void pmi_get_local_ranks(int **local_ranks, int *local_cnt)
{
int rc, found;
char *pmapping = (char*)malloc(PMI2_MAX_VALLEN);
if( pmapping == NULL ){
fprintf(stderr,"pmi_get_local_ranks (pmi2): cannot mallic for pmappig\n");
abort();
}
rc = PMI2_Info_GetJobAttr("PMI_process_mapping", pmapping, PMI2_MAX_VALLEN, &found);
if( !found || PMI2_SUCCESS != rc ) {
fprintf(stderr,"pmi_get_local_ranks (pmi2): PMI2_Info_GetJobAttr rc = %d\n", rc);
abort();
}
*local_ranks = NULL;
*local_ranks = mca_common_pmi2_parse_pmap(pmapping, my_rank, &my_node, local_cnt);
if (NULL == *local_ranks) {
fprintf(stderr,"mca_common_pmi2_parse_pmap: error\n");
abort();
}
free(pmapping);
}
int mca_common_pmi_local_info(int vpid, int **ranks_ret,
int *procs_ret, char **error)
{
int *ranks;
int procs = -1;
int rc;
#if WANT_PMI2_SUPPORT
if(mca_common_pmi_version == 2){
{
char *pmapping = (char*)malloc(PMI2_MAX_VALLEN);
if( pmapping == NULL ){
*error = "mca_common_pmi_local_info: could not get memory for PMIv2 process mapping";
return OPAL_ERR_OUT_OF_RESOURCE;
}
int found;
int my_node;
rc = PMI2_Info_GetJobAttr("PMI_process_mapping", pmapping, PMI2_MAX_VALLEN, &found);
if( !found || PMI2_SUCCESS != rc ) {
/* can't check PMI2_SUCCESS as some folks (i.e., Cray) don't define it */
OPAL_PMI_ERROR(rc,"PMI2_Info_GetJobAttr");
*error = "mca_common_pmi_local_info: could not get PMI_process_mapping";
return OPAL_ERROR;
}
ranks = mca_common_pmi2_parse_pmap(pmapping, vpid, &my_node, &procs);
if (NULL == ranks) {
*error = "mca_common_pmi_local_info: could not get memory for PMIv2 local ranks";
return OPAL_ERR_OUT_OF_RESOURCE;
}
free(pmapping);
}
}
else
#endif
{
/* get our local proc info to find our local rank */
if (PMI_SUCCESS != (rc = PMI_Get_clique_size(&procs))) {
OPAL_PMI_ERROR(rc, "PMI_Get_clique_size");
*error = "mca_common_pmi_local_info: could not get PMI clique size";
return OPAL_ERROR;
}
/* now get the specific ranks */
ranks = (int*)calloc(procs, sizeof(int));
if (NULL == ranks) {
*error = "mca_common_pmi_local_info: could not get memory for local ranks";
return OPAL_ERR_OUT_OF_RESOURCE;
}
if (PMI_SUCCESS != (rc = PMI_Get_clique_ranks(ranks, procs))) {
OPAL_PMI_ERROR(rc, "PMI_Get_clique_ranks");
*error = "mca_common_pmi_local_info: could not get clique ranks";
return OPAL_ERROR;
}
}
*ranks_ret = ranks;
*procs_ret = procs;
return OPAL_SUCCESS;
}
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 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;
}
