/* gasneti_bootstrapFini
*/
void gasneti_bootstrapFini_pmi(void) {
#if USE_PMI2_API
PMI2_Finalize();
#else
if (PMI_FALSE == gasneti_pmi_initialized) {
PMI_Finalize();
}
#endif
}
void mca_common_pmi_finalize (void) {
if (0 == mca_common_pmi_init_count) {
return;
}
if (0 == --mca_common_pmi_init_count) {
#if WANT_PMI2_SUPPORT
PMI2_Finalize ();
#else
PMI_Finalize ();
#endif
}
}
void mca_common_pmi_finalize (void) {
if (0 == mca_common_pmi_init_count) {
return;
}
if (0 == --mca_common_pmi_init_count) {
#if WANT_PMI2_SUPPORT
if( mca_common_pmi_version == 2){
PMI2_Finalize ();
}
else
#endif
{
PMI_Finalize ();
}
}
}
static int s2_fini(void) {
if (0 == pmix_init_count) {
return OPAL_SUCCESS;
}
if (0 == --pmix_init_count) {
PMI2_Finalize();
}
if (NULL != pmix_kvs_name) {
free(pmix_kvs_name);
pmix_kvs_name = NULL;
}
if (NULL != s2_lranks) {
free(s2_lranks);
}
return OPAL_SUCCESS;
}
int orte_grpcomm_pmi_close(void)
{
#if WANT_CRAY_PMI2_EXT
if (started_by_me && PMI2_Initialized()) {
PMI2_Finalize();
}
#else
PMI_BOOL initialized;
/* if we weren't selected, cleanup */
if (started_by_me &&
PMI_SUCCESS == PMI_Initialized(&initialized) &&
PMI_TRUE == initialized) {
PMI_Finalize();
}
#endif
return ORTE_SUCCESS;
}
int
main(int argc, char **argv)
{
int spawned, size, rank, appnum;
struct timeval tv, tv2;
int ring_rank, ring_size;
char val[128];
char buf[128];
char left[128];
char right[128];
{
int x = 0;
while (x) {
fprintf(stderr, "attachme %d\n", getpid());
sleep(2);
}
}
gettimeofday(&tv, NULL);
PMI2_Init(&spawned, &size, &rank, &appnum);
/* test PMIX_Ring */
snprintf(val, sizeof(val), "pmi_rank=%d", rank);
PMIX_Ring(val, &ring_rank, &ring_size, left, right, 128);
printf("pmi_rank:%d ring_rank:%d ring_size:%d left:%s mine:%s right:%s\n",
rank, ring_rank, ring_size, left, val, right);
PMI2_Finalize();
gettimeofday(&tv2, NULL);
printf("%f\n",
((tv2.tv_sec - tv.tv_sec) * 1000.0
+ (tv2.tv_usec - tv.tv_usec) / 1000.0));
return 0;
}
int main(int argc, char **argv)
{
int ret = 0;
int rc;
char *str = NULL;
int ti = (argc > 1 ? atoi(argv[1]) : 0);
srand(time(NULL));
str = getenv("VERBOSE");
_verbose = (str ? atoi(str) : _verbose);
str = getenv("LEGACY");
_legacy = (str ? atoi(str) : _legacy);
spawned = random_value(10, 20);
size = random_value(10, 20);
rank = random_value(10, 20);
appnum = random_value(10, 20);
if (PMI2_SUCCESS != (rc = PMI2_Init(&spawned, &size, &rank, &appnum))) {
log_fatal("PMI2_Init failed: %d\n", rc);
return rc;
}
if (!ti || 1 == ti) {
rc = test_item1();
ret += (rc ? 1 : 0);
log_info("TI1 : %s\n", (rc ? "FAIL" : "PASS"));
}
if (!ti || 2 == ti) {
rc = test_item2();
ret += (rc ? 1 : 0);
log_info("TI2 : %s\n", (rc ? "FAIL" : "PASS"));
}
if (!ti || 3 == ti) {
rc = test_item3();
ret += (rc ? 1 : 0);
log_info("TI3 : %s\n", (rc ? "FAIL" : "PASS"));
}
if (!ti || 4 == ti) {
rc = test_item4();
ret += (rc ? 1 : 0);
log_info("TI4 : %s\n", (rc ? "FAIL" : "PASS"));
}
if (!ti || 5 == ti) {
rc = test_item5();
ret += (rc ? 1 : 0);
log_info("TI5 : %s\n", (rc ? "FAIL" : "PASS"));
}
if (!ti || 6 == ti) {
rc = test_item6();
ret += (rc ? 1 : 0);
log_info("TI6 : %s\n", (rc ? "FAIL" : "PASS"));
}
if (!ti || 7 == ti) {
rc = test_item7();
ret += (rc ? 1 : 0);
log_info("TI7 : %s\n", (rc ? "FAIL" : "PASS"));
}
if (!ti || 8 == ti) {
rc = test_item8();
ret += (rc ? 1 : 0);
log_info("TI8 : %s\n", (rc ? "FAIL" : "PASS"));
}
if (!ti || 9 == ti) {
rc = test_item9();
ret += (rc ? 1 : 0);
log_info("TI9 : %s\n", (rc ? "FAIL" : "PASS"));
}
if (PMI2_SUCCESS != (rc = PMI2_Finalize())) {
log_fatal("PMI2_Finalize failed: %d\n", rc);
return rc;
}
return ret;
}
int
main(int argc, char **argv)
{
int rank;
int size;
int appnum;
int spawned;
int flag;
int len;
int i;
struct timeval tv;
struct timeval tv2;
char jobid[128];
char key[128];
char val[128];
char buf[128];
{
int x = 1;
while (x == 0) {
sleep(2);
}
}
gettimeofday(&tv, NULL);
srand(tv.tv_sec);
PMI2_Init(&spawned, &size, &rank, &appnum);
PMI2_Job_GetId(jobid, sizeof(buf));
memset(val, 0, sizeof(val));
PMI2_Info_GetJobAttr("mpi_reserved_ports",
val,
PMI2_MAX_ATTRVALUE,
&flag);
sprintf(key, "mpi_reserved_ports");
PMI2_KVS_Put(key, val);
memset(val, 0, sizeof(val));
sprintf(buf, "PMI_netinfo_of_task");
PMI2_Info_GetJobAttr(buf,
val,
PMI2_MAX_ATTRVALUE,
&flag);
sprintf(key, buf);
PMI2_KVS_Put(key, val);
memset(val, 0, sizeof(val));
sprintf(key, "david@%d", rank);
sprintf(val, "%s", mrand(97, 122));
PMI2_KVS_Put(key, val);
PMI2_KVS_Fence();
for (i = 0; i < size; i++) {
memset(val, 0, sizeof(val));
sprintf(key, "PMI_netinfo_of_task");
PMI2_KVS_Get(jobid,
PMI2_ID_NULL,
key,
val,
sizeof(val),
&len);
printf("rank: %d key:%s val:%s\n", rank, key, val);
memset(val, 0, sizeof(val));
sprintf(key, "david@%d", rank);
PMI2_KVS_Get(jobid,
PMI2_ID_NULL,
key,
val,
sizeof(val),
&len);
printf("rank: %d key:%s val:%s\n", rank, key, val);
memset(val, 0, sizeof(val));
sprintf(key, "mpi_reserved_ports");
PMI2_KVS_Get(jobid,
PMI2_ID_NULL,
key,
val,
sizeof(val),
&len);
printf("rank: %d key:%s val:%s\n", rank, key, val);
}
PMI2_Finalize();
gettimeofday(&tv2, NULL);
printf("%f\n",
((tv2.tv_sec - tv.tv_sec) * 1000.0
+ (tv2.tv_usec - tv.tv_usec) / 1000.0));
return 0;
}
void pmi_fini()
{
PMI2_Finalize();
}
static int cray_init(void)
{
int i, spawned, size, rank, appnum, my_node;
int rc, ret = OPAL_ERROR;
char *pmapping = NULL;
char buf[PMI2_MAX_ATTRVALUE];
int found;
uint32_t jobfam;
++pmix_init_count;
/* if we can't startup PMI, we can't be used */
if ( PMI2_Initialized () ) {
return OPAL_SUCCESS;
}
size = -1;
rank = -1;
appnum = -1;
if (PMI_SUCCESS != (rc = PMI2_Init(&spawned, &size, &rank, &appnum))) {
opal_show_help("help-pmix-base.txt", "pmix2-init-failed", true, rc);
return OPAL_ERROR;
}
if( size < 0 || rank < 0 ){
opal_show_help("help-pmix-base.txt", "pmix2-init-returned-bad-values", true);
goto err_exit;
}
pmix_size = size;
pmix_rank = rank;
pmix_appnum = appnum;
pmix_vallen_max = PMI2_MAX_VALLEN;
pmix_kvslen_max = PMI2_MAX_VALLEN; // FIX ME: What to put here for versatility?
pmix_keylen_max = PMI2_MAX_KEYLEN;
rc = PMI2_Info_GetJobAttr("universeSize", buf, 16, &found);
if( PMI_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI_Get_universe_size");
goto err_exit;
}
pmix_usize = atoi(buf);
pmix_kvs_name = (char*)malloc(pmix_kvslen_max);
if( pmix_kvs_name == NULL ){
PMI2_Finalize();
ret = OPAL_ERR_OUT_OF_RESOURCE;
goto err_exit;
}
rc = PMI2_Job_GetId(pmix_kvs_name, pmix_kvslen_max);
if( PMI_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI2_Job_GetId");
goto err_exit;
}
rc = sscanf(pmix_kvs_name,"kvs_%u",&jobfam);
if (rc != 1) {
OPAL_PMI_ERROR(rc, "PMI2_Job_GetId");
rc = OPAL_ERROR;
goto err_exit;
}
pmix_jobid = jobfam << 16;
/* store our name in the opal_proc_t so that
* debug messages will make sense - an upper
* layer will eventually overwrite it, but that
* won't do any harm */
pmix_pname.jid = pmix_jobid;
pmix_pname.vid = pmix_rank;
opal_proc_set_name((opal_process_name_t*)&pmix_pname);
opal_output_verbose(10, opal_pmix_base_framework.framework_output,
"%s pmix:cray: assigned tmp name %d %d pmix_kvs_name %s",
OPAL_NAME_PRINT(*(opal_process_name_t*)&pmix_pname),pmix_pname.jid,pmix_pname.vid,pmix_kvs_name);
pmapping = (char*)malloc(PMI2_MAX_VALLEN);
if( pmapping == NULL ){
rc = OPAL_ERR_OUT_OF_RESOURCE;
OPAL_ERROR_LOG(rc);
return rc;
}
rc = PMI2_Info_GetJobAttr("PMI_process_mapping", pmapping, PMI2_MAX_VALLEN, &found);
if( !found || PMI_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc,"PMI2_Info_GetJobAttr");
return OPAL_ERROR;
}
pmix_lranks = pmix_cray_parse_pmap(pmapping, pmix_rank, &my_node, &pmix_nlranks);
if (NULL == pmix_lranks) {
rc = OPAL_ERR_OUT_OF_RESOURCE;
OPAL_ERROR_LOG(rc);
return rc;
}
free(pmapping);
/* find ourselves */
for (i=0; i < pmix_nlranks; i++) {
if (pmix_rank == pmix_lranks[i]) {
pmix_lrank = i;
pmix_nrank = my_node;
break;
}
}
return OPAL_SUCCESS;
err_exit:
PMI2_Finalize();
return ret;
}
int
shmem_runtime_fini(void)
{
PMI2_Finalize();
return 0;
}
void chpl_comm_ofi_oob_fini(void) {
if (PMI2_Initialized() == PMI_TRUE) {
DBG_PRINTF(DBG_OOB, "OOB finalize");
PMI_CHK(PMI2_Finalize());
}
}
static int s2_init(void)
{
int spawned, size, rank, appnum;
int rc, ret = OPAL_ERROR;
char buf[16];
int found;
int my_node;
char *tmp;
uint32_t jobfam, stepid;
int i;
/* if we can't startup PMI, we can't be used */
if ( PMI2_Initialized () ) {
return OPAL_SUCCESS;
}
size = -1;
rank = -1;
appnum = -1;
if (PMI2_SUCCESS != (rc = PMI2_Init(&spawned, &size, &rank, &appnum))) {
opal_show_help("help-pmix-base.txt", "pmix2-init-failed", true, rc);
return OPAL_ERROR;
}
if( size < 0 || rank < 0 ){
opal_show_help("help-pmix-base.txt", "pmix2-init-returned-bad-values", true);
goto err_exit;
}
s2_jsize = size;
s2_rank = rank;
s2_appnum = appnum;
pmix_vallen_max = PMI2_MAX_VALLEN;
pmix_kvslen_max = PMI2_MAX_VALLEN; // FIX ME: What to put here for versatility?
pmix_keylen_max = PMI2_MAX_KEYLEN;
pmix_vallen_threshold = PMI2_MAX_VALLEN * 3;
pmix_vallen_threshold >>= 2;
rc = PMI2_Info_GetJobAttr("universeSize", buf, 16, &found);
if( PMI2_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI_Get_universe_size");
goto err_exit;
}
s2_usize = atoi(buf);
pmix_kvs_name = (char*)malloc(pmix_kvslen_max);
if( pmix_kvs_name == NULL ){
PMI2_Finalize();
ret = OPAL_ERR_OUT_OF_RESOURCE;
goto err_exit;
}
rc = PMI2_Job_GetId(pmix_kvs_name, pmix_kvslen_max);
if( PMI2_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI2_Job_GetId");
goto err_exit;
}
/* Slurm PMI provides the job id as an integer followed
* by a '.', followed by essentially a stepid. The first integer
* defines an overall job number. The second integer is the number of
* individual jobs we have run within that allocation. So we translate
* this as the overall job number equating to our job family, and
* the individual number equating to our local jobid
*/
jobfam = strtoul(pmix_kvs_name, &tmp, 10);
if (NULL == tmp) {
/* hmmm - no '.', so let's just use zero */
stepid = 0;
} else {
tmp++; /* step over the '.' */
stepid = strtoul(tmp, NULL, 10);
}
/* now build the jobid */
s2_jobid = (jobfam << 16) | stepid;
/* store our name in the opal_proc_t so that
* debug messages will make sense - an upper
* layer will eventually overwrite it, but that
* won't do any harm */
s2_pname.jobid = s2_jobid;
s2_pname.vpid = s2_rank;
opal_proc_set_name(&s2_pname);
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s pmix:s2: assigned tmp name",
OPAL_NAME_PRINT(s2_pname));
char *pmapping = (char*)malloc(PMI2_MAX_VALLEN);
if( pmapping == NULL ){
rc = OPAL_ERR_OUT_OF_RESOURCE;
OPAL_ERROR_LOG(rc);
return rc;
}
rc = PMI2_Info_GetJobAttr("PMI_process_mapping", pmapping, PMI2_MAX_VALLEN, &found);
if( !found || PMI2_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc,"PMI2_Info_GetJobAttr");
return OPAL_ERROR;
}
s2_lranks = mca_common_pmi2_parse_pmap(pmapping, s2_pname.vpid, &my_node, &s2_nlranks);
if (NULL == s2_lranks) {
rc = OPAL_ERR_OUT_OF_RESOURCE;
OPAL_ERROR_LOG(rc);
return rc;
}
free(pmapping);
/* find ourselves */
for (i=0; i < s2_nlranks; i++) {
if (s2_rank == s2_lranks[i]) {
s2_lrank = i;
s2_nrank = i;
break;
}
}
/* increment the init count */
++pmix_init_count;
return OPAL_SUCCESS;
err_exit:
PMI2_Finalize();
return ret;
}
static int mca_initialize_pmi_v2(void)
{
int spawned, size, rank, appnum;
int rc, ret = OPAL_ERROR;
/* deal with a Slurm bug by first checking if we were
* even launched by a PMI server before attempting
* to use PMI */
if (NULL == getenv("PMI_FD")) {
return OPAL_ERROR;
}
/* if we can't startup PMI, we can't be used */
if ( PMI2_Initialized () ) {
return OPAL_SUCCESS;
}
size = -1;
rank = -1;
appnum = -1;
if (PMI2_SUCCESS != (rc = PMI2_Init(&spawned, &size, &rank, &appnum))) {
opal_show_help("help-common-pmi.txt", "pmi2-init-failed", true, rc);
return OPAL_ERROR;
}
if( size < 0 || rank < 0 ){
opal_output(0, "SIZE %d RANK %d", size, rank);
opal_show_help("help-common-pmi.txt", "pmi2-init-returned-bad-values", true);
goto err_exit;
}
pmi_size = size;
pmi_rank = rank;
pmi_appnum = appnum;
pmi_vallen_max = PMI2_MAX_VALLEN;
pmi_kvslen_max = PMI2_MAX_VALLEN; // FIX ME: What to put here for versatility?
pmi_keylen_max = PMI2_MAX_KEYLEN;
char buf[16];
int found;
rc = PMI2_Info_GetJobAttr("universeSize", buf, 16, &found);
if( PMI2_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI_Get_universe_size");
goto err_exit;
}
pmi_usize = atoi(buf);
pmi_kvs_name = (char*)malloc(pmi_kvslen_max);
if( pmi_kvs_name == NULL ){
PMI2_Finalize();
ret = OPAL_ERR_OUT_OF_RESOURCE;
goto err_exit;
}
rc = PMI2_Job_GetId(pmi_kvs_name, pmi_kvslen_max);
if( PMI2_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI2_Job_GetId");
goto err_exit;
}
return OPAL_SUCCESS;
err_exit:
PMI2_Finalize();
return ret;
}
/*@
MPIDI_PG_Finalize - Finalize the process groups, including freeing all
process group structures
@*/
int MPIDI_PG_Finalize(void)
{
int mpi_errno = MPI_SUCCESS;
MPIDI_PG_t *pg, *pgNext;
MPIR_FUNC_VERBOSE_STATE_DECL(MPID_STATE_MPIDI_PG_FINALIZE);
MPIR_FUNC_VERBOSE_ENTER(MPID_STATE_MPIDI_PG_FINALIZE);
/* Print the state of the process groups */
if (verbose) {
MPIU_PG_Printall( stdout );
}
/* FIXME - straighten out the use of PMI_Finalize - no use after
PG_Finalize */
if (pg_world->connData) {
#ifdef USE_PMI2_API
mpi_errno = PMI2_Finalize();
if (mpi_errno) MPIR_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**ch3|pmi_finalize");
#else
int rc;
rc = PMI_Finalize();
if (rc) {
MPIR_ERR_SET1(mpi_errno,MPI_ERR_OTHER,
"**ch3|pmi_finalize",
"**ch3|pmi_finalize %d", rc);
}
#endif
}
/* Free the storage associated with the process groups */
pg = MPIDI_PG_list;
while (pg) {
pgNext = pg->next;
/* In finalize, we free all process group information, even if
the ref count is not zero. This can happen if the user
fails to use MPI_Comm_disconnect on communicators that
were created with the dynamic process routines.*/
/* XXX DJG FIXME-MT should we be checking this? */
if (MPIR_Object_get_ref(pg) == 0 || 1) {
if (pg == MPIDI_Process.my_pg)
MPIDI_Process.my_pg = NULL;
MPIR_Object_set_ref(pg, 0); /* satisfy assertions in PG_Destroy */
MPIDI_PG_Destroy( pg );
}
pg = pgNext;
}
/* If COMM_WORLD is still around (it normally should be),
try to free it here. The reason that we need to free it at this
point is that comm_world (and comm_self) still exist, and
hence the usual process to free the related VC structures will
not be invoked. */
if (MPIDI_Process.my_pg) {
MPIDI_PG_Destroy(MPIDI_Process.my_pg);
}
MPIDI_Process.my_pg = NULL;
/* ifdefing out this check because the list will not be NULL in
Ch3_finalize because
one additional reference is retained in MPIDI_Process.my_pg.
That reference is released
only after ch3_finalize returns. If I release it before ch3_finalize,
the ssm channel crashes. */
#if 0
if (MPIDI_PG_list != NULL)
{
/* --BEGIN ERROR HANDLING-- */
mpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_FATAL, FCNAME, __LINE__, MPI_ERR_INTERN,
"**dev|pg_finalize|list_not_empty", NULL);
/* --END ERROR HANDLING-- */
}
#endif
MPIR_FUNC_VERBOSE_EXIT(MPID_STATE_MPIDI_PG_FINALIZE);
return mpi_errno;
}
static int cray_init(void)
{
int i, spawned, size, rank, appnum, my_node;
int rc, ret = OPAL_ERROR;
char *pmapping = NULL;
char buf[PMI2_MAX_ATTRVALUE];
int found;
int major, minor, revision;
uint32_t jobfam;
opal_value_t kv;
opal_process_name_t ldr;
char nmtmp[64];
char *str, **localranks = NULL;
++pmix_init_count;
/* if we can't startup PMI, we can't be used */
if ( PMI2_Initialized () ) {
opal_output_verbose(10, opal_pmix_base_framework.framework_output,
"%s pmix:cray: pmi already initialized",
OPAL_NAME_PRINT(pmix_pname));
return OPAL_SUCCESS;
}
size = -1;
rank = -1;
appnum = -1;
if (PMI_SUCCESS != (rc = PMI2_Init(&spawned, &size, &rank, &appnum))) {
opal_show_help("help-pmix-base.txt", "pmix2-init-failed", true, rc);
return OPAL_ERROR;
}
if( size < 0 || rank < 0 ){
opal_show_help("help-pmix-base.txt", "pmix2-init-returned-bad-values", true);
goto err_exit;
}
pmix_size = size;
pmix_rank = rank;
pmix_appnum = appnum;
pmix_vallen_max = PMI2_MAX_VALLEN;
pmix_kvslen_max = PMI2_MAX_VALLEN; // FIX ME: What to put here for versatility?
pmix_keylen_max = PMI2_MAX_KEYLEN;
pmix_vallen_threshold = PMI2_MAX_VALLEN * 3;
pmix_vallen_threshold >>= 2;
/*
* get the version info
*/
if (PMI_SUCCESS != PMI_Get_version_info(&major,&minor,&revision)) {
return OPAL_ERROR;
}
snprintf(cray_pmi_version, sizeof(cray_pmi_version),
"%d.%d.%d", major, minor, revision);
pmix_kvs_name = (char*)malloc(pmix_kvslen_max);
if( pmix_kvs_name == NULL ){
PMI2_Finalize();
ret = OPAL_ERR_OUT_OF_RESOURCE;
goto err_exit;
}
rc = PMI2_Job_GetId(pmix_kvs_name, pmix_kvslen_max);
if( PMI_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI2_Job_GetId");
goto err_exit;
}
rc = sscanf(pmix_kvs_name,"kvs_%u",&jobfam);
if (rc != 1) {
opal_output_verbose(10, opal_pmix_base_framework.framework_output,
"%s pmix:cray: pmix_kvs_name %s",
OPAL_NAME_PRINT(pmix_pname), pmix_kvs_name);
rc = OPAL_ERROR;
goto err_exit;
}
pmix_jobid = jobfam << 16;
/* store our name in the opal_proc_t so that
* debug messages will make sense - an upper
* layer will eventually overwrite it, but that
* won't do any harm */
pmix_pname.jobid = pmix_jobid;
pmix_pname.vpid = pmix_rank;
opal_proc_set_name(&pmix_pname);
opal_output_verbose(10, opal_pmix_base_framework.framework_output,
"%s pmix:cray: assigned tmp name %d %d pmix_kvs_name %s",
OPAL_NAME_PRINT(pmix_pname),pmix_pname.jobid,pmix_pname.vpid,pmix_kvs_name);
pmapping = (char*)malloc(PMI2_MAX_VALLEN);
if( pmapping == NULL ){
rc = OPAL_ERR_OUT_OF_RESOURCE;
OPAL_ERROR_LOG(rc);
return rc;
}
rc = PMI2_Info_GetJobAttr("PMI_process_mapping", pmapping, PMI2_MAX_VALLEN, &found);
if( !found || PMI_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc,"PMI2_Info_GetJobAttr");
return OPAL_ERROR;
}
pmix_lranks = pmix_cray_parse_pmap(pmapping, pmix_rank, &my_node, &pmix_nlranks);
if (NULL == pmix_lranks) {
rc = OPAL_ERR_OUT_OF_RESOURCE;
OPAL_ERROR_LOG(rc);
return rc;
}
free(pmapping);
// setup hash table
opal_pmix_base_hash_init();
/* save the job size */
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_JOB_SIZE);
kv.type = OPAL_UINT32;
kv.data.uint32 = pmix_size;
if (OPAL_SUCCESS != (rc = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(rc);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
/* save the appnum */
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_APPNUM);
kv.type = OPAL_UINT32;
kv.data.uint32 = pmix_appnum;
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
rc = PMI2_Info_GetJobAttr("universeSize", buf, 16, &found);
if( PMI_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI_Get_universe_size");
goto err_exit;
}
pmix_usize = atoi(buf);
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_UNIV_SIZE);
kv.type = OPAL_UINT32;
kv.data.uint32 = pmix_usize;
if (OPAL_SUCCESS != (rc = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(rc);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_JOBID);
kv.type = OPAL_UINT32;
kv.data.uint32 = pmix_jobid;
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
/* save the local size */
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_LOCAL_SIZE);
kv.type = OPAL_UINT16;
kv.data.uint16 = pmix_nlranks;
if (OPAL_SUCCESS != (rc = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(rc);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
ldr.vpid = pmix_lranks[0];
ldr.jobid = pmix_pname.jobid;
/* find ourselves and build up a string for local peer info */
memset(nmtmp, 0, 64);
for (i=0; i < pmix_nlranks; i++) {
ret = snprintf(nmtmp, 64, "%d", pmix_lranks[i]);
opal_argv_append_nosize(&localranks, nmtmp);
if (pmix_rank == pmix_lranks[i]) {
pmix_lrank = i;
pmix_nrank = i;
}
}
str = opal_argv_join(localranks, ',');
opal_argv_free(localranks);
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_LOCAL_PEERS);
kv.type = OPAL_STRING;
kv.data.string = str;
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
/* save the local leader */
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_LOCALLDR);
kv.type = OPAL_UINT64;
kv.data.uint64 = *(uint64_t*)&ldr;
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
/* save our local rank */
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_LOCAL_RANK);
kv.type = OPAL_UINT16;
kv.data.uint16 = pmix_lrank;
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
/* and our node rank */
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_NODE_RANK);
kv.type = OPAL_UINT16;
kv.data.uint16 = pmix_nrank;
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
return OPAL_SUCCESS;
err_exit:
PMI2_Finalize();
return ret;
}
static int s2_init(void)
{
int spawned, size, rank, appnum;
int rc, ret = OPAL_ERROR;
char buf[16];
int found;
int my_node;
uint32_t stepid;
int i;
opal_process_name_t ldr;
opal_value_t kv;
char **localranks;
char *str;
char nmtmp[64];
opal_process_name_t wildcard_rank;
/* if we can't startup PMI, we can't be used */
if ( PMI2_Initialized () ) {
return OPAL_SUCCESS;
}
size = -1;
rank = -1;
appnum = -1;
// setup hash table so we always can finalize it
opal_pmix_base_hash_init();
if (PMI2_SUCCESS != (rc = PMI2_Init(&spawned, &size, &rank, &appnum))) {
opal_show_help("help-pmix-base.txt", "pmix2-init-failed", true, rc);
return OPAL_ERROR;
}
if( size < 0 || rank < 0 ){
opal_show_help("help-pmix-base.txt", "pmix2-init-returned-bad-values", true);
goto err_exit;
}
s2_jsize = size;
s2_rank = rank;
s2_appnum = appnum;
pmix_vallen_max = PMI2_MAX_VALLEN;
pmix_kvslen_max = PMI2_MAX_VALLEN; // FIX ME: What to put here for versatility?
pmix_keylen_max = PMI2_MAX_KEYLEN;
pmix_vallen_threshold = PMI2_MAX_VALLEN * 3;
pmix_vallen_threshold >>= 2;
pmix_kvs_name = (char*)malloc(pmix_kvslen_max);
if( pmix_kvs_name == NULL ){
PMI2_Finalize();
ret = OPAL_ERR_OUT_OF_RESOURCE;
goto err_exit;
}
rc = PMI2_Job_GetId(pmix_kvs_name, pmix_kvslen_max);
if( PMI2_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI2_Job_GetId");
free(pmix_kvs_name);
goto err_exit;
}
/* store our name in the opal_proc_t so that
* debug messages will make sense - an upper
* layer will eventually overwrite it, but that
* won't do any harm */
s2_pname.jobid = strtoul(pmix_kvs_name, &str, 10);
s2_pname.jobid = (s2_pname.jobid << 16) & 0xffff0000;
if (NULL != str) {
stepid = strtoul(str, NULL, 10);
s2_pname.jobid |= (stepid & 0x0000ffff);
}
s2_pname.vpid = s2_rank;
opal_proc_set_name(&s2_pname);
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s pmix:s2: assigned tmp name",
OPAL_NAME_PRINT(s2_pname));
/* setup wildcard rank*/
wildcard_rank = OPAL_PROC_MY_NAME;
wildcard_rank.vpid = OPAL_VPID_WILDCARD;
/* Slurm PMI provides the job id as an integer followed
* by a '.', followed by essentially a stepid. The first integer
* defines an overall job number. The second integer is the number of
* individual jobs we have run within that allocation.
*/
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_JOBID);
kv.type = OPAL_UINT32;
kv.data.uint32 = s2_pname.jobid;
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&wildcard_rank, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv); // frees pmix_kvs_name
/* save the job size */
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_JOB_SIZE);
kv.type = OPAL_UINT32;
kv.data.uint32 = size;
if (OPAL_SUCCESS != (rc = opal_pmix_base_store(&wildcard_rank, &kv))) {
OPAL_ERROR_LOG(rc);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
/* save the appnum */
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_APPNUM);
kv.type = OPAL_UINT32;
kv.data.uint32 = appnum;
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
rc = PMI2_Info_GetJobAttr("universeSize", buf, 16, &found);
if( PMI2_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc, "PMI_Get_universe_size");
goto err_exit;
}
/* save it */
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_UNIV_SIZE);
kv.type = OPAL_UINT32;
kv.data.uint32 = atoi(buf);
if (OPAL_SUCCESS != (rc = opal_pmix_base_store(&wildcard_rank, &kv))) {
OPAL_ERROR_LOG(rc);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
/* push this into the dstore for subsequent fetches */
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_MAX_PROCS);
kv.type = OPAL_UINT32;
kv.data.uint32 = atoi(buf);
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&wildcard_rank, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
char *pmapping = (char*)malloc(PMI2_MAX_VALLEN);
if( pmapping == NULL ){
rc = OPAL_ERR_OUT_OF_RESOURCE;
OPAL_ERROR_LOG(rc);
return rc;
}
rc = PMI2_Info_GetJobAttr("PMI_process_mapping", pmapping, PMI2_MAX_VALLEN, &found);
if( !found || PMI2_SUCCESS != rc ) {
OPAL_PMI_ERROR(rc,"PMI2_Info_GetJobAttr");
return OPAL_ERROR;
}
s2_lranks = mca_common_pmi2_parse_pmap(pmapping, s2_pname.vpid, &my_node, &s2_nlranks);
if (NULL == s2_lranks) {
rc = OPAL_ERR_OUT_OF_RESOURCE;
OPAL_ERROR_LOG(rc);
return rc;
}
free(pmapping);
/* save the local size */
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_LOCAL_SIZE);
kv.type = OPAL_UINT32;
kv.data.uint32 = s2_nlranks;
if (OPAL_SUCCESS != (rc = opal_pmix_base_store(&wildcard_rank, &kv))) {
OPAL_ERROR_LOG(rc);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
s2_lrank = 0;
s2_nrank = 0;
ldr.vpid = rank;
localranks = NULL;
if (0 < s2_nlranks && NULL != s2_lranks) {
/* note the local ldr */
ldr.vpid = s2_lranks[0];
/* find ourselves */
ldr.jobid = s2_pname.jobid;
ldr.vpid = s2_pname.vpid;
memset(nmtmp, 0, 64);
for (i=0; i < s2_nlranks; i++) {
(void)snprintf(nmtmp, 64, "%d", s2_lranks[i]);
opal_argv_append_nosize(&localranks, nmtmp);
if (s2_rank == s2_lranks[i]) {
s2_lrank = i;
s2_nrank = i;
}
}
str = opal_argv_join(localranks, ',');
opal_argv_free(localranks);
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_LOCAL_PEERS);
kv.type = OPAL_STRING;
kv.data.string = str;
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&wildcard_rank, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
}
/* save the local leader */
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_LOCALLDR);
kv.type = OPAL_UINT64;
kv.data.uint64 = *(uint64_t*)&ldr;
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
/* save our local rank */
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_LOCAL_RANK);
kv.type = OPAL_UINT16;
kv.data.uint16 = s2_lrank;
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
/* and our node rank */
OBJ_CONSTRUCT(&kv, opal_value_t);
kv.key = strdup(OPAL_PMIX_NODE_RANK);
kv.type = OPAL_UINT16;
kv.data.uint16 = s2_nrank;
if (OPAL_SUCCESS != (ret = opal_pmix_base_store(&OPAL_PROC_MY_NAME, &kv))) {
OPAL_ERROR_LOG(ret);
OBJ_DESTRUCT(&kv);
goto err_exit;
}
OBJ_DESTRUCT(&kv);
/* increment the init count */
++pmix_init_count;
return OPAL_SUCCESS;
err_exit:
PMI2_Finalize();
return ret;
}
