/*
* Indexed
*/
MPID_Dataloop *MPID_Dataloop_init_indexed( int count, int *blocksize,
MPI_Aint *offset )
{
MPID_Dataloop *it;
MPI_Aint extent = 0;
int i;
it = (MPID_Dataloop *)MPIU_Malloc( sizeof(MPID_Dataloop) );
it->kind = MPID_DTYPE_INDEXED | DATALOOP_FINAL_MASK;
it->loop_params.i_t.count = count;
it->loop_params.i_t.blocksize = (int *)MPIU_Malloc( sizeof(int) * count );
it->loop_params.i_t.offset =
(MPI_Aint *)MPIU_Malloc( sizeof(MPI_Aint) * count );
for (i=0; i<count; i++) {
it->loop_params.i_t.offset[i] = offset[i];
it->loop_params.i_t.blocksize[i] = blocksize[i];
if (offset[i] + blocksize[i] > extent)
extent = offset[i] + blocksize[i];
}
it->loop_params.i_t.dataloop = 0;
it->extent = extent;
it->handle = 0;
return it;
}
/*
* Add an enviroinment variable to the global list of variables
*/
int MPIE_Putenv( ProcessWorld *pWorld, const char *env_string )
{
EnvInfo *genv;
EnvData *p;
/* FIXME: This should be getGenv (so allocation/init in one place) */
if (!pWorld->genv) {
genv = (EnvInfo *)MPIU_Malloc( sizeof(EnvInfo) );
genv->includeAll = 1;
genv->envPairs = 0;
genv->envNames = 0;
pWorld->genv = genv;
}
genv = pWorld->genv;
p = (EnvData *)MPIU_Malloc( sizeof(EnvData) );
if (!p) return 1;
p->name = 0;
p->value = 0;
p->envvalue = (const char *)MPIU_Strdup( env_string );
if (!p->envvalue) return 1;
p->nextData = genv->envPairs;
genv->envPairs = p;
return 0;
}
int
MPID_nem_ib_allocate_memory(int pg_rank, int pg_size)
{
int mpi_errno = MPI_SUCCESS;
process_info.polling_group_size = 0;
if (rdma_polling_set_limit > 0)
{
process_info.polling_set = (MPIDI_VC_t**) MPIU_Malloc(rdma_polling_set_limit * sizeof(MPIDI_VC_t*));
}
else
{
process_info.polling_set = (MPIDI_VC_t**) MPIU_Malloc(pg_size * sizeof(MPIDI_VC_t*));
}
if (!process_info.polling_set)
{
fprintf(
stderr,
"[%s:%d]: %s\n",
__FILE__,
__LINE__,
"unable to allocate space for polling set\n");
return 0;
}
/* We need to allocate vbufs for send/recv path */
if ((mpi_errno = allocate_vbufs(rdma_vbuf_pool_size)))
{
return mpi_errno;
}
return mpi_errno;
}
int MPID_NS_Create( const MPID_Info *info_ptr, MPID_NS_Handle *handle_ptr )
{
static const char FCNAME[] = "MPID_NS_Create";
int err;
int length;
char *pmi_namepub_kvs;
*handle_ptr = (MPID_NS_Handle)MPIU_Malloc( sizeof(struct MPID_NS_Handle) );
/* --BEGIN ERROR HANDLING-- */
if (!*handle_ptr)
{
err = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**nomem", 0);
return err;
}
/* --END ERROR HANDLING-- */
err = PMI_KVS_Get_name_length_max(&length);
/* --BEGIN ERROR HANDLING-- */
if (err != PMI_SUCCESS)
{
err = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**fail", 0);
}
/* --END ERROR HANDLING-- */
(*handle_ptr)->kvsname = (char*)MPIU_Malloc(length);
/* --BEGIN ERROR HANDLING-- */
if (!(*handle_ptr)->kvsname)
{
err = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**nomem", 0);
return err;
}
/* --END ERROR HANDLING-- */
pmi_namepub_kvs = getenv("PMI_NAMEPUB_KVS");
if (pmi_namepub_kvs)
{
MPIU_Strncpy((*handle_ptr)->kvsname, pmi_namepub_kvs, length);
}
else
{
err = PMI_KVS_Get_my_name((*handle_ptr)->kvsname, length);
/* --BEGIN ERROR HANDLING-- */
if (err != PMI_SUCCESS)
{
err = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**fail", 0);
}
/* --END ERROR HANDLING-- */
}
/*printf("namepub kvs: <%s>\n", (*handle_ptr)->kvsname);fflush(stdout);*/
return 0;
}
static int reinit_pmi(void)
{
int ret;
int has_parent = 0;
int pg_rank, pg_size;
int kvs_name_sz, pg_id_sz;
MPIDI_STATE_DECL(MPID_STATE_REINIT_PMI);
MPIDI_FUNC_ENTER(MPID_STATE_REINIT_PMI);
/* Init pmi and do some sanity checks */
ret = PMI_Init(&has_parent);
CHECK_ERR(ret, "pmi_init");
ret = PMI_Get_rank(&pg_rank);
CHECK_ERR(ret, "pmi_get_rank");
ret = PMI_Get_size(&pg_size);
CHECK_ERR(ret, "pmi_get_size");
CHECK_ERR(pg_size != MPIDI_Process.my_pg->size, "pg size differs after restart");
CHECK_ERR(pg_rank != MPIDI_Process.my_pg_rank, "pg rank differs after restart");
/* get new pg_id */
ret = PMI_KVS_Get_name_length_max(&pg_id_sz);
CHECK_ERR(ret, "pmi_get_id_length_max");
MPIU_Free(MPIDI_Process.my_pg->id);
MPIDI_Process.my_pg->id = MPIU_Malloc(pg_id_sz + 1);
CHECK_ERR(MPIDI_Process.my_pg->id == NULL, "malloc failed");
ret = PMI_KVS_Get_my_name(MPIDI_Process.my_pg->id, pg_id_sz);
CHECK_ERR(ret, "pmi_kvs_get_my_name");
/* get new kvsname */
ret = PMI_KVS_Get_name_length_max(&kvs_name_sz);
CHECK_ERR(ret, "PMI_KVS_Get_name_length_max");
MPIU_Free(MPIDI_Process.my_pg->connData);
MPIDI_Process.my_pg->connData = MPIU_Malloc(kvs_name_sz + 1);
CHECK_ERR(MPIDI_Process.my_pg->connData == NULL, "malloc failed");
ret = PMI_KVS_Get_my_name(MPIDI_Process.my_pg->connData, kvs_name_sz);
CHECK_ERR(ret, "PMI_Get_my_name");
MPIDI_FUNC_EXIT(MPID_STATE_REINIT_PMI);
return 0;
}
int MPIDI_CH3U_Post_data_receive_unexpected(MPID_Request * rreq)
{
int mpi_errno = MPI_SUCCESS;
MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3U_POST_DATA_RECEIVE_UNEXPECTED);
MPIDI_FUNC_ENTER(MPID_STATE_MPIDI_CH3U_POST_DATA_RECEIVE_UNEXPECTED);
/* FIXME: to improve performance, allocate temporary buffer from a
specialized buffer pool. */
/* FIXME: to avoid memory exhaustion, integrate buffer pool management
with flow control */
MPIU_DBG_MSG(CH3_OTHER,VERBOSE,"unexpected request allocated");
rreq->dev.tmpbuf = MPIU_Malloc(rreq->dev.recv_data_sz);
if (!rreq->dev.tmpbuf) {
MPIU_ERR_SETANDJUMP1(mpi_errno,MPI_ERR_OTHER,"**nomem","**nomem %d",
rreq->dev.recv_data_sz);
}
rreq->dev.tmpbuf_sz = rreq->dev.recv_data_sz;
rreq->dev.iov[0].MPID_IOV_BUF = (MPID_IOV_BUF_CAST)rreq->dev.tmpbuf;
rreq->dev.iov[0].MPID_IOV_LEN = rreq->dev.recv_data_sz;
rreq->dev.iov_count = 1;
rreq->dev.OnDataAvail = MPIDI_CH3_ReqHandler_UnpackUEBufComplete;
rreq->dev.recv_pending_count = 2;
fn_fail:
MPIDI_FUNC_EXIT(MPID_STATE_MPIDI_CH3U_POST_DATA_RECEIVE_UNEXPECTED);
return mpi_errno;
}
static void context_id_init(void)
{
int i;
#if defined(FINEGRAIN_MPI)
MPIU_Assert (NULL == context_mask);
context_mask = (uint32_t *)MPIU_Malloc(MPIR_MAX_CONTEXT_MASK * sizeof(uint32_t));
MPIU_Assert (NULL != context_mask);
#endif
for (i = 1; i < MPIR_MAX_CONTEXT_MASK; i++) {
context_mask[i] = 0xFFFFFFFF;
}
/* The first two values are already used (comm_world, comm_self).
* The third value is also used for the internal-only copy of
* comm_world, if needed by mpid. */
#ifdef MPID_NEEDS_ICOMM_WORLD
context_mask[0] = 0xFFFFFFF8;
#else
context_mask[0] = 0xFFFFFFFC;
#endif
initialize_context_mask = 0;
#ifdef MPICH_DEBUG_HANDLEALLOC
/* check for context ID leaks in MPI_Finalize. Use (_PRIO-1) to make sure
* that we run after MPID_Finalize. */
MPIR_Add_finalize(check_context_ids_on_finalize, context_mask, MPIR_FINALIZE_CALLBACK_PRIO - 1); /* FG: TODO IMPORTANT */
#endif
}
/*
* MPIU_Thread_create()
*/
void MPIU_Thread_create(MPIU_Thread_func_t func, void * data, MPIU_Thread_id_t * idp, int * errp)
{
struct MPEI_Thread_info * thread_info;
int err = MPIU_THREAD_SUCCESS;
/* FIXME: faster allocation, or avoid it all together? */
thread_info = (struct MPEI_Thread_info *) MPIU_Malloc(sizeof(struct MPEI_Thread_info));
if (thread_info != NULL)
{
pthread_attr_t attr;
thread_info->func = func;
thread_info->data = data;
pthread_attr_init(&attr);
pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED);
err = pthread_create(idp, &attr, MPEI_Thread_start, thread_info);
/* FIXME: convert error to an MPIU_THREAD_ERR value */
pthread_attr_destroy(&attr);
}
else
{
err = 1000000000;
}
if (errp != NULL)
{
*errp = err;
}
}
/*
* MPIU_Thread_create()
*/
void MPIU_Thread_create(MPIU_Thread_func_t func, void * data, MPIU_Thread_id_t * idp, int * errp)
{
struct MPEI_Thread_info * thread_info;
int err = MPIU_THREAD_SUCCESS;
thread_info = (struct MPEI_Thread_info *) MPIU_Malloc(sizeof(struct MPEI_Thread_info));
if (thread_info != NULL)
{
thread_info->func = func;
thread_info->data = data;
*idp = CreateThread(NULL, 0, MPEI_Thread_start, thread_info, 0, NULL);
if (*idp == NULL)
{
err = GetLastError();
}
}
else
{
err = 1000000000;
}
if (errp != NULL)
{
*errp = err;
}
}
/*
* MPIU_Thread_create()
*/
void MPIU_Thread_create(MPIU_Thread_func_t func, void * data, MPIU_Thread_id_t * idp, int * errp)
{
struct MPEI_Thread_info * thread_info;
int err = MPIU_THREAD_SUCCESS;
/* FIXME: faster allocation, or avoid it all together? */
thread_info = (struct MPEI_Thread_info *) MPIU_Malloc(sizeof(struct MPEI_Thread_info));
if (thread_info != NULL)
{
thread_info->func = func;
thread_info->data = data;
err = thr_create(NULL, 0, MPEI_Thread_start, thread_info, THR_DETACHED, idp);
/* FIXME: convert error to an MPIU_THREAD_ERR value */
}
else
{
err = 1000000000;
}
if (errp != NULL)
{
*errp = err;
}
}
int MPID_nem_ib_init_hash_table(
MPID_nem_ib_hash_table_ptr_t table,
uint32_t nentries)
{
int mpi_errno = MPI_SUCCESS;
table->entries = MPIU_Malloc(
sizeof(MPID_nem_ib_hash_elem_t) * nentries);
table->num_entries = nentries;
if(NULL == table->entries) {
MPIU_CHKMEM_SETERR(mpi_errno,
sizeof(MPID_nem_ib_hash_elem_t) * nentries,
"IB Module Hash Table");
}
memset(table->entries, 0,
sizeof(MPID_nem_ib_hash_elem_t) * nentries);
pthread_mutex_init(&table->hash_table_lock, NULL);
/* fn_exit: */
return mpi_errno;
/* fn_fail:
goto fn_exit;
*/
}
static int create_r_cookie (char *hostname, int port, int data_sz, char **cookie, int *len)
{
int mpi_errno = MPI_SUCCESS;
int hostname_len;
int cookie_len;
r_cookie_t *c;
hostname_len = strnlen (hostname, MAX_HOSTNAME_LEN) + 1;
cookie_len = sizeof (r_cookie_t) - 1 + hostname_len;
c = MPIU_Malloc (cookie_len);
MPIU_ERR_CHKANDJUMP (c == NULL, mpi_errno, MPI_ERR_OTHER, "**nomem");
c->port = port;
c->data_sz = data_sz;
MPIU_Strncpy (c->hostname, hostname, hostname_len);
*cookie = (char *)c;
*len = sizeof (r_cookie_t) - 1 + hostname_len;
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
void
MPIDI_Win_datatype_map(MPIDI_Datatype * dt)
{
if (dt->contig)
{
dt->num_contig = 1;
dt->map = &dt->__map;
dt->map[0].DLOOP_VECTOR_BUF = (void*)(size_t)dt->true_lb;
dt->map[0].DLOOP_VECTOR_LEN = dt->size;
}
else
{
unsigned map_size = dt->pointer->max_contig_blocks*dt->count + 1;
dt->num_contig = map_size;
dt->map = (DLOOP_VECTOR*)MPIU_Malloc(map_size * sizeof(DLOOP_VECTOR));
MPID_assert(dt->map != NULL);
DLOOP_Offset last = dt->pointer->size*dt->count;
MPID_Segment seg;
MPID_Segment_init(NULL, dt->count, dt->type, &seg, 0);
MPID_Segment_pack_vector(&seg, 0, &last, dt->map, &dt->num_contig);
MPID_assert((unsigned)dt->num_contig <= map_size);
#ifdef TRACE_ON
TRACE_ERR("dt->pointer->size=%d num_contig: orig=%u new=%d\n", dt->pointer->size, map_size, dt->num_contig);
int i;
for(i=0; i<dt->num_contig; ++i)
TRACE_ERR(" %d: BUF=%zu LEN=%zu\n", i, (size_t)dt->map[i].DLOOP_VECTOR_BUF, (size_t)dt->map[i].DLOOP_VECTOR_LEN);
#endif
}
}
int MPIR_Ibsend_impl(const void *buf, int count, MPI_Datatype datatype, int dest, int tag,
MPID_Comm *comm_ptr, MPI_Request *request)
{
int mpi_errno = MPI_SUCCESS;
MPID_Request *request_ptr, *new_request_ptr;
ibsend_req_info *ibinfo=0;
/* We don't try tbsend in for MPI_Ibsend because we must create a
request even if we can send the message */
mpi_errno = MPIR_Bsend_isend( buf, count, datatype, dest, tag, comm_ptr,
IBSEND, &request_ptr );
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
/* FIXME: use the memory management macros */
ibinfo = (ibsend_req_info *)MPIU_Malloc( sizeof(ibsend_req_info) );
ibinfo->req = request_ptr;
ibinfo->cancelled = 0;
mpi_errno = MPIR_Grequest_start_impl( MPIR_Ibsend_query, MPIR_Ibsend_free,
MPIR_Ibsend_cancel, ibinfo, &new_request_ptr );
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* The request is immediately complete because the MPIR_Bsend_isend has
already moved the data out of the user's buffer */
MPIR_Request_add_ref( request_ptr );
/* Request count is now 2 (set to 1 in Grequest_start) */
MPIR_Grequest_complete_impl(new_request_ptr);
MPIU_OBJ_PUBLISH_HANDLE(*request, new_request_ptr->handle);
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
/* This routine is used to establish a queue of send requests to allow the
debugger easier access to the active requests. Some devices may be able
to provide this information without requiring this separate queue. */
void MPIR_Sendq_remember( MPID_Request *req,
int rank, int tag, int context_id )
{
MPIR_Sendq *p;
MPID_THREAD_CS_ENTER(POBJ, req->pobj_mutex);
if (pool) {
p = pool;
pool = p->next;
}
else {
p = (MPIR_Sendq *)MPIU_Malloc( sizeof(MPIR_Sendq) );
if (!p) {
/* Just ignore it */
req->dbg_next = NULL;
goto fn_exit;
}
}
p->sreq = req;
p->tag = tag;
p->rank = rank;
p->context_id = context_id;
p->next = MPIR_Sendq_head;
p->prev = NULL;
MPIR_Sendq_head = p;
if (p->next) p->next->prev = p;
req->dbg_next = p;
fn_exit:
MPID_THREAD_CS_EXIT(POBJ, req->pobj_mutex);
}
/* --BEGIN DEBUG-- */
void MPITEST_Group_create( int nproc, int myrank, MPI_Group *new_group )
{
MPID_Group *new_group_ptr;
int i;
new_group_ptr = (MPID_Group *)MPIU_Handle_obj_alloc( &MPID_Group_mem );
if (!new_group_ptr) {
fprintf( stderr, "Could not create a new group\n" );
PMPI_Abort( MPI_COMM_WORLD, 1 );
}
MPIU_Object_set_ref( new_group_ptr, 1 );
new_group_ptr->lrank_to_lpid = (MPID_Group_pmap_t *)MPIU_Malloc( nproc * sizeof(MPID_Group_pmap_t) );
if (!new_group_ptr->lrank_to_lpid) {
fprintf( stderr, "Could not create lrank map for new group\n" );
PMPI_Abort( MPI_COMM_WORLD, 1 );
}
new_group_ptr->rank = MPI_UNDEFINED;
for (i=0; i<nproc; i++) {
new_group_ptr->lrank_to_lpid[i].lrank = i;
new_group_ptr->lrank_to_lpid[i].lpid = i;
}
new_group_ptr->size = nproc;
new_group_ptr->rank = myrank;
new_group_ptr->idx_of_first_lpid = -1;
*new_group = new_group_ptr->handle;
}
IRLOG_IOStruct *IRLOG_CreateOutputStruct(const char *filename)
{
IRLOG_IOStruct *pOutput = NULL;
/* allocate a data structure */
pOutput = (IRLOG_IOStruct*)MPIU_Malloc(sizeof(IRLOG_IOStruct));
if (pOutput == NULL)
{
MPIU_Error_printf("malloc failed - %s\n", strerror(errno));
return NULL;
}
/* open the output clog file */
pOutput->f = fopen(filename, "wb");
if (pOutput->f == NULL)
{
MPIU_Error_printf("Unable to open output file '%s' - %s\n", filename, strerror(errno));
MPIU_Free(pOutput);
return NULL;
}
/* set all the data fields */
pOutput->header.type = RLOG_INVALID_TYPE;
pOutput->pCurHeader = pOutput->buffer;
pOutput->pNextHeader = pOutput->buffer;
pOutput->pEnd = &pOutput->buffer[RLOG_BUFFSIZE];
return pOutput;
}
int MPIR_Group_create( int nproc, MPID_Group **new_group_ptr )
{
int mpi_errno = MPI_SUCCESS;
*new_group_ptr = (MPID_Group *)MPIU_Handle_obj_alloc( &MPID_Group_mem );
/* --BEGIN ERROR HANDLING-- */
if (!*new_group_ptr) {
mpi_errno = MPIR_Err_create_code( MPI_SUCCESS, MPIR_ERR_RECOVERABLE, "MPIR_Group_create", __LINE__, MPI_ERR_OTHER, "**nomem", 0 );
return mpi_errno;
}
/* --END ERROR HANDLING-- */
MPIU_Object_set_ref( *new_group_ptr, 1 );
(*new_group_ptr)->lrank_to_lpid =
(MPID_Group_pmap_t *)MPIU_Malloc( nproc * sizeof(MPID_Group_pmap_t) );
/* --BEGIN ERROR HANDLING-- */
if (!(*new_group_ptr)->lrank_to_lpid) {
MPIU_Handle_obj_free( &MPID_Group_mem, *new_group_ptr );
*new_group_ptr = NULL;
MPIU_CHKMEM_SETERR(mpi_errno,nproc*sizeof(MPID_Group_pmap_t),
"newgroup->lrank_to_lpid");
return mpi_errno;
}
/* --END ERROR HANDLING-- */
(*new_group_ptr)->size = nproc;
/* Make sure that there is no question that the list of ranks sorted
by pids is marked as uninitialized */
(*new_group_ptr)->idx_of_first_lpid = -1;
(*new_group_ptr)->is_local_dense_monotonic = FALSE;
return mpi_errno;
}
void MPIDI_CH3U_Buffer_allocate(
const void * const sbuf, MPI_Aint scount, MPI_Datatype sdt, int * smpi_errno,
void ** rbuf_handle, MPI_Aint rcount, MPI_Datatype rdt, MPIDI_msg_sz_t * rsz,
int * rmpi_errno)
{
int sdt_contig;
int rdt_contig;
MPI_Aint sdt_true_lb, rdt_true_lb;
MPIDI_msg_sz_t sdata_sz;
MPIDI_msg_sz_t rdata_sz;
MPID_Datatype * sdt_ptr;
MPID_Datatype * rdt_ptr;
MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3U_BUFFER_ALLOCATE);
MPIDI_FUNC_ENTER(MPID_STATE_MPIDI_CH3U_BUFFER_ALLOCATE);
*smpi_errno = MPI_SUCCESS;
*rmpi_errno = MPI_SUCCESS;
MPIDI_Datatype_get_info(scount, sdt, sdt_contig, sdata_sz, sdt_ptr, sdt_true_lb);
MPIDI_Datatype_get_info(rcount, rdt, rdt_contig, rdata_sz, rdt_ptr, rdt_true_lb);
/* --BEGIN ERROR HANDLING-- */
if (sdata_sz > rdata_sz)
{
MPIU_DBG_MSG_FMT(CH3_OTHER,TYPICAL,(MPIU_DBG_FDEST,
"message truncated, sdata_sz=" MPIDI_MSG_SZ_FMT " rdata_sz=" MPIDI_MSG_SZ_FMT,
sdata_sz, rdata_sz));
sdata_sz = rdata_sz;
*rmpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_TRUNCATE, "**truncate", "**truncate %d %d", sdata_sz, rdata_sz );
}
/* --END ERROR HANDLING-- */
if (sdata_sz == 0)
{
*rsz = 0;
goto fn_exit;
}
if (sdt_contig && rdt_contig)
{
*rbuf_handle = (void *)MPIU_Malloc(sdata_sz);
MPIU_Assert(*rbuf_handle);
*rsz = sdata_sz;
}
else
{
/* --BEGIN ERROR HANDLING-- */
MPIU_DBG_MSG(CH3_OTHER,TYPICAL,"Sender and receiver datatypes are not contiguous");
*smpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_FATAL, FCNAME, __LINE__, MPI_ERR_OTHER, "**zcopybufalloc", "**zcopybufalloc %d %d", scount, rcount);
*rmpi_errno = *smpi_errno;
*rsz = 0;
goto fn_exit;
/* --END ERROR HANDLING-- */
}
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_MPIDI_CH3U_BUFFER_ALLOCATE);
}
int
MPID_nem_gm_lmt_init()
{
int mpi_errno = MPI_SUCCESS;
int i;
MPID_nem_gm_lmt_queue.head = NULL;
MPID_nem_gm_lmt_queue.tail = NULL;
MPID_nem_gm_lmt_free_queue = NULL;
for (i = 0; i < FREE_LMT_QUEUE_ELEMENTS; ++i)
{
MPID_nem_gm_lmt_queue_t *e;
e = MPIU_Malloc (sizeof (MPID_nem_gm_lmt_queue_t));
if (e == NULL) MPIU_CHKMEM_SETERR (mpi_errno, sizeof (MPID_nem_gm_send_queue_t), "gm module lmt queue");
e->next = MPID_nem_gm_lmt_free_queue;
MPID_nem_gm_lmt_free_queue = e;
}
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIR_Comm_register_hint(const char *hint_key, MPIR_Comm_hint_fn_t fn, void *state)
{
int mpi_errno = MPI_SUCCESS;
struct MPIR_Comm_hint_fn_elt *hint_elt = NULL;
MPID_MPI_STATE_DECL(MPID_STATE_MPIR_COMM_REGISTER_HINT);
MPID_MPI_FUNC_ENTER(MPID_STATE_MPIR_COMM_REGISTER_HINT);
if (MPID_hint_fns == NULL) {
MPIR_Add_finalize(free_hint_handles, NULL, MPIR_FINALIZE_CALLBACK_PRIO - 1);
}
hint_elt = MPIU_Malloc(sizeof(struct MPIR_Comm_hint_fn_elt));
strncpy(hint_elt->name, hint_key, MPI_MAX_INFO_KEY);
hint_elt->state = state;
hint_elt->fn = fn;
HASH_ADD_STR(MPID_hint_fns, name, hint_elt);
fn_exit:
MPID_MPI_FUNC_EXIT(MPID_STATE_MPIR_COMM_REGISTER_HINT);
return mpi_errno;
fn_fail:
goto fn_exit;
}
/* MSGQUEUE lock is not held */
void
MPIDI_Callback_process_trunc(pami_context_t context,
MPID_Request *rreq,
pami_recv_t *recv,
const void *sndbuf)
{
rreq->status.MPI_ERROR = MPI_ERR_TRUNCATE;
/* -------------------------------------------------------------- */
/* The data is already available, so we can just unpack it now. */
/* -------------------------------------------------------------- */
if (recv)
{
MPIDI_Request_setCA(rreq, MPIDI_CA_UNPACK_UEBUF_AND_COMPLETE);
rreq->mpid.uebuflen = MPIR_STATUS_GET_COUNT(rreq->status);
rreq->mpid.uebuf = MPIU_Malloc(MPIR_STATUS_GET_COUNT(rreq->status));
MPID_assert(rreq->mpid.uebuf != NULL);
rreq->mpid.uebuf_malloc = mpiuMalloc;
recv->addr = rreq->mpid.uebuf;
}
else
{
MPIDI_Request_setCA(rreq, MPIDI_CA_UNPACK_UEBUF_AND_COMPLETE);
rreq->mpid.uebuflen = MPIR_STATUS_GET_COUNT(rreq->status);
rreq->mpid.uebuf = (void*)sndbuf;
MPIDI_RecvDoneCB(context, rreq, PAMI_SUCCESS);
MPID_Request_release(rreq);
}
}
void GenerateNewArgv(int *pargc, char ***pargv, int n)
{
int argc;
char **argv;
int length, i;
char *buffer, *str;
length = (sizeof(char*) * (n+3)) +strlen((*pargv)[0]) + 1 + strlen((*pargv)[1]) + 1 + (15 * n);
buffer = (char*)MPIU_Malloc(length);
argc = n+2;
argv = (char**)buffer;
str = buffer + (sizeof(char*) * (n+4));
argv[0] = str;
str += sprintf(str, "%s", (*pargv)[0]);
*str++ = '\0';
argv[1] = str;
str += sprintf(str, "%s", (*pargv)[1]);
*str++ = '\0';
for (i=0; i<n; i++)
{
argv[i+2] = str;
str += sprintf(str, "log%d.irlog", i);
*str++ = '\0';
}
argv[n+3] = NULL;
*pargc = argc;
*pargv = argv;
s_bFreeArgv = TRUE;
}
int MPID_nem_ib_insert_hash_elem(
MPID_nem_ib_hash_table_ptr_t table,
uint64_t key, void *data, uint32_t uniq)
{
int mpi_errno = MPI_SUCCESS;
uint32_t hash_index;
MPID_nem_ib_hash_elem_ptr_t start_elem;
MPID_nem_ib_hash_elem_ptr_t new_elem;
MPID_nem_ib_hash_elem_ptr_t elem;
MPIU_Assert(NULL != table);
pthread_mutex_lock(&table->hash_table_lock);
hash_index = hash(key, table->num_entries);
/* Note that the first element is allocated
* at the beginning, so this is guaranteed
* to be non-null */
start_elem = &table->entries[hash_index];
MPIU_Assert(start_elem != NULL);
/* Walk to end of list in this hash slot */
elem = start_elem;
while(elem->next != NULL) {
elem = elem->next;
}
/* Insert the element */
new_elem = MPIU_Malloc(sizeof(MPID_nem_ib_hash_elem_t));
if(NULL == new_elem) {
MPIU_CHKMEM_SETERR(mpi_errno,
sizeof(MPID_nem_ib_hash_elem_t),
"IB Module Hash Table New Element");
}
memset(new_elem, 0, sizeof(MPID_nem_ib_hash_elem_t));
new_elem->data = data;
new_elem->uniq = uniq;
new_elem->key = key;
new_elem->next = NULL;
new_elem->prev = elem;
elem->next = new_elem;
NEM_IB_DBG("Inserted elem key %lu, uniq %u, hash index %u",
key, uniq, hash_index);
pthread_mutex_unlock(&table->hash_table_lock);
/* fn_exit: */
return mpi_errno;
/* fn_fail:
goto fn_exit;
*/
}
int MPID_nem_mxm_SendNoncontig(MPIDI_VC_t * vc, MPID_Request * sreq, void *hdr,
MPIDI_msg_sz_t hdr_sz)
{
int mpi_errno = MPI_SUCCESS;
MPIDI_msg_sz_t last;
MPID_nem_mxm_vc_area *vc_area = NULL;
MPID_nem_mxm_req_area *req_area = NULL;
MPIDI_STATE_DECL(MPID_STATE_MPID_NEM_MXM_SENDNONCONTIGMSG);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_NEM_MXM_SENDNONCONTIGMSG);
MPIU_Assert(hdr_sz <= sizeof(MPIDI_CH3_Pkt_t));
MPIU_DBG_MSG(CH3_CHANNEL, VERBOSE, "MPID_nem_mxm_iSendNoncontig");
MPIU_Memcpy(&(sreq->dev.pending_pkt), (char *) hdr, sizeof(MPIDI_CH3_Pkt_t));
_dbg_mxm_output(5,
"SendNoncontig ========> Sending ADI msg (to=%d type=%d) for req %p (data_size %d, %d) \n",
vc->pg_rank, sreq->dev.pending_pkt.type, sreq, sizeof(MPIDI_CH3_Pkt_t),
sreq->dev.segment_size);
vc_area = VC_BASE(vc);
req_area = REQ_BASE(sreq);
req_area->ctx = sreq;
req_area->iov_buf = req_area->tmp_buf;
req_area->iov_count = 1;
req_area->iov_buf[0].ptr = (void *) &(sreq->dev.pending_pkt);
req_area->iov_buf[0].length = sizeof(MPIDI_CH3_Pkt_t);
MPIU_Assert(sreq->dev.segment_first == 0);
last = sreq->dev.segment_size;
if (last > 0) {
sreq->dev.tmpbuf = MPIU_Malloc((size_t) sreq->dev.segment_size);
MPIU_Assert(sreq->dev.tmpbuf);
MPID_Segment_pack(sreq->dev.segment_ptr, sreq->dev.segment_first, &last, sreq->dev.tmpbuf);
MPIU_Assert(last == sreq->dev.segment_size);
req_area->iov_count = 2;
req_area->iov_buf[1].ptr = sreq->dev.tmpbuf;
req_area->iov_buf[1].length = last;
}
vc_area->pending_sends += 1;
sreq->ch.vc = vc;
sreq->ch.noncontig = TRUE;
mpi_errno = _mxm_isend(vc_area->mxm_ep, req_area, MXM_MPICH_ISEND_AM,
mxm_obj->mxm_mq, mxm_obj->mxm_rank, MXM_MPICH_HID_ADI_MSG, 0, 0);
if (mpi_errno)
MPIU_ERR_POP(mpi_errno);
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_MPID_NEM_MXM_SENDNONCONTIGMSG);
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPID_nem_mxm_vc_init(MPIDI_VC_t * vc)
{
int mpi_errno = MPI_SUCCESS;
MPIDI_CH3I_VC *vc_ch = &vc->ch;
MPID_nem_mxm_vc_area *vc_area = VC_BASE(vc);
MPIDI_STATE_DECL(MPID_STATE_MXM_VC_INIT);
MPIDI_FUNC_ENTER(MPID_STATE_MXM_VC_INIT);
/* local connection is used for any source communication */
MPIU_Assert(MPID_nem_mem_region.rank != vc->lpid);
MPIU_DBG_MSG_FMT(CH3_CHANNEL, VERBOSE,
(MPIU_DBG_FDEST,
"[%i]=== connecting to %i \n", MPID_nem_mem_region.rank, vc->lpid));
{
char *business_card;
int val_max_sz;
#ifdef USE_PMI2_API
val_max_sz = PMI2_MAX_VALLEN;
#else
mpi_errno = PMI_KVS_Get_value_length_max(&val_max_sz);
if (mpi_errno)
MPIU_ERR_POP(mpi_errno);
#endif
business_card = (char *) MPIU_Malloc(val_max_sz);
mpi_errno = vc->pg->getConnInfo(vc->pg_rank, business_card, val_max_sz, vc->pg);
if (mpi_errno)
MPIU_ERR_POP(mpi_errno);
vc_area->ctx = vc;
vc_area->mxm_ep = &_mxm_obj.endpoint[vc->pg_rank];
mpi_errno = _mxm_connect(&_mxm_obj.endpoint[vc->pg_rank], business_card, vc_area);
if (mpi_errno)
MPIU_ERR_POP(mpi_errno);
MPIU_Free(business_card);
}
MPIDI_CHANGE_VC_STATE(vc, ACTIVE);
vc_area->pending_sends = 0;
vc->rndvSend_fn = NULL;
vc->rndvRecv_fn = NULL;
vc->sendNoncontig_fn = MPID_nem_mxm_SendNoncontig;
vc->comm_ops = &comm_ops;
vc_ch->iStartContigMsg = MPID_nem_mxm_iStartContigMsg;
vc_ch->iSendContig = MPID_nem_mxm_iSendContig;
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_MXM_VC_INIT);
return mpi_errno;
fn_fail:
goto fn_exit;
}
int
MPID_nem_mx_vc_init (MPIDI_VC_t *vc)
{
uint32_t threshold;
MPIDI_CH3I_VC *vc_ch = VC_CH(vc);
int mpi_errno = MPI_SUCCESS;
/* first make sure that our private fields in the vc fit into the area provided */
MPIU_Assert(sizeof(MPID_nem_mx_vc_area) <= MPID_NEM_VC_NETMOD_AREA_LEN);
#ifdef ONDEMAND
VC_FIELD(vc, local_connected) = 0;
VC_FIELD(vc, remote_connected) = 0;
#else
{
char *business_card;
int val_max_sz;
int ret;
#ifdef USE_PMI2_API
val_max_sz = PMI2_MAX_VALLEN;
#else
mpi_errno = PMI_KVS_Get_value_length_max(&val_max_sz);
#endif
business_card = (char *)MPIU_Malloc(val_max_sz);
mpi_errno = vc->pg->getConnInfo(vc->pg_rank, business_card,val_max_sz, vc->pg);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPID_nem_mx_get_from_bc (business_card, &VC_FIELD(vc, remote_endpoint_id), &VC_FIELD(vc, remote_nic_id));
if (mpi_errno) MPIU_ERR_POP (mpi_errno);
MPIU_Free(business_card);
ret = mx_connect(MPID_nem_mx_local_endpoint,VC_FIELD(vc, remote_nic_id),VC_FIELD(vc, remote_endpoint_id),
MPID_NEM_MX_FILTER,MX_INFINITE,&(VC_FIELD(vc, remote_endpoint_addr)));
MPIU_ERR_CHKANDJUMP1 (ret != MX_SUCCESS, mpi_errno, MPI_ERR_OTHER, "**mx_connect", "**mx_connect %s", mx_strerror (ret));
mx_set_endpoint_addr_context(VC_FIELD(vc, remote_endpoint_addr),(void *)vc);
MPIDI_CHANGE_VC_STATE(vc, ACTIVE);
}
#endif
mx_get_info(MPID_nem_mx_local_endpoint, MX_COPY_SEND_MAX, NULL, 0, &threshold, sizeof(uint32_t));
vc->eager_max_msg_sz = threshold;
vc->rndvSend_fn = NULL;
vc->sendNoncontig_fn = MPID_nem_mx_SendNoncontig;
vc->comm_ops = &comm_ops;
vc_ch->iStartContigMsg = MPID_nem_mx_iStartContigMsg;
vc_ch->iSendContig = MPID_nem_mx_iSendContig;
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
void *MPIDI_Alloc_mem( size_t size, MPID_Info *info_ptr )
{
void *ap;
MPIDI_STATE_DECL(MPID_STATE_MPIDI_ALLOC_MEM);
MPIDI_FUNC_ENTER(MPID_STATE_MPIDI_ALLOC_MEM);
ap = MPIU_Malloc(size);
MPIDI_FUNC_EXIT(MPID_STATE_MPIDI_ALLOC_MEM);
return ap;
}
static int mpi_to_pmi_keyvals( MPID_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 *)MPIU_Malloc( nkeys * sizeof(PMI_keyval_t) );
if (!kv) { MPIU_ERR_POP(mpi_errno); }
for (i=0; i<nkeys; i++) {
mpi_errno = MPIR_Info_get_nthkey_impl( info_ptr, i, key );
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
MPIR_Info_get_valuelen_impl( info_ptr, key, &vallen, &flag );
MPIU_ERR_CHKANDJUMP1(!flag, mpi_errno, MPI_ERR_OTHER,"**infonokey", "**infonokey %s", key);
kv[i].key = MPIU_Strdup(key);
kv[i].val = MPIU_Malloc( vallen + 1 );
if (!kv[i].key || !kv[i].val) {
MPIU_ERR_SETANDJUMP(mpi_errno,MPI_ERR_OTHER,"**nomem" );
}
MPIR_Info_get_impl( info_ptr, key, vallen+1, kv[i].val, &flag );
MPIU_ERR_CHKANDJUMP1(!flag, mpi_errno, MPI_ERR_OTHER,"**infonokey", "**infonokey %s", key);
MPIU_DBG_PRINTF(("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 smpd_hpc_js_init(smpd_hpc_js_ctxt_t *pctxt)
{
HRESULT hr;
int result;
smpd_enter_fn(FCNAME);
if(pctxt == NULL){
smpd_err_printf("ERROR: Invalid pointer to js handle\n");
smpd_exit_fn(FCNAME);
return SMPD_FAIL;
}
/* Alloc memory for scheduler object */
*pctxt = (smpd_hpc_js_ctxt_t )MPIU_Malloc(sizeof(smpd_hpc_js_ctxt_));
if(*pctxt == NULL){
smpd_err_printf("ERROR: Unable to allocate memory for js handle\n");
smpd_exit_fn(FCNAME);
return SMPD_FAIL;
}
(*pctxt)->pscheduler = NULL;
(*pctxt)->pnode_names = NULL;
CoInitializeEx(NULL, COINIT_APARTMENTTHREADED);
/* Get an instance of the Scheduler object */
hr = CoCreateInstance( __uuidof(Scheduler),
NULL,
CLSCTX_INPROC_SERVER,
__uuidof(IScheduler),
reinterpret_cast<void **>(&((*pctxt)->pscheduler)));
CoUninitialize();
if (FAILED(hr)){
smpd_err_printf("ERROR: CoCreateInstance(IScheduler) failed, 0x%x\n", hr);
smpd_exit_fn(FCNAME);
return SMPD_FAIL;
}
/* Connect to the head node */
result = smpd_hpc_js_connect_to_head_node(*pctxt, NULL);
if(result != SMPD_SUCCESS){
smpd_err_printf("Unable to connect to head node \n");
smpd_exit_fn(FCNAME);
return SMPD_FAIL;
}
smpd_exit_fn(FCNAME);
return SMPD_SUCCESS;
}
