static void component_shutdown(void)
{
    opal_output_verbose(2, orte_oob_base_framework.framework_output,
                        "%s ALPS SHUTDOWN",
                        ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
}

static int setup_fork(orte_job_t *jdata,
                      orte_app_context_t *app)
{
    int i;
    char *param;
    bool oversubscribed;
    orte_node_t *node;
    char **envcpy, **nps, **firstranks;
    char *npstring, *firstrankstring;
    char *num_app_ctx;
    bool takeus = false;

    opal_output_verbose(1, orte_schizo_base_framework.framework_output,
                        "%s schizo:ompi: setup_fork",
                        ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));

    if (NULL != orte_schizo_base.personalities) {
    /* see if we are included */
        for (i=0; NULL != jdata->personality[i]; i++) {
            if (0 == strcmp(jdata->personality[i], "ompi")) {
                takeus = true;
                break;
            }
        }
        if (!takeus) {
            return ORTE_ERR_TAKE_NEXT_OPTION;
        }
    }

    /* see if the mapper thinks we are oversubscribed */
    oversubscribed = false;
    if (NULL == (node = (orte_node_t*)opal_pointer_array_get_item(orte_node_pool, ORTE_PROC_MY_NAME->vpid))) {
        ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
        return ORTE_ERR_NOT_FOUND;
    }
    if (ORTE_FLAG_TEST(node, ORTE_NODE_FLAG_OVERSUBSCRIBED)) {
        oversubscribed = true;
    }

    /* setup base environment: copy the current environ and merge
       in the app context environ */
    if (NULL != app->env) {
        /* manually free original context->env to avoid a memory leak */
        char **tmp = app->env;
        envcpy = opal_environ_merge(orte_launch_environ, app->env);
        if (NULL != tmp) {
            opal_argv_free(tmp);
        }
    } else {
        envcpy = opal_argv_copy(orte_launch_environ);
    }
    app->env = envcpy;

    /* special case handling for --prefix: this is somewhat icky,
       but at least some users do this.  :-\ It is possible that
       when using --prefix, the user will also "-x PATH" and/or
       "-x LD_LIBRARY_PATH", which would therefore clobber the
       work that was done in the prior pls to ensure that we have
       the prefix at the beginning of the PATH and
       LD_LIBRARY_PATH.  So examine the context->env and see if we
       find PATH or LD_LIBRARY_PATH.  If found, that means the
       prior work was clobbered, and we need to re-prefix those
       variables. */
    param = NULL;
    orte_get_attribute(&app->attributes, ORTE_APP_PREFIX_DIR, (void**)&param, OPAL_STRING);
    for (i = 0; NULL != param && NULL != app->env && NULL != app->env[i]; ++i) {
        char *newenv;

        /* Reset PATH */
        if (0 == strncmp("PATH=", app->env[i], 5)) {
            asprintf(&newenv, "%s/bin:%s", param, app->env[i] + 5);
            opal_setenv("PATH", newenv, true, &app->env);
            free(newenv);
        }

        /* Reset LD_LIBRARY_PATH */
        else if (0 == strncmp("LD_LIBRARY_PATH=", app->env[i], 16)) {
            asprintf(&newenv, "%s/lib:%s", param, app->env[i] + 16);
            opal_setenv("LD_LIBRARY_PATH", newenv, true, &app->env);
            free(newenv);
        }
    }
    if (NULL != param) {
        free(param);
    }

    /* pass my contact info to the local proc so we can talk */
    opal_setenv("OMPI_MCA_orte_local_daemon_uri", orte_process_info.my_daemon_uri, true, &app->env);

    /* pass the hnp's contact info to the local proc in case it
     * needs it
     */
    if (NULL != orte_process_info.my_hnp_uri) {
        opal_setenv("OMPI_MCA_orte_hnp_uri", orte_process_info.my_hnp_uri, true, &app->env);
    }

    /* setup yield schedule - do not override any user-supplied directive! */
    if (oversubscribed) {
        opal_setenv("OMPI_MCA_mpi_yield_when_idle", "1", false, &app->env);
    } else {
//.........这里部分代码省略.........

/* called when a receive should be progressed */
static int
ompi_mtl_portals4_recv_progress(ptl_event_t *ev,
                                ompi_mtl_portals4_base_request_t* ptl_base_request)
{
    int ret;
    ompi_mtl_portals4_recv_request_t* ptl_request =
        (ompi_mtl_portals4_recv_request_t*) ptl_base_request;
    size_t msg_length = 0;
    ptl_match_bits_t read_match_bits;

    /* as soon as we've seen any event associated with a request, it's
       started */
    ptl_request->req_started = true;

    switch (ev->type) {
    case PTL_EVENT_PUT:
        OPAL_OUTPUT_VERBOSE((50, ompi_mtl_base_framework.framework_output,
                             "Recv %lu (0x%lx) got put event",
                             ptl_request->opcount, ev->hdr_data));

        if (ev->ni_fail_type != PTL_NI_OK) {
            opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
                                "%s:%d: PTL_EVENT_PUT with ni_fail_type: %d",
                                __FILE__, __LINE__, ev->ni_fail_type);
            goto callback_error;
        }

        ptl_request->me_h = PTL_INVALID_HANDLE;

        msg_length = MTL_PORTALS4_GET_LENGTH(ev->hdr_data);
        ptl_request->super.super.ompi_req->req_status.MPI_SOURCE =
                MTL_PORTALS4_GET_SOURCE(ev->match_bits);
        ptl_request->super.super.ompi_req->req_status.MPI_TAG =
                MTL_PORTALS4_GET_TAG(ev->match_bits);
        if (OPAL_UNLIKELY(msg_length > ptl_request->delivery_len)) {
            opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
                    "truncate expected: %ld %ld",
                    msg_length, ptl_request->delivery_len);
            ptl_request->super.super.ompi_req->req_status.MPI_ERROR = MPI_ERR_TRUNCATE;
        }

#if OPAL_ENABLE_DEBUG
        ptl_request->hdr_data = ev->hdr_data;
#endif

        if (!MTL_PORTALS4_IS_SHORT_MSG(ev->match_bits) && ompi_mtl_portals4.protocol == rndv) {
            /* If it's not a short message and we're doing rndv, we
               only have the first part of the message.  Issue the get
               to pull the second part of the message. */
            if (ptl_request->is_triggered) {
                ptl_request->super.super.ompi_req->req_status._ucount = 0;
            }
            else {

                ptl_request->super.super.ompi_req->req_status._ucount = ompi_mtl_portals4.eager_limit;

                MTL_PORTALS4_SET_READ_BITS(read_match_bits,
                        MTL_PORTALS4_GET_CONTEXT(ev->match_bits),
                        MTL_PORTALS4_GET_TAG(ev->match_bits));

                ret = read_msg((char*) ptl_request->delivery_ptr + ompi_mtl_portals4.eager_limit,
                        ((msg_length > ptl_request->delivery_len) ?
                                ptl_request->delivery_len : msg_length) - ompi_mtl_portals4.eager_limit,
                                ev->initiator,
                                read_match_bits,
                                ompi_mtl_portals4.eager_limit,
                                ptl_request);
                if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
                    if (NULL != ptl_request->buffer_ptr) free(ptl_request->buffer_ptr);
                    goto callback_error;
                }
            }
        } else {
            /* If we're either using the eager protocol or were a
               short message, all data has been received, so complete
               the message. */
            ret = ompi_mtl_datatype_unpack(ptl_request->convertor,
                    ev->start,
                    ev->mlength);
            if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
                opal_output_verbose(1, ompi_mtl_base_framework.framework_output,
                        "%s:%d: ompi_mtl_datatype_unpack failed: %d",
                        __FILE__, __LINE__, ret);
                ptl_request->super.super.ompi_req->req_status.MPI_ERROR = ret;
            }
            ptl_request->super.super.ompi_req->req_status._ucount = ev->mlength;

            OPAL_OUTPUT_VERBOSE((50, ompi_mtl_base_framework.framework_output,
                    "Recv %lu (0x%lx) completed, expected",
                    ptl_request->opcount, ptl_request->hdr_data));
            ptl_request->super.super.completion_callback(&ptl_request->super.super);
        }
        break;

    case PTL_EVENT_REPLY:
        OPAL_OUTPUT_VERBOSE((50, ompi_mtl_base_framework.framework_output,
                             "Recv %lu (0x%lx) got reply event",
                             ptl_request->opcount, ptl_request->hdr_data));

//.........这里部分代码省略.........

static int initialize(int argc, char *argv[])
{
    int ret, exit_status = OPAL_SUCCESS;
    char * tmp_env_var = NULL;

    /*
     * Make sure to init util before parse_args
     * to ensure installdirs is setup properly
     * before calling mca_base_open();
     */
    if( OPAL_SUCCESS != (ret = opal_init_util(&argc, &argv)) ) {
        return ret;
    }

    /*
     * Parse Command line arguments
     */
    if (OPAL_SUCCESS != (ret = parse_args(argc, argv))) {
        exit_status = ret;
        goto cleanup;
    }

    /*
     * Setup OPAL Output handle from the verbose argument
     */
    if( opal_restart_globals.verbose ) {
        opal_restart_globals.output = opal_output_open(NULL);
        opal_output_set_verbosity(opal_restart_globals.output, 10);
    } else {
        opal_restart_globals.output = 0; /* Default=STDOUT */
    }

    /*
     * Turn off the selection of the CRS component,
     * we need to do that later
     */
    (void) mca_base_var_env_name("crs_base_do_not_select", &tmp_env_var);
    opal_setenv(tmp_env_var,
                "1", /* turn off the selection */
                true, &environ);
    free(tmp_env_var);
    tmp_env_var = NULL;

    /*
     * Make sure we select the proper compress component.
     */
    if( NULL != opal_restart_globals.snapshot_compress ) {
        (void) mca_base_var_env_name("compress", &tmp_env_var);
        opal_setenv(tmp_env_var,
                    opal_restart_globals.snapshot_compress,
                    true, &environ);
        free(tmp_env_var);
        tmp_env_var = NULL;
    }

    /*
     * Initialize the OPAL layer
     */
    if (OPAL_SUCCESS != (ret = opal_init(&argc, &argv))) {
        exit_status = ret;
        goto cleanup;
    }

    /*
     * If the checkpoint was compressed, then decompress it before continuing
     */
    if( NULL != opal_restart_globals.snapshot_compress ) {
        char * zip_dir = NULL;
        char * tmp_str = NULL;

        /* Make sure to clear the selection for the restart,
         * this way the user can swich compression mechanism
         * across restart
         */
        (void) mca_base_var_env_name("compress", &tmp_env_var);
        opal_unsetenv(tmp_env_var, &environ);
        free(tmp_env_var);
        tmp_env_var = NULL;

        opal_asprintf(&zip_dir, "%s/%s%s",
                 opal_restart_globals.snapshot_loc,
                 opal_restart_globals.snapshot_ref,
                 opal_restart_globals.snapshot_compress_postfix);

        if (0 >  (ret = access(zip_dir, F_OK)) ) {
            opal_output(opal_restart_globals.output,
                        "Error: Unable to access the file [%s]!",
                        zip_dir);
            exit_status = OPAL_ERROR;
            goto cleanup;
        }

        opal_output_verbose(10, opal_restart_globals.output,
                            "Decompressing (%s)",
                            zip_dir);

        opal_compress.decompress(zip_dir, &tmp_str);

        if( NULL != zip_dir ) {
            free(zip_dir);
//.........这里部分代码省略.........

static int parse_cli(int argc, int start, char **argv)
{
    int i, j, k;
    bool ignore;
    char *no_dups[] = {
        "grpcomm",
        "odls",
        "rml",
        "routed",
        NULL
    };
    bool takeus = false;

    opal_output_verbose(1, orte_schizo_base_framework.framework_output,
                        "%s schizo:ompi: parse_cli",
                        ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));

    /* if they gave us a list of personalities,
     * see if we are included */
    if (NULL != orte_schizo_base.personalities) {
        for (i=0; NULL != orte_schizo_base.personalities[i]; i++) {
            if (0 == strcmp(orte_schizo_base.personalities[i], "ompi")) {
                takeus = true;
                break;
            }
        }
        if (!takeus) {
            return ORTE_ERR_TAKE_NEXT_OPTION;
        }
    } else {
        /* attempt to auto-detect CLI options that
         * we recognize */
    }

    for (i = 0; i < (argc-start); ++i) {
        if (0 == strcmp("-mca",  argv[i]) ||
            0 == strcmp("--mca", argv[i]) ) {
            /* ignore this one */
            if (0 == strcmp(argv[i+1], "mca_base_env_list")) {
                i += 2;
                continue;
            }
            /* It would be nice to avoid increasing the length
             * of the orted cmd line by removing any non-ORTE
             * params. However, this raises a problem since
             * there could be OPAL directives that we really
             * -do- want the orted to see - it's only the OMPI
             * related directives we could ignore. This becomes
             * a very complicated procedure, however, since
             * the OMPI mca params are not cleanly separated - so
             * filtering them out is nearly impossible.
             *
             * see if this is already present so we at least can
             * avoid growing the cmd line with duplicates
             */
            ignore = false;
            if (NULL != orted_cmd_line) {
                for (j=0; NULL != orted_cmd_line[j]; j++) {
                    if (0 == strcmp(argv[i+1], orted_cmd_line[j])) {
                        /* already here - if the value is the same,
                         * we can quitely ignore the fact that they
                         * provide it more than once. However, some
                         * frameworks are known to have problems if the
                         * value is different. We don't have a good way
                         * to know this, but we at least make a crude
                         * attempt here to protect ourselves.
                         */
                        if (0 == strcmp(argv[i+2], orted_cmd_line[j+1])) {
                            /* values are the same */
                            ignore = true;
                            break;
                        } else {
                            /* values are different - see if this is a problem */
                            for (k=0; NULL != no_dups[k]; k++) {
                                if (0 == strcmp(no_dups[k], argv[i+1])) {
                                    /* print help message
                                     * and abort as we cannot know which one is correct
                                     */
                                    orte_show_help("help-orterun.txt", "orterun:conflicting-params",
                                                   true, orte_basename, argv[i+1],
                                                   argv[i+2], orted_cmd_line[j+1]);
                                    return ORTE_ERR_BAD_PARAM;
                                }
                            }
                            /* this passed muster - just ignore it */
                            ignore = true;
                            break;
                        }
                    }
                }
            }
            if (!ignore) {
                opal_argv_append_nosize(&orted_cmd_line, argv[i]);
                opal_argv_append_nosize(&orted_cmd_line, argv[i+1]);
                opal_argv_append_nosize(&orted_cmd_line, argv[i+2]);
            }
            i += 2;
        }
    }
    return ORTE_SUCCESS;
//.........这里部分代码省略.........

static int read_bytes(mca_oob_usock_peer_t* peer)
{
    int rc;

    /* read until all bytes recvd or error */
    while (0 < peer->recv_msg->rdbytes) {
        rc = read(peer->sd, peer->recv_msg->rdptr, peer->recv_msg->rdbytes);
        if (rc < 0) {
            if(opal_socket_errno == EINTR) {
                continue;
            } else if (opal_socket_errno == EAGAIN) {
                /* tell the caller to keep this message on active,
                 * but let the event lib cycle so other messages
                 * can progress while this socket is busy
                 */
                return ORTE_ERR_RESOURCE_BUSY;
            } else if (opal_socket_errno == EWOULDBLOCK) {
                /* tell the caller to keep this message on active,
                 * but let the event lib cycle so other messages
                 * can progress while this socket is busy
                 */
                return ORTE_ERR_WOULD_BLOCK;
            }
            /* we hit an error and cannot progress this message - report
             * the error back to the RML and let the caller know
             * to abort this message
             */
            opal_output_verbose(OOB_USOCK_DEBUG_FAIL, orte_oob_base_framework.framework_output,
                                "%s-%s mca_oob_usock_msg_recv: readv failed: %s (%d)",
                                ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
                                ORTE_NAME_PRINT(&(peer->name)),
                                strerror(opal_socket_errno),
                                opal_socket_errno);
            // mca_oob_usock_peer_close(peer);
            // if (NULL != mca_oob_usock.oob_exception_callback) {
            // mca_oob_usock.oob_exception_callback(&peer->name, ORTE_RML_PEER_DISCONNECTED);
            //}
            return ORTE_ERR_COMM_FAILURE;
        } else if (rc == 0)  {
            /* the remote peer closed the connection - report that condition
             * and let the caller know
             */
            opal_output_verbose(OOB_USOCK_DEBUG_FAIL, orte_oob_base_framework.framework_output,
                                "%s-%s mca_oob_usock_msg_recv: peer closed connection",
                                ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
                                ORTE_NAME_PRINT(&(peer->name)));
            /* stop all events */
            if (peer->recv_ev_active) {
                opal_event_del(&peer->recv_event);
                peer->recv_ev_active = false;
            }
            if (peer->timer_ev_active) {
                opal_event_del(&peer->timer_event);
                peer->timer_ev_active = false;
            }
            if (peer->send_ev_active) {
                opal_event_del(&peer->send_event);
                peer->send_ev_active = false;
            }
            if (NULL != peer->recv_msg) {
                OBJ_RELEASE(peer->recv_msg);
                peer->recv_msg = NULL;
            }
            mca_oob_usock_peer_close(peer);
            //if (NULL != mca_oob_usock.oob_exception_callback) {
            //   mca_oob_usock.oob_exception_callback(&peer->peer_name, ORTE_RML_PEER_DISCONNECTED);
            //}
            return ORTE_ERR_WOULD_BLOCK;
        }
        /* we were able to read something, so adjust counters and location */
        peer->recv_msg->rdbytes -= rc;
        peer->recv_msg->rdptr += rc;
    }

    /* we read the full data block */
    return ORTE_SUCCESS;
}

/*
 * Function to find as many components of a given type as possible.  This
 * includes statically-linked in components as well as opening up a
 * directory and looking for shared-library MCA components of the
 * appropriate type (load them if available).
 *
 * Return one consolidated array of (mca_base_component_t*) pointing to all
 * available components.
 */
int mca_base_component_find(const char *directory, const char *type, 
                            const mca_base_component_t *static_components[], 
                            char **requested_component_names,
                            bool include_mode,
                            opal_list_t *found_components,
                            bool open_dso_components)
{
    int i;
    opal_list_item_t *item;
    mca_base_component_list_item_t *cli;

    /* Find all the components that were statically linked in */
    OBJ_CONSTRUCT(found_components, opal_list_t);
    for (i = 0; NULL != static_components[i]; ++i) {
        if ( use_component(include_mode,
                           (const char**)requested_component_names,
                           static_components[i]->mca_component_name) ) {
            cli = OBJ_NEW(mca_base_component_list_item_t);
            if (NULL == cli) {
                return OPAL_ERR_OUT_OF_RESOURCE;
            }
            cli->cli_component = static_components[i];
            opal_list_append(found_components, (opal_list_item_t *) cli);
        }
    }

#if OMPI_WANT_LIBLTDL
    /* Find any available dynamic components in the specified directory */
    if (open_dso_components) {
        int param, param_disable_dlopen;
        param = mca_base_param_find("mca", NULL, "component_disable_dlopen");
        mca_base_param_lookup_int(param, &param_disable_dlopen);

        if (0 == param_disable_dlopen) {
            find_dyn_components(directory, type,
                                (const char**)requested_component_names,
                                include_mode, found_components); 
        }
    } else {
        opal_output_verbose(40, 0, 
                            "mca: base: component_find: dso loading for %s MCA components disabled", 
                            type);
    }
#endif

    /* Ensure that *all* requested components exist.  Print a warning
       and abort if they do not. */
    for (i = 0; include_mode && NULL != requested_component_names && 
             NULL != requested_component_names[i]; ++i) {
        for (item = opal_list_get_first(found_components);
             opal_list_get_end(found_components) != item; 
             item = opal_list_get_next(item)) {
            cli = (mca_base_component_list_item_t*) item;
            if (0 == strcmp(requested_component_names[i], 
                            cli->cli_component->mca_component_name)) {
                break;
            }
        }

        if (opal_list_get_end(found_components) == item) {
            char h[MAXHOSTNAMELEN];
            gethostname(h, sizeof(h));
            opal_show_help("help-mca-base.txt", 
                           "find-available:not-valid", true,
                           h, type, requested_component_names[i]);
            return OPAL_ERR_NOT_FOUND;
        }
    }

    /* All done */

    return OPAL_SUCCESS;
}

/* Compare the addresses of the local interface corresponding to module and the
 * remote interface corresponding to proc_modex_addr.  Returns a weight value
 * (higher values indicate more desirable connections). */
static uint64_t compute_weight(
    ompi_btl_usnic_module_t *module,
    ompi_btl_usnic_addr_t *proc_modex_addr)
{
    char my_ip_string[INET_ADDRSTRLEN], peer_ip_string[INET_ADDRSTRLEN];
    uint32_t mynet, peernet;
    int err, metric;
    uint32_t min_link_speed_gbps;

    inet_ntop(AF_INET, &module->if_ipv4_addr,
              my_ip_string, sizeof(my_ip_string));
    inet_ntop(AF_INET, &proc_modex_addr->ipv4_addr,
              peer_ip_string, sizeof(peer_ip_string));

    /* Just compare the CIDR-masked IP address to see if they're on
       the same network.  If so, we're good. */
    mynet   = ompi_btl_usnic_get_ipv4_subnet(module->if_ipv4_addr,
                                             module->if_cidrmask);
    peernet = ompi_btl_usnic_get_ipv4_subnet(proc_modex_addr->ipv4_addr,
                                             proc_modex_addr->cidrmask);
    opal_output_verbose(5, USNIC_OUT,
                        "btl:usnic:%s: checking my IP address/subnet (%s/%d) vs. peer (%s/%d): %s",
                        __func__, my_ip_string, module->if_cidrmask,
                        peer_ip_string, proc_modex_addr->cidrmask,
                        (mynet == peernet ? "match" : "DO NOT match"));

    if (!mca_btl_usnic_component.use_udp) {
        if (mynet != peernet) {
            return WEIGHT_UNREACHABLE;
        } else {
            return 1; /* any positive weight is fine */
        }
    }

    min_link_speed_gbps = MIN(module->super.btl_bandwidth,
                              proc_modex_addr->link_speed_mbps) / 1000;

    metric = 0;
    err = ompi_btl_usnic_nl_ip_rt_lookup(mca_btl_usnic_component.unlsk,
                                         module->if_ipv4_addr,
                                         proc_modex_addr->ipv4_addr,
                                         &metric);
    if (0 != err) {
        return 0; /* no connectivity */
    }
    else {
        /* Format in binary    MSB                             LSB
         * most sig. 32-bits:  00000000 0000000A BBBBBBBB 00000001
         * least sig. 32-bits: CCCCCCCC CCCCCCCC CCCCCCCC CCCCCCCC
         *
         * A = 1 iff same subnet
         * B = min link speed (in Gbps) between iface pair
         * C = metric from routing table
         *
         * That is, this prioritizes interfaces in the same subnet first,
         * followed by having the same link speed.  The extra literal "1" is in
         * there to help prioritize over any zero-cost links that might
         * otherwise make their way into the graph.  It is not strictly
         * necessary and could be eliminated if the extra byte is needed.
         *
         * TODO add an MCA parameter to optionally swap the offsets of A and
         * B, thereby prioritizing link speed over same subnet reachability.
         */
        /* FIXME how can we check that the metric is the same before we have
         * communication with this host?  Mismatched metrics could cause the
         * remote peer to make a different pairing decision... */
        if (min_link_speed_gbps > 0xff) {
            opal_output_verbose(20, USNIC_OUT, "clamping min_link_speed_gbps=%u to 255",
                                min_link_speed_gbps);
            min_link_speed_gbps = 0xff;
        }
        return ((uint64_t)(mynet == peernet) << 48) |
               ((uint64_t)(min_link_speed_gbps & 0xff) << 40) |
               ((uint64_t)0x1 << 32) |
               (/*metric=*/0);
    }
}

/**
 * Constructs an interface graph from all local modules and the given proc's
 * remote interfaces.  The resulting vertices will always have the module
 * vertices appear before the proc vertices.
 */
static int create_proc_module_graph(
    ompi_btl_usnic_proc_t *proc,
    bool proc_is_left,
    ompi_btl_usnic_graph_t **g_out)
{
    int err;
    int i, j;
    int u, v;
    int num_modules;
    ompi_btl_usnic_graph_t *g = NULL;

    if (NULL == g_out) {
        return OMPI_ERR_BAD_PARAM;
    }
    *g_out = NULL;

    num_modules = (int)mca_btl_usnic_component.num_modules;

    /* Construct a bipartite graph with remote interfaces on the one side and
     * local interfaces (modules) on the other. */
    err = ompi_btl_usnic_gr_create(NULL, NULL, &g);
    if (OMPI_SUCCESS != err) {
        OMPI_ERROR_LOG(err);
        goto out;
    }

    /* create vertices for each interface (local and remote) */
    for (i = 0; i < num_modules; ++i) {
        int idx = -1;
        err = ompi_btl_usnic_gr_add_vertex(g,
                                           mca_btl_usnic_component.usnic_active_modules[i],
                                           &idx);
        if (OMPI_SUCCESS != err) {
            OMPI_ERROR_LOG(err);
            goto out_free_graph;
        }
        assert(idx == MODULE_VERTEX(i));
    }
    for (i = 0; i < (int)proc->proc_modex_count; ++i) {
        int idx = -1;
        err = ompi_btl_usnic_gr_add_vertex(g, &proc->proc_modex[i], &idx);
        if (OMPI_SUCCESS != err) {
            OMPI_ERROR_LOG(err);
            goto out_free_graph;
        }
        assert(idx == (int)PROC_VERTEX(i));
    }

    /* now add edges between interfaces that can communicate */
    for (i = 0; i < num_modules; ++i) {
        for (j = 0; j < (int)proc->proc_modex_count; ++j) {
            int64_t weight, cost;

            /* assumption: compute_weight returns the same weight on the
             * remote process with these arguments (effectively) transposed */
            weight = compute_weight(mca_btl_usnic_component.usnic_active_modules[i],
                                    &proc->proc_modex[j]);

            opal_output_verbose(20, USNIC_OUT,
                                "btl:usnic:%s: weight=0x%016" PRIx64 " for edge module[%d] (%p) <--> endpoint[%d] on proc %p",
                                __func__,
                                weight, i,
                                (void *)mca_btl_usnic_component.usnic_active_modules[i],
                                j, (void *)proc);

            if (WEIGHT_UNREACHABLE == weight) {
                continue;
            } else {
                /* the graph code optimizes for minimum *cost*, but we have
                 * been computing weights (negative costs) */
                cost = -weight;
            }
            assert(INT64_MAX != cost);
            assert(INT64_MIN != cost);

            if (proc_is_left) {
                u = PROC_VERTEX(j);
                v = MODULE_VERTEX(i);
            } else {
                u = MODULE_VERTEX(i);
                v = PROC_VERTEX(j);
            }
            opal_output_verbose(20, USNIC_OUT,
                                "btl:usnic:%s: adding edge (%d,%d) with cost=%" PRIi64 " for edge module[%d] <--> endpoint[%d]",
                                __func__, u, v, cost, i, j);
            err = ompi_btl_usnic_gr_add_edge(g, u, v, cost,
                                             /*capacity=*/1,
                                             /*e_data=*/NULL);
            if (OMPI_SUCCESS != err) {
                OMPI_ERROR_LOG(err);
                goto out_free_graph;
            }
        }
    }

//.........这里部分代码省略.........

//.........这里部分代码省略.........
     * have a field in the jdata structure for "mandatory" vs "optional"
     * allocations, so we'll have to add that someday. Likewise, you may
     * want to provide a param to adjust the timeout value
     */
    /* construct the cmd string */
    opal_argv_append_nosize(&cmd, "allocate");
    /* add the jobid */
    orte_util_convert_jobid_to_string(&jstring, jdata->jobid);
    opal_asprintf(&tmp, "jobid=%s", jstring);
    opal_argv_append_nosize(&cmd, tmp);
    free(tmp);
    free(jstring);
    /* if we want the allocation for all apps in one shot,
     * then tell slurm
     *
     * RHC: we don't currently have the ability to handle
     * rolling allocations in the rest of the code base
     */
#if 0
    if (!mca_ras_slurm_component.rolling_alloc) {
        opal_argv_append_nosize(&cmd, "return=all");
    }
#else
    opal_argv_append_nosize(&cmd, "return=all");
#endif

    /* pass the timeout */
    opal_asprintf(&tmp, "timeout=%d", mca_ras_slurm_component.timeout);
    opal_argv_append_nosize(&cmd, tmp);
    free(tmp);

    /* for each app, add its allocation request info */
    i64ptr = &i64;
    for (i=0; i < jdata->apps->size; i++) {
        if (NULL == (app = (orte_app_context_t*)opal_pointer_array_get_item(jdata->apps, i))) {
            continue;
        }
        /* add the app id, preceded by a colon separator */
        opal_asprintf(&tmp, ": app=%d", (int)app->idx);
        opal_argv_append_nosize(&cmd, tmp);
        free(tmp);
        /* add the number of process "slots" we need */
        opal_asprintf(&tmp, "np=%d", app->num_procs);
        opal_argv_append_nosize(&cmd, tmp);
        free(tmp);
        /* if we were given a minimum number of nodes, pass it along */
        if (orte_get_attribute(&app->attributes, ORTE_APP_MIN_NODES, (void**)&i64ptr, OPAL_INT64)) {
            opal_asprintf(&tmp, "N=%ld", (long int)i64);
            opal_argv_append_nosize(&cmd, tmp);
            free(tmp);
        }
        /* add the list of nodes, if one was given, ensuring
         * that each node only appears once
         */
        node_list =  get_node_list(app);
        if (NULL != node_list) {
            opal_asprintf(&tmp, "node_list=%s", node_list);
            opal_argv_append_nosize(&cmd, tmp);
            free(node_list);
            free(tmp);
        }
        /* add the mandatory/optional flag */
        if (orte_get_attribute(&app->attributes, ORTE_APP_MANDATORY, NULL, OPAL_BOOL)) {
            opal_argv_append_nosize(&cmd, "flag=mandatory");
        } else {
            opal_argv_append_nosize(&cmd, "flag=optional");
        }
    }

    /* assemble it into the final cmd to be sent */
    cmd_str = opal_argv_join(cmd, ' ');
    opal_argv_free(cmd);

    /* start a timer - if the response to our request doesn't appear
     * in the defined time, then we will error out as Slurm isn't
     * responding to us
     */
    opal_event_evtimer_set(orte_event_base, &jtrk->timeout_ev, timeout, jtrk);
    tv.tv_sec = mca_ras_slurm_component.timeout * 2;
    tv.tv_usec = 0;
    opal_event_evtimer_add(&jtrk->timeout_ev, &tv);

    opal_output_verbose(2, orte_ras_base_framework.framework_output,
                        "%s slurm:dynalloc cmd_str = %s",
                        ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
                        cmd_str);

    if (send(socket_fd, cmd_str, strlen(cmd_str)+1, 0) < 0) {
        ORTE_ERROR_LOG(ORTE_ERR_COMM_FAILURE);
    }
    free(cmd_str);

    /* we cannot wait here for a response as we
     * are already in an event. So return a value
     * that indicates we are waiting for an
     * allocation so the base functions know
     * that they shouldn't progress the job
     */
    return ORTE_ERR_ALLOCATION_PENDING;
}

static void recv_data(int fd, short args, void *cbdata)
{
    bool found;
    int i, rc;
    orte_node_t *nd, *nd2;
    opal_list_t nds, ndtmp;
    opal_list_item_t *item, *itm;
    char recv_msg[8192];
    int nbytes, idx, sjob;
    char **alloc, *nodelist, *tpn;
    local_jobtracker_t *ptr, *jtrk;
    local_apptracker_t *aptrk;
    orte_app_context_t *app;
    orte_jobid_t jobid;
    orte_job_t *jdata;
    char **dash_host = NULL;

    opal_output_verbose(2, orte_ras_base_framework.framework_output,
                        "%s ras:slurm: dynamic allocation - data recvd",
                        ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));

    /* read the data from the socket and put it in the
     * nodes field of op
     */
    memset(recv_msg, 0, sizeof(recv_msg));
    nbytes = read(fd, recv_msg, sizeof(recv_msg) - 1);

    opal_output_verbose(2, orte_ras_base_framework.framework_output,
                        "%s ras:slurm: dynamic allocation msg: %s",
                        ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), recv_msg);

    /* check if we got something */
    if (0 == nbytes || 0 == strlen(recv_msg) || strstr(recv_msg, "failure") != NULL) {
        /* show an error here - basically, a "nothing was available"
         * message
         */
        orte_show_help("help-ras-slurm.txt", "slurm-dyn-alloc-failed", true,
                       (0 == strlen(recv_msg)) ? "NO MSG" : recv_msg);
        ORTE_ACTIVATE_JOB_STATE(NULL, ORTE_JOB_STATE_ALLOC_FAILED);
        return;
    }

    /* break the message into its component parts, separated by colons */
    alloc = opal_argv_split(recv_msg, ':');

    /* the first section contains the ORTE jobid for this allocation */
    tpn = strchr(alloc[0], '=');
    orte_util_convert_string_to_jobid(&jobid, tpn+1);
    /* get the corresponding job object */
    jdata = orte_get_job_data_object(jobid);
    jtrk = NULL;
    /* find the associated tracking object */
    for (item = opal_list_get_first(&jobs);
         item != opal_list_get_end(&jobs);
         item = opal_list_get_next(item)) {
        ptr = (local_jobtracker_t*)item;
        if (ptr->jobid == jobid) {
            jtrk = ptr;
            break;
        }
    }
    if (NULL == jtrk) {
        orte_show_help("help-ras-slurm.txt", "slurm-dyn-alloc-failed", true, "NO JOB TRACKER");
        ORTE_ACTIVATE_JOB_STATE(NULL, ORTE_JOB_STATE_ALLOC_FAILED);
        opal_argv_free(alloc);
        return;
    }

    /* stop the timeout event */
    opal_event_del(&jtrk->timeout_ev);

    /* cycle across all the remaining parts - each is the allocation for
     * an app in this job
     */
    OBJ_CONSTRUCT(&nds, opal_list_t);
    OBJ_CONSTRUCT(&ndtmp, opal_list_t);
    idx = -1;
    sjob = -1;
    nodelist = NULL;
    tpn = NULL;
    for (i=1; NULL != alloc[i]; i++) {
        if (ORTE_SUCCESS != parse_alloc_msg(alloc[i], &idx, &sjob, &nodelist, &tpn)) {
            orte_show_help("help-ras-slurm.txt", "slurm-dyn-alloc-failed", true, jtrk->cmd);
            ORTE_ACTIVATE_JOB_STATE(jdata, ORTE_JOB_STATE_ALLOC_FAILED);
            opal_argv_free(alloc);
            if (NULL != nodelist) {
                free(nodelist);
            }
            if (NULL != tpn) {
                free(tpn);
            }
            return;
        }
        if (idx < 0) {
            orte_show_help("help-ras-slurm.txt", "slurm-dyn-alloc-failed", true, jtrk->cmd);
            ORTE_ACTIVATE_JOB_STATE(jdata, ORTE_JOB_STATE_ALLOC_FAILED);
            opal_argv_free(alloc);
            free(nodelist);
            free(tpn);
            return;
//.........这里部分代码省略.........

/**
 * Discover available (pre-allocated) nodes.  Allocate the
 * requested number of nodes/process slots to the job.
 *
 */
static int orte_ras_slurm_allocate(orte_job_t *jdata, opal_list_t *nodes)
{
    int ret, cpus_per_task;
    char *slurm_node_str, *regexp;
    char *tasks_per_node, *node_tasks;
    char *tmp;
    char *slurm_jobid;

    if (NULL == (slurm_jobid = getenv("SLURM_JOBID"))) {
        /* we are not in a slurm allocation - see if dyn alloc
         * is enabled
         */
        if (!mca_ras_slurm_component.dyn_alloc_enabled) {
            /* nope - nothing we can do */
            opal_output_verbose(2, orte_ras_base_framework.framework_output,
                                "%s ras:slurm: no prior allocation and dynamic alloc disabled",
                                ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
            return ORTE_ERR_TAKE_NEXT_OPTION;
        }
    } else {
        /* save this value in the global job ident string for
         * later use in any error reporting
         */
        orte_job_ident = strdup(slurm_jobid);
    }

    slurm_node_str = getenv("SLURM_NODELIST");
    if (NULL == slurm_node_str) {
        /* see if dynamic allocation is enabled */
        if (mca_ras_slurm_component.dyn_alloc_enabled) {
            /* attempt to get the allocation - the function
             * dyn_allocate will return as ORTE_ERR_ALLOCATION_PENDING
             * if it succeeds in sending the allocation request
             */
            ret = dyn_allocate(jdata);
            /* return to the above layer in ras/base/ras_base_allocate.c
             * to wait for event (libevent) happening
             */
            return ret;
        }
        orte_show_help("help-ras-slurm.txt", "slurm-env-var-not-found", 1,
                       "SLURM_NODELIST");
        return ORTE_ERR_NOT_FOUND;
    }
    regexp = strdup(slurm_node_str);
    if(NULL == regexp) {
        ORTE_ERROR_LOG(ORTE_ERR_OUT_OF_RESOURCE);
        return ORTE_ERR_OUT_OF_RESOURCE;
    }

    if (mca_ras_slurm_component.use_all) {
        /* this is an oddball case required for debug situations where
         * a tool is started that will then call mpirun. In this case,
         * Slurm will assign only 1 tasks/per node to the tool, but
         * we want mpirun to use the entire allocation. They don't give
         * us a specific variable for this purpose, so we have to fudge
         * a bit - but this is a special edge case, and we'll live with it */
        tasks_per_node = getenv("SLURM_JOB_CPUS_PER_NODE");
        if (NULL == tasks_per_node) {
            /* couldn't find any version - abort */
            orte_show_help("help-ras-slurm.txt", "slurm-env-var-not-found", 1,
                           "SLURM_JOB_CPUS_PER_NODE");
            free(regexp);
            return ORTE_ERR_NOT_FOUND;
        }
        node_tasks = strdup(tasks_per_node);
        if (NULL == node_tasks) {
            ORTE_ERROR_LOG(ORTE_ERR_OUT_OF_RESOURCE);
            free(regexp);
            return ORTE_ERR_OUT_OF_RESOURCE;
        }
        cpus_per_task = 1;
    } else {
        /* get the number of process slots we were assigned on each node */
        tasks_per_node = getenv("SLURM_TASKS_PER_NODE");
        if (NULL == tasks_per_node) {
            /* couldn't find any version - abort */
            orte_show_help("help-ras-slurm.txt", "slurm-env-var-not-found", 1,
                           "SLURM_TASKS_PER_NODE");
            free(regexp);
            return ORTE_ERR_NOT_FOUND;
        }
        node_tasks = strdup(tasks_per_node);
        if (NULL == node_tasks) {
            ORTE_ERROR_LOG(ORTE_ERR_OUT_OF_RESOURCE);
            free(regexp);
            return ORTE_ERR_OUT_OF_RESOURCE;
        }

        /* get the number of CPUs per task that the user provided to slurm */
        tmp = getenv("SLURM_CPUS_PER_TASK");
        if(NULL != tmp) {
            cpus_per_task = atoi(tmp);
            if(0 >= cpus_per_task) {
                opal_output(0, "ras:slurm:allocate: Got bad value from SLURM_CPUS_PER_TASK. "
//.........这里部分代码省略.........

int
ompi_osc_pt2pt_sendreq_recv_accum(ompi_osc_pt2pt_module_t *module,
                                  ompi_osc_pt2pt_send_header_t *header,
                                  void *payload)
{
    int ret = OMPI_SUCCESS;
    struct ompi_op_t *op = ompi_osc_pt2pt_op_create(header->hdr_target_op);
    ompi_proc_t *proc = ompi_comm_peer_lookup( module->p2p_comm, header->hdr_origin );
    struct ompi_datatype_t *datatype = 
        ompi_osc_pt2pt_datatype_create(proc, &payload);

    if (header->hdr_msg_length > 0) {
        /* lock the window for accumulates */
        OPAL_THREAD_LOCK(&module->p2p_acc_lock);

        /* copy the data from the temporary buffer into the user window */
        ret = ompi_osc_pt2pt_process_op(module, header, datatype, op, payload, 
                                        header->hdr_msg_length);

        /* unlock the window for accumulates */
        OPAL_THREAD_UNLOCK(&module->p2p_acc_lock);

        /* Release datatype & op */
        OBJ_RELEASE(datatype);
        OBJ_RELEASE(op);

        OPAL_THREAD_ADD32(&(module->p2p_num_pending_in), -1);

        opal_output_verbose(50, ompi_osc_base_output,
                            "%d received accum message from %d",
                            module->p2p_comm->c_my_rank,
                            header->hdr_origin);
        
    } else {
        ompi_osc_pt2pt_longreq_t *longreq;
        ptrdiff_t lb, extent, true_lb, true_extent;
        size_t buflen;

        /* figure out how big a buffer we need */
        ompi_ddt_get_extent(datatype, &lb, &extent);
        ompi_ddt_get_true_extent(datatype, &true_lb, &true_extent);
        buflen = true_extent + (header->hdr_target_count - 1) * extent;

        /* get a longreq and fill it in */
        ompi_osc_pt2pt_longreq_alloc(&longreq);

        longreq->req_comp_cb = ompi_osc_pt2pt_sendreq_recv_accum_long_cb;
        longreq->req_datatype = datatype;
        longreq->req_op = op;
        longreq->req_module = module;

        /* allocate a buffer to receive into ... */
        longreq->req_comp_cbdata = malloc(buflen + sizeof(ompi_osc_pt2pt_send_header_t));
        
        if (NULL == longreq->req_comp_cbdata) return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
        /* fill in tmp header */
        memcpy(longreq->req_comp_cbdata, header,
               sizeof(ompi_osc_pt2pt_send_header_t));
        ((ompi_osc_pt2pt_send_header_t*) longreq->req_comp_cbdata)->hdr_msg_length = buflen;

        ret = mca_pml.pml_irecv(((char*) longre