btl_gm_component.c

来自「MPI stands for the Message Passing Inter」· C语言 代码 · 共 730 行 · 第 1/2 页

 * Scan all ports on the boards. As it's difficult to find the total number of boards
 * so we use a predefined maximum.  
 */
static int mca_btl_gm_discover( void )
{
    uint32_t board_no;
    uint32_t  port_no;
    uint32_t node_id;
    struct gm_port* port;
#if GM_API_VERSION > 0x200
    uint32_t global_id;
#else
    char  global_id[GM_MAX_HOST_NAME_LEN];
#endif  /* GM_API_VERSION > 0x200 */
    int rc;
    
    for( board_no = 0; board_no < mca_btl_gm_component.gm_max_boards; board_no++ ) {
        mca_btl_gm_module_t *btl;

        /* open the first available gm port for this board  */
        for( port_no = 4; port_no < mca_btl_gm_component.gm_max_ports; port_no++ ) {
            if (3 == port_no) {
                continue;  /* port 0,1,3 reserved  */
            } else if (GM_SUCCESS == gm_open(&port, board_no, port_no,
                mca_btl_gm_component.gm_port_name, GM_API_VERSION) ) {
                break;
            }
        }
        if( port_no == mca_btl_gm_component.gm_max_ports ) {
            continue;
        }

        /*  Get node local Id */
        if( GM_SUCCESS != gm_get_node_id( port, &node_id) ) {
            opal_output (0, " failure to get node_id \n");
            continue;
        }
        /* Gather an unique id for the node */
#if GM_API_VERSION > 0x200
        if (GM_SUCCESS != gm_node_id_to_global_id( port, node_id, &global_id) ) {
            opal_output (0, "[%s:%d] Unable to get my GM global unique id", __FILE__, __LINE__);
            continue;
        }
#else
        if( GM_SUCCESS != gm_get_host_name( port, global_id ) ) {
            opal_output( 0, "[%s:%d] Unable to get the GM host name\n", __FILE__, __LINE__);
            continue;
        }
#endif  /* GM_API_VERSION > 0x200 */

        /* create the btl module */
        btl = (mca_btl_gm_module_t *)malloc( sizeof(mca_btl_gm_module_t) );
        if (NULL == btl) {
            opal_output( 0, "[%s:%d] out of resources", __FILE__, __LINE__);
            return OMPI_ERR_OUT_OF_RESOURCE;
        }
        /* copy the basic informations into the new BTL */
        memcpy (btl, &mca_btl_gm_module, sizeof(mca_btl_gm_module_t));

        /* setup local address */
        btl->port = port;
        btl->gm_addr.port_id  = port_no;
        btl->gm_addr.node_id = node_id;
#if GM_API_VERSION > 0x200
        btl->gm_addr.global_id = global_id;
#else
        strncpy( btl->gm_addr.global_id, global_id, GM_MAX_HOST_NAME_LEN );
#endif  /* GM_API_VERSION > 0x200 */

        if(mca_btl_gm_component.gm_debug > 0) {
            opal_output(0, "[%d,%d,%d] gm_port %08X, board %lu, global %lu node %lu port %lu\n", 
                ORTE_NAME_ARGS(orte_process_info.my_name), port, board_no, global_id, node_id, port_no);
        }

        if((rc = mca_btl_gm_module_init(btl)) != OMPI_SUCCESS) {
            opal_output(0, "[%s:%d] unable to initialze gm port", __FILE__, __LINE__);
            return rc;
        }

        /* everything is OK let's mark it as usable and go to the next one */
        mca_btl_gm_component.gm_btls[mca_btl_gm_component.gm_num_btls] = btl;
        if(++mca_btl_gm_component.gm_num_btls >= mca_btl_gm_component.gm_max_btls ) {
            break;
        }
    }
    return OMPI_SUCCESS;
}

/*
 *  Register GM component addressing information. The MCA framework
 *  will make this available to all peers.
 */

static int
mca_btl_gm_modex_send(void)
{
    int         rc;
    size_t      i;
    size_t      size;
    mca_btl_gm_addr_t *addrs = NULL;

    size = mca_btl_gm_component.gm_num_btls * sizeof (mca_btl_gm_addr_t);
    if (0 != size) {
        addrs = (mca_btl_gm_addr_t *)malloc (size);
        if (NULL == addrs) {
            return OMPI_ERR_OUT_OF_RESOURCE;
        }

        for (i = 0; i < mca_btl_gm_component.gm_num_btls; i++) {
            mca_btl_gm_module_t *btl = mca_btl_gm_component.gm_btls[i];
            addrs[i] = btl->gm_addr;
            MCA_BTL_GM_ADDR_HTON(addrs[i]);
        }
    }
    rc = mca_pml_base_modex_send (&mca_btl_gm_component.super.btl_version, addrs, size);
    if (NULL != addrs) {
        free (addrs);
    }
    return rc;
}
                                                                                                                

/*
 * Initialize the GM component,
 * check how many boards are available and open ports on them.
 */
                                                                                                                
mca_btl_base_module_t **
mca_btl_gm_component_init (int *num_btl_modules,
                           bool enable_progress_threads,
                           bool enable_mpi_threads)
{
    mca_btl_base_module_t **btls;
    *num_btl_modules = 0;

    OPAL_THREAD_LOCK(&mca_btl_gm_component.gm_lock);

    /* try to initialize GM */
    if( GM_SUCCESS != gm_init() ) {
        opal_output( 0, "[%s:%d] error in initializing the gm library\n", __FILE__, __LINE__ );
        mca_btl_gm_component.gm_num_btls = 0;
        mca_btl_gm_modex_send();
        OPAL_THREAD_UNLOCK(&mca_btl_gm_component.gm_lock);
        return NULL;
    }

    /* First discover all available boards. For each board we create a unique BTL */
    mca_btl_gm_component.gm_btls = malloc( mca_btl_gm_component.gm_max_btls * sizeof (mca_btl_gm_module_t *));
    if (NULL == mca_btl_gm_component.gm_btls) {
        opal_output( 0, "[%s:%d] out of resources.", __FILE__, __LINE__ );
        OPAL_THREAD_UNLOCK(&mca_btl_gm_component.gm_lock);
        return NULL;
    }

    /* initialize gm */
    if (OMPI_SUCCESS != mca_btl_gm_discover()) {
        mca_btl_base_error_no_nics("Myrinet/GM", "NIC");
        mca_btl_gm_component.gm_num_btls = 0;
        mca_btl_gm_modex_send();
        OPAL_THREAD_UNLOCK(&mca_btl_gm_component.gm_lock);
        return NULL;
    }
    if (mca_btl_gm_component.gm_num_btls == 0) {
        mca_btl_base_error_no_nics("Myrinet/GM", "NIC");
        mca_btl_gm_component.gm_num_btls = 0;
        mca_btl_gm_modex_send();
        OPAL_THREAD_UNLOCK(&mca_btl_gm_component.gm_lock);
        return NULL;
    }

    /* publish GM parameters with the MCA framework */
    if (OMPI_SUCCESS != mca_btl_gm_modex_send()) {
        OPAL_THREAD_UNLOCK(&mca_btl_gm_component.gm_lock);
        return NULL;
    }

    /* return array of BTLs */
    btls = (mca_btl_base_module_t**) malloc (
                mca_btl_gm_component.gm_num_btls * sizeof(mca_btl_base_module_t *));
    if (NULL == btls) {
        OPAL_THREAD_UNLOCK(&mca_btl_gm_component.gm_lock);
        return NULL;
    }

    memcpy(btls, mca_btl_gm_component.gm_btls,
           mca_btl_gm_component.gm_num_btls * sizeof(mca_btl_gm_module_t *));
    *num_btl_modules = mca_btl_gm_component.gm_num_btls;
    OPAL_THREAD_UNLOCK(&mca_btl_gm_component.gm_lock);
    return btls;
}
                                                                                                                

/*
 *  GM component progress.
 */


int mca_btl_gm_component_progress()
{
    static int32_t inprogress = 0;
    int count = 0;
    size_t i;
    
    /* could get into deadlock in this case as we post recvs after callback completes */
    if(OPAL_THREAD_ADD32(&inprogress, 1) > 1) {
        OPAL_THREAD_ADD32(&inprogress, -1);
        return OMPI_SUCCESS;
    }

    OPAL_THREAD_LOCK(&mca_btl_gm_component.gm_lock);
    for( i = 0; i < mca_btl_gm_component.gm_num_btls; ) {
        mca_btl_gm_module_t* btl = mca_btl_gm_component.gm_btls[i];
        gm_recv_event_t* event = gm_receive(btl->port);

        /* If there are no receive events just skip the function call */
        switch(gm_ntohc(event->recv.type)) {
            case GM_FAST_RECV_EVENT:
            case GM_FAST_PEER_RECV_EVENT:
            case GM_FAST_HIGH_RECV_EVENT:
            case GM_FAST_HIGH_PEER_RECV_EVENT:
                {
                unsigned char* buffer = (unsigned char*)gm_ntohp(event->recv.buffer);
                mca_btl_gm_frag_t* frag = (mca_btl_gm_frag_t*)(buffer - sizeof(mca_btl_gm_frag_t));
                mca_btl_base_header_t* hdr = (mca_btl_base_header_t *)gm_ntohp(event->recv.message);
                mca_btl_base_recv_reg_t* reg;
                frag->segment.seg_addr.pval = (hdr+1);
                frag->segment.seg_len = gm_ntohl(event->recv.length) - sizeof(mca_btl_base_header_t);
                reg = &btl->gm_reg[hdr->tag];

                /* cbfunc may be null if interface goes down.. */
                if(reg->cbfunc) { 
                    OPAL_THREAD_UNLOCK(&mca_btl_gm_component.gm_lock);
                    reg->cbfunc(&btl->super, hdr->tag, &frag->base, reg->cbdata);
                    OPAL_THREAD_LOCK(&mca_btl_gm_component.gm_lock);
                } else { 
                    btl->error_cb(&btl->super, 
                                  MCA_BTL_ERROR_FLAGS_FATAL);
                    return 0;
                }
                MCA_BTL_GM_FRAG_POST(btl,frag);
                count++;
                break;
                }
            case GM_RECV_EVENT:
            case GM_PEER_RECV_EVENT:
            case GM_HIGH_RECV_EVENT:
            case GM_HIGH_PEER_RECV_EVENT:
                {
                unsigned char* buffer = (unsigned char*)gm_ntohp(event->recv.buffer);
                mca_btl_gm_frag_t* frag = (mca_btl_gm_frag_t*)(buffer - sizeof(mca_btl_gm_frag_t));
                mca_btl_base_header_t* hdr = (mca_btl_base_header_t*)buffer;
                mca_btl_base_recv_reg_t* reg;
                frag->segment.seg_addr.pval = (hdr+1);
                frag->segment.seg_len = gm_ntohl(event->recv.length) - sizeof(mca_btl_base_header_t);
                reg = &btl->gm_reg[hdr->tag];
                if(reg->cbfunc) { 
                    OPAL_THREAD_UNLOCK(&mca_btl_gm_component.gm_lock);
                    reg->cbfunc(&btl->super, hdr->tag, &frag->base, reg->cbdata);
                    OPAL_THREAD_LOCK(&mca_btl_gm_component.gm_lock);
                    
                    MCA_BTL_GM_FRAG_POST(btl,frag);
                } else { 
                    btl->error_cb(&btl->super, 
                                  MCA_BTL_ERROR_FLAGS_FATAL);
                    return 0;
                }
                count++;
                break;
                }
            case GM_NO_RECV_EVENT:
                i++;
                break;
            default:
                gm_unknown(btl->port, event);
                break;
        }
    }
    OPAL_THREAD_UNLOCK(&mca_btl_gm_component.gm_lock);
    OPAL_THREAD_ADD32(&inprogress, -1);
    return count;
}


#if OMPI_ENABLE_PROGRESS_THREADS
static void* mca_btl_gm_progress_thread( opal_object_t* arg )
{
    opal_thread_t* thread = (opal_thread_t*)arg;
    mca_btl_gm_module_t* btl = thread->t_arg;

    /* This thread enter in a cancel enabled state */
    pthread_setcancelstate( PTHREAD_CANCEL_ENABLE, NULL );
    pthread_setcanceltype( PTHREAD_CANCEL_ASYNCHRONOUS, NULL );

    OPAL_THREAD_LOCK(&mca_btl_gm_component.gm_lock);
    while(btl->gm_progress) {
        gm_recv_event_t* event;

        /* dont process events while the app is in the library */
        while(opal_progress_threads()) {
            OPAL_THREAD_UNLOCK(&mca_btl_gm_component.gm_lock);
            while(opal_progress_threads())
               sched_yield();
            usleep(100); /* give app a chance to re-enter library */
            OPAL_THREAD_LOCK(&mca_btl_gm_component.gm_lock);
        }

        /* otherwise processes any pending events */
        event = gm_blocking_receive_no_spin(btl->port);
        switch(gm_ntohc(event->recv.type)) {
            case GM_FAST_RECV_EVENT:
            case GM_FAST_PEER_RECV_EVENT:
            case GM_FAST_HIGH_RECV_EVENT:
            case GM_FAST_HIGH_PEER_RECV_EVENT:
                {
                unsigned char* buffer = (unsigned char*)gm_ntohp(event->recv.buffer);
                mca_btl_gm_frag_t* frag = (mca_btl_gm_frag_t*)(buffer - sizeof(mca_btl_gm_frag_t));
                mca_btl_base_header_t* hdr = (mca_btl_base_header_t *)gm_ntohp(event->recv.message);
                mca_btl_base_recv_reg_t* reg;
                frag->segment.seg_addr.pval = (hdr+1);
                frag->segment.seg_len = gm_ntohl(event->recv.length) - sizeof(mca_btl_base_header_t);
                reg = &btl->gm_reg[hdr->tag];

                OPAL_THREAD_UNLOCK(&mca_btl_gm_component.gm_lock);
                reg->cbfunc(&btl->super, hdr->tag, &frag->base, reg->cbdata);
                OPAL_THREAD_LOCK(&mca_btl_gm_component.gm_lock);

                MCA_BTL_GM_FRAG_POST(btl,frag);
                break;
                }
            case GM_RECV_EVENT:
            case GM_PEER_RECV_EVENT:
            case GM_HIGH_RECV_EVENT:
            case GM_HIGH_PEER_RECV_EVENT:
                {
                unsigned char* buffer = (unsigned char*)gm_ntohp(event->recv.buffer);
                mca_btl_gm_frag_t* frag = (mca_btl_gm_frag_t*)(buffer - sizeof(mca_btl_gm_frag_t));
                mca_btl_base_header_t* hdr = (mca_btl_base_header_t*)buffer;
                mca_btl_base_recv_reg_t* reg;
                frag->segment.seg_addr.pval = (hdr+1);
                frag->segment.seg_len = gm_ntohl(event->recv.length) - sizeof(mca_btl_base_header_t);
                reg = &btl->gm_reg[hdr->tag];

                OPAL_THREAD_UNLOCK(&mca_btl_gm_component.gm_lock);
                reg->cbfunc(&btl->super, hdr->tag, &frag->base, reg->cbdata);
                OPAL_THREAD_LOCK(&mca_btl_gm_component.gm_lock);

                MCA_BTL_GM_FRAG_POST(btl,frag);
                break;
                }
            case _GM_SLEEP_EVENT:
                OPAL_THREAD_UNLOCK(&mca_btl_gm_component.gm_lock);
                gm_unknown(btl->port, event);
                OPAL_THREAD_LOCK(&mca_btl_gm_component.gm_lock);
                break;
            default:
                gm_unknown(btl->port, event);
                break;
        }
    }
    OPAL_THREAD_UNLOCK(&mca_btl_gm_component.gm_lock);
    return PTHREAD_CANCELED;
}
#endif