rank && comm->comm_kind != MPID_INTERCOMM)
{
mpi_errno = MPIDI_Isend_self(buf, count, datatype, rank, tag, comm,
context_offset, MPIDI_REQUEST_TYPE_SEND, &sreq);
goto fn_exit;
}
if (rank != MPI_PROC_NULL) {
MPIDI_Comm_get_vc_set_active(comm, rank, &vc);
#ifdef ENABLE_COMM_OVERRIDES
/* this needs to come before the sreq is created, since the override
* function is responsible for creating its own request */
if (vc->comm_ops && vc->comm_ops->isend)
{
mpi_errno = vc->comm_ops->isend( vc, buf, count, datatype, rank, tag, comm, context_offset, &sreq);
goto fn_exit;
}
#endif
}
MPIDI_Request_create_sreq(sreq, mpi_errno, goto fn_exit);
MPIDI_Request_set_type(sreq, MPIDI_REQUEST_TYPE_SEND);
if (rank == MPI_PROC_NULL)
{
MPIU_Object_set_ref(sreq, 1);
MPID_cc_set(&sreq->cc, 0);
goto fn_exit;
}
MPIDI_Datatype_get_info(count, datatype, dt_contig, data_sz, dt_ptr,
dt_true_lb);
if (data_sz == 0)
{
MPIDI_CH3_Pkt_t upkt;
MPIDI_CH3_Pkt_eager_send_t * const eager_pkt = &upkt.eager_send;
MPIDI_Request_set_msg_type(sreq, MPIDI_REQUEST_EAGER_MSG);
sreq->dev.OnDataAvail = 0;
MPIU_DBG_MSG(CH3_OTHER,VERBOSE,"sending zero length message");
MPIDI_Pkt_init(eager_pkt, MPIDI_CH3_PKT_EAGER_SEND);
eager_pkt->match.parts.rank = comm->rank;
eager_pkt->match.parts.tag = tag;
eager_pkt->match.parts.context_id = comm->context_id + context_offset;
eager_pkt->sender_req_id = sreq->handle;
eager_pkt->data_sz = 0;
MPIDI_VC_FAI_send_seqnum(vc, seqnum);
MPIDI_Pkt_set_seqnum(eager_pkt, seqnum);
MPIDI_Request_set_seqnum(sreq, seqnum);
MPIU_THREAD_CS_ENTER(CH3COMM,vc);
mpi_errno = MPIDI_CH3_iSend(vc, sreq, eager_pkt, sizeof(*eager_pkt));
MPIU_THREAD_CS_EXIT(CH3COMM,vc);
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno != MPI_SUCCESS)
{
MPIU_Object_set_ref(sreq, 0);
MPIDI_CH3_Request_destroy(sreq);
sreq = NULL;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**ch3|eagermsg");
goto fn_exit;
}
/* --END ERROR HANDLING-- */
goto fn_exit;
}
MPIDI_CH3_GET_EAGER_THRESHOLD(&eager_threshold, comm, vc);
/* FIXME: flow control: limit number of outstanding eager messages
containing data and need to be buffered by the receiver */
if (data_sz + sizeof(MPIDI_CH3_Pkt_eager_send_t) <= eager_threshold)
{
if (dt_contig)
{
mpi_errno = MPIDI_CH3_EagerContigIsend( &sreq,
MPIDI_CH3_PKT_EAGER_SEND,
(char*)buf + dt_true_lb,
data_sz, rank, tag,
comm, context_offset );
}
else
{
mpi_errno = MPIDI_CH3_EagerNoncontigSend( &sreq,
MPIDI_CH3_PKT_EAGER_SEND,
buf, count, datatype,
data_sz, rank, tag,
comm, context_offset );
/* If we're not complete, then add a reference to the datatype */
if (sreq && sreq->dev.OnDataAvail) {
sreq->dev.datatype_ptr = dt_ptr;
MPID_Datatype_add_ref(dt_ptr);
}
}
}
else
{
/* Note that the sreq was created above */
MPIDI_Request_set_msg_type( sreq, MPIDI_REQUEST_RNDV_MSG );
mpi_errno = vc->rndvSend_fn( &sreq, buf, count, datatype, dt_contig,
data_sz, dt_true_lb, rank, tag, comm,
context_offset );
/* FIXME: fill temporary IOV or pack temporary buffer after send to
hide some latency. This requires synchronization
because the CTS packet could arrive and be processed before the
above iStartmsg completes (depending on the progress
engine, threads, etc.). */
if (sreq && dt_ptr != NULL)
{
sreq->dev.datatype_ptr = dt_ptr;
MPID_Datatype_add_ref(dt_ptr);
}
}
fn_exit:
*request = sreq;
MPIU_DBG_STMT(CH3_OTHER,VERBOSE,
{
if (sreq != NULL)
{
MPIU_DBG_MSG_P(CH3_OTHER,VERBOSE,"request allocated, handle=0x%08x", sreq->handle);
}
}
);
int MPIDI_Isend_self(const void * buf, int count, MPI_Datatype datatype, int rank, int tag, MPID_Comm * comm, int context_offset,
int type, MPID_Request ** request)
{
MPIDI_Message_match match;
MPID_Request * sreq;
MPID_Request * rreq;
MPIDI_VC_t * vc;
#if defined(MPID_USE_SEQUENCE_NUMBERS)
MPID_Seqnum_t seqnum;
#endif
int found;
int mpi_errno = MPI_SUCCESS;
MPIU_DBG_MSG(CH3_OTHER,VERBOSE,"sending message to self");
MPIDI_Request_create_sreq(sreq, mpi_errno, goto fn_exit);
MPIDI_Request_set_type(sreq, type);
MPIDI_Request_set_msg_type(sreq, MPIDI_REQUEST_SELF_MSG);
match.parts.rank = rank;
match.parts.tag = tag;
match.parts.context_id = comm->context_id + context_offset;
MPIU_THREAD_CS_ENTER(MSGQUEUE,);
rreq = MPIDI_CH3U_Recvq_FDP_or_AEU(&match, &found);
/* --BEGIN ERROR HANDLING-- */
if (rreq == NULL)
{
MPIU_Object_set_ref(sreq, 0);
MPIDI_CH3_Request_destroy(sreq);
sreq = NULL;
MPIU_ERR_SET1(mpi_errno, MPI_ERR_OTHER, "**nomem",
"**nomemuereq %d", MPIDI_CH3U_Recvq_count_unexp());
goto fn_exit;
}
/* --END ERROR HANDLING-- */
MPIDI_Comm_get_vc_set_active(comm, rank, &vc);
MPIDI_VC_FAI_send_seqnum(vc, seqnum);
MPIDI_Request_set_seqnum(sreq, seqnum);
MPIDI_Request_set_seqnum(rreq, seqnum);
rreq->status.MPI_SOURCE = rank;
rreq->status.MPI_TAG = tag;
if (found)
{
MPIDI_msg_sz_t data_sz;
/* we found a posted req, which we now own, so we can release the CS */
MPIU_THREAD_CS_EXIT(MSGQUEUE,);
MPIU_DBG_MSG(CH3_OTHER,VERBOSE,
"found posted receive request; copying data");
MPIDI_CH3U_Buffer_copy(buf, count, datatype, &sreq->status.MPI_ERROR,
rreq->dev.user_buf, rreq->dev.user_count, rreq->dev.datatype, &data_sz, &rreq->status.MPI_ERROR);
rreq->status.count = (int)data_sz;
MPID_REQUEST_SET_COMPLETED(rreq);
MPID_Request_release(rreq);
/* sreq has never been seen by the user or outside this thread, so it is safe to reset ref_count and cc */
MPIU_Object_set_ref(sreq, 1);
MPID_cc_set(&sreq->cc, 0);
}
else
{
if (type != MPIDI_REQUEST_TYPE_RSEND)
int MPID_Issend(const void * buf, int count, MPI_Datatype datatype, int rank, int tag, MPID_Comm * comm, int context_offset,
MPID_Request ** request)
{
MPIDI_msg_sz_t data_sz;
int dt_contig;
MPI_Aint dt_true_lb;
MPID_Datatype * dt_ptr;
MPID_Request * sreq;
MPIDI_VC_t * vc=0;
#if defined(MPID_USE_SEQUENCE_NUMBERS)
MPID_Seqnum_t seqnum;
#endif
int eager_threshold = -1;
int mpi_errno = MPI_SUCCESS;
MPIDI_STATE_DECL(MPID_STATE_MPID_ISSEND);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_ISSEND);
MPIU_DBG_MSG_FMT(CH3_OTHER,VERBOSE,(MPIU_DBG_FDEST,
"rank=%d, tag=%d, context=%d",
rank, tag, comm->context_id + context_offset));
if (rank == comm->rank && comm->comm_kind != MPID_INTERCOMM)
{
mpi_errno = MPIDI_Isend_self(buf, count, datatype, rank, tag, comm, context_offset, MPIDI_REQUEST_TYPE_SSEND, &sreq);
goto fn_exit;
}
if (rank != MPI_PROC_NULL)
{
MPIDI_Comm_get_vc_set_active(comm, rank, &vc);
/* this needs to come before the sreq is created, since the override */
/* function is responsible for creating its own request */
#ifdef ENABLE_COMM_OVERRIDES
if (vc->comm_ops && vc->comm_ops->issend)
{
mpi_errno = vc->comm_ops->issend( vc, buf, count, datatype, rank, tag, comm, context_offset, &sreq);
goto fn_exit;
}
#endif
}
MPIDI_Request_create_sreq(sreq, mpi_errno, goto fn_exit);
MPIDI_Request_set_type(sreq, MPIDI_REQUEST_TYPE_SSEND);
if (rank == MPI_PROC_NULL)
{
MPIU_Object_set_ref(sreq, 1);
MPID_cc_set(&sreq->cc, 0);
goto fn_exit;
}
MPIDI_Datatype_get_info(count, datatype, dt_contig, data_sz, dt_ptr, dt_true_lb);
if (data_sz == 0)
{
mpi_errno = MPIDI_CH3_EagerSyncZero( &sreq, rank, tag, comm,
context_offset );
goto fn_exit;
}
MPIDI_CH3_GET_EAGER_THRESHOLD(&eager_threshold, comm, vc);
if (data_sz + sizeof(MPIDI_CH3_Pkt_eager_sync_send_t) <= eager_threshold)
{
mpi_errno = MPIDI_CH3_EagerSyncNoncontigSend( &sreq, buf, count,
datatype, data_sz,
dt_contig, dt_true_lb,
rank, tag, comm,
context_offset );
/* If we're not complete, then add a reference to the datatype */
if (sreq && sreq->dev.OnDataAvail) {
sreq->dev.datatype_ptr = dt_ptr;
MPID_Datatype_add_ref(dt_ptr);
}
}
else
{
/* Note that the sreq was created above */
MPIDI_Request_set_msg_type(sreq, MPIDI_REQUEST_RNDV_MSG);
mpi_errno = vc->rndvSend_fn( &sreq, buf, count, datatype, dt_contig,
data_sz, dt_true_lb, rank, tag, comm,
context_offset );
/* FIXME: fill temporary IOV or pack temporary buffer after send to
hide some latency. This requires synchronization
because the CTS packet could arrive and be processed before the
above iStartmsg completes (depending on the progress
engine, threads, etc.). */
if (sreq && dt_ptr != NULL)
{
sreq->dev.datatype_ptr = dt_ptr;
MPID_Datatype_add_ref(dt_ptr);
}
}
fn_exit:
*request = sreq;
MPIU_DBG_STMT(CH3_OTHER,VERBOSE,
{
if (sreq != NULL) {
MPIU_DBG_MSG_P(CH3_OTHER,VERBOSE,
"request allocated, handle=0x%08x", sreq->handle);
}
}
)
int MPID_Irsend(const void * buf, int count, MPI_Datatype datatype, int rank, int tag, MPID_Comm * comm, int context_offset,
MPID_Request ** request)
{
MPIDI_CH3_Pkt_t upkt;
MPIDI_CH3_Pkt_ready_send_t * const ready_pkt = &upkt.ready_send;
MPIDI_msg_sz_t data_sz;
int dt_contig;
MPI_Aint dt_true_lb;
MPID_Datatype * dt_ptr;
MPID_Request * sreq;
MPIDI_VC_t * vc;
#if defined(MPID_USE_SEQUENCE_NUMBERS)
MPID_Seqnum_t seqnum;
#endif
int mpi_errno = MPI_SUCCESS;
MPIDI_STATE_DECL(MPID_STATE_MPID_IRSEND);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_IRSEND);
MPIU_DBG_MSG_FMT(CH3_OTHER,VERBOSE,(MPIU_DBG_FDEST,
"rank=%d, tag=%d, context=%d",
rank, tag, comm->context_id + context_offset));
if (rank == comm->rank && comm->comm_kind != MPID_INTERCOMM)
{
mpi_errno = MPIDI_Isend_self(buf, count, datatype, rank, tag, comm, context_offset, MPIDI_REQUEST_TYPE_RSEND, &sreq);
goto fn_exit;
}
MPIDI_Request_create_sreq(sreq, mpi_errno, goto fn_exit);
MPIDI_Request_set_type(sreq, MPIDI_REQUEST_TYPE_RSEND);
MPIDI_Request_set_msg_type(sreq, MPIDI_REQUEST_EAGER_MSG);
if (rank == MPI_PROC_NULL)
{
MPIU_Object_set_ref(sreq, 1);
MPID_cc_set(&sreq->cc, 0);
goto fn_exit;
}
MPIDI_Comm_get_vc_set_active(comm, rank, &vc);
#ifdef ENABLE_COMM_OVERRIDES
if (vc->comm_ops && vc->comm_ops->irsend)
{
mpi_errno = vc->comm_ops->irsend( vc, buf, count, datatype, rank, tag, comm, context_offset, &sreq);
goto fn_exit;
}
#endif
MPIDI_Datatype_get_info(count, datatype, dt_contig, data_sz, dt_ptr, dt_true_lb);
MPIDI_Pkt_init(ready_pkt, MPIDI_CH3_PKT_READY_SEND);
ready_pkt->match.parts.rank = comm->rank;
ready_pkt->match.parts.tag = tag;
ready_pkt->match.parts.context_id = comm->context_id + context_offset;
ready_pkt->sender_req_id = MPI_REQUEST_NULL;
ready_pkt->data_sz = data_sz;
if (data_sz == 0)
{
MPIU_DBG_MSG(CH3_OTHER,VERBOSE,"sending zero length message");
sreq->dev.OnDataAvail = 0;
MPIDI_VC_FAI_send_seqnum(vc, seqnum);
MPIDI_Pkt_set_seqnum(ready_pkt, seqnum);
MPIDI_Request_set_seqnum(sreq, seqnum);
MPIU_THREAD_CS_ENTER(CH3COMM,vc);
mpi_errno = MPIU_CALL(MPIDI_CH3,iSend(vc, sreq, ready_pkt, sizeof(*ready_pkt)));
MPIU_THREAD_CS_EXIT(CH3COMM,vc);
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno != MPI_SUCCESS)
{
MPIU_Object_set_ref(sreq, 0);
MPIDI_CH3_Request_destroy(sreq);
sreq = NULL;
MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**ch3|eagermsg");
goto fn_exit;
}
/* --END ERROR HANDLING-- */
goto fn_exit;
}
if (dt_contig) {
mpi_errno = MPIDI_CH3_EagerContigIsend( &sreq,
MPIDI_CH3_PKT_READY_SEND,
(char*)buf + dt_true_lb,
data_sz, rank, tag,
comm, context_offset );
}
else {
mpi_errno = MPIDI_CH3_EagerNoncontigSend( &sreq,
MPIDI_CH3_PKT_READY_SEND,
buf, count, datatype,
data_sz, rank, tag,
comm, context_offset );
/* If we're not complete, then add a reference to the datatype */
if (sreq && sreq->dev.OnDataAvail) {
sreq->dev.datatype_ptr = dt_ptr;
MPID_Datatype_add_ref(dt_ptr);
}
}
fn_exit:
*request = sreq;
MPIU_DBG_STMT(CH3_OTHER,VERBOSE,{
if (sreq != NULL)
{
MPIU_DBG_MSG_P(CH3_OTHER,VERBOSE,"request allocated, handle=0x%08x", sreq->handle);
}
}
);
/* Complete a request, saving the status data if necessary.
"active" has meaning only if the request is a persistent request; this
allows the completion routines to indicate that a persistent request
was inactive and did not require any extra completion operation.
If debugger information is being provided for pending (user-initiated)
send operations, the macros MPIR_SENDQ_FORGET will be defined to
call the routine MPIR_Sendq_forget; otherwise that macro will be a no-op.
The implementation of the MPIR_Sendq_xxx is in src/mpi/debugger/dbginit.c .
*/
int MPIR_Request_complete(MPI_Request * request, MPID_Request * request_ptr,
MPI_Status * status, int * active)
{
int mpi_errno = MPI_SUCCESS;
*active = TRUE;
switch(request_ptr->kind)
{
case MPID_REQUEST_SEND:
{
if (status != MPI_STATUS_IGNORE)
{
MPIR_STATUS_SET_CANCEL_BIT(*status, MPIR_STATUS_GET_CANCEL_BIT(request_ptr->status));
}
mpi_errno = request_ptr->status.MPI_ERROR;
MPIR_SENDQ_FORGET(request_ptr);
MPID_Request_release(request_ptr);
if (NULL != request) *request = MPI_REQUEST_NULL;
break;
}
case MPID_REQUEST_RECV:
{
MPIR_Request_extract_status(request_ptr, status);
mpi_errno = request_ptr->status.MPI_ERROR;
MPID_Request_release(request_ptr);
if (NULL != request) *request = MPI_REQUEST_NULL;
break;
}
case MPID_PREQUEST_SEND:
{
if (request_ptr->partner_request != NULL)
{
MPID_Request * prequest_ptr = request_ptr->partner_request;
/* reset persistent request to inactive state */
MPID_cc_set(&request_ptr->cc, 0);
request_ptr->cc_ptr = &request_ptr->cc;
request_ptr->partner_request = NULL;
if (prequest_ptr->kind != MPID_UREQUEST)
{
if (status != MPI_STATUS_IGNORE)
{
MPIR_STATUS_SET_CANCEL_BIT(*status, MPIR_STATUS_GET_CANCEL_BIT(prequest_ptr->status));
}
mpi_errno = prequest_ptr->status.MPI_ERROR;
}
else
{
/* This is needed for persistent Bsend requests */
int rc;
rc = MPIR_Grequest_query(prequest_ptr);
if (mpi_errno == MPI_SUCCESS)
{
mpi_errno = rc;
}
if (status != MPI_STATUS_IGNORE)
{
MPIR_STATUS_SET_CANCEL_BIT(*status, MPIR_STATUS_GET_CANCEL_BIT(prequest_ptr->status));
}
if (mpi_errno == MPI_SUCCESS)
{
mpi_errno = prequest_ptr->status.MPI_ERROR;
}
rc = MPIR_Grequest_free(prequest_ptr);
if (mpi_errno == MPI_SUCCESS)
{
mpi_errno = rc;
}
}
MPID_Request_release(prequest_ptr);
}
else
{
if (request_ptr->status.MPI_ERROR != MPI_SUCCESS)
{
/* if the persistent request failed to start then make the
error code available */
if (status != MPI_STATUS_IGNORE)
{
MPIR_STATUS_SET_CANCEL_BIT(*status, FALSE);
}
mpi_errno = request_ptr->status.MPI_ERROR;
}
else
{
MPIR_Status_set_empty(status);
*active = FALSE;
}
}
break;
}
case MPID_PREQUEST_RECV:
{
if (request_ptr->partner_request != NULL)
{
MPID_Request * prequest_ptr = request_ptr->partner_request;
/* reset persistent request to inactive state */
MPID_cc_set(&request_ptr->cc, 0);
request_ptr->cc_ptr = &request_ptr->cc;
request_ptr->partner_request = NULL;
MPIR_Request_extract_status(prequest_ptr, status);
mpi_errno = prequest_ptr->status.MPI_ERROR;
MPID_Request_release(prequest_ptr);
}
else
{
MPIR_Status_set_empty(status);
/* --BEGIN ERROR HANDLING-- */
if (request_ptr->status.MPI_ERROR != MPI_SUCCESS)
{
/* if the persistent request failed to start then make the
error code available */
mpi_errno = request_ptr->status.MPI_ERROR;
}
else
{
*active = FALSE;
}
/* --END ERROR HANDLING-- */
}
break;
}
case MPID_UREQUEST:
{
int rc;
rc = MPIR_Grequest_query(request_ptr);
if (mpi_errno == MPI_SUCCESS)
{
mpi_errno = rc;
}
MPIR_Request_extract_status(request_ptr, status);
rc = MPIR_Grequest_free(request_ptr);
if (mpi_errno == MPI_SUCCESS)
{
mpi_errno = rc;
}
MPID_Request_release(request_ptr);
if (NULL != request) *request = MPI_REQUEST_NULL;
break;
}
case MPID_COLL_REQUEST:
case MPID_WIN_REQUEST:
{
mpi_errno = request_ptr->status.MPI_ERROR;
MPIR_Request_extract_status(request_ptr, status);
MPID_Request_release(request_ptr);
if (NULL != request) *request = MPI_REQUEST_NULL;
break;
}
default:
{
/* --BEGIN ERROR HANDLING-- */
/* This should not happen */
MPIR_ERR_SETANDSTMT1(mpi_errno, MPI_ERR_INTERN,;, "**badcase",
"**badcase %d", request_ptr->kind);
break;
/* --END ERROR HANDLING-- */
}
}
int MPID_Startall(int count, MPID_Request * requests[])
{
int i;
int rc;
int mpi_errno = MPI_SUCCESS;
MPIDI_STATE_DECL(MPID_STATE_MPID_STARTALL);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_STARTALL);
for (i = 0; i < count; i++)
{
MPID_Request * const preq = requests[i];
/* FIXME: The odd 7th arg (match.context_id - comm->context_id)
is probably to get the context offset. Do we really need the
context offset? Is there any case where the offset isn't zero? */
switch (MPIDI_Request_get_type(preq))
{
case MPIDI_REQUEST_TYPE_RECV:
{
rc = MPID_Irecv(preq->dev.user_buf, preq->dev.user_count, preq->dev.datatype, preq->dev.match.parts.rank,
preq->dev.match.parts.tag, preq->comm, preq->dev.match.parts.context_id - preq->comm->recvcontext_id,
&preq->partner_request);
break;
}
case MPIDI_REQUEST_TYPE_SEND:
{
rc = MPID_Isend(preq->dev.user_buf, preq->dev.user_count, preq->dev.datatype, preq->dev.match.parts.rank,
preq->dev.match.parts.tag, preq->comm, preq->dev.match.parts.context_id - preq->comm->context_id,
&preq->partner_request);
break;
}
case MPIDI_REQUEST_TYPE_RSEND:
{
rc = MPID_Irsend(preq->dev.user_buf, preq->dev.user_count, preq->dev.datatype, preq->dev.match.parts.rank,
preq->dev.match.parts.tag, preq->comm, preq->dev.match.parts.context_id - preq->comm->context_id,
&preq->partner_request);
break;
}
case MPIDI_REQUEST_TYPE_SSEND:
{
rc = MPID_Issend(preq->dev.user_buf, preq->dev.user_count, preq->dev.datatype, preq->dev.match.parts.rank,
preq->dev.match.parts.tag, preq->comm, preq->dev.match.parts.context_id - preq->comm->context_id,
&preq->partner_request);
break;
}
case MPIDI_REQUEST_TYPE_BSEND:
{
MPI_Request sreq_handle;
rc = MPIR_Ibsend_impl(preq->dev.user_buf, preq->dev.user_count,
preq->dev.datatype, preq->dev.match.parts.rank,
preq->dev.match.parts.tag, preq->comm,
&sreq_handle);
if (rc == MPI_SUCCESS)
{
MPID_Request_get_ptr(sreq_handle, preq->partner_request);
}
break;
}
default:
{
/* --BEGIN ERROR HANDLING-- */
rc = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_FATAL, FCNAME, __LINE__, MPI_ERR_INTERN, "**ch3|badreqtype",
"**ch3|badreqtype %d", MPIDI_Request_get_type(preq));
/* --END ERROR HANDLING-- */
}
}
if (rc == MPI_SUCCESS)
{
preq->status.MPI_ERROR = MPI_SUCCESS;
preq->cc_ptr = &preq->partner_request->cc;
}
/* --BEGIN ERROR HANDLING-- */
else
{
/* If a failure occurs attempting to start the request, then we
assume that partner request was not created, and stuff
the error code in the persistent request. The wait and test
routines will look at the error code in the persistent
request if a partner request is not present. */
preq->partner_request = NULL;
preq->status.MPI_ERROR = rc;
preq->cc_ptr = &preq->cc;
MPID_cc_set(&preq->cc, 0);
}
/* --END ERROR HANDLING-- */
}
MPIDI_FUNC_EXIT(MPID_STATE_MPID_STARTALL);
return mpi_errno;
}
/* MPIDI_CH3_EagerSyncNoncontigSend - Eagerly send noncontiguous data in
synchronous mode.
Some implementations may choose to use Rendezvous sends (see ch3u_rndv.c)
for all Synchronous sends (MPI_Issend and MPI_Ssend). An eager
synchronous send eliminates one of the handshake messages, but
most application codes should not be using synchronous sends in
performance-critical operations.
*/
int MPIDI_CH3_EagerSyncNoncontigSend( MPID_Request **sreq_p,
const void * buf, int count,
MPI_Datatype datatype, MPIDI_msg_sz_t data_sz,
int dt_contig, MPI_Aint dt_true_lb,
int rank,
int tag, MPID_Comm * comm,
int context_offset )
{
int mpi_errno = MPI_SUCCESS;
MPIDI_CH3_Pkt_t upkt;
MPIDI_CH3_Pkt_eager_sync_send_t * const es_pkt = &upkt.eager_sync_send;
MPIDI_VC_t * vc;
MPID_Request *sreq = *sreq_p;
/* MT FIXME what are the two operations we are waiting for? the send and
* the sync response? */
MPID_cc_set(&sreq->cc, 2);
sreq->dev.OnDataAvail = 0;
sreq->dev.OnFinal = 0;
MPIDI_Pkt_init(es_pkt, MPIDI_CH3_PKT_EAGER_SYNC_SEND);
es_pkt->match.parts.rank = comm->rank;
es_pkt->match.parts.tag = tag;
es_pkt->match.parts.context_id = comm->context_id + context_offset;
es_pkt->sender_req_id = sreq->handle;
es_pkt->data_sz = data_sz;
MPIDI_Comm_get_vc_set_active(comm, rank, &vc);
MPIDI_VC_FAI_send_seqnum(vc, seqnum);
MPIDI_Pkt_set_seqnum(es_pkt, seqnum);
MPIDI_Request_set_seqnum(sreq, seqnum);
MPIU_DBG_MSGPKT(vc,tag,es_pkt->match.parts.context_id,rank,data_sz,"EagerSync");
if (dt_contig)
{
MPL_IOV iov[2];
MPIU_DBG_MSG_FMT(CH3_OTHER,VERBOSE,(MPIU_DBG_FDEST,
"sending contiguous sync eager message, data_sz=" MPIDI_MSG_SZ_FMT,
data_sz));
iov[0].MPL_IOV_BUF = (MPL_IOV_BUF_CAST)es_pkt;
iov[0].MPL_IOV_LEN = sizeof(*es_pkt);
iov[1].MPL_IOV_BUF = (MPL_IOV_BUF_CAST) ((char *)buf + dt_true_lb);
iov[1].MPL_IOV_LEN = data_sz;
MPID_THREAD_CS_ENTER(POBJ, vc->pobj_mutex);
mpi_errno = MPIDI_CH3_iSendv(vc, sreq, iov, 2);
MPID_THREAD_CS_EXIT(POBJ, vc->pobj_mutex);
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno != MPI_SUCCESS)
{
/* Make sure to destroy the request before setting the pointer to
* NULL, otherwise we lose the handle on the request */
MPID_Request_release(sreq);
*sreq_p = NULL;
MPIR_ERR_SETANDJUMP(mpi_errno, MPI_ERR_OTHER, "**ch3|eagermsg");
}
/* --END ERROR HANDLING-- */
}
else
{
MPIU_DBG_MSG_D(CH3_OTHER,VERBOSE,
"sending non-contiguous sync eager message, data_sz=" MPIDI_MSG_SZ_FMT,
data_sz);
sreq->dev.segment_ptr = MPID_Segment_alloc( );
MPIR_ERR_CHKANDJUMP1((sreq->dev.segment_ptr == NULL), mpi_errno, MPI_ERR_OTHER, "**nomem", "**nomem %s", "MPID_Segment_alloc");
MPID_Segment_init(buf, count, datatype, sreq->dev.segment_ptr, 0);
sreq->dev.segment_first = 0;
sreq->dev.segment_size = data_sz;
MPID_THREAD_CS_ENTER(POBJ, vc->pobj_mutex);
mpi_errno = vc->sendNoncontig_fn(vc, sreq, es_pkt, sizeof(MPIDI_CH3_Pkt_eager_sync_send_t));
MPID_THREAD_CS_EXIT(POBJ, vc->pobj_mutex);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
fn_exit:
return mpi_errno;
fn_fail:
*sreq_p = NULL;
goto fn_exit;
}
