/* * Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana * University Research and Technology * Corporation. All rights reserved. * Copyright (c) 2004-2006 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) 2004-2005 The Regents of the University of California. * All rights reserved. * $COPYRIGHT$ * * Additional copyrights may follow * * $HEADER$ */ #include "ompi_config.h" #include "opal/prefetch.h" #include "ompi/constants.h" #include "ompi/mca/pml/pml.h" #include "ompi/mca/btl/btl.h" #include "orte/mca/errmgr/errmgr.h" #include "ompi/mca/mpool/mpool.h" #include "pml_ob1.h" #include "pml_ob1_hdr.h" #include "pml_ob1_sendreq.h" #include "pml_ob1_rdmafrag.h" #include "pml_ob1_recvreq.h" #include "ompi/mca/bml/base/base.h" #include "ompi/datatype/dt_arch.h" OBJ_CLASS_INSTANCE(mca_pml_ob1_send_range_t, ompi_free_list_item_t, NULL, NULL); void mca_pml_ob1_send_request_process_pending(mca_bml_base_btl_t *bml_btl) { int i, s = opal_list_get_size(&mca_pml_ob1.send_pending); /* advance pending requests */ for(i = 0; i < s; i++) { mca_pml_ob1_send_pending_t pending_type; mca_pml_ob1_send_request_t* sendreq; mca_bml_base_btl_t *send_dst; OPAL_THREAD_LOCK(&mca_pml_ob1.lock); sendreq = (mca_pml_ob1_send_request_t*) opal_list_remove_first(&mca_pml_ob1.send_pending); OPAL_THREAD_UNLOCK(&mca_pml_ob1.lock); if(NULL == sendreq) break; pending_type = sendreq->req_pending; sendreq->req_pending = MCA_PML_OB1_SEND_PENDING_NONE; switch(pending_type) { case MCA_PML_OB1_SEND_PENDING_SCHEDULE: if(mca_pml_ob1_send_request_schedule_exclusive(sendreq) == OMPI_ERR_OUT_OF_RESOURCE) { return; } break; case MCA_PML_OB1_SEND_PENDING_START: send_dst = mca_bml_base_btl_array_find( &sendreq->req_endpoint->btl_eager, bml_btl->btl); if(NULL == send_dst || mca_pml_ob1_send_request_start_btl(sendreq, send_dst) == OMPI_ERR_OUT_OF_RESOURCE) { /* if dst of this sendreq cannot be reached through the * endpoint or no resources put request back on the list */ OPAL_THREAD_LOCK(&mca_pml_ob1.lock); sendreq->req_pending = MCA_PML_OB1_SEND_PENDING_START; if(NULL == send_dst) { opal_list_append(&mca_pml_ob1.send_pending, (opal_list_item_t*)sendreq); } else { /* prepend to the pending list to minimize reordering */ opal_list_prepend(&mca_pml_ob1.send_pending, (opal_list_item_t*)sendreq); } OPAL_THREAD_UNLOCK(&mca_pml_ob1.lock); /* if no destination try next request otherwise give up, * no more resources on this btl */ if(send_dst != NULL) return; } break; default: opal_output(0, "[%s:%d] wrong send request type\n", __FILE__, __LINE__); break; } } } /* * 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; assert( false == sendreq->req_send.req_base.req_free_called ); OPAL_THREAD_LOCK(&ompi_request_lock); 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 ) { MCA_PML_OB1_SEND_REQUEST_RETURN( sendreq ); } OPAL_THREAD_UNLOCK(&ompi_request_lock); *request = MPI_REQUEST_NULL; return OMPI_SUCCESS; } static int mca_pml_ob1_send_request_cancel(struct ompi_request_t* request, int complete) { /* we dont cancel send requests by now */ return OMPI_SUCCESS; } static void mca_pml_ob1_send_request_construct(mca_pml_ob1_send_request_t* req) { req->req_send.req_base.req_type = MCA_PML_REQUEST_SEND; req->req_send.req_base.req_ompi.req_free = mca_pml_ob1_send_request_free; req->req_send.req_base.req_ompi.req_cancel = mca_pml_ob1_send_request_cancel; req->req_rdma_cnt = 0; req->req_throttle_sends = false; OBJ_CONSTRUCT(&req->req_send_ranges, opal_list_t); } OBJ_CLASS_INSTANCE( mca_pml_ob1_send_request_t, mca_pml_base_send_request_t, mca_pml_ob1_send_request_construct, NULL); /** * Completion of a short message - nothing left to schedule. Note that this * function is only called for 0 sized messages. */ void mca_pml_ob1_match_completion_cache( struct mca_btl_base_module_t* btl, struct mca_btl_base_endpoint_t* ep, struct mca_btl_base_descriptor_t* descriptor, int status) { mca_pml_ob1_send_request_t* sendreq = (mca_pml_ob1_send_request_t*)descriptor->des_cbdata; mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*) descriptor->des_context; if( sendreq->req_send.req_bytes_packed > 0 ) { PERUSE_TRACE_COMM_EVENT( PERUSE_COMM_REQ_XFER_BEGIN, &(sendreq->req_send.req_base), PERUSE_SEND ); } /* check completion status */ if(OMPI_SUCCESS != status) { /* TSW - FIX */ opal_output(0, "%s:%d FATAL", __FILE__, __LINE__); orte_errmgr.abort(); } /* attempt to cache the descriptor */ MCA_BML_BASE_BTL_DES_RETURN( bml_btl, descriptor ); /* signal request completion */ MCA_PML_OB1_SEND_REQUEST_PML_COMPLETE(sendreq); /* check for pending requests */ MCA_PML_OB1_PROGRESS_PENDING(bml_btl); } /** * Completion of a short message - nothing left to schedule. */ void mca_pml_ob1_match_completion_free( struct mca_btl_base_module_t* btl, struct mca_btl_base_endpoint_t* ep, struct mca_btl_base_descriptor_t* descriptor, int status) { mca_pml_ob1_send_request_t* sendreq = (mca_pml_ob1_send_request_t*)descriptor->des_cbdata; mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*) descriptor->des_context; if( sendreq->req_send.req_bytes_packed > 0 ) { PERUSE_TRACE_COMM_EVENT( PERUSE_COMM_REQ_XFER_BEGIN, &(sendreq->req_send.req_base), PERUSE_SEND ); } /* check completion status */ if(OMPI_SUCCESS != status) { /* TSW - FIX */ opal_output(0, "%s:%d FATAL", __FILE__, __LINE__); orte_errmgr.abort(); } /* free the descriptor */ mca_bml_base_free( bml_btl, descriptor ); /* signal request completion */ MCA_PML_OB1_SEND_REQUEST_PML_COMPLETE(sendreq); /* check for pending requests */ MCA_PML_OB1_PROGRESS_PENDING(bml_btl); } /* * Completion of the first fragment of a long message that * requires an acknowledgement */ static void mca_pml_ob1_rndv_completion( mca_btl_base_module_t* btl, struct mca_btl_base_endpoint_t* ep, struct mca_btl_base_descriptor_t* descriptor, int status) { mca_pml_ob1_send_request_t* sendreq = (mca_pml_ob1_send_request_t*)descriptor->des_cbdata; mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*)descriptor->des_context; size_t req_bytes_delivered = 0; if( sendreq->req_send.req_bytes_packed > 0 ) { PERUSE_TRACE_COMM_EVENT( PERUSE_COMM_REQ_XFER_BEGIN, &(sendreq->req_send.req_base), PERUSE_SEND ); } /* check completion status */ if(OMPI_SUCCESS != status) { /* TSW - FIX */ opal_output(0, "%s:%d FATAL", __FILE__, __LINE__); orte_errmgr.abort(); } /* count bytes of user data actually delivered. As the rndv completion only * happens in one thread, the increase of the req_bytes_delivered does not * have to be atomic. */ MCA_PML_OB1_COMPUTE_SEGMENT_LENGTH( descriptor->des_src, descriptor->des_src_cnt, sizeof(mca_pml_ob1_rendezvous_hdr_t), req_bytes_delivered ); OPAL_THREAD_ADD_SIZE_T(&sendreq->req_bytes_delivered, req_bytes_delivered); /* return the descriptor */ mca_bml_base_free(bml_btl, descriptor); /* advance the request */ if(OPAL_THREAD_ADD32(&sendreq->req_state, 1) == 2 && sendreq->req_bytes_delivered >= sendreq->req_send.req_bytes_packed) { if(!sendreq->req_send.req_base.req_pml_complete){ /* We must check that completion hasn't already occured */ /* for the self BTL we may choose the RDMA PUT protocol */ /* on the send side, in this case we send no eager data */ /* if, on the receiver side the data is not contiguous we */ /* may choose to use the copy in/out protocol */ /* if this occurs, the entire request can be completed in a */ /* single call to mca_pml_ob1_recv_request_ack */ /* as soon as the last fragment of the copy in/out protocol */ /* gets local completion. This doesn't occur in the general */ /* case of the copy in/out protocol because when both sender */ /* and receiver agree on the copy in/out protoocol we eagerly */ /* send data, we don't update the request with this eagerly sent */ /* data until here in this function, so completion could not have */ /* yet occurred. */ MCA_PML_OB1_SEND_REQUEST_PML_COMPLETE(sendreq); } } /* check for pending requests */ MCA_PML_OB1_PROGRESS_PENDING(bml_btl); } /** * Completion of a get request. */ static void mca_pml_ob1_rget_completion( mca_btl_base_module_t* btl, struct mca_btl_base_endpoint_t* ep, struct mca_btl_base_descriptor_t* des, int status) { mca_pml_ob1_send_request_t* sendreq = (mca_pml_ob1_send_request_t*)des->des_cbdata; mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*)des->des_context; size_t req_bytes_delivered = 0; /* count bytes of user data actually delivered and check for request completion */ MCA_PML_OB1_COMPUTE_SEGMENT_LENGTH( des->des_src, des->des_src_cnt, 0, req_bytes_delivered ); if( OPAL_THREAD_ADD_SIZE_T( &sendreq->req_bytes_delivered, req_bytes_delivered ) == sendreq->req_send.req_bytes_packed ) { MCA_PML_OB1_SEND_REQUEST_PML_COMPLETE(sendreq); } /* release resources */ btl->btl_free(btl,des); MCA_PML_OB1_PROGRESS_PENDING(bml_btl); } /** * Completion of a control message - return resources. */ static void mca_pml_ob1_send_ctl_completion( mca_btl_base_module_t* btl, struct mca_btl_base_endpoint_t* ep, struct mca_btl_base_descriptor_t* descriptor, int status) { /* return the descriptor */ mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*) descriptor->des_context; mca_bml_base_free(bml_btl, descriptor); /* check for pending requests */ MCA_PML_OB1_PROGRESS_PENDING(bml_btl); } /** * Completion of additional fragments of a large message - may need * to schedule additional fragments. */ static void mca_pml_ob1_frag_completion( mca_btl_base_module_t* btl, struct mca_btl_base_endpoint_t* ep, struct mca_btl_base_descriptor_t* descriptor, int status) { mca_pml_ob1_send_request_t* sendreq = (mca_pml_ob1_send_request_t*)descriptor->des_cbdata; mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*) descriptor->des_context; size_t req_bytes_delivered = 0; /* check completion status */ if(OMPI_SUCCESS != status) { /* TSW - FIX */ opal_output(0, "%s:%d FATAL", __FILE__, __LINE__); orte_errmgr.abort(); } /* count bytes of user data actually delivered */ MCA_PML_OB1_COMPUTE_SEGMENT_LENGTH( descriptor->des_src, descriptor->des_src_cnt, sizeof(mca_pml_ob1_frag_hdr_t), req_bytes_delivered ); OPAL_THREAD_ADD_SIZE_T(&sendreq->req_pipeline_depth, -1); /* return the descriptor */ mca_bml_base_free(bml_btl, descriptor); if(OPAL_THREAD_ADD_SIZE_T(&sendreq->req_bytes_delivered, req_bytes_delivered) == sendreq->req_send.req_bytes_packed) { MCA_PML_OB1_SEND_REQUEST_PML_COMPLETE(sendreq); } else { mca_pml_ob1_send_request_schedule(sendreq); } /* check for pending requests */ MCA_PML_OB1_PROGRESS_PENDING(bml_btl); } /** * Buffer the entire message and mark as complete. */ int mca_pml_ob1_send_request_start_buffered( mca_pml_ob1_send_request_t* sendreq, mca_bml_base_btl_t* bml_btl, size_t size) { mca_btl_base_descriptor_t* descriptor; mca_btl_base_segment_t* segment; mca_pml_ob1_hdr_t* hdr; struct iovec iov; unsigned int iov_count; size_t max_data; int rc; /* allocate descriptor */ mca_bml_base_alloc(bml_btl, &descriptor, MCA_BTL_NO_ORDER, sizeof(mca_pml_ob1_rendezvous_hdr_t) + size); if(NULL == descriptor) { return OMPI_ERR_OUT_OF_RESOURCE; } segment = descriptor->des_src; /* pack the data into the BTL supplied buffer */ iov.iov_base = (IOVBASE_TYPE*)((unsigned char*)segment->seg_addr.pval + sizeof(mca_pml_ob1_rendezvous_hdr_t)); iov.iov_len = size; iov_count = 1; max_data = size; if((rc = ompi_convertor_pack( &sendreq->req_send.req_convertor, &iov, &iov_count, &max_data)) < 0) { mca_bml_base_free(bml_btl, descriptor); return rc; } /* build rendezvous header */ hdr = (mca_pml_ob1_hdr_t*)segment->seg_addr.pval; hdr->hdr_common.hdr_flags = 0; hdr->hdr_common.hdr_type = MCA_PML_OB1_HDR_TYPE_RNDV; hdr->hdr_match.hdr_ctx = sendreq->req_send.req_base.req_comm->c_contextid; hdr->hdr_match.hdr_src = sendreq->req_send.req_base.req_comm->c_my_rank; hdr->hdr_match.hdr_tag = sendreq->req_send.req_base.req_tag; hdr->hdr_match.hdr_seq = (uint16_t)sendreq->req_send.req_base.req_sequence; hdr->hdr_rndv.hdr_msg_length = sendreq->req_send.req_bytes_packed; hdr->hdr_rndv.hdr_src_req.pval = sendreq; #if OMPI_ENABLE_HETEROGENEOUS_SUPPORT #ifdef WORDS_BIGENDIAN hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO; #else /* if we are little endian and the remote side is big endian, we're responsible for making sure the data is in network byte order */ if (sendreq->req_send.req_base.req_proc->proc_arch & OMPI_ARCH_ISBIGENDIAN) { hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO; MCA_PML_OB1_RNDV_HDR_HTON(hdr->hdr_rndv); } #endif #endif /* update lengths */ segment->seg_len = sizeof(mca_pml_ob1_rendezvous_hdr_t) + max_data; descriptor->des_cbfunc = mca_pml_ob1_rndv_completion; descriptor->des_flags |= MCA_BTL_DES_FLAGS_PRIORITY; descriptor->des_cbdata = sendreq; /* buffer the remainder of the message */ rc = mca_pml_base_bsend_request_alloc((ompi_request_t*)sendreq); if(OMPI_SUCCESS != rc) { mca_bml_base_free(bml_btl, descriptor); return rc; } iov.iov_base = (IOVBASE_TYPE*)(((unsigned char*)sendreq->req_send.req_addr) + max_data); iov.iov_len = max_data = sendreq->req_send.req_bytes_packed - max_data; if((rc = ompi_convertor_pack( &sendreq->req_send.req_convertor, &iov, &iov_count, &max_data)) < 0) { mca_bml_base_free(bml_btl, descriptor); return rc; } /* re-init convertor for packed data */ ompi_convertor_prepare_for_send( &sendreq->req_send.req_convertor, MPI_BYTE, sendreq->req_send.req_bytes_packed, sendreq->req_send.req_addr ); /* request is complete at mpi level */ OPAL_THREAD_LOCK(&ompi_request_lock); MCA_PML_OB1_SEND_REQUEST_MPI_COMPLETE(sendreq); OPAL_THREAD_UNLOCK(&ompi_request_lock); /* send */ rc = mca_bml_base_send(bml_btl, descriptor, MCA_BTL_TAG_PML); if(OMPI_SUCCESS != rc) { mca_bml_base_free(bml_btl, descriptor ); } return rc; } /** * We work on a buffered request with a size smaller than the eager size * or the BTL is not able to send the data IN_PLACE. Request a segment * that is used for initial hdr and any eager data. This is used only * from the _START macro. */ int mca_pml_ob1_send_request_start_copy( mca_pml_ob1_send_request_t* sendreq, mca_bml_base_btl_t* bml_btl, size_t size ) { mca_btl_base_descriptor_t* descriptor; mca_btl_base_segment_t* segment; mca_pml_ob1_hdr_t* hdr; struct iovec iov; unsigned int iov_count; size_t max_data = size; int rc; /* allocate descriptor */ if( 0 == size ) { MCA_PML_OB1_DES_ALLOC( bml_btl, descriptor, MCA_BTL_NO_ORDER, sizeof(mca_pml_ob1_match_hdr_t) ); if( OPAL_UNLIKELY(NULL == descriptor) ) { return OMPI_ERR_OUT_OF_RESOURCE; } segment = descriptor->des_src; descriptor->des_cbfunc = mca_pml_ob1_match_completion_cache; } else { mca_bml_base_alloc( bml_btl, &descriptor, MCA_BTL_NO_ORDER, sizeof(mca_pml_ob1_match_hdr_t) + size ); if( OPAL_UNLIKELY(NULL == descriptor) ) { return OMPI_ERR_OUT_OF_RESOURCE; } segment = descriptor->des_src; /* pack the data into the supplied buffer */ iov.iov_base = (IOVBASE_TYPE*)((unsigned char*)segment->seg_addr.pval + sizeof(mca_pml_ob1_match_hdr_t)); iov.iov_len = size; iov_count = 1; (void)ompi_convertor_pack( &sendreq->req_send.req_convertor, &iov, &iov_count, &max_data ); descriptor->des_cbfunc = mca_pml_ob1_match_completion_free; } /* build match header */ hdr = (mca_pml_ob1_hdr_t*)segment->seg_addr.pval; hdr->hdr_common.hdr_flags = 0; hdr->hdr_common.hdr_type = MCA_PML_OB1_HDR_TYPE_MATCH; hdr->hdr_match.hdr_ctx = sendreq->req_send.req_base.req_comm->c_contextid; hdr->hdr_match.hdr_src = sendreq->req_send.req_base.req_comm->c_my_rank; hdr->hdr_match.hdr_tag = sendreq->req_send.req_base.req_tag; hdr->hdr_match.hdr_seq = (uint16_t)sendreq->req_send.req_base.req_sequence; #if OMPI_ENABLE_HETEROGENEOUS_SUPPORT #ifdef WORDS_BIGENDIAN hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO; #else /* if we are little endian and the remote side is big endian, we're responsible for making sure the data is in network byte order */ if (sendreq->req_send.req_base.req_proc->proc_arch & OMPI_ARCH_ISBIGENDIAN) { hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO; MCA_PML_OB1_MATCH_HDR_HTON(hdr->hdr_match); } #endif #endif /* update lengths */ segment->seg_len = sizeof(mca_pml_ob1_match_hdr_t) + max_data; /* short message */ descriptor->des_flags |= MCA_BTL_DES_FLAGS_PRIORITY; descriptor->des_cbdata = sendreq; /* signal request completion */ OPAL_THREAD_LOCK(&ompi_request_lock); MCA_PML_OB1_SEND_REQUEST_MPI_COMPLETE(sendreq); OPAL_THREAD_UNLOCK(&ompi_request_lock); /* send */ rc = mca_bml_base_send(bml_btl, descriptor, MCA_BTL_TAG_PML); if( OPAL_UNLIKELY(OMPI_SUCCESS != rc) ) { if( 0 == size ) MCA_BML_BASE_BTL_DES_RETURN(bml_btl, descriptor); else mca_bml_base_free(bml_btl, descriptor ); } return rc; } /** * BTL can send directly from user buffer so allow the BTL * to prepare the segment list. Start sending a small message. */ int mca_pml_ob1_send_request_start_prepare( mca_pml_ob1_send_request_t* sendreq, mca_bml_base_btl_t* bml_btl, size_t size ) { mca_btl_base_descriptor_t* descriptor; mca_btl_base_segment_t* segment; mca_pml_ob1_hdr_t* hdr; int rc; /* prepare descriptor */ mca_bml_base_prepare_src( bml_btl, NULL, &sendreq->req_send.req_convertor, MCA_BTL_NO_ORDER, sizeof(mca_pml_ob1_match_hdr_t), &size, &descriptor); if( OPAL_UNLIKELY(NULL == descriptor) ) { return OMPI_ERR_OUT_OF_RESOURCE; } segment = descriptor->des_src; /* build match header */ hdr = (mca_pml_ob1_hdr_t*)segment->seg_addr.pval; hdr->hdr_common.hdr_flags = 0; hdr->hdr_common.hdr_type = MCA_PML_OB1_HDR_TYPE_MATCH; hdr->hdr_match.hdr_ctx = sendreq->req_send.req_base.req_comm->c_contextid; hdr->hdr_match.hdr_src = sendreq->req_send.req_base.req_comm->c_my_rank; hdr->hdr_match.hdr_tag = sendreq->req_send.req_base.req_tag; hdr->hdr_match.hdr_seq = (uint16_t)sendreq->req_send.req_base.req_sequence; #if OMPI_ENABLE_HETEROGENEOUS_SUPPORT #ifdef WORDS_BIGENDIAN hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO; #else /* if we are little endian and the remote side is big endian, we're responsible for making sure the data is in network byte order */ if (sendreq->req_send.req_base.req_proc->proc_arch & OMPI_ARCH_ISBIGENDIAN) { hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO; MCA_PML_OB1_MATCH_HDR_HTON(hdr->hdr_match); } #endif #endif /* short message */ descriptor->des_cbfunc = mca_pml_ob1_match_completion_free; descriptor->des_flags |= MCA_BTL_DES_FLAGS_PRIORITY; descriptor->des_cbdata = sendreq; /* send */ rc = mca_bml_base_send(bml_btl, descriptor, MCA_BTL_TAG_PML); if( OPAL_UNLIKELY(OMPI_SUCCESS != rc) ) { mca_bml_base_free(bml_btl, descriptor ); } return rc; } /** * We have contigous data that is registered - schedule across * available nics. */ int mca_pml_ob1_send_request_start_rdma( mca_pml_ob1_send_request_t* sendreq, mca_bml_base_btl_t* bml_btl, size_t size) { /* * When req_rdma array is constructed the firs element of the array always * assigned different btl in round robin fashion (if there are more than * one RDMA capable BTLs). This way round robin distribution of RDMA * operation is achieved. */ mca_mpool_base_registration_t* reg = sendreq->req_rdma[0].btl_reg; mca_btl_base_descriptor_t* src; mca_btl_base_descriptor_t* des; mca_btl_base_segment_t* segment; mca_pml_ob1_hdr_t* hdr; size_t i; int rc; bml_btl = sendreq->req_rdma[0].bml_btl; if(sendreq->req_rdma_cnt == 1 && bml_btl->btl_flags & MCA_BTL_FLAGS_GET) { size_t old_position = sendreq->req_send.req_convertor.bConverted; /* prepare source descriptor/segment(s) */ mca_bml_base_prepare_src( bml_btl, reg, &sendreq->req_send.req_convertor, MCA_BTL_NO_ORDER, 0, &size, &src); if(NULL == src) { ompi_convertor_set_position(&sendreq->req_send.req_convertor, &old_position); return OMPI_ERR_OUT_OF_RESOURCE; } src->des_cbfunc = mca_pml_ob1_rget_completion; src->des_cbdata = sendreq; /* allocate space for get hdr + segment list */ mca_bml_base_alloc(bml_btl, &des, MCA_BTL_NO_ORDER, sizeof(mca_pml_ob1_rget_hdr_t) + (sizeof(mca_btl_base_segment_t)*(src->des_src_cnt-1))); if(NULL == des) { ompi_convertor_set_position(&sendreq->req_send.req_convertor, &old_position); mca_bml_base_free(bml_btl, src); return OMPI_ERR_OUT_OF_RESOURCE; } segment = des->des_src; /* build match header */ hdr = (mca_pml_ob1_hdr_t*)segment->seg_addr.pval; hdr->hdr_common.hdr_flags = MCA_PML_OB1_HDR_FLAGS_CONTIG|MCA_PML_OB1_HDR_FLAGS_PIN; hdr->hdr_common.hdr_type = MCA_PML_OB1_HDR_TYPE_RGET; hdr->hdr_match.hdr_ctx = sendreq->req_send.req_base.req_comm->c_contextid; hdr->hdr_match.hdr_src = sendreq->req_send.req_base.req_comm->c_my_rank; hdr->hdr_match.hdr_tag = sendreq->req_send.req_base.req_tag; hdr->hdr_match.hdr_seq = (uint16_t)sendreq->req_send.req_base.req_sequence; hdr->hdr_rndv.hdr_msg_length = sendreq->req_send.req_bytes_packed; hdr->hdr_rndv.hdr_src_req.pval = sendreq; hdr->hdr_rget.hdr_des.pval = src; hdr->hdr_rget.hdr_seg_cnt = src->des_src_cnt; #if OMPI_ENABLE_HETEROGENEOUS_SUPPORT #ifdef WORDS_BIGENDIAN hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO; #else /* if we are little endian and the remote side is big endian, we're responsible for making sure the data is in network byte order */ /* RDMA is currently disabled by bml if arch doesn't match, so this shouldn't be needed. here to make sure we remember if we ever change the bml. */ assert(0 == (sendreq->req_send.req_base.req_proc->proc_arch & OMPI_ARCH_ISBIGENDIAN)); #endif #endif for( i = 0; i < src->des_src_cnt; i++ ) { hdr->hdr_rget.hdr_segs[i].seg_addr.lval = ompi_ptr_ptol(src->des_src[i].seg_addr.pval); hdr->hdr_rget.hdr_segs[i].seg_len = src->des_src[i].seg_len; hdr->hdr_rget.hdr_segs[i].seg_key.key64 = src->des_src[i].seg_key.key64; } des->des_cbfunc = mca_pml_ob1_send_ctl_completion; /** * Well, it's a get so we will not know when the peer get the data anyway. * If we generate the PERUSE event here, at least we will know when do we * sent the GET message ... */ if( sendreq->req_send.req_bytes_packed > 0 ) { PERUSE_TRACE_COMM_EVENT( PERUSE_COMM_REQ_XFER_BEGIN, &(sendreq->req_send.req_base), PERUSE_SEND ); } } else { /* allocate a rendezvous header - dont eager send any data * receiver will schedule rdma put(s) of the entire message */ mca_bml_base_alloc(bml_btl, &des, MCA_BTL_NO_ORDER, sizeof(mca_pml_ob1_rendezvous_hdr_t)); if(NULL == des) { return OMPI_ERR_OUT_OF_RESOURCE; } segment = des->des_src; /* build hdr */ hdr = (mca_pml_ob1_hdr_t*)segment->seg_addr.pval; hdr->hdr_common.hdr_flags = MCA_PML_OB1_HDR_FLAGS_CONTIG|MCA_PML_OB1_HDR_FLAGS_PIN; hdr->hdr_common.hdr_type = MCA_PML_OB1_HDR_TYPE_RNDV; hdr->hdr_match.hdr_ctx = sendreq->req_send.req_base.req_comm->c_contextid; hdr->hdr_match.hdr_src = sendreq->req_send.req_base.req_comm->c_my_rank; hdr->hdr_match.hdr_tag = sendreq->req_send.req_base.req_tag; hdr->hdr_match.hdr_seq = (uint16_t)sendreq->req_send.req_base.req_sequence; hdr->hdr_rndv.hdr_msg_length = sendreq->req_send.req_bytes_packed; hdr->hdr_rndv.hdr_src_req.pval = sendreq; #if OMPI_ENABLE_HETEROGENEOUS_SUPPORT #ifdef WORDS_BIGENDIAN hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO; #else /* if we are little endian and the remote side is big endian, we're responsible for making sure the data is in network byte order */ if (sendreq->req_send.req_base.req_proc->proc_arch & OMPI_ARCH_ISBIGENDIAN) { hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO; MCA_PML_OB1_RNDV_HDR_HTON(hdr->hdr_rndv); } #endif #endif /* update lengths with number of bytes actually packed */ segment->seg_len = sizeof(mca_pml_ob1_rendezvous_hdr_t); /* first fragment of a long message */ des->des_cbfunc = mca_pml_ob1_rndv_completion; } des->des_flags |= MCA_BTL_DES_FLAGS_PRIORITY; des->des_cbdata = sendreq; /* send */ rc = mca_bml_base_send(bml_btl, des, MCA_BTL_TAG_PML); if(OMPI_SUCCESS != rc) { mca_bml_base_free(bml_btl, des); } return rc; } /** * Rendezvous is required. Not doing rdma so eager send up to * the btls eager limit. */ int mca_pml_ob1_send_request_start_rndv( mca_pml_ob1_send_request_t* sendreq, mca_bml_base_btl_t* bml_btl, size_t size, int flags) { mca_btl_base_descriptor_t* des; mca_btl_base_segment_t* segment; mca_pml_ob1_hdr_t* hdr; int rc; /* prepare descriptor */ if(size == 0) { mca_bml_base_alloc( bml_btl, &des, MCA_BTL_NO_ORDER, sizeof(mca_pml_ob1_rendezvous_hdr_t) ); } else { mca_bml_base_prepare_src( bml_btl, NULL, &sendreq->req_send.req_convertor, MCA_BTL_NO_ORDER, sizeof(mca_pml_ob1_rendezvous_hdr_t), &size, &des); } if(NULL == des) { return OMPI_ERR_OUT_OF_RESOURCE; } segment = des->des_src; /* build hdr */ hdr = (mca_pml_ob1_hdr_t*)segment->seg_addr.pval; hdr->hdr_common.hdr_flags = flags; hdr->hdr_common.hdr_type = MCA_PML_OB1_HDR_TYPE_RNDV; hdr->hdr_match.hdr_ctx = sendreq->req_send.req_base.req_comm->c_contextid; hdr->hdr_match.hdr_src = sendreq->req_send.req_base.req_comm->c_my_rank; hdr->hdr_match.hdr_tag = sendreq->req_send.req_base.req_tag; hdr->hdr_match.hdr_seq = (uint16_t)sendreq->req_send.req_base.req_sequence; hdr->hdr_rndv.hdr_msg_length = sendreq->req_send.req_bytes_packed; hdr->hdr_rndv.hdr_src_req.pval = sendreq; #if OMPI_ENABLE_HETEROGENEOUS_SUPPORT #ifdef WORDS_BIGENDIAN hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO; #else /* if we are little endian and the remote side is big endian, we're responsible for making sure the data is in network byte order */ if (sendreq->req_send.req_base.req_proc->proc_arch & OMPI_ARCH_ISBIGENDIAN) { hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO; MCA_PML_OB1_RNDV_HDR_HTON(hdr->hdr_rndv); } #endif #endif /* first fragment of a long message */ des->des_flags |= MCA_BTL_DES_FLAGS_PRIORITY; des->des_cbdata = sendreq; des->des_cbfunc = mca_pml_ob1_rndv_completion; /* send */ rc = mca_bml_base_send(bml_btl, des, MCA_BTL_TAG_PML); if(OMPI_SUCCESS != rc) { mca_bml_base_free(bml_btl, des ); } return rc; } void mca_pml_ob1_send_requst_copy_in_out(mca_pml_ob1_send_request_t *sendreq, uint64_t send_offset, uint64_t send_length) { mca_pml_ob1_send_range_t *sr; ompi_free_list_item_t *i; int rc = OMPI_SUCCESS; if(0 == send_length) return; OMPI_FREE_LIST_WAIT(&mca_pml_ob1.send_ranges, i, rc); sr = (mca_pml_ob1_send_range_t*)i; sr->range_send_offset = send_offset; sr->range_send_length = send_length; OPAL_THREAD_LOCK(&sendreq->req_send_range_lock); opal_list_append(&sendreq->req_send_ranges, (opal_list_item_t*)sr); OPAL_THREAD_UNLOCK(&sendreq->req_send_range_lock); } /** * Schedule pipeline of send descriptors for the given request. * Up to the rdma threshold. If this is a send based protocol, * the rdma threshold is the end of the message. Otherwise, schedule * fragments up to the threshold to overlap initial registration/setup * costs of the rdma. Only one thread can be inside this function. */ int mca_pml_ob1_send_request_schedule_exclusive( mca_pml_ob1_send_request_t* sendreq) { mca_bml_base_endpoint_t* bml_endpoint = sendreq->req_endpoint; size_t num_btl_avail = mca_bml_base_btl_array_get_size(&bml_endpoint->btl_send); do { size_t prev_bytes_remaining = 0, num_fail = 0; mca_pml_ob1_send_range_t *range = NULL; while(true) { mca_pml_ob1_frag_hdr_t* hdr; mca_btl_base_descriptor_t* des; int rc; size_t size, offset; opal_list_item_t *item; mca_bml_base_btl_t* bml_btl = mca_bml_base_btl_array_get_next(&bml_endpoint->btl_send); if(NULL == range || 0 == range->range_send_length) { OPAL_THREAD_LOCK(&sendreq->req_send_range_lock); if(range) { opal_list_remove_first(&sendreq->req_send_ranges); OMPI_FREE_LIST_RETURN(&mca_pml_ob1.send_ranges, &range->base); } item = opal_list_get_first(&sendreq->req_send_ranges); OPAL_THREAD_UNLOCK(&sendreq->req_send_range_lock); if(opal_list_get_end(&sendreq->req_send_ranges) == item) break; range = (mca_pml_ob1_send_range_t*)item; prev_bytes_remaining = 0; } if(true == sendreq->req_throttle_sends && sendreq->req_pipeline_depth >= mca_pml_ob1.send_pipeline_depth) break; if(prev_bytes_remaining == range->range_send_length) num_fail++; else num_fail = 0; prev_bytes_remaining = range->range_send_length; if (num_fail == num_btl_avail) { assert(sendreq->req_pending == MCA_PML_OB1_SEND_PENDING_NONE); OPAL_THREAD_LOCK(&mca_pml_ob1.lock); sendreq->req_pending = MCA_PML_OB1_SEND_PENDING_SCHEDULE; opal_list_append(&mca_pml_ob1.send_pending, (opal_list_item_t*)sendreq); OPAL_THREAD_UNLOCK(&mca_pml_ob1.lock); return OMPI_ERR_OUT_OF_RESOURCE; } if(num_btl_avail == 1 || range->range_send_length < bml_btl->btl_min_send_size) { size = range->range_send_length; } else { /* otherwise attempt to give the BTL a percentage of the message * based on a weighting factor. for simplicity calculate this as * a percentage of the overall message length (regardless of * amount previously assigned) */ size = (size_t)(bml_btl->btl_weight * range->range_send_length); } /* makes sure that we don't exceed BTL max send size */ if (bml_btl->btl_max_send_size != 0 && size > (bml_btl->btl_max_send_size - sizeof(mca_pml_ob1_frag_hdr_t))) { size = bml_btl->btl_max_send_size - sizeof(mca_pml_ob1_frag_hdr_t); } /* pack into a descriptor */ offset = (size_t)range->range_send_offset; ompi_convertor_set_position(&sendreq->req_send.req_convertor, &offset); range->range_send_offset = (uint64_t)offset; mca_bml_base_prepare_src(bml_btl, NULL, &sendreq->req_send.req_convertor, MCA_BTL_NO_ORDER, sizeof(mca_pml_ob1_frag_hdr_t), &size, &des); if(des == NULL) { continue; } des->des_cbfunc = mca_pml_ob1_frag_completion; des->des_cbdata = sendreq; /* setup header */ hdr = (mca_pml_ob1_frag_hdr_t*)des->des_src->seg_addr.pval; hdr->hdr_common.hdr_flags = 0; hdr->hdr_common.hdr_type = MCA_PML_OB1_HDR_TYPE_FRAG; hdr->hdr_frag_offset = range->range_send_offset; hdr->hdr_src_req.pval = sendreq; hdr->hdr_dst_req = sendreq->req_recv; #if OMPI_ENABLE_HETEROGENEOUS_SUPPORT #ifdef WORDS_BIGENDIAN hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO; #else /* * if we are little endian and the remote side is big endian, * we're responsible for making sure the data is in network byte * order */ if(sendreq->req_send.req_base.req_proc->proc_arch & OMPI_ARCH_ISBIGENDIAN) { hdr->hdr_common.hdr_flags |= MCA_PML_OB1_HDR_FLAGS_NBO; MCA_PML_OB1_FRAG_HDR_HTON(*hdr); } #endif #endif #if OMPI_WANT_PERUSE PERUSE_TRACE_COMM_OMPI_EVENT(PERUSE_COMM_REQ_XFER_CONTINUE, &(sendreq->req_send.req_base), size, PERUSE_SEND); #endif /* OMPI_WANT_PERUSE */ /* initiate send - note that this may complete before the call returns */ rc = mca_bml_base_send(bml_btl, des, MCA_BTL_TAG_PML); if(rc == OMPI_SUCCESS) { range->range_send_length -= size; /* update state */ range->range_send_offset += size; OPAL_THREAD_ADD_SIZE_T(&sendreq->req_pipeline_depth, 1); } else { mca_bml_base_free(bml_btl,des); continue; } mca_bml.bml_progress(); } } while (OPAL_THREAD_ADD32(&sendreq->req_lock,-1) > 0); return OMPI_SUCCESS; } /** * An RDMA put operation has completed: * (1) Update request status and if required set completed * (2) Send FIN control message to the destination */ static void mca_pml_ob1_put_completion( mca_btl_base_module_t* btl, struct mca_btl_base_endpoint_t* ep, struct mca_btl_base_descriptor_t* des, int status ) { mca_pml_ob1_rdma_frag_t* frag = (mca_pml_ob1_rdma_frag_t*)des->des_cbdata; mca_pml_ob1_send_request_t* sendreq = (mca_pml_ob1_send_request_t*)frag->rdma_req; mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*) des->des_context; /* check completion status */ if(OMPI_SUCCESS != status) { /* TSW - FIX */ ORTE_ERROR_LOG(status); orte_errmgr.abort(); } mca_pml_ob1_send_fin(sendreq->req_send.req_base.req_proc, bml_btl, frag->rdma_hdr.hdr_rdma.hdr_des.pval, des->order, 0); /* check for request completion */ if( OPAL_THREAD_ADD_SIZE_T(&sendreq->req_bytes_delivered, frag->rdma_length) >= sendreq->req_send.req_bytes_packed) { /* if we've got completion on rndv packet */ if (sendreq->req_state == 2) { MCA_PML_OB1_SEND_REQUEST_PML_COMPLETE(sendreq); } } MCA_PML_OB1_RDMA_FRAG_RETURN(frag); /* return rdma descriptor - do this after queuing the fin message - as * release rdma resources (unpin memory) can take some time. */ des->des_dst = NULL; des->des_dst_cnt = 0; mca_bml_base_free(bml_btl, des); MCA_PML_OB1_PROGRESS_PENDING(bml_btl); } int mca_pml_ob1_send_request_put_frag( mca_pml_ob1_rdma_frag_t* frag ) { mca_mpool_base_registration_t* reg = NULL; mca_bml_base_btl_t* bml_btl = frag->rdma_bml; mca_btl_base_descriptor_t* des; size_t save_size = frag->rdma_length; int rc; /* setup descriptor */ mca_bml_base_prepare_src( bml_btl, reg, &frag->convertor, MCA_BTL_NO_ORDER, 0, &frag->rdma_length, &des ); if(NULL == des) { if(frag->retries < mca_pml_ob1.rdma_put_retries_limit) { size_t offset = (size_t)frag->rdma_hdr.hdr_rdma.hdr_rdma_offset; frag->rdma_length = save_size; ompi_convertor_set_position(&frag->convertor, &offset); OPAL_THREAD_LOCK(&mca_pml_ob1.lock); opal_list_append(&mca_pml_ob1.rdma_pending, (opal_list_item_t*)frag); OPAL_THREAD_UNLOCK(&mca_pml_ob1.lock); } else { mca_pml_ob1_send_request_t *sendreq = (mca_pml_ob1_send_request_t*)frag->rdma_req; /* tell receiver to unregister memory */ mca_pml_ob1_send_fin(sendreq->req_send.req_base.req_proc, frag->rdma_hdr.hdr_rdma.hdr_des.pval, bml_btl, MCA_BTL_NO_ORDER, 1); /* send fragment by copy in/out */ mca_pml_ob1_send_requst_copy_in_out(sendreq, frag->rdma_hdr.hdr_rdma.hdr_rdma_offset, frag->rdma_length); mca_pml_ob1_send_request_schedule(sendreq); } return OMPI_ERR_OUT_OF_RESOURCE; } des->des_dst = frag->rdma_segs; des->des_dst_cnt = frag->rdma_hdr.hdr_rdma.hdr_seg_cnt; des->des_cbfunc = mca_pml_ob1_put_completion; des->des_cbdata = frag; PERUSE_TRACE_COMM_OMPI_EVENT( PERUSE_COMM_REQ_XFER_CONTINUE, &(((mca_pml_ob1_send_request_t*)frag->rdma_req)->req_send.req_base), save_size, PERUSE_SEND ); if(OMPI_SUCCESS != (rc = mca_bml_base_put(bml_btl, des))) { mca_bml_base_free(bml_btl, des); frag->rdma_length = save_size; if(OMPI_ERR_OUT_OF_RESOURCE == rc) { OPAL_THREAD_LOCK(&mca_pml_ob1.lock); opal_list_append(&mca_pml_ob1.rdma_pending, (opal_list_item_t*)frag); OPAL_THREAD_UNLOCK(&mca_pml_ob1.lock); return OMPI_ERR_OUT_OF_RESOURCE; } else { /* TSW - FIX */ ORTE_ERROR_LOG(rc); orte_errmgr.abort(); } } return OMPI_SUCCESS; } /** * Receiver has scheduled an RDMA operation: * (1) Allocate an RDMA fragment to maintain the state of the operation * (2) Call BTL prepare_src to pin/prepare source buffers * (3) Queue the RDMA put */ void mca_pml_ob1_send_request_put( mca_pml_ob1_send_request_t* sendreq, mca_btl_base_module_t* btl, mca_pml_ob1_rdma_hdr_t* hdr) { mca_bml_base_endpoint_t *bml_endpoint = sendreq->req_endpoint; mca_pml_ob1_rdma_frag_t* frag; int rc; size_t i, size = 0; if(hdr->hdr_common.hdr_flags & MCA_PML_OB1_HDR_TYPE_ACK) { OPAL_THREAD_ADD32(&sendreq->req_state, 1); } MCA_PML_OB1_RDMA_FRAG_ALLOC(frag, rc); if(NULL == frag) { /* TSW - FIX */ ORTE_ERROR_LOG(rc); orte_errmgr.abort(); } /* setup fragment */ for( i = 0; i < hdr->hdr_seg_cnt; i++ ) { frag->rdma_segs[i].seg_addr.lval = hdr->hdr_segs[i].seg_addr.lval; frag->rdma_segs[i].seg_len = hdr->hdr_segs[i].seg_len; frag->rdma_segs[i].seg_key.key64 = hdr->hdr_segs[i].seg_key.key64; size += frag->rdma_segs[i].seg_len; } frag->rdma_bml = mca_bml_base_btl_array_find(&bml_endpoint->btl_rdma, btl); frag->rdma_hdr.hdr_rdma = *hdr; frag->rdma_req = sendreq; frag->rdma_ep = bml_endpoint; frag->rdma_length = size; frag->rdma_state = MCA_PML_OB1_RDMA_PUT; frag->reg = NULL; frag->retries = 0; /* lookup the corresponding registration */ for(i=0; ireq_rdma_cnt; i++) { if(sendreq->req_rdma[i].bml_btl == frag->rdma_bml) { frag->reg = sendreq->req_rdma[i].btl_reg; break; } } /* RDMA writes may proceed in parallel to send and to each other, so * create clone of the convertor for each RDMA fragment */ size = hdr->hdr_rdma_offset; ompi_convertor_clone_with_position(&sendreq->req_send.req_convertor, &frag->convertor, 0, &size); mca_pml_ob1_send_request_put_frag(frag); }