/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */ /* * Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana * University Research and Technology * Corporation. All rights reserved. * Copyright (c) 2004-2013 The University of Tennessee and The University * of Tennessee Research Foundation. All rights * reserved. * Copyright (c) 2004-2008 High Performance Computing Center Stuttgart, * University of Stuttgart. All rights reserved. * Copyright (c) 2004-2005 The Regents of the University of California. * All rights reserved. * Copyright (c) 2008 UT-Battelle, LLC. All rights reserved. * Copyright (c) 2010-2012 Oracle and/or its affiliates. All rights reserved. * Copyright (c) 2011-2012 Los Alamos National Security, LLC. * All rights reserved. * $COPYRIGHT$ * * Additional copyrights may follow * * $HEADER$ */ #include "ompi_config.h" #include "ompi/mca/pml/pml.h" #include "ompi/mca/bml/bml.h" #include "ompi/mca/btl/btl.h" #include "ompi/mca/mpool/mpool.h" #include "pml_bfo_comm.h" #include "pml_bfo_recvreq.h" #include "pml_bfo_recvfrag.h" #include "pml_bfo_sendreq.h" #include "pml_bfo_rdmafrag.h" #include "ompi/mca/bml/base/base.h" #include "opal/util/arch.h" #include "ompi/memchecker.h" #if OPAL_CUDA_SUPPORT int mca_pml_bfo_cuda_need_buffers(mca_pml_bfo_recv_request_t* recvreq, mca_btl_base_module_t* btl); #endif /* OPAL_CUDA_SUPPORT */ #if PML_BFO #include "pml_bfo_failover.h" #endif /* PML_BFO */ void mca_pml_bfo_recv_request_process_pending(void) { mca_pml_bfo_recv_request_t* recvreq; int rc, i, s = (int)opal_list_get_size(&mca_pml_bfo.recv_pending); for(i = 0; i < s; i++) { OPAL_THREAD_LOCK(&mca_pml_bfo.lock); recvreq = (mca_pml_bfo_recv_request_t*) opal_list_remove_first(&mca_pml_bfo.recv_pending); OPAL_THREAD_UNLOCK(&mca_pml_bfo.lock); if( OPAL_UNLIKELY(NULL == recvreq) ) break; recvreq->req_pending = false; rc = mca_pml_bfo_recv_request_schedule_exclusive(recvreq, NULL); if(OMPI_ERR_OUT_OF_RESOURCE == rc) break; } } static int mca_pml_bfo_recv_request_free(struct ompi_request_t** request) { mca_pml_bfo_recv_request_t* recvreq = *(mca_pml_bfo_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); ); MCA_PML_BFO_RECV_REQUEST_RETURN( recvreq ); } OPAL_THREAD_UNLOCK(&ompi_request_lock); *request = MPI_REQUEST_NULL; return OMPI_SUCCESS; } static int mca_pml_bfo_recv_request_cancel(struct ompi_request_t* ompi_request, int complete) { mca_pml_bfo_recv_request_t* request = (mca_pml_bfo_recv_request_t*)ompi_request; mca_pml_bfo_comm_t* comm = request->req_recv.req_base.req_comm->c_pml_comm; if( true == ompi_request->req_complete ) { /* way to late to cancel this one */ /* * Receive request completed, make user buffer accessable. */ MEMCHECKER( memchecker_call(&opal_memchecker_base_mem_defined, request->req_recv.req_base.req_addr, request->req_recv.req_base.req_count, request->req_recv.req_base.req_datatype); ); return OMPI_SUCCESS; } /* The rest should be protected behind the match logic lock */ OPAL_THREAD_LOCK(&comm->matching_lock); if( OMPI_ANY_TAG == ompi_request->req_status.MPI_TAG ) { /* the match has not been already done */ if( request->req_recv.req_base.req_peer == OMPI_ANY_SOURCE ) { opal_list_remove_item( &comm->wild_receives, (opal_list_item_t*)request ); } else { mca_pml_bfo_comm_proc_t* proc = comm->procs + request->req_recv.req_base.req_peer; opal_list_remove_item(&proc->specific_receives, (opal_list_item_t*)request); } PERUSE_TRACE_COMM_EVENT( PERUSE_COMM_REQ_REMOVE_FROM_POSTED_Q, &(request->req_recv.req_base), PERUSE_RECV ); /** * As now the PML is done with this request we have to force the pml_complete * to true. Otherwise, the request will never be freed. */ request->req_recv.req_base.req_pml_complete = true; } OPAL_THREAD_UNLOCK(&comm->matching_lock); OPAL_THREAD_LOCK(&ompi_request_lock); ompi_request->req_status._cancelled = true; /* This macro will set the req_complete to true so the MPI Test/Wait* functions * on this request will be able to complete. As the status is marked as * cancelled the cancel state will be detected. */ MCA_PML_BFO_RECV_REQUEST_MPI_COMPLETE(request); OPAL_THREAD_UNLOCK(&ompi_request_lock); /* * Receive request cancelled, make user buffer accessable. */ MEMCHECKER( memchecker_call(&opal_memchecker_base_mem_defined, request->req_recv.req_base.req_addr, request->req_recv.req_base.req_count, request->req_recv.req_base.req_datatype); ); return OMPI_SUCCESS; } static void mca_pml_bfo_recv_request_construct(mca_pml_bfo_recv_request_t* request) { request->req_recv.req_base.req_type = MCA_PML_REQUEST_RECV; request->req_recv.req_base.req_ompi.req_free = mca_pml_bfo_recv_request_free; request->req_recv.req_base.req_ompi.req_cancel = mca_pml_bfo_recv_request_cancel; request->req_rdma_cnt = 0; OBJ_CONSTRUCT(&request->lock, opal_mutex_t); } OBJ_CLASS_INSTANCE( mca_pml_bfo_recv_request_t, mca_pml_base_recv_request_t, mca_pml_bfo_recv_request_construct, NULL); /* * Release resources. */ static void mca_pml_bfo_recv_ctl_completion( mca_btl_base_module_t* btl, struct mca_btl_base_endpoint_t* ep, struct mca_btl_base_descriptor_t* des, int status ) { mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*)des->des_context; #if PML_BFO if (btl->btl_flags & MCA_BTL_FLAGS_FAILOVER_SUPPORT) { mca_pml_bfo_check_recv_ctl_completion_status(btl, des, status); } MCA_PML_BFO_CHECK_RECVREQ_EAGER_BML_BTL_RECV_CTL(bml_btl, btl, des); #endif /* PML_BFO */ MCA_PML_BFO_PROGRESS_PENDING(bml_btl); } /* * Put operation has completed remotely - update request status */ static void mca_pml_bfo_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_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*)des->des_context; mca_pml_bfo_recv_request_t* recvreq = (mca_pml_bfo_recv_request_t*)des->des_cbdata; size_t bytes_received = 0; if( OPAL_LIKELY(status == OMPI_SUCCESS) ) { bytes_received = mca_pml_bfo_compute_segment_length (btl->btl_seg_size, (void *) des->des_dst, des->des_dst_cnt, 0); } OPAL_THREAD_ADD_SIZE_T(&recvreq->req_pipeline_depth,-1); #if PML_BFO btl->btl_free(btl, des); MCA_PML_BFO_ERROR_CHECK_ON_FIN_FOR_PUT(recvreq); MCA_PML_BFO_CHECK_RECVREQ_EAGER_BML_BTL(bml_btl, btl, recvreq, "PUT"); #else /* PML_BFO */ mca_bml_base_free(bml_btl, des); #endif /* PML_BFO */ /* check completion status */ OPAL_THREAD_ADD_SIZE_T(&recvreq->req_bytes_received, bytes_received); if(recv_request_pml_complete_check(recvreq) == false && recvreq->req_rdma_offset < recvreq->req_send_offset) { /* schedule additional rdma operations */ mca_pml_bfo_recv_request_schedule(recvreq, bml_btl); } MCA_PML_BFO_PROGRESS_PENDING(bml_btl); } /* * */ int mca_pml_bfo_recv_request_ack_send_btl( ompi_proc_t* proc, mca_bml_base_btl_t* bml_btl, uint64_t hdr_src_req, void *hdr_dst_req, uint64_t hdr_send_offset, bool nordma) { mca_btl_base_descriptor_t* des; mca_pml_bfo_ack_hdr_t* ack; int rc; /* allocate descriptor */ mca_bml_base_alloc(bml_btl, &des, MCA_BTL_NO_ORDER, sizeof(mca_pml_bfo_ack_hdr_t), MCA_BTL_DES_FLAGS_PRIORITY | MCA_BTL_DES_FLAGS_BTL_OWNERSHIP | MCA_BTL_DES_SEND_ALWAYS_CALLBACK); if( OPAL_UNLIKELY(NULL == des) ) { return OMPI_ERR_OUT_OF_RESOURCE; } /* fill out header */ ack = (mca_pml_bfo_ack_hdr_t*)des->des_src->seg_addr.pval; ack->hdr_common.hdr_type = MCA_PML_BFO_HDR_TYPE_ACK; ack->hdr_common.hdr_flags = nordma ? MCA_PML_BFO_HDR_FLAGS_NORDMA : 0; ack->hdr_src_req.lval = hdr_src_req; ack->hdr_dst_req.pval = hdr_dst_req; ack->hdr_send_offset = hdr_send_offset; bfo_hdr_hton(ack, MCA_PML_BFO_HDR_TYPE_ACK, proc); /* initialize descriptor */ des->des_cbfunc = mca_pml_bfo_recv_ctl_completion; #if PML_BFO des->des_cbdata = hdr_dst_req; #endif /* PML_BFO */ rc = mca_bml_base_send(bml_btl, des, MCA_PML_BFO_HDR_TYPE_ACK); if( OPAL_LIKELY( rc >= 0 ) ) { #if PML_BFO if ((bml_btl->btl_flags & MCA_BTL_FLAGS_FAILOVER_SUPPORT) && (des->des_flags & MCA_BTL_DES_SEND_ALWAYS_CALLBACK)) { ((mca_pml_bfo_recv_request_t *)hdr_dst_req)->req_events++; } #endif /* PML_BFO */ return OMPI_SUCCESS; } mca_bml_base_free(bml_btl, des); return OMPI_ERR_OUT_OF_RESOURCE; } static int mca_pml_bfo_recv_request_ack( mca_pml_bfo_recv_request_t* recvreq, mca_pml_bfo_rendezvous_hdr_t* hdr, size_t bytes_received) { ompi_proc_t* proc = (ompi_proc_t*)recvreq->req_recv.req_base.req_proc; mca_bml_base_endpoint_t* bml_endpoint = NULL; bml_endpoint = (mca_bml_base_endpoint_t*) proc->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_BML]; /* by default copy everything */ recvreq->req_send_offset = bytes_received; if(hdr->hdr_msg_length > bytes_received) { size_t rdma_num = mca_bml_base_btl_array_get_size(&bml_endpoint->btl_rdma); /* * lookup request buffer to determine if memory is already * registered. */ if(opal_convertor_need_buffers(&recvreq->req_recv.req_base.req_convertor) == 0 && hdr->hdr_match.hdr_common.hdr_flags & MCA_PML_BFO_HDR_FLAGS_CONTIG && rdma_num != 0) { unsigned char *base; opal_convertor_get_current_pointer( &recvreq->req_recv.req_base.req_convertor, (void**)&(base) ); if(hdr->hdr_match.hdr_common.hdr_flags & MCA_PML_BFO_HDR_FLAGS_PIN) recvreq->req_rdma_cnt = mca_pml_bfo_rdma_btls(bml_endpoint, base, recvreq->req_recv.req_bytes_packed, recvreq->req_rdma ); else recvreq->req_rdma_cnt = 0; /* memory is already registered on both sides */ if (recvreq->req_rdma_cnt != 0) { recvreq->req_send_offset = hdr->hdr_msg_length; /* are rdma devices available for long rdma protocol */ } else if(bml_endpoint->btl_send_limit < hdr->hdr_msg_length) { /* use convertor to figure out the rdma offset for this request */ recvreq->req_send_offset = hdr->hdr_msg_length - bml_endpoint->btl_pipeline_send_length; if(recvreq->req_send_offset < bytes_received) recvreq->req_send_offset = bytes_received; /* use converter to figure out the rdma offset for this * request */ opal_convertor_set_position(&recvreq->req_recv.req_base.req_convertor, &recvreq->req_send_offset); recvreq->req_rdma_cnt = mca_pml_bfo_rdma_pipeline_btls(bml_endpoint, recvreq->req_send_offset - bytes_received, recvreq->req_rdma); } } /* nothing to send by copy in/out - no need to ack */ if(recvreq->req_send_offset == hdr->hdr_msg_length) return OMPI_SUCCESS; } /* let know to shedule function there is no need to put ACK flag */ recvreq->req_ack_sent = true; return mca_pml_bfo_recv_request_ack_send(proc, hdr->hdr_src_req.lval, recvreq, recvreq->req_send_offset, recvreq->req_send_offset == bytes_received); } /** * Return resources used by the RDMA */ static void mca_pml_bfo_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_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*)des->des_context; mca_pml_bfo_rdma_frag_t* frag = (mca_pml_bfo_rdma_frag_t*)des->des_cbdata; mca_pml_bfo_recv_request_t* recvreq = (mca_pml_bfo_recv_request_t*)frag->rdma_req; #if PML_BFO if (btl->btl_flags & MCA_BTL_FLAGS_FAILOVER_SUPPORT) { recvreq->req_events--; } #endif /* PML_BFO */ /* check completion status */ if( OPAL_UNLIKELY(OMPI_SUCCESS != status) ) { #if PML_BFO MCA_PML_BFO_ERROR_CHECK_ON_RDMA_READ_COMPLETION(recvreq); #else /* PML_BFO */ /* TSW - FIX */ OMPI_ERROR_LOG(status); ompi_rte_abort(-1, NULL); #endif /* PML_BFO */ } #if PML_BFO MCA_PML_BFO_SECOND_ERROR_CHECK_ON_RDMA_READ_COMPLETION(recvreq, status, btl); MCA_PML_BFO_CHECK_RECVREQ_RDMA_BML_BTL(bml_btl, btl, recvreq, "RDMA write"); #endif /* PML_BFO */ mca_pml_bfo_send_fin(recvreq->req_recv.req_base.req_proc, bml_btl, frag->rdma_hdr.hdr_rget.hdr_des, #if PML_BFO des->order, 0, (uint16_t)recvreq->req_msgseq, recvreq->req_restartseq, recvreq->req_recv.req_base.req_comm->c_contextid, recvreq->req_recv.req_base.req_comm->c_my_rank); #else /* PML_BFO */ des->order, 0); #endif /* PML_BFO */ /* is receive request complete */ OPAL_THREAD_ADD_SIZE_T(&recvreq->req_bytes_received, frag->rdma_length); recv_request_pml_complete_check(recvreq); MCA_PML_BFO_RDMA_FRAG_RETURN(frag); MCA_PML_BFO_PROGRESS_PENDING(bml_btl); } /* * */ int mca_pml_bfo_recv_request_get_frag( mca_pml_bfo_rdma_frag_t* frag ) { mca_pml_bfo_recv_request_t* recvreq = (mca_pml_bfo_recv_request_t*)frag->rdma_req; mca_bml_base_btl_t* bml_btl = frag->rdma_bml; mca_btl_base_descriptor_t* descriptor; size_t save_size = frag->rdma_length; int rc; /* prepare descriptor */ mca_bml_base_prepare_dst( bml_btl, NULL, &recvreq->req_recv.req_base.req_convertor, MCA_BTL_NO_ORDER, 0, &frag->rdma_length, MCA_BTL_DES_FLAGS_BTL_OWNERSHIP | MCA_BTL_DES_SEND_ALWAYS_CALLBACK | MCA_BTL_DES_FLAGS_GET, &descriptor ); if( OPAL_UNLIKELY(NULL == descriptor) ) { frag->rdma_length = save_size; OPAL_THREAD_LOCK(&mca_pml_bfo.lock); opal_list_append(&mca_pml_bfo.rdma_pending, (opal_list_item_t*)frag); OPAL_THREAD_UNLOCK(&mca_pml_bfo.lock); return OMPI_ERR_OUT_OF_RESOURCE; } descriptor->des_src = (mca_btl_base_segment_t *) frag->rdma_segs; descriptor->des_src_cnt = frag->rdma_hdr.hdr_rdma.hdr_seg_cnt; descriptor->des_cbfunc = mca_pml_bfo_rget_completion; descriptor->des_cbdata = frag; PERUSE_TRACE_COMM_OMPI_EVENT(PERUSE_COMM_REQ_XFER_CONTINUE, &(recvreq->req_recv.req_base), frag->rdma_length, PERUSE_RECV); /* queue up get request */ rc = mca_bml_base_get(bml_btl,descriptor); if( OPAL_UNLIKELY(OMPI_SUCCESS != rc) ) { if(OMPI_ERR_OUT_OF_RESOURCE == rc) { mca_bml_base_free(bml_btl, descriptor); OPAL_THREAD_LOCK(&mca_pml_bfo.lock); opal_list_append(&mca_pml_bfo.rdma_pending, (opal_list_item_t*)frag); OPAL_THREAD_UNLOCK(&mca_pml_bfo.lock); return OMPI_ERR_OUT_OF_RESOURCE; } else { OMPI_ERROR_LOG(rc); ompi_rte_abort(-1, NULL); } } #if PML_BFO if ((bml_btl->btl_flags & MCA_BTL_FLAGS_FAILOVER_SUPPORT) && (descriptor->des_flags & MCA_BTL_DES_SEND_ALWAYS_CALLBACK)) { recvreq->req_events++; } #endif /* PML_BFO */ return OMPI_SUCCESS; } /* * Update the recv request status to reflect the number of bytes * received and actually delivered to the application. */ void mca_pml_bfo_recv_request_progress_frag( mca_pml_bfo_recv_request_t* recvreq, mca_btl_base_module_t* btl, mca_btl_base_segment_t* segments, size_t num_segments ) { size_t bytes_received, data_offset = 0; size_t bytes_delivered __opal_attribute_unused__; /* is being set to zero in MCA_PML_BFO_RECV_REQUEST_UNPACK */ mca_pml_bfo_hdr_t* hdr = (mca_pml_bfo_hdr_t*)segments->seg_addr.pval; bytes_received = mca_pml_bfo_compute_segment_length_base (segments, num_segments, sizeof(mca_pml_bfo_frag_hdr_t)); data_offset = hdr->hdr_frag.hdr_frag_offset; /* * Make user buffer accessable(defined) before unpacking. */ 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); ); MCA_PML_BFO_RECV_REQUEST_UNPACK( recvreq, segments, num_segments, sizeof(mca_pml_bfo_frag_hdr_t), data_offset, bytes_received, bytes_delivered ); /* * Unpacking finished, make the user buffer unaccessable again. */ MEMCHECKER( memchecker_call(&opal_memchecker_base_mem_noaccess, recvreq->req_recv.req_base.req_addr, recvreq->req_recv.req_base.req_count, recvreq->req_recv.req_base.req_datatype); ); OPAL_THREAD_ADD_SIZE_T(&recvreq->req_bytes_received, bytes_received); /* check completion status */ if(recv_request_pml_complete_check(recvreq) == false && recvreq->req_rdma_offset < recvreq->req_send_offset) { /* schedule additional rdma operations */ mca_pml_bfo_recv_request_schedule(recvreq, NULL); } } /* * Update the recv request status to reflect the number of bytes * received and actually delivered to the application. */ void mca_pml_bfo_recv_request_progress_rget( mca_pml_bfo_recv_request_t* recvreq, mca_btl_base_module_t* btl, mca_btl_base_segment_t* segments, size_t num_segments ) { mca_pml_bfo_rget_hdr_t* hdr = (mca_pml_bfo_rget_hdr_t*)segments->seg_addr.pval; mca_bml_base_endpoint_t* bml_endpoint = NULL; mca_pml_bfo_rdma_frag_t* frag; size_t i, size = 0; recvreq->req_recv.req_bytes_packed = hdr->hdr_rndv.hdr_msg_length; #if PML_BFO recvreq->remote_req_send = hdr->hdr_rndv.hdr_src_req; #endif /* PML_BFO */ MCA_PML_BFO_RECV_REQUEST_MATCHED(recvreq, &hdr->hdr_rndv.hdr_match); /* if receive buffer is not contiguous we can't just RDMA read into it, so * fall back to copy in/out protocol. It is a pity because buffer on the * sender side is already registered. We need to be smarter here, perhaps * do couple of RDMA reads */ if(opal_convertor_need_buffers(&recvreq->req_recv.req_base.req_convertor) == true) { #if OPAL_CUDA_SUPPORT if (mca_pml_bfo_cuda_need_buffers(recvreq, btl)) { mca_pml_bfo_recv_request_ack(recvreq, &hdr->hdr_rndv, 0); return; } #else /* OPAL_CUDA_SUPPORT */ mca_pml_bfo_recv_request_ack(recvreq, &hdr->hdr_rndv, 0); return; #endif /* OPAL_CUDA_SUPPORT */ } MCA_PML_BFO_RDMA_FRAG_ALLOC(frag); if( OPAL_UNLIKELY(NULL == frag) ) { /* GLB - FIX */ OMPI_ERROR_LOG(OMPI_ERR_OUT_OF_RESOURCE); ompi_rte_abort(-1, NULL); } /* lookup bml datastructures */ bml_endpoint = (mca_bml_base_endpoint_t*)recvreq->req_recv.req_base.req_proc->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_BML]; assert (btl->btl_seg_size * hdr->hdr_seg_cnt <= sizeof (frag->rdma_segs)); /* allocate/initialize a fragment */ memmove (frag->rdma_segs, hdr + 1, btl->btl_seg_size * hdr->hdr_seg_cnt); for(i = 0; i < hdr->hdr_seg_cnt; i++) { mca_btl_base_segment_t *seg = (mca_btl_base_segment_t *)(frag->rdma_segs + i * btl->btl_seg_size); #if OPAL_ENABLE_HETEROGENEOUS_SUPPORT if ((recvreq->req_recv.req_base.req_proc->proc_arch & OPAL_ARCH_ISBIGENDIAN) != (ompi_proc_local()->proc_arch & OPAL_ARCH_ISBIGENDIAN)) { size += opal_swap_bytes4(seg->seg_len); } else #endif { size += seg->seg_len; } } #if PML_BFO frag->rdma_btl = btl; #endif /* PML_BFO */ frag->rdma_bml = mca_bml_base_btl_array_find(&bml_endpoint->btl_rdma, btl); #if OPAL_CUDA_SUPPORT if( OPAL_UNLIKELY(NULL == frag->rdma_bml) ) { if (recvreq->req_recv.req_base.req_convertor.flags & CONVERTOR_CUDA) { /* Check to see if this is a CUDA get */ if (btl->btl_flags & MCA_BTL_FLAGS_CUDA_GET) { frag->rdma_bml = mca_bml_base_btl_array_find(&bml_endpoint->btl_send, btl); } if( OPAL_UNLIKELY(NULL == frag->rdma_bml) ) { opal_output(0, "[%s:%d] invalid bml for rdma get", __FILE__, __LINE__); ompi_rte_abort(-1, NULL); } } else { /* Just default back to send and receive. Must be mix of GPU and HOST memory. */ mca_pml_bfo_recv_request_ack(recvreq, &hdr->hdr_rndv, 0); return; } } #else /* OPAL_CUDA_SUPPORT */ if( OPAL_UNLIKELY(NULL == frag->rdma_bml) ) { opal_output(0, "[%s:%d] invalid bml for rdma get", __FILE__, __LINE__); ompi_rte_abort(-1, NULL); } #endif /* OPAL_CUDA_SUPPORT */ frag->rdma_hdr.hdr_rget = *hdr; frag->rdma_req = recvreq; frag->rdma_ep = bml_endpoint; frag->rdma_length = size; frag->rdma_state = MCA_PML_BFO_RDMA_GET; frag->reg = NULL; mca_pml_bfo_recv_request_get_frag(frag); return; } /* * Update the recv request status to reflect the number of bytes * received and actually delivered to the application. */ void mca_pml_bfo_recv_request_progress_rndv( mca_pml_bfo_recv_request_t* recvreq, mca_btl_base_module_t* btl, mca_btl_base_segment_t* segments, size_t num_segments ) { size_t bytes_received; size_t bytes_delivered __opal_attribute_unused__; /* is being set to zero in MCA_PML_BFO_RECV_REQUEST_UNPACK */ size_t data_offset = 0; mca_pml_bfo_hdr_t* hdr = (mca_pml_bfo_hdr_t*)segments->seg_addr.pval; bytes_received = mca_pml_bfo_compute_segment_length_base (segments, num_segments, sizeof(mca_pml_bfo_rendezvous_hdr_t)); recvreq->req_recv.req_bytes_packed = hdr->hdr_rndv.hdr_msg_length; recvreq->remote_req_send = hdr->hdr_rndv.hdr_src_req; recvreq->req_rdma_offset = bytes_received; MCA_PML_BFO_RECV_REQUEST_MATCHED(recvreq, &hdr->hdr_match); mca_pml_bfo_recv_request_ack(recvreq, &hdr->hdr_rndv, bytes_received); /** * The PUT protocol do not attach any data to the original request. * Therefore, we might want to avoid unpacking if there is nothing to * unpack. */ if( 0 < bytes_received ) { 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); ); MCA_PML_BFO_RECV_REQUEST_UNPACK( recvreq, segments, num_segments, sizeof(mca_pml_bfo_rendezvous_hdr_t), data_offset, bytes_received, bytes_delivered ); MEMCHECKER( memchecker_call(&opal_memchecker_base_mem_noaccess, recvreq->req_recv.req_base.req_addr, recvreq->req_recv.req_base.req_count, recvreq->req_recv.req_base.req_datatype); ); } OPAL_THREAD_ADD_SIZE_T(&recvreq->req_bytes_received, bytes_received); /* check completion status */ if(recv_request_pml_complete_check(recvreq) == false && recvreq->req_rdma_offset < recvreq->req_send_offset) { /* schedule additional rdma operations */ mca_pml_bfo_recv_request_schedule(recvreq, NULL); } } /* * Update the recv request status to reflect the number of bytes * received and actually delivered to the application. */ void mca_pml_bfo_recv_request_progress_match( mca_pml_bfo_recv_request_t* recvreq, mca_btl_base_module_t* btl, mca_btl_base_segment_t* segments, size_t num_segments ) { size_t bytes_received, data_offset = 0; size_t bytes_delivered __opal_attribute_unused__; /* is being set to zero in MCA_PML_BFO_RECV_REQUEST_UNPACK */ mca_pml_bfo_hdr_t* hdr = (mca_pml_bfo_hdr_t*)segments->seg_addr.pval; bytes_received = mca_pml_bfo_compute_segment_length_base (segments, num_segments, OMPI_PML_BFO_MATCH_HDR_LEN); recvreq->req_recv.req_bytes_packed = bytes_received; MCA_PML_BFO_RECV_REQUEST_MATCHED(recvreq, &hdr->hdr_match); /* * Make user buffer accessable(defined) before unpacking. */ 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); ); MCA_PML_BFO_RECV_REQUEST_UNPACK( recvreq, segments, num_segments, OMPI_PML_BFO_MATCH_HDR_LEN, data_offset, bytes_received, bytes_delivered); /* * Unpacking finished, make the user buffer unaccessable again. */ MEMCHECKER( memchecker_call(&opal_memchecker_base_mem_noaccess, recvreq->req_recv.req_base.req_addr, recvreq->req_recv.req_base.req_count, recvreq->req_recv.req_base.req_datatype); ); /* * No need for atomic here, as we know there is only one fragment * for this request. */ recvreq->req_bytes_received += bytes_received; recv_request_pml_complete(recvreq); } /** * Handle completion of a probe request */ void mca_pml_bfo_recv_request_matched_probe( mca_pml_bfo_recv_request_t* recvreq, mca_btl_base_module_t* btl, mca_btl_base_segment_t* segments, size_t num_segments ) { size_t bytes_packed = 0; mca_pml_bfo_hdr_t* hdr = (mca_pml_bfo_hdr_t*)segments->seg_addr.pval; switch(hdr->hdr_common.hdr_type) { case MCA_PML_BFO_HDR_TYPE_MATCH: bytes_packed = mca_pml_bfo_compute_segment_length_base (segments, num_segments, OMPI_PML_BFO_MATCH_HDR_LEN); break; case MCA_PML_BFO_HDR_TYPE_RNDV: case MCA_PML_BFO_HDR_TYPE_RGET: bytes_packed = hdr->hdr_rndv.hdr_msg_length; break; } /* set completion status */ recvreq->req_recv.req_base.req_ompi.req_status.MPI_TAG = hdr->hdr_match.hdr_tag; recvreq->req_recv.req_base.req_ompi.req_status.MPI_SOURCE = hdr->hdr_match.hdr_src; recvreq->req_bytes_received = bytes_packed; recvreq->req_bytes_expected = bytes_packed; recv_request_pml_complete(recvreq); } /* * Schedule RDMA protocol. * */ int mca_pml_bfo_recv_request_schedule_once( mca_pml_bfo_recv_request_t* recvreq, mca_bml_base_btl_t *start_bml_btl ) { mca_bml_base_btl_t* bml_btl; int num_tries = recvreq->req_rdma_cnt, num_fail = 0; size_t i, prev_bytes_remaining = 0; size_t bytes_remaining = recvreq->req_send_offset - recvreq->req_rdma_offset; /* if starting bml_btl is provided schedule next fragment on it first */ if(start_bml_btl != NULL) { for(i = 0; i < recvreq->req_rdma_cnt; i++) { if(recvreq->req_rdma[i].bml_btl != start_bml_btl) continue; /* something left to be send? */ if( OPAL_LIKELY(recvreq->req_rdma[i].length) ) recvreq->req_rdma_idx = i; break; } } while(bytes_remaining > 0 && recvreq->req_pipeline_depth < mca_pml_bfo.recv_pipeline_depth) { size_t size, seg_size; mca_pml_bfo_rdma_hdr_t* hdr; mca_btl_base_descriptor_t* dst; mca_btl_base_descriptor_t* ctl; mca_mpool_base_registration_t * reg = NULL; mca_btl_base_module_t* btl; int rc, rdma_idx; if(prev_bytes_remaining == bytes_remaining) { if(++num_fail == num_tries) { OPAL_THREAD_LOCK(&mca_pml_bfo.lock); if(false == recvreq->req_pending) { opal_list_append(&mca_pml_bfo.recv_pending, (opal_list_item_t*)recvreq); recvreq->req_pending = true; } OPAL_THREAD_UNLOCK(&mca_pml_bfo.lock); return OMPI_ERR_OUT_OF_RESOURCE; } } else { num_fail = 0; prev_bytes_remaining = bytes_remaining; } do { rdma_idx = recvreq->req_rdma_idx; bml_btl = recvreq->req_rdma[rdma_idx].bml_btl; reg = recvreq->req_rdma[rdma_idx].btl_reg; size = recvreq->req_rdma[rdma_idx].length; if(++recvreq->req_rdma_idx >= recvreq->req_rdma_cnt) recvreq->req_rdma_idx = 0; } while(!size); btl = bml_btl->btl; /* makes sure that we don't exceed BTL max rdma size * if memory is not pinned already */ if( (NULL == reg) && (btl->btl_rdma_pipeline_frag_size != 0) && (size > btl->btl_rdma_pipeline_frag_size)) { size = btl->btl_rdma_pipeline_frag_size; } /* take lock to protect converter against concurrent access * from unpack */ OPAL_THREAD_LOCK(&recvreq->lock); opal_convertor_set_position( &recvreq->req_recv.req_base.req_convertor, &recvreq->req_rdma_offset ); /* prepare a descriptor for RDMA */ mca_bml_base_prepare_dst(bml_btl, reg, &recvreq->req_recv.req_base.req_convertor, MCA_BTL_NO_ORDER, 0, &size, MCA_BTL_DES_FLAGS_BTL_OWNERSHIP | MCA_BTL_DES_FLAGS_PUT, &dst); OPAL_THREAD_UNLOCK(&recvreq->lock); if(OPAL_UNLIKELY(dst == NULL)) { continue; } dst->des_cbfunc = mca_pml_bfo_put_completion; dst->des_cbdata = recvreq; seg_size = btl->btl_seg_size * dst->des_dst_cnt; /* prepare a descriptor for rdma control message */ mca_bml_base_alloc(bml_btl, &ctl, MCA_BTL_NO_ORDER, sizeof(mca_pml_bfo_rdma_hdr_t) + seg_size, MCA_BTL_DES_FLAGS_PRIORITY | MCA_BTL_DES_FLAGS_BTL_OWNERSHIP | MCA_BTL_DES_SEND_ALWAYS_CALLBACK); if( OPAL_UNLIKELY(NULL == ctl) ) { mca_bml_base_free(bml_btl,dst); continue; } ctl->des_cbfunc = mca_pml_bfo_recv_ctl_completion; #if PML_BFO ctl->des_cbdata = recvreq; #endif /* PML_BFO */ /* fill in rdma header */ hdr = (mca_pml_bfo_rdma_hdr_t*)ctl->des_src->seg_addr.pval; hdr->hdr_common.hdr_type = MCA_PML_BFO_HDR_TYPE_PUT; hdr->hdr_common.hdr_flags = (!recvreq->req_ack_sent) ? MCA_PML_BFO_HDR_TYPE_ACK : 0; hdr->hdr_req = recvreq->remote_req_send; #if PML_BFO hdr->hdr_dst_req.pval = recvreq; /* only needed in the first put message */ #endif /* PML_BFO */ hdr->hdr_des.pval = dst; hdr->hdr_rdma_offset = recvreq->req_rdma_offset; hdr->hdr_seg_cnt = dst->des_dst_cnt; /* copy segments */ memmove (hdr + 1, dst->des_dst, seg_size); if(!recvreq->req_ack_sent) recvreq->req_ack_sent = true; bfo_hdr_hton(hdr, MCA_PML_BFO_HDR_TYPE_PUT, recvreq->req_recv.req_base.req_proc); PERUSE_TRACE_COMM_OMPI_EVENT( PERUSE_COMM_REQ_XFER_CONTINUE, &(recvreq->req_recv.req_base), size, PERUSE_RECV); /* send rdma request to peer */ rc = mca_bml_base_send(bml_btl, ctl, MCA_PML_BFO_HDR_TYPE_PUT); if( OPAL_LIKELY( rc >= 0 ) ) { #if PML_BFO if ((btl->btl_flags & MCA_BTL_FLAGS_FAILOVER_SUPPORT) && (ctl->des_flags & MCA_BTL_DES_SEND_ALWAYS_CALLBACK)) { recvreq->req_events++; } #endif /* PML_BFO */ /* update request state */ recvreq->req_rdma_offset += size; OPAL_THREAD_ADD_SIZE_T(&recvreq->req_pipeline_depth, 1); recvreq->req_rdma[rdma_idx].length -= size; bytes_remaining -= size; } else { mca_bml_base_free(bml_btl,ctl); mca_bml_base_free(bml_btl,dst); } } return OMPI_SUCCESS; } #define IS_PROB_REQ(R) \ ((MCA_PML_REQUEST_IPROBE == (R)->req_recv.req_base.req_type) || \ (MCA_PML_REQUEST_PROBE == (R)->req_recv.req_base.req_type) || \ (MCA_PML_REQUEST_IMPROBE == (R)->req_recv.req_base.req_type) || \ (MCA_PML_REQUEST_MPROBE == (R)->req_recv.req_base.req_type)) #define IS_MPROB_REQ(R) \ ((MCA_PML_REQUEST_IMPROBE == (R)->req_recv.req_base.req_type) || \ (MCA_PML_REQUEST_MPROBE == (R)->req_recv.req_base.req_type)) static inline void append_recv_req_to_queue(opal_list_t *queue, mca_pml_bfo_recv_request_t *req) { if(OPAL_UNLIKELY(req->req_recv.req_base.req_type == MCA_PML_REQUEST_IPROBE || req->req_recv.req_base.req_type == MCA_PML_REQUEST_IMPROBE)) return; opal_list_append(queue, (opal_list_item_t*)req); /** * We don't want to generate this kind of event for MPI_Probe. Hopefully, * the compiler will optimize out the empty if loop in the case where PERUSE * support is not required by the user. */ if(req->req_recv.req_base.req_type != MCA_PML_REQUEST_PROBE || req->req_recv.req_base.req_type != MCA_PML_REQUEST_MPROBE) { PERUSE_TRACE_COMM_EVENT(PERUSE_COMM_REQ_INSERT_IN_POSTED_Q, &(req->req_recv.req_base), PERUSE_RECV); } } /* * this routine tries to match a posted receive. If a match is found, * it places the request in the appropriate matched receive list. This * function has to be called with the communicator matching lock held. */ static mca_pml_bfo_recv_frag_t* recv_req_match_specific_proc( const mca_pml_bfo_recv_request_t *req, mca_pml_bfo_comm_proc_t *proc ) { opal_list_t* unexpected_frags = &proc->unexpected_frags; opal_list_item_t *i; mca_pml_bfo_recv_frag_t* frag; int tag = req->req_recv.req_base.req_tag; if(opal_list_get_size(unexpected_frags) == 0) return NULL; if( OMPI_ANY_TAG == tag ) { for (i = opal_list_get_first(unexpected_frags); i != opal_list_get_end(unexpected_frags); i = opal_list_get_next(i)) { frag = (mca_pml_bfo_recv_frag_t*)i; if( frag->hdr.hdr_match.hdr_tag >= 0 ) return frag; } } else { for (i = opal_list_get_first(unexpected_frags); i != opal_list_get_end(unexpected_frags); i = opal_list_get_next(i)) { frag = (mca_pml_bfo_recv_frag_t*)i; if( frag->hdr.hdr_match.hdr_tag == tag ) return frag; } } return NULL; } /* * this routine is used to try and match a wild posted receive - where * wild is determined by the value assigned to the source process */ static mca_pml_bfo_recv_frag_t* recv_req_match_wild( mca_pml_bfo_recv_request_t* req, mca_pml_bfo_comm_proc_t **p) { mca_pml_bfo_comm_t* comm = req->req_recv.req_base.req_comm->c_pml_comm; mca_pml_bfo_comm_proc_t* proc = comm->procs; size_t i; /* * Loop over all the outstanding messages to find one that matches. * There is an outer loop over lists of messages from each * process, then an inner loop over the messages from the * process. * * In order to avoid starvation do this in a round-robin fashion. */ for (i = comm->last_probed + 1; i < comm->num_procs; i++) { mca_pml_bfo_recv_frag_t* frag; /* loop over messages from the current proc */ if((frag = recv_req_match_specific_proc(req, &proc[i]))) { *p = &proc[i]; comm->last_probed = i; req->req_recv.req_base.req_proc = proc[i].ompi_proc; prepare_recv_req_converter(req); return frag; /* match found */ } } for (i = 0; i <= comm->last_probed; i++) { mca_pml_bfo_recv_frag_t* frag; /* loop over messages from the current proc */ if((frag = recv_req_match_specific_proc(req, &proc[i]))) { *p = &proc[i]; comm->last_probed = i; req->req_recv.req_base.req_proc = proc[i].ompi_proc; prepare_recv_req_converter(req); return frag; /* match found */ } } *p = NULL; return NULL; } void mca_pml_bfo_recv_req_start(mca_pml_bfo_recv_request_t *req) { mca_pml_bfo_comm_t* comm = req->req_recv.req_base.req_comm->c_pml_comm; mca_pml_bfo_comm_proc_t* proc; mca_pml_bfo_recv_frag_t* frag; opal_list_t *queue; mca_pml_bfo_hdr_t* hdr; /* init/re-init the request */ req->req_lock = 0; req->req_pipeline_depth = 0; req->req_bytes_received = 0; req->req_bytes_expected = 0; /* What about req_rdma_cnt ? */ #if PML_BFO req->req_rdma_cnt = 0; req->req_events = 0; req->req_restartseq = 0; req->req_errstate = 0; #endif /* PML_BFO */ req->req_rdma_idx = 0; req->req_pending = false; req->req_ack_sent = false; MCA_PML_BASE_RECV_START(&req->req_recv.req_base); OPAL_THREAD_LOCK(&comm->matching_lock); /** * The laps of time between the ACTIVATE event and the SEARCH_UNEX one include * the cost of the request lock. */ PERUSE_TRACE_COMM_EVENT(PERUSE_COMM_SEARCH_UNEX_Q_BEGIN, &(req->req_recv.req_base), PERUSE_RECV); /* assign sequence number */ req->req_recv.req_base.req_sequence = comm->recv_sequence++; /* attempt to match posted recv */ if(req->req_recv.req_base.req_peer == OMPI_ANY_SOURCE) { frag = recv_req_match_wild(req, &proc); queue = &comm->wild_receives; #if !OPAL_ENABLE_HETEROGENEOUS_SUPPORT /* As we are in a homogeneous environment we know that all remote * architectures are exactly the same as the local one. Therefore, * we can safely construct the convertor based on the proc * information of rank 0. */ if( NULL == frag ) { req->req_recv.req_base.req_proc = ompi_proc_local_proc; prepare_recv_req_converter(req); } #endif /* !OPAL_ENABLE_HETEROGENEOUS_SUPPORT */ } else { proc = &comm->procs[req->req_recv.req_base.req_peer]; req->req_recv.req_base.req_proc = proc->ompi_proc; frag = recv_req_match_specific_proc(req, proc); queue = &proc->specific_receives; /* wild cardrecv will be prepared on match */ prepare_recv_req_converter(req); } if(OPAL_UNLIKELY(NULL == frag)) { PERUSE_TRACE_COMM_EVENT(PERUSE_COMM_SEARCH_UNEX_Q_END, &(req->req_recv.req_base), PERUSE_RECV); /* We didn't find any matches. Record this irecv so we can match it when the message comes in. */ append_recv_req_to_queue(queue, req); req->req_match_received = false; OPAL_THREAD_UNLOCK(&comm->matching_lock); } else { if(OPAL_LIKELY(!IS_PROB_REQ(req))) { PERUSE_TRACE_COMM_EVENT(PERUSE_COMM_REQ_MATCH_UNEX, &(req->req_recv.req_base), PERUSE_RECV); hdr = (mca_pml_bfo_hdr_t*)frag->segments->seg_addr.pval; PERUSE_TRACE_MSG_EVENT(PERUSE_COMM_MSG_REMOVE_FROM_UNEX_Q, req->req_recv.req_base.req_comm, hdr->hdr_match.hdr_src, hdr->hdr_match.hdr_tag, PERUSE_RECV); PERUSE_TRACE_COMM_EVENT(PERUSE_COMM_SEARCH_UNEX_Q_END, &(req->req_recv.req_base), PERUSE_RECV); opal_list_remove_item(&proc->unexpected_frags, (opal_list_item_t*)frag); OPAL_THREAD_UNLOCK(&comm->matching_lock); switch(hdr->hdr_common.hdr_type) { case MCA_PML_BFO_HDR_TYPE_MATCH: mca_pml_bfo_recv_request_progress_match(req, frag->btl, frag->segments, frag->num_segments); break; case MCA_PML_BFO_HDR_TYPE_RNDV: mca_pml_bfo_recv_request_progress_rndv(req, frag->btl, frag->segments, frag->num_segments); break; case MCA_PML_BFO_HDR_TYPE_RGET: mca_pml_bfo_recv_request_progress_rget(req, frag->btl, frag->segments, frag->num_segments); break; default: assert(0); } MCA_PML_BFO_RECV_FRAG_RETURN(frag); } else if (OPAL_UNLIKELY(IS_MPROB_REQ(req))) { /* Remove the fragment from the match list, as it's now matched. Stash it somewhere in the request (which, yes, is a complete hack), where it will be plucked out during the end of mprobe. The request will then be "recreated" as a receive request, and the frag will be restarted with this request during mrecv */ opal_list_remove_item(&proc->unexpected_frags, (opal_list_item_t*)frag); OPAL_THREAD_UNLOCK(&comm->matching_lock); req->req_recv.req_base.req_addr = frag; mca_pml_bfo_recv_request_matched_probe(req, frag->btl, frag->segments, frag->num_segments); } else { OPAL_THREAD_UNLOCK(&comm->matching_lock); mca_pml_bfo_recv_request_matched_probe(req, frag->btl, frag->segments, frag->num_segments); } } }