int MPIDU_Ftb_init(void)
{
int mpi_errno = MPI_SUCCESS;
int ret;
FTB_client_t ci;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIDU_FTB_INIT);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPIDU_FTB_INIT);
MPL_strncpy(ci.event_space, "ftb.mpi.mpich", sizeof(ci.event_space));
MPL_strncpy(ci.client_name, "mpich " MPICH_VERSION, sizeof(ci.client_name));
MPL_strncpy(ci.client_subscription_style, "FTB_SUBSCRIPTION_NONE", sizeof(ci.client_subscription_style));
ci.client_polling_queue_len = -1;
#ifdef USE_PMI2_API
ret = PMI2_Job_GetId(ci.client_jobid, sizeof(ci.client_jobid));
MPIR_ERR_CHKANDJUMP(ret, mpi_errno, MPI_ERR_OTHER, "**pmi_jobgetid");
#else
ret = PMI_KVS_Get_my_name(ci.client_jobid, sizeof(ci.client_jobid));
MPIR_ERR_CHKANDJUMP(ret, mpi_errno, MPI_ERR_OTHER, "**pmi_get_id");
#endif
ret = FTB_Connect(&ci, &client_handle);
MPIR_ERR_CHKANDJUMP(ret, mpi_errno, MPI_ERR_OTHER, "**ftb_connect");
ret = FTB_Declare_publishable_events(client_handle, NULL, event_info, sizeof(event_info) / sizeof(event_info[0]));
MPIR_ERR_CHKANDJUMP(ret, mpi_errno, MPI_ERR_OTHER, "**ftb_declare_publishable_events");
fn_exit:
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPIDU_FTB_INIT);
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPID_Rget_accumulate(const void *origin_addr, int origin_count,
MPI_Datatype origin_datatype, void *result_addr, int result_count,
MPI_Datatype result_datatype, int target_rank, MPI_Aint target_disp,
int target_count, MPI_Datatype target_datatype, MPI_Op op,
MPIR_Win * win_ptr, MPIR_Request ** request)
{
int mpi_errno = MPI_SUCCESS;
int dt_contig ATTRIBUTE((unused));
MPIR_Datatype*dtp;
MPI_Aint dt_true_lb ATTRIBUTE((unused));
intptr_t data_sz, trg_data_sz;
MPIR_Request *ureq;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPID_RGET_ACCUMULATE);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPID_RGET_ACCUMULATE);
/* request-based RMA operations are only valid within a passive epoch */
MPIR_ERR_CHKANDJUMP(win_ptr->states.access_state != MPIDI_RMA_PER_TARGET &&
win_ptr->states.access_state != MPIDI_RMA_LOCK_ALL_CALLED &&
win_ptr->states.access_state != MPIDI_RMA_LOCK_ALL_ISSUED &&
win_ptr->states.access_state != MPIDI_RMA_LOCK_ALL_GRANTED,
mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
/* Create user request, initially cc=1, ref=1 */
ureq = MPIR_Request_create(MPIR_REQUEST_KIND__RMA);
MPIR_ERR_CHKANDJUMP(ureq == NULL, mpi_errno, MPI_ERR_OTHER, "**nomemreq");
/* This request is referenced by user and ch3 by default. */
MPIR_Object_set_ref(ureq, 2);
/* Note that GACC is only a no-op if no data goes in both directions */
MPIDI_Datatype_get_info(origin_count, origin_datatype, dt_contig, data_sz, dtp, dt_true_lb);
MPIDI_Datatype_get_info(origin_count, origin_datatype, dt_contig, trg_data_sz, dtp, dt_true_lb);
/* Enqueue or perform the RMA operation */
if (target_rank != MPI_PROC_NULL && (data_sz != 0 || trg_data_sz != 0)) {
mpi_errno = MPIDI_CH3I_Get_accumulate(origin_addr, origin_count,
origin_datatype, result_addr,
result_count, result_datatype,
target_rank, target_disp,
target_count, target_datatype, op, win_ptr, ureq);
if (mpi_errno != MPI_SUCCESS) {
MPIR_ERR_POP(mpi_errno);
}
}
else {
mpi_errno = MPID_Request_complete(ureq);
if (mpi_errno != MPI_SUCCESS) {
MPIR_ERR_POP(mpi_errno);
}
}
*request = ureq;
fn_exit:
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPID_RGET_ACCUMULATE);
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;
}
int MPIDI_check_for_failed_procs(void)
{
int mpi_errno = MPI_SUCCESS;
int pmi_errno;
int len;
char *kvsname = MPIDI_global.jobid;
char *failed_procs_string = NULL;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIDI_CHECK_FOR_FAILED_PROCS);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPIDI_CHECK_FOR_FAILED_PROCS);
/* FIXME: Currently this only handles failed processes in
* comm_world. We need to fix hydra to include the pgid along
* with the rank, then we need to create the failed group from
* something bigger than comm_world. */
#ifdef USE_PMIX_API
MPIR_Assert(0);
#elif defined(USE_PMI2_API)
{
int vallen = 0;
len = PMI2_MAX_VALLEN;
failed_procs_string = MPL_malloc(len, MPL_MEM_OTHER);
MPIR_Assert(failed_procs_string);
pmi_errno =
PMI2_KVS_Get(kvsname, PMI2_ID_NULL, "PMI_dead_processes", failed_procs_string,
len, &vallen);
MPIR_ERR_CHKANDJUMP(pmi_errno, mpi_errno, MPI_ERR_OTHER, "**pmi_kvs_get");
MPL_free(failed_procs_string);
}
#else
pmi_errno = PMI_KVS_Get_value_length_max(&len);
MPIR_ERR_CHKANDJUMP(pmi_errno, mpi_errno, MPI_ERR_OTHER, "**pmi_kvs_get_value_length_max");
failed_procs_string = MPL_malloc(len, MPL_MEM_OTHER);
MPIR_Assert(failed_procs_string);
pmi_errno = PMI_KVS_Get(kvsname, "PMI_dead_processes", failed_procs_string, len);
MPIR_ERR_CHKANDJUMP(pmi_errno, mpi_errno, MPI_ERR_OTHER, "**pmi_kvs_get");
MPL_free(failed_procs_string);
#endif
MPL_DBG_MSG_FMT(MPIDI_CH4_DBG_GENERAL, VERBOSE,
(MPL_DBG_FDEST, "Received proc fail notification: %s", failed_procs_string));
/* FIXME: handle ULFM failed groups here */
fn_exit:
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPIDI_CHECK_FOR_FAILED_PROCS);
return mpi_errno;
fn_fail:
MPL_free(failed_procs_string);
goto fn_exit;
}
int MPIR_Dist_graph_neighbors_impl(MPIR_Comm * comm_ptr,
int maxindegree, int sources[], int sourceweights[],
int maxoutdegree, int destinations[], int destweights[])
{
int mpi_errno = MPI_SUCCESS;
MPIR_Topology *topo_ptr = NULL;
topo_ptr = MPIR_Topology_get(comm_ptr);
MPIR_ERR_CHKANDJUMP(!topo_ptr ||
topo_ptr->kind != MPI_DIST_GRAPH, mpi_errno, MPI_ERR_TOPOLOGY,
"**notdistgraphtopo");
MPIR_Memcpy(sources, topo_ptr->topo.dist_graph.in, maxindegree * sizeof(int));
MPIR_Memcpy(destinations, topo_ptr->topo.dist_graph.out, maxoutdegree * sizeof(int));
if (sourceweights != MPI_UNWEIGHTED && topo_ptr->topo.dist_graph.is_weighted) {
MPIR_Memcpy(sourceweights, topo_ptr->topo.dist_graph.in_weights, maxindegree * sizeof(int));
}
if (destweights != MPI_UNWEIGHTED && topo_ptr->topo.dist_graph.is_weighted) {
MPIR_Memcpy(destweights, topo_ptr->topo.dist_graph.out_weights, maxoutdegree * sizeof(int));
}
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Graph_neighbors_impl(MPIR_Comm * comm_ptr, int rank, int maxneighbors, int neighbors[])
{
int mpi_errno = MPI_SUCCESS;
MPIR_Topology *graph_ptr;
int i, is, ie;
graph_ptr = MPIR_Topology_get(comm_ptr);
MPIR_ERR_CHKANDJUMP((!graph_ptr ||
graph_ptr->kind != MPI_GRAPH), mpi_errno, MPI_ERR_TOPOLOGY,
"**notgraphtopo");
MPIR_ERR_CHKANDJUMP2((rank < 0 ||
rank >= graph_ptr->topo.graph.nnodes), mpi_errno, MPI_ERR_RANK, "**rank",
"**rank %d %d", rank, graph_ptr->topo.graph.nnodes);
/* Get location in edges array of the neighbors of the specified rank */
if (rank == 0)
is = 0;
else
is = graph_ptr->topo.graph.index[rank - 1];
ie = graph_ptr->topo.graph.index[rank];
/* Get neighbors */
for (i = is; i < ie; i++)
*neighbors++ = graph_ptr->topo.graph.edges[i];
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Comm_create(MPID_Comm ** newcomm_ptr)
{
int mpi_errno = MPI_SUCCESS;
MPID_Comm *newptr;
MPID_MPI_STATE_DECL(MPID_STATE_MPIR_COMM_CREATE);
MPID_MPI_FUNC_ENTER(MPID_STATE_MPIR_COMM_CREATE);
newptr = (MPID_Comm *) MPIU_Handle_obj_alloc(&MPID_Comm_mem);
MPIR_ERR_CHKANDJUMP(!newptr, mpi_errno, MPI_ERR_OTHER, "**nomem");
*newcomm_ptr = newptr;
mpi_errno = MPIR_Comm_init(newptr);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
/* Insert this new communicator into the list of known communicators.
* Make this conditional on debugger support to match the test in
* MPIR_Comm_release . */
MPIR_COMML_REMEMBER(newptr);
fn_fail:
MPID_MPI_FUNC_EXIT(MPID_STATE_MPIR_COMM_CREATE);
return mpi_errno;
}
static int do_readv(MPID_Request *rreq, int pipe_fd, MPL_IOV iov[],
int *iov_offset, int *iov_count, int *complete)
{
int mpi_errno = MPI_SUCCESS;
ssize_t nread;
nread = readv(pipe_fd, &rreq->dev.iov[rreq->dev.iov_offset], rreq->dev.iov_count);
MPIR_ERR_CHKANDJUMP2(nread < 0 && errno != EAGAIN, mpi_errno, MPI_ERR_OTHER, "**read",
"**readv %d %s", errno, MPIU_Strerror(errno));
if (nread < 0) {
if (errno == EAGAIN) goto fn_exit;
MPIR_ERR_CHKANDJUMP2(errno != EAGAIN, mpi_errno, MPI_ERR_OTHER, "**vmsplice",
"**vmsplice %d %s", errno, MPIU_Strerror(errno));
}
*complete = adjust_partially_xferred_iov(iov, iov_offset, iov_count, nread);
if (*complete) {
/* look for additional data to send and reload IOV if there is more */
mpi_errno = check_req_complete(rreq->ch.vc, rreq, complete);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
if (*complete) {
nread = close(pipe_fd);
MPIR_ERR_CHKANDJUMP(nread < 0, mpi_errno, MPI_ERR_OTHER, "**close");
MPL_DBG_MSG(MPIDI_CH3_DBG_CHANNEL, VERBOSE, ".... complete");
}
}
fn_fail:
fn_exit:
return mpi_errno;
}
static int do_vmsplice(MPID_Request *sreq, int pipe_fd, MPL_IOV iov[],
int *iov_offset, int *iov_count, int *complete)
{
int mpi_errno = MPI_SUCCESS;
ssize_t err;
#if 1
err = vmsplice(pipe_fd, &iov[*iov_offset], *iov_count, SPLICE_F_NONBLOCK);
#else
err = writev(pipe_fd, &iov[*iov_offset], *iov_count);
#endif
if (err < 0) {
if (errno == EAGAIN) goto fn_exit;
MPIR_ERR_CHKANDJUMP2(errno != EAGAIN, mpi_errno, MPI_ERR_OTHER, "**vmsplice",
"**vmsplice %d %s", errno, MPIU_Strerror(errno));
}
*complete = adjust_partially_xferred_iov(iov, iov_offset, iov_count, err);
if (*complete) {
/* look for additional data to send and reload IOV if there is more */
mpi_errno = check_req_complete(sreq->ch.vc, sreq, complete);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
if (*complete) {
err = close(pipe_fd);
MPIR_ERR_CHKANDJUMP(err < 0, mpi_errno, MPI_ERR_OTHER, "**close");
MPL_DBG_MSG(MPIDI_CH3_DBG_CHANNEL, VERBOSE, ".... complete");
}
}
fn_fail:
fn_exit:
return mpi_errno;
}
/*@
MPIDI_CH3U_VC_SendClose - Initiate a close on a virtual connection
Input Parameters:
+ vc - Virtual connection to close
- i - rank of virtual connection within a process group (used for debugging)
Notes:
The current state of this connection must be either 'MPIDI_VC_STATE_ACTIVE'
or 'MPIDI_VC_STATE_REMOTE_CLOSE'.
@*/
int MPIDI_CH3U_VC_SendClose( MPIDI_VC_t *vc, int rank )
{
MPIDI_CH3_Pkt_t upkt;
MPIDI_CH3_Pkt_close_t * close_pkt = &upkt.close;
MPIR_Request * sreq;
int mpi_errno = MPI_SUCCESS;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIDI_CH3U_VC_SENDCLOSE);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPIDI_CH3U_VC_SENDCLOSE);
MPID_THREAD_CS_ENTER(POBJ, vc->pobj_mutex);
MPIR_Assert( vc->state == MPIDI_VC_STATE_ACTIVE ||
vc->state == MPIDI_VC_STATE_REMOTE_CLOSE );
MPIDI_Pkt_init(close_pkt, MPIDI_CH3_PKT_CLOSE);
close_pkt->ack = (vc->state == MPIDI_VC_STATE_ACTIVE) ? FALSE : TRUE;
/* MT: this is not thread safe, the POBJ CS is scoped to the vc and
* doesn't protect this global correctly */
MPIDI_Outstanding_close_ops += 1;
MPL_DBG_MSG_FMT(MPIDI_CH3_DBG_DISCONNECT,TYPICAL,(MPL_DBG_FDEST,
"sending close(%s) on vc (pg=%p) %p to rank %d, ops = %d",
close_pkt->ack ? "TRUE" : "FALSE", vc->pg, vc,
rank, MPIDI_Outstanding_close_ops));
/*
* A close packet acknowledging this close request could be
* received during iStartMsg, therefore the state must
* be changed before the close packet is sent.
*/
if (vc->state == MPIDI_VC_STATE_ACTIVE) {
MPIDI_CHANGE_VC_STATE(vc, LOCAL_CLOSE);
}
else {
MPIR_Assert( vc->state == MPIDI_VC_STATE_REMOTE_CLOSE );
MPIDI_CHANGE_VC_STATE(vc, CLOSE_ACKED);
}
mpi_errno = MPIDI_CH3_iStartMsg(vc, close_pkt, sizeof(*close_pkt), &sreq);
MPIR_ERR_CHKANDJUMP(mpi_errno, mpi_errno, MPI_ERR_OTHER, "**ch3|send_close_ack");
if (sreq != NULL) {
/* There is still another reference being held by the channel. It
will not be released until the pkt is actually sent. */
MPIR_Request_free(sreq);
}
fn_exit:
MPID_THREAD_CS_EXIT(POBJ, vc->pobj_mutex);
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPIDI_CH3U_VC_SENDCLOSE);
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Get_intercomm_contextid(MPIR_Comm * comm_ptr, MPIR_Context_id_t * context_id,
MPIR_Context_id_t * recvcontext_id)
{
MPIR_Context_id_t mycontext_id, remote_context_id;
int mpi_errno = MPI_SUCCESS;
int tag = 31567; /* FIXME - we need an internal tag or
* communication channel. Can we use a different
* context instead?. Or can we use the tag
* provided in the intercomm routine? (not on a dup,
* but in that case it can use the collective context) */
MPIR_Errflag_t errflag = MPIR_ERR_NONE;
MPIR_FUNC_TERSE_STATE_DECL(MPID_STATE_MPIR_GET_INTERCOMM_CONTEXTID);
MPIR_FUNC_TERSE_ENTER(MPID_STATE_MPIR_GET_INTERCOMM_CONTEXTID);
if (!comm_ptr->local_comm) {
/* Manufacture the local communicator */
mpi_errno = MPII_Setup_intercomm_localcomm(comm_ptr);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
}
mpi_errno = MPIR_Get_contextid_sparse(comm_ptr->local_comm, &mycontext_id, FALSE);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_Assert(mycontext_id != 0);
/* MPIC routine uses an internal context id. The local leads (process 0)
* exchange data */
remote_context_id = -1;
if (comm_ptr->rank == 0) {
mpi_errno = MPIC_Sendrecv(&mycontext_id, 1, MPIR_CONTEXT_ID_T_DATATYPE, 0, tag,
&remote_context_id, 1, MPIR_CONTEXT_ID_T_DATATYPE, 0, tag,
comm_ptr, MPI_STATUS_IGNORE, &errflag);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
}
/* Make sure that all of the local processes now have this
* id */
mpi_errno = MPID_Bcast(&remote_context_id, 1, MPIR_CONTEXT_ID_T_DATATYPE,
0, comm_ptr->local_comm, &errflag);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIR_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail");
/* The recvcontext_id must be the one that was allocated out of the local
* group, not the remote group. Otherwise we could end up posting two
* MPI_ANY_SOURCE,MPI_ANY_TAG recvs on the same context IDs even though we
* are attempting to post them for two separate communicators. */
*context_id = remote_context_id;
*recvcontext_id = mycontext_id;
fn_fail:
MPIR_FUNC_TERSE_EXIT(MPID_STATE_MPIR_GET_INTERCOMM_CONTEXTID);
return mpi_errno;
}
int MPID_nem_tcp_get_addr_port_from_bc(const char *business_card, struct in_addr *addr, in_port_t *port)
{
int mpi_errno = MPI_SUCCESS;
int ret;
int port_int;
/*char desc_str[256];*/
char ifname[256];
MPIDI_STATE_DECL(MPID_STATE_MPID_NEM_TCP_GET_ADDR_PORT_FROM_BC);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_NEM_TCP_GET_ADDR_PORT_FROM_BC);
/* fprintf(stdout, FCNAME " Enter\n"); fflush(stdout); */
/* desc_str is only used for debugging
ret = MPIU_Str_get_string_arg (business_card, MPIDI_CH3I_HOST_DESCRIPTION_KEY, desc_str, sizeof(desc_str));
MPIR_ERR_CHKANDJUMP (ret != MPIU_STR_SUCCESS, mpi_errno, MPI_ERR_OTHER, "**argstr_missinghost");
*/
/* sizeof(in_port_t) != sizeof(int) on most platforms, so we need to use
* port_int as the arg to MPIU_Str_get_int_arg. */
ret = MPIU_Str_get_int_arg (business_card, MPIDI_CH3I_PORT_KEY, &port_int);
/* MPIU_STR_FAIL is not a valid MPI error code so we store the result in ret
* instead of mpi_errno. */
MPIR_ERR_CHKANDJUMP (ret != MPIU_STR_SUCCESS, mpi_errno, MPI_ERR_OTHER, "**argstr_missingport");
MPIU_Assert((port_int >> (8*sizeof(*port))) == 0); /* ensure port_int isn't too large for *port */
*port = htons((in_port_t)port_int);
ret = MPIU_Str_get_string_arg(business_card, MPIDI_CH3I_IFNAME_KEY, ifname, sizeof(ifname));
MPIR_ERR_CHKANDJUMP (ret != MPIU_STR_SUCCESS, mpi_errno, MPI_ERR_OTHER, "**argstr_missingifname");
ret = inet_pton (AF_INET, (const char *)ifname, addr);
MPIR_ERR_CHKANDJUMP(ret == 0, mpi_errno,MPI_ERR_OTHER,"**ifnameinvalid");
MPIR_ERR_CHKANDJUMP(ret < 0, mpi_errno, MPI_ERR_OTHER, "**afinetinvalid");
fn_exit:
/* fprintf(stdout, FCNAME " Exit\n"); fflush(stdout); */
MPIDI_FUNC_EXIT(MPID_STATE_MPID_NEM_TCP_GET_ADDR_PORT_FROM_BC);
return mpi_errno;
fn_fail:
/* fprintf(stdout, "failure. mpi_errno = %d\n", mpi_errno); */
MPIU_DBG_MSG_FMT(NEM_SOCK_DET, VERBOSE, (MPIU_DBG_FDEST, "failure. mpi_errno = %d", mpi_errno));
goto fn_exit;
}
/*@
MPI_Add_error_class - Add an MPI error class to the known classes
Output Parameters:
. errorclass - New error class
.N ThreadSafe
.N Fortran
.N Errors
.N MPI_SUCCESS
.N MPI_ERR_OTHER
@*/
int MPI_Add_error_class(int *errorclass)
{
int mpi_errno = MPI_SUCCESS;
int new_class;
MPIR_FUNC_TERSE_STATE_DECL(MPID_STATE_MPI_ADD_ERROR_CLASS);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPID_THREAD_CS_ENTER(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
MPIR_FUNC_TERSE_ENTER(MPID_STATE_MPI_ADD_ERROR_CLASS);
/* Validate parameters, especially handles needing to be converted */
#ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_ARGNULL(errorclass, "errorclass", mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
#endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
new_class = MPIR_Err_add_class();
MPIR_ERR_CHKANDJUMP(new_class < 0, mpi_errno, MPI_ERR_OTHER, "**noerrclasses");
*errorclass = ERROR_DYN_MASK | new_class;
/* FIXME why isn't this done in MPIR_Err_add_class? */
if (*errorclass > MPIR_Process.attrs.lastusedcode) {
MPIR_Process.attrs.lastusedcode = *errorclass;
}
/* ... end of body of routine ... */
fn_exit:
MPIR_FUNC_TERSE_EXIT(MPID_STATE_MPI_ADD_ERROR_CLASS);
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_add_error_class", "**mpi_add_error_class %p", errorclass);
}
#endif
mpi_errno = MPIR_Err_return_comm(NULL, FCNAME, mpi_errno);
goto fn_exit;
/* --END ERROR HANDLING-- */
}
int MPIR_Setup_intercomm_localcomm(MPID_Comm * intercomm_ptr)
{
MPID_Comm *localcomm_ptr;
int mpi_errno = MPI_SUCCESS;
MPID_MPI_STATE_DECL(MPID_STATE_MPIR_SETUP_INTERCOMM_LOCALCOMM);
MPID_MPI_FUNC_ENTER(MPID_STATE_MPIR_SETUP_INTERCOMM_LOCALCOMM);
localcomm_ptr = (MPID_Comm *) MPIU_Handle_obj_alloc(&MPID_Comm_mem);
MPIR_ERR_CHKANDJUMP(!localcomm_ptr, mpi_errno, MPI_ERR_OTHER, "**nomem");
/* get sensible default values for most fields (usually zeros) */
mpi_errno = MPIR_Comm_init(localcomm_ptr);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
/* use the parent intercomm's recv ctx as the basis for our ctx */
localcomm_ptr->recvcontext_id =
MPID_CONTEXT_SET_FIELD(IS_LOCALCOMM, intercomm_ptr->recvcontext_id, 1);
localcomm_ptr->context_id = localcomm_ptr->recvcontext_id;
MPIU_DBG_MSG_FMT(COMM, TYPICAL,
(MPIU_DBG_FDEST,
"setup_intercomm_localcomm ic=%p ic->context_id=%d ic->recvcontext_id=%d lc->recvcontext_id=%d",
intercomm_ptr, intercomm_ptr->context_id, intercomm_ptr->recvcontext_id,
localcomm_ptr->recvcontext_id));
/* Save the kind of the communicator */
localcomm_ptr->comm_kind = MPID_INTRACOMM;
/* Set the sizes and ranks */
localcomm_ptr->remote_size = intercomm_ptr->local_size;
localcomm_ptr->local_size = intercomm_ptr->local_size;
localcomm_ptr->rank = intercomm_ptr->rank;
MPIR_Comm_map_dup(localcomm_ptr, intercomm_ptr, MPIR_COMM_MAP_DIR_L2L);
/* TODO More advanced version: if the group is available, dup it by
* increasing the reference count instead of recreating it later */
/* FIXME : No coll_fns functions for the collectives */
/* FIXME : No local functions for the topology routines */
intercomm_ptr->local_comm = localcomm_ptr;
/* sets up the SMP-aware sub-communicators and tables */
mpi_errno = MPIR_Comm_commit(localcomm_ptr);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
fn_fail:
MPID_MPI_FUNC_EXIT(MPID_STATE_MPIR_SETUP_INTERCOMM_LOCALCOMM);
return mpi_errno;
}
/*
* MPIDI_Open_port()
*/
static int MPIDI_Open_port(MPIR_Info *info_ptr, char *port_name)
{
int mpi_errno = MPI_SUCCESS;
int str_errno = MPL_STR_SUCCESS;
int len;
int port_name_tag = 0; /* this tag is added to the business card,
which is then returned as the port name */
int myRank = MPIR_Process.comm_world->rank;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIDI_OPEN_PORT);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPIDI_OPEN_PORT);
mpi_errno = get_port_name_tag(&port_name_tag);
MPIR_ERR_CHKANDJUMP(mpi_errno,mpi_errno,MPI_ERR_OTHER,"**argstr_port_name_tag");
len = MPI_MAX_PORT_NAME;
str_errno = MPL_str_add_int_arg(&port_name, &len,
MPIDI_CH3I_PORT_NAME_TAG_KEY, port_name_tag);
MPIR_ERR_CHKANDJUMP(str_errno, mpi_errno, MPI_ERR_OTHER, "**argstr_port_name_tag");
/* This works because Get_business_card uses the same MPL_str_xxx
functions as above to add the business card to the input string */
/* FIXME: We should instead ask the mpid_pg routines to give us
a connection string. There may need to be a separate step to
restrict us to a connection information that is only valid for
connections between processes that are started separately (e.g.,
may not use shared memory). We may need a channel-specific
function to create an exportable connection string. */
mpi_errno = MPIDI_CH3_Get_business_card(myRank, port_name, len);
MPL_DBG_MSG_FMT(MPIDI_CH3_DBG_OTHER, VERBOSE, (MPL_DBG_FDEST, "port_name = %s", port_name));
mpi_errno = MPIDI_CH3I_Port_init(port_name_tag);
fn_exit:
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPIDI_OPEN_PORT);
return mpi_errno;
fn_fail:
goto fn_exit;
}
/*@
MPI_Add_error_code - Add an MPI error code to an MPI error class
Input Parameters:
. errorclass - Error class to add an error code.
Output Parameters:
. errorcode - New error code for this error class.
.N ThreadSafe
.N Fortran
.N Errors
.N MPI_SUCCESS
.N MPI_ERR_OTHER
@*/
int MPI_Add_error_code(int errorclass, int *errorcode)
{
int mpi_errno = MPI_SUCCESS;
int new_code;
MPIR_FUNC_TERSE_STATE_DECL(MPID_STATE_MPI_ADD_ERROR_CODE);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPID_THREAD_CS_ENTER(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
MPIR_FUNC_TERSE_ENTER(MPID_STATE_MPI_ADD_ERROR_CODE);
/* Validate parameters, especially handles needing to be converted */
#ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
/* FIXME: verify that errorclass is a dynamic class */
MPIR_ERRTEST_ARGNULL(errorcode, "errorcode", mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
#endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
new_code = MPIR_Err_add_code(errorclass);
MPIR_ERR_CHKANDJUMP(new_code < 0, mpi_errno, MPI_ERR_OTHER, "**noerrcodes");
*errorcode = new_code;
/* ... end of body of routine ... */
fn_exit:
MPIR_FUNC_TERSE_EXIT(MPID_STATE_MPI_ADD_ERROR_CODE);
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_add_error_code", "**mpi_add_error_code %d %p", errorclass,
errorcode);
}
#endif
mpi_errno = MPIR_Err_return_comm(NULL, __func__, mpi_errno);
goto fn_exit;
/* --END ERROR HANDLING-- */
}
/*
* The connect and accept routines use this routine to get the port tag
* from the port name.
*/
int MPIDI_GetTagFromPort( const char *port_name, int *port_name_tag )
{
int mpi_errno = MPI_SUCCESS;
int str_errno = MPIU_STR_SUCCESS;
str_errno = MPIU_Str_get_int_arg(port_name, MPIDI_CH3I_PORT_NAME_TAG_KEY,
port_name_tag);
MPIR_ERR_CHKANDJUMP(str_errno, mpi_errno, MPI_ERR_OTHER, "**argstr_no_port_name_tag");
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPID_nem_llc_finalize(void)
{
int mpi_errno = MPI_SUCCESS;
int rc;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPID_NEM_LLC_FINALIZE);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPID_NEM_LLC_FINALIZE);
rc = LLC_finalize();
MPIR_ERR_CHKANDJUMP(rc != 0, mpi_errno, MPI_ERR_OTHER, "**fail");
fn_exit:
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPID_NEM_LLC_FINALIZE);
return mpi_errno;
fn_fail:
goto fn_exit;
}
static int MPIDI_Close_port(const char *port_name)
{
int mpi_errno = MPI_SUCCESS;
int port_name_tag;
MPIDI_STATE_DECL(MPID_STATE_MPIDI_CLOSE_PORT);
MPIDI_FUNC_ENTER(MPID_STATE_MPIDI_CLOSE_PORT);
mpi_errno = MPIDI_GetTagFromPort(port_name, &port_name_tag);
MPIR_ERR_CHKANDJUMP(mpi_errno, mpi_errno, MPI_ERR_OTHER,"**argstr_port_name_tag");
free_port_name_tag(port_name_tag);
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_MPIDI_CLOSE_PORT);
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPID_nem_tcp_bind (int sockfd)
{
int mpi_errno = MPI_SUCCESS;
int ret;
struct sockaddr_in sin;
int port;
MPIDI_STATE_DECL(MPID_STATE_MPID_NEM_TCP_BIND);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_NEM_TCP_BIND);
MPIR_ERR_CHKANDJUMP(MPIR_CVAR_CH3_PORT_RANGE.low < 0 || MPIR_CVAR_CH3_PORT_RANGE.low > MPIR_CVAR_CH3_PORT_RANGE.high, mpi_errno, MPI_ERR_OTHER, "**badportrange");
/* default MPICH_PORT_RANGE is {0,0} so bind will use any available port */
ret = 0;
for (port = MPIR_CVAR_CH3_PORT_RANGE.low; port <= MPIR_CVAR_CH3_PORT_RANGE.high; ++port)
{
memset ((void *)&sin, 0, sizeof(sin));
sin.sin_family = AF_INET;
sin.sin_addr.s_addr = htonl(INADDR_ANY);
sin.sin_port = htons(port);
ret = bind (sockfd, (struct sockaddr *)&sin, sizeof(sin));
if (ret == 0)
break;
/* check for real error */
MPIR_ERR_CHKANDJUMP3 (errno != EADDRINUSE && errno != EADDRNOTAVAIL, mpi_errno, MPI_ERR_OTHER, "**sock|poll|bind", "**sock|poll|bind %d %d %s", port, errno, MPIU_Strerror (errno));
}
/* check if an available port was found */
MPIR_ERR_CHKANDJUMP3 (ret == -1, mpi_errno, MPI_ERR_OTHER, "**sock|poll|bind", "**sock|poll|bind %d %d %s", port-1, errno, MPIU_Strerror (errno));
fn_exit:
/* if (ret == 0) */
/* fprintf(stdout, "sockfd=%d port=%d bound\n", sockfd, port); */
/* fprintf(stdout, FCNAME " Exit\n"); fflush(stdout); */
MPIDI_FUNC_EXIT(MPID_STATE_MPID_NEM_TCP_BIND);
return mpi_errno;
fn_fail:
/* fprintf(stdout, "failure. mpi_errno = %d\n", mpi_errno); */
MPIU_DBG_MSG_FMT(NEM_SOCK_DET, VERBOSE, (MPIU_DBG_FDEST, "failure. mpi_errno = %d", mpi_errno));
goto fn_exit;
}
int MPID_nem_ptl_get_id_from_bc(const char *business_card, ptl_process_t *id, ptl_pt_index_t *pt, ptl_pt_index_t *ptg, ptl_pt_index_t *ptc, ptl_pt_index_t *ptr, ptl_pt_index_t *ptrg, ptl_pt_index_t *ptrc)
{
int mpi_errno = MPI_SUCCESS;
int ret;
int len;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPID_NEM_PTL_GET_ID_FROM_BC);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPID_NEM_PTL_GET_ID_FROM_BC);
ret = MPL_str_get_binary_arg(business_card, NID_KEY, (char *)&id->phys.nid, sizeof(id->phys.nid), &len);
MPIR_ERR_CHKANDJUMP(ret != MPL_STR_SUCCESS || len != sizeof(id->phys.nid), mpi_errno, MPI_ERR_OTHER, "**badbusinesscard");
ret = MPL_str_get_binary_arg(business_card, PID_KEY, (char *)&id->phys.pid, sizeof(id->phys.pid), &len);
MPIR_ERR_CHKANDJUMP(ret != MPL_STR_SUCCESS || len != sizeof(id->phys.pid), mpi_errno, MPI_ERR_OTHER, "**badbusinesscard");
ret = MPL_str_get_binary_arg(business_card, PTI_KEY, (char *)pt, sizeof(*pt), &len);
MPIR_ERR_CHKANDJUMP(ret != MPL_STR_SUCCESS || len != sizeof(*pt), mpi_errno, MPI_ERR_OTHER, "**badbusinesscard");
ret = MPL_str_get_binary_arg(business_card, PTIG_KEY, (char *)ptg, sizeof(*ptg), &len);
MPIR_ERR_CHKANDJUMP(ret != MPL_STR_SUCCESS || len != sizeof(*ptg), mpi_errno, MPI_ERR_OTHER, "**badbusinesscard");
ret = MPL_str_get_binary_arg(business_card, PTIC_KEY, (char *)ptc, sizeof(*ptc), &len);
MPIR_ERR_CHKANDJUMP(ret != MPL_STR_SUCCESS || len != sizeof(*ptc), mpi_errno, MPI_ERR_OTHER, "**badbusinesscard");
ret = MPL_str_get_binary_arg(business_card, PTIR_KEY, (char *)ptr, sizeof(*ptr), &len);
MPIR_ERR_CHKANDJUMP(ret != MPL_STR_SUCCESS || len != sizeof(*ptr), mpi_errno, MPI_ERR_OTHER, "**badbusinesscard");
ret = MPL_str_get_binary_arg(business_card, PTIRG_KEY, (char *)ptrg, sizeof(*ptrg), &len);
MPIR_ERR_CHKANDJUMP(ret != MPL_STR_SUCCESS || len != sizeof(*ptrg), mpi_errno, MPI_ERR_OTHER, "**badbusinesscard");
ret = MPL_str_get_binary_arg(business_card, PTIRC_KEY, (char *)ptrc, sizeof(*ptrc), &len);
MPIR_ERR_CHKANDJUMP(ret != MPL_STR_SUCCESS || len != sizeof(*ptrc), mpi_errno, MPI_ERR_OTHER, "**badbusinesscard");
fn_exit:
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPID_NEM_PTL_GET_ID_FROM_BC);
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPID_nem_mxm_get_business_card(int my_rank, char **bc_val_p, int *val_max_sz_p)
{
int mpi_errno = MPI_SUCCESS;
int str_errno = MPL_STR_SUCCESS;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MXM_GET_BUSINESS_CARD);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MXM_GET_BUSINESS_CARD);
str_errno = MPL_str_add_binary_arg(bc_val_p, val_max_sz_p, MXM_MPICH_ENDPOINT_KEY,
_mxm_obj.mxm_ep_addr, _mxm_obj.mxm_ep_addr_size);
if (str_errno) {
MPIR_ERR_CHKANDJUMP(str_errno == MPL_STR_NOMEM, mpi_errno, MPI_ERR_OTHER, "**buscard_len");
MPIR_ERR_SETANDJUMP(mpi_errno, MPI_ERR_OTHER, "**buscard");
}
fn_exit:
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MXM_GET_BUSINESS_CARD);
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Graph_neighbors_count_impl(MPID_Comm *comm_ptr, int rank, int *nneighbors)
{
int mpi_errno = MPI_SUCCESS;
MPIR_Topology *graph_ptr;
graph_ptr = MPIR_Topology_get( comm_ptr );
MPIR_ERR_CHKANDJUMP((!graph_ptr || graph_ptr->kind != MPI_GRAPH), mpi_errno, MPI_ERR_TOPOLOGY, "**notgraphtopo");
MPIR_ERR_CHKANDJUMP2((rank < 0 || rank >= graph_ptr->topo.graph.nnodes), mpi_errno, MPI_ERR_RANK, "**rank",
"**rank %d %d", rank, graph_ptr->topo.graph.nnodes );
if ( rank == 0 )
*nneighbors = graph_ptr->topo.graph.index[rank];
else
*nneighbors = graph_ptr->topo.graph.index[rank] -
graph_ptr->topo.graph.index[rank-1];
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Comm_create_keyval_impl(MPI_Comm_copy_attr_function *comm_copy_attr_fn,
MPI_Comm_delete_attr_function *comm_delete_attr_fn,
int *comm_keyval, void *extra_state)
{
int mpi_errno = MPI_SUCCESS;
MPII_Keyval *keyval_ptr;
keyval_ptr = (MPII_Keyval *)MPIR_Handle_obj_alloc( &MPII_Keyval_mem );
MPIR_ERR_CHKANDJUMP(!keyval_ptr, mpi_errno, MPI_ERR_OTHER,"**nomem");
/* Initialize the attribute dup function */
if (!MPIR_Process.attr_dup) {
MPIR_Process.attr_dup = MPIR_Attr_dup_list;
MPIR_Process.attr_free = MPIR_Attr_delete_list;
}
/* The handle encodes the keyval kind. Modify it to have the correct
field */
keyval_ptr->handle = (keyval_ptr->handle & ~(0x03c00000)) |
(MPIR_COMM << 22);
MPIR_Object_set_ref(keyval_ptr,1);
keyval_ptr->was_freed = 0;
keyval_ptr->kind = MPIR_COMM;
keyval_ptr->extra_state = extra_state;
keyval_ptr->copyfn.user_function = comm_copy_attr_fn;
keyval_ptr->copyfn.proxy = MPII_Attr_copy_c_proxy;
keyval_ptr->delfn.user_function = comm_delete_attr_fn;
keyval_ptr->delfn.proxy = MPII_Attr_delete_c_proxy;
MPIR_OBJ_PUBLISH_HANDLE(*comm_keyval, keyval_ptr->handle);
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPID_nem_mxm_init(MPIDI_PG_t * pg_p, int pg_rank, char **bc_val_p, int *val_max_sz_p)
{
int r;
int mpi_errno = MPI_SUCCESS;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MXM_INIT);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MXM_INIT);
/* first make sure that our private fields in the vc and req fit into the area provided */
MPIR_Assert(sizeof(MPID_nem_mxm_vc_area) <= MPIDI_NEM_VC_NETMOD_AREA_LEN);
MPIR_Assert(sizeof(MPID_nem_mxm_req_area) <= MPIDI_NEM_REQ_NETMOD_AREA_LEN);
/* mpich-specific initialization of mxm */
/* check if the user is not trying to override the tls setting
* before resetting it */
if (getenv("MXM_TLS") == NULL) {
r = MPL_putenv("MXM_TLS=rc,dc,ud");
MPIR_ERR_CHKANDJUMP(r, mpi_errno, MPI_ERR_OTHER, "**putenv");
}
/* [PB @ 2014-10-06] If hugepage support is not enabled, we force
* memory allocation to go through mmap. This is mainly to
* workaround issues in MXM with Ubuntu where the detection has
* some issues (either because of bugs on the platform or within
* MXM) causing errors. This can probably be deleted eventually
* when this issue is resolved. */
if (MPIR_CVAR_NEMESIS_MXM_HUGEPAGE == 0) {
if (getenv("MXM_MEM_ALLOC") == NULL) {
r = MPL_putenv("MXM_MEM_ALLOC=mmap,libc,sysv");
MPIR_ERR_CHKANDJUMP(r, mpi_errno, MPI_ERR_OTHER, "**putenv");
}
}
mpi_errno = _mxm_init(pg_rank, pg_p->size);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
mpi_errno = MPID_nem_mxm_get_business_card(pg_rank, bc_val_p, val_max_sz_p);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
mpi_errno =
MPIDI_CH3I_Register_anysource_notification(MPID_nem_mxm_anysource_posted,
MPID_nem_mxm_anysource_matched);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
mpi_errno = MPID_nem_register_initcomp_cb(_mxm_post_init);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIDI_CH3U_Comm_register_create_hook(_mxm_add_comm, NULL);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIDI_CH3U_Comm_register_destroy_hook(_mxm_del_comm, NULL);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPIDI_Anysource_improbe_fn = MPID_nem_mxm_anysource_improbe;
fn_exit:
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MXM_INIT);
return mpi_errno;
fn_fail:
goto fn_exit;
}
/*@
MPI_Cart_rank - Determines process rank in communicator given Cartesian
location
Input Parameters:
+ comm - communicator with cartesian structure (handle)
- coords - integer array (of size 'ndims', the number of dimensions of
the Cartesian topology associated with 'comm') specifying the cartesian
coordinates of a process
Output Parameters:
. rank - rank of specified process (integer)
Notes:
Out-of-range coordinates are erroneous for non-periodic dimensions.
Versions of MPICH before 1.2.2 returned 'MPI_PROC_NULL' for the rank in this
case.
.N SignalSafe
.N Fortran
.N Errors
.N MPI_SUCCESS
.N MPI_ERR_TOPOLOGY
.N MPI_ERR_RANK
.N MPI_ERR_ARG
@*/
int MPI_Cart_rank(MPI_Comm comm, const int coords[], int *rank)
{
int mpi_errno = MPI_SUCCESS;
MPIR_Comm *comm_ptr = NULL;
MPIR_Topology *cart_ptr;
MPIR_FUNC_TERSE_STATE_DECL(MPID_STATE_MPI_CART_RANK);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPIR_FUNC_TERSE_ENTER(MPID_STATE_MPI_CART_RANK);
/* Validate parameters, especially handles needing to be converted */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_COMM(comm, mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
# endif
/* Convert MPI object handles to object pointers */
MPIR_Comm_get_ptr( comm, comm_ptr );
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
/* Validate comm_ptr */
MPIR_Comm_valid_ptr( comm_ptr, mpi_errno, TRUE );
if (mpi_errno) goto fn_fail;
/* If comm_ptr is not valid, it will be reset to null */
MPIR_ERRTEST_ARGNULL(rank,"rank",mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
cart_ptr = MPIR_Topology_get( comm_ptr );
MPIR_ERR_CHKANDJUMP((!cart_ptr || cart_ptr->kind != MPI_CART), mpi_errno, MPI_ERR_TOPOLOGY, "**notcarttopo");
/* Validate coordinates */
# ifdef HAVE_ERROR_CHECKING
{
int i, ndims, coord;
MPID_BEGIN_ERROR_CHECKS;
{
ndims = cart_ptr->topo.cart.ndims;
if (ndims != 0) {
MPIR_ERRTEST_ARGNULL(coords,"coords",mpi_errno);
}
for (i=0; i<ndims; i++) {
if (!cart_ptr->topo.cart.periodic[i]) {
coord = coords[i];
MPIR_ERR_CHKANDJUMP3(
(coord < 0 || coord >= cart_ptr->topo.cart.dims[i] ), mpi_errno, MPI_ERR_ARG, "**cartcoordinvalid",
"**cartcoordinvalid %d %d %d",i, coords[i], cart_ptr->topo.cart.dims[i]-1 );
}
}
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
MPIR_Cart_rank_impl(cart_ptr, coords, rank);
/* ... end of body of routine ... */
fn_exit:
MPIR_FUNC_TERSE_EXIT(MPID_STATE_MPI_CART_RANK);
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_cart_rank",
"**mpi_cart_rank %C %p %p", comm, coords, rank);
}
# endif
mpi_errno = MPIR_Err_return_comm( comm_ptr, FCNAME, mpi_errno );
goto fn_exit;
/* --END ERROR HANDLING-- */
}
int MPID_Win_free(MPIR_Win ** win_ptr)
{
int mpi_errno = MPI_SUCCESS;
int in_use;
MPIR_Comm *comm_ptr;
MPIR_Errflag_t errflag = MPIR_ERR_NONE;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPID_WIN_FREE);
MPIR_FUNC_VERBOSE_RMA_ENTER(MPID_STATE_MPID_WIN_FREE);
MPIR_ERR_CHKANDJUMP(((*win_ptr)->states.access_state != MPIDI_RMA_NONE &&
(*win_ptr)->states.access_state != MPIDI_RMA_FENCE_ISSUED &&
(*win_ptr)->states.access_state != MPIDI_RMA_FENCE_GRANTED) ||
((*win_ptr)->states.exposure_state != MPIDI_RMA_NONE),
mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
/* 1. Here we must wait until all passive locks are released on this target,
* because for some UNLOCK messages, we do not send ACK back to origin,
* we must wait until lock is released so that we can free window.
* 2. We also need to wait until AT completion counter being zero, because
* this counter is increment everytime we meet a GET-like operation, it is
* possible that when target entering Win_free, passive epoch is not finished
* yet and there are still GETs doing on this target.
* 3. We also need to wait until lock queue becomes empty. It is possible
* that some lock requests is still waiting in the queue when target is
* entering Win_free. */
while ((*win_ptr)->current_lock_type != MPID_LOCK_NONE ||
(*win_ptr)->at_completion_counter != 0 ||
(*win_ptr)->target_lock_queue_head != NULL ||
(*win_ptr)->current_target_lock_data_bytes != 0 || (*win_ptr)->sync_request_cnt != 0) {
mpi_errno = wait_progress_engine();
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
}
mpi_errno = MPID_Barrier((*win_ptr)->comm_ptr, &errflag);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
/* Free window resources in lower layer. */
if (MPIDI_CH3U_Win_hooks.win_free != NULL) {
mpi_errno = MPIDI_CH3U_Win_hooks.win_free(win_ptr);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
}
/* dequeue window from the global list */
MPIR_Assert((*win_ptr)->active == FALSE);
DL_DELETE(MPIDI_RMA_Win_inactive_list_head, (*win_ptr));
if (MPIDI_RMA_Win_inactive_list_head == NULL && MPIDI_RMA_Win_active_list_head == NULL) {
/* this is the last window, de-register RMA progress hook */
mpi_errno = MPID_Progress_deregister_hook(MPIDI_CH3I_RMA_Progress_hook_id);
if (mpi_errno != MPI_SUCCESS) {
MPIR_ERR_POP(mpi_errno);
}
}
comm_ptr = (*win_ptr)->comm_ptr;
mpi_errno = MPIR_Comm_free_impl(comm_ptr);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
if ((*win_ptr)->basic_info_table != NULL)
MPL_free((*win_ptr)->basic_info_table);
MPL_free((*win_ptr)->op_pool_start);
MPL_free((*win_ptr)->target_pool_start);
MPL_free((*win_ptr)->slots);
MPL_free((*win_ptr)->target_lock_entry_pool_start);
MPIR_Assert((*win_ptr)->current_target_lock_data_bytes == 0);
/* Free the attached buffer for windows created with MPI_Win_allocate() */
if ((*win_ptr)->create_flavor == MPI_WIN_FLAVOR_ALLOCATE ||
(*win_ptr)->create_flavor == MPI_WIN_FLAVOR_SHARED) {
if ((*win_ptr)->shm_allocated == FALSE && (*win_ptr)->size > 0) {
MPL_free((*win_ptr)->base);
}
}
MPIR_Object_release_ref(*win_ptr, &in_use);
/* MPI windows don't have reference count semantics, so this should always be true */
MPIR_Assert(!in_use);
MPIR_Handle_obj_free(&MPIR_Win_mem, *win_ptr);
fn_exit:
MPIR_FUNC_VERBOSE_RMA_EXIT(MPID_STATE_MPID_WIN_FREE);
return mpi_errno;
fn_fail:
goto fn_exit;
}
/*@
MPI_Dist_graph_neighbors - Provides adjacency information for a distributed graph topology.
Input Parameters:
+ comm - communicator with distributed graph topology (handle)
. maxindegree - size of sources and sourceweights arrays (non-negative integer)
- maxoutdegree - size of destinations and destweights arrays (non-negative integer)
Output Parameters:
+ sources - processes for which the calling process is a destination (array of non-negative integers)
. sourceweights - weights of the edges into the calling process (array of non-negative integers)
. destinations - processes for which the calling process is a source (array of non-negative integers)
- destweights - weights of the edges out of the calling process (array of non-negative integers)
.N ThreadSafe
.N Fortran
.N Errors
.N MPI_SUCCESS
@*/
int MPI_Dist_graph_neighbors(MPI_Comm comm,
int maxindegree, int sources[], int sourceweights[],
int maxoutdegree, int destinations[], int destweights[])
{
int mpi_errno = MPI_SUCCESS;
MPIR_Comm *comm_ptr = NULL;
MPIR_FUNC_TERSE_STATE_DECL(MPID_STATE_MPI_DIST_GRAPH_NEIGHBORS);
MPIR_ERRTEST_INITIALIZED_ORDIE();
/* FIXME: Why does this routine need a CS */
MPID_THREAD_CS_ENTER(GLOBAL, MPIR_THREAD_GLOBAL_ALLFUNC_MUTEX);
MPIR_FUNC_TERSE_ENTER(MPID_STATE_MPI_DIST_GRAPH_NEIGHBORS);
/* Validate parameters, especially handles needing to be converted */
#ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_COMM(comm, mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
#endif
/* Convert MPI object handles to object pointers */
MPIR_Comm_get_ptr(comm, comm_ptr);
/* Validate parameters */
#ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_Topology *topo_ptr = NULL;
topo_ptr = MPIR_Topology_get(comm_ptr);
MPIR_ERR_CHKANDJUMP(!topo_ptr ||
topo_ptr->kind != MPI_DIST_GRAPH, mpi_errno, MPI_ERR_TOPOLOGY,
"**notdistgraphtopo");
MPIR_ERRTEST_ARGNEG(maxindegree, "maxindegree", mpi_errno);
MPIR_ERRTEST_ARGNEG(maxoutdegree, "maxoutdegree", mpi_errno);
MPIR_ERR_CHKANDJUMP3((maxindegree < topo_ptr->topo.dist_graph.indegree), mpi_errno,
MPI_ERR_ARG, "**argtoosmall", "**argtoosmall %s %d %d",
"maxindegree", maxindegree, topo_ptr->topo.dist_graph.indegree);
MPIR_ERR_CHKANDJUMP3((maxoutdegree < topo_ptr->topo.dist_graph.outdegree), mpi_errno,
MPI_ERR_ARG, "**argtoosmall", "**argtoosmall %s %d %d",
"maxoutdegree", maxoutdegree, topo_ptr->topo.dist_graph.outdegree);
}
MPID_END_ERROR_CHECKS;
}
#endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
mpi_errno = MPIR_Dist_graph_neighbors_impl(comm_ptr,
maxindegree, sources, sourceweights,
maxoutdegree, destinations, destweights);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
/* ... end of body of routine ... */
fn_exit:
MPIR_FUNC_TERSE_EXIT(MPID_STATE_MPI_DIST_GRAPH_NEIGHBORS);
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_dist_graph_neighbors",
"**mpi_dist_graph_neighbors %C %d %p %p %d %p %p", comm, maxindegree,
sources, sourceweights, maxoutdegree, destinations, destweights);
#endif
mpi_errno = MPIR_Err_return_comm(comm_ptr, FCNAME, mpi_errno);
goto fn_exit;
/* --END ERROR HANDLING-- */
}
int MPID_Compare_and_swap(const void *origin_addr, const void *compare_addr,
void *result_addr, MPI_Datatype datatype, int target_rank,
MPI_Aint target_disp, MPID_Win * win_ptr)
{
int mpi_errno = MPI_SUCCESS;
int rank;
MPIDI_VC_t *orig_vc = NULL, *target_vc = NULL;
int made_progress = 0;
MPIDI_STATE_DECL(MPID_STATE_MPID_COMPARE_AND_SWAP);
MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPID_COMPARE_AND_SWAP);
MPIR_ERR_CHKANDJUMP(win_ptr->states.access_state == MPIDI_RMA_NONE,
mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
if (target_rank == MPI_PROC_NULL) {
goto fn_exit;
}
rank = win_ptr->comm_ptr->rank;
if (win_ptr->shm_allocated == TRUE && target_rank != rank &&
win_ptr->create_flavor != MPI_WIN_FLAVOR_SHARED) {
/* check if target is local and shared memory is allocated on window,
* if so, we directly perform this operation on shared memory region. */
/* FIXME: Here we decide whether to perform SHM operations by checking if origin and target are on
* the same node. However, in ch3:sock, even if origin and target are on the same node, they do
* not within the same SHM region. Here we filter out ch3:sock by checking shm_allocated flag first,
* which is only set to TRUE when SHM region is allocated in nemesis.
* In future we need to figure out a way to check if origin and target are in the same "SHM comm".
*/
MPIDI_Comm_get_vc(win_ptr->comm_ptr, rank, &orig_vc);
MPIDI_Comm_get_vc(win_ptr->comm_ptr, target_rank, &target_vc);
}
/* The datatype must be predefined, and one of: C integer, Fortran integer,
* Logical, Multi-language types, or Byte. This is checked above the ADI,
* so there's no need to check it again here. */
/* FIXME: For shared memory windows, we should provide an implementation
* that uses a processor atomic operation. */
if (target_rank == rank || win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED ||
(win_ptr->shm_allocated == TRUE && orig_vc->node_id == target_vc->node_id)) {
mpi_errno = MPIDI_CH3I_Shm_cas_op(origin_addr, compare_addr, result_addr,
datatype, target_rank, target_disp, win_ptr);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
}
else {
MPIDI_RMA_Op_t *op_ptr = NULL;
MPIDI_CH3_Pkt_cas_t *cas_pkt = NULL;
MPI_Aint type_size;
void *src = NULL, *dest = NULL;
/* Append this operation to the RMA ops queue */
mpi_errno = MPIDI_CH3I_Win_get_op(win_ptr, &op_ptr);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
MPIR_T_PVAR_TIMER_START(RMA, rma_rmaqueue_set);
/******************** Setting operation struct areas ***********************/
op_ptr->origin_addr = (void *) origin_addr;
op_ptr->origin_count = 1;
op_ptr->origin_datatype = datatype;
op_ptr->result_addr = result_addr;
op_ptr->result_datatype = datatype;
op_ptr->compare_addr = (void *) compare_addr;
op_ptr->compare_datatype = datatype;
op_ptr->target_rank = target_rank;
op_ptr->piggyback_lock_candidate = 1; /* CAS is always able to piggyback LOCK */
/************** Setting packet struct areas in operation ****************/
cas_pkt = &(op_ptr->pkt.cas);
MPIDI_Pkt_init(cas_pkt, MPIDI_CH3_PKT_CAS_IMMED);
cas_pkt->addr = (char *) win_ptr->basic_info_table[target_rank].base_addr +
win_ptr->basic_info_table[target_rank].disp_unit * target_disp;
cas_pkt->datatype = datatype;
cas_pkt->target_win_handle = win_ptr->basic_info_table[target_rank].win_handle;
cas_pkt->flags = MPIDI_CH3_PKT_FLAG_NONE;
/* REQUIRE: All datatype arguments must be of the same, builtin
* type and counts must be 1. */
MPID_Datatype_get_size_macro(datatype, type_size);
MPIU_Assert(type_size <= sizeof(MPIDI_CH3_CAS_Immed_u));
src = (void *) origin_addr, dest = (void *) (&(cas_pkt->origin_data));
mpi_errno = immed_copy(src, dest, type_size);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
src = (void *) compare_addr, dest = (void *) (&(cas_pkt->compare_data));
mpi_errno = immed_copy(src, dest, type_size);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
MPIR_T_PVAR_TIMER_END(RMA, rma_rmaqueue_set);
mpi_errno = MPIDI_CH3I_Win_enqueue_op(win_ptr, op_ptr);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIDI_CH3I_RMA_Make_progress_target(win_ptr, target_rank, &made_progress);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
if (MPIR_CVAR_CH3_RMA_ACTIVE_REQ_THRESHOLD >= 0 &&
MPIDI_CH3I_RMA_Active_req_cnt >= MPIR_CVAR_CH3_RMA_ACTIVE_REQ_THRESHOLD) {
while (MPIDI_CH3I_RMA_Active_req_cnt >= MPIR_CVAR_CH3_RMA_ACTIVE_REQ_THRESHOLD) {
mpi_errno = wait_progress_engine();
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
}
}
}
fn_exit:
MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPID_COMPARE_AND_SWAP);
return mpi_errno;
/* --BEGIN ERROR HANDLING-- */
fn_fail:
goto fn_exit;
/* --END ERROR HANDLING-- */
}
int MPIDI_CH3I_Get_accumulate(const void *origin_addr, int origin_count,
MPI_Datatype origin_datatype, void *result_addr, int result_count,
MPI_Datatype result_datatype, int target_rank, MPI_Aint target_disp,
int target_count, MPI_Datatype target_datatype, MPI_Op op,
MPID_Win * win_ptr, MPID_Request * ureq)
{
int mpi_errno = MPI_SUCCESS;
MPIDI_msg_sz_t orig_data_sz, target_data_sz;
int rank;
int dt_contig ATTRIBUTE((unused));
MPI_Aint dt_true_lb ATTRIBUTE((unused));
MPID_Datatype *dtp;
MPIDI_VC_t *orig_vc = NULL, *target_vc = NULL;
int made_progress = 0;
MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3I_GET_ACCUMULATE);
MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_CH3I_GET_ACCUMULATE);
MPIR_ERR_CHKANDJUMP(win_ptr->states.access_state == MPIDI_RMA_NONE,
mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
if (target_rank == MPI_PROC_NULL) {
goto fn_exit;
}
MPIDI_Datatype_get_info(target_count, target_datatype, dt_contig, target_data_sz, dtp,
dt_true_lb);
if (target_data_sz == 0) {
goto fn_exit;
}
rank = win_ptr->comm_ptr->rank;
if (win_ptr->shm_allocated == TRUE && target_rank != rank &&
win_ptr->create_flavor != MPI_WIN_FLAVOR_SHARED) {
/* check if target is local and shared memory is allocated on window,
* if so, we directly perform this operation on shared memory region. */
/* FIXME: Here we decide whether to perform SHM operations by checking if origin and target are on
* the same node. However, in ch3:sock, even if origin and target are on the same node, they do
* not within the same SHM region. Here we filter out ch3:sock by checking shm_allocated flag first,
* which is only set to TRUE when SHM region is allocated in nemesis.
* In future we need to figure out a way to check if origin and target are in the same "SHM comm".
*/
MPIDI_Comm_get_vc(win_ptr->comm_ptr, rank, &orig_vc);
MPIDI_Comm_get_vc(win_ptr->comm_ptr, target_rank, &target_vc);
}
/* Do =! rank first (most likely branch?) */
if (target_rank == rank || win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED ||
(win_ptr->shm_allocated == TRUE && orig_vc->node_id == target_vc->node_id)) {
mpi_errno = MPIDI_CH3I_Shm_get_acc_op(origin_addr, origin_count, origin_datatype,
result_addr, result_count, result_datatype,
target_rank, target_disp, target_count,
target_datatype, op, win_ptr);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
if (ureq) {
/* Complete user request and release the ch3 ref */
mpi_errno = MPID_Request_complete(ureq);
if (mpi_errno != MPI_SUCCESS) {
MPIR_ERR_POP(mpi_errno);
}
}
}
else {
MPIDI_RMA_Op_t *op_ptr = NULL;
MPIDI_CH3_Pkt_get_accum_t *get_accum_pkt;
MPI_Aint origin_type_size;
MPI_Aint target_type_size;
int use_immed_pkt = FALSE, i;
int is_origin_contig, is_target_contig, is_result_contig;
MPI_Aint stream_elem_count, stream_unit_count;
MPI_Aint predefined_dtp_size, predefined_dtp_count, predefined_dtp_extent;
MPID_Datatype *origin_dtp = NULL, *target_dtp = NULL, *result_dtp = NULL;
int is_empty_origin = FALSE;
/* Judge if origin buffer is empty */
if (op == MPI_NO_OP)
is_empty_origin = TRUE;
/* Append the operation to the window's RMA ops queue */
mpi_errno = MPIDI_CH3I_Win_get_op(win_ptr, &op_ptr);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
/* TODO: Can we use the MPIDI_RMA_ACC_CONTIG optimization? */
MPIR_T_PVAR_TIMER_START(RMA, rma_rmaqueue_set);
/******************** Setting operation struct areas ***********************/
op_ptr->origin_addr = (void *) origin_addr;
op_ptr->origin_count = origin_count;
op_ptr->origin_datatype = origin_datatype;
op_ptr->result_addr = result_addr;
op_ptr->result_count = result_count;
op_ptr->result_datatype = result_datatype;
op_ptr->target_rank = target_rank;
/* Remember user request */
op_ptr->ureq = ureq;
/* if source or target datatypes are derived, increment their
* reference counts */
if (is_empty_origin == FALSE && !MPIR_DATATYPE_IS_PREDEFINED(origin_datatype)) {
MPID_Datatype_get_ptr(origin_datatype, origin_dtp);
}
if (!MPIR_DATATYPE_IS_PREDEFINED(result_datatype)) {
MPID_Datatype_get_ptr(result_datatype, result_dtp);
}
if (!MPIR_DATATYPE_IS_PREDEFINED(target_datatype)) {
MPID_Datatype_get_ptr(target_datatype, target_dtp);
}
if (is_empty_origin == FALSE) {
MPID_Datatype_get_size_macro(origin_datatype, origin_type_size);
MPIU_Assign_trunc(orig_data_sz, origin_count * origin_type_size, MPIDI_msg_sz_t);
}
else {
/* If origin buffer is empty, set origin data size to 0 */
orig_data_sz = 0;
}
MPID_Datatype_get_size_macro(target_datatype, target_type_size);
/* Get size and count for predefined datatype elements */
if (MPIR_DATATYPE_IS_PREDEFINED(target_datatype)) {
predefined_dtp_size = target_type_size;
predefined_dtp_count = target_count;
MPID_Datatype_get_extent_macro(target_datatype, predefined_dtp_extent);
}
else {
MPIU_Assert(target_dtp->basic_type != MPI_DATATYPE_NULL);
MPID_Datatype_get_size_macro(target_dtp->basic_type, predefined_dtp_size);
predefined_dtp_count = target_data_sz / predefined_dtp_size;
MPID_Datatype_get_extent_macro(target_dtp->basic_type, predefined_dtp_extent);
}
MPIU_Assert(predefined_dtp_count > 0 && predefined_dtp_size > 0 &&
predefined_dtp_extent > 0);
/* Calculate number of predefined elements in each stream unit, and
* total number of stream units. */
stream_elem_count = MPIDI_CH3U_Acc_stream_size / predefined_dtp_extent;
stream_unit_count = (predefined_dtp_count - 1) / stream_elem_count + 1;
MPIU_Assert(stream_elem_count > 0 && stream_unit_count > 0);
for (i = 0; i < stream_unit_count; i++) {
if (origin_dtp != NULL) {
MPID_Datatype_add_ref(origin_dtp);
}
if (target_dtp != NULL) {
MPID_Datatype_add_ref(target_dtp);
}
if (result_dtp != NULL) {
MPID_Datatype_add_ref(result_dtp);
}
}
if (is_empty_origin == FALSE) {
MPID_Datatype_is_contig(origin_datatype, &is_origin_contig);
}
else {
/* If origin buffer is empty, mark origin data as contig data */
is_origin_contig = 1;
}
MPID_Datatype_is_contig(target_datatype, &is_target_contig);
MPID_Datatype_is_contig(result_datatype, &is_result_contig);
/* Judge if we can use IMMED data packet */
if ((is_empty_origin == TRUE || MPIR_DATATYPE_IS_PREDEFINED(origin_datatype)) &&
MPIR_DATATYPE_IS_PREDEFINED(result_datatype) &&
MPIR_DATATYPE_IS_PREDEFINED(target_datatype) &&
is_origin_contig && is_target_contig && is_result_contig) {
if (target_data_sz <= MPIDI_RMA_IMMED_BYTES)
use_immed_pkt = TRUE;
}
/* Judge if this operation is a piggyback candidate */
if ((is_empty_origin == TRUE || MPIR_DATATYPE_IS_PREDEFINED(origin_datatype)) &&
MPIR_DATATYPE_IS_PREDEFINED(result_datatype) &&
MPIR_DATATYPE_IS_PREDEFINED(target_datatype)) {
/* FIXME: currently we only piggyback LOCK flag with op using predefined datatypes
* for origin, target and result data. We should extend this optimization to derived
* datatypes as well. */
if (orig_data_sz <= MPIR_CVAR_CH3_RMA_OP_PIGGYBACK_LOCK_DATA_SIZE)
op_ptr->piggyback_lock_candidate = 1;
}
/************** Setting packet struct areas in operation ****************/
get_accum_pkt = &(op_ptr->pkt.get_accum);
if (use_immed_pkt) {
MPIDI_Pkt_init(get_accum_pkt, MPIDI_CH3_PKT_GET_ACCUM_IMMED);
}
else {
MPIDI_Pkt_init(get_accum_pkt, MPIDI_CH3_PKT_GET_ACCUM);
}
get_accum_pkt->addr = (char *) win_ptr->basic_info_table[target_rank].base_addr +
win_ptr->basic_info_table[target_rank].disp_unit * target_disp;
get_accum_pkt->count = target_count;
get_accum_pkt->datatype = target_datatype;
get_accum_pkt->info.dataloop_size = 0;
get_accum_pkt->op = op;
get_accum_pkt->target_win_handle = win_ptr->basic_info_table[target_rank].win_handle;
get_accum_pkt->flags = MPIDI_CH3_PKT_FLAG_NONE;
if (use_immed_pkt) {
void *src = (void *) origin_addr, *dest = (void *) (get_accum_pkt->info.data);
mpi_errno = immed_copy(src, dest, orig_data_sz);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
}
MPIR_T_PVAR_TIMER_END(RMA, rma_rmaqueue_set);
mpi_errno = MPIDI_CH3I_Win_enqueue_op(win_ptr, op_ptr);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
mpi_errno = MPIDI_CH3I_RMA_Make_progress_target(win_ptr, target_rank, &made_progress);
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
if (MPIR_CVAR_CH3_RMA_ACTIVE_REQ_THRESHOLD >= 0 &&
MPIDI_CH3I_RMA_Active_req_cnt >= MPIR_CVAR_CH3_RMA_ACTIVE_REQ_THRESHOLD) {
while (MPIDI_CH3I_RMA_Active_req_cnt >= MPIR_CVAR_CH3_RMA_ACTIVE_REQ_THRESHOLD) {
mpi_errno = wait_progress_engine();
if (mpi_errno != MPI_SUCCESS)
MPIR_ERR_POP(mpi_errno);
}
}
}
fn_exit:
MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_CH3I_GET_ACCUMULATE);
return mpi_errno;
/* --BEGIN ERROR HANDLING-- */
fn_fail:
goto fn_exit;
/* --END ERROR HANDLING-- */
}
