/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */ /* * Copyright (c) 2004-2010 The Trustees of Indiana University and Indiana * University Research and Technology * Corporation. All rights reserved. * Copyright (c) 2004-2022 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 (c) 2008 UT-Battelle, LLC. All rights reserved. * Copyright (c) 2006-2008 University of Houston. All rights reserved. * Copyright (c) 2009-2010 Oracle and/or its affiliates. All rights reserved * Copyright (c) 2011 Sandia National Laboratories. All rights reserved. * Copyright (c) 2011-2015 Los Alamos National Security, LLC. All rights * reserved. * Copyright (c) 2012 Cisco Systems, Inc. All rights reserved. * Copyright (c) 2015 FUJITSU LIMITED. All rights reserved. * Copyright (c) 2018 Sandia National Laboratories * All rights reserved. * Copyright (c) 2018 IBM Corporation. All rights reserved. * Copyright (c) 2019-2020 Intel, Inc. All rights reserved. * Copyright (c) 2021 Nanook Consulting. All rights reserved. * Copyright (c) 2018-2021 Triad National Security, LLC. All rights * reserved. * Copyright (c) 2022 IBM Corporation. All rights reserved * Copyright (c) 2023 Jeffrey M. Squyres. All rights reserved. * $COPYRIGHT$ * * Additional copyrights may follow * * $HEADER$ */ #include "ompi_config.h" #include #include #include "opal/class/opal_bitmap.h" #include "opal/util/output.h" #include "opal/util/show_help.h" #include "opal_stdint.h" #include "opal/mca/btl/btl.h" #include "opal/mca/btl/base/base.h" #include "ompi/mca/pml/pml.h" #include "ompi/mca/pml/base/base.h" #include "ompi/mca/pml/base/base.h" #include "ompi/mca/bml/base/base.h" #include "ompi/errhandler/errhandler.h" #include "opal/mca/pmix/pmix-internal.h" #include "ompi/runtime/ompi_spc.h" #include "pml_ob1.h" #include "pml_ob1_component.h" #include "pml_ob1_comm.h" #include "pml_ob1_hdr.h" #include "pml_ob1_recvfrag.h" #include "pml_ob1_sendreq.h" #include "pml_ob1_recvreq.h" #include "pml_ob1_rdmafrag.h" #include "pml_ob1_accelerator.h" mca_pml_ob1_t mca_pml_ob1 = { { .pml_add_procs = mca_pml_ob1_add_procs, .pml_del_procs = mca_pml_ob1_del_procs, .pml_enable = mca_pml_ob1_enable, .pml_progress = NULL, /* mca_pml_ob1_progress, */ .pml_add_comm = mca_pml_ob1_add_comm, .pml_del_comm = mca_pml_ob1_del_comm, #if OPAL_ENABLE_FT_MPI .pml_revoke_comm = mca_pml_ob1_revoke_comm, #else .pml_revoke_comm = NULL, #endif .pml_irecv_init = mca_pml_ob1_irecv_init, .pml_irecv = mca_pml_ob1_irecv, .pml_recv = mca_pml_ob1_recv, .pml_isend_init = mca_pml_ob1_isend_init, .pml_isend = mca_pml_ob1_isend, .pml_send = mca_pml_ob1_send, .pml_iprobe = mca_pml_ob1_iprobe, .pml_probe = mca_pml_ob1_probe, .pml_start = mca_pml_ob1_start, .pml_improbe = mca_pml_ob1_improbe, .pml_mprobe = mca_pml_ob1_mprobe, .pml_imrecv = mca_pml_ob1_imrecv, .pml_mrecv = mca_pml_ob1_mrecv, .pml_dump = mca_pml_ob1_dump, .pml_max_contextid = 65535, .pml_max_tag = INT_MAX, .pml_flags = 0 /* flags */ } }; extern void mca_pml_ob1_accelerator_add_ipc_support(struct mca_btl_base_module_t* btl, int32_t flags, ompi_proc_t* errproc, char* btlinfo); void mca_pml_ob1_error_handler( struct mca_btl_base_module_t* btl, int32_t flags, opal_proc_t* errproc, char* btlinfo ); int mca_pml_ob1_enable(bool enable) { if( false == enable ) { return OMPI_SUCCESS; } OBJ_CONSTRUCT(&mca_pml_ob1.lock, opal_mutex_t); /* fragments */ OBJ_CONSTRUCT(&mca_pml_ob1.rdma_frags, opal_free_list_t); opal_free_list_init ( &mca_pml_ob1.rdma_frags, sizeof(mca_pml_ob1_rdma_frag_t), opal_cache_line_size, OBJ_CLASS(mca_pml_ob1_rdma_frag_t), 0,opal_cache_line_size, mca_pml_ob1.free_list_num, mca_pml_ob1.free_list_max, mca_pml_ob1.free_list_inc, NULL, 0, NULL, NULL, NULL); OBJ_CONSTRUCT(&mca_pml_ob1.recv_frags, opal_free_list_t); opal_free_list_init ( &mca_pml_ob1.recv_frags, sizeof(mca_pml_ob1_recv_frag_t) + mca_pml_ob1.unexpected_limit, opal_cache_line_size, OBJ_CLASS(mca_pml_ob1_recv_frag_t), 0,opal_cache_line_size, mca_pml_ob1.free_list_num, mca_pml_ob1.free_list_max, mca_pml_ob1.free_list_inc, NULL, 0, NULL, NULL, NULL); OBJ_CONSTRUCT(&mca_pml_ob1.pending_pckts, opal_free_list_t); opal_free_list_init ( &mca_pml_ob1.pending_pckts, sizeof(mca_pml_ob1_pckt_pending_t), opal_cache_line_size, OBJ_CLASS(mca_pml_ob1_pckt_pending_t), 0,opal_cache_line_size, mca_pml_ob1.free_list_num, mca_pml_ob1.free_list_max, mca_pml_ob1.free_list_inc, NULL, 0, NULL, NULL, NULL); OBJ_CONSTRUCT(&mca_pml_ob1.buffers, opal_free_list_t); OBJ_CONSTRUCT(&mca_pml_ob1.send_ranges, opal_free_list_t); opal_free_list_init ( &mca_pml_ob1.send_ranges, sizeof(mca_pml_ob1_send_range_t) + sizeof(mca_pml_ob1_com_btl_t[mca_pml_ob1.max_send_per_range]), opal_cache_line_size, OBJ_CLASS(mca_pml_ob1_send_range_t), 0,opal_cache_line_size, mca_pml_ob1.free_list_num, mca_pml_ob1.free_list_max, mca_pml_ob1.free_list_inc, NULL, 0, NULL, NULL, NULL); /* pending operations */ OBJ_CONSTRUCT(&mca_pml_ob1.send_pending, opal_list_t); OBJ_CONSTRUCT(&mca_pml_ob1.recv_pending, opal_list_t); OBJ_CONSTRUCT(&mca_pml_ob1.pckt_pending, opal_list_t); OBJ_CONSTRUCT(&mca_pml_ob1.rdma_pending, opal_list_t); /* missing communicator pending list */ OBJ_CONSTRUCT(&mca_pml_ob1.non_existing_communicator_pending, opal_list_t); /** * If we get here this is the PML who get selected for the run. We * should get ownership for the send and receive requests list, and * initialize them with the size of our own requests. */ opal_free_list_init ( &mca_pml_base_send_requests, sizeof(mca_pml_ob1_send_request_t) + sizeof(mca_pml_ob1_com_btl_t[mca_pml_ob1.max_rdma_per_request]), opal_cache_line_size, OBJ_CLASS(mca_pml_ob1_send_request_t), 0,opal_cache_line_size, mca_pml_ob1.free_list_num, mca_pml_ob1.free_list_max, mca_pml_ob1.free_list_inc, NULL, 0, NULL, NULL, NULL); opal_free_list_init ( &mca_pml_base_recv_requests, sizeof(mca_pml_ob1_recv_request_t) + sizeof(mca_pml_ob1_com_btl_t[mca_pml_ob1.max_rdma_per_request]), opal_cache_line_size, OBJ_CLASS(mca_pml_ob1_recv_request_t), 0,opal_cache_line_size, mca_pml_ob1.free_list_num, mca_pml_ob1.free_list_max, mca_pml_ob1.free_list_inc, NULL, 0, NULL, NULL, NULL); mca_pml_ob1.accelerator_enabled = (0 == mca_pml_ob1_accelerator_init()) ? true : false; mca_pml_ob1.enabled = true; return OMPI_SUCCESS; } static const char* mca_pml_ob1_set_allow_overtake(opal_infosubscriber_t* obj, const char* key, const char* value) { ompi_communicator_t *ompi_comm = (ompi_communicator_t *) obj; bool allow_overtake_was_set = OMPI_COMM_CHECK_ASSERT_ALLOW_OVERTAKE(ompi_comm); /* As we keep the out-of-sequence messages ordered by their sequence, as a receiver we * can just move the previously considered out-of-order messages into the unexpected queue, * and we maintain some form of logical consistency with the message order. */ if (opal_str_to_bool(value)) { if (!allow_overtake_was_set) { ompi_comm->c_flags |= OMPI_COMM_ASSERT_ALLOW_OVERTAKE; mca_pml_ob1_merge_cant_match(ompi_comm); } return "true"; } if (allow_overtake_was_set) { /* However, in the case we are trying to turn off allow_overtake, it is not clear what * should be done with the previous messages that are pending on our peers, nor with * the messages currently in the network. Similarly, if one process turns off allow * overtake, before any potential sender start sending valid sequence numbers there * is no way to order the messages in a sensible order. * The possible solution is cumbersome, it would force a network quiescence followed by * a synchronization of all processes in the communicator, and then all peers will * start sending messages starting with sequence number 0. * A lot of code for minimal benefit, especially taking in account that the MPI standard * does not define this. Instead, refuse to disable allow overtake, and at least the * user has the opportunity to check if we accepted to change it. */ return "true"; } return "false"; } int mca_pml_ob1_add_comm(ompi_communicator_t* comm) { /* allocate pml specific comm data */ mca_pml_ob1_comm_t* pml_comm = OBJ_NEW(mca_pml_ob1_comm_t); mca_pml_ob1_recv_frag_t *frag, *next_frag; mca_pml_ob1_comm_proc_t* pml_proc; mca_pml_ob1_match_hdr_t* hdr; if (NULL == pml_comm) { return OMPI_ERR_OUT_OF_RESOURCE; } /* should never happen, but it was, so check */ if (comm->c_index > mca_pml_ob1.super.pml_max_contextid) { OBJ_RELEASE(pml_comm); return OMPI_ERR_OUT_OF_RESOURCE; } ompi_comm_assert_subscribe (comm, OMPI_COMM_ASSERT_NO_ANY_SOURCE); mca_pml_ob1_comm_init_size(pml_comm, comm->c_remote_group->grp_proc_count); comm->c_pml_comm = pml_comm; /* Register the subscriber alert for the mpi_assert_allow_overtaking info. */ opal_infosubscribe_subscribe (&comm->super, "mpi_assert_allow_overtaking", "false", mca_pml_ob1_set_allow_overtake); /* Grab all related messages from the non_existing_communicator pending queue */ OPAL_LIST_FOREACH_SAFE(frag, next_frag, &mca_pml_ob1.non_existing_communicator_pending, mca_pml_ob1_recv_frag_t) { hdr = &frag->hdr.hdr_match; if (MCA_PML_OB1_HDR_TYPE_CID == frag->hdr.hdr_common.hdr_type) { if (!ompi_comm_cid_compare (comm, frag->hdr.hdr_cid.hdr_cid)) { continue; } /* handle this CID*/ mca_pml_ob1_handle_cid (comm, frag->hdr.hdr_ext_match.hdr_match.hdr_src, &frag->hdr.hdr_cid); hdr = &frag->hdr.hdr_ext_match.hdr_match; hdr->hdr_ctx = comm->c_index; /* NTH: this is ok because the pointer that will be freed is stored in frag->addr[] */ frag->segments[0].seg_addr.pval = (void *)((uintptr_t) frag->segments[0].seg_addr.pval + sizeof (frag->hdr.hdr_cid)); } /* Is this fragment for the current communicator ? */ if (hdr->hdr_ctx != comm->c_index) { continue; } /* As we now know we work on a fragment for this communicator * we should remove it from the * non_existing_communicator_pending list. */ opal_list_remove_item (&mca_pml_ob1.non_existing_communicator_pending, (opal_list_item_t *) frag); /* We generate the MSG_ARRIVED event as soon as the PML is aware * of a matching fragment arrival. Independing if it is received * on the correct order or not. This will allow the tools to * figure out if the messages are not received in the correct * order (if multiple network interfaces). */ PERUSE_TRACE_MSG_EVENT(PERUSE_COMM_MSG_ARRIVED, comm, hdr->hdr_src, hdr->hdr_tag, PERUSE_RECV); /* There is no matching to be done, and no lock to be held on the communicator as * we know at this point that the communicator has not yet been returned to the user. * The only required protection is around the non_existing_communicator_pending queue. * We just have to push the fragment into the unexpected list of the corresponding * proc, or into the out-of-order (cant_match) list. */ pml_proc = mca_pml_ob1_peer_lookup(comm, hdr->hdr_src); if (OMPI_COMM_CHECK_ASSERT_ALLOW_OVERTAKE(comm)) { #if !MCA_PML_OB1_CUSTOM_MATCH opal_list_append( &pml_proc->unexpected_frags, (opal_list_item_t*)frag ); #else custom_match_umq_append(pml_comm->umq, hdr->hdr_tag, hdr->hdr_src, frag); #endif PERUSE_TRACE_MSG_EVENT(PERUSE_COMM_MSG_INSERT_IN_UNEX_Q, comm, hdr->hdr_src, hdr->hdr_tag, PERUSE_RECV); continue; } if (((uint16_t)hdr->hdr_seq) == ((uint16_t)pml_proc->expected_sequence) ) { add_fragment_to_unexpected: /* We're now expecting the next sequence number. */ pml_proc->expected_sequence++; #if !MCA_PML_OB1_CUSTOM_MATCH opal_list_append( &pml_proc->unexpected_frags, (opal_list_item_t*)frag ); #else custom_match_umq_append(pml_comm->umq, hdr->hdr_tag, hdr->hdr_src, frag); #endif PERUSE_TRACE_MSG_EVENT(PERUSE_COMM_MSG_INSERT_IN_UNEX_Q, comm, hdr->hdr_src, hdr->hdr_tag, PERUSE_RECV); /* And now the ugly part. As some fragments can be inserted in the cant_match list, * every time we successfully add a fragment in the unexpected list we have to make * sure the next one is not in the cant_match. Otherwise, we will endup in a deadlock * situation as the cant_match is only checked when a new fragment is received from * the network. */ if( NULL != pml_proc->frags_cant_match ) { frag = ompi_pml_ob1_check_cantmatch_for_match(pml_proc); if( NULL != frag ) { hdr = &frag->hdr.hdr_match; goto add_fragment_to_unexpected; } } } else { ompi_pml_ob1_append_frag_to_ordered_list(&pml_proc->frags_cant_match, frag, pml_proc->expected_sequence); } } return OMPI_SUCCESS; } int mca_pml_ob1_del_comm(ompi_communicator_t* comm) { OBJ_RELEASE(comm->c_pml_comm); comm->c_pml_comm = NULL; return OMPI_SUCCESS; } /* * For each proc setup a datastructure that indicates the BTLs * that can be used to reach the destination. * */ int mca_pml_ob1_add_procs(ompi_proc_t** procs, size_t nprocs) { mca_btl_base_selected_module_t *sm; opal_bitmap_t reachable; int rc; if(nprocs == 0) return OMPI_SUCCESS; /* make sure remote procs are using the same PML as us */ if (OMPI_SUCCESS != (rc = mca_pml_base_pml_check_selected("ob1", procs, nprocs))) { return rc; } if (NULL == mca_bml.bml_add_procs) { return OMPI_ERR_UNREACH; } OBJ_CONSTRUCT(&reachable, opal_bitmap_t); rc = opal_bitmap_init(&reachable, (int)nprocs); if (OMPI_SUCCESS != rc) { return rc; } rc = mca_bml.bml_add_procs (nprocs, procs, &reachable); OBJ_DESTRUCT(&reachable); if (OMPI_SUCCESS != rc) { return rc; } /* Check that values supplied by all initialized btls will work for us. Note that this is the list of all initialized BTLs, not the ones used for the just added procs. This is a little overkill and inaccurate, as we may end up not using the BTL in question and all add_procs calls after the first one are duplicating an already completed check. But the final initialization of the PML occurs before the final initialization of the BTLs, and iterating through the in-use BTLs requires iterating over the procs, as the BML does not expose all currently in use btls. */ OPAL_LIST_FOREACH(sm, &mca_btl_base_modules_initialized, mca_btl_base_selected_module_t) { if ((MCA_BTL_FLAGS_SEND & sm->btl_module->btl_flags) && sm->btl_module->btl_eager_limit < sizeof(mca_pml_ob1_hdr_t)) { opal_show_help("help-mpi-pml-ob1.txt", "eager_limit_too_small", true, sm->btl_component->btl_version.mca_component_name, ompi_process_info.nodename, sm->btl_component->btl_version.mca_component_name, sm->btl_module->btl_eager_limit, sm->btl_component->btl_version.mca_component_name, sizeof(mca_pml_ob1_hdr_t), sm->btl_component->btl_version.mca_component_name); return OMPI_ERR_BAD_PARAM; } #if OPAL_CUDA_GDR_SUPPORT /* If size is SIZE_MAX, then we know we want to set this to the minimum possible * value which is the size of the PML header. */ if (SIZE_MAX == sm->btl_module->btl_accelerator_eager_limit) { sm->btl_module->btl_accelerator_eager_limit = sizeof(mca_pml_ob1_hdr_t); } /* If size is 0, then this value is unused. If it is non-zero then do some * extra checking of it. */ if (0 != sm->btl_module->btl_accelerator_eager_limit) { if (sm->btl_module->btl_accelerator_eager_limit < sizeof(mca_pml_ob1_hdr_t)) { opal_show_help("help-mpi-pml-ob1.txt", "cuda_eager_limit_too_small", true, sm->btl_component->btl_version.mca_component_name, ompi_process_info.nodename, sm->btl_component->btl_version.mca_component_name, sm->btl_module->btl_accelerator_eager_limit, sm->btl_component->btl_version.mca_component_name, sizeof(mca_pml_ob1_hdr_t), sm->btl_component->btl_version.mca_component_name); return OMPI_ERR_BAD_PARAM; } } if (0 == sm->btl_module->btl_accelerator_rdma_limit) { /* All is fine. 0 means to ignore value so set to SIZE_MAX */ sm->btl_module->btl_accelerator_rdma_limit = SIZE_MAX; } else { if (sm->btl_module->btl_accelerator_rdma_limit < sm->btl_module->btl_accelerator_eager_limit) { opal_show_help("help-mpi-pml-ob1.txt", "cuda_rdma_limit_too_small", true, sm->btl_component->btl_version.mca_component_name, ompi_process_info.nodename, sm->btl_component->btl_version.mca_component_name, sm->btl_module->btl_accelerator_rdma_limit, sm->btl_component->btl_version.mca_component_name, sm->btl_module->btl_accelerator_eager_limit, sm->btl_component->btl_version.mca_component_name); return OMPI_ERR_BAD_PARAM; } } #endif /* OPAL_CUDA_GDR_SUPPORT */ } /* TODO: Move these callback registration to another place */ rc = mca_bml.bml_register( MCA_PML_OB1_HDR_TYPE_MATCH, mca_pml_ob1_recv_frag_callback_match, NULL ); if (OMPI_SUCCESS != rc) { return rc; } rc = mca_bml.bml_register( MCA_PML_OB1_HDR_TYPE_RNDV, mca_pml_ob1_recv_frag_callback_rndv, NULL ); if (OMPI_SUCCESS != rc) { return rc; } rc = mca_bml.bml_register( MCA_PML_OB1_HDR_TYPE_RGET, mca_pml_ob1_recv_frag_callback_rget, NULL ); if (OMPI_SUCCESS != rc) { return rc; } rc = mca_bml.bml_register( MCA_PML_OB1_HDR_TYPE_ACK, mca_pml_ob1_recv_frag_callback_ack, NULL ); if (OMPI_SUCCESS != rc) { return rc; } rc = mca_bml.bml_register( MCA_PML_OB1_HDR_TYPE_FRAG, mca_pml_ob1_recv_frag_callback_frag, NULL ); if (OMPI_SUCCESS != rc) { return rc; } rc = mca_bml.bml_register( MCA_PML_OB1_HDR_TYPE_PUT, mca_pml_ob1_recv_frag_callback_put, NULL ); if (OMPI_SUCCESS != rc) { return rc; } rc = mca_bml.bml_register( MCA_PML_OB1_HDR_TYPE_FIN, mca_pml_ob1_recv_frag_callback_fin, NULL ); if (OMPI_SUCCESS != rc) { return rc; } rc = mca_bml.bml_register (MCA_PML_OB1_HDR_TYPE_CID, mca_pml_ob1_recv_frag_callback_cid, NULL); if (OMPI_SUCCESS != rc) { return rc; } /* register error handlers */ return mca_bml.bml_register_error(mca_pml_ob1_error_handler); } /* * iterate through each proc and notify any PTLs associated * with the proc that it is/has gone away */ int mca_pml_ob1_del_procs(ompi_proc_t** procs, size_t nprocs) { return mca_bml.bml_del_procs(nprocs, procs); } /* * diagnostics */ static void mca_pml_ob1_dump_hdr(mca_pml_ob1_hdr_t* hdr) { char *type, header[128]; switch(hdr->hdr_common.hdr_type) { case MCA_PML_OB1_HDR_TYPE_MATCH: type = "MATCH"; snprintf( header, 128, "ctx %5d src %d tag %d seq %d", hdr->hdr_match.hdr_ctx, hdr->hdr_match.hdr_src, hdr->hdr_match.hdr_tag, hdr->hdr_match.hdr_seq); break; case MCA_PML_OB1_HDR_TYPE_RNDV: type = "RNDV"; snprintf( header, 128, "ctx %5d src %d tag %d seq %d msg_length %" PRIu64, hdr->hdr_rndv.hdr_match.hdr_ctx, hdr->hdr_rndv.hdr_match.hdr_src, hdr->hdr_rndv.hdr_match.hdr_tag, hdr->hdr_rndv.hdr_match.hdr_seq, hdr->hdr_rndv.hdr_msg_length); break; case MCA_PML_OB1_HDR_TYPE_RGET: type = "RGET"; snprintf( header, 128, "ctx %5d src %d tag %d seq %d msg_length %" PRIu64 "frag %" PRIu64 " src_ptr %" PRIu64, hdr->hdr_rndv.hdr_match.hdr_ctx, hdr->hdr_rndv.hdr_match.hdr_src, hdr->hdr_rndv.hdr_match.hdr_tag, hdr->hdr_rndv.hdr_match.hdr_seq, hdr->hdr_rndv.hdr_msg_length, hdr->hdr_rget.hdr_frag.lval, hdr->hdr_rget.hdr_src_ptr); break; case MCA_PML_OB1_HDR_TYPE_ACK: type = "ACK"; snprintf( header, 128, "src_req %p dst_req %p offset %" PRIu64 " size %" PRIu64, hdr->hdr_ack.hdr_src_req.pval, hdr->hdr_ack.hdr_dst_req.pval, hdr->hdr_ack.hdr_send_offset, hdr->hdr_ack.hdr_send_size); break; case MCA_PML_OB1_HDR_TYPE_FRAG: type = "FRAG"; snprintf( header, 128, "offset %" PRIu64 " src_req %p dst_req %p", hdr->hdr_frag.hdr_frag_offset, hdr->hdr_frag.hdr_src_req.pval, hdr->hdr_frag.hdr_dst_req.pval); break; case MCA_PML_OB1_HDR_TYPE_PUT: type = "PUT"; snprintf( header, 128, "dst_req %p src_frag %p recv_req %p offset %" PRIu64 " dst_ptr %" PRIu64 " dst_size %" PRIu64, hdr->hdr_rdma.hdr_req.pval, hdr->hdr_rdma.hdr_frag.pval, hdr->hdr_rdma.hdr_recv_req.pval, hdr->hdr_rdma.hdr_rdma_offset, hdr->hdr_rdma.hdr_dst_ptr, hdr->hdr_rdma.hdr_dst_size); break; case MCA_PML_OB1_HDR_TYPE_FIN: type = "FIN"; header[0] = '\0'; break; default: type = "UNKWN"; header[0] = '\0'; break; } opal_output(0,"hdr %s [%s] %s", type, (hdr->hdr_common.hdr_flags & MCA_PML_OB1_HDR_FLAGS_NBO ? "nbo" : " "), header); } #if !MCA_PML_OB1_CUSTOM_MATCH static void mca_pml_ob1_dump_frag_list(opal_list_t* queue, bool is_req) { opal_list_item_t* item; char cpeer[64], ctag[64]; for( item = opal_list_get_first(queue); item != opal_list_get_end(queue); item = opal_list_get_next(item) ) { if( is_req ) { mca_pml_base_request_t *req = &(((mca_pml_ob1_recv_request_t*)item)->req_recv.req_base); if( OMPI_ANY_SOURCE == req->req_peer ) snprintf(cpeer, 64, "%s", "ANY_SOURCE"); else snprintf(cpeer, 64, "%d", req->req_peer); if( OMPI_ANY_TAG == req->req_tag ) snprintf(ctag, 64, "%s", "ANY_TAG"); else snprintf(ctag, 64, "%d", req->req_tag); opal_output(0, "req %p peer %s tag %s addr %p count %lu datatype %s [%p] [%s %s] req_seq %" PRIu64, (void*) req, cpeer, ctag, (void*) req->req_addr, req->req_count, (0 != req->req_count ? req->req_datatype->name : "N/A"), (void*) req->req_datatype, (req->req_pml_complete ? "pml_complete" : ""), (req->req_free_called ? "freed" : ""), req->req_sequence); } else { mca_pml_ob1_recv_frag_t* frag = (mca_pml_ob1_recv_frag_t*)item; mca_pml_ob1_dump_hdr( &frag->hdr ); } } } #endif void mca_pml_ob1_dump_cant_match(mca_pml_ob1_recv_frag_t* queue) { mca_pml_ob1_recv_frag_t* item = queue; do { mca_pml_ob1_dump_hdr( &item->hdr ); if( NULL != item->range ) { mca_pml_ob1_recv_frag_t* frag = item->range; do { mca_pml_ob1_dump_hdr( &frag->hdr ); frag = (mca_pml_ob1_recv_frag_t*)frag->super.super.opal_list_next; } while( frag != item->range ); } item = (mca_pml_ob1_recv_frag_t*)item->super.super.opal_list_next; } while( item != queue ); } int mca_pml_ob1_dump(struct ompi_communicator_t* comm, int verbose) { struct mca_pml_comm_t* pml_comm = comm->c_pml_comm; int i; /* TODO: don't forget to dump mca_pml_ob1.non_existing_communicator_pending */ opal_output(0, "Communicator %s [%p](%s) rank %d recv_seq %d num_procs %lu last_probed %lu\n", comm->c_name, (void*) comm, ompi_comm_print_cid (comm), comm->c_my_rank, pml_comm->recv_sequence, pml_comm->num_procs, pml_comm->last_probed); #if !MCA_PML_OB1_CUSTOM_MATCH if( opal_list_get_size(&pml_comm->wild_receives) ) { opal_output(0, "expected MPI_ANY_SOURCE fragments\n"); mca_pml_ob1_dump_frag_list(&pml_comm->wild_receives, true); } #endif #if MCA_PML_OB1_CUSTOM_MATCH opal_output(0, "expected receives\n"); custom_match_prq_dump(pml_comm->prq); opal_output(0, "unexpected frag\n"); custom_match_umq_dump(pml_comm->umq); #endif /* iterate through all procs on communicator */ for( i = 0; i < (int)pml_comm->num_procs; i++ ) { mca_pml_ob1_comm_proc_t* proc = pml_comm->procs[i]; if (NULL == proc) { continue; } mca_bml_base_endpoint_t* ep = mca_bml_base_get_endpoint(proc->ompi_proc); size_t n; opal_output(0, "[Rank %d] expected_seq %d ompi_proc %p send_seq %d\n", i, proc->expected_sequence, (void*) proc->ompi_proc, proc->send_sequence); /* dump all receive queues */ #if !MCA_PML_OB1_CUSTOM_MATCH if( opal_list_get_size(&proc->specific_receives) ) { opal_output(0, "expected specific receives\n"); mca_pml_ob1_dump_frag_list(&proc->specific_receives, true); } #endif if( NULL != proc->frags_cant_match ) { opal_output(0, "out of sequence\n"); mca_pml_ob1_dump_cant_match(proc->frags_cant_match); } #if !MCA_PML_OB1_CUSTOM_MATCH if( opal_list_get_size(&proc->unexpected_frags) ) { opal_output(0, "unexpected frag\n"); mca_pml_ob1_dump_frag_list(&proc->unexpected_frags, false); } #endif /* dump all btls used for eager messages */ for( n = 0; n < ep->btl_eager.arr_size; n++ ) { mca_bml_base_btl_t* bml_btl = &ep->btl_eager.bml_btls[n]; bml_btl->btl->btl_dump(bml_btl->btl, bml_btl->btl_endpoint, verbose); } } return OMPI_SUCCESS; } static void mca_pml_ob1_control_completion (mca_btl_base_module_t* btl, struct mca_btl_base_endpoint_t *endpoint, mca_btl_base_descriptor_t *des, int status) { mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*) des->des_context; /* check for pending requests */ MCA_PML_OB1_PROGRESS_PENDING(bml_btl); } int mca_pml_ob1_send_control_btl (mca_bml_base_btl_t *bml_btl, int order, mca_pml_ob1_hdr_t *hdr, size_t hdr_size, bool add_to_pending) { int des_flags = MCA_BTL_DES_FLAGS_PRIORITY | MCA_BTL_DES_FLAGS_BTL_OWNERSHIP | MCA_BTL_DES_FLAGS_SIGNAL; mca_btl_base_descriptor_t *des; int rc; if (NULL != bml_btl->btl->btl_sendi) { rc = mca_bml_base_sendi (bml_btl, NULL, hdr, hdr_size, 0, order, des_flags, hdr->hdr_common.hdr_type, &des); if (OPAL_LIKELY(OPAL_SUCCESS == rc)) { return rc; } } else { (void) mca_bml_base_alloc (bml_btl, &des, order, hdr_size, des_flags); } if (OPAL_UNLIKELY(NULL == des)) { if (add_to_pending) { mca_pml_ob1_add_to_pending (NULL, bml_btl, order, hdr, hdr_size); } return OMPI_ERR_OUT_OF_RESOURCE; } des->des_cbfunc = mca_pml_ob1_control_completion; memcpy (des->des_segments->seg_addr.pval, hdr, hdr_size); /* queue request */ rc = mca_bml_base_send (bml_btl, des, hdr->hdr_common.hdr_type); if( OPAL_LIKELY( rc >= 0 ) ) { if( OPAL_LIKELY( 1 == rc ) ) { MCA_PML_OB1_PROGRESS_PENDING(bml_btl); } SPC_RECORD(OMPI_SPC_BYTES_SENT_MPI, (ompi_spc_value_t)sizeof(mca_pml_ob1_fin_hdr_t)); return OMPI_SUCCESS; } mca_bml_base_free(bml_btl, des); if (add_to_pending) { mca_pml_ob1_add_to_pending (NULL, bml_btl, order, hdr, hdr_size); } return OMPI_ERR_OUT_OF_RESOURCE; } int mca_pml_ob1_send_control_any (ompi_proc_t *proc, int order, mca_pml_ob1_hdr_t *hdr, size_t hdr_size, bool add_to_pending) { mca_bml_base_endpoint_t* endpoint = mca_bml_base_get_endpoint (proc); int rc; assert (NULL != endpoint); for (size_t i = 0 ; i < mca_bml_base_btl_array_get_size(&endpoint->btl_eager) ; ++i) { mca_bml_base_btl_t *bml_btl = mca_bml_base_btl_array_get_next (&endpoint->btl_eager); rc = mca_pml_ob1_send_control_btl (bml_btl, order, hdr, hdr_size, false); if (OMPI_SUCCESS == rc) { return OMPI_SUCCESS; } } if (add_to_pending) { mca_pml_ob1_add_to_pending (proc, NULL, order, hdr, hdr_size); } return OMPI_ERR_OUT_OF_RESOURCE; } /** * Send an FIN to the peer. If we fail to send this ack (no more available * fragments or the send failed) this function automatically add the FIN * to the list of pending FIN, Which guarantee that the FIN will be sent * later. */ int mca_pml_ob1_send_fin (ompi_proc_t* proc, mca_bml_base_btl_t* bml_btl, opal_ptr_t hdr_frag, uint64_t rdma_size, uint8_t order, int status) { mca_pml_ob1_fin_hdr_t fin; /* fill in header */ mca_pml_ob1_fin_hdr_prepare (&fin, 0, hdr_frag.lval, status ? status : (int64_t) rdma_size); ob1_hdr_hton((mca_pml_ob1_hdr_t *) &fin, MCA_PML_OB1_HDR_TYPE_FIN, proc); return mca_pml_ob1_send_control_btl (bml_btl, order, (mca_pml_ob1_hdr_t *) &fin, sizeof (fin), true); } int mca_pml_ob1_send_cid (ompi_proc_t *proc, ompi_communicator_t *comm) { mca_pml_ob1_cid_hdr_t cid; mca_pml_ob1_cid_hdr_prepare (&cid, comm); ob1_hdr_hton ((mca_pml_ob1_hdr_t *) &cid, cid.hdr_common.hdr_type, proc); return mca_pml_ob1_send_control_any (proc, MCA_BTL_NO_ORDER, (mca_pml_ob1_hdr_t *) &cid, sizeof (cid), true); } void mca_pml_ob1_process_pending_packets(mca_bml_base_btl_t* bml_btl) { mca_pml_ob1_pckt_pending_t *pckt; int32_t rc, max = (int32_t) opal_list_get_size (&mca_pml_ob1.pckt_pending); for (int32_t i = 0; i < max ; ++i) { OPAL_THREAD_SCOPED_LOCK(&mca_pml_ob1.lock, { pckt = (mca_pml_ob1_pckt_pending_t*) opal_list_remove_first(&mca_pml_ob1.pckt_pending); }); if (NULL == pckt) { break; } if (pckt->bml_btl) { rc = mca_pml_ob1_send_control_btl (pckt->bml_btl, pckt->order, &pckt->hdr, pckt->hdr_size, false); } else { rc = mca_pml_ob1_send_control_any (pckt->proc, pckt->order, &pckt->hdr, pckt->hdr_size, false); } if (OPAL_SUCCESS != rc) { /* could not send the packet. readd it to the pending list */ OPAL_THREAD_SCOPED_LOCK(&mca_pml_ob1.lock, { opal_list_append(&mca_pml_ob1.pckt_pending, (opal_list_item_t*)pckt); }); } else { /* We're done with this packet, return it back to the free list */ MCA_PML_OB1_PCKT_PENDING_RETURN(pckt); } } } void mca_pml_ob1_process_pending_rdma(void) { mca_pml_ob1_rdma_frag_t* frag; int32_t i, rc, s = (int32_t)opal_list_get_size(&mca_pml_ob1.rdma_pending); for(i = 0; i < s; i++) { OPAL_THREAD_LOCK(&mca_pml_ob1.lock); frag = (mca_pml_ob1_rdma_frag_t*) opal_list_remove_first(&mca_pml_ob1.rdma_pending); OPAL_THREAD_UNLOCK(&mca_pml_ob1.lock); if(NULL == frag) break; frag->retries++; if(frag->rdma_state == MCA_PML_OB1_RDMA_PUT) { rc = mca_pml_ob1_send_request_put_frag(frag); } else { rc = mca_pml_ob1_recv_request_get_frag(frag); } if(OMPI_ERR_OUT_OF_RESOURCE == rc) break; } } void mca_pml_ob1_error_handler( struct mca_btl_base_module_t* btl, int32_t flags, opal_proc_t* errproc, char* btlinfo ) { if (flags & MCA_BTL_ERROR_FLAGS_ADD_ACCELERATOR_IPC) { mca_pml_ob1_accelerator_add_ipc_support(btl, flags, (struct ompi_proc_t*)errproc, btlinfo); return; } /* Some BTL report unreachable errors during normal MPI_Finalize * termination. Lets simply ignore such errors after MPI is not supposed to * be operational anyway. */ if(ompi_mpi_state >= OMPI_MPI_STATE_FINALIZE_PAST_COMM_SELF_DESTRUCT) { return; } #if OPAL_ENABLE_FT_MPI opal_output_verbose( 1, mca_pml_ob1_output, "PML:OB1: the error handler was invoked by the %s BTL for proc %s with info %s", btl->btl_component->btl_version.mca_component_name, (NULL == errproc ? "null" : OMPI_NAME_PRINT(&errproc->proc_name)), btlinfo); if( ompi_ftmpi_enabled && (NULL != errproc) ) { /* It's safe to upgrade to the OMPI type */ ompi_errhandler_proc_failed((ompi_proc_t*)errproc); return; } #endif /* OPAL_ENABLE_FT_MPI */ /* TODO: this error should return to the caller and invoke an error * handler from the MPI API call. * For now, it is fatal. */ ompi_mpi_errors_are_fatal_comm_handler(NULL, NULL, btlinfo); } int mca_pml_ob1_com_btl_comp(const void *v1, const void *v2) { const mca_pml_ob1_com_btl_t *b1 = (const mca_pml_ob1_com_btl_t *) v1; const mca_pml_ob1_com_btl_t *b2 = (const mca_pml_ob1_com_btl_t *) v2; if(b1->bml_btl->btl_weight < b2->bml_btl->btl_weight) return 1; if(b1->bml_btl->btl_weight > b2->bml_btl->btl_weight) return -1; return 0; }