int MPIR_Init_async_thread(void)
{
#if MPICH_THREAD_LEVEL == MPI_THREAD_MULTIPLE
int mpi_errno = MPI_SUCCESS;
MPIR_Comm *comm_self_ptr;
int err = 0;
MPIR_FUNC_TERSE_STATE_DECL(MPID_STATE_MPIR_INIT_ASYNC_THREAD);
MPIR_FUNC_TERSE_ENTER(MPID_STATE_MPIR_INIT_ASYNC_THREAD);
/* Dup comm world for the progress thread */
MPIR_Comm_get_ptr(MPI_COMM_SELF, comm_self_ptr);
mpi_errno = MPIR_Comm_dup_impl(comm_self_ptr, &progress_comm_ptr);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
MPID_Thread_cond_create(&progress_cond, &err);
MPIR_ERR_CHKANDJUMP1(err, mpi_errno, MPI_ERR_OTHER, "**cond_create", "**cond_create %s", strerror(err));
MPID_Thread_mutex_create(&progress_mutex, &err);
MPIR_ERR_CHKANDJUMP1(err, mpi_errno, MPI_ERR_OTHER, "**mutex_create", "**mutex_create %s", strerror(err));
MPID_Thread_create((MPID_Thread_func_t) progress_fn, NULL, &progress_thread_id, &err);
MPIR_ERR_CHKANDJUMP1(err, mpi_errno, MPI_ERR_OTHER, "**mutex_create", "**mutex_create %s", strerror(err));
MPIR_FUNC_TERSE_EXIT(MPID_STATE_MPIR_INIT_ASYNC_THREAD);
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
#else
return MPI_SUCCESS;
#endif /* MPICH_THREAD_LEVEL == MPI_THREAD_MULTIPLE */
}
int MPID_nem_tcp_init (MPIDI_PG_t *pg_p, int pg_rank, char **bc_val_p, int *val_max_sz_p)
{
int mpi_errno = MPI_SUCCESS;
MPIDI_STATE_DECL(MPID_STATE_MPID_NEM_TCP_INIT);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_NEM_TCP_INIT);
MPID_nem_net_module_vc_dbg_print_sendq = MPID_nem_tcp_vc_dbg_print_sendq;
/* first make sure that our private fields in the vc fit into the area provided */
MPIU_Assert(sizeof(MPID_nem_tcp_vc_area) <= MPIDI_NEM_VC_NETMOD_AREA_LEN);
/* set up listener socket */
mpi_errno = set_up_listener();
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
/* create business card */
mpi_errno = MPID_nem_tcp_get_business_card(pg_rank, bc_val_p, val_max_sz_p);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPID_nem_tcp_sm_init();
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPID_nem_tcp_send_init();
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
#ifdef HAVE_SIGNAL
{
/* In order to be able to handle socket errors on our own, we need
to ignore SIGPIPE. This may cause problems for programs that
intend to handle SIGPIPE or count on being killed, but I expect
such programs are very rare, and I'm not sure what the best
solution would be anyway. */
void *ret;
ret = signal(SIGPIPE, SIG_IGN);
MPIR_ERR_CHKANDJUMP1(ret == SIG_ERR, mpi_errno, MPI_ERR_OTHER, "**signal", "**signal %s", MPIU_Strerror(errno));
if (ret != SIG_DFL && ret != SIG_IGN) {
/* The app has set its own signal handler. Replace the previous handler. */
ret = signal(SIGPIPE, ret);
MPIR_ERR_CHKANDJUMP1(ret == SIG_ERR, mpi_errno, MPI_ERR_OTHER, "**signal", "**signal %s", MPIU_Strerror(errno));
}
}
#endif
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_MPID_NEM_TCP_INIT);
/* fprintf(stdout, FCNAME " Exit\n"); fflush(stdout); */
return mpi_errno;
fn_fail:
/* fprintf(stdout, "failure. mpi_errno = %d\n", mpi_errno); */
goto fn_exit;
}
int MPID_nem_ptl_init_id(MPIDI_VC_t *vc)
{
int mpi_errno = MPI_SUCCESS;
MPID_nem_ptl_vc_area *const vc_ptl = VC_PTL(vc);
char *bc;
int pmi_errno;
int val_max_sz;
MPIR_CHKLMEM_DECL(1);
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPID_NEM_PTL_INIT_ID);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPID_NEM_PTL_INIT_ID);
pmi_errno = PMI_KVS_Get_value_length_max(&val_max_sz);
MPIR_ERR_CHKANDJUMP1(pmi_errno, mpi_errno, MPI_ERR_OTHER, "**fail", "**fail %d", pmi_errno);
MPIR_CHKLMEM_MALLOC(bc, char *, val_max_sz, mpi_errno, "bc");
mpi_errno = vc->pg->getConnInfo(vc->pg_rank, bc, val_max_sz, vc->pg);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = MPID_nem_ptl_get_id_from_bc(bc, &vc_ptl->id, &vc_ptl->pt, &vc_ptl->ptg, &vc_ptl->ptc, &vc_ptl->ptr, &vc_ptl->ptrg, &vc_ptl->ptrc);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
vc_ptl->id_initialized = TRUE;
MPIDI_CHANGE_VC_STATE(vc, ACTIVE);
fn_exit:
MPIR_CHKLMEM_FREEALL();
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPID_NEM_PTL_INIT_ID);
return mpi_errno;
fn_fail:
goto fn_exit;
}
static int handler_recv_complete(const ptl_event_t *e)
{
int mpi_errno = MPI_SUCCESS;
MPID_Request *const rreq = e->user_ptr;
int ret;
int i;
MPIDI_STATE_DECL(MPID_STATE_HANDLER_RECV_COMPLETE);
MPIDI_FUNC_ENTER(MPID_STATE_HANDLER_RECV_COMPLETE);
MPIU_Assert(e->type == PTL_EVENT_REPLY || e->type == PTL_EVENT_PUT || e->type == PTL_EVENT_PUT_OVERFLOW);
if (REQ_PTL(rreq)->md != PTL_INVALID_HANDLE) {
ret = PtlMDRelease(REQ_PTL(rreq)->md);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlmdrelease", "**ptlmdrelease %s", MPID_nem_ptl_strerror(ret));
}
for (i = 0; i < MPID_NEM_PTL_NUM_CHUNK_BUFFERS; ++i)
if (REQ_PTL(rreq)->chunk_buffer[i])
MPIU_Free(REQ_PTL(rreq)->chunk_buffer[i]);
mpi_errno = MPID_Request_complete(rreq);
if (mpi_errno) {
MPIR_ERR_POP(mpi_errno);
}
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_HANDLER_RECV_COMPLETE);
return mpi_errno;
fn_fail:
goto fn_exit;
}
/* this function is not used in pmi2 */
static int publish_node_id(MPIDI_PG_t *pg, int our_pg_rank)
{
int mpi_errno = MPI_SUCCESS;
int pmi_errno;
int ret;
char *key;
int key_max_sz;
char *kvs_name;
MPIU_CHKLMEM_DECL(1);
/* set MPIU_hostname */
ret = gethostname(MPIU_hostname, MAX_HOSTNAME_LEN);
MPIR_ERR_CHKANDJUMP2(ret == -1, mpi_errno, MPI_ERR_OTHER, "**sock_gethost", "**sock_gethost %s %d", MPIU_Strerror(errno), errno);
MPIU_hostname[MAX_HOSTNAME_LEN-1] = '\0';
/* Allocate space for pmi key */
pmi_errno = PMI_KVS_Get_key_length_max(&key_max_sz);
MPIR_ERR_CHKANDJUMP1(pmi_errno, mpi_errno, MPI_ERR_OTHER, "**fail", "**fail %d", pmi_errno);
MPIU_CHKLMEM_MALLOC(key, char *, key_max_sz, mpi_errno, "key");
mpi_errno = MPIDI_PG_GetConnKVSname(&kvs_name);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
/* Put my hostname id */
if (pg->size > 1)
{
memset(key, 0, key_max_sz);
MPL_snprintf(key, key_max_sz, "hostname[%d]", our_pg_rank);
pmi_errno = PMI_KVS_Put(kvs_name, key, MPIU_hostname);
MPIR_ERR_CHKANDJUMP1(pmi_errno != PMI_SUCCESS, mpi_errno, MPI_ERR_OTHER, "**pmi_kvs_put", "**pmi_kvs_put %d", pmi_errno);
pmi_errno = PMI_KVS_Commit(kvs_name);
MPIR_ERR_CHKANDJUMP1(pmi_errno != PMI_SUCCESS, mpi_errno, MPI_ERR_OTHER, "**pmi_kvs_commit", "**pmi_kvs_commit %d", pmi_errno);
pmi_errno = PMI_Barrier();
MPIR_ERR_CHKANDJUMP1(pmi_errno != PMI_SUCCESS, mpi_errno, MPI_ERR_OTHER, "**pmi_barrier", "**pmi_barrier %d", pmi_errno);
}
fn_exit:
MPIU_CHKLMEM_FREEALL();
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPID_nem_tcp_get_business_card (int my_rank, char **bc_val_p, int *val_max_sz_p)
{
int mpi_errno = MPI_SUCCESS;
int str_errno = MPIU_STR_SUCCESS;
MPIDU_Sock_ifaddr_t ifaddr;
char ifname[MAX_HOST_DESCRIPTION_LEN];
int ret;
struct sockaddr_in sock_id;
socklen_t len;
MPIDI_STATE_DECL(MPID_STATE_MPID_NEM_TCP_GET_BUSINESS_CARD);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_NEM_TCP_GET_BUSINESS_CARD);
mpi_errno = GetSockInterfaceAddr(my_rank, ifname, sizeof(ifname), &ifaddr);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
str_errno = MPIU_Str_add_string_arg(bc_val_p, val_max_sz_p, MPIDI_CH3I_HOST_DESCRIPTION_KEY, ifname);
if (str_errno) {
MPIR_ERR_CHKANDJUMP(str_errno == MPIU_STR_NOMEM, mpi_errno, MPI_ERR_OTHER, "**buscard_len");
MPIR_ERR_SETANDJUMP(mpi_errno, MPI_ERR_OTHER, "**buscard");
}
len = sizeof(sock_id);
ret = getsockname (MPID_nem_tcp_g_lstn_sc.fd, (struct sockaddr *)&sock_id, &len);
MPIR_ERR_CHKANDJUMP1 (ret == -1, mpi_errno, MPI_ERR_OTHER, "**getsockname", "**getsockname %s", MPIU_Strerror (errno));
str_errno = MPIU_Str_add_int_arg (bc_val_p, val_max_sz_p, MPIDI_CH3I_PORT_KEY, ntohs(sock_id.sin_port));
if (str_errno) {
MPIR_ERR_CHKANDJUMP(str_errno == MPIU_STR_NOMEM, mpi_errno, MPI_ERR_OTHER, "**buscard_len");
MPIR_ERR_SETANDJUMP(mpi_errno, MPI_ERR_OTHER, "**buscard");
}
if (ifaddr.len > 0 && ifaddr.type == AF_INET)
{
unsigned char *p;
p = (unsigned char *)(ifaddr.ifaddr);
MPL_snprintf( ifname, sizeof(ifname), "%u.%u.%u.%u", p[0], p[1], p[2], p[3] );
MPIU_DBG_MSG_S(CH3_CONNECT,VERBOSE,"ifname = %s",ifname );
str_errno = MPIU_Str_add_string_arg(bc_val_p, val_max_sz_p, MPIDI_CH3I_IFNAME_KEY, ifname);
if (str_errno) {
MPIR_ERR_CHKANDJUMP(str_errno == MPIU_STR_NOMEM, mpi_errno, MPI_ERR_OTHER, "**buscard_len");
MPIR_ERR_SETANDJUMP(mpi_errno, MPI_ERR_OTHER, "**buscard");
}
}
/* printf("MPID_nem_tcp_get_business_card. port=%d\n", sock_id.sin_port); */
fn_exit:
/* fprintf(stdout, "MPID_nem_tcp_get_business_card Exit, mpi_errno=%d\n", mpi_errno); fflush(stdout); */
MPIDI_FUNC_EXIT(MPID_STATE_MPID_NEM_TCP_GET_BUSINESS_CARD);
return mpi_errno;
fn_fail:
goto fn_exit;
}
/* MPIDI_CH3_EagerNoncontigSend - Eagerly send noncontiguous data */
int MPIDI_CH3_EagerNoncontigSend( MPID_Request **sreq_p,
MPIDI_CH3_Pkt_type_t reqtype,
const void * buf, MPI_Aint count,
MPI_Datatype datatype, MPIDI_msg_sz_t data_sz,
int rank,
int tag, MPID_Comm * comm,
int context_offset )
{
int mpi_errno = MPI_SUCCESS;
MPIDI_VC_t * vc;
MPID_Request *sreq = *sreq_p;
MPIDI_CH3_Pkt_t upkt;
MPIDI_CH3_Pkt_eager_send_t * const eager_pkt = &upkt.eager_send;
MPIU_DBG_MSG_FMT(CH3_OTHER,VERBOSE,(MPIU_DBG_FDEST,
"sending non-contiguous eager message, data_sz=" MPIDI_MSG_SZ_FMT,
data_sz));
sreq->dev.OnDataAvail = 0;
sreq->dev.OnFinal = 0;
MPIDI_Pkt_init(eager_pkt, reqtype);
eager_pkt->match.parts.rank = comm->rank;
eager_pkt->match.parts.tag = tag;
eager_pkt->match.parts.context_id = comm->context_id + context_offset;
eager_pkt->sender_req_id = MPI_REQUEST_NULL;
eager_pkt->data_sz = data_sz;
MPIDI_Comm_get_vc_set_active(comm, rank, &vc);
MPIDI_VC_FAI_send_seqnum(vc, seqnum);
MPIDI_Pkt_set_seqnum(eager_pkt, seqnum);
MPIDI_Request_set_seqnum(sreq, seqnum);
MPIU_DBG_MSGPKT(vc,tag,eager_pkt->match.parts.context_id,rank,data_sz,
"Eager");
sreq->dev.segment_ptr = MPID_Segment_alloc( );
MPIR_ERR_CHKANDJUMP1((sreq->dev.segment_ptr == NULL), mpi_errno, MPI_ERR_OTHER, "**nomem", "**nomem %s", "MPID_Segment_alloc");
MPID_Segment_init(buf, count, datatype, sreq->dev.segment_ptr, 0);
sreq->dev.segment_first = 0;
sreq->dev.segment_size = data_sz;
MPID_THREAD_CS_ENTER(POBJ, vc->pobj_mutex);
mpi_errno = vc->sendNoncontig_fn(vc, sreq, eager_pkt,
sizeof(MPIDI_CH3_Pkt_eager_send_t));
MPID_THREAD_CS_EXIT(POBJ, vc->pobj_mutex);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
fn_exit:
return mpi_errno;
fn_fail:
*sreq_p = NULL;
goto fn_exit;
}
static int mpi_to_pmi_keyvals( MPIR_Info *info_ptr, PMI_keyval_t **kv_ptr,
int *nkeys_ptr )
{
char key[MPI_MAX_INFO_KEY];
PMI_keyval_t *kv = 0;
int i, nkeys = 0, vallen, flag, mpi_errno=MPI_SUCCESS;
if (!info_ptr || info_ptr->handle == MPI_INFO_NULL) {
goto fn_exit;
}
MPIR_Info_get_nkeys_impl( info_ptr, &nkeys );
if (nkeys == 0) {
goto fn_exit;
}
kv = (PMI_keyval_t *)MPL_malloc( nkeys * sizeof(PMI_keyval_t) );
if (!kv) { MPIR_ERR_POP(mpi_errno); }
for (i=0; i<nkeys; i++) {
mpi_errno = MPIR_Info_get_nthkey_impl( info_ptr, i, key );
if (mpi_errno) { MPIR_ERR_POP(mpi_errno); }
MPIR_Info_get_valuelen_impl( info_ptr, key, &vallen, &flag );
MPIR_ERR_CHKANDJUMP1(!flag, mpi_errno, MPI_ERR_OTHER,"**infonokey", "**infonokey %s", key);
kv[i].key = MPL_strdup(key);
kv[i].val = MPL_malloc( vallen + 1 );
if (!kv[i].key || !kv[i].val) {
MPIR_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,"**nomem" );
}
MPIR_Info_get_impl( info_ptr, key, vallen+1, kv[i].val, &flag );
MPIR_ERR_CHKANDJUMP1(!flag, mpi_errno, MPI_ERR_OTHER,"**infonokey", "**infonokey %s", key);
MPL_DBG_MSG_FMT(MPIDI_CH3_DBG_OTHER,TERSE,(MPL_DBG_FDEST,"key: <%s>, value: <%s>\n", kv[i].key, kv[i].val));
}
fn_fail:
fn_exit:
*kv_ptr = kv;
*nkeys_ptr = nkeys;
return mpi_errno;
}
int MPID_nem_lmt_vmsplice_start_recv(MPIDI_VC_t *vc, MPID_Request *rreq, MPL_IOV s_cookie)
{
int mpi_errno = MPI_SUCCESS;
int i;
int complete = 0;
struct lmt_vmsplice_node *node = NULL;
MPIDI_CH3I_VC *vc_ch = &vc->ch;
int pipe_fd;
MPIDI_STATE_DECL(MPID_STATE_MPID_NEM_LMT_VMSPLICE_START_RECV);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_NEM_LMT_VMSPLICE_START_RECV);
if (vc_ch->lmt_recv_copy_buf_handle == NULL) {
MPIU_Assert(s_cookie.MPL_IOV_BUF != NULL);
vc_ch->lmt_recv_copy_buf_handle = MPL_strdup(s_cookie.MPL_IOV_BUF);
}
/* XXX DJG FIXME in a real version we would want to cache the fd on the vc
so that we don't have two open's on the critical path every time. */
pipe_fd = open(vc_ch->lmt_recv_copy_buf_handle, O_NONBLOCK|O_RDONLY);
MPIR_ERR_CHKANDJUMP1(pipe_fd < 0, mpi_errno, MPI_ERR_OTHER, "**open",
"**open %s", MPIU_Strerror(errno));
MPID_nem_lmt_send_CTS(vc, rreq, NULL, 0);
mpi_errno = populate_iov_from_req(rreq);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = do_readv(rreq, pipe_fd, rreq->dev.iov, &rreq->dev.iov_offset,
&rreq->dev.iov_count, &complete);
/* push request if not complete for progress checks later */
if (!complete) {
node = MPL_malloc(sizeof(struct lmt_vmsplice_node));
node->pipe_fd = pipe_fd;
node->req = rreq;
node->next = outstanding_head;
outstanding_head = node;
++MPID_nem_local_lmt_pending;
}
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_MPID_NEM_LMT_VMSPLICE_START_RECV);
return mpi_errno;
fn_fail:
goto fn_exit;
}
/* fills in req->dev.iov{,_offset,_count} based on the datatype info in the
request, creating a segment if necessary */
static int populate_iov_from_req(MPID_Request *req)
{
int mpi_errno = MPI_SUCCESS;
int dt_contig;
MPI_Aint dt_true_lb;
MPIDI_msg_sz_t data_sz;
MPID_Datatype * dt_ptr;
/* find out contig/noncontig, size, and lb for the datatype */
MPIDI_Datatype_get_info(req->dev.user_count, req->dev.datatype,
dt_contig, data_sz, dt_ptr, dt_true_lb);
if (dt_contig) {
/* handle the iov creation ourselves */
req->dev.iov[0].MPL_IOV_BUF = (char *)req->dev.user_buf + dt_true_lb;
req->dev.iov[0].MPL_IOV_LEN = data_sz;
req->dev.iov_count = 1;
}
else {
/* use the segment routines to handle the iovec creation */
MPIU_Assert(req->dev.segment_ptr == NULL);
req->dev.iov_count = MPL_IOV_LIMIT;
req->dev.iov_offset = 0;
/* XXX DJG FIXME where is this segment freed? */
req->dev.segment_ptr = MPID_Segment_alloc();
MPIR_ERR_CHKANDJUMP1((req->dev.segment_ptr == NULL), mpi_errno,
MPI_ERR_OTHER, "**nomem",
"**nomem %s", "MPID_Segment_alloc");
MPID_Segment_init(req->dev.user_buf, req->dev.user_count,
req->dev.datatype, req->dev.segment_ptr, 0);
req->dev.segment_first = 0;
req->dev.segment_size = data_sz;
/* FIXME we should write our own function that isn't dependent on
the in-request iov array. This will let us use IOVs that are
larger than MPL_IOV_LIMIT. */
mpi_errno = MPIDI_CH3U_Request_load_send_iov(req, &req->dev.iov[0],
&req->dev.iov_count);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
fn_fail:
return mpi_errno;
}
int MPID_nem_lmt_vmsplice_start_send(MPIDI_VC_t *vc, MPID_Request *sreq, MPL_IOV r_cookie)
{
int mpi_errno = MPI_SUCCESS;
MPIDI_STATE_DECL(MPID_STATE_MPID_NEM_LMT_VMSPLICE_START_SEND);
int pipe_fd;
int complete;
struct lmt_vmsplice_node *node = NULL;
int (*reqFn)(MPIDI_VC_t *, MPID_Request *, int *);
MPIDI_CH3I_VC *vc_ch = &vc->ch;
MPIDI_FUNC_ENTER(MPID_STATE_MPID_NEM_LMT_VMSPLICE_START_SEND);
/* Must do this after the other side has opened for reading, otherwise we
will error out with ENXIO. This will be indicated by the receipt of a
CTS message. */
pipe_fd = open(vc_ch->lmt_copy_buf_handle, O_NONBLOCK|O_WRONLY);
MPIR_ERR_CHKANDJUMP1(pipe_fd < 0, mpi_errno, MPI_ERR_OTHER, "**open",
"**open %s", MPIU_Strerror(errno));
mpi_errno = populate_iov_from_req(sreq);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
/* send the first flight */
sreq->ch.vc = vc; /* XXX DJG is this already assigned? */
complete = 0;
mpi_errno = do_vmsplice(sreq, pipe_fd, sreq->dev.iov,
&sreq->dev.iov_offset, &sreq->dev.iov_count, &complete);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
if (!complete) {
/* push for later progress */
node = MPL_malloc(sizeof(struct lmt_vmsplice_node));
node->pipe_fd = pipe_fd;
node->req = sreq;
node->next = outstanding_head;
outstanding_head = node;
++MPID_nem_local_lmt_pending;
}
fn_fail:
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_MPID_NEM_LMT_VMSPLICE_START_SEND);
return mpi_errno;
}
int MPID_nem_ptl_pkt_cancel_send_resp_handler(MPIDI_VC_t *vc, MPIDI_CH3_Pkt_t *pkt,
MPIDI_msg_sz_t *buflen, MPID_Request **rreqp)
{
int mpi_errno = MPI_SUCCESS;
MPID_Request *sreq;
MPIDI_nem_ptl_pkt_cancel_send_resp_t *resp_pkt = (MPIDI_nem_ptl_pkt_cancel_send_resp_t *)pkt;
int i, ret;
MPID_Request_get_ptr(resp_pkt->sender_req_id, sreq);
if (resp_pkt->ack) {
MPIR_STATUS_SET_CANCEL_BIT(sreq->status, TRUE);
/* remove/free any remaining get MEs and handles */
for (i = 0; i < REQ_PTL(sreq)->num_gets; i++) {
ret = PtlMEUnlink(REQ_PTL(sreq)->get_me_p[i]);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlmeunlink", "**ptlmeunlink %s", MPID_nem_ptl_strerror(ret));
mpi_errno = MPID_Request_complete(sreq);
if (mpi_errno != MPI_SUCCESS) {
MPIR_ERR_POP(mpi_errno);
}
}
if (REQ_PTL(sreq)->get_me_p)
MPIU_Free(REQ_PTL(sreq)->get_me_p);
MPIU_DBG_MSG(CH3_OTHER,TYPICAL,"message cancelled");
} else {
MPIR_STATUS_SET_CANCEL_BIT(sreq->status, FALSE);
MPIU_DBG_MSG(CH3_OTHER,TYPICAL,"unable to cancel message");
}
mpi_errno = MPID_Request_complete(sreq);
if (mpi_errno != MPI_SUCCESS) {
MPIR_ERR_POP(mpi_errno);
}
*rreqp = NULL;
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPID_NS_Lookup( MPID_NS_Handle handle, const MPID_Info *info_ptr,
const char service_name[], char port[] )
{
int mpi_errno = MPI_SUCCESS;
int rc;
MPIU_UNREFERENCED_ARG(info_ptr);
MPIU_UNREFERENCED_ARG(handle);
#ifdef USE_PMI2_API
/* release the global CS for PMI calls */
MPID_THREAD_CS_EXIT(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
rc = PMI2_Nameserv_lookup(service_name, info_ptr, port, MPI_MAX_PORT_NAME);
MPID_THREAD_CS_ENTER(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
#else
rc = PMI_Lookup_name( service_name, port );
#endif
MPIR_ERR_CHKANDJUMP1(rc, mpi_errno, MPI_ERR_NAME, "**namepubnotfound", "**namepubnotfound %s", service_name);
fn_fail:
return mpi_errno;
}
int MPID_NS_Unpublish( MPID_NS_Handle handle, const MPID_Info *info_ptr,
const char service_name[] )
{
int mpi_errno = MPI_SUCCESS;
int rc;
MPIU_UNREFERENCED_ARG(info_ptr);
MPIU_UNREFERENCED_ARG(handle);
#ifdef USE_PMI2_API
/* release the global CS for PMI calls */
MPID_THREAD_CS_EXIT(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
rc = PMI2_Nameserv_unpublish(service_name, info_ptr);
MPID_THREAD_CS_ENTER(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
#else
rc = PMI_Unpublish_name( service_name );
#endif
MPIR_ERR_CHKANDJUMP1(rc, mpi_errno, MPI_ERR_SERVICE, "**namepubnotunpub", "**namepubnotunpub %s", service_name);
fn_fail:
return mpi_errno;
}
static int handler_send(const ptl_event_t *e)
{
int mpi_errno = MPI_SUCCESS;
MPID_Request *const sreq = e->user_ptr;
int i, ret;
MPIDI_STATE_DECL(MPID_STATE_HANDLER_SEND);
MPIDI_FUNC_ENTER(MPID_STATE_HANDLER_SEND);
MPIU_Assert(e->type == PTL_EVENT_SEND || e->type == PTL_EVENT_GET);
/* if we are done, release all netmod resources */
if (MPID_cc_get(sreq->cc) == 1) {
if (REQ_PTL(sreq)->md != PTL_INVALID_HANDLE) {
ret = PtlMDRelease(REQ_PTL(sreq)->md);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlmdrelease", "**ptlmdrelease %s", MPID_nem_ptl_strerror(ret));
}
for (i = 0; i < MPID_NEM_PTL_NUM_CHUNK_BUFFERS; ++i)
if (REQ_PTL(sreq)->chunk_buffer[i])
MPIU_Free(REQ_PTL(sreq)->chunk_buffer[i]);
if (REQ_PTL(sreq)->get_me_p)
MPIU_Free(REQ_PTL(sreq)->get_me_p);
}
mpi_errno = MPID_Request_complete(sreq);
if (mpi_errno != MPI_SUCCESS) {
MPIR_ERR_POP(mpi_errno);
}
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_HANDLER_SEND);
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPID_nem_ptl_poll_finalize(void)
{
int mpi_errno = MPI_SUCCESS;
int i;
int ret;
MPIDI_STATE_DECL(MPID_STATE_MPID_NEM_PTL_POLL_FINALIZE);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_NEM_PTL_POLL_FINALIZE);
for (i = 0; i < NUM_OVERFLOW_ME; ++i) {
if (overflow_me_handle[i] != PTL_INVALID_HANDLE) {
ret = PtlMEUnlink(overflow_me_handle[i]);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlmeunlink", "**ptlmeunlink %s", MPID_nem_ptl_strerror(ret));
}
MPL_free(overflow_buf[i]);
}
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_MPID_NEM_PTL_POLL_FINALIZE);
return mpi_errno;
fn_fail:
goto fn_exit;
}
static int append_overflow(int i)
{
int mpi_errno = MPI_SUCCESS;
int ret;
ptl_me_t me;
ptl_process_t id_any;
MPIDI_STATE_DECL(MPID_STATE_APPEND_OVERFLOW);
MPIDI_FUNC_ENTER(MPID_STATE_APPEND_OVERFLOW);
MPIU_Assert(i >= 0 && i < NUM_OVERFLOW_ME);
id_any.phys.pid = PTL_PID_ANY;
id_any.phys.nid = PTL_NID_ANY;
me.start = overflow_buf[i];
me.length = OVERFLOW_LENGTH;
me.ct_handle = PTL_CT_NONE;
me.uid = PTL_UID_ANY;
me.options = ( PTL_ME_OP_PUT | PTL_ME_MANAGE_LOCAL | PTL_ME_NO_TRUNCATE | PTL_ME_MAY_ALIGN |
PTL_ME_IS_ACCESSIBLE | PTL_ME_EVENT_LINK_DISABLE );
me.match_id = id_any;
me.match_bits = 0;
me.ignore_bits = ~((ptl_match_bits_t)0);
me.min_free = PTL_LARGE_THRESHOLD;
/* if there is no space to append the entry, process outstanding events and try again */
ret = PtlMEAppend(MPIDI_nem_ptl_ni, MPIDI_nem_ptl_pt, &me, PTL_OVERFLOW_LIST, (void *)(size_t)i,
&overflow_me_handle[i]);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlmeappend", "**ptlmeappend %s", MPID_nem_ptl_strerror(ret));
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_APPEND_OVERFLOW);
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPID_nem_ptl_poll(int is_blocking_poll)
{
int mpi_errno = MPI_SUCCESS;
ptl_event_t event;
int ret;
MPIDI_STATE_DECL(MPID_STATE_MPID_NEM_PTL_POLL);
/* MPIDI_FUNC_ENTER(MPID_STATE_MPID_NEM_PTL_POLL); */
while (1) {
int ctl_event = FALSE;
/* Check the rptls EQ first. It should never return an event. */
ret = MPID_nem_ptl_rptl_eqget(MPIDI_nem_ptl_rpt_eq, &event);
MPIU_Assert(ret == PTL_EQ_EMPTY);
/* check EQs for events */
ret = MPID_nem_ptl_rptl_eqget(MPIDI_nem_ptl_eq, &event);
MPIR_ERR_CHKANDJUMP(ret == PTL_EQ_DROPPED, mpi_errno, MPI_ERR_OTHER, "**eqdropped");
if (ret == PTL_EQ_EMPTY) {
ret = MPID_nem_ptl_rptl_eqget(MPIDI_nem_ptl_get_eq, &event);
MPIR_ERR_CHKANDJUMP(ret == PTL_EQ_DROPPED, mpi_errno, MPI_ERR_OTHER, "**eqdropped");
if (ret == PTL_EQ_EMPTY) {
ret = MPID_nem_ptl_rptl_eqget(MPIDI_nem_ptl_control_eq, &event);
MPIR_ERR_CHKANDJUMP(ret == PTL_EQ_DROPPED, mpi_errno, MPI_ERR_OTHER, "**eqdropped");
if (ret == PTL_EQ_EMPTY) {
ret = MPID_nem_ptl_rptl_eqget(MPIDI_nem_ptl_origin_eq, &event);
MPIR_ERR_CHKANDJUMP(ret == PTL_EQ_DROPPED, mpi_errno, MPI_ERR_OTHER, "**eqdropped");
} else {
ctl_event = TRUE;
}
/* all EQs are empty */
if (ret == PTL_EQ_EMPTY)
break;
}
}
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptleqget", "**ptleqget %s", MPID_nem_ptl_strerror(ret));
MPL_DBG_MSG_FMT(MPIDI_CH3_DBG_CHANNEL, VERBOSE, (MPL_DBG_FDEST, "Received event %s pt_idx=%d ni_fail=%s list=%s user_ptr=%p hdr_data=%#lx mlength=%lu rlength=%lu",
MPID_nem_ptl_strevent(&event), event.pt_index, MPID_nem_ptl_strnifail(event.ni_fail_type),
MPID_nem_ptl_strlist(event.ptl_list), event.user_ptr, event.hdr_data, event.mlength, event.rlength));
MPIR_ERR_CHKANDJUMP2(event.ni_fail_type != PTL_NI_OK && event.ni_fail_type != PTL_NI_NO_MATCH, mpi_errno, MPI_ERR_OTHER, "**ptlni_fail", "**ptlni_fail %s %s", MPID_nem_ptl_strevent(&event), MPID_nem_ptl_strnifail(event.ni_fail_type));
/* special case for events on the control portal */
if (ctl_event) {
mpi_errno = MPID_nem_ptl_nm_ctl_event_handler(&event);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
continue;
}
switch (event.type) {
case PTL_EVENT_PUT:
if (event.ptl_list == PTL_OVERFLOW_LIST)
break;
case PTL_EVENT_PUT_OVERFLOW:
case PTL_EVENT_GET:
case PTL_EVENT_SEND:
case PTL_EVENT_REPLY:
case PTL_EVENT_SEARCH: {
MPID_Request * const req = event.user_ptr;
MPL_DBG_MSG_P(MPIDI_CH3_DBG_CHANNEL, VERBOSE, "req = %p", req);
MPL_DBG_MSG_P(MPIDI_CH3_DBG_CHANNEL, VERBOSE, "REQ_PTL(req)->event_handler = %p", REQ_PTL(req)->event_handler);
if (REQ_PTL(req)->event_handler) {
mpi_errno = REQ_PTL(req)->event_handler(&event);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
break;
}
case PTL_EVENT_AUTO_FREE:
mpi_errno = append_overflow((size_t)event.user_ptr);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
break;
case PTL_EVENT_AUTO_UNLINK:
overflow_me_handle[(size_t)event.user_ptr] = PTL_INVALID_HANDLE;
break;
case PTL_EVENT_LINK:
/* ignore */
break;
case PTL_EVENT_ACK:
default:
MPL_error_printf("Received unexpected event type: %d %s", event.type, MPID_nem_ptl_strevent(&event));
MPIR_ERR_INTERNALANDJUMP(mpi_errno, "Unexpected event type");
}
}
fn_exit:
/* MPIDI_FUNC_EXIT(MPID_STATE_MPID_NEM_PTL_POLL); */
return mpi_errno;
fn_fail:
goto fn_exit;
}
/*@
MPI_Type_create_darray - Create a datatype representing a distributed array
Input Parameters:
+ size - size of process group (positive integer)
. rank - rank in process group (nonnegative integer)
. ndims - number of array dimensions as well as process grid dimensions (positive integer)
. array_of_gsizes - number of elements of type oldtype in each dimension of global array (array of positive integers)
. array_of_distribs - distribution of array in each dimension (array of state)
. array_of_dargs - distribution argument in each dimension (array of positive integers)
. array_of_psizes - size of process grid in each dimension (array of positive integers)
. order - array storage order flag (state)
- oldtype - old datatype (handle)
Output Parameters:
. newtype - new datatype (handle)
.N ThreadSafe
.N Fortran
.N Errors
.N MPI_SUCCESS
.N MPI_ERR_TYPE
.N MPI_ERR_ARG
@*/
int MPI_Type_create_darray(int size,
int rank,
int ndims,
const int array_of_gsizes[],
const int array_of_distribs[],
const int array_of_dargs[],
const int array_of_psizes[],
int order, MPI_Datatype oldtype, MPI_Datatype * newtype)
{
int mpi_errno = MPI_SUCCESS, i;
MPI_Datatype new_handle;
int procs, tmp_rank, tmp_size, *coords;
MPI_Aint *st_offsets, orig_extent, disps[3];
MPI_Datatype type_old, type_new = MPI_DATATYPE_NULL, tmp_type;
#ifdef HAVE_ERROR_CHECKING
MPI_Aint size_with_aint;
MPI_Offset size_with_offset;
#endif
int *ints;
MPIR_Datatype *datatype_ptr = NULL;
MPIR_CHKLMEM_DECL(3);
MPIR_FUNC_TERSE_STATE_DECL(MPID_STATE_MPI_TYPE_CREATE_DARRAY);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPID_THREAD_CS_ENTER(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
MPIR_FUNC_TERSE_ENTER(MPID_STATE_MPI_TYPE_CREATE_DARRAY);
/* Validate parameters, especially handles needing to be converted */
#ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_DATATYPE(oldtype, "datatype", mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
#endif
/* Convert MPI object handles to object pointers */
MPIR_Datatype_get_ptr(oldtype, datatype_ptr);
MPIR_Datatype_get_extent_macro(oldtype, orig_extent);
/* Validate parameters and objects (post conversion) */
#ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
/* Check parameters */
MPIR_ERRTEST_ARGNONPOS(size, "size", mpi_errno, MPI_ERR_ARG);
/* use MPI_ERR_RANK class for PE-MPI compatibility */
MPIR_ERR_CHKANDJUMP3((rank < 0 || rank >= size), mpi_errno, MPI_ERR_RANK,
"**argrange", "**argrange %s %d %d", "rank", rank, (size - 1));
MPIR_ERRTEST_ARGNONPOS(ndims, "ndims", mpi_errno, MPI_ERR_DIMS);
MPIR_ERRTEST_ARGNULL(array_of_gsizes, "array_of_gsizes", mpi_errno);
MPIR_ERRTEST_ARGNULL(array_of_distribs, "array_of_distribs", mpi_errno);
MPIR_ERRTEST_ARGNULL(array_of_dargs, "array_of_dargs", mpi_errno);
MPIR_ERRTEST_ARGNULL(array_of_psizes, "array_of_psizes", mpi_errno);
if (order != MPI_ORDER_C && order != MPI_ORDER_FORTRAN) {
mpi_errno = MPIR_Err_create_code(MPI_SUCCESS,
MPIR_ERR_RECOVERABLE,
__func__,
__LINE__,
MPI_ERR_ARG, "**arg", "**arg %s", "order");
goto fn_fail;
}
tmp_size = 1;
for (i = 0; mpi_errno == MPI_SUCCESS && i < ndims; i++) {
MPIR_ERRTEST_ARGNONPOS(array_of_gsizes[i], "gsize", mpi_errno, MPI_ERR_ARG);
MPIR_ERRTEST_ARGNONPOS(array_of_psizes[i], "psize", mpi_errno, MPI_ERR_ARG);
if ((array_of_distribs[i] != MPI_DISTRIBUTE_NONE) &&
(array_of_distribs[i] != MPI_DISTRIBUTE_BLOCK) &&
(array_of_distribs[i] != MPI_DISTRIBUTE_CYCLIC)) {
mpi_errno = MPIR_Err_create_code(MPI_SUCCESS,
MPIR_ERR_RECOVERABLE,
__func__,
__LINE__, MPI_ERR_ARG, "**darrayunknown", 0);
goto fn_fail;
}
if ((array_of_dargs[i] != MPI_DISTRIBUTE_DFLT_DARG) && (array_of_dargs[i] <= 0)) {
mpi_errno = MPIR_Err_create_code(MPI_SUCCESS,
MPIR_ERR_RECOVERABLE,
__func__,
__LINE__,
MPI_ERR_ARG,
"**arg", "**arg %s", "array_of_dargs");
goto fn_fail;
}
if ((array_of_distribs[i] == MPI_DISTRIBUTE_NONE) && (array_of_psizes[i] != 1)) {
mpi_errno = MPIR_Err_create_code(MPI_SUCCESS,
MPIR_ERR_RECOVERABLE,
__func__,
__LINE__,
MPI_ERR_ARG,
"**darraydist",
"**darraydist %d %d", i, array_of_psizes[i]);
goto fn_fail;
}
tmp_size *= array_of_psizes[i];
}
MPIR_ERR_CHKANDJUMP1((tmp_size != size), mpi_errno, MPI_ERR_ARG,
"**arg", "**arg %s", "array_of_psizes");
/* TODO: GET THIS CHECK IN ALSO */
/* check if MPI_Aint is large enough for size of global array.
* if not, complain. */
size_with_aint = orig_extent;
for (i = 0; i < ndims; i++)
size_with_aint *= array_of_gsizes[i];
size_with_offset = orig_extent;
for (i = 0; i < ndims; i++)
size_with_offset *= array_of_gsizes[i];
if (size_with_aint != size_with_offset) {
mpi_errno = MPIR_Err_create_code(MPI_SUCCESS,
MPIR_ERR_FATAL,
__func__,
__LINE__,
MPI_ERR_ARG,
"**darrayoverflow",
"**darrayoverflow %L", size_with_offset);
goto fn_fail;
}
/* Validate datatype_ptr */
MPIR_Datatype_valid_ptr(datatype_ptr, mpi_errno);
/* If datatype_ptr is not valid, it will be reset to null */
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno)
goto fn_fail;
/* --END ERROR HANDLING-- */
}
MPID_END_ERROR_CHECKS;
}
#endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
/* calculate position in Cartesian grid as MPI would (row-major
ordering) */
MPIR_CHKLMEM_MALLOC_ORJUMP(coords, int *, ndims * sizeof(int), mpi_errno,
"position is Cartesian grid", MPL_MEM_COMM);
procs = size;
tmp_rank = rank;
for (i = 0; i < ndims; i++) {
procs = procs / array_of_psizes[i];
coords[i] = tmp_rank / procs;
tmp_rank = tmp_rank % procs;
}
MPIR_CHKLMEM_MALLOC_ORJUMP(st_offsets, MPI_Aint *, ndims * sizeof(MPI_Aint), mpi_errno,
"st_offsets", MPL_MEM_COMM);
type_old = oldtype;
if (order == MPI_ORDER_FORTRAN) {
/* dimension 0 changes fastest */
for (i = 0; i < ndims; i++) {
switch (array_of_distribs[i]) {
case MPI_DISTRIBUTE_BLOCK:
mpi_errno = MPIR_Type_block(array_of_gsizes,
i,
ndims,
array_of_psizes[i],
coords[i],
array_of_dargs[i],
order,
orig_extent, type_old, &type_new, st_offsets + i);
break;
case MPI_DISTRIBUTE_CYCLIC:
mpi_errno = MPIR_Type_cyclic(array_of_gsizes,
i,
ndims,
array_of_psizes[i],
coords[i],
array_of_dargs[i],
order,
orig_extent, type_old, &type_new, st_offsets + i);
break;
case MPI_DISTRIBUTE_NONE:
/* treat it as a block distribution on 1 process */
mpi_errno = MPIR_Type_block(array_of_gsizes,
i,
ndims,
1,
0,
MPI_DISTRIBUTE_DFLT_DARG,
order,
orig_extent, type_old, &type_new, st_offsets + i);
break;
}
if (i) {
MPIR_Type_free_impl(&type_old);
}
type_old = type_new;
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno != MPI_SUCCESS)
goto fn_fail;
/* --END ERROR HANDLING-- */
}
/* add displacement and UB */
disps[1] = st_offsets[0];
tmp_size = 1;
for (i = 1; i < ndims; i++) {
tmp_size *= array_of_gsizes[i - 1];
disps[1] += (MPI_Aint) tmp_size *st_offsets[i];
}
/* rest done below for both Fortran and C order */
}
else { /* order == MPI_ORDER_C */
/* dimension ndims-1 changes fastest */
for (i = ndims - 1; i >= 0; i--) {
switch (array_of_distribs[i]) {
case MPI_DISTRIBUTE_BLOCK:
mpi_errno = MPIR_Type_block(array_of_gsizes,
i,
ndims,
array_of_psizes[i],
coords[i],
array_of_dargs[i],
order,
orig_extent, type_old, &type_new, st_offsets + i);
break;
case MPI_DISTRIBUTE_CYCLIC:
mpi_errno = MPIR_Type_cyclic(array_of_gsizes,
i,
ndims,
array_of_psizes[i],
coords[i],
array_of_dargs[i],
order,
orig_extent, type_old, &type_new, st_offsets + i);
break;
case MPI_DISTRIBUTE_NONE:
/* treat it as a block distribution on 1 process */
mpi_errno = MPIR_Type_block(array_of_gsizes,
i,
ndims,
array_of_psizes[i],
coords[i],
MPI_DISTRIBUTE_DFLT_DARG,
order,
orig_extent, type_old, &type_new, st_offsets + i);
break;
}
if (i != ndims - 1) {
MPIR_Type_free_impl(&type_old);
}
type_old = type_new;
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno != MPI_SUCCESS)
goto fn_fail;
/* --END ERROR HANDLING-- */
}
/* add displacement and UB */
disps[1] = st_offsets[ndims - 1];
tmp_size = 1;
for (i = ndims - 2; i >= 0; i--) {
tmp_size *= array_of_gsizes[i + 1];
disps[1] += (MPI_Aint) tmp_size *st_offsets[i];
}
}
disps[1] *= orig_extent;
disps[2] = orig_extent;
for (i = 0; i < ndims; i++)
disps[2] *= (MPI_Aint) (array_of_gsizes[i]);
disps[0] = 0;
/* Instead of using MPI_LB/MPI_UB, which have been removed from MPI in MPI-3,
use MPI_Type_create_resized. Use hindexed_block to set the starting displacement
of the datatype (disps[1]) and type_create_resized to set lb to 0 (disps[0])
and extent to disps[2], which makes ub = disps[2].
*/
mpi_errno = MPIR_Type_blockindexed(1, 1, &disps[1], 1, /* 1 means disp is in bytes */
type_new, &tmp_type);
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno != MPI_SUCCESS)
goto fn_fail;
/* --END ERROR HANDLING-- */
mpi_errno = MPIR_Type_create_resized(tmp_type, 0, disps[2], &new_handle);
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno != MPI_SUCCESS)
goto fn_fail;
/* --END ERROR HANDLING-- */
MPIR_Type_free_impl(&tmp_type);
MPIR_Type_free_impl(&type_new);
/* at this point we have the new type, and we've cleaned up any
* intermediate types created in the process. we just need to save
* all our contents/envelope information.
*/
/* Save contents */
MPIR_CHKLMEM_MALLOC_ORJUMP(ints, int *, (4 * ndims + 4) * sizeof(int), mpi_errno,
"content description", MPL_MEM_BUFFER);
ints[0] = size;
ints[1] = rank;
ints[2] = ndims;
for (i = 0; i < ndims; i++) {
ints[i + 3] = array_of_gsizes[i];
}
for (i = 0; i < ndims; i++) {
ints[i + ndims + 3] = array_of_distribs[i];
}
for (i = 0; i < ndims; i++) {
ints[i + 2 * ndims + 3] = array_of_dargs[i];
}
for (i = 0; i < ndims; i++) {
ints[i + 3 * ndims + 3] = array_of_psizes[i];
}
ints[4 * ndims + 3] = order;
MPIR_Datatype_get_ptr(new_handle, datatype_ptr);
mpi_errno = MPIR_Datatype_set_contents(datatype_ptr,
MPI_COMBINER_DARRAY,
4 * ndims + 4, 0, 1, ints, NULL, &oldtype);
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno != MPI_SUCCESS)
goto fn_fail;
/* --END ERROR HANDLING-- */
MPIR_OBJ_PUBLISH_HANDLE(*newtype, new_handle);
/* ... end of body of routine ... */
fn_exit:
MPIR_CHKLMEM_FREEALL();
MPIR_FUNC_TERSE_EXIT(MPID_STATE_MPI_TYPE_CREATE_DARRAY);
MPID_THREAD_CS_EXIT(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
#ifdef HAVE_ERROR_CHECKING
{
mpi_errno =
MPIR_Err_create_code(mpi_errno, MPIR_ERR_RECOVERABLE, __func__, __LINE__, MPI_ERR_OTHER,
"**mpi_type_create_darray",
"**mpi_type_create_darray %d %d %d %p %p %p %p %d %D %p", size,
rank, ndims, array_of_gsizes, array_of_distribs, array_of_dargs,
array_of_psizes, order, oldtype, newtype);
}
#endif
mpi_errno = MPIR_Err_return_comm(NULL, __func__, mpi_errno);
goto fn_exit;
/* --END ERROR HANDLING-- */
}
static int handler_recv_dequeue_large(const ptl_event_t *e)
{
int mpi_errno = MPI_SUCCESS;
MPID_Request *const rreq = e->user_ptr;
MPIDI_VC_t *vc;
MPID_nem_ptl_vc_area *vc_ptl;
int ret;
int dt_contig;
MPIDI_msg_sz_t data_sz;
MPID_Datatype *dt_ptr;
MPI_Aint dt_true_lb;
MPI_Aint last;
MPIU_CHKPMEM_DECL(1);
MPIDI_STATE_DECL(MPID_STATE_HANDLER_RECV_DEQUEUE_LARGE);
MPIDI_FUNC_ENTER(MPID_STATE_HANDLER_RECV_DEQUEUE_LARGE);
MPIU_Assert(e->type == PTL_EVENT_PUT || e->type == PTL_EVENT_PUT_OVERFLOW);
MPIDI_Comm_get_vc(rreq->comm, NPTL_MATCH_GET_RANK(e->match_bits), &vc);
vc_ptl = VC_PTL(vc);
dequeue_req(e);
MPIDI_Datatype_get_info(rreq->dev.user_count, rreq->dev.datatype, dt_contig, data_sz, dt_ptr, dt_true_lb);
/* unpack data from unexpected buffer first */
if (e->type == PTL_EVENT_PUT_OVERFLOW) {
if (dt_contig) {
MPIU_Memcpy((char *)rreq->dev.user_buf + dt_true_lb, e->start, e->mlength);
} else {
last = e->mlength;
MPID_Segment_unpack(rreq->dev.segment_ptr, 0, &last, e->start);
MPIU_Assert(last == e->mlength);
rreq->dev.segment_first = e->mlength;
}
}
if (!(e->hdr_data & NPTL_LARGE)) {
/* all data has already been received; we're done */
mpi_errno = handler_recv_complete(e);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
goto fn_exit;
}
MPIU_Assert (e->mlength == PTL_LARGE_THRESHOLD);
/* we need to GET the rest of the data from the sender's buffer */
if (dt_contig) {
big_get((char *)rreq->dev.user_buf + dt_true_lb + PTL_LARGE_THRESHOLD, data_sz - PTL_LARGE_THRESHOLD,
vc, e->match_bits, rreq);
goto fn_exit;
}
/* noncontig recv buffer */
last = rreq->dev.segment_size;
rreq->dev.iov_count = MPL_IOV_LIMIT;
MPID_Segment_pack_vector(rreq->dev.segment_ptr, rreq->dev.segment_first, &last, rreq->dev.iov, &rreq->dev.iov_count);
if (last == rreq->dev.segment_size && rreq->dev.segment_size <= MPIDI_nem_ptl_ni_limits.max_msg_size + PTL_LARGE_THRESHOLD) {
/* Rest of message fits in one IOV */
ptl_md_t md;
md.start = rreq->dev.iov;
md.length = rreq->dev.iov_count;
md.options = PTL_IOVEC;
md.eq_handle = MPIDI_nem_ptl_origin_eq;
md.ct_handle = PTL_CT_NONE;
ret = PtlMDBind(MPIDI_nem_ptl_ni, &md, &REQ_PTL(rreq)->md);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlmdbind", "**ptlmdbind %s", MPID_nem_ptl_strerror(ret));
REQ_PTL(rreq)->event_handler = handler_recv_complete;
ret = MPID_nem_ptl_rptl_get(REQ_PTL(rreq)->md, 0, rreq->dev.segment_size - rreq->dev.segment_first, vc_ptl->id, vc_ptl->ptg,
e->match_bits, 0, rreq);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlget", "**ptlget %s", MPID_nem_ptl_strerror(ret));
goto fn_exit;
}
/* message won't fit in a single IOV, allocate buffer and unpack when received */
/* FIXME: For now, allocate a single large buffer to hold entire message */
MPIU_CHKPMEM_MALLOC(REQ_PTL(rreq)->chunk_buffer[0], void *, data_sz - PTL_LARGE_THRESHOLD,
mpi_errno, "chunk_buffer");
big_get(REQ_PTL(rreq)->chunk_buffer[0], data_sz - PTL_LARGE_THRESHOLD, vc, e->match_bits, rreq);
fn_exit:
MPIU_CHKPMEM_COMMIT();
fn_exit2:
MPIDI_FUNC_EXIT(MPID_STATE_HANDLER_RECV_DEQUEUE_LARGE);
return mpi_errno;
fn_fail:
MPIU_CHKPMEM_REAP();
goto fn_exit2;
}
int MPII_Comm_create_calculate_mapping(MPIR_Group *group_ptr,
MPIR_Comm *comm_ptr,
int **mapping_out,
MPIR_Comm **mapping_comm)
{
int mpi_errno = MPI_SUCCESS;
int subsetOfWorld = 0;
int i, j;
int n;
int *mapping=0;
MPIR_CHKPMEM_DECL(1);
MPIR_FUNC_TERSE_STATE_DECL(MPID_STATE_MPIR_COMM_CREATE_CALCULATE_MAPPING);
MPIR_FUNC_TERSE_ENTER(MPID_STATE_MPIR_COMM_CREATE_CALCULATE_MAPPING);
*mapping_out = NULL;
*mapping_comm = comm_ptr;
n = group_ptr->size;
MPIR_CHKPMEM_MALLOC(mapping,int*,n*sizeof(int),mpi_errno,"mapping",MPL_MEM_ADDRESS);
/* Make sure that the processes for this group are contained within
the input communicator. Also identify the mapping from the ranks of
the old communicator to the new communicator.
We do this by matching the lpids of the members of the group
with the lpids of the members of the input communicator.
It is an error if the group contains a reference to an lpid that
does not exist in the communicator.
An important special case is groups (and communicators) that
are subsets of MPI_COMM_WORLD. In this case, the lpids are
exactly the same as the ranks in comm world.
*/
/* we examine the group's lpids in both the intracomm and non-comm_world cases */
MPII_Group_setup_lpid_list( group_ptr );
/* Optimize for groups contained within MPI_COMM_WORLD. */
if (comm_ptr->comm_kind == MPIR_COMM_KIND__INTRACOMM) {
int wsize;
subsetOfWorld = 1;
wsize = MPIR_Process.comm_world->local_size;
for (i=0; i<n; i++) {
int g_lpid = group_ptr->lrank_to_lpid[i].lpid;
/* This mapping is relative to comm world */
MPL_DBG_MSG_FMT(MPIR_DBG_COMM,VERBOSE,
(MPL_DBG_FDEST,
"comm-create - mapping into world[%d] = %d",
i, g_lpid ));
if (g_lpid < wsize) {
mapping[i] = g_lpid;
}
else {
subsetOfWorld = 0;
break;
}
}
}
MPL_DBG_MSG_D(MPIR_DBG_COMM,VERBOSE, "subsetOfWorld=%d", subsetOfWorld );
if (subsetOfWorld) {
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
mpi_errno = MPIR_Group_check_subset( group_ptr, comm_ptr );
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
# endif
/* Override the comm to be used with the mapping array. */
*mapping_comm = MPIR_Process.comm_world;
}
else {
for (i=0; i<n; i++) {
/* mapping[i] is the rank in the communicator of the process
that is the ith element of the group */
/* FIXME : BUBBLE SORT */
mapping[i] = -1;
for (j=0; j<comm_ptr->local_size; j++) {
int comm_lpid;
MPID_Comm_get_lpid( comm_ptr, j, &comm_lpid, FALSE );
if (comm_lpid == group_ptr->lrank_to_lpid[i].lpid) {
mapping[i] = j;
break;
}
}
MPIR_ERR_CHKANDJUMP1(mapping[i] == -1,mpi_errno,MPI_ERR_GROUP,
"**groupnotincomm", "**groupnotincomm %d", i );
}
}
MPIR_Assert(mapping != NULL);
*mapping_out = mapping;
MPL_VG_CHECK_MEM_IS_DEFINED(*mapping_out, n * sizeof(**mapping_out));
MPIR_CHKPMEM_COMMIT();
fn_exit:
MPIR_FUNC_TERSE_EXIT(MPID_STATE_MPIR_COMM_CREATE_CALCULATE_MAPPING);
return mpi_errno;
fn_fail:
MPIR_CHKPMEM_REAP();
goto fn_exit;
}
int MPID_nem_ptl_recv_posted(MPIDI_VC_t *vc, MPID_Request *rreq)
{
int mpi_errno = MPI_SUCCESS;
MPID_nem_ptl_vc_area *const vc_ptl = VC_PTL(vc);
ptl_me_t me;
int dt_contig;
MPIDI_msg_sz_t data_sz;
MPID_Datatype *dt_ptr;
MPI_Aint dt_true_lb;
MPI_Aint last;
ptl_process_t id_any;
int ret;
MPIU_CHKPMEM_DECL(1);
MPIDI_STATE_DECL(MPID_STATE_MPID_NEM_PTL_RECV_POSTED);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_NEM_PTL_RECV_POSTED);
id_any.phys.nid = PTL_NID_ANY;
id_any.phys.pid = PTL_PID_ANY;
MPID_nem_ptl_init_req(rreq);
me.ct_handle = PTL_CT_NONE;
me.uid = PTL_UID_ANY;
me.options = ( PTL_ME_OP_PUT | PTL_ME_IS_ACCESSIBLE | PTL_ME_EVENT_LINK_DISABLE |
PTL_ME_EVENT_UNLINK_DISABLE | PTL_ME_USE_ONCE );
if (vc == NULL) {
/* MPI_ANY_SOURCE receive */
me.match_id = id_any;
} else {
if (!vc_ptl->id_initialized) {
mpi_errno = MPID_nem_ptl_init_id(vc);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
me.match_id = vc_ptl->id;
}
MPIU_DBG_MSG_FMT(CH3_CHANNEL, VERBOSE, (MPIU_DBG_FDEST, "tag=%#x ctx=%#x rank=%#x", rreq->dev.match.parts.tag, rreq->dev.match.parts.context_id, rreq->dev.match.parts.rank));
me.match_bits = NPTL_MATCH(rreq->dev.match.parts.tag, rreq->dev.match.parts.context_id,
rreq->dev.match.parts.rank);
if (rreq->dev.match.parts.tag == MPI_ANY_TAG)
me.ignore_bits = NPTL_MATCH_IGNORE_ANY_TAG;
else
me.ignore_bits = NPTL_MATCH_IGNORE;
me.min_free = 0;
MPIDI_Datatype_get_info(rreq->dev.user_count, rreq->dev.datatype, dt_contig, data_sz, dt_ptr, dt_true_lb);
MPIU_DBG_MSG_FMT(CH3_CHANNEL, VERBOSE, (MPIU_DBG_FDEST, "count="MPI_AINT_FMT_DEC_SPEC" datatype=%#x contig=%d data_sz=%lu", rreq->dev.user_count, rreq->dev.datatype, dt_contig, data_sz));
if (data_sz <= PTL_LARGE_THRESHOLD) {
if (dt_contig) {
/* small contig message */
void *start = (char *)rreq->dev.user_buf + dt_true_lb;
MPIU_DBG_MSG(CH3_CHANNEL, VERBOSE, "Small contig message");
if (start == NULL)
me.start = &dummy;
else
me.start = start;
me.length = data_sz;
REQ_PTL(rreq)->event_handler = handler_recv_dequeue_complete;
} else {
/* small noncontig */
MPIU_DBG_MSG(CH3_CHANNEL, VERBOSE, "Small noncontig message");
rreq->dev.segment_ptr = MPID_Segment_alloc();
MPIR_ERR_CHKANDJUMP1(rreq->dev.segment_ptr == NULL, mpi_errno, MPI_ERR_OTHER, "**nomem", "**nomem %s", "MPID_Segment_alloc");
MPID_Segment_init(rreq->dev.user_buf, rreq->dev.user_count, rreq->dev.datatype, rreq->dev.segment_ptr, 0);
rreq->dev.segment_first = 0;
rreq->dev.segment_size = data_sz;
last = rreq->dev.segment_size;
rreq->dev.iov_count = MPL_IOV_LIMIT;
MPID_Segment_pack_vector(rreq->dev.segment_ptr, rreq->dev.segment_first, &last, rreq->dev.iov, &rreq->dev.iov_count);
if (last == rreq->dev.segment_size) {
/* entire message fits in IOV */
MPIU_DBG_MSG(CH3_CHANNEL, VERBOSE, " entire message fits in IOV");
me.start = rreq->dev.iov;
me.length = rreq->dev.iov_count;
me.options |= PTL_IOVEC;
REQ_PTL(rreq)->event_handler = handler_recv_dequeue_complete;
} else {
/* IOV is not long enough to describe entire message: recv into
buffer and unpack later */
MPIU_DBG_MSG(CH3_CHANNEL, VERBOSE, " IOV too long: using bounce buffer");
MPIU_CHKPMEM_MALLOC(REQ_PTL(rreq)->chunk_buffer[0], void *, data_sz, mpi_errno, "chunk_buffer");
me.start = REQ_PTL(rreq)->chunk_buffer[0];
me.length = data_sz;
REQ_PTL(rreq)->event_handler = handler_recv_dequeue_unpack_complete;
}
}
} else {
/* Large message: Create an ME for the first chunk of data, then do a GET for the rest */
if (dt_contig) {
/*@
MPI_Win_allocate_shared - Create an MPI Window object for one-sided
communication and shared memory access, and allocate memory at each process.
This is a collective call executed by all processes in the group of comm. On
each process i, it allocates memory of at least size bytes that is shared among
all processes in comm, and returns a pointer to the locally allocated segment
in baseptr that can be used for load/store accesses on the calling process. The
locally allocated memory can be the target of load/store accesses by remote
processes; the base pointers for other processes can be queried using the
function 'MPI_Win_shared_query'.
The call also returns a window object that can be used by all processes in comm
to perform RMA operations. The size argument may be different at each process
and size = 0 is valid. It is the user''s responsibility to ensure that the
communicator comm represents a group of processes that can create a shared
memory segment that can be accessed by all processes in the group. The
allocated memory is contiguous across process ranks unless the info key
alloc_shared_noncontig is specified. Contiguous across process ranks means that
the first address in the memory segment of process i is consecutive with the
last address in the memory segment of process i − 1. This may enable the user
to calculate remote address offsets with local information only.
Input Parameters:
. size - size of window in bytes (nonnegative integer)
. disp_unit - local unit size for displacements, in bytes (positive integer)
. info - info argument (handle)
- comm - communicator (handle)
Output Parameters:
. baseptr - initial address of window (choice)
- win - window object returned by the call (handle)
.N ThreadSafe
.N Fortran
.N Errors
.N MPI_SUCCESS
.N MPI_ERR_ARG
.N MPI_ERR_COMM
.N MPI_ERR_INFO
.N MPI_ERR_OTHER
.N MPI_ERR_SIZE
.seealso: MPI_Win_allocate MPI_Win_create MPI_Win_create_dynamic MPI_Win_free MPI_Win_shared_query
@*/
int MPI_Win_allocate_shared(MPI_Aint size, int disp_unit, MPI_Info info, MPI_Comm comm,
void *baseptr, MPI_Win *win)
{
int mpi_errno = MPI_SUCCESS;
MPID_Win *win_ptr = NULL;
MPID_Comm *comm_ptr = NULL;
MPID_Info *info_ptr = NULL;
MPID_MPI_STATE_DECL(MPID_STATE_MPI_WIN_ALLOCATE_SHARED);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPID_THREAD_CS_ENTER(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
MPID_MPI_RMA_FUNC_ENTER(MPID_STATE_MPI_WIN_ALLOCATE_SHARED);
/* Validate parameters, especially handles needing to be converted */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_COMM(comm, mpi_errno);
MPIR_ERRTEST_INFO_OR_NULL(info, mpi_errno);
MPIR_ERRTEST_ARGNULL(win, "win", mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* Convert MPI object handles to object pointers */
MPID_Comm_get_ptr( comm, comm_ptr );
MPID_Info_get_ptr( info, info_ptr );
/* Validate parameters and objects (post conversion) */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
/* Validate pointers */
MPID_Comm_valid_ptr( comm_ptr, mpi_errno, FALSE );
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
MPIR_ERR_CHKANDJUMP1(disp_unit <= 0, mpi_errno, MPI_ERR_ARG,
"**arg", "**arg %s", "disp_unit must be positive");
MPIR_ERR_CHKANDJUMP1(size < 0, mpi_errno, MPI_ERR_SIZE,
"**rmasize", "**rmasize %d", size);
MPIR_ERR_CHKANDJUMP1(size > 0 && baseptr == NULL, mpi_errno, MPI_ERR_ARG,
"**nullptr", "**nullptr %s",
"NULL base pointer is invalid when size is nonzero");
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
mpi_errno = MPID_Win_allocate_shared(size, disp_unit, info_ptr, comm_ptr, baseptr, &win_ptr);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
/* Initialize a few fields that have specific defaults */
win_ptr->name[0] = 0;
win_ptr->errhandler = 0;
/* return the handle of the window object to the user */
MPIR_OBJ_PUBLISH_HANDLE(*win, win_ptr->handle);
/* ... end of body of routine ... */
fn_exit:
MPID_MPI_RMA_FUNC_EXIT(MPID_STATE_MPI_WIN_ALLOCATE_SHARED);
MPID_THREAD_CS_EXIT(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
# ifdef HAVE_ERROR_CHECKING
{
mpi_errno = MPIR_Err_create_code(
mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**mpi_win_allocate_shared",
"**mpi_win_allocate_shared %d %I %C %p %p", size, info, comm, baseptr, win);
}
# endif
mpi_errno = MPIR_Err_return_comm( comm_ptr, FCNAME, mpi_errno );
goto fn_exit;
/* --END ERROR HANDLING-- */
}
static int GetSockInterfaceAddr(int myRank, char *ifname, int maxIfname,
MPIDU_Sock_ifaddr_t *ifaddr)
{
const char *ifname_string;
int mpi_errno = MPI_SUCCESS;
int ifaddrFound = 0;
MPIU_Assert(maxIfname);
ifname[0] = '\0';
MPIR_ERR_CHKANDJUMP(MPIR_CVAR_CH3_INTERFACE_HOSTNAME && MPIR_CVAR_NEMESIS_TCP_NETWORK_IFACE, mpi_errno, MPI_ERR_OTHER, "**ifname_and_hostname");
/* Set "not found" for ifaddr */
ifaddr->len = 0;
/* Check if user specified ethernet interface name, e.g., ib0, eth1 */
if (MPIR_CVAR_NEMESIS_TCP_NETWORK_IFACE) {
int len;
mpi_errno = MPIDI_Get_IP_for_iface(MPIR_CVAR_NEMESIS_TCP_NETWORK_IFACE, ifaddr, &ifaddrFound);
MPIR_ERR_CHKANDJUMP1(mpi_errno || !ifaddrFound, mpi_errno, MPI_ERR_OTHER, "**iface_notfound", "**iface_notfound %s", MPIR_CVAR_NEMESIS_TCP_NETWORK_IFACE);
MPIU_DBG_MSG_FMT(CH3_CONNECT, VERBOSE, (MPIU_DBG_FDEST,
"ifaddrFound=TRUE ifaddr->type=%d ifaddr->len=%d ifaddr->ifaddr[0-3]=%d.%d.%d.%d",
ifaddr->type, ifaddr->len, ifaddr->ifaddr[0], ifaddr->ifaddr[1], ifaddr->ifaddr[2],
ifaddr->ifaddr[3]));
/* In this case, ifname is only used for debugging purposes */
mpi_errno = MPID_Get_processor_name(ifname, maxIfname, &len );
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
goto fn_exit;
}
/* Check for a host name supplied through an environment variable */
ifname_string = MPIR_CVAR_CH3_INTERFACE_HOSTNAME;
if (!ifname_string) {
/* See if there is a per-process name for the interfaces (e.g.,
the process manager only delievers the same values for the
environment to each process. There's no way to do this with
the param interface, so we need to use getenv() here. */
char namebuf[1024];
MPL_snprintf( namebuf, sizeof(namebuf),
"MPICH_INTERFACE_HOSTNAME_R%d", myRank );
ifname_string = getenv( namebuf );
if (DBG_IFNAME && ifname_string) {
fprintf( stdout, "Found interface name %s from %s\n",
ifname_string, namebuf );
fflush( stdout );
}
}
else if (DBG_IFNAME) {
fprintf( stdout,
"Found interface name %s from MPICH_INTERFACE_HOSTNAME\n",
ifname_string );
fflush( stdout );
}
if (!ifname_string) {
int len;
/* User did not specify a hostname. Look it up. */
mpi_errno = MPID_Get_processor_name(ifname, maxIfname, &len );
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
ifname_string = ifname;
/* If we didn't find a specific name, then try to get an IP address
directly from the available interfaces, if that is supported on
this platform. Otherwise, we'll drop into the next step that uses
the ifname */
mpi_errno = MPIDI_GetIPInterface( ifaddr, &ifaddrFound );
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
else {
/* Copy this name into the output name */
MPIU_Strncpy( ifname, ifname_string, maxIfname );
}
/* If we don't have an IP address, try to get it from the name */
if (!ifaddrFound) {
int i;
struct hostent *info = NULL;
for (i = 0; i < MPIR_CVAR_NEMESIS_TCP_HOST_LOOKUP_RETRIES; ++i) {
info = gethostbyname( ifname_string );
if (info || h_errno != TRY_AGAIN)
break;
}
MPIR_ERR_CHKANDJUMP2(!info || !info->h_addr_list, mpi_errno, MPI_ERR_OTHER, "**gethostbyname", "**gethostbyname %s %d", ifname_string, h_errno);
/* Use the primary address */
ifaddr->len = info->h_length;
ifaddr->type = info->h_addrtype;
if (ifaddr->len > sizeof(ifaddr->ifaddr)) {
/* If the address won't fit in the field, reset to
no address */
ifaddr->len = 0;
ifaddr->type = -1;
MPIR_ERR_INTERNAL(mpi_errno, "Address too long to fit in field");
} else {
MPIU_Memcpy( ifaddr->ifaddr, info->h_addr_list[0], ifaddr->len );
}
}
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPID_Send(const void * buf, MPI_Aint count, MPI_Datatype datatype, int rank,
int tag, MPID_Comm * comm, int context_offset,
MPID_Request ** request)
{
MPIDI_msg_sz_t data_sz;
int dt_contig;
MPI_Aint dt_true_lb;
MPID_Datatype * dt_ptr;
MPID_Request * sreq = NULL;
MPIDI_VC_t * vc;
#if defined(MPID_USE_SEQUENCE_NUMBERS)
MPID_Seqnum_t seqnum;
#endif
int eager_threshold = -1;
int mpi_errno = MPI_SUCCESS;
#if defined(FINEGRAIN_MPI)
int destpid=-1, destworldrank=-1;
#endif
MPIDI_STATE_DECL(MPID_STATE_MPID_SEND);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_SEND);
MPIU_DBG_MSG_FMT(CH3_OTHER,VERBOSE,(MPIU_DBG_FDEST,
"rank=%d, tag=%d, context=%d",
rank, tag, comm->context_id + context_offset));
/* Check to make sure the communicator hasn't already been revoked */
if (comm->revoked &&
MPIR_AGREE_TAG != MPIR_TAG_MASK_ERROR_BITS(tag & ~MPIR_Process.tagged_coll_mask) &&
MPIR_SHRINK_TAG != MPIR_TAG_MASK_ERROR_BITS(tag & ~MPIR_Process.tagged_coll_mask)) {
MPIR_ERR_SETANDJUMP(mpi_errno,MPIX_ERR_REVOKED,"**revoked");
}
#if defined(FINEGRAIN_MPI)
MPIDI_Comm_get_pid_worldrank(comm, rank, &destpid, &destworldrank);
if (COMPARE_RANKS(rank,comm,destpid) && comm->comm_kind != MPID_INTERCOMM)
{
mpi_errno = MPIDI_Isend_self(&buf, count, datatype, rank, tag, comm,
context_offset, MPIDI_REQUEST_TYPE_SEND,
&sreq);
if (rank == comm->rank)
{
printf("my_fgrank=%d: %s, self send DEADLOCK\n", my_fgrank, __FUNCTION__);
if (sreq != NULL && sreq->cc != 0) {
MPIR_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,
"**dev|selfsenddeadlock");
}
}
#else
if (rank == comm->rank && comm->comm_kind != MPID_INTERCOMM)
{
mpi_errno = MPIDI_Isend_self(buf, count, datatype, rank, tag, comm,
context_offset, MPIDI_REQUEST_TYPE_SEND,
&sreq);
/* In the single threaded case, sending to yourself will cause
deadlock. Note that in the runtime-thread case, this check
will not be made (long-term FIXME) */
# ifndef MPICH_IS_THREADED
{
if (sreq != NULL && MPID_cc_get(sreq->cc) != 0) {
MPIR_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,
"**dev|selfsenddeadlock");
}
}
# endif
#endif
if (mpi_errno != MPI_SUCCESS) { MPIR_ERR_POP(mpi_errno); }
goto fn_exit;
}
if (rank == MPI_PROC_NULL)
{
goto fn_exit;
}
#if defined(FINEGRAIN_MPI)
MPIDI_Comm_get_vc_set_active_direct(comm, destpid, &vc);
#else
MPIDI_Comm_get_vc_set_active(comm, rank, &vc);
#endif
MPIR_ERR_CHKANDJUMP1(vc->state == MPIDI_VC_STATE_MORIBUND, mpi_errno, MPIX_ERR_PROC_FAILED, "**comm_fail", "**comm_fail %d", rank);
#ifdef ENABLE_COMM_OVERRIDES
if (vc->comm_ops && vc->comm_ops->send)
{
mpi_errno = vc->comm_ops->send( vc, buf, count, datatype, rank, tag, comm, context_offset, &sreq);
goto fn_exit;
}
#endif
MPIDI_Datatype_get_info(count, datatype, dt_contig, data_sz, dt_ptr,
dt_true_lb);
if (data_sz == 0)
{
MPIDI_CH3_Pkt_t upkt;
MPIDI_CH3_Pkt_eager_send_t * const eager_pkt = &upkt.eager_send;
MPIU_DBG_MSG(CH3_OTHER,VERBOSE,"sending zero length message");
MPIDI_Pkt_init(eager_pkt, MPIDI_CH3_PKT_EAGER_SEND);
#if defined(FINEGRAIN_MPI)
eager_pkt->match.parts.dest_rank = destworldrank;
#endif
eager_pkt->match.parts.rank = comm->rank;
eager_pkt->match.parts.tag = tag;
eager_pkt->match.parts.context_id = comm->context_id + context_offset;
eager_pkt->sender_req_id = MPI_REQUEST_NULL;
eager_pkt->data_sz = 0;
MPIDI_VC_FAI_send_seqnum(vc, seqnum);
MPIDI_Pkt_set_seqnum(eager_pkt, seqnum);
MPID_THREAD_CS_ENTER(POBJ, vc->pobj_mutex);
mpi_errno = MPIDI_CH3_iStartMsg(vc, eager_pkt, sizeof(*eager_pkt), &sreq);
MPID_THREAD_CS_EXIT(POBJ, vc->pobj_mutex);
/* --BEGIN ERROR HANDLING-- */
if (mpi_errno != MPI_SUCCESS)
{
MPIR_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,"**ch3|eagermsg");
}
/* --END ERROR HANDLING-- */
if (sreq != NULL)
{
MPIDI_Request_set_seqnum(sreq, seqnum);
MPIDI_Request_set_type(sreq, MPIDI_REQUEST_TYPE_SEND);
/* sreq->comm = comm;
MPIR_Comm_add_ref(comm); -- not necessary for blocking functions */
}
goto fn_exit;
}
MPIDI_CH3_GET_EAGER_THRESHOLD(&eager_threshold, comm, vc);
/* FIXME: flow control: limit number of outstanding eager messages
containing data and need to be buffered by the receiver */
#ifdef USE_EAGER_SHORT
if (dt_contig && data_sz <= MPIDI_EAGER_SHORT_SIZE) {
mpi_errno = MPIDI_CH3_EagerContigShortSend( &sreq,
MPIDI_CH3_PKT_EAGERSHORT_SEND,
(char *)buf + dt_true_lb,
data_sz, rank, tag, comm,
context_offset );
}
else
#endif
if (data_sz + sizeof(MPIDI_CH3_Pkt_eager_send_t) <= eager_threshold)
{
if (dt_contig)
{
mpi_errno = MPIDI_CH3_EagerContigSend( &sreq,
MPIDI_CH3_PKT_EAGER_SEND,
(char *)buf + dt_true_lb,
data_sz, rank, tag, comm,
context_offset );
}
else
{
MPIDI_Request_create_sreq(sreq, mpi_errno, goto fn_exit);
MPIDI_Request_set_type(sreq, MPIDI_REQUEST_TYPE_SEND);
mpi_errno = MPIDI_CH3_EagerNoncontigSend( &sreq,
MPIDI_CH3_PKT_EAGER_SEND,
buf, count, datatype,
data_sz, rank, tag,
comm, context_offset );
}
}
else
{
int MPID_nem_ptl_improbe(MPIDI_VC_t *vc, int source, int tag, MPID_Comm *comm, int context_offset, int *flag,
MPID_Request **message, MPI_Status *status)
{
int mpi_errno = MPI_SUCCESS;
MPID_nem_ptl_vc_area *const vc_ptl = VC_PTL(vc);
int ret;
ptl_process_t id_any;
ptl_me_t me;
MPID_Request *req;
MPIDI_STATE_DECL(MPID_STATE_MPID_NEM_PTL_IMPROBE);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_NEM_PTL_IMPROBE);
id_any.phys.nid = PTL_NID_ANY;
id_any.phys.pid = PTL_PID_ANY;
/* create a request */
req = MPID_Request_create();
MPID_nem_ptl_init_req(req);
MPIR_ERR_CHKANDJUMP1(!req, mpi_errno, MPI_ERR_OTHER, "**nomem", "**nomem %s", "MPID_Request_create");
MPIU_Object_set_ref(req, 2); /* 1 ref for progress engine and 1 ref for us */
REQ_PTL(req)->event_handler = handle_mprobe;
req->kind = MPID_REQUEST_MPROBE;
/* create a dummy ME to use for searching the list */
me.start = NULL;
me.length = 0;
me.ct_handle = PTL_CT_NONE;
me.uid = PTL_UID_ANY;
me.options = ( PTL_ME_OP_PUT | PTL_ME_USE_ONCE );
me.min_free = 0;
me.match_bits = NPTL_MATCH(tag, comm->context_id + context_offset, source);
if (source == MPI_ANY_SOURCE)
me.match_id = id_any;
else {
if (!vc_ptl->id_initialized) {
mpi_errno = MPID_nem_ptl_init_id(vc);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
me.match_id = vc_ptl->id;
}
if (tag == MPI_ANY_TAG)
me.ignore_bits = NPTL_MATCH_IGNORE_ANY_TAG;
else
me.ignore_bits = NPTL_MATCH_IGNORE;
/* submit a search request */
ret = PtlMESearch(MPIDI_nem_ptl_ni, MPIDI_nem_ptl_pt, &me, PTL_SEARCH_DELETE, req);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlmesearch", "**ptlmesearch %s", MPID_nem_ptl_strerror(ret));
DBG_MSG_MESearch("REG", vc ? vc->pg_rank : 0, me, req);
/* wait for search request to complete */
do {
mpi_errno = MPID_nem_ptl_poll(FALSE);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
} while (!MPID_Request_is_complete(req));
*flag = REQ_PTL(req)->found;
if (*flag) {
req->comm = comm;
MPIR_Comm_add_ref(comm);
MPIR_Request_extract_status(req, status);
*message = req;
}
else {
MPID_Request_release(req);
}
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_MPID_NEM_PTL_IMPROBE);
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPID_nem_ptl_pkt_cancel_send_req_handler(MPIDI_VC_t *vc, MPIDI_CH3_Pkt_t *pkt,
MPIDI_msg_sz_t *buflen, MPID_Request **rreqp)
{
int ret, mpi_errno = MPI_SUCCESS;
MPIDI_nem_ptl_pkt_cancel_send_req_t *req_pkt = (MPIDI_nem_ptl_pkt_cancel_send_req_t *)pkt;
MPID_PKT_DECL_CAST(upkt, MPIDI_nem_ptl_pkt_cancel_send_resp_t, resp_pkt);
MPID_Request *search_req, *resp_req;
ptl_me_t me;
MPID_nem_ptl_vc_area *const vc_ptl = VC_PTL(vc);
MPIU_DBG_MSG_FMT(CH3_OTHER,VERBOSE,(MPIU_DBG_FDEST,
"received cancel send req pkt, sreq=0x%08x, rank=%d, tag=%d, context=%d",
req_pkt->sender_req_id, req_pkt->match.parts.rank,
req_pkt->match.parts.tag, req_pkt->match.parts.context_id));
/* create a dummy request and search for the message */
/* create a request */
search_req = MPID_Request_create();
MPID_nem_ptl_init_req(search_req);
MPIR_ERR_CHKANDJUMP1(!search_req, mpi_errno, MPI_ERR_OTHER, "**nomem", "**nomem %s", "MPID_Request_create");
MPIU_Object_set_ref(search_req, 2); /* 1 ref for progress engine and 1 ref for us */
search_req->kind = MPID_REQUEST_MPROBE;
/* create a dummy ME to use for searching the list */
me.start = NULL;
me.length = 0;
me.ct_handle = PTL_CT_NONE;
me.uid = PTL_UID_ANY;
me.options = ( PTL_ME_OP_PUT | PTL_ME_USE_ONCE );
me.min_free = 0;
me.match_bits = NPTL_MATCH(req_pkt->match.parts.tag, req_pkt->match.parts.context_id, req_pkt->match.parts.rank);
me.match_id = vc_ptl->id;
me.ignore_bits = NPTL_MATCH_IGNORE;
/* FIXME: this should use a custom handler that throws the data away inline */
REQ_PTL(search_req)->event_handler = handle_mprobe;
/* submit a search request */
ret = PtlMESearch(MPIDI_nem_ptl_ni, MPIDI_nem_ptl_pt, &me, PTL_SEARCH_DELETE, search_req);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlmesearch", "**ptlmesearch %s", MPID_nem_ptl_strerror(ret));
DBG_MSG_MESearch("REG", vc ? vc->pg_rank : 0, me, search_req);
/* wait for search request to complete */
do {
mpi_errno = MPID_nem_ptl_poll(FALSE);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
} while (!MPID_Request_is_complete(search_req));
/* send response */
resp_pkt->type = MPIDI_NEM_PKT_NETMOD;
resp_pkt->subtype = MPIDI_NEM_PTL_PKT_CANCEL_SEND_RESP;
resp_pkt->ack = REQ_PTL(search_req)->found;
resp_pkt->sender_req_id = req_pkt->sender_req_id;
MPID_nem_ptl_iStartContigMsg(vc, resp_pkt, sizeof(*resp_pkt), NULL,
0, &resp_req);
/* if the message was found, free the temporary buffer used to copy the data */
if (REQ_PTL(search_req)->found)
MPIU_Free(search_req->dev.tmpbuf);
MPID_Request_release(search_req);
if (resp_req != NULL)
MPID_Request_release(resp_req);
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
static int win_init(MPI_Aint size, int disp_unit, int create_flavor, int model, MPIR_Info * info,
MPIR_Comm * comm_ptr, MPIR_Win ** win_ptr)
{
int mpi_errno = MPI_SUCCESS;
int i;
MPIR_Comm *win_comm_ptr;
int win_target_pool_size;
MPIR_CHKPMEM_DECL(5);
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_WIN_INIT);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_WIN_INIT);
MPID_THREAD_CS_ENTER(POBJ, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
if (initRMAoptions) {
MPIDI_CH3_RMA_Init_sync_pvars();
MPIDI_CH3_RMA_Init_pkthandler_pvars();
initRMAoptions = 0;
}
MPID_THREAD_CS_EXIT(POBJ, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
*win_ptr = (MPIR_Win *) MPIR_Handle_obj_alloc(&MPIR_Win_mem);
MPIR_ERR_CHKANDJUMP1(!(*win_ptr), mpi_errno, MPI_ERR_OTHER, "**nomem",
"**nomem %s", "MPIR_Win_mem");
mpi_errno = MPIR_Comm_dup_impl(comm_ptr, &win_comm_ptr);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_Object_set_ref(*win_ptr, 1);
/* (*win_ptr)->errhandler is set by upper level; */
/* (*win_ptr)->base is set by caller; */
(*win_ptr)->size = size;
(*win_ptr)->disp_unit = disp_unit;
(*win_ptr)->create_flavor = create_flavor;
(*win_ptr)->model = model;
(*win_ptr)->attributes = NULL;
(*win_ptr)->comm_ptr = win_comm_ptr;
(*win_ptr)->at_completion_counter = 0;
(*win_ptr)->shm_base_addrs = NULL;
/* (*win_ptr)->basic_info_table[] is set by caller; */
(*win_ptr)->current_lock_type = MPID_LOCK_NONE;
(*win_ptr)->shared_lock_ref_cnt = 0;
(*win_ptr)->target_lock_queue_head = NULL;
(*win_ptr)->shm_allocated = FALSE;
(*win_ptr)->states.access_state = MPIDI_RMA_NONE;
(*win_ptr)->states.exposure_state = MPIDI_RMA_NONE;
(*win_ptr)->num_targets_with_pending_net_ops = 0;
(*win_ptr)->start_ranks_in_win_grp = NULL;
(*win_ptr)->start_grp_size = 0;
(*win_ptr)->lock_all_assert = 0;
(*win_ptr)->lock_epoch_count = 0;
(*win_ptr)->outstanding_locks = 0;
(*win_ptr)->current_target_lock_data_bytes = 0;
(*win_ptr)->sync_request_cnt = 0;
(*win_ptr)->active = FALSE;
(*win_ptr)->next = NULL;
(*win_ptr)->prev = NULL;
(*win_ptr)->outstanding_acks = 0;
/* Initialize the info flags */
(*win_ptr)->info_args.no_locks = 0;
(*win_ptr)->info_args.accumulate_ordering = MPIDI_ACC_ORDER_RAR | MPIDI_ACC_ORDER_RAW |
MPIDI_ACC_ORDER_WAR | MPIDI_ACC_ORDER_WAW;
(*win_ptr)->info_args.accumulate_ops = MPIDI_ACC_OPS_SAME_OP_NO_OP;
(*win_ptr)->info_args.same_size = 0;
(*win_ptr)->info_args.same_disp_unit = FALSE;
(*win_ptr)->info_args.alloc_shared_noncontig = 0;
(*win_ptr)->info_args.alloc_shm = FALSE;
if ((*win_ptr)->create_flavor == MPI_WIN_FLAVOR_ALLOCATE ||
(*win_ptr)->create_flavor == MPI_WIN_FLAVOR_SHARED) {
(*win_ptr)->info_args.alloc_shm = TRUE;
}
/* Set info_args on window based on info provided by user */
mpi_errno = MPID_Win_set_info((*win_ptr), info);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
MPIR_CHKPMEM_MALLOC((*win_ptr)->op_pool_start, MPIDI_RMA_Op_t *,
sizeof(MPIDI_RMA_Op_t) * MPIR_CVAR_CH3_RMA_OP_WIN_POOL_SIZE, mpi_errno,
"RMA op pool", MPL_MEM_RMA);
(*win_ptr)->op_pool_head = NULL;
for (i = 0; i < MPIR_CVAR_CH3_RMA_OP_WIN_POOL_SIZE; i++) {
(*win_ptr)->op_pool_start[i].pool_type = MPIDI_RMA_POOL_WIN;
DL_APPEND((*win_ptr)->op_pool_head, &((*win_ptr)->op_pool_start[i]));
}
win_target_pool_size =
MPL_MIN(MPIR_CVAR_CH3_RMA_TARGET_WIN_POOL_SIZE, MPIR_Comm_size(win_comm_ptr));
MPIR_CHKPMEM_MALLOC((*win_ptr)->target_pool_start, MPIDI_RMA_Target_t *,
sizeof(MPIDI_RMA_Target_t) * win_target_pool_size, mpi_errno,
"RMA target pool", MPL_MEM_RMA);
(*win_ptr)->target_pool_head = NULL;
for (i = 0; i < win_target_pool_size; i++) {
(*win_ptr)->target_pool_start[i].pool_type = MPIDI_RMA_POOL_WIN;
DL_APPEND((*win_ptr)->target_pool_head, &((*win_ptr)->target_pool_start[i]));
}
(*win_ptr)->num_slots = MPL_MIN(MPIR_CVAR_CH3_RMA_SLOTS_SIZE, MPIR_Comm_size(win_comm_ptr));
MPIR_CHKPMEM_MALLOC((*win_ptr)->slots, MPIDI_RMA_Slot_t *,
sizeof(MPIDI_RMA_Slot_t) * (*win_ptr)->num_slots, mpi_errno, "RMA slots",
MPL_MEM_RMA);
for (i = 0; i < (*win_ptr)->num_slots; i++) {
(*win_ptr)->slots[i].target_list_head = NULL;
}
MPIR_CHKPMEM_MALLOC((*win_ptr)->target_lock_entry_pool_start,
MPIDI_RMA_Target_lock_entry_t *,
sizeof(MPIDI_RMA_Target_lock_entry_t) *
MPIR_CVAR_CH3_RMA_TARGET_LOCK_ENTRY_WIN_POOL_SIZE, mpi_errno,
"RMA lock entry pool", MPL_MEM_RMA);
(*win_ptr)->target_lock_entry_pool_head = NULL;
for (i = 0; i < MPIR_CVAR_CH3_RMA_TARGET_LOCK_ENTRY_WIN_POOL_SIZE; i++) {
DL_APPEND((*win_ptr)->target_lock_entry_pool_head,
&((*win_ptr)->target_lock_entry_pool_start[i]));
}
if (MPIDI_RMA_Win_inactive_list_head == NULL && MPIDI_RMA_Win_active_list_head == NULL) {
/* this is the first window, register RMA progress hook */
mpi_errno = MPID_Progress_register_hook(MPIDI_CH3I_RMA_Make_progress_global,
&MPIDI_CH3I_RMA_Progress_hook_id);
if (mpi_errno) {
MPIR_ERR_POP(mpi_errno);
}
}
DL_APPEND(MPIDI_RMA_Win_inactive_list_head, (*win_ptr));
if (MPIDI_CH3U_Win_hooks.win_init != NULL) {
mpi_errno =
MPIDI_CH3U_Win_hooks.win_init(size, disp_unit, create_flavor, model, info, comm_ptr,
win_ptr);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
}
fn_exit:
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_WIN_INIT);
return mpi_errno;
fn_fail:
MPIR_CHKPMEM_REAP();
goto fn_exit;
}
int MPIDI_PG_SetConnInfo( int rank, const char *connString )
{
#ifdef USE_PMI2_API
int mpi_errno = MPI_SUCCESS;
int len;
char key[PMI2_MAX_KEYLEN];
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIDI_PG_SetConnInfo);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPIDI_PG_SetConnInfo);
len = MPL_snprintf(key, sizeof(key), "P%d-businesscard", rank);
MPIR_ERR_CHKANDJUMP1(len < 0 || len > sizeof(key), mpi_errno, MPI_ERR_OTHER, "**snprintf", "**snprintf %d", len);
mpi_errno = PMI2_KVS_Put(key, connString);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
mpi_errno = PMI2_KVS_Fence();
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
fn_exit:
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPIDI_PG_SetConnInfo);
return mpi_errno;
fn_fail:
goto fn_exit;
#else
int mpi_errno = MPI_SUCCESS;
int pmi_errno;
int len;
char key[128];
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIDI_PG_SetConnInfo);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPIDI_PG_SetConnInfo);
MPIR_Assert(pg_world->connData);
len = MPL_snprintf(key, sizeof(key), "P%d-businesscard", rank);
if (len < 0 || len > sizeof(key)) {
MPIR_ERR_SETANDJUMP1(mpi_errno,MPI_ERR_OTHER, "**snprintf",
"**snprintf %d", len);
}
pmi_errno = PMI_KVS_Put(pg_world->connData, key, connString );
if (pmi_errno != PMI_SUCCESS) {
MPIR_ERR_SETANDJUMP1(mpi_errno,MPI_ERR_OTHER, "**pmi_kvs_put",
"**pmi_kvs_put %d", pmi_errno);
}
pmi_errno = PMI_KVS_Commit(pg_world->connData);
if (pmi_errno != PMI_SUCCESS) {
MPIR_ERR_SETANDJUMP1(mpi_errno,MPI_ERR_OTHER, "**pmi_kvs_commit",
"**pmi_kvs_commit %d", pmi_errno);
}
pmi_errno = PMI_Barrier();
if (pmi_errno != PMI_SUCCESS) {
MPIR_ERR_SETANDJUMP1(mpi_errno,MPI_ERR_OTHER, "**pmi_barrier",
"**pmi_barrier %d", pmi_errno);
}
fn_exit:
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPIDI_PG_SetConnInfo);
return mpi_errno;
fn_fail:
goto fn_exit;
#endif
}
static int _mxm_process_rdtype(MPID_Request ** rreq_p, MPI_Datatype datatype,
MPID_Datatype * dt_ptr, MPIDI_msg_sz_t data_sz, const void *buf,
int count, mxm_req_buffer_t ** iov_buf, int *iov_count)
{
int mpi_errno = MPI_SUCCESS;
MPID_Request *rreq = *rreq_p;
MPIDI_msg_sz_t last;
MPL_IOV *iov;
int n_iov = 0;
int index;
if (rreq->dev.segment_ptr == NULL) {
rreq->dev.segment_ptr = MPID_Segment_alloc();
MPIR_ERR_CHKANDJUMP1((rreq->dev.segment_ptr == NULL), mpi_errno, MPI_ERR_OTHER, "**nomem",
"**nomem %s", "MPID_Segment_alloc");
}
MPID_Segment_init(buf, count, datatype, rreq->dev.segment_ptr, 0);
rreq->dev.segment_first = 0;
rreq->dev.segment_size = data_sz;
last = rreq->dev.segment_size;
MPID_Segment_count_contig_blocks(rreq->dev.segment_ptr, rreq->dev.segment_first, &last,
(MPI_Aint *) & n_iov);
MPIU_Assert(n_iov > 0);
iov = MPIU_Malloc(n_iov * sizeof(*iov));
MPIU_Assert(iov);
last = rreq->dev.segment_size;
MPID_Segment_unpack_vector(rreq->dev.segment_ptr, rreq->dev.segment_first, &last, iov, &n_iov);
MPIU_Assert(last == rreq->dev.segment_size);
#if defined(MXM_DEBUG) && (MXM_DEBUG > 0)
_dbg_mxm_output(7, "Recv Noncontiguous data vector %i entries (free slots : %i)\n", n_iov,
MXM_REQ_DATA_MAX_IOV);
for (index = 0; index < n_iov; index++) {
_dbg_mxm_output(7, "======= Recv iov[%i] = ptr : %p, len : %i \n",
index, iov[index].MPL_IOV_BUF, iov[index].MPL_IOV_LEN);
}
#endif
if (n_iov <= MXM_REQ_DATA_MAX_IOV) {
if (n_iov > MXM_MPICH_MAX_IOV) {
*iov_buf = (mxm_req_buffer_t *) MPIU_Malloc(n_iov * sizeof(**iov_buf));
MPIU_Assert(*iov_buf);
}
for (index = 0; index < n_iov; index++) {
(*iov_buf)[index].ptr = iov[index].MPL_IOV_BUF;
(*iov_buf)[index].length = iov[index].MPL_IOV_LEN;
}
rreq->dev.tmpbuf = NULL;
rreq->dev.tmpbuf_sz = 0;
*iov_count = n_iov;
}
else {
MPI_Aint packsize = 0;
MPIR_Pack_size_impl(rreq->dev.user_count, rreq->dev.datatype, &packsize);
rreq->dev.tmpbuf = MPIU_Malloc((size_t) packsize);
MPIU_Assert(rreq->dev.tmpbuf);
rreq->dev.tmpbuf_sz = packsize;
(*iov_buf)[0].ptr = rreq->dev.tmpbuf;
(*iov_buf)[0].length = (size_t) packsize;
*iov_count = 1;
}
MPIU_Free(iov);
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
