int main(int argc, char **argv)
{
pmix_status_t rc;
pmix_value_t value;
pmix_value_t *val = &value;
pmix_proc_t proc;
uint32_t nprocs, n;
pmix_info_t *info, *iptr;
bool flag;
mylock_t mylock;
pmix_data_array_t *dptr;
/* init us - note that the call to "init" includes the return of
* any job-related info provided by the RM. */
if (PMIX_SUCCESS != (rc = PMIx_Init(&myproc, NULL, 0))) {
fprintf(stderr, "Client ns %s rank %d: PMIx_Init failed: %d\n", myproc.nspace, myproc.rank, rc);
exit(0);
}
fprintf(stderr, "Client ns %s rank %d: Running\n", myproc.nspace, myproc.rank);
/* register our default event handler - again, this isn't strictly
* required, but is generally good practice */
DEBUG_CONSTRUCT_LOCK(&mylock);
PMIx_Register_event_handler(NULL, 0, NULL, 0,
notification_fn, evhandler_reg_callbk, (void*)&mylock);
/* wait for registration to complete */
DEBUG_WAIT_THREAD(&mylock);
rc = mylock.status;
DEBUG_DESTRUCT_LOCK(&mylock);
if (PMIX_SUCCESS != rc) {
fprintf(stderr, "[%s:%d] Default handler registration failed\n", myproc.nspace, myproc.rank);
goto done;
}
/* job-related info is found in our nspace, assigned to the
* wildcard rank as it doesn't relate to a specific rank. Setup
* a name to retrieve such values */
PMIX_PROC_CONSTRUCT(&proc);
(void)strncpy(proc.nspace, myproc.nspace, PMIX_MAX_NSLEN);
proc.rank = PMIX_RANK_WILDCARD;
/* get our universe size */
if (PMIX_SUCCESS != (rc = PMIx_Get(&proc, PMIX_UNIV_SIZE, NULL, 0, &val))) {
fprintf(stderr, "Client ns %s rank %d: PMIx_Get universe size failed: %d\n", myproc.nspace, myproc.rank, rc);
goto done;
}
nprocs = val->data.uint32;
PMIX_VALUE_RELEASE(val);
fprintf(stderr, "Client %s:%d universe size %d\n", myproc.nspace, myproc.rank, nprocs);
/* inform the RM that we are preemptible, and that our checkpoint methods are
* "signal" on SIGUSR2 and event on PMIX_JCTRL_CHECKPOINT */
PMIX_INFO_CREATE(info, 2);
flag = true;
PMIX_INFO_LOAD(&info[0], PMIX_JOB_CTRL_PREEMPTIBLE, (void*)&flag, PMIX_BOOL);
/* can't use "load" to load a pmix_data_array_t */
(void)strncpy(info[1].key, PMIX_JOB_CTRL_CHECKPOINT_METHOD, PMIX_MAX_KEYLEN);
PMIX_DATA_ARRAY_CREATE(info[1].value.data.darray, 2, PMIX_INFO);
dptr = info[1].value.data.darray;
rc = SIGUSR2;
iptr = (pmix_info_t*)dptr->array;
PMIX_INFO_LOAD(&iptr[0], PMIX_JOB_CTRL_CHECKPOINT_SIGNAL, &rc, PMIX_INT);
rc = PMIX_JCTRL_CHECKPOINT;
PMIX_INFO_LOAD(&iptr[1], PMIX_JOB_CTRL_CHECKPOINT_EVENT, &rc, PMIX_STATUS);
/* since this is informational and not a requested operation, the target parameter
* doesn't mean anything and can be ignored */
DEBUG_CONSTRUCT_LOCK(&mylock);
if (PMIX_SUCCESS != (rc = PMIx_Job_control_nb(NULL, 0, info, 2, infocbfunc, (void*)&mylock))) {
fprintf(stderr, "Client ns %s rank %d: PMIx_Job_control_nb failed: %d\n", myproc.nspace, myproc.rank, rc);
DEBUG_DESTRUCT_LOCK(&mylock);
goto done;
}
DEBUG_WAIT_THREAD(&mylock);
PMIX_INFO_FREE(info, 2);
rc = mylock.status;
DEBUG_DESTRUCT_LOCK(&mylock);
if (PMIX_SUCCESS != rc) {
fprintf(stderr, "Client ns %s rank %d: PMIx_Job_control_nb failed: %d\n", myproc.nspace, myproc.rank, rc);
goto done;
}
/* now request that this process be monitored using heartbeats */
PMIX_INFO_CREATE(iptr, 1);
PMIX_INFO_LOAD(&iptr[0], PMIX_MONITOR_HEARTBEAT, NULL, PMIX_POINTER);
PMIX_INFO_CREATE(info, 3);
PMIX_INFO_LOAD(&info[0], PMIX_MONITOR_ID, "MONITOR1", PMIX_STRING);
n = 5; // require a heartbeat every 5 seconds
PMIX_INFO_LOAD(&info[1], PMIX_MONITOR_HEARTBEAT_TIME, &n, PMIX_UINT32);
n = 2; // two heartbeats can be missed before declaring us "stalled"
PMIX_INFO_LOAD(&info[2], PMIX_MONITOR_HEARTBEAT_DROPS, &n, PMIX_UINT32);
/* make the request */
DEBUG_CONSTRUCT_LOCK(&mylock);
if (PMIX_SUCCESS != (rc = PMIx_Process_monitor_nb(iptr, PMIX_MONITOR_HEARTBEAT_ALERT,
info, 3, infocbfunc, (void*)&mylock))) {
fprintf(stderr, "Client ns %s rank %d: PMIx_Process_monitor_nb failed: %d\n", myproc.nspace, myproc.rank, rc);
DEBUG_DESTRUCT_LOCK(&mylock);
goto done;
}
DEBUG_WAIT_THREAD(&mylock);
PMIX_INFO_FREE(iptr, 1);
PMIX_INFO_FREE(info, 3);
rc = mylock.status;
DEBUG_DESTRUCT_LOCK(&mylock);
if (PMIX_SUCCESS != rc) {
fprintf(stderr, "Client ns %s rank %d: PMIx_Process_monitor_nb failed: %d\n", myproc.nspace, myproc.rank, rc);
goto done;
}
/* send a heartbeat */
PMIx_Heartbeat();
/* call fence to synchronize with our peers - no need to
* collect any info as we didn't "put" anything */
PMIX_INFO_CREATE(info, 1);
flag = false;
PMIX_INFO_LOAD(info, PMIX_COLLECT_DATA, &flag, PMIX_BOOL);
if (PMIX_SUCCESS != (rc = PMIx_Fence(&proc, 1, info, 1))) {
fprintf(stderr, "Client ns %s rank %d: PMIx_Fence failed: %d\n", myproc.nspace, myproc.rank, rc);
goto done;
}
PMIX_INFO_FREE(info, 1);
done:
/* finalize us */
fprintf(stderr, "Client ns %s rank %d: Finalizing\n", myproc.nspace, myproc.rank);
if (PMIX_SUCCESS != (rc = PMIx_Finalize(NULL, 0))) {
fprintf(stderr, "Client ns %s rank %d:PMIx_Finalize failed: %d\n", myproc.nspace, myproc.rank, rc);
} else {
fprintf(stderr, "Client ns %s rank %d:PMIx_Finalize successfully completed\n", myproc.nspace, myproc.rank);
}
fflush(stderr);
return(0);
}
int main(int argc, char **argv)
{
int rc;
pmix_value_t value;
pmix_value_t *val = &value;
char *tmp;
pmix_proc_t proc;
uint32_t n, num_gets;
bool active;
/* init us */
if (PMIX_SUCCESS != (rc = PMIx_Init(&myproc, NULL, 0))) {
pmix_output(0, "Client ns %s rank %d: PMIx_Init failed: %d", myproc.nspace, myproc.rank, rc);
exit(0);
}
pmix_output(0, "Client ns %s rank %d: Running", myproc.nspace, myproc.rank);
/* get our universe size */
if (PMIX_SUCCESS != (rc = PMIx_Get(&myproc, PMIX_UNIV_SIZE, NULL, 0, &val))) {
pmix_output(0, "Client ns %s rank %d: PMIx_Get universe size failed: %d", myproc.nspace, myproc.rank, rc);
goto done;
}
nprocs = val->data.uint32;
PMIX_VALUE_RELEASE(val);
pmix_output(0, "Client %s:%d universe size %d", myproc.nspace, myproc.rank, nprocs);
/* put a few values */
(void)asprintf(&tmp, "%s-%d-internal", myproc.nspace, myproc.rank);
value.type = PMIX_UINT32;
value.data.uint32 = 1234;
if (PMIX_SUCCESS != (rc = PMIx_Store_internal(&myproc, tmp, &value))) {
pmix_output(0, "Client ns %s rank %d: PMIx_Store_internal failed: %d", myproc.nspace, myproc.rank, rc);
goto done;
}
(void)asprintf(&tmp, "%s-%d-local", myproc.nspace, myproc.rank);
value.type = PMIX_UINT64;
value.data.uint64 = 1234;
if (PMIX_SUCCESS != (rc = PMIx_Put(PMIX_LOCAL, tmp, &value))) {
pmix_output(0, "Client ns %s rank %d: PMIx_Put internal failed: %d", myproc.nspace, myproc.rank, rc);
goto done;
}
(void)asprintf(&tmp, "%s-%d-remote", myproc.nspace, myproc.rank);
value.type = PMIX_STRING;
value.data.string = "1234";
if (PMIX_SUCCESS != (rc = PMIx_Put(PMIX_REMOTE, tmp, &value))) {
pmix_output(0, "Client ns %s rank %d: PMIx_Put internal failed: %d", myproc.nspace, myproc.rank, rc);
goto done;
}
/* introduce a delay by one rank so we can check what happens
* if a "get" is received prior to data being provided */
if (0 == myproc.rank) {
sleep(2);
}
/* commit the data to the server */
if (PMIX_SUCCESS != (rc = PMIx_Commit())) {
pmix_output(0, "Client ns %s rank %d: PMIx_Commit failed: %d", myproc.nspace, myproc.rank, rc);
goto done;
}
/* call fence_nb, but don't return any data */
PMIX_PROC_CONSTRUCT(&proc);
(void)strncpy(proc.nspace, myproc.nspace, PMIX_MAX_NSLEN);
proc.rank = PMIX_RANK_WILDCARD;
active = true;
if (PMIX_SUCCESS != (rc = PMIx_Fence_nb(&proc, 1, NULL, 0, opcbfunc, &active))) {
pmix_output(0, "Client ns %s rank %d: PMIx_Fence failed: %d", myproc.nspace, myproc.rank, rc);
goto done;
}
/* get the committed data - ask for someone who doesn't exist as well */
num_gets = 0;
for (n=0; n < nprocs; n++) {
(void)asprintf(&tmp, "%s-%d-local", myproc.nspace, n);
proc.rank = n;
if (PMIX_SUCCESS != (rc = PMIx_Get_nb(&proc, tmp,
NULL, 0, valcbfunc, tmp))) {
pmix_output(0, "Client ns %s rank %d: PMIx_Get %s failed: %d", myproc.nspace, n, tmp, rc);
goto done;
}
++num_gets;
(void)asprintf(&tmp, "%s-%d-remote", myproc.nspace, n);
if (PMIX_SUCCESS != (rc = PMIx_Get_nb(&proc, tmp,
NULL, 0, valcbfunc, tmp))) {
pmix_output(0, "Client ns %s rank %d: PMIx_Get %s failed: %d", myproc.nspace, n, tmp, rc);
goto done;
}
++num_gets;
}
/* wait for the first fence to finish */
PMIX_WAIT_FOR_COMPLETION(active);
/* wait for all my "get" calls to complete */
while (getcount < num_gets) {
struct timespec ts;
ts.tv_sec = 0;
ts.tv_nsec = 100000;
nanosleep(&ts, NULL);
}
/* call fence again so everyone waits before leaving */
proc.rank = PMIX_RANK_WILDCARD;
if (PMIX_SUCCESS != (rc = PMIx_Fence(&proc, 1, NULL, 0))) {
pmix_output(0, "Client ns %s rank %d: PMIx_Fence failed: %d", myproc.nspace, myproc.rank, rc);
goto done;
}
done:
/* finalize us */
pmix_output(0, "Client ns %s rank %d: Finalizing", myproc.nspace, myproc.rank);
if (PMIX_SUCCESS != (rc = PMIx_Finalize(NULL, 0))) {
fprintf(stderr, "Client ns %s rank %d:PMIx_Finalize failed: %d\n", myproc.nspace, myproc.rank, rc);
} else {
fprintf(stderr, "Client ns %s rank %d:PMIx_Finalize successfully completed\n", myproc.nspace, myproc.rank);
}
fflush(stderr);
return(0);
}
int MPIDU_bc_table_create(int rank, int size, int *nodemap, void *bc, int bc_len, int same_len,
int roots_only, void **bc_table, size_t ** bc_indices)
{
int rc, mpi_errno = MPI_SUCCESS;
int start, end, i;
char *val = NULL, *val_p;
int out_len, val_len, rem, flag;
pmix_value_t value, *pvalue;
pmix_info_t *info;
pmix_proc_t proc;
int local_rank, local_leader;
size_t my_bc_len = bc_len;
MPIR_NODEMAP_get_local_info(rank, size, nodemap, &local_size, &local_rank, &local_leader);
/* if business cards can be different length, use the max value length */
if (!same_len)
bc_len = VALLEN;
mpi_errno = MPIDU_shm_seg_alloc(bc_len * size, (void **) &segment, MPL_MEM_ADDRESS);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
mpi_errno =
MPIDU_shm_seg_commit(&memory, &barrier, local_size, local_rank, local_leader, rank,
MPL_MEM_ADDRESS);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
if (size == 1) {
memcpy(segment, bc, my_bc_len);
goto single;
}
val = MPL_malloc(VALLEN, MPL_MEM_ADDRESS);
memset(val, 0, VALLEN);
val_p = val;
rem = VALLEN;
rc = MPL_str_add_binary_arg(&val_p, &rem, "mpi", (char *) bc, my_bc_len);
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**buscard");
MPIR_Assert(rem >= 0);
if (!roots_only || rank == local_leader) {
value.type = PMIX_STRING;
value.data.string = val;
rc = PMIx_Put(PMIX_LOCAL, "bc", &value);
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**pmix_put");
rc = PMIx_Put(PMIX_REMOTE, "bc", &value);
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**pmix_put");
rc = PMIx_Commit();
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**pmix_commit");
}
PMIX_INFO_CREATE(info, 1);
PMIX_INFO_LOAD(info, PMIX_COLLECT_DATA, &flag, PMIX_BOOL);
rc = PMIx_Fence(&MPIR_Process.pmix_wcproc, 1, info, 1);
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**pmix_fence");
PMIX_INFO_FREE(info, 1);
if (!roots_only) {
start = local_rank * (size / local_size);
end = start + (size / local_size);
if (local_rank == local_size - 1)
end += size % local_size;
for (i = start; i < end; i++) {
PMIX_PROC_CONSTRUCT(&proc);
MPL_strncpy(proc.nspace, MPIR_Process.pmix_proc.nspace, PMIX_MAX_NSLEN);
proc.rank = i;
rc = PMIx_Get(&proc, "bc", NULL, 0, &pvalue);
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**pmix_get");
rc = MPL_str_get_binary_arg(val, "mpi", &segment[i * bc_len], bc_len, &out_len);
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**argstr_missinghost");
PMIX_VALUE_RELEASE(pvalue);
}
} else {
int num_nodes, *node_roots;
MPIR_NODEMAP_get_node_roots(nodemap, size, &node_roots, &num_nodes);
start = local_rank * (num_nodes / local_size);
end = start + (num_nodes / local_size);
if (local_rank == local_size - 1)
end += num_nodes % local_size;
for (i = start; i < end; i++) {
PMIX_PROC_CONSTRUCT(&proc);
MPL_strncpy(proc.nspace, MPIR_Process.pmix_proc.nspace, PMIX_MAX_NSLEN);
proc.rank = i;
rc = PMIx_Get(&proc, "bc", NULL, 0, &pvalue);
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**pmix_get");
rc = MPL_str_get_binary_arg(val, "mpi", &segment[i * bc_len], bc_len, &out_len);
MPIR_ERR_CHKANDJUMP(rc, mpi_errno, MPI_ERR_OTHER, "**argstr_missinghost");
PMIX_VALUE_RELEASE(pvalue);
}
}
mpi_errno = MPIDU_shm_barrier(barrier, local_size);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
single:
if (!same_len) {
indices = MPL_malloc(size * sizeof(size_t), MPL_MEM_ADDRESS);
for (i = 0; i < size; i++)
indices[i] = bc_len * i;
*bc_indices = indices;
}
fn_exit:
MPL_free(val);
*bc_table = segment;
return mpi_errno;
fn_fail:
goto fn_exit;
}
int main(int argc, char **argv)
{
int rc;
pmix_value_t value;
pmix_value_t *val = &value;
pmix_proc_t proc;
uint32_t nprocs;
char nsp2[PMIX_MAX_NSLEN+1];
pmix_app_t *app;
char hostname[1024];
pmix_proc_t *peers;
size_t npeers, ntmp=0;
char *nodelist;
gethostname(hostname, 1024);
/* init us */
if (PMIX_SUCCESS != (rc = PMIx_Init(&myproc))) {
fprintf(stderr, "Client ns %s rank %d: PMIx_Init failed: %d\n", myproc.nspace, myproc.rank, rc);
exit(0);
}
fprintf(stderr, "Client ns %s rank %d: Running\n", myproc.nspace, myproc.rank);
/* get our universe size */
if (PMIX_SUCCESS != (rc = PMIx_Get(&myproc, PMIX_UNIV_SIZE, NULL, 0, &val))) {
fprintf(stderr, "Client ns %s rank %d: PMIx_Get universe size failed: %d\n", myproc.nspace, myproc.rank, rc);
goto done;
}
nprocs = val->data.uint32;
PMIX_VALUE_RELEASE(val);
fprintf(stderr, "Client %s:%d universe size %d\n", myproc.nspace, myproc.rank, nprocs);
/* call fence to sync */
PMIX_PROC_CONSTRUCT(&proc);
(void)strncpy(proc.nspace, myproc.nspace, PMIX_MAX_NSLEN);
proc.rank = PMIX_RANK_WILDCARD;
if (PMIX_SUCCESS != (rc = PMIx_Fence(&proc, 1, NULL, 0))) {
fprintf(stderr, "Client ns %s rank %d: PMIx_Fence failed: %d\n", myproc.nspace, myproc.rank, rc);
goto done;
}
/* rank=0 calls spawn */
if (0 == myproc.rank) {
PMIX_APP_CREATE(app, 1);
app->cmd = strdup("gumby");
app->maxprocs = 2;
app->argc = 3;
app->argv = (char**)malloc(4 * sizeof(char*));
app->argv[0] = strdup("gumby");
app->argv[1] = strdup("-n");
app->argv[2] = strdup("2");
app->argv[3] = NULL;
app->env = (char**)malloc(2 * sizeof(char*));
app->env[0] = strdup("PMIX_ENV_VALUE=3");
app->env[1] = NULL;
PMIX_INFO_CREATE(app->info, 2);
(void)strncpy(app->info[0].key, "DARTH", PMIX_MAX_KEYLEN);
app->info[0].value.type = PMIX_INT8;
app->info[0].value.data.int8 = 12;
(void)strncpy(app->info[1].key, "VADER", PMIX_MAX_KEYLEN);
app->info[1].value.type = PMIX_DOUBLE;
app->info[1].value.data.dval = 12.34;
fprintf(stderr, "Client ns %s rank %d: calling PMIx_Spawn\n", myproc.nspace, myproc.rank);
if (PMIX_SUCCESS != (rc = PMIx_Spawn(NULL, 0, app, 1, nsp2))) {
fprintf(stderr, "Client ns %s rank %d: PMIx_Spawn failed: %d\n", myproc.nspace, myproc.rank, rc);
goto done;
}
PMIX_APP_FREE(app, 1);
/* check to see if we got the expected info back */
if (0 != strncmp(nsp2, "DYNSPACE", PMIX_MAX_NSLEN)) {
fprintf(stderr, "Client ns %s rank %d: PMIx_Spawn returned incorrect nspace: %s\n", myproc.nspace, myproc.rank, nsp2);
goto done;
} else {
fprintf(stderr, "Client ns %s rank %d: PMIx_Spawn succeeded returning nspace: %s\n", myproc.nspace, myproc.rank, nsp2);
}
/* get their universe size */
val = NULL;
(void)strncpy(proc.nspace, nsp2, PMIX_MAX_NSLEN);
proc.rank = PMIX_RANK_WILDCARD;
if (PMIX_SUCCESS != (rc = PMIx_Get(&proc, PMIX_UNIV_SIZE, NULL, 0, &val)) ||
NULL == val) {
fprintf(stderr, "Client ns %s rank %d: PMIx_Get universe size failed: %d\n", myproc.nspace, myproc.rank, rc);
goto done;
}
ntmp = val->data.uint32;
PMIX_VALUE_RELEASE(val);
fprintf(stderr, "Client %s:%d universe %s size %d\n", myproc.nspace, myproc.rank, nsp2, (int)ntmp);
}
/* just cycle the connect/disconnect functions */
(void)strncpy(proc.nspace, myproc.nspace, PMIX_MAX_NSLEN);
proc.rank = PMIX_RANK_WILDCARD;
if (PMIX_SUCCESS != (rc = PMIx_Connect(&proc, 1, NULL, 0))) {
fprintf(stderr, "Client ns %s rank %d: PMIx_Connect failed: %d\n", myproc.nspace, myproc.rank, rc);
goto done;
}
fprintf(stderr, "Client ns %s rank %d: PMIx_Connect succeeded\n", myproc.nspace, myproc.rank);
if (PMIX_SUCCESS != (rc = PMIx_Disconnect(&proc, 1, NULL, 0))) {
fprintf(stderr, "Client ns %s rank %d: PMIx_Disonnect failed: %d\n", myproc.nspace, myproc.rank, rc);
goto done;
}
fprintf(stderr, "Client ns %s rank %d: PMIx_Disconnect succeeded\n", myproc.nspace, myproc.rank);
/* finally, test the resolve functions */
if (0 == myproc.rank) {
if (PMIX_SUCCESS != (rc = PMIx_Resolve_peers(hostname, NULL, &peers, &npeers))) {
fprintf(stderr, "Client ns %s rank %d: PMIx_Resolve_peers failed for nspace %s: %d\n", myproc.nspace, myproc.rank, nsp2, rc);
goto done;
}
if ((nprocs+ntmp) != npeers) {
fprintf(stderr, "Client ns %s rank %d: PMIx_Resolve_peers returned incorrect npeers: %d vs %d\n", myproc.nspace, myproc.rank, (int)(nprocs+ntmp), (int)npeers);
goto done;
}
fprintf(stderr, "Client ns %s rank %d: PMIx_Resolve_peers returned %d npeers\n", myproc.nspace, myproc.rank, (int)npeers);
if (PMIX_SUCCESS != (rc = PMIx_Resolve_nodes(nsp2, &nodelist))) {
fprintf(stderr, "Client ns %s rank %d: PMIx_Resolve_nodes failed for nspace %s: %d\n", myproc.nspace, myproc.rank, nsp2, rc);
goto done;
}
fprintf(stderr, "Client ns %s rank %d: PMIx_Resolve_nodes %s", myproc.nspace, myproc.rank, nodelist);
} else {
if (PMIX_SUCCESS != (rc = PMIx_Resolve_peers(hostname, myproc.nspace, &peers, &npeers))) {
fprintf(stderr, "Client ns %s rank %d: PMIx_Resolve_peers failed for nspace %s: %d\n", myproc.nspace, myproc.rank, myproc.nspace, rc);
goto done;
}
if (nprocs != npeers) {
fprintf(stderr, "Client ns %s rank %d: PMIx_Resolve_peers returned incorrect npeers: %d vs %d\n", myproc.nspace, myproc.rank, nprocs, (int)npeers);
goto done;
}
fprintf(stderr, "Client ns %s rank %d: PMIx_Resolve_peers returned %d npeers\n", myproc.nspace, myproc.rank, (int)npeers);
if (PMIX_SUCCESS != (rc = PMIx_Resolve_nodes(myproc.nspace, &nodelist))) {
fprintf(stderr, "Client ns %s rank %d: PMIx_Resolve_nodes failed: %d\n", myproc.nspace, myproc.rank, rc);
goto done;
}
fprintf(stderr, "Client ns %s rank %d: PMIx_Resolve_nodes %s\n", myproc.nspace, myproc.rank, nodelist);
}
PMIX_PROC_FREE(peers, npeers);
free(nodelist);
done:
/* call fence to sync */
PMIX_PROC_CONSTRUCT(&proc);
(void)strncpy(proc.nspace, myproc.nspace, PMIX_MAX_NSLEN);
proc.rank = PMIX_RANK_WILDCARD;
if (PMIX_SUCCESS != (rc = PMIx_Fence(&proc, 1, NULL, 0))) {
fprintf(stderr, "Client ns %s rank %d: PMIx_Fence failed: %d\n", myproc.nspace, myproc.rank, rc);
goto done;
}
/* finalize us */
fprintf(stderr, "Client ns %s rank %d: Finalizing\n", myproc.nspace, myproc.rank);
if (PMIX_SUCCESS != (rc = PMIx_Finalize())) {
fprintf(stderr, "Client ns %s rank %d:PMIx_Finalize failed: %d\n", myproc.nspace, myproc.rank, rc);
} else {
fprintf(stderr, "Client ns %s rank %d:PMIx_Finalize successfully completed\n", myproc.nspace, myproc.rank);
}
fflush(stderr);
return(0);
}
int main(int argc, char **argv)
{
int rc;
pmix_value_t value;
pmix_value_t *val = &value;
pmix_proc_t proc;
uint32_t nprocs;
/* init us */
if (PMIX_SUCCESS != (rc = PMIx_Init(&myproc, NULL, 0))) {
fprintf(stderr, "Client ns %s rank %d: PMIx_Init failed: %d\n", myproc.nspace, myproc.rank, rc);
exit(0);
}
fprintf(stderr, "Client ns %s rank %d: Running\n", myproc.nspace, myproc.rank);
/* get our universe size */
if (PMIX_SUCCESS != (rc = PMIx_Get(&myproc, PMIX_UNIV_SIZE, NULL, 0, &val))) {
fprintf(stderr, "Client ns %s rank %d: PMIx_Get universe size failed: %d\n", myproc.nspace, myproc.rank, rc);
goto done;
}
nprocs = val->data.uint32;
PMIX_VALUE_RELEASE(val);
fprintf(stderr, "Client %s:%d universe size %d\n", myproc.nspace, myproc.rank, nprocs);
completed = false;
/* register our errhandler */
PMIx_Register_event_handler(NULL, 0, NULL, 0,
notification_fn, errhandler_reg_callbk, NULL);
/* call fence to sync */
PMIX_PROC_CONSTRUCT(&proc);
(void)strncpy(proc.nspace, myproc.nspace, PMIX_MAX_NSLEN);
proc.rank = PMIX_RANK_WILDCARD;
if (PMIX_SUCCESS != (rc = PMIx_Fence(&proc, 1, NULL, 0))) {
fprintf(stderr, "Client ns %s rank %d: PMIx_Fence failed: %d\n", myproc.nspace, myproc.rank, rc);
goto done;
}
/* rank=0 calls abort */
if (0 == myproc.rank) {
PMIx_Abort(PMIX_ERR_OUT_OF_RESOURCE, "Eat rocks",
&proc, 1);
fprintf(stderr, "Client ns %s rank %d: Abort called\n", myproc.nspace, myproc.rank);
}
/* everyone simply waits */
while (!completed) {
struct timespec ts;
ts.tv_sec = 0;
ts.tv_nsec = 100000;
nanosleep(&ts, NULL);
}
done:
/* finalize us */
fprintf(stderr, "Client ns %s rank %d: Finalizing\n", myproc.nspace, myproc.rank);
PMIx_Deregister_event_handler(1, op_callbk, NULL);
if (PMIX_SUCCESS != (rc = PMIx_Finalize(NULL, 0))) {
fprintf(stderr, "Client ns %s rank %d:PMIx_Finalize failed: %d\n", myproc.nspace, myproc.rank, rc);
} else {
fprintf(stderr, "Client ns %s rank %d:PMIx_Finalize successfully completed\n", myproc.nspace, myproc.rank);
}
fflush(stderr);
return(0);
}
int main(int argc, char **argv)
{
pmix_status_t rc;
pmix_value_t *val;
pmix_proc_t proc;
pmix_info_t *info;
size_t ninfo;
pmix_query_t *query;
size_t nq, n;
myquery_data_t myquery_data;
pid_t pid;
pmix_status_t code = PMIX_ERR_JOB_TERMINATED;
mylock_t mylock;
myrel_t myrel;
uint16_t localrank;
char *target = NULL;
pid = getpid();
/* init us - since we were launched by the RM, our connection info
* will have been provided at startup. */
if (PMIX_SUCCESS != (rc = PMIx_tool_init(&myproc, NULL, 0))) {
fprintf(stderr, "Debugger daemon: PMIx_tool_init failed: %d\n", rc);
exit(0);
}
fprintf(stderr, "Debugger daemon ns %s rank %d pid %lu: Running\n", myproc.nspace, myproc.rank, (unsigned long)pid);
/* register our default event handler */
DEBUG_CONSTRUCT_LOCK(&mylock);
PMIx_Register_event_handler(NULL, 0, NULL, 0,
notification_fn, evhandler_reg_callbk, (void*)&mylock);
DEBUG_WAIT_THREAD(&mylock);
if (PMIX_SUCCESS != mylock.status) {
rc = mylock.status;
DEBUG_DESTRUCT_LOCK(&mylock);
goto done;
}
DEBUG_DESTRUCT_LOCK(&mylock);
/* get the nspace of the job we are to debug - it will be in our JOB info */
#ifdef PMIX_LOAD_PROCID
PMIX_LOAD_PROCID(&proc, myproc.nspace, PMIX_RANK_WILDCARD);
#else
PMIX_PROC_CONSTRUCT(&proc);
(void)strncpy(proc.nspace, myproc.nspace, PMIX_MAX_KEYLEN);
proc.rank = PMIX_RANK_WILDCARD;
#endif
if (PMIX_SUCCESS != (rc = PMIx_Get(&proc, PMIX_DEBUG_JOB, NULL, 0, &val))) {
fprintf(stderr, "[%s:%d:%lu] Failed to get job being debugged - error %s\n",
myproc.nspace, myproc.rank,
(unsigned long)pid, PMIx_Error_string(rc));
goto done;
}
if (NULL == val || PMIX_STRING != val->type || NULL == val->data.string) {
fprintf(stderr, "[%s:%d:%lu] Failed to get job being debugged - NULL data returned\n",
myproc.nspace, myproc.rank, (unsigned long)pid);
goto done;
}
/* save it for later */
target = strdup(val->data.string);
PMIX_VALUE_RELEASE(val);
fprintf(stderr, "[%s:%d:%lu] Debugging %s\n", myproc.nspace, myproc.rank,
(unsigned long)pid, target);
/* get my local rank so I can determine which local proc is "mine"
* to debug */
val = NULL;
if (PMIX_SUCCESS != (rc = PMIx_Get(&myproc, PMIX_LOCAL_RANK, NULL, 0, &val))) {
fprintf(stderr, "[%s:%d:%lu] Failed to get my local rank - error %s\n",
myproc.nspace, myproc.rank,
(unsigned long)pid, PMIx_Error_string(rc));
goto done;
}
if (NULL == val) {
fprintf(stderr, "[%s:%d:%lu] Failed to get my local rank - NULL data returned\n",
myproc.nspace, myproc.rank, (unsigned long)pid);
goto done;
}
if (PMIX_UINT16 != val->type) {
fprintf(stderr, "[%s:%d:%lu] Failed to get my local rank - returned wrong type %s\n",
myproc.nspace, myproc.rank, (unsigned long)pid, PMIx_Data_type_string(val->type));
goto done;
}
/* save the data */
localrank = val->data.uint16;
PMIX_VALUE_RELEASE(val);
fprintf(stderr, "[%s:%d:%lu] my local rank %d\n", myproc.nspace, myproc.rank,
(unsigned long)pid, (int)localrank);
/* register another handler specifically for when the target
* job completes */
DEBUG_CONSTRUCT_LOCK(&myrel.lock);
myrel.nspace = strdup(proc.nspace);
PMIX_INFO_CREATE(info, 2);
PMIX_INFO_LOAD(&info[0], PMIX_EVENT_RETURN_OBJECT, &myrel, PMIX_POINTER);
/* only call me back when this specific job terminates */
PMIX_LOAD_PROCID(&proc, target, PMIX_RANK_WILDCARD);
PMIX_INFO_LOAD(&info[1], PMIX_EVENT_AFFECTED_PROC, &proc, PMIX_PROC);
fprintf(stderr, "[%s:%d:%lu] registering for termination of %s\n", myproc.nspace, myproc.rank,
(unsigned long)pid, proc.nspace);
DEBUG_CONSTRUCT_LOCK(&mylock);
PMIx_Register_event_handler(&code, 1, info, 2,
release_fn, evhandler_reg_callbk, (void*)&mylock);
DEBUG_WAIT_THREAD(&mylock);
if (PMIX_SUCCESS != mylock.status) {
rc = mylock.status;
DEBUG_DESTRUCT_LOCK(&mylock);
PMIX_INFO_FREE(info, 2);
goto done;
}
DEBUG_DESTRUCT_LOCK(&mylock);
PMIX_INFO_FREE(info, 2);
/* get our local proctable - for scalability reasons, we don't want to
* have our "root" debugger process get the proctable for everybody and
* send it out to us. So ask the local PMIx server for the pid's of
* our local target processes */
nq = 1;
PMIX_QUERY_CREATE(query, nq);
PMIX_ARGV_APPEND(rc, query[0].keys, PMIX_QUERY_LOCAL_PROC_TABLE);
query[0].nqual = 1;
PMIX_INFO_CREATE(query[0].qualifiers, 1);
PMIX_INFO_LOAD(&query[0].qualifiers[0], PMIX_NSPACE, target, PMIX_STRING); // the nspace we are enquiring about
/* setup the caddy to retrieve the data */
DEBUG_CONSTRUCT_LOCK(&myquery_data.lock);
myquery_data.info = NULL;
myquery_data.ninfo = 0;
/* execute the query */
if (PMIX_SUCCESS != (rc = PMIx_Query_info_nb(query, nq, cbfunc, (void*)&myquery_data))) {
fprintf(stderr, "PMIx_Query_info failed: %d\n", rc);
goto done;
}
DEBUG_WAIT_THREAD(&myquery_data.lock);
DEBUG_DESTRUCT_LOCK(&myquery_data.lock);
PMIX_QUERY_FREE(query, nq);
if (PMIX_SUCCESS != myquery_data.status) {
rc = myquery_data.status;
goto done;
}
fprintf(stderr, "[%s:%d:%lu] Local proctable received\n", myproc.nspace, myproc.rank, (unsigned long)pid);
/* now that we have the proctable for our local processes, we can do our
* magic debugger stuff and attach to them. We then send a "release" event
* to them - i.e., it's the equivalent to setting the MPIR breakpoint. We
* do this with the event notification system. For this example, we just
* send it to all local procs of the job being debugged */
(void)strncpy(proc.nspace, target, PMIX_MAX_NSLEN);
proc.rank = PMIX_RANK_WILDCARD;
ninfo = 2;
PMIX_INFO_CREATE(info, ninfo);
PMIX_INFO_LOAD(&info[0], PMIX_EVENT_CUSTOM_RANGE, &proc, PMIX_PROC); // deliver to the target nspace
PMIX_INFO_LOAD(&info[1], PMIX_EVENT_NON_DEFAULT, NULL, PMIX_BOOL); // deliver to the target nspace
fprintf(stderr, "[%s:%u:%lu] Sending release\n", myproc.nspace, myproc.rank, (unsigned long)pid);
rc = PMIx_Notify_event(PMIX_ERR_DEBUGGER_RELEASE,
NULL, PMIX_RANGE_CUSTOM,
info, ninfo, NULL, NULL);
if (PMIX_SUCCESS != rc) {
fprintf(stderr, "%s[%s:%u:%lu] Sending release failed with error %s(%d)\n",
myproc.nspace, myproc.rank, (unsigned long)pid, PMIx_Error_string(rc), rc);
goto done;
}
/* do some debugger magic while waiting for the job to terminate */
DEBUG_WAIT_THREAD(&myrel.lock);
done:
if (NULL != target) {
free(target);
}
/* finalize us */
fprintf(stderr, "Debugger daemon ns %s rank %d pid %lu: Finalizing\n", myproc.nspace, myproc.rank, (unsigned long)pid);
if (PMIX_SUCCESS != (rc = PMIx_Finalize(NULL, 0))) {
fprintf(stderr, "Debugger daemon ns %s rank %d:PMIx_Finalize failed: %d\n", myproc.nspace, myproc.rank, rc);
} else {
fprintf(stderr, "Debugger daemon ns %s rank %d pid %lu:PMIx_Finalize successfully completed\n", myproc.nspace, myproc.rank, (unsigned long)pid);
}
fflush(stderr);
return(0);
}
