void * orte_qos_base_create (opal_list_t *qos_attributes)
{
orte_qos_base_channel_t * base_chan;
int32_t num_attributes;
int32_t rc, *window;
orte_qos_type_t *type;
orte_attribute_t *kv;
base_chan = OBJ_NEW (orte_qos_base_channel_t);
*type = orte_qos_noop;
// TBD _ we ignore inapplicable attributes for now - need to return error?
// get attributes of interest to the base and store them locally.
if (ORTE_SUCCESS == (rc = orte_set_attribute( &base_chan->attributes, ORTE_QOS_TYPE, ORTE_ATTR_GLOBAL, (void*)type, OPAL_UINT8))) {
// window size??
if( orte_get_attribute (qos_attributes, ORTE_QOS_WINDOW_SIZE, (void**)&window, OPAL_UINT32)) {
if ( ORTE_QOS_MAX_WINDOW_SIZE > (*window)) {
ORTE_ERROR_LOG(OPAL_ERR_VALUE_OUT_OF_BOUNDS);
OBJ_RELEASE(base_chan);
}
else {
if (ORTE_SUCCESS != (rc = orte_set_attribute(&base_chan->attributes, ORTE_QOS_WINDOW_SIZE,
ORTE_ATTR_GLOBAL, (void*)window, OPAL_UINT32))) {
ORTE_ERROR_LOG(rc);
OBJ_RELEASE(base_chan);
}
}
} else
OBJ_RELEASE(base_chan);
} else {
ORTE_ERROR_LOG(rc);
OBJ_RELEASE(base_chan);
}
return base_chan;
}
int orte_pre_condition_transports(orte_job_t *jdata, char **key)
{
uint64_t unique_key[2];
int n;
orte_app_context_t *app;
char *string_key, *cs_env;
int fd_rand;
size_t bytes_read;
struct stat buf;
/* put the number here - or else create an appropriate string. this just needs to
* eventually be a string variable
*/
if(0 != stat("/dev/urandom", &buf)) {
/* file doesn't exist! */
orte_pre_condition_transports_use_rand(unique_key);
}
if(-1 == (fd_rand = open("/dev/urandom", O_RDONLY))) {
orte_pre_condition_transports_use_rand(unique_key);
} else {
bytes_read = read(fd_rand, (char *) unique_key, 16);
if(bytes_read != 16) {
orte_pre_condition_transports_use_rand(unique_key);
}
close(fd_rand);
}
if (NULL == (string_key = orte_pre_condition_transports_print(unique_key))) {
ORTE_ERROR_LOG(ORTE_ERR_OUT_OF_RESOURCE);
return ORTE_ERR_OUT_OF_RESOURCE;
}
/* record it in case this job executes a dynamic spawn */
if (NULL != jdata) {
orte_set_attribute(&jdata->attributes, ORTE_JOB_TRANSPORT_KEY, ORTE_ATTR_LOCAL, string_key, OPAL_STRING);
if (OPAL_SUCCESS != mca_base_var_env_name ("orte_precondition_transports", &cs_env)) {
ORTE_ERROR_LOG(ORTE_ERR_OUT_OF_RESOURCE);
free(string_key);
return ORTE_ERR_OUT_OF_RESOURCE;
}
for (n=0; n < jdata->apps->size; n++) {
if (NULL == (app = (orte_app_context_t*)opal_pointer_array_get_item(jdata->apps, n))) {
continue;
}
opal_setenv(cs_env, string_key, true, &app->env);
}
free(cs_env);
free(string_key);
} else if (NULL != key) {
*key = string_key;
} else {
free(string_key);
}
return ORTE_SUCCESS;
}
/* after we allocate, we need to map the processes
* so we know what nodes will be used
*/
static void allocation_complete(int fd, short args, void *cbdata)
{
orte_state_caddy_t *state = (orte_state_caddy_t*)cbdata;
orte_job_t *jdata;
orte_job_t *daemons;
orte_topology_t *t;
orte_node_t *node;
int i;
ORTE_ACQUIRE_OBJECT(caddy);
jdata = state->jdata;
jdata->state = ORTE_JOB_STATE_ALLOCATION_COMPLETE;
/* get the daemon job object */
if (NULL == (daemons = orte_get_job_data_object(ORTE_PROC_MY_NAME->jobid))) {
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
goto done;
}
/* mark that we are not using a VM */
orte_set_attribute(&daemons->attributes, ORTE_JOB_NO_VM, ORTE_ATTR_GLOBAL, NULL, OPAL_BOOL);
/* ensure that all nodes point to our topology - we
* cannot support hetero nodes with this state machine
*/
t = (orte_topology_t*)opal_pointer_array_get_item(orte_node_topologies, 0);
for (i=1; i < orte_node_pool->size; i++) {
if (NULL == (node = (orte_node_t*)opal_pointer_array_get_item(orte_node_pool, i))) {
continue;
}
node->topology = t;
}
if (!orte_managed_allocation) {
if (NULL != orte_set_slots &&
0 != strncmp(orte_set_slots, "none", strlen(orte_set_slots))) {
for (i=0; i < orte_node_pool->size; i++) {
if (NULL == (node = (orte_node_t*)opal_pointer_array_get_item(orte_node_pool, i))) {
continue;
}
if (!ORTE_FLAG_TEST(node, ORTE_NODE_FLAG_SLOTS_GIVEN)) {
OPAL_OUTPUT_VERBOSE((5, orte_plm_base_framework.framework_output,
"%s plm:base:setting slots for node %s by %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), node->name, orte_set_slots));
orte_plm_base_set_slots(node);
}
}
}
}
/* move to the map stage */
ORTE_ACTIVATE_JOB_STATE(jdata, ORTE_JOB_STATE_MAP);
done:
/* cleanup */
OBJ_RELEASE(state);
}
/* after we allocate, we need to map the processes
* so we know what nodes will be used
*/
static void allocation_complete(int fd, short args, void *cbdata)
{
orte_state_caddy_t *state = (orte_state_caddy_t*)cbdata;
orte_job_t *jdata = state->jdata;
orte_job_t *daemons;
jdata->state = ORTE_JOB_STATE_ALLOCATION_COMPLETE;
/* get the daemon job object */
if (NULL == (daemons = orte_get_job_data_object(ORTE_PROC_MY_NAME->jobid))) {
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
goto done;
}
/* mark that we are not using a VM */
orte_set_attribute(&daemons->attributes, ORTE_JOB_NO_VM, ORTE_ATTR_GLOBAL, NULL, OPAL_BOOL);
#if OPAL_HAVE_HWLOC
{
orte_topology_t *t;
orte_node_t *node;
int i;
/* ensure that all nodes point to our topology - we
* cannot support hetero nodes with this state machine
*/
t = (orte_topology_t*)opal_pointer_array_get_item(orte_node_topologies, 0);
for (i=1; i < orte_node_pool->size; i++) {
if (NULL == (node = (orte_node_t*)opal_pointer_array_get_item(orte_node_pool, i))) {
continue;
}
node->topology = t->topo;
}
}
#endif
/* move to the map stage */
ORTE_ACTIVATE_JOB_STATE(jdata, ORTE_JOB_STATE_MAP);
done:
/* cleanup */
OBJ_RELEASE(state);
}
static void assign(orte_job_t *jdata)
{
#if HWLOC_API_VERSION >= 0x20000
opal_list_t *cache;
opal_value_t *kv;
if (VM_HOLE_NONE == mca_rtc_hwloc_component.kind ||
NULL == shmemfile) {
return;
}
/* add the shmem address and size to the job-level info that
* will be provided to the proc upon registration */
cache = NULL;
if (!orte_get_attribute(&jdata->attributes, ORTE_JOB_INFO_CACHE, (void**)&cache, OPAL_PTR) ||
NULL == cache) {
cache = OBJ_NEW(opal_list_t);
orte_set_attribute(&jdata->attributes, ORTE_JOB_INFO_CACHE, ORTE_ATTR_LOCAL, cache, OPAL_PTR);
}
opal_output_verbose(2, orte_rtc_base_framework.framework_output,
"FILE %s ADDR %lx SIZE %lx", shmemfile,
(unsigned long)shmemaddr,
(unsigned long)shmemsize);
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_HWLOC_SHMEM_FILE);
kv->type = OPAL_STRING;
kv->data.string = strdup(shmemfile);
opal_list_append(cache, &kv->super);
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_HWLOC_SHMEM_ADDR);
kv->type = OPAL_SIZE;
kv->data.size = shmemaddr;
opal_list_append(cache, &kv->super);
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_HWLOC_SHMEM_SIZE);
kv->type = OPAL_SIZE;
kv->data.size = shmemsize;
opal_list_append(cache, &kv->super);
#endif
}
static int
orte_ras_alps_read_appinfo_file(opal_list_t *nodes, char *filename,
unsigned int *uMe)
{
int iq;
int ix;
int iFd; /* file descriptor for appinfo */
int iTrips; /* counter appinfo read attempts */
int max_appinfo_read_attempts;
struct stat ssBuf; /* stat buffer */
size_t szLen; /* size of appinfo (file) */
off_t oNow; /* current appinfo data offset */
off_t oInfo=sizeof(appInfoHdr_t);
off_t oDet=sizeof(appInfo_t);
off_t oSlots;
off_t oEntry;
int32_t sNodes=0;
char *cpBuf;
char *hostname;
orte_node_t *node = NULL;
appInfoHdr_t *apHdr; /* ALPS header structure */
appInfo_t *apInfo; /* ALPS table info structure */
#if ALPS_APPINFO_VERSION==0
placeList_t *apSlots; /* ALPS node specific info */
#else
orte_ras_alps_placeNodeList_t *apNodes;
#endif
orte_ras_alps_get_appinfo_attempts(&max_appinfo_read_attempts);
oNow=0;
iTrips=0;
opal_output_verbose(1, orte_ras_base_framework.framework_output,
"ras:alps:allocate: begin processing appinfo file");
while(!oNow) { /* Until appinfo read is complete */
iTrips++; /* Increment trip count */
iFd=open( filename, O_RDONLY );
if( iFd==-1 ) { /* If file absent, ALPS is down */
opal_output_verbose(1, orte_ras_base_framework.framework_output,
"ras:alps:allocate: ALPS information open failure");
usleep(iTrips*50000); /* Increasing delays, .05 s/try */
/* Fail only when number of attempts have been exhausted. */
if( iTrips <= max_appinfo_read_attempts ) continue;
ORTE_ERROR_LOG(ORTE_ERR_FILE_OPEN_FAILURE);
return ORTE_ERR_FILE_OPEN_FAILURE;
}
if( fstat( iFd, &ssBuf )==-1 ) { /* If stat fails, access denied */
ORTE_ERROR_LOG(ORTE_ERR_NOT_AVAILABLE);
return ORTE_ERR_NOT_AVAILABLE;
}
szLen=ssBuf.st_size; /* Get buffer size */
cpBuf=malloc(szLen+1); /* Allocate buffer */
if (NULL == cpBuf) {
ORTE_ERROR_LOG(ORTE_ERR_OUT_OF_RESOURCE);
return ORTE_ERR_OUT_OF_RESOURCE;
}
/* Repeated attempts to read appinfo, with an increasing delay between *
* successive attempts to allow scheduler I/O a chance to complete. */
if( (oNow=read( iFd, cpBuf, szLen ))!=(off_t)szLen ) {
/* This is where apstat fails; we will record it and try again. */
opal_output_verbose(1, orte_ras_base_framework.framework_output,
"ras:alps:allocate: ALPS information read failure: %ld bytes", (long int)oNow);
free(cpBuf); /* Free (old) buffer */
close(iFd); /* Close (old) descriptor */
oNow=0; /* Reset byte count */
usleep(iTrips*50000); /* Increasing delays, .05 s/try */
/* Fail only when number of attempts have been exhausted. */
if( iTrips<=max_appinfo_read_attempts ) continue;
ORTE_ERROR_LOG(ORTE_ERR_FILE_READ_FAILURE);
return ORTE_ERR_FILE_READ_FAILURE;
}
}
close(iFd);
opal_output_verbose(1, orte_ras_base_framework.framework_output,
"ras:alps:allocate: file %s read", filename);
/* Now that we have the scheduler information, we just have to parse it for *
* the data that we seek. */
oNow=0;
apHdr=(appInfoHdr_t *)cpBuf;
opal_output_verbose(1, orte_ras_base_framework.framework_output,
"ras:alps:allocate: %d entries in file", apHdr->apNum);
/* Header info (apHdr) tells us how many entries are in the file: *
* *
* apHdr->apNum */
for( iq=0; iq<apHdr->apNum; iq++ ) { /* Parse all entries in file */
/* Just at this level, a lot of information is available: *
* *
* apInfo->apid ... ALPS job ID *
* apInfo->resId ... ALPS reservation ID *
* apInfo->numCmds ... Number of executables *
* apInfo->numPlaces ... Number of PEs */
apInfo=(appInfo_t *)(cpBuf+oNow+oInfo);
/* Calculate the dependent offsets. */
oSlots=sizeof(cmdDetail_t)*apInfo->numCmds;
opal_output_verbose(1, orte_ras_base_framework.framework_output,
"ras:alps:allocate: read data for resId %u - myId %u",
apInfo->resId, *uMe);
#if ALPS_APPINFO_VERSION==0
/* Finally, we get to the actual node-specific information: *
* *
* apSlots[ix].cmdIx ... index of apDet[].cmd *
* apSlots[ix].nid ... NodeID (NID) *
* apSlots[ix].procMask ... mask for processors... need 16-bit shift */
apSlots=(placeList_t *)(cpBuf+oNow+oInfo+oDet+oSlots);
oEntry=sizeof(placeList_t)*apInfo->numPlaces;
oNow+=(oDet+oSlots+oEntry); /* Target next slot */
if( apInfo->resId != *uMe ) continue; /* Filter to our reservation Id */
/* in this early version of alps, there is one entry for each PE in the
* allocation - so cycle across the numPlaces entries, assigning a slot
* for each time a node is named
*/
for( ix=0; ix<apInfo->numPlaces; ix++ ) {
opal_output_verbose(5, orte_ras_base_framework.framework_output,
"ras:alps:read_appinfo: got NID %d", apSlots[ix].nid);
asprintf( &hostname, "%d", apSlots[ix].nid );
if (NULL == hostname) {
ORTE_ERROR_LOG(ORTE_ERR_OUT_OF_RESOURCE);
return ORTE_ERR_OUT_OF_RESOURCE;
}
/* If this matches the prior nodename, just add to the slot count. */
if( NULL!=node && !strcmp(node->name, hostname) ) {
free(hostname); /* free hostname since not needed */
++node->slots;
} else { /* must be new, so add to list */
opal_output_verbose(1, orte_ras_base_framework.framework_output,
"ras:alps:read_appinfo: added NID %d to list", apSlots[ix].nid);
node = OBJ_NEW(orte_node_t);
node->name = hostname;
orte_set_attribute(&node->attributes, ORTE_NODE_LAUNCH_ID, ORTE_ATTR_LOCAL, &apSlots[ix].nid, OPAL_INT32);
node->slots_inuse = 0;
node->slots_max = 0;
node->slots = 1;
node->state = ORTE_NODE_STATE_UP;
/* need to order these node ids so the regex generator
* can properly function
*/
/* add it to the end */
opal_list_append(nodes, &node->super);
sNodes++; /* Increment the node count */
}
}
#else
/* in newer versions of alps, there is one entry for each node in the
* allocation, and that struct directly carries the number of PEs
* allocated on that node to this job.
*/
apNodes=(orte_ras_alps_placeNodeList_t *)(cpBuf+oNow+oInfo+oDet+oSlots);
oEntry=sizeof(orte_ras_alps_placeNodeList_t)*apInfo->numPlaces;
oNow+=(oDet+oSlots+oEntry); /* Target next entry */
if( apInfo->resId != *uMe ) continue; /* Filter to our reservation Id */
for( ix=0; ix<apInfo->numPlaces; ix++ ) {
opal_output_verbose(5, orte_ras_base_framework.framework_output,
"ras:alps:read_appinfo(modern): processing NID %d with %d slots",
apNodes[ix].nid, apNodes[ix].numPEs);
asprintf( &hostname, "%d", apNodes[ix].nid );
if (NULL == hostname) {
ORTE_ERROR_LOG(ORTE_ERR_OUT_OF_RESOURCE);
return ORTE_ERR_OUT_OF_RESOURCE;
}
node = OBJ_NEW(orte_node_t);
node->name = hostname;
orte_set_attribute(&node->attributes, ORTE_NODE_LAUNCH_ID, ORTE_ATTR_LOCAL, &apNodes[ix].nid, OPAL_INT32);
node->slots_inuse = 0;
node->slots_max = 0;
node->slots = apNodes[ix].numPEs;
node->state = ORTE_NODE_STATE_UP;
/* need to order these node ids so the regex generator
* can properly function
*/
/* add it to the end */
opal_list_append(nodes, &node->super);
sNodes++; /* Increment the node count */
}
#endif
break; /* Extended details ignored */
}
opal_list_sort (nodes, compare_nodes);
free(cpBuf); /* Free the buffer */
return ORTE_SUCCESS;
}
/* stuff proc attributes for sending back to a proc */
int orte_pmix_server_register_nspace(orte_job_t *jdata)
{
int rc;
orte_proc_t *pptr;
int i, k, n;
opal_list_t *info, *pmap;
opal_value_t *kv;
orte_node_t *node, *mynode;
opal_vpid_t vpid;
char **list, **procs, **micro, *tmp, *regex, *cpulist, *peerlist;
orte_job_t *dmns;
orte_job_map_t *map;
orte_app_context_t *app;
uid_t uid;
gid_t gid;
opal_list_t *cache;
opal_output_verbose(2, orte_pmix_server_globals.output,
"%s register nspace for %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_JOBID_PRINT(jdata->jobid));
/* setup the info list */
info = OBJ_NEW(opal_list_t);
uid = geteuid();
gid = getegid();
/* jobid */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_JOBID);
kv->data.string = strdup(ORTE_JOBID_PRINT(jdata->jobid));
kv->type = OPAL_STRING;
opal_list_append(info, &kv->super);
/* offset */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_NPROC_OFFSET);
kv->data.uint32 = jdata->offset;
kv->type = OPAL_UINT32;
opal_list_append(info, &kv->super);
/* check for cached values to add to the job info */
cache = NULL;
if (orte_get_attribute(&jdata->attributes, ORTE_JOB_INFO_CACHE, (void**)&cache, OPAL_PTR) &&
NULL != cache) {
while (NULL != (kv = (opal_value_t*)opal_list_remove_first(cache))) {
opal_list_append(info, &kv->super);
}
orte_remove_attribute(&jdata->attributes, ORTE_JOB_INFO_CACHE);
OBJ_RELEASE(cache);
}
/* assemble the node and proc map info */
list = NULL;
procs = NULL;
map = jdata->map;
for (i=0; i < map->nodes->size; i++) {
micro = NULL;
if (NULL != (node = (orte_node_t*)opal_pointer_array_get_item(map->nodes, i))) {
opal_argv_append_nosize(&list, node->name);
/* assemble all the ranks for this job that are on this node */
for (k=0; k < node->procs->size; k++) {
if (NULL != (pptr = (orte_proc_t*)opal_pointer_array_get_item(node->procs, k))) {
if (jdata->jobid == pptr->name.jobid) {
opal_argv_append_nosize(µ, ORTE_VPID_PRINT(pptr->name.vpid));
}
}
}
/* assemble the rank/node map */
if (NULL != micro) {
tmp = opal_argv_join(micro, ',');
opal_argv_free(micro);
opal_argv_append_nosize(&procs, tmp);
free(tmp);
}
}
}
/* let the PMIx server generate the nodemap regex */
if (NULL != list) {
tmp = opal_argv_join(list, ',');
opal_argv_free(list);
list = NULL;
if (OPAL_SUCCESS != (rc = opal_pmix.generate_regex(tmp, ®ex))) {
ORTE_ERROR_LOG(rc);
free(tmp);
OPAL_LIST_RELEASE(info);
return rc;
}
free(tmp);
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_NODE_MAP);
kv->type = OPAL_STRING;
kv->data.string = regex;
opal_list_append(info, &kv->super);
}
/* let the PMIx server generate the procmap regex */
if (NULL != procs) {
tmp = opal_argv_join(procs, ';');
opal_argv_free(procs);
procs = NULL;
if (OPAL_SUCCESS != (rc = opal_pmix.generate_ppn(tmp, ®ex))) {
ORTE_ERROR_LOG(rc);
free(tmp);
OPAL_LIST_RELEASE(info);
return rc;
}
free(tmp);
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_PROC_MAP);
kv->type = OPAL_STRING;
kv->data.string = regex;
opal_list_append(info, &kv->super);
}
/* get our local node */
if (NULL == (dmns = orte_get_job_data_object(ORTE_PROC_MY_NAME->jobid))) {
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
OPAL_LIST_RELEASE(info);
return ORTE_ERR_NOT_FOUND;
}
if (NULL == (pptr = (orte_proc_t*)opal_pointer_array_get_item(dmns->procs, ORTE_PROC_MY_NAME->vpid))) {
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
OPAL_LIST_RELEASE(info);
return ORTE_ERR_NOT_FOUND;
}
mynode = pptr->node;
if (NULL == mynode) {
/* cannot happen */
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
OPAL_LIST_RELEASE(info);
return ORTE_ERR_NOT_FOUND;
}
/* pass our node ID */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_NODEID);
kv->type = OPAL_UINT32;
kv->data.uint32 = mynode->index;
opal_list_append(info, &kv->super);
/* pass our node size */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_NODE_SIZE);
kv->type = OPAL_UINT32;
kv->data.uint32 = mynode->num_procs;
opal_list_append(info, &kv->super);
/* pass the number of nodes in the job */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_NUM_NODES);
kv->type = OPAL_UINT32;
kv->data.uint32 = map->num_nodes;
opal_list_append(info, &kv->super);
/* univ size */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_UNIV_SIZE);
kv->type = OPAL_UINT32;
kv->data.uint32 = jdata->total_slots_alloc;
opal_list_append(info, &kv->super);
/* job size */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_JOB_SIZE);
kv->type = OPAL_UINT32;
kv->data.uint32 = jdata->num_procs;
opal_list_append(info, &kv->super);
/* number of apps in this job */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_JOB_NUM_APPS);
kv->type = OPAL_UINT32;
kv->data.uint32 = jdata->num_apps;
opal_list_append(info, &kv->super);
/* local size */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_LOCAL_SIZE);
kv->type = OPAL_UINT32;
kv->data.uint32 = jdata->num_local_procs;
opal_list_append(info, &kv->super);
/* max procs */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_MAX_PROCS);
kv->type = OPAL_UINT32;
kv->data.uint32 = jdata->total_slots_alloc;
opal_list_append(info, &kv->super);
/* register any local clients */
vpid = ORTE_VPID_MAX;
for (i=0; i < mynode->procs->size; i++) {
if (NULL == (pptr = (orte_proc_t*)opal_pointer_array_get_item(mynode->procs, i))) {
continue;
}
if (pptr->name.jobid == jdata->jobid) {
if (pptr->name.vpid < vpid) {
vpid = pptr->name.vpid;
}
/* go ahead and register this client */
if (OPAL_SUCCESS != (rc = opal_pmix.server_register_client(&pptr->name, uid, gid,
(void*)pptr, NULL, NULL))) {
ORTE_ERROR_LOG(rc);
}
}
}
/* pass the local ldr */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_LOCALLDR);
kv->type = OPAL_VPID;
kv->data.name.vpid = vpid;
opal_list_append(info, &kv->super);
/* for each proc in this job, create an object that
* includes the info describing the proc so the recipient has a complete
* picture. This allows procs to connect to each other without
* any further info exchange, assuming the underlying transports
* support it. We also pass all the proc-specific data here so
* that each proc can lookup info about every other proc in the job */
for (n=0; n < map->nodes->size; n++) {
if (NULL == (node = (orte_node_t*)opal_pointer_array_get_item(map->nodes, n))) {
continue;
}
/* construct the list of local peers, while adding
* each proc's locality info */
list = NULL;
procs = NULL;
cpulist = NULL;
peerlist = NULL;
vpid = ORTE_VPID_MAX;
for (i=0; i < node->procs->size; i++) {
if (NULL == (pptr = (orte_proc_t*)opal_pointer_array_get_item(node->procs, i))) {
continue;
}
if (pptr->name.jobid == jdata->jobid) {
opal_argv_append_nosize(&list, ORTE_VPID_PRINT(pptr->name.vpid));
if (pptr->name.vpid < vpid) {
vpid = pptr->name.vpid;
}
/* note that we have to pass the cpuset for each local
* peer so locality can be computed */
tmp = NULL;
if (orte_get_attribute(&pptr->attributes, ORTE_PROC_CPU_BITMAP, (void**)&tmp, OPAL_STRING)) {
if (NULL != tmp) {
opal_argv_append_nosize(&procs, tmp);
free(tmp);
} else {
opal_argv_append_nosize(&procs, "UNBOUND");
}
} else {
opal_argv_append_nosize(&procs, "UNBOUND");
}
}
}
/* construct the list of peers for transmission */
if (NULL != list) {
peerlist = opal_argv_join(list, ',');
opal_argv_free(list);
list = NULL;
}
/* construct the list of cpusets for transmission */
if (NULL != procs) {
cpulist = opal_argv_join(procs, ':');
opal_argv_free(procs);
procs = NULL;
}
/* if this is me, then pass the peers and cpusets to myself
* in order to maintain backward compatibility for the non-pmix
* components in OPAL/pmix */
if (node == mynode) {
/* pass the list of peers */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_LOCAL_PEERS);
kv->type = OPAL_STRING;
kv->data.string = strdup(peerlist);
opal_list_append(info, &kv->super);
/* pass the list of cpusets */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_LOCAL_CPUSETS);
kv->type = OPAL_STRING;
kv->data.string = strdup(cpulist);
opal_list_append(info, &kv->super);
}
/* now cycle across each proc on this node, passing all data that
* varies by proc */
for (i=0; i < node->procs->size; i++) {
if (NULL == (pptr = (orte_proc_t*)opal_pointer_array_get_item(node->procs, i))) {
continue;
}
/* only consider procs from this job */
if (pptr->name.jobid != jdata->jobid) {
continue;
}
/* setup the proc map object */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_PROC_DATA);
kv->type = OPAL_PTR;
kv->data.ptr = OBJ_NEW(opal_list_t);
opal_list_append(info, &kv->super);
pmap = kv->data.ptr;
/* must start with rank */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_RANK);
kv->type = OPAL_VPID;
kv->data.name.vpid = pptr->name.vpid;
opal_list_append(pmap, &kv->super);
/* pass the list of peers */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_LOCAL_PEERS);
kv->type = OPAL_STRING;
kv->data.string = strdup(peerlist);
opal_list_append(pmap, &kv->super);
/* pass the list of cpusets */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_LOCAL_CPUSETS);
kv->type = OPAL_STRING;
kv->data.string = strdup(cpulist);
opal_list_append(pmap, &kv->super);
/* appnum */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_APPNUM);
kv->type = OPAL_UINT32;
kv->data.uint32 = pptr->app_idx;
opal_list_append(pmap, &kv->super);
/* app ldr */
app = (orte_app_context_t*)opal_pointer_array_get_item(jdata->apps, pptr->app_idx);
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_APPLDR);
kv->type = OPAL_VPID;
kv->data.name.vpid = app->first_rank;
opal_list_append(pmap, &kv->super);
/* global/univ rank */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_GLOBAL_RANK);
kv->type = OPAL_VPID;
kv->data.name.vpid = pptr->name.vpid + jdata->offset;
opal_list_append(pmap, &kv->super);
/* app rank */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_APP_RANK);
kv->type = OPAL_VPID;
kv->data.name.vpid = pptr->app_rank;
opal_list_append(pmap, &kv->super);
/* app size */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_APP_SIZE);
kv->type = OPAL_UINT32;
kv->data.uint32 = app->num_procs;
opal_list_append(info, &kv->super);
/* local rank */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_LOCAL_RANK);
kv->type = OPAL_UINT16;
kv->data.uint16 = pptr->local_rank;
opal_list_append(pmap, &kv->super);
/* node rank */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_NODE_RANK);
kv->type = OPAL_UINT16;
kv->data.uint32 = pptr->node_rank;
opal_list_append(pmap, &kv->super);
/* hostname */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_HOSTNAME);
kv->type = OPAL_STRING;
kv->data.string = strdup(pptr->node->name);
opal_list_append(pmap, &kv->super);
/* node ID */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_NODEID);
kv->type = OPAL_UINT32;
kv->data.uint32 = pptr->node->index;
opal_list_append(pmap, &kv->super);
}
/* cleanup */
if (NULL != cpulist) {
free(cpulist);
}
if (NULL != peerlist) {
free(peerlist);
}
}
/* mark the job as registered */
orte_set_attribute(&jdata->attributes, ORTE_JOB_NSPACE_REGISTERED, ORTE_ATTR_LOCAL, NULL, OPAL_BOOL);
/* pass it down */
/* we are in an event, so no need to callback */
rc = opal_pmix.server_register_nspace(jdata->jobid,
jdata->num_local_procs,
info, NULL, NULL);
OPAL_LIST_RELEASE(info);
return rc;
}
int orte_ess_base_orted_setup(char **hosts)
{
int ret = ORTE_ERROR;
int fd;
char log_file[PATH_MAX];
char *jobidstring;
char *error = NULL;
orte_job_t *jdata;
orte_proc_t *proc;
orte_app_context_t *app;
orte_node_t *node;
char *param;
hwloc_obj_t obj;
unsigned i, j;
opal_list_t transports;
/* my name is set, xfer it to the OPAL layer */
orte_process_info.super.proc_name = *(opal_process_name_t*)ORTE_PROC_MY_NAME;
orte_process_info.super.proc_hostname = strdup(orte_process_info.nodename);
orte_process_info.super.proc_flags = OPAL_PROC_ALL_LOCAL;
orte_process_info.super.proc_arch = opal_local_arch;
opal_proc_local_set(&orte_process_info.super);
plm_in_use = false;
/* setup callback for SIGPIPE */
setup_sighandler(SIGPIPE, &epipe_handler, epipe_signal_callback);
/* Set signal handlers to catch kill signals so we can properly clean up
* after ourselves.
*/
setup_sighandler(SIGTERM, &term_handler, shutdown_signal);
setup_sighandler(SIGINT, &int_handler, shutdown_signal);
/** setup callbacks for signals we should ignore */
setup_sighandler(SIGUSR1, &sigusr1_handler, signal_callback);
setup_sighandler(SIGUSR2, &sigusr2_handler, signal_callback);
signals_set = true;
/* get the local topology */
if (NULL == opal_hwloc_topology) {
if (OPAL_SUCCESS != (ret = opal_hwloc_base_get_topology())) {
error = "topology discovery";
goto error;
}
}
/* generate the signature */
orte_topo_signature = opal_hwloc_base_get_topo_signature(opal_hwloc_topology);
/* remove the hostname from the topology. Unfortunately, hwloc
* decided to add the source hostname to the "topology", thus
* rendering it unusable as a pure topological description. So
* we remove that information here.
*/
obj = hwloc_get_root_obj(opal_hwloc_topology);
for (i=0; i < obj->infos_count; i++) {
if (NULL == obj->infos[i].name ||
NULL == obj->infos[i].value) {
continue;
}
if (0 == strncmp(obj->infos[i].name, "HostName", strlen("HostName"))) {
free(obj->infos[i].name);
free(obj->infos[i].value);
/* left justify the array */
for (j=i; j < obj->infos_count-1; j++) {
obj->infos[j] = obj->infos[j+1];
}
obj->infos[obj->infos_count-1].name = NULL;
obj->infos[obj->infos_count-1].value = NULL;
obj->infos_count--;
break;
}
}
if (15 < opal_output_get_verbosity(orte_ess_base_framework.framework_output)) {
opal_output(0, "%s Topology Info:", ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
opal_dss.dump(0, opal_hwloc_topology, OPAL_HWLOC_TOPO);
}
/* open and setup the opal_pstat framework so we can provide
* process stats if requested
*/
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&opal_pstat_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "opal_pstat_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = opal_pstat_base_select())) {
ORTE_ERROR_LOG(ret);
error = "opal_pstat_base_select";
goto error;
}
/* define the HNP name */
ORTE_PROC_MY_HNP->jobid = ORTE_PROC_MY_NAME->jobid;
ORTE_PROC_MY_HNP->vpid = 0;
/* open and setup the state machine */
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_state_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_state_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_state_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_state_base_select";
goto error;
}
/* open the errmgr */
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_errmgr_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_errmgr_base_open";
goto error;
}
/* some environments allow remote launches - e.g., ssh - so
* open and select something -only- if we are given
* a specific module to use
*/
(void) mca_base_var_env_name("plm", ¶m);
plm_in_use = !!(getenv(param));
free (param);
if (plm_in_use) {
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_plm_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_plm_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_plm_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_plm_base_select";
goto error;
}
}
/* setup my session directory here as the OOB may need it */
if (orte_create_session_dirs) {
OPAL_OUTPUT_VERBOSE((2, orte_ess_base_framework.framework_output,
"%s setting up session dir with\n\ttmpdir: %s\n\thost %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
(NULL == orte_process_info.tmpdir_base) ? "UNDEF" : orte_process_info.tmpdir_base,
orte_process_info.nodename));
/* take a pass thru the session directory code to fillin the
* tmpdir names - don't create anything yet
*/
if (ORTE_SUCCESS != (ret = orte_session_dir(false, ORTE_PROC_MY_NAME))) {
ORTE_ERROR_LOG(ret);
error = "orte_session_dir define";
goto error;
}
/* clear the session directory just in case there are
* stale directories laying around
*/
orte_session_dir_cleanup(ORTE_JOBID_WILDCARD);
/* now actually create the directory tree */
if (ORTE_SUCCESS != (ret = orte_session_dir(true, ORTE_PROC_MY_NAME))) {
ORTE_ERROR_LOG(ret);
error = "orte_session_dir";
goto error;
}
/* set the opal_output env file location to be in the
* proc-specific session directory. */
opal_output_set_output_file_info(orte_process_info.proc_session_dir,
"output-", NULL, NULL);
/* setup stdout/stderr */
if (orte_debug_daemons_file_flag) {
/* if we are debugging to a file, then send stdout/stderr to
* the orted log file
*/
/* get my jobid */
if (ORTE_SUCCESS != (ret = orte_util_convert_jobid_to_string(&jobidstring,
ORTE_PROC_MY_NAME->jobid))) {
ORTE_ERROR_LOG(ret);
error = "convert_jobid";
goto error;
}
/* define a log file name in the session directory */
snprintf(log_file, PATH_MAX, "output-orted-%s-%s.log",
jobidstring, orte_process_info.nodename);
log_path = opal_os_path(false, orte_process_info.top_session_dir,
log_file, NULL);
fd = open(log_path, O_RDWR|O_CREAT|O_TRUNC, 0640);
if (fd < 0) {
/* couldn't open the file for some reason, so
* just connect everything to /dev/null
*/
fd = open("/dev/null", O_RDWR|O_CREAT|O_TRUNC, 0666);
} else {
dup2(fd, STDOUT_FILENO);
dup2(fd, STDERR_FILENO);
if(fd != STDOUT_FILENO && fd != STDERR_FILENO) {
close(fd);
}
}
}
}
/* setup the global job and node arrays */
orte_job_data = OBJ_NEW(opal_hash_table_t);
if (ORTE_SUCCESS != (ret = opal_hash_table_init(orte_job_data, 128))) {
ORTE_ERROR_LOG(ret);
error = "setup job array";
goto error;
}
orte_node_pool = OBJ_NEW(opal_pointer_array_t);
if (ORTE_SUCCESS != (ret = opal_pointer_array_init(orte_node_pool,
ORTE_GLOBAL_ARRAY_BLOCK_SIZE,
ORTE_GLOBAL_ARRAY_MAX_SIZE,
ORTE_GLOBAL_ARRAY_BLOCK_SIZE))) {
ORTE_ERROR_LOG(ret);
error = "setup node array";
goto error;
}
orte_node_topologies = OBJ_NEW(opal_pointer_array_t);
if (ORTE_SUCCESS != (ret = opal_pointer_array_init(orte_node_topologies,
ORTE_GLOBAL_ARRAY_BLOCK_SIZE,
ORTE_GLOBAL_ARRAY_MAX_SIZE,
ORTE_GLOBAL_ARRAY_BLOCK_SIZE))) {
ORTE_ERROR_LOG(ret);
error = "setup node topologies array";
goto error;
}
/* Setup the job data object for the daemons */
/* create and store the job data object */
jdata = OBJ_NEW(orte_job_t);
jdata->jobid = ORTE_PROC_MY_NAME->jobid;
opal_hash_table_set_value_uint32(orte_job_data, jdata->jobid, jdata);
/* every job requires at least one app */
app = OBJ_NEW(orte_app_context_t);
opal_pointer_array_set_item(jdata->apps, 0, app);
jdata->num_apps++;
/* create and store a node object where we are */
node = OBJ_NEW(orte_node_t);
node->name = strdup(orte_process_info.nodename);
node->index = opal_pointer_array_set_item(orte_node_pool, ORTE_PROC_MY_NAME->vpid, node);
/* point our topology to the one detected locally */
node->topology = opal_hwloc_topology;
/* create and store a proc object for us */
proc = OBJ_NEW(orte_proc_t);
proc->name.jobid = ORTE_PROC_MY_NAME->jobid;
proc->name.vpid = ORTE_PROC_MY_NAME->vpid;
proc->pid = orte_process_info.pid;
proc->state = ORTE_PROC_STATE_RUNNING;
opal_pointer_array_set_item(jdata->procs, proc->name.vpid, proc);
/* record that the daemon (i.e., us) is on this node
* NOTE: we do not add the proc object to the node's
* proc array because we are not an application proc.
* Instead, we record it in the daemon field of the
* node object
*/
OBJ_RETAIN(proc); /* keep accounting straight */
node->daemon = proc;
ORTE_FLAG_SET(node, ORTE_NODE_FLAG_DAEMON_LAUNCHED);
node->state = ORTE_NODE_STATE_UP;
/* now point our proc node field to the node */
OBJ_RETAIN(node); /* keep accounting straight */
proc->node = node;
/* record that the daemon job is running */
jdata->num_procs = 1;
jdata->state = ORTE_JOB_STATE_RUNNING;
/* obviously, we have "reported" */
jdata->num_reported = 1;
/* setup the PMIx framework - ensure it skips all non-PMIx components,
* but do not override anything we were given */
opal_setenv("OMPI_MCA_pmix", "^s1,s2,cray,isolated", false, &environ);
if (OPAL_SUCCESS != (ret = mca_base_framework_open(&opal_pmix_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_pmix_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = opal_pmix_base_select())) {
ORTE_ERROR_LOG(ret);
error = "opal_pmix_base_select";
goto error;
}
/* set the event base */
opal_pmix_base_set_evbase(orte_event_base);
/* setup the PMIx server */
if (ORTE_SUCCESS != (ret = pmix_server_init())) {
/* the server code already barked, so let's be quiet */
ret = ORTE_ERR_SILENT;
error = "pmix_server_init";
goto error;
}
/* Setup the communication infrastructure */
/* Routed system */
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_routed_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_routed_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_routed_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_routed_base_select";
goto error;
}
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_oob_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_oob_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_oob_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_oob_base_select";
goto error;
}
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_rml_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_rml_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_rml_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_rml_base_select";
goto error;
}
if (NULL != orte_process_info.my_hnp_uri) {
/* extract the HNP's name so we can update the routing table */
if (ORTE_SUCCESS != (ret = orte_rml_base_parse_uris(orte_process_info.my_hnp_uri,
ORTE_PROC_MY_HNP, NULL))) {
ORTE_ERROR_LOG(ret);
error = "orte_rml_parse_HNP";
goto error;
}
/* Set the contact info in the RML - this won't actually establish
* the connection, but just tells the RML how to reach the HNP
* if/when we attempt to send to it
*/
orte_rml.set_contact_info(orte_process_info.my_hnp_uri);
}
/* select the errmgr */
if (ORTE_SUCCESS != (ret = orte_errmgr_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_errmgr_base_select";
goto error;
}
/* get a conduit for our use - we never route IO over fabric */
OBJ_CONSTRUCT(&transports, opal_list_t);
orte_set_attribute(&transports, ORTE_RML_TRANSPORT_TYPE,
ORTE_ATTR_LOCAL, orte_mgmt_transport, OPAL_STRING);
orte_mgmt_conduit = orte_rml.open_conduit(&transports);
OPAL_LIST_DESTRUCT(&transports);
OBJ_CONSTRUCT(&transports, opal_list_t);
orte_set_attribute(&transports, ORTE_RML_TRANSPORT_TYPE,
ORTE_ATTR_LOCAL, orte_coll_transport, OPAL_STRING);
orte_coll_conduit = orte_rml.open_conduit(&transports);
OPAL_LIST_DESTRUCT(&transports);
/* add our contact info to our proc object */
proc->rml_uri = orte_rml.get_contact_info();
/*
* Group communications
*/
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_grpcomm_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_grpcomm_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_grpcomm_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_grpcomm_base_select";
goto error;
}
/* Open/select the odls */
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_odls_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_odls_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_odls_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_odls_base_select";
goto error;
}
/* Open/select the rtc */
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_rtc_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_rtc_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_rtc_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_rtc_base_select";
goto error;
}
/* be sure to update the routing tree so the initial "phone home"
* to mpirun goes through the tree if static ports were enabled - still
* need to do it anyway just to initialize things
*/
orte_routed.update_routing_plan(NULL);
/* if we are using static ports, then we need to setup
* the daemon info so the RML can function properly
* without requiring a wireup stage. This must be done
* after we enable_comm as that function determines our
* own port, which we need in order to construct the nidmap
*/
if (orte_static_ports) {
/* extract the node info from the environment and
* build a nidmap from it - this will update the
* routing plan as well
*/
if (ORTE_SUCCESS != (ret = orte_util_build_daemon_nidmap(hosts))) {
ORTE_ERROR_LOG(ret);
error = "construct daemon map from static ports";
goto error;
}
}
/* Now provide a chance for the PLM
* to perform any module-specific init functions. This
* needs to occur AFTER the communications are setup
* as it may involve starting a non-blocking recv
* Do this only if a specific PLM was given to us - the
* orted has no need of the proxy PLM at all
*/
if (plm_in_use) {
if (ORTE_SUCCESS != (ret = orte_plm.init())) {
ORTE_ERROR_LOG(ret);
error = "orte_plm_init";
goto error;
}
}
/* setup I/O forwarding system - must come after we init routes */
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_iof_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_iof_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_iof_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_iof_base_select";
goto error;
}
/* setup the FileM */
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_filem_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_filem_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_filem_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_filem_base_select";
goto error;
}
#if OPAL_ENABLE_FT_CR == 1
/*
* Setup the SnapC
*/
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_snapc_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_snapc_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_sstore_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_sstore_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_snapc_base_select(!ORTE_PROC_IS_HNP, ORTE_PROC_IS_DAEMON))) {
ORTE_ERROR_LOG(ret);
error = "orte_snapc_base_select";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_sstore_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_sstore_base_select";
goto error;
}
/* For daemons, ORTE doesn't need the OPAL CR stuff */
opal_cr_set_enabled(false);
#else
opal_cr_set_enabled(false);
#endif
/*
* Initalize the CR setup
* Note: Always do this, even in non-FT builds.
* If we don't some user level tools may hang.
*/
if (ORTE_SUCCESS != (ret = orte_cr_init())) {
ORTE_ERROR_LOG(ret);
error = "orte_cr_init";
goto error;
}
/* setup the DFS framework */
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_dfs_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_dfs_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_dfs_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_dfs_select";
goto error;
}
return ORTE_SUCCESS;
error:
orte_show_help("help-orte-runtime.txt",
"orte_init:startup:internal-failure",
true, error, ORTE_ERROR_NAME(ret), ret);
/* remove our use of the session directory tree */
orte_session_dir_finalize(ORTE_PROC_MY_NAME);
/* ensure we scrub the session directory tree */
orte_session_dir_cleanup(ORTE_JOBID_WILDCARD);
return ORTE_ERR_SILENT;
}
int orte_rmaps_rr_byslot(orte_job_t *jdata,
orte_app_context_t *app,
opal_list_t *node_list,
orte_std_cntr_t num_slots,
orte_vpid_t num_procs)
{
int rc, i, nprocs_mapped;
orte_node_t *node;
orte_proc_t *proc;
int num_procs_to_assign, extra_procs_to_assign=0, nxtra_nodes=0;
hwloc_obj_t obj=NULL;
float balance;
bool add_one=false;
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr: mapping by slot for job %s slots %d num_procs %lu",
ORTE_JOBID_PRINT(jdata->jobid), (int)num_slots, (unsigned long)num_procs);
/* check to see if we can map all the procs */
if (num_slots < ((int)app->num_procs * orte_rmaps_base.cpus_per_rank)) {
if (ORTE_MAPPING_NO_OVERSUBSCRIBE & ORTE_GET_MAPPING_DIRECTIVE(jdata->map->mapping)) {
orte_show_help("help-orte-rmaps-base.txt", "orte-rmaps-base:alloc-error",
true, app->num_procs, app->app);
return ORTE_ERR_SILENT;
}
}
/* first pass: map the number of procs to each node until we
* map all specified procs or use all allocated slots
*/
nprocs_mapped = 0;
OPAL_LIST_FOREACH(node, node_list, orte_node_t) {
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr:slot working node %s",
node->name);
/* get the root object as we are not assigning
* locale here except at the node level
*/
if (NULL != node->topology) {
obj = hwloc_get_root_obj(node->topology);
}
if (node->slots <= node->slots_inuse) {
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr:slot node %s is full - skipping",
node->name);
continue;
}
/* assign a number of procs equal to the number of available
* slots divided by the number of cpus/rank the user
* requested
*/
num_procs_to_assign = (node->slots - node->slots_inuse) / orte_rmaps_base.cpus_per_rank;
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr:slot assigning %d procs to node %s",
(int)num_procs_to_assign, node->name);
for (i=0; i < num_procs_to_assign && nprocs_mapped < app->num_procs; i++) {
/* add this node to the map - do it only once */
if (!ORTE_FLAG_TEST(node, ORTE_NODE_FLAG_MAPPED)) {
if (ORTE_SUCCESS > (rc = opal_pointer_array_add(jdata->map->nodes, (void*)node))) {
ORTE_ERROR_LOG(rc);
return rc;
}
ORTE_FLAG_SET(node, ORTE_NODE_FLAG_MAPPED);
OBJ_RETAIN(node); /* maintain accounting on object */
++(jdata->map->num_nodes);
}
if (NULL == (proc = orte_rmaps_base_setup_proc(jdata, node, app->idx))) {
return ORTE_ERR_OUT_OF_RESOURCE;
}
nprocs_mapped++;
orte_set_attribute(&proc->attributes, ORTE_PROC_HWLOC_LOCALE, ORTE_ATTR_LOCAL, obj, OPAL_PTR);
}
}
int
main(int argc, char *argv[]){
int count;
int msgsize;
uint8_t *msg;
int i, j, rc;
orte_process_name_t peer;
double maxpower;
opal_buffer_t *buf;
orte_rml_recv_cb_t blob;
int conduit_id = 0; //use the first available conduit
struct timeval start, end;
opal_list_t *conduit_attr;
/*
* Init
*/
orte_init(&argc, &argv, ORTE_PROC_NON_MPI);
conduit_attr = OBJ_NEW(opal_list_t);
if( ORTE_SUCCESS ==
( orte_set_attribute( conduit_attr, ORTE_RML_PROVIDER_ATTRIB, ORTE_ATTR_GLOBAL,"sockets",OPAL_STRING))) {
if( ORTE_SUCCESS ==
( orte_set_attribute( conduit_attr, ORTE_RML_INCLUDE_COMP_ATTRIB, ORTE_ATTR_GLOBAL,"ofi",OPAL_STRING))) {
opal_output(0, "%s calling open_conduit with ORTE_RML_INCLUDE_COMP_ATTRIB and ORTE_RML_OFI_PROV_NAME_ATTRIB",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
conduit_id = orte_rml_API_open_conduit(conduit_attr);
if (0 > conduit_id ) {
opal_output(0, "Conduit could not be opened for OFI, exiting");
return;
}
}
}
opal_output(0, "Using conduit-id %d ", conduit_id);
if (argc > 1) {
count = atoi(argv[1]);
if (count < 0) {
count = INT_MAX-1;
}
} else {
count = MAX_COUNT;
}
peer.jobid = ORTE_PROC_MY_NAME->jobid;
peer.vpid = ORTE_PROC_MY_NAME->vpid + 1;
if (peer.vpid == orte_process_info.num_procs) {
peer.vpid = 0;
}
gettimeofday(&start, NULL);
for (j=1; j < count+1; j++) {
/* rank0 starts ring */
if (ORTE_PROC_MY_NAME->vpid == 0) {
/* setup the initiating buffer - put random sized message in it */
buf = OBJ_NEW(opal_buffer_t);
maxpower = (double)(j%7);
msgsize = (int)pow(10.0, maxpower);
opal_output(0, "Ring %d message size %d bytes", j, msgsize);
msg = (uint8_t*)malloc(msgsize);
opal_dss.pack(buf, msg, msgsize, OPAL_BYTE);
free(msg);
orte_rml.send_buffer_nb(conduit_id,&peer, buf, MY_TAG, orte_rml_send_callback, NULL);
/* wait for it to come around */
OBJ_CONSTRUCT(&blob, orte_rml_recv_cb_t);
blob.active = true;
orte_rml.recv_buffer_nb(ORTE_NAME_WILDCARD, MY_TAG,
ORTE_RML_NON_PERSISTENT,
orte_rml_recv_callback, &blob);
ORTE_WAIT_FOR_COMPLETION(blob.active);
OBJ_DESTRUCT(&blob);
opal_output(0, "%s Ring %d completed", ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), j);
} else {
/* wait for msg */
OBJ_CONSTRUCT(&blob, orte_rml_recv_cb_t);
blob.active = true;
orte_rml.recv_buffer_nb(ORTE_NAME_WILDCARD, MY_TAG,
ORTE_RML_NON_PERSISTENT,
orte_rml_recv_callback, &blob);
ORTE_WAIT_FOR_COMPLETION(blob.active);
opal_output(0, "%s received message %d from %s", ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), j, ORTE_NAME_PRINT(&blob.name));
/* send it along */
buf = OBJ_NEW(opal_buffer_t);
opal_dss.copy_payload(buf, &blob.data);
OBJ_DESTRUCT(&blob);
msg_active = true;
orte_rml.send_buffer_nb(conduit_id,&peer, buf, MY_TAG, send_callback, NULL);
ORTE_WAIT_FOR_COMPLETION(msg_active);
}
}
gettimeofday(&end, NULL);
orte_finalize();
printf("start: %d secs, %d usecs\n",start.tv_sec,start.tv_usec);
printf("end: %d secs, %d usecs\n",end.tv_sec,end.tv_usec);
printf("Total minutes = %d, Total seconds = %d", (end.tv_sec - start.tv_sec)/60, (end.tv_sec - start.tv_sec) );
return 0;
}
static int rte_init(void)
{
int ret;
char *error = NULL;
char *contact_path, *jobfam_dir;
orte_job_t *jdata;
orte_node_t *node;
orte_proc_t *proc;
orte_app_context_t *app;
char **aliases, *aptr;
/* run the prolog */
if (ORTE_SUCCESS != (ret = orte_ess_base_std_prolog())) {
error = "orte_ess_base_std_prolog";
goto error;
}
/* setup callback for SIGPIPE */
setup_sighandler(SIGPIPE, &epipe_handler, epipe_signal_callback);
/** setup callbacks for abort signals - from this point
* forward, we need to abort in a manner that allows us
* to cleanup. However, we cannot directly use libevent
* to trap these signals as otherwise we cannot respond
* to them if we are stuck in an event! So instead use
* the basic POSIX trap functions to handle the signal,
* and then let that signal handler do some magic to
* avoid the hang
*
* NOTE: posix traps don't allow us to do anything major
* in them, so use a pipe tied to a libevent event to
* reach a "safe" place where the termination event can
* be created
*/
pipe(term_pipe);
/* setup an event to attempt normal termination on signal */
opal_event_set(orte_event_base, &term_handler, term_pipe[0], OPAL_EV_READ, clean_abort, NULL);
opal_event_set_priority(&term_handler, ORTE_ERROR_PRI);
opal_event_add(&term_handler, NULL);
/* Set both ends of this pipe to be close-on-exec so that no
children inherit it */
if (opal_fd_set_cloexec(term_pipe[0]) != OPAL_SUCCESS ||
opal_fd_set_cloexec(term_pipe[1]) != OPAL_SUCCESS) {
error = "unable to set the pipe to CLOEXEC";
goto error;
}
/* point the signal trap to a function that will activate that event */
signal(SIGTERM, abort_signal_callback);
signal(SIGINT, abort_signal_callback);
signal(SIGHUP, abort_signal_callback);
/** setup callbacks for signals we should foward */
setup_sighandler(SIGUSR1, &sigusr1_handler, signal_forward_callback);
setup_sighandler(SIGUSR2, &sigusr2_handler, signal_forward_callback);
setup_sighandler(SIGTSTP, &sigtstp_handler, signal_forward_callback);
setup_sighandler(SIGCONT, &sigcont_handler, signal_forward_callback);
signals_set = true;
#if OPAL_HAVE_HWLOC
{
hwloc_obj_t obj;
unsigned i, j;
/* get the local topology */
if (NULL == opal_hwloc_topology) {
if (OPAL_SUCCESS != opal_hwloc_base_get_topology()) {
error = "topology discovery";
goto error;
}
}
/* remove the hostname from the topology. Unfortunately, hwloc
* decided to add the source hostname to the "topology", thus
* rendering it unusable as a pure topological description. So
* we remove that information here.
*/
obj = hwloc_get_root_obj(opal_hwloc_topology);
for (i=0; i < obj->infos_count; i++) {
if (NULL == obj->infos[i].name ||
NULL == obj->infos[i].value) {
continue;
}
if (0 == strncmp(obj->infos[i].name, "HostName", strlen("HostName"))) {
free(obj->infos[i].name);
free(obj->infos[i].value);
/* left justify the array */
for (j=i; j < obj->infos_count-1; j++) {
obj->infos[j] = obj->infos[j+1];
}
obj->infos[obj->infos_count-1].name = NULL;
obj->infos[obj->infos_count-1].value = NULL;
obj->infos_count--;
break;
}
}
if (4 < opal_output_get_verbosity(orte_ess_base_framework.framework_output)) {
opal_output(0, "%s Topology Info:", ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
opal_dss.dump(0, opal_hwloc_topology, OPAL_HWLOC_TOPO);
}
}
#endif
/* if we are using xml for output, put an mpirun start tag */
if (orte_xml_output) {
fprintf(orte_xml_fp, "<mpirun>\n");
fflush(orte_xml_fp);
}
/* open and setup the opal_pstat framework so we can provide
* process stats if requested
*/
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&opal_pstat_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "opal_pstat_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = opal_pstat_base_select())) {
ORTE_ERROR_LOG(ret);
error = "opal_pstat_base_select";
goto error;
}
/* open and setup the state machine */
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_state_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_state_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_state_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_state_base_select";
goto error;
}
/* open the errmgr */
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_errmgr_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_errmgr_base_open";
goto error;
}
/* Since we are the HNP, then responsibility for
* defining the name falls to the PLM component for our
* respective environment - hence, we have to open the PLM
* first and select that component.
*/
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_plm_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_plm_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_plm_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_plm_base_select";
goto error;
}
/* if we were spawned by a singleton, our jobid was given to us */
if (NULL != orte_ess_base_jobid) {
if (ORTE_SUCCESS != (ret = orte_util_convert_string_to_jobid(&ORTE_PROC_MY_NAME->jobid, orte_ess_base_jobid))) {
ORTE_ERROR_LOG(ret);
error = "convert_string_to_jobid";
goto error;
}
ORTE_PROC_MY_NAME->vpid = 0;
} else {
if (ORTE_SUCCESS != (ret = orte_plm.set_hnp_name())) {
ORTE_ERROR_LOG(ret);
error = "orte_plm_set_hnp_name";
goto error;
}
}
/* Setup the communication infrastructure */
/*
* OOB Layer
*/
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_oob_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_oob_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_oob_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_oob_base_select";
goto error;
}
/*
* Runtime Messaging Layer
*/
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_rml_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_rml_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_rml_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_rml_base_select";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_errmgr_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_errmgr_base_select";
goto error;
}
/* setup the global job and node arrays */
orte_job_data = OBJ_NEW(opal_pointer_array_t);
if (ORTE_SUCCESS != (ret = opal_pointer_array_init(orte_job_data,
1,
ORTE_GLOBAL_ARRAY_MAX_SIZE,
1))) {
ORTE_ERROR_LOG(ret);
error = "setup job array";
goto error;
}
orte_node_pool = OBJ_NEW(opal_pointer_array_t);
if (ORTE_SUCCESS != (ret = opal_pointer_array_init(orte_node_pool,
ORTE_GLOBAL_ARRAY_BLOCK_SIZE,
ORTE_GLOBAL_ARRAY_MAX_SIZE,
ORTE_GLOBAL_ARRAY_BLOCK_SIZE))) {
ORTE_ERROR_LOG(ret);
error = "setup node array";
goto error;
}
orte_node_topologies = OBJ_NEW(opal_pointer_array_t);
if (ORTE_SUCCESS != (ret = opal_pointer_array_init(orte_node_topologies,
ORTE_GLOBAL_ARRAY_BLOCK_SIZE,
ORTE_GLOBAL_ARRAY_MAX_SIZE,
ORTE_GLOBAL_ARRAY_BLOCK_SIZE))) {
ORTE_ERROR_LOG(ret);
error = "setup node topologies array";
goto error;
}
/* init the nidmap - just so we register that verbosity */
orte_util_nidmap_init(NULL);
/* Setup the job data object for the daemons */
/* create and store the job data object */
jdata = OBJ_NEW(orte_job_t);
jdata->jobid = ORTE_PROC_MY_NAME->jobid;
opal_pointer_array_set_item(orte_job_data, 0, jdata);
/* mark that the daemons have reported as we are the
* only ones in the system right now, and we definitely
* are running!
*/
jdata->state = ORTE_JOB_STATE_DAEMONS_REPORTED;
/* every job requires at least one app */
app = OBJ_NEW(orte_app_context_t);
opal_pointer_array_set_item(jdata->apps, 0, app);
jdata->num_apps++;
/* create and store a node object where we are */
node = OBJ_NEW(orte_node_t);
node->name = strdup(orte_process_info.nodename);
node->index = opal_pointer_array_set_item(orte_node_pool, 0, node);
#if OPAL_HAVE_HWLOC
/* add it to the array of known topologies */
opal_pointer_array_add(orte_node_topologies, opal_hwloc_topology);
#endif
/* create and store a proc object for us */
proc = OBJ_NEW(orte_proc_t);
proc->name.jobid = ORTE_PROC_MY_NAME->jobid;
proc->name.vpid = ORTE_PROC_MY_NAME->vpid;
proc->pid = orte_process_info.pid;
proc->rml_uri = orte_rml.get_contact_info();
proc->state = ORTE_PROC_STATE_RUNNING;
OBJ_RETAIN(node); /* keep accounting straight */
proc->node = node;
opal_pointer_array_set_item(jdata->procs, proc->name.vpid, proc);
/* record that the daemon (i.e., us) is on this node
* NOTE: we do not add the proc object to the node's
* proc array because we are not an application proc.
* Instead, we record it in the daemon field of the
* node object
*/
OBJ_RETAIN(proc); /* keep accounting straight */
node->daemon = proc;
ORTE_FLAG_SET(node, ORTE_NODE_FLAG_DAEMON_LAUNCHED);
node->state = ORTE_NODE_STATE_UP;
/* if we are to retain aliases, get ours */
if (orte_retain_aliases) {
aliases = NULL;
opal_ifgetaliases(&aliases);
/* add our own local name to it */
opal_argv_append_nosize(&aliases, orte_process_info.nodename);
aptr = opal_argv_join(aliases, ',');
opal_argv_free(aliases);
orte_set_attribute(&node->attributes, ORTE_NODE_ALIAS, ORTE_ATTR_LOCAL, aptr, OPAL_STRING);
free(aptr);
}
/* record that the daemon job is running */
jdata->num_procs = 1;
jdata->state = ORTE_JOB_STATE_RUNNING;
/* obviously, we have "reported" */
jdata->num_reported = 1;
/*
* Routed system
*/
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_routed_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_rml_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_routed_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_routed_base_select";
goto error;
}
/* datastore - ensure we don't pickup the pmi component, but
* don't override anything set by user
*/
if (NULL == getenv("OMPI_MCA_dstore")) {
putenv("OMPI_MCA_dstore=^pmi");
}
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&opal_dstore_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "opal_dstore_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = opal_dstore_base_select())) {
ORTE_ERROR_LOG(ret);
error = "opal_dstore_base_select";
goto error;
}
/* create the handles */
if (0 > (opal_dstore_peer = opal_dstore.open("PEER"))) {
error = "opal dstore global";
ret = ORTE_ERR_FATAL;
goto error;
}
if (0 > (opal_dstore_internal = opal_dstore.open("INTERNAL"))) {
error = "opal dstore internal";
ret = ORTE_ERR_FATAL;
goto error;
}
if (0 > (opal_dstore_nonpeer = opal_dstore.open("NONPEER"))) {
error = "opal dstore nonpeer";
ret = ORTE_ERR_FATAL;
goto error;
}
/*
* Group communications
*/
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_grpcomm_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_grpcomm_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_grpcomm_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_grpcomm_base_select";
goto error;
}
/* Now provide a chance for the PLM
* to perform any module-specific init functions. This
* needs to occur AFTER the communications are setup
* as it may involve starting a non-blocking recv
*/
if (ORTE_SUCCESS != (ret = orte_plm.init())) {
ORTE_ERROR_LOG(ret);
error = "orte_plm_init";
goto error;
}
/*
* Setup the remaining resource
* management and errmgr frameworks - application procs
* and daemons do not open these frameworks as they only use
* the hnp proxy support in the PLM framework.
*/
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_ras_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_ras_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_ras_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_ras_base_find_available";
goto error;
}
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_rmaps_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_rmaps_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_rmaps_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_rmaps_base_find_available";
goto error;
}
#if OPAL_HAVE_HWLOC
{
char *coprocessors, **sns;
uint32_t h;
int idx;
/* if a topology file was given, then the rmaps framework open
* will have reset our topology. Ensure we always get the right
* one by setting our node topology afterwards
*/
node->topology = opal_hwloc_topology;
/* init the hash table, if necessary */
if (NULL == orte_coprocessors) {
orte_coprocessors = OBJ_NEW(opal_hash_table_t);
opal_hash_table_init(orte_coprocessors, orte_process_info.num_procs);
}
/* detect and add any coprocessors */
coprocessors = opal_hwloc_base_find_coprocessors(opal_hwloc_topology);
if (NULL != coprocessors) {
/* separate the serial numbers of the coprocessors
* on this host
*/
sns = opal_argv_split(coprocessors, ',');
for (idx=0; NULL != sns[idx]; idx++) {
/* compute the hash */
OPAL_HASH_STR(sns[idx], h);
/* mark that this coprocessor is hosted by this node */
opal_hash_table_set_value_uint32(orte_coprocessors, h, (void*)&(ORTE_PROC_MY_NAME->vpid));
}
opal_argv_free(sns);
free(coprocessors);
orte_coprocessors_detected = true;
}
/* see if I am on a coprocessor */
coprocessors = opal_hwloc_base_check_on_coprocessor();
if (NULL != coprocessors) {
orte_set_attribute(&node->attributes, ORTE_NODE_SERIAL_NUMBER, ORTE_ATTR_LOCAL, coprocessors, OPAL_STRING);
free(coprocessors);
orte_coprocessors_detected = true;
}
}
#endif
/* Open/select the odls */
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_odls_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_odls_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_odls_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_odls_base_select";
goto error;
}
/* Open/select the rtc */
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_rtc_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_rtc_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_rtc_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_rtc_base_select";
goto error;
}
/* enable communication with the rml */
if (ORTE_SUCCESS != (ret = orte_rml.enable_comm())) {
ORTE_ERROR_LOG(ret);
error = "orte_rml.enable_comm";
goto error;
}
/* we are an hnp, so update the contact info field for later use */
orte_process_info.my_hnp_uri = orte_rml.get_contact_info();
proc->rml_uri = strdup(orte_process_info.my_hnp_uri);
/* we are also officially a daemon, so better update that field too */
orte_process_info.my_daemon_uri = strdup(orte_process_info.my_hnp_uri);
/* setup the orte_show_help system to recv remote output */
orte_rml.recv_buffer_nb(ORTE_NAME_WILDCARD, ORTE_RML_TAG_SHOW_HELP,
ORTE_RML_PERSISTENT, orte_show_help_recv, NULL);
/* setup my session directory */
if (orte_create_session_dirs) {
OPAL_OUTPUT_VERBOSE((2, orte_debug_output,
"%s setting up session dir with\n\ttmpdir: %s\n\thost %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
(NULL == orte_process_info.tmpdir_base) ? "UNDEF" : orte_process_info.tmpdir_base,
orte_process_info.nodename));
/* take a pass thru the session directory code to fillin the
* tmpdir names - don't create anything yet
*/
if (ORTE_SUCCESS != (ret = orte_session_dir(false,
orte_process_info.tmpdir_base,
orte_process_info.nodename, NULL,
ORTE_PROC_MY_NAME))) {
ORTE_ERROR_LOG(ret);
error = "orte_session_dir define";
goto error;
}
/* clear the session directory just in case there are
* stale directories laying around
*/
orte_session_dir_cleanup(ORTE_JOBID_WILDCARD);
/* now actually create the directory tree */
if (ORTE_SUCCESS != (ret = orte_session_dir(true,
orte_process_info.tmpdir_base,
orte_process_info.nodename, NULL,
ORTE_PROC_MY_NAME))) {
ORTE_ERROR_LOG(ret);
error = "orte_session_dir";
goto error;
}
/* Once the session directory location has been established, set
the opal_output hnp file location to be in the
proc-specific session directory. */
opal_output_set_output_file_info(orte_process_info.proc_session_dir,
"output-", NULL, NULL);
/* save my contact info in a file for others to find */
jobfam_dir = opal_dirname(orte_process_info.job_session_dir);
contact_path = opal_os_path(false, jobfam_dir, "contact.txt", NULL);
free(jobfam_dir);
OPAL_OUTPUT_VERBOSE((2, orte_debug_output,
"%s writing contact file %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
contact_path));
if (ORTE_SUCCESS != (ret = orte_write_hnp_contact_file(contact_path))) {
OPAL_OUTPUT_VERBOSE((2, orte_debug_output,
"%s writing contact file failed with error %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_ERROR_NAME(ret)));
} else {
OPAL_OUTPUT_VERBOSE((2, orte_debug_output,
"%s wrote contact file",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
}
free(contact_path);
}
/* setup the routed info - the selected routed component
* will know what to do.
*/
if (ORTE_SUCCESS != (ret = orte_routed.init_routes(ORTE_PROC_MY_NAME->jobid, NULL))) {
ORTE_ERROR_LOG(ret);
error = "orte_routed.init_routes";
goto error;
}
/* setup I/O forwarding system - must come after we init routes */
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_iof_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_iof_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_iof_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_iof_base_select";
goto error;
}
/* setup the FileM */
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_filem_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_filem_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_filem_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_filem_base_select";
goto error;
}
#if OPAL_ENABLE_FT_CR == 1
/*
* Setup the SnapC
*/
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_snapc_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_snapc_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_sstore_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_sstore_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_snapc_base_select(ORTE_PROC_IS_HNP, ORTE_PROC_IS_APP))) {
ORTE_ERROR_LOG(ret);
error = "orte_snapc_base_select";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_sstore_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_sstore_base_select";
goto error;
}
/* For HNP, ORTE doesn't need the OPAL CR stuff */
opal_cr_set_enabled(false);
#else
opal_cr_set_enabled(false);
#endif
/*
* Initalize the CR setup
* Note: Always do this, even in non-FT builds.
* If we don't some user level tools may hang.
*/
if (ORTE_SUCCESS != (ret = orte_cr_init())) {
ORTE_ERROR_LOG(ret);
error = "orte_cr_init";
goto error;
}
/* setup the dfs framework */
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_dfs_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_dfs_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_dfs_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_dfs_select";
goto error;
}
/* if a tool has launched us and is requesting event reports,
* then set its contact info into the comm system
*/
if (orte_report_events) {
if (ORTE_SUCCESS != (ret = orte_util_comm_connect_tool(orte_report_events_uri))) {
error = "could not connect to tool";
goto error;
}
}
/* We actually do *not* want an HNP to voluntarily yield() the
processor more than necessary. Orterun already blocks when
it is doing nothing, so it doesn't use any more CPU cycles than
it should; but when it *is* doing something, we do not want it
to be unnecessarily delayed because it voluntarily yielded the
processor in the middle of its work.
For example: when a message arrives at orterun, we want the
OS to wake us up in a timely fashion (which most OS's
seem good about doing) and then we want orterun to process
the message as fast as possible. If orterun yields and lets
aggressive MPI applications get the processor back, it may be a
long time before the OS schedules orterun to run again
(particularly if there is no IO event to wake it up). Hence,
routed OOB messages (for example) may be significantly delayed
before being delivered to MPI processes, which can be
problematic in some scenarios (e.g., COMM_SPAWN, BTL's that
require OOB messages for wireup, etc.). */
opal_progress_set_yield_when_idle(false);
return ORTE_SUCCESS;
error:
if (ORTE_ERR_SILENT != ret && !orte_report_silent_errors) {
orte_show_help("help-orte-runtime.txt",
"orte_init:startup:internal-failure",
true, error, ORTE_ERROR_NAME(ret), ret);
}
return ORTE_ERR_SILENT;
}
static int bind_in_place(orte_job_t *jdata,
hwloc_obj_type_t target,
unsigned cache_level)
{
/* traverse the hwloc topology tree on each node downwards
* until we find an unused object of type target - and then bind
* the process to that target
*/
int i, j;
orte_job_map_t *map;
orte_node_t *node;
orte_proc_t *proc;
hwloc_cpuset_t cpus;
unsigned int idx, ncpus;
struct hwloc_topology_support *support;
opal_hwloc_obj_data_t *data;
hwloc_obj_t locale, sib;
char *cpu_bitmap;
bool found;
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps: bind in place for job %s with bindings %s",
ORTE_JOBID_PRINT(jdata->jobid),
opal_hwloc_base_print_binding(jdata->map->binding));
/* initialize */
map = jdata->map;
for (i=0; i < map->nodes->size; i++) {
if (NULL == (node = (orte_node_t*)opal_pointer_array_get_item(map->nodes, i))) {
continue;
}
if (!orte_do_not_launch) {
/* if we don't want to launch, then we are just testing the system,
* so ignore questions about support capabilities
*/
support = (struct hwloc_topology_support*)hwloc_topology_get_support(node->topology);
/* check if topology supports cpubind - have to be careful here
* as Linux doesn't currently support thread-level binding. This
* may change in the future, though, and it isn't clear how hwloc
* interprets the current behavior. So check both flags to be sure.
*/
if (!support->cpubind->set_thisproc_cpubind &&
!support->cpubind->set_thisthread_cpubind) {
if (!OPAL_BINDING_REQUIRED(map->binding) ||
!OPAL_BINDING_POLICY_IS_SET(map->binding)) {
/* we are not required to bind, so ignore this */
continue;
}
orte_show_help("help-orte-rmaps-base.txt", "rmaps:cpubind-not-supported", true, node->name);
return ORTE_ERR_SILENT;
}
/* check if topology supports membind - have to be careful here
* as hwloc treats this differently than I (at least) would have
* expected. Per hwloc, Linux memory binding is at the thread,
* and not process, level. Thus, hwloc sets the "thisproc" flag
* to "false" on all Linux systems, and uses the "thisthread" flag
* to indicate binding capability - don't warn if the user didn't
* specifically request binding
*/
if (!support->membind->set_thisproc_membind &&
!support->membind->set_thisthread_membind &&
OPAL_BINDING_POLICY_IS_SET(map->binding)) {
if (OPAL_HWLOC_BASE_MBFA_WARN == opal_hwloc_base_mbfa && !membind_warned) {
orte_show_help("help-orte-rmaps-base.txt", "rmaps:membind-not-supported", true, node->name);
membind_warned = true;
} else if (OPAL_HWLOC_BASE_MBFA_ERROR == opal_hwloc_base_mbfa) {
orte_show_help("help-orte-rmaps-base.txt", "rmaps:membind-not-supported-fatal", true, node->name);
return ORTE_ERR_SILENT;
}
}
}
/* some systems do not report cores, and so we can get a situation where our
* default binding policy will fail for no necessary reason. So if we are
* computing a binding due to our default policy, and no cores are found
* on this node, just silently skip it - we will not bind
*/
if (!OPAL_BINDING_POLICY_IS_SET(map->binding) &&
HWLOC_TYPE_DEPTH_UNKNOWN == hwloc_get_type_depth(node->topology, HWLOC_OBJ_CORE)) {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"Unable to bind-to core by default on node %s as no cores detected",
node->name);
continue;
}
/* we share topologies in order
* to save space, so we need to reset the usage info to reflect
* our own current state
*/
reset_usage(node, jdata->jobid);
/* cycle thru the procs */
for (j=0; j < node->procs->size; j++) {
if (NULL == (proc = (orte_proc_t*)opal_pointer_array_get_item(node->procs, j))) {
continue;
}
/* ignore procs from other jobs */
if (proc->name.jobid != jdata->jobid) {
continue;
}
/* bozo check */
if (!orte_get_attribute(&proc->attributes, ORTE_PROC_HWLOC_LOCALE, (void**)&locale, OPAL_PTR)) {
orte_show_help("help-orte-rmaps-base.txt", "rmaps:no-locale", true, ORTE_NAME_PRINT(&proc->name));
return ORTE_ERR_SILENT;
}
/* get the index of this location */
if (UINT_MAX == (idx = opal_hwloc_base_get_obj_idx(node->topology, locale, OPAL_HWLOC_AVAILABLE))) {
ORTE_ERROR_LOG(ORTE_ERR_BAD_PARAM);
return ORTE_ERR_SILENT;
}
data = (opal_hwloc_obj_data_t*)locale->userdata;
/* get the number of cpus under this location */
if (0 == (ncpus = opal_hwloc_base_get_npus(node->topology, locale))) {
orte_show_help("help-orte-rmaps-base.txt", "rmaps:no-available-cpus", true, node->name);
return ORTE_ERR_SILENT;
}
/* if we don't have enough cpus to support this additional proc, try
* shifting the location to a cousin that can support it - the important
* thing is that we maintain the same level in the topology */
if (ncpus < (data->num_bound+1)) {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"%s bind_in_place: searching right",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
sib = locale;
found = false;
while (NULL != (sib = sib->next_cousin)) {
data = (opal_hwloc_obj_data_t*)sib->userdata;
ncpus = opal_hwloc_base_get_npus(node->topology, sib);
if (data->num_bound < ncpus) {
found = true;
locale = sib;
break;
}
}
if (!found) {
/* try the other direction */
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"%s bind_in_place: searching left",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
sib = locale;
while (NULL != (sib = sib->prev_cousin)) {
data = (opal_hwloc_obj_data_t*)sib->userdata;
ncpus = opal_hwloc_base_get_npus(node->topology, sib);
if (data->num_bound < ncpus) {
found = true;
locale = sib;
break;
}
}
}
if (!found) {
/* no place to put this - see if overload is allowed */
if (!OPAL_BIND_OVERLOAD_ALLOWED(jdata->map->binding)) {
if (OPAL_BINDING_POLICY_IS_SET(jdata->map->binding)) {
/* if the user specified a binding policy, then we cannot meet
* it since overload isn't allowed, so error out - have the
* message indicate that setting overload allowed will remove
* this restriction */
orte_show_help("help-orte-rmaps-base.txt", "rmaps:binding-overload", true,
opal_hwloc_base_print_binding(map->binding), node->name,
data->num_bound, ncpus);
return ORTE_ERR_SILENT;
} else {
/* if we have the default binding policy, then just don't bind */
OPAL_SET_BINDING_POLICY(map->binding, OPAL_BIND_TO_NONE);
unbind_procs(jdata);
return ORTE_SUCCESS;
}
}
}
}
/* track the number bound */
data = (opal_hwloc_obj_data_t*)locale->userdata; // just in case it changed
data->num_bound++;
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"BINDING PROC %s TO %s NUMBER %u",
ORTE_NAME_PRINT(&proc->name),
hwloc_obj_type_string(locale->type), idx);
/* bind the proc here */
cpus = opal_hwloc_base_get_available_cpus(node->topology, locale);
hwloc_bitmap_list_asprintf(&cpu_bitmap, cpus);
orte_set_attribute(&proc->attributes, ORTE_PROC_CPU_BITMAP, ORTE_ATTR_GLOBAL, cpu_bitmap, OPAL_STRING);
/* update the location, in case it changed */
orte_set_attribute(&proc->attributes, ORTE_PROC_HWLOC_BOUND, ORTE_ATTR_LOCAL, locale, OPAL_PTR);
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"%s BOUND PROC %s TO %s[%s:%u] on node %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&proc->name),
cpu_bitmap, hwloc_obj_type_string(locale->type),
idx, node->name);
if (NULL != cpu_bitmap) {
free(cpu_bitmap);
}
}
}
return ORTE_SUCCESS;
}
static int bind_downwards(orte_job_t *jdata,
orte_node_t *node,
hwloc_obj_type_t target,
unsigned cache_level)
{
int j;
orte_job_map_t *map;
orte_proc_t *proc;
hwloc_obj_t trg_obj, nxt_obj;
hwloc_cpuset_t cpus;
unsigned int ncpus;
opal_hwloc_obj_data_t *data;
int total_cpus;
hwloc_cpuset_t totalcpuset;
hwloc_obj_t locale;
char *cpu_bitmap;
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps: bind downward for job %s with bindings %s",
ORTE_JOBID_PRINT(jdata->jobid),
opal_hwloc_base_print_binding(jdata->map->binding));
/* initialize */
map = jdata->map;
totalcpuset = hwloc_bitmap_alloc();
/* cycle thru the procs */
for (j=0; j < node->procs->size; j++) {
if (NULL == (proc = (orte_proc_t*)opal_pointer_array_get_item(node->procs, j))) {
continue;
}
/* ignore procs from other jobs */
if (proc->name.jobid != jdata->jobid) {
continue;
}
/* bozo check */
locale = NULL;
if (!orte_get_attribute(&proc->attributes, ORTE_PROC_HWLOC_LOCALE, (void**)&locale, OPAL_PTR)) {
orte_show_help("help-orte-rmaps-base.txt", "rmaps:no-locale", true, ORTE_NAME_PRINT(&proc->name));
hwloc_bitmap_free(totalcpuset);
return ORTE_ERR_SILENT;
}
/* we don't know if the target is a direct child of this locale,
* or if it is some depth below it, so we have to conduct a bit
* of a search. Let hwloc find the min usage one for us.
*/
trg_obj = opal_hwloc_base_find_min_bound_target_under_obj(node->topology, locale,
target, cache_level);
if (NULL == trg_obj) {
/* there aren't any such targets under this object */
orte_show_help("help-orte-rmaps-base.txt", "rmaps:no-available-cpus", true, node->name);
hwloc_bitmap_free(totalcpuset);
return ORTE_ERR_SILENT;
}
/* record the location */
orte_set_attribute(&proc->attributes, ORTE_PROC_HWLOC_BOUND, ORTE_ATTR_LOCAL, trg_obj, OPAL_PTR);
/* start with a clean slate */
hwloc_bitmap_zero(totalcpuset);
total_cpus = 0;
nxt_obj = trg_obj;
do {
if (NULL == nxt_obj) {
/* could not find enough cpus to meet request */
orte_show_help("help-orte-rmaps-base.txt", "rmaps:no-available-cpus", true, node->name);
hwloc_bitmap_free(totalcpuset);
return ORTE_ERR_SILENT;
}
trg_obj = nxt_obj;
/* get the number of cpus under this location */
ncpus = opal_hwloc_base_get_npus(node->topology, trg_obj);
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"%s GOT %d CPUS",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), ncpus);
/* track the number bound */
if (NULL == (data = (opal_hwloc_obj_data_t*)trg_obj->userdata)) {
data = OBJ_NEW(opal_hwloc_obj_data_t);
trg_obj->userdata = data;
}
data->num_bound++;
/* error out if adding a proc would cause overload and that wasn't allowed,
* and it wasn't a default binding policy (i.e., the user requested it)
*/
if (ncpus < data->num_bound &&
!OPAL_BIND_OVERLOAD_ALLOWED(jdata->map->binding)) {
if (OPAL_BINDING_POLICY_IS_SET(jdata->map->binding)) {
/* if the user specified a binding policy, then we cannot meet
* it since overload isn't allowed, so error out - have the
* message indicate that setting overload allowed will remove
* this restriction */
orte_show_help("help-orte-rmaps-base.txt", "rmaps:binding-overload", true,
opal_hwloc_base_print_binding(map->binding), node->name,
data->num_bound, ncpus);
hwloc_bitmap_free(totalcpuset);
return ORTE_ERR_SILENT;
} else {
/* if we have the default binding policy, then just don't bind */
OPAL_SET_BINDING_POLICY(map->binding, OPAL_BIND_TO_NONE);
unbind_procs(jdata);
hwloc_bitmap_zero(totalcpuset);
return ORTE_SUCCESS;
}
}
/* bind the proc here */
cpus = opal_hwloc_base_get_available_cpus(node->topology, trg_obj);
hwloc_bitmap_or(totalcpuset, totalcpuset, cpus);
/* track total #cpus */
total_cpus += ncpus;
/* move to the next location, in case we need it */
nxt_obj = trg_obj->next_cousin;
} while (total_cpus < orte_rmaps_base.cpus_per_rank);
hwloc_bitmap_list_asprintf(&cpu_bitmap, totalcpuset);
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"%s PROC %s BITMAP %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&proc->name), cpu_bitmap);
orte_set_attribute(&proc->attributes, ORTE_PROC_CPU_BITMAP, ORTE_ATTR_GLOBAL, cpu_bitmap, OPAL_STRING);
if (NULL != cpu_bitmap) {
free(cpu_bitmap);
}
if (4 < opal_output_get_verbosity(orte_rmaps_base_framework.framework_output)) {
char tmp1[1024], tmp2[1024];
if (OPAL_ERR_NOT_BOUND == opal_hwloc_base_cset2str(tmp1, sizeof(tmp1),
node->topology, totalcpuset)) {
opal_output(orte_rmaps_base_framework.framework_output,
"%s PROC %s ON %s IS NOT BOUND",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&proc->name), node->name);
} else {
opal_hwloc_base_cset2mapstr(tmp2, sizeof(tmp2), node->topology, totalcpuset);
opal_output(orte_rmaps_base_framework.framework_output,
"%s BOUND PROC %s[%s] TO %s: %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&proc->name), node->name,
tmp1, tmp2);
}
}
}
hwloc_bitmap_free(totalcpuset);
return ORTE_SUCCESS;
}
static int bind_upwards(orte_job_t *jdata,
orte_node_t *node,
hwloc_obj_type_t target,
unsigned cache_level)
{
/* traverse the hwloc topology tree on each node upwards
* until we find an object of type target - and then bind
* the process to that target
*/
int j;
orte_job_map_t *map;
orte_proc_t *proc;
hwloc_obj_t obj;
hwloc_cpuset_t cpus;
unsigned int idx, ncpus;
opal_hwloc_obj_data_t *data;
hwloc_obj_t locale;
char *cpu_bitmap;
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps: bind upwards for job %s with bindings %s",
ORTE_JOBID_PRINT(jdata->jobid),
opal_hwloc_base_print_binding(jdata->map->binding));
/* initialize */
map = jdata->map;
/* cycle thru the procs */
for (j=0; j < node->procs->size; j++) {
if (NULL == (proc = (orte_proc_t*)opal_pointer_array_get_item(node->procs, j))) {
continue;
}
/* ignore procs from other jobs */
if (proc->name.jobid != jdata->jobid) {
continue;
}
/* bozo check */
if (!orte_get_attribute(&proc->attributes, ORTE_PROC_HWLOC_LOCALE, (void**)&locale, OPAL_PTR)) {
orte_show_help("help-orte-rmaps-base.txt", "rmaps:no-locale", true, ORTE_NAME_PRINT(&proc->name));
return ORTE_ERR_SILENT;
}
/* starting at the locale, move up thru the parents
* to find the target object type
*/
cpu_bitmap = NULL;
for (obj = locale->parent; NULL != obj; obj = obj->parent) {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"%s bind:upward target %s type %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
hwloc_obj_type_string(target),
hwloc_obj_type_string(obj->type));
if (target == obj->type) {
if (HWLOC_OBJ_CACHE == target && cache_level != obj->attr->cache.depth) {
continue;
}
/* get its index */
if (UINT_MAX == (idx = opal_hwloc_base_get_obj_idx(node->topology, obj, OPAL_HWLOC_AVAILABLE))) {
ORTE_ERROR_LOG(ORTE_ERR_BAD_PARAM);
return ORTE_ERR_SILENT;
}
/* track the number bound */
data = (opal_hwloc_obj_data_t*)obj->userdata;
data->num_bound++;
/* get the number of cpus under this location */
if (0 == (ncpus = opal_hwloc_base_get_npus(node->topology, obj))) {
orte_show_help("help-orte-rmaps-base.txt", "rmaps:no-available-cpus", true, node->name);
return ORTE_ERR_SILENT;
}
/* error out if adding a proc would cause overload and that wasn't allowed,
* and it wasn't a default binding policy (i.e., the user requested it)
*/
if (ncpus < data->num_bound &&
!OPAL_BIND_OVERLOAD_ALLOWED(jdata->map->binding)) {
if (OPAL_BINDING_POLICY_IS_SET(jdata->map->binding)) {
/* if the user specified a binding policy, then we cannot meet
* it since overload isn't allowed, so error out - have the
* message indicate that setting overload allowed will remove
* this restriction */
orte_show_help("help-orte-rmaps-base.txt", "rmaps:binding-overload", true,
opal_hwloc_base_print_binding(map->binding), node->name,
data->num_bound, ncpus);
return ORTE_ERR_SILENT;
} else {
/* if we have the default binding policy, then just don't bind */
OPAL_SET_BINDING_POLICY(map->binding, OPAL_BIND_TO_NONE);
unbind_procs(jdata);
return ORTE_SUCCESS;
}
}
/* bind it here */
cpus = opal_hwloc_base_get_available_cpus(node->topology, obj);
hwloc_bitmap_list_asprintf(&cpu_bitmap, cpus);
orte_set_attribute(&proc->attributes, ORTE_PROC_CPU_BITMAP, ORTE_ATTR_GLOBAL, cpu_bitmap, OPAL_STRING);
/* record the location */
orte_set_attribute(&proc->attributes, ORTE_PROC_HWLOC_BOUND, ORTE_ATTR_LOCAL, obj, OPAL_PTR);
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"%s BOUND PROC %s TO %s[%s:%u] on node %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&proc->name),
cpu_bitmap,
hwloc_obj_type_string(target),
idx, node->name);
break;
}
}
if (NULL == cpu_bitmap && OPAL_BINDING_REQUIRED(jdata->map->binding)) {
/* didn't find anyone to bind to - this is an error
* unless the user specified if-supported
*/
orte_show_help("help-orte-rmaps-base.txt", "rmaps:binding-target-not-found", true,
opal_hwloc_base_print_binding(map->binding), node->name);
return ORTE_ERR_SILENT;
}
if (NULL != cpu_bitmap) {
free(cpu_bitmap);
}
}
return ORTE_SUCCESS;
}
int main(int argc, char *argv[])
{
int rc, i, j;
opal_cmd_line_t cmd_line;
char *param, *value;
orte_job_t *jdata=NULL;
orte_app_context_t *app;
char *uri, *ptr;
/* Setup and parse the command line */
memset(&myglobals, 0, sizeof(myglobals));
/* find our basename (the name of the executable) so that we can
use it in pretty-print error messages */
myglobals.basename = opal_basename(argv[0]);
opal_cmd_line_create(&cmd_line, cmd_line_init);
mca_base_cmd_line_setup(&cmd_line);
if (OPAL_SUCCESS != (rc = opal_cmd_line_parse(&cmd_line, true,
argc, argv)) ) {
if (OPAL_ERR_SILENT != rc) {
fprintf(stderr, "%s: command line error (%s)\n", argv[0],
opal_strerror(rc));
}
return rc;
}
/* print version if requested. Do this before check for help so
that --version --help works as one might expect. */
if (myglobals.version) {
char *str;
str = opal_info_make_version_str("all",
OPAL_MAJOR_VERSION, OPAL_MINOR_VERSION,
OPAL_RELEASE_VERSION,
OPAL_GREEK_VERSION,
OPAL_REPO_REV);
if (NULL != str) {
fprintf(stdout, "%s %s\n\nReport bugs to %s\n",
myglobals.basename, str, PACKAGE_BUGREPORT);
free(str);
}
exit(0);
}
/* check if we are running as root - if we are, then only allow
* us to proceed if the allow-run-as-root flag was given. Otherwise,
* exit with a giant warning flag
*/
if (0 == geteuid() && !myglobals.run_as_root) {
fprintf(stderr, "--------------------------------------------------------------------------\n");
if (myglobals.help) {
fprintf(stderr, "%s cannot provide the help message when run as root\n", myglobals.basename);
} else {
/* show_help is not yet available, so print an error manually */
fprintf(stderr, "%s has detected an attempt to run as root.\n", myglobals.basename);
}
fprintf(stderr, " This is *strongly* discouraged as any mistake (e.g., in defining TMPDIR) or bug can\n");
fprintf(stderr, "result in catastrophic damage to the OS file system, leaving\n");
fprintf(stderr, "your system in an unusable state.\n\n");
fprintf(stderr, "You can override this protection by adding the --allow-run-as-root\n");
fprintf(stderr, "option to your cmd line. However, we reiterate our strong advice\n");
fprintf(stderr, "against doing so - please do so at your own risk.\n");
fprintf(stderr, "--------------------------------------------------------------------------\n");
exit(1);
}
/*
* Since this process can now handle MCA/GMCA parameters, make sure to
* process them.
* NOTE: It is "safe" to call mca_base_cmd_line_process_args() before
* opal_init_util() since mca_base_cmd_line_process_args() does *not*
* depend upon opal_init_util() functionality.
*/
if (OPAL_SUCCESS != mca_base_cmd_line_process_args(&cmd_line, &environ, &environ)) {
exit(1);
}
/* Need to initialize OPAL so that install_dirs are filled in */
if (OPAL_SUCCESS != opal_init(&argc, &argv)) {
exit(1);
}
/* Check for help request */
if (myglobals.help) {
char *str, *args = NULL;
char *project_name = NULL;
if (0 == strcmp(myglobals.basename, "mpirun")) {
project_name = "Open MPI";
} else {
project_name = "OpenRTE";
}
args = opal_cmd_line_get_usage_msg(&cmd_line);
str = opal_show_help_string("help-orterun.txt", "orterun:usage", false,
myglobals.basename, project_name, OPAL_VERSION,
myglobals.basename, args,
PACKAGE_BUGREPORT);
if (NULL != str) {
printf("%s", str);
free(str);
}
free(args);
/* If someone asks for help, that should be all we do */
exit(0);
}
/* flag that I am the HNP */
orte_process_info.proc_type = ORTE_PROC_HNP;
/* Setup MCA params */
orte_register_params();
/* specify the DVM state machine */
opal_setenv("OMPI_MCA_state", "dvm", true, &environ);
/* Intialize our Open RTE environment */
if (ORTE_SUCCESS != (rc = orte_init(&argc, &argv, ORTE_PROC_HNP))) {
/* cannot call ORTE_ERROR_LOG as it could be the errmgr
* never got loaded!
*/
return rc;
}
/* finalize OPAL. As it was opened again from orte_init->opal_init
* we continue to have a reference count on it. So we have to finalize it twice...
*/
opal_finalize();
/* check for request to report uri */
uri = orte_rml.get_contact_info();
if (NULL != myglobals.report_uri) {
FILE *fp;
if (0 == strcmp(myglobals.report_uri, "-")) {
/* if '-', then output to stdout */
printf("VMURI: %s\n", uri);
} else if (0 == strcmp(myglobals.report_uri, "+")) {
/* if '+', output to stderr */
fprintf(stderr, "VMURI: %s\n", uri);
} else if (0 == strncasecmp(myglobals.report_uri, "file:", strlen("file:"))) {
ptr = strchr(myglobals.report_uri, ':');
++ptr;
fp = fopen(ptr, "w");
if (NULL == fp) {
orte_show_help("help-orterun.txt", "orterun:write_file", false,
myglobals.basename, "pid", ptr);
exit(0);
}
fprintf(fp, "%s\n", uri);
fclose(fp);
} else {
fp = fopen(myglobals.report_uri, "w");
if (NULL == fp) {
orte_show_help("help-orterun.txt", "orterun:write_file", false,
myglobals.basename, "pid", myglobals.report_uri);
exit(0);
}
fprintf(fp, "%s\n", uri);
fclose(fp);
}
free(uri);
} else {
printf("VMURI: %s\n", uri);
}
/* get the daemon job object - was created by ess/hnp component */
if (NULL == (jdata = orte_get_job_data_object(ORTE_PROC_MY_NAME->jobid))) {
orte_show_help("help-orterun.txt", "bad-job-object", true,
myglobals.basename);
exit(0);
}
/* also should have created a daemon "app" */
if (NULL == (app = (orte_app_context_t*)opal_pointer_array_get_item(jdata->apps, 0))) {
orte_show_help("help-orterun.txt", "bad-app-object", true,
myglobals.basename);
exit(0);
}
/* Did the user specify a prefix, or want prefix by default? */
if (opal_cmd_line_is_taken(&cmd_line, "prefix") || want_prefix_by_default) {
size_t param_len;
/* if both the prefix was given and we have a prefix
* given above, check to see if they match
*/
if (opal_cmd_line_is_taken(&cmd_line, "prefix") &&
NULL != myglobals.prefix) {
/* if they don't match, then that merits a warning */
param = strdup(opal_cmd_line_get_param(&cmd_line, "prefix", 0, 0));
/* ensure we strip any trailing '/' */
if (0 == strcmp(OPAL_PATH_SEP, &(param[strlen(param)-1]))) {
param[strlen(param)-1] = '\0';
}
value = strdup(myglobals.prefix);
if (0 == strcmp(OPAL_PATH_SEP, &(value[strlen(value)-1]))) {
value[strlen(value)-1] = '\0';
}
if (0 != strcmp(param, value)) {
orte_show_help("help-orterun.txt", "orterun:app-prefix-conflict",
true, myglobals.basename, value, param);
/* let the global-level prefix take precedence since we
* know that one is being used
*/
free(param);
param = strdup(myglobals.prefix);
}
free(value);
} else if (NULL != myglobals.prefix) {
param = myglobals.prefix;
} else if (opal_cmd_line_is_taken(&cmd_line, "prefix")){
/* must be --prefix alone */
param = strdup(opal_cmd_line_get_param(&cmd_line, "prefix", 0, 0));
} else {
/* --enable-orterun-prefix-default was given to orterun */
param = strdup(opal_install_dirs.prefix);
}
if (NULL != param) {
/* "Parse" the param, aka remove superfluous path_sep. */
param_len = strlen(param);
while (0 == strcmp (OPAL_PATH_SEP, &(param[param_len-1]))) {
param[param_len-1] = '\0';
param_len--;
if (0 == param_len) {
orte_show_help("help-orterun.txt", "orterun:empty-prefix",
true, myglobals.basename, myglobals.basename);
return ORTE_ERR_FATAL;
}
}
orte_set_attribute(&app->attributes, ORTE_APP_PREFIX_DIR, ORTE_ATTR_GLOBAL, param, OPAL_STRING);
free(param);
}
}
/* Did the user specify a hostfile. Need to check for both
* hostfile and machine file.
* We can only deal with one hostfile per app context, otherwise give an error.
*/
if (0 < (j = opal_cmd_line_get_ninsts(&cmd_line, "hostfile"))) {
if(1 < j) {
orte_show_help("help-orterun.txt", "orterun:multiple-hostfiles",
true, myglobals.basename, NULL);
return ORTE_ERR_FATAL;
} else {
value = opal_cmd_line_get_param(&cmd_line, "hostfile", 0, 0);
orte_set_attribute(&app->attributes, ORTE_APP_HOSTFILE, ORTE_ATTR_LOCAL, value, OPAL_STRING);
}
}
if (0 < (j = opal_cmd_line_get_ninsts(&cmd_line, "machinefile"))) {
if(1 < j || orte_get_attribute(&app->attributes, ORTE_APP_HOSTFILE, NULL, OPAL_STRING)) {
orte_show_help("help-orterun.txt", "orterun:multiple-hostfiles",
true, myglobals.basename, NULL);
return ORTE_ERR_FATAL;
} else {
value = opal_cmd_line_get_param(&cmd_line, "machinefile", 0, 0);
orte_set_attribute(&app->attributes, ORTE_APP_HOSTFILE, ORTE_ATTR_LOCAL, value, OPAL_STRING);
}
}
/* Did the user specify any hosts? */
if (0 < (j = opal_cmd_line_get_ninsts(&cmd_line, "host"))) {
char **targ=NULL, *tval;
for (i = 0; i < j; ++i) {
value = opal_cmd_line_get_param(&cmd_line, "host", i, 0);
opal_argv_append_nosize(&targ, value);
}
tval = opal_argv_join(targ, ',');
orte_set_attribute(&app->attributes, ORTE_APP_DASH_HOST, ORTE_ATTR_LOCAL, tval, OPAL_STRING);
opal_argv_free(targ);
free(tval);
}
OBJ_DESTRUCT(&cmd_line);
/* setup to listen for commands sent specifically to me, even though I would probably
* be the one sending them! Unfortunately, since I am a participating daemon,
* there are times I need to send a command to "all daemons", and that means *I* have
* to receive it too
*/
orte_rml.recv_buffer_nb(ORTE_NAME_WILDCARD, ORTE_RML_TAG_DAEMON,
ORTE_RML_PERSISTENT, orte_daemon_recv, NULL);
/* override the notify_completed state so we can send a message
* back to anyone who submits a job to us telling them the job
* completed */
if (ORTE_SUCCESS != (rc = orte_state.set_job_state_callback(ORTE_JOB_STATE_NOTIFY_COMPLETED, notify_requestor))) {
ORTE_ERROR_LOG(rc);
ORTE_UPDATE_EXIT_STATUS(rc);
exit(orte_exit_status);
}
/* spawn the DVM - we skip the initial steps as this
* isn't a user-level application */
ORTE_ACTIVATE_JOB_STATE(jdata, ORTE_JOB_STATE_ALLOCATE);
/* loop the event lib until an exit event is detected */
while (orte_event_base_active) {
opal_event_loop(orte_event_base, OPAL_EVLOOP_ONCE);
}
/* cleanup and leave */
orte_finalize();
if (orte_debug_flag) {
fprintf(stderr, "exiting with status %d\n", orte_exit_status);
}
exit(orte_exit_status);
}
static void track_procs(int fd, short argc, void *cbdata)
{
orte_state_caddy_t *caddy = (orte_state_caddy_t*)cbdata;
orte_process_name_t *proc = &caddy->name;
orte_proc_state_t state = caddy->proc_state;
orte_job_t *jdata;
orte_proc_t *pdata, *pptr;
opal_buffer_t *alert;
int rc, i;
orte_plm_cmd_flag_t cmd;
int8_t flag;
OPAL_OUTPUT_VERBOSE((5, orte_state_base_framework.framework_output,
"%s state:orcm:track_procs called for proc %s state %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(proc),
orte_proc_state_to_str(state)));
/* get the job object for this proc */
if (NULL == (jdata = orte_get_job_data_object(proc->jobid))) {
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
goto cleanup;
}
pdata = (orte_proc_t*)opal_pointer_array_get_item(jdata->procs, proc->vpid);
if (NULL == pdata) {
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
goto cleanup;
}
if (ORTE_PROC_STATE_RUNNING == state) {
/* update the proc state */
pdata->state = state;
jdata->num_launched++;
/* don't update until we are told that all are done */
} else if (ORTE_PROC_STATE_REGISTERED == state) {
/* update the proc state */
pdata->state = state;
jdata->num_reported++;
if (jdata->num_reported == jdata->num_local_procs) {
/* once everyone registers, send their contact info to
* the HNP so it is available to debuggers and anyone
* else that needs it
*/
OPAL_OUTPUT_VERBOSE((5, orte_state_base_framework.framework_output,
"%s state:orcm: notifying HNP all local registered",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
alert = OBJ_NEW(opal_buffer_t);
/* pack registered command */
cmd = ORTE_PLM_REGISTERED_CMD;
if (ORTE_SUCCESS != (rc = opal_dss.pack(alert, &cmd, 1, ORTE_PLM_CMD))) {
ORTE_ERROR_LOG(rc);
goto cleanup;
}
/* pack the jobid */
if (ORTE_SUCCESS != (rc = opal_dss.pack(alert, &proc->jobid, 1, ORTE_JOBID))) {
ORTE_ERROR_LOG(rc);
goto cleanup;
}
/* pack all the local child vpids */
for (i=0; i < orte_local_children->size; i++) {
if (NULL == (pptr = (orte_proc_t*)opal_pointer_array_get_item(orte_local_children, i))) {
continue;
}
if (pptr->name.jobid == proc->jobid) {
if (ORTE_SUCCESS != (rc = opal_dss.pack(alert, &pptr->name.vpid, 1, ORTE_VPID))) {
ORTE_ERROR_LOG(rc);
goto cleanup;
}
if (ORTE_FLAG_TEST(pptr, ORTE_PROC_FLAG_AS_MPI)) {
flag = 1;
} else {
flag = 0;
}
if (ORTE_SUCCESS != (rc = opal_dss.pack(alert, &flag, 1, OPAL_INT8))) {
ORTE_ERROR_LOG(rc);
goto cleanup;
}
}
}
/* send it */
if (ORTE_SUCCESS != (rc = orte_rml.send_buffer_nb(ORTE_PROC_MY_HNP, alert,
ORTE_RML_TAG_PLM,
orte_rml_send_callback, NULL))) {
ORTE_ERROR_LOG(rc);
} else {
rc = ORTE_SUCCESS;
}
}
} else if (ORTE_PROC_STATE_IOF_COMPLETE == state) {
/* do NOT update the proc state as this can hit
* while we are still trying to notify the HNP of
* successful launch for short-lived procs
*/
ORTE_FLAG_SET(pdata, ORTE_PROC_FLAG_IOF_COMPLETE);
if (ORTE_FLAG_TEST(pdata, ORTE_PROC_FLAG_WAITPID) &&
!ORTE_FLAG_TEST(pdata, ORTE_PROC_FLAG_RECORDED)) {
ORTE_ACTIVATE_PROC_STATE(proc, ORTE_PROC_STATE_TERMINATED);
}
} else if (ORTE_PROC_STATE_WAITPID_FIRED == state) {
/* do NOT update the proc state as this can hit
* while we are still trying to notify the HNP of
* successful launch for short-lived procs
*/
ORTE_FLAG_SET(pdata, ORTE_PROC_FLAG_WAITPID);
if (ORTE_FLAG_TEST(pdata, ORTE_PROC_FLAG_IOF_COMPLETE) &&
!ORTE_FLAG_TEST(pdata, ORTE_PROC_FLAG_RECORDED)) {
ORTE_ACTIVATE_PROC_STATE(proc, ORTE_PROC_STATE_TERMINATED);
}
} else if (ORTE_PROC_STATE_TERMINATED == state) {
/* if this proc has not already recorded as terminated, then
* update the accounting here */
if (!ORTE_FLAG_TEST(pdata, ORTE_PROC_FLAG_RECORDED)) {
jdata->num_terminated++;
}
/* update the proc state */
ORTE_FLAG_SET(pdata, ORTE_PROC_FLAG_RECORDED);
ORTE_FLAG_UNSET(pdata, ORTE_PROC_FLAG_ALIVE);
pdata->state = state;
/* Clean up the session directory as if we were the process
* itself. This covers the case where the process died abnormally
* and didn't cleanup its own session directory.
*/
orte_session_dir_finalize(proc);
/* track job status */
if (jdata->num_terminated == jdata->num_local_procs &&
!orte_get_attribute(&jdata->attributes, ORTE_JOB_TERM_NOTIFIED, NULL, OPAL_BOOL)) {
/* pack update state command */
cmd = ORTE_PLM_UPDATE_PROC_STATE;
alert = OBJ_NEW(opal_buffer_t);
if (ORTE_SUCCESS != (rc = opal_dss.pack(alert, &cmd, 1, ORTE_PLM_CMD))) {
ORTE_ERROR_LOG(rc);
goto cleanup;
}
/* pack the job info */
if (ORTE_SUCCESS != (rc = pack_state_update(alert, jdata))) {
ORTE_ERROR_LOG(rc);
}
/* send it */
OPAL_OUTPUT_VERBOSE((5, orte_state_base_framework.framework_output,
"%s state:orcm: SENDING JOB LOCAL TERMINATION UPDATE FOR JOB %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_JOBID_PRINT(jdata->jobid)));
if (0 > (rc = orte_rml.send_buffer_nb(ORTE_PROC_MY_HNP, alert,
ORTE_RML_TAG_PLM,
orte_rml_send_callback, NULL))) {
ORTE_ERROR_LOG(rc);
}
/* mark that we sent it so we ensure we don't do it again */
orte_set_attribute(&jdata->attributes, ORTE_JOB_TERM_NOTIFIED, ORTE_ATTR_LOCAL, NULL, OPAL_BOOL);
}
}
cleanup:
OBJ_RELEASE(caddy);
}
/*
* JOB
* NOTE: We do not pack all of the job object's fields as many of them have no
* value in sending them to another location. The only purpose in packing and
* sending a job object is to communicate the data required to dynamically
* spawn another job - so we only pack that limited set of required data.
* Therefore, only unpack what was packed
*/
int orte_dt_unpack_job(opal_buffer_t *buffer, void *dest,
int32_t *num_vals, opal_data_type_t type)
{
int rc;
int32_t i, k, n, count, bookmark;
orte_job_t **jobs;
orte_app_idx_t j;
orte_attribute_t *kv;
char *tmp;
opal_value_t *val;
opal_list_t *cache;
/* unpack into array of orte_job_t objects */
jobs = (orte_job_t**) dest;
for (i=0; i < *num_vals; i++) {
/* create the orte_job_t object */
jobs[i] = OBJ_NEW(orte_job_t);
if (NULL == jobs[i]) {
ORTE_ERROR_LOG(ORTE_ERR_OUT_OF_RESOURCE);
return ORTE_ERR_OUT_OF_RESOURCE;
}
/* unpack the jobid */
n = 1;
if (ORTE_SUCCESS != (rc = opal_dss_unpack_buffer(buffer,
&(jobs[i]->jobid), &n, ORTE_JOBID))) {
ORTE_ERROR_LOG(rc);
return rc;
}
/* unpack the flags */
n = 1;
if (ORTE_SUCCESS != (rc = opal_dss_unpack_buffer(buffer,
(&(jobs[i]->flags)), &n, ORTE_JOB_FLAGS_T))) {
ORTE_ERROR_LOG(rc);
return rc;
}
/* unpack the attributes */
n=1;
if (ORTE_SUCCESS != (rc = opal_dss_unpack_buffer(buffer, &count,
&n, ORTE_STD_CNTR))) {
ORTE_ERROR_LOG(rc);
return rc;
}
for (k=0; k < count; k++) {
n=1;
if (ORTE_SUCCESS != (rc = opal_dss_unpack_buffer(buffer, &kv,
&n, ORTE_ATTRIBUTE))) {
ORTE_ERROR_LOG(rc);
return rc;
}
kv->local = ORTE_ATTR_GLOBAL; // obviously not a local value
opal_list_append(&jobs[i]->attributes, &kv->super);
}
/* unpack any job info */
n=1;
if (ORTE_SUCCESS != (rc = opal_dss_unpack_buffer(buffer, &count,
&n, ORTE_STD_CNTR))) {
ORTE_ERROR_LOG(rc);
return rc;
}
if (0 < count){
cache = OBJ_NEW(opal_list_t);
orte_set_attribute(&jobs[i]->attributes, ORTE_JOB_INFO_CACHE, ORTE_ATTR_LOCAL, (void*)cache, OPAL_PTR);
for (k=0; k < count; k++) {
n=1;
if (ORTE_SUCCESS != (rc = opal_dss_unpack_buffer(buffer, &val,
&n, OPAL_VALUE))) {
ORTE_ERROR_LOG(rc);
return rc;
}
opal_list_append(cache, &val->super);
}
}
/* unpack the personality */
n=1;
if (ORTE_SUCCESS != (rc = opal_dss_unpack_buffer(buffer, &count, &n, OPAL_INT32))) {
ORTE_ERROR_LOG(rc);
return rc;
}
for (k=0; k < count; k++) {
n=1;
if (ORTE_SUCCESS != (rc = opal_dss_unpack_buffer(buffer, &tmp, &n, OPAL_STRING))) {
ORTE_ERROR_LOG(rc);
return rc;
}
opal_argv_append_nosize(&jobs[i]->personality, tmp);
free(tmp);
}
/* unpack the num apps */
n = 1;
if (ORTE_SUCCESS != (rc = opal_dss_unpack_buffer(buffer,
(&(jobs[i]->num_apps)), &n, ORTE_APP_IDX))) {
ORTE_ERROR_LOG(rc);
return rc;
}
/* if there are apps, unpack them */
if (0 < jobs[i]->num_apps) {
orte_app_context_t *app;
for (j=0; j < jobs[i]->num_apps; j++) {
n = 1;
if (ORTE_SUCCESS != (rc = opal_dss_unpack_buffer(buffer,
&app, &n, ORTE_APP_CONTEXT))) {
ORTE_ERROR_LOG(rc);
return rc;
}
opal_pointer_array_add(jobs[i]->apps, app);
}
}
/* unpack num procs and offset */
n = 1;
if (ORTE_SUCCESS != (rc = opal_dss_unpack_buffer(buffer,
(&(jobs[i]->num_procs)), &n, ORTE_VPID))) {
ORTE_ERROR_LOG(rc);
return rc;
}
n = 1;
if (ORTE_SUCCESS != (rc = opal_dss_unpack_buffer(buffer,
(&(jobs[i]->offset)), &n, ORTE_VPID))) {
ORTE_ERROR_LOG(rc);
return rc;
}
if (0 < jobs[i]->num_procs) {
/* check attributes to see if this job was fully
* described in the launch msg */
if (orte_get_attribute(&jobs[i]->attributes, ORTE_JOB_FULLY_DESCRIBED, NULL, OPAL_BOOL)) {
orte_proc_t *proc;
for (j=0; j < jobs[i]->num_procs; j++) {
n = 1;
if (ORTE_SUCCESS != (rc = opal_dss_unpack_buffer(buffer,
&proc, &n, ORTE_PROC))) {
ORTE_ERROR_LOG(rc);
return rc;
}
opal_pointer_array_add(jobs[i]->procs, proc);
}
}
}
/* unpack stdin target */
n = 1;
if (ORTE_SUCCESS != (rc = opal_dss_unpack_buffer(buffer,
(&(jobs[i]->stdin_target)), &n, ORTE_VPID))) {
ORTE_ERROR_LOG(rc);
return rc;
}
/* unpack the total slots allocated to the job */
n = 1;
if (ORTE_SUCCESS != (rc = opal_dss_unpack_buffer(buffer,
(&(jobs[i]->total_slots_alloc)), &n, ORTE_STD_CNTR))) {
ORTE_ERROR_LOG(rc);
return rc;
}
/* if the map is NULL, then we didn't pack it as there was
* nothing to pack. Instead, we packed a flag to indicate whether or not
* the map is included */
n = 1;
if (ORTE_SUCCESS != (rc = opal_dss_unpack_buffer(buffer,
&j, &n, ORTE_STD_CNTR))) {
ORTE_ERROR_LOG(rc);
return rc;
}
if (0 < j) {
/* unpack the map */
n = 1;
if (ORTE_SUCCESS != (rc = opal_dss_unpack_buffer(buffer,
(&(jobs[i]->map)), &n, ORTE_JOB_MAP))) {
ORTE_ERROR_LOG(rc);
return rc;
}
}
/* unpack the bookmark */
n = 1;
if (ORTE_SUCCESS != (rc = opal_dss_unpack_buffer(buffer,
&bookmark, &n, OPAL_INT32))) {
ORTE_ERROR_LOG(rc);
return rc;
}
if (0 <= bookmark) {
/* retrieve it */
jobs[i]->bookmark = (orte_node_t*)opal_pointer_array_get_item(orte_node_pool, bookmark);
}
/* unpack the job state */
n = 1;
if (ORTE_SUCCESS != (rc = opal_dss_unpack_buffer(buffer,
(&(jobs[i]->state)), &n, ORTE_JOB_STATE))) {
ORTE_ERROR_LOG(rc);
return rc;
}
}
return ORTE_SUCCESS;
}
static int bind_to_cpuset(orte_job_t *jdata)
{
/* bind each process to opal_hwloc_base_cpu_set */
int i, j;
orte_job_map_t *map;
orte_node_t *node;
orte_proc_t *proc;
struct hwloc_topology_support *support;
opal_hwloc_topo_data_t *sum;
hwloc_obj_t root;
char *cpu_bitmap;
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps: bind job %s to cpuset %s",
ORTE_JOBID_PRINT(jdata->jobid),
opal_hwloc_base_cpu_set);
/* initialize */
map = jdata->map;
for (i=0; i < map->nodes->size; i++) {
if (NULL == (node = (orte_node_t*)opal_pointer_array_get_item(map->nodes, i))) {
continue;
}
if (!orte_do_not_launch) {
/* if we don't want to launch, then we are just testing the system,
* so ignore questions about support capabilities
*/
support = (struct hwloc_topology_support*)hwloc_topology_get_support(node->topology);
/* check if topology supports cpubind - have to be careful here
* as Linux doesn't currently support thread-level binding. This
* may change in the future, though, and it isn't clear how hwloc
* interprets the current behavior. So check both flags to be sure.
*/
if (!support->cpubind->set_thisproc_cpubind &&
!support->cpubind->set_thisthread_cpubind) {
if (!OPAL_BINDING_REQUIRED(opal_hwloc_binding_policy)) {
/* we are not required to bind, so ignore this */
continue;
}
orte_show_help("help-orte-rmaps-base.txt", "rmaps:cpubind-not-supported", true, node->name);
return ORTE_ERR_SILENT;
}
/* check if topology supports membind - have to be careful here
* as hwloc treats this differently than I (at least) would have
* expected. Per hwloc, Linux memory binding is at the thread,
* and not process, level. Thus, hwloc sets the "thisproc" flag
* to "false" on all Linux systems, and uses the "thisthread" flag
* to indicate binding capability
*/
if (!support->membind->set_thisproc_membind &&
!support->membind->set_thisthread_membind) {
if (OPAL_HWLOC_BASE_MBFA_WARN == opal_hwloc_base_mbfa && !membind_warned) {
orte_show_help("help-orte-rmaps-base.txt", "rmaps:membind-not-supported", true, node->name);
membind_warned = true;
} else if (OPAL_HWLOC_BASE_MBFA_ERROR == opal_hwloc_base_mbfa) {
orte_show_help("help-orte-rmaps-base.txt", "rmaps:membind-not-supported-fatal", true, node->name);
return ORTE_ERR_SILENT;
}
}
}
root = hwloc_get_root_obj(node->topology);
if (NULL == root->userdata) {
/* something went wrong */
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
return ORTE_ERR_NOT_FOUND;
}
sum = (opal_hwloc_topo_data_t*)root->userdata;
if (NULL == sum->available) {
/* another error */
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
return ORTE_ERR_NOT_FOUND;
}
for (j=0; j < node->procs->size; j++) {
if (NULL == (proc = (orte_proc_t*)opal_pointer_array_get_item(node->procs, j))) {
continue;
}
/* ignore procs from other jobs */
if (proc->name.jobid != jdata->jobid) {
continue;
}
hwloc_bitmap_list_asprintf(&cpu_bitmap, sum->available);
orte_set_attribute(&proc->attributes, ORTE_PROC_CPU_BITMAP, ORTE_ATTR_GLOBAL, cpu_bitmap, OPAL_STRING);
if (NULL != cpu_bitmap) {
free(cpu_bitmap);
}
}
}
return ORTE_SUCCESS;
}
static int hostfile_parse_line(int token, opal_list_t* updates,
opal_list_t* exclude, bool keep_all)
{
int rc;
orte_node_t* node;
bool got_max = false;
char* value;
char **argv;
char* node_name = NULL;
char* username = NULL;
int cnt;
int number_of_slots = 0;
char buff[64];
if (ORTE_HOSTFILE_STRING == token ||
ORTE_HOSTFILE_HOSTNAME == token ||
ORTE_HOSTFILE_INT == token ||
ORTE_HOSTFILE_IPV4 == token ||
ORTE_HOSTFILE_IPV6 == token) {
if(ORTE_HOSTFILE_INT == token) {
snprintf(buff, 64, "%d", orte_util_hostfile_value.ival);
value = buff;
} else {
value = orte_util_hostfile_value.sval;
}
argv = opal_argv_split (value, '@');
cnt = opal_argv_count (argv);
if (1 == cnt) {
node_name = strdup(argv[0]);
} else if (2 == cnt) {
username = argv[0];
node_name = strdup(argv[1]);
} else {
opal_output(0, "WARNING: Unhandled user@host-combination\n"); /* XXX */
}
opal_argv_free (argv);
// Strip off the FQDN if present, ignore IP addresses
if( !orte_keep_fqdn_hostnames && !opal_net_isaddr(node_name) ) {
char *ptr;
if (NULL != (ptr = strchr(node_name, '.'))) {
*ptr = '\0';
}
}
/* if the first letter of the name is '^', then this is a node
* to be excluded. Remove the ^ character so the nodename is
* usable, and put it on the exclude list
*/
if ('^' == node_name[0]) {
int i, len;
len = strlen(node_name);
for (i=1; i < len; i++) {
node_name[i-1] = node_name[i];
}
node_name[len-1] = '\0'; /* truncate */
OPAL_OUTPUT_VERBOSE((3, orte_ras_base_framework.framework_output,
"%s hostfile: node %s is being excluded",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), node_name));
/* see if this is another name for us */
if (orte_ifislocal(node_name)) {
/* Nodename has been allocated, that is for sure */
free (node_name);
node_name = strdup(orte_process_info.nodename);
}
/* Do we need to make a new node object? First check to see
if it's already in the exclude list */
if (NULL == (node = hostfile_lookup(exclude, node_name))) {
node = OBJ_NEW(orte_node_t);
node->name = node_name;
if (NULL != username) {
orte_set_attribute(&node->attributes, ORTE_NODE_USERNAME, ORTE_ATTR_LOCAL, username, OPAL_STRING);
}
opal_list_append(exclude, &node->super);
} else {
free(node_name);
}
return ORTE_SUCCESS;
}
/* this is not a node to be excluded, so we need to process it and
* add it to the "include" list. See if this host is actually us.
*/
if (orte_ifislocal(node_name)) {
/* Nodename has been allocated, that is for sure */
free (node_name);
node_name = strdup(orte_process_info.nodename);
}
OPAL_OUTPUT_VERBOSE((3, orte_ras_base_framework.framework_output,
"%s hostfile: node %s is being included - keep all is %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), node_name,
keep_all ? "TRUE" : "FALSE"));
/* Do we need to make a new node object? */
if (keep_all || NULL == (node = hostfile_lookup(updates, node_name))) {
node = OBJ_NEW(orte_node_t);
node->name = node_name;
node->slots = 1;
if (NULL != username) {
orte_set_attribute(&node->attributes, ORTE_NODE_USERNAME, ORTE_ATTR_LOCAL, username, OPAL_STRING);
}
opal_list_append(updates, &node->super);
} else {
/* this node was already found once - add a slot and mark slots as "given" */
node->slots++;
ORTE_FLAG_SET(node, ORTE_NODE_FLAG_SLOTS_GIVEN);
free(node_name);
}
} else if (ORTE_HOSTFILE_RELATIVE == token) {
/* store this for later processing */
node = OBJ_NEW(orte_node_t);
node->name = strdup(orte_util_hostfile_value.sval);
opal_list_append(updates, &node->super);
} else if (ORTE_HOSTFILE_RANK == token) {
/* we can ignore the rank, but we need to extract the node name. we
* first need to shift over to the other side of the equal sign as
* this is where the node name will be
*/
while (!orte_util_hostfile_done &&
ORTE_HOSTFILE_EQUAL != token) {
token = orte_util_hostfile_lex();
}
if (orte_util_hostfile_done) {
/* bad syntax somewhere */
return ORTE_ERROR;
}
/* next position should be the node name */
token = orte_util_hostfile_lex();
if(ORTE_HOSTFILE_INT == token) {
snprintf(buff, 64, "%d", orte_util_hostfile_value.ival);
value = buff;
} else {
value = orte_util_hostfile_value.sval;
}
argv = opal_argv_split (value, '@');
cnt = opal_argv_count (argv);
if (1 == cnt) {
node_name = strdup(argv[0]);
} else if (2 == cnt) {
username = argv[0];
node_name = strdup(argv[1]);
} else {
opal_output(0, "WARNING: Unhandled user@host-combination\n"); /* XXX */
}
opal_argv_free (argv);
// Strip off the FQDN if present, ignore IP addresses
if( !orte_keep_fqdn_hostnames && !opal_net_isaddr(node_name) ) {
char *ptr;
if (NULL != (ptr = strchr(node_name, '.'))) {
*ptr = '\0';
}
}
/* Do we need to make a new node object? */
if (NULL == (node = hostfile_lookup(updates, node_name))) {
node = OBJ_NEW(orte_node_t);
node->name = node_name;
node->slots = 1;
if (NULL != username) {
orte_set_attribute(&node->attributes, ORTE_NODE_USERNAME, ORTE_ATTR_LOCAL, username, OPAL_STRING);
}
opal_list_append(updates, &node->super);
} else {
/* add a slot */
node->slots++;
free(node_name);
}
OPAL_OUTPUT_VERBOSE((1, orte_ras_base_framework.framework_output,
"%s hostfile: node %s slots %d nodes-given %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), node->name, node->slots,
ORTE_FLAG_TEST(node, ORTE_NODE_FLAG_SLOTS_GIVEN) ? "TRUE" : "FALSE"));
/* mark the slots as "given" since we take them as being the
* number specified via the rankfile
*/
ORTE_FLAG_SET(node, ORTE_NODE_FLAG_SLOTS_GIVEN);
/* skip to end of line */
while (!orte_util_hostfile_done &&
ORTE_HOSTFILE_NEWLINE != token) {
token = orte_util_hostfile_lex();
}
return ORTE_SUCCESS;
} else {
hostfile_parse_error(token);
return ORTE_ERROR;
}
free(username);
while (!orte_util_hostfile_done) {
token = orte_util_hostfile_lex();
switch (token) {
case ORTE_HOSTFILE_DONE:
goto done;
case ORTE_HOSTFILE_NEWLINE:
goto done;
case ORTE_HOSTFILE_USERNAME:
username = hostfile_parse_string();
if (NULL != username) {
orte_set_attribute(&node->attributes, ORTE_NODE_USERNAME, ORTE_ATTR_LOCAL, username, OPAL_STRING);
free(username);
}
break;
case ORTE_HOSTFILE_PORT:
rc = hostfile_parse_int();
if (rc < 0) {
orte_show_help("help-hostfile.txt", "port",
true,
cur_hostfile_name, rc);
return ORTE_ERROR;
}
orte_set_attribute(&node->attributes, ORTE_NODE_PORT, ORTE_ATTR_LOCAL, &rc, OPAL_INT);
break;
case ORTE_HOSTFILE_COUNT:
case ORTE_HOSTFILE_CPU:
case ORTE_HOSTFILE_SLOTS:
rc = hostfile_parse_int();
if (rc < 0) {
orte_show_help("help-hostfile.txt", "slots",
true,
cur_hostfile_name, rc);
opal_list_remove_item(updates, &node->super);
OBJ_RELEASE(node);
return ORTE_ERROR;
}
if (ORTE_FLAG_TEST(node, ORTE_NODE_FLAG_SLOTS_GIVEN)) {
/* multiple definitions were given for the
* slot count - this is not allowed
*/
orte_show_help("help-hostfile.txt", "slots-given",
true,
cur_hostfile_name, node->name);
opal_list_remove_item(updates, &node->super);
OBJ_RELEASE(node);
return ORTE_ERROR;
}
node->slots = rc;
ORTE_FLAG_SET(node, ORTE_NODE_FLAG_SLOTS_GIVEN);
/* Ensure that slots_max >= slots */
if (node->slots_max != 0 && node->slots_max < node->slots) {
node->slots_max = node->slots;
}
break;
case ORTE_HOSTFILE_SLOTS_MAX:
rc = hostfile_parse_int();
if (rc < 0) {
orte_show_help("help-hostfile.txt", "max_slots",
true,
cur_hostfile_name, ((size_t) rc));
opal_list_remove_item(updates, &node->super);
OBJ_RELEASE(node);
return ORTE_ERROR;
}
/* Only take this update if it puts us >= node_slots */
if (rc >= node->slots) {
if (node->slots_max != rc) {
node->slots_max = rc;
got_max = true;
}
} else {
orte_show_help("help-hostfile.txt", "max_slots_lt",
true,
cur_hostfile_name, node->slots, rc);
ORTE_ERROR_LOG(ORTE_ERR_BAD_PARAM);
opal_list_remove_item(updates, &node->super);
OBJ_RELEASE(node);
return ORTE_ERROR;
}
break;
case ORTE_HOSTFILE_STRING:
case ORTE_HOSTFILE_INT:
/* just ignore it */
break;
default:
hostfile_parse_error(token);
opal_list_remove_item(updates, &node->super);
OBJ_RELEASE(node);
return ORTE_ERROR;
}
if (number_of_slots > node->slots) {
ORTE_ERROR_LOG(ORTE_ERR_BAD_PARAM);
opal_list_remove_item(updates, &node->super);
OBJ_RELEASE(node);
return ORTE_ERROR;
}
}
done:
if (got_max && !ORTE_FLAG_TEST(node, ORTE_NODE_FLAG_SLOTS_GIVEN)) {
node->slots = node->slots_max;
ORTE_FLAG_SET(node, ORTE_NODE_FLAG_SLOTS_GIVEN);
}
return ORTE_SUCCESS;
}
static int allocate(orte_job_t *jdata, opal_list_t *nodes)
{
char **nodelist;
orte_node_t *node;
int i, num_nodes;
char *affinity_file;
struct stat buf;
char *ptr;
/* get the list of allocated nodes */
if ((num_nodes = lsb_getalloc(&nodelist)) < 0) {
orte_show_help("help-ras-lsf.txt", "nodelist-failed", true);
return ORTE_ERR_NOT_AVAILABLE;
}
node = NULL;
/* step through the list */
for (i = 0; i < num_nodes; i++) {
if( !orte_keep_fqdn_hostnames && !opal_net_isaddr(nodelist[i]) ) {
if (NULL != (ptr = strchr(nodelist[i], '.'))) {
*ptr = '\0';
}
}
/* is this a repeat of the current node? */
if (NULL != node && 0 == strcmp(nodelist[i], node->name)) {
/* it is a repeat - just bump the slot count */
++node->slots;
opal_output_verbose(10, orte_ras_base_framework.framework_output,
"ras/lsf: +++ Node (%s) [slots=%d]", node->name, node->slots);
continue;
}
/* not a repeat - create a node entry for it */
node = OBJ_NEW(orte_node_t);
node->name = strdup(nodelist[i]);
node->slots_inuse = 0;
node->slots_max = 0;
node->slots = 1;
node->state = ORTE_NODE_STATE_UP;
opal_list_append(nodes, &node->super);
opal_output_verbose(10, orte_ras_base_framework.framework_output,
"ras/lsf: New Node (%s) [slots=%d]", node->name, node->slots);
}
/* release the nodelist from lsf */
opal_argv_free(nodelist);
/* check for an affinity file */
if (NULL != (affinity_file = getenv("LSB_AFFINITY_HOSTFILE"))) {
/* check to see if the file is empty - if it is,
* then affinity wasn't actually set for this job */
if (0 != stat(affinity_file, &buf)) {
orte_show_help("help-ras-lsf.txt", "affinity-file-not-found", true, affinity_file);
return ORTE_ERR_SILENT;
}
if (0 == buf.st_size) {
/* no affinity, so just return */
return ORTE_SUCCESS;
}
/* the affinity file sequentially lists rank locations, with
* cpusets given as physical cpu-ids. Setup the job object
* so it knows to process this accordingly */
if (NULL == jdata->map) {
jdata->map = OBJ_NEW(orte_job_map_t);
}
ORTE_SET_MAPPING_POLICY(jdata->map->mapping, ORTE_MAPPING_SEQ);
jdata->map->req_mapper = strdup("seq"); // need sequential mapper
/* tell the sequential mapper that all cpusets are to be treated as "physical" */
orte_set_attribute(&jdata->attributes, ORTE_JOB_PHYSICAL_CPUIDS, true, NULL, OPAL_BOOL);
/* LSF provides its info as hwthreads, so set the hwthread-as-cpus flag */
opal_hwloc_use_hwthreads_as_cpus = true;
/* don't override something provided by the user, but default to bind-to hwthread */
if (!OPAL_BINDING_POLICY_IS_SET(opal_hwloc_binding_policy)) {
OPAL_SET_BINDING_POLICY(opal_hwloc_binding_policy, OPAL_BIND_TO_HWTHREAD);
}
/*
* Do not set the hostfile attribute on each app_context since that
* would confuse the sequential mapper when it tries to assign bindings
* when running an MPMD job.
* Instead just overwrite the orte_default_hostfile so it will be
* general for all of the app_contexts.
*/
if( NULL != orte_default_hostfile ) {
free(orte_default_hostfile);
orte_default_hostfile = NULL;
}
orte_default_hostfile = strdup(affinity_file);
opal_output_verbose(10, orte_ras_base_framework.framework_output,
"ras/lsf: Set default_hostfile to %s",orte_default_hostfile);
return ORTE_SUCCESS;
}
return ORTE_SUCCESS;
}
static int setup_app(orte_app_context_t *app)
{
int i;
char *newenv, *pth, *t2;
bool takeus = false;
if (NULL != orte_schizo_base.personalities) {
/* see if we are included */
for (i=0; NULL != orte_schizo_base.personalities[i]; i++) {
if (0 == strcmp(orte_schizo_base.personalities[i], "singularity")) {
takeus = true;
break;
}
}
}
if (!takeus) {
/* even if they didn't specify, check to see if
* this involves a singularity container */
if (0 != strcmp(app->argv[0],"singularity") &&
0 != strcmp(app->argv[0],"sapprun") &&
NULL == strstr(app->argv[0], ".sapp")) {
/* guess not! */
return ORTE_ERR_TAKE_NEXT_OPTION;
}
}
opal_output_verbose(1, orte_schizo_base_framework.framework_output,
"%s schizo:singularity: checking app %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), app->argv[0]);
if (0 < strlen(OPAL_SINGULARITY_PATH)) {
asprintf(&pth, "%s/singularity", OPAL_SINGULARITY_PATH);
} else {
/* since we allow for detecting singularity's presence, it
* is possible that we found it in the PATH, but not in a
* standard location. Check for that here */
pth = opal_path_findv("singularity", X_OK, app->env, NULL);
if (NULL == pth) {
/* cannot execute */
return ORTE_ERR_TAKE_NEXT_OPTION;
}
}
/* find the path and prepend it with the path to Singularity */
for (i = 0; NULL != app->env && NULL != app->env[i]; ++i) {
/* add to PATH */
if (0 == strncmp("PATH=", app->env[i], 5)) {
t2 = opal_dirname(pth);
asprintf(&newenv, "%s:%s", t2, app->env[i] + 5);
opal_setenv("PATH", newenv, true, &app->env);
free(newenv);
free(t2);
break;
}
}
free(pth);
if (0 == strcmp(app->argv[0], "singularity")) {
/* we don't want the backend to setup a cache dir */
orte_set_attribute(&app->attributes, ORTE_APP_NO_CACHEDIR, ORTE_ATTR_GLOBAL, NULL, OPAL_BOOL);
}
/* export an envar to permit shared memory operations */
opal_setenv("SINGULARITY_NO_NAMESPACE_PID", "1", true, &app->env);
return ORTE_SUCCESS;
}
static void* ack_create (opal_list_t *qos_attributes, uint32_t channel_num) {
orte_qos_ack_channel_t * ack_chan;
int32_t rc;
uint32_t *type, type_val, *attribute, attribute_val;
type_val = orte_qos_ack;
ack_chan = OBJ_NEW (orte_qos_ack_channel_t);
ack_chan->channel_num = channel_num;
type = &type_val;
attribute = &attribute_val;
/* validate and store ack specific channel attributes */
/* set channel type */
if (ORTE_SUCCESS == (rc = orte_set_attribute( &ack_chan->attributes, ORTE_QOS_TYPE, ORTE_ATTR_GLOBAL, (void*)type, OPAL_UINT8))) {
if( orte_get_attribute (qos_attributes, ORTE_QOS_WINDOW_SIZE, (void**)&attribute, OPAL_UINT32)) {
if ( QOS_ACK_MAX_WINDOW < (*attribute)) {
ORTE_ERROR_LOG(OPAL_ERR_VALUE_OUT_OF_BOUNDS);
OBJ_RELEASE(ack_chan);
}
else {
ack_chan->window = *attribute;
if (ORTE_SUCCESS != (rc = orte_set_attribute(&ack_chan->attributes, ORTE_QOS_WINDOW_SIZE,
ORTE_ATTR_GLOBAL, (void*)attribute, OPAL_UINT32))) {
ORTE_ERROR_LOG(rc);
OBJ_RELEASE(ack_chan);
} else {
if( orte_get_attribute (qos_attributes, ORTE_QOS_ACK_NACK_TIMEOUT, (void**)&attribute, OPAL_UINT32)) {
ack_chan->timeout_secs = *attribute;
if (ORTE_SUCCESS != (rc = orte_set_attribute(&ack_chan->attributes, ORTE_QOS_ACK_NACK_TIMEOUT,
ORTE_ATTR_GLOBAL, (void*)attribute, OPAL_UINT32))) {
ORTE_ERROR_LOG(rc);
OBJ_RELEASE(ack_chan);
} else {
if( orte_get_attribute (qos_attributes, ORTE_QOS_MSG_RETRY, NULL, OPAL_BOOL)) {
OPAL_OUTPUT_VERBOSE((1, orte_qos_base_framework.framework_output,
"%s ack_create created channel = %p window = %d timeout =%d retry = %d",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
(void*)ack_chan,
ack_chan->window,
ack_chan->timeout_secs,
ack_chan->retry));
ack_chan->retry = true;
if (ORTE_SUCCESS != (rc = orte_set_attribute(&ack_chan->attributes, ORTE_QOS_MSG_RETRY,
ORTE_ATTR_GLOBAL, NULL, OPAL_BOOL))) {
ORTE_ERROR_LOG(rc);
OBJ_RELEASE(ack_chan);
}
} else {
ack_chan->retry = false;
OPAL_OUTPUT_VERBOSE((1, orte_qos_base_framework.framework_output,
"%s ack_create created channel = %p window = %d timeout =%d retry = %d",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
(void*)ack_chan,
ack_chan->window,
ack_chan->timeout_secs,
ack_chan->retry));
}
}
}else {
ORTE_ERROR_LOG(rc);
OBJ_RELEASE(ack_chan);
}
}
}
}else
OBJ_RELEASE(ack_chan);
}else {
ORTE_ERROR_LOG(rc);
OBJ_RELEASE(ack_chan);
}
return ack_chan;
}
static int ppr_mapper(orte_job_t *jdata)
{
int rc = ORTE_SUCCESS, j, n;
mca_base_component_t *c=&mca_rmaps_ppr_component.base_version;
orte_node_t *node;
orte_proc_t *proc;
orte_app_context_t *app;
orte_vpid_t total_procs, nprocs_mapped;
opal_hwloc_level_t start=OPAL_HWLOC_NODE_LEVEL;
#if OPAL_HAVE_HWLOC
hwloc_obj_t obj;
hwloc_obj_type_t lowest;
unsigned cache_level=0;
unsigned int nobjs, i;
bool pruning_reqd = false;
opal_hwloc_level_t level;
#endif
opal_list_t node_list;
opal_list_item_t *item;
orte_std_cntr_t num_slots;
orte_app_idx_t idx;
char **ppr_req, **ck;
size_t len;
bool initial_map=true;
/* only handle initial launch of loadbalanced
* or NPERxxx jobs - allow restarting of failed apps
*/
if (ORTE_FLAG_TEST(jdata, ORTE_JOB_FLAG_RESTART)) {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:ppr: job %s being restarted - ppr cannot map",
ORTE_JOBID_PRINT(jdata->jobid));
return ORTE_ERR_TAKE_NEXT_OPTION;
}
if (NULL != jdata->map->req_mapper &&
0 != strcasecmp(jdata->map->req_mapper, c->mca_component_name)) {
/* a mapper has been specified, and it isn't me */
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:ppr: job %s not using ppr mapper",
ORTE_JOBID_PRINT(jdata->jobid));
return ORTE_ERR_TAKE_NEXT_OPTION;
}
if (NULL == jdata->map->ppr ||
ORTE_MAPPING_PPR != ORTE_GET_MAPPING_POLICY(jdata->map->mapping)) {
/* not for us */
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:ppr: job %s not using ppr mapper",
ORTE_JOBID_PRINT(jdata->jobid));
return ORTE_ERR_TAKE_NEXT_OPTION;
}
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:ppr: mapping job %s with ppr %s",
ORTE_JOBID_PRINT(jdata->jobid), jdata->map->ppr);
/* flag that I did the mapping */
if (NULL != jdata->map->last_mapper) {
free(jdata->map->last_mapper);
}
jdata->map->last_mapper = strdup(c->mca_component_name);
/* initialize */
memset(ppr, 0, OPAL_HWLOC_HWTHREAD_LEVEL * sizeof(opal_hwloc_level_t));
/* parse option */
n=0;
ppr_req = opal_argv_split(jdata->map->ppr, ',');
for (j=0; NULL != ppr_req[j]; j++) {
/* split on the colon */
ck = opal_argv_split(ppr_req[j], ':');
if (2 != opal_argv_count(ck)) {
/* must provide a specification */
orte_show_help("help-orte-rmaps-ppr.txt", "invalid-ppr", true, jdata->map->ppr);
opal_argv_free(ppr_req);
opal_argv_free(ck);
return ORTE_ERR_SILENT;
}
len = strlen(ck[1]);
if (0 == strncasecmp(ck[1], "node", len)) {
ppr[OPAL_HWLOC_NODE_LEVEL] = strtol(ck[0], NULL, 10);
ORTE_SET_MAPPING_POLICY(jdata->map->mapping, ORTE_MAPPING_BYNODE);
start = OPAL_HWLOC_NODE_LEVEL;
n++;
#if OPAL_HAVE_HWLOC
} else if (0 == strncasecmp(ck[1], "hwthread", len) ||
0 == strncasecmp(ck[1], "thread", len)) {
ppr[OPAL_HWLOC_HWTHREAD_LEVEL] = strtol(ck[0], NULL, 10);
start = OPAL_HWLOC_HWTHREAD_LEVEL;
ORTE_SET_MAPPING_POLICY(jdata->map->mapping, ORTE_MAPPING_BYHWTHREAD);
n++;
} else if (0 == strncasecmp(ck[1], "core", len)) {
ppr[OPAL_HWLOC_CORE_LEVEL] = strtol(ck[0], NULL, 10);
if (start < OPAL_HWLOC_CORE_LEVEL) {
start = OPAL_HWLOC_CORE_LEVEL;
ORTE_SET_MAPPING_POLICY(jdata->map->mapping, ORTE_MAPPING_BYCORE);
}
n++;
} else if (0 == strncasecmp(ck[1], "socket", len) ||
0 == strncasecmp(ck[1], "skt", len)) {
ppr[OPAL_HWLOC_SOCKET_LEVEL] = strtol(ck[0], NULL, 10);
if (start < OPAL_HWLOC_SOCKET_LEVEL) {
start = OPAL_HWLOC_SOCKET_LEVEL;
ORTE_SET_MAPPING_POLICY(jdata->map->mapping, ORTE_MAPPING_BYSOCKET);
}
n++;
} else if (0 == strncasecmp(ck[1], "l1cache", len)) {
ppr[OPAL_HWLOC_L1CACHE_LEVEL] = strtol(ck[0], NULL, 10);
if (start < OPAL_HWLOC_L1CACHE_LEVEL) {
start = OPAL_HWLOC_L1CACHE_LEVEL;
cache_level = 1;
ORTE_SET_MAPPING_POLICY(jdata->map->mapping, ORTE_MAPPING_BYL1CACHE);
}
n++;
} else if (0 == strncasecmp(ck[1], "l2cache", len)) {
ppr[OPAL_HWLOC_L2CACHE_LEVEL] = strtol(ck[0], NULL, 10);
if (start < OPAL_HWLOC_L2CACHE_LEVEL) {
start = OPAL_HWLOC_L2CACHE_LEVEL;
cache_level = 2;
ORTE_SET_MAPPING_POLICY(jdata->map->mapping, ORTE_MAPPING_BYL2CACHE);
}
n++;
} else if (0 == strncasecmp(ck[1], "l3cache", len)) {
ppr[OPAL_HWLOC_L3CACHE_LEVEL] = strtol(ck[0], NULL, 10);
if (start < OPAL_HWLOC_L3CACHE_LEVEL) {
start = OPAL_HWLOC_L3CACHE_LEVEL;
cache_level = 3;
ORTE_SET_MAPPING_POLICY(jdata->map->mapping, ORTE_MAPPING_BYL3CACHE);
}
n++;
} else if (0 == strncasecmp(ck[1], "numa", len)) {
ppr[OPAL_HWLOC_NUMA_LEVEL] = strtol(ck[0], NULL, 10);
if (start < OPAL_HWLOC_NUMA_LEVEL) {
start = OPAL_HWLOC_NUMA_LEVEL;
ORTE_SET_MAPPING_POLICY(jdata->map->mapping, ORTE_MAPPING_BYNUMA);
}
n++;
#endif
} else {
/* unknown spec */
orte_show_help("help-orte-rmaps-ppr.txt", "unrecognized-ppr-option", true, ck[1], jdata->map->ppr);
opal_argv_free(ppr_req);
opal_argv_free(ck);
return ORTE_ERR_SILENT;
}
opal_argv_free(ck);
}
opal_argv_free(ppr_req);
/* if nothing was given, that's an error */
if (0 == n) {
opal_output(0, "NOTHING GIVEN");
return ORTE_ERR_SILENT;
}
#if OPAL_HAVE_HWLOC
/* if more than one level was specified, then pruning will be reqd */
if (1 < n) {
pruning_reqd = true;
}
#endif
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:ppr: job %s assigned policy %s",
ORTE_JOBID_PRINT(jdata->jobid),
orte_rmaps_base_print_mapping(jdata->map->mapping));
#if OPAL_HAVE_HWLOC
/* convenience */
level = start;
lowest = opal_hwloc_levels[start];
#endif
for (idx=0; idx < (orte_app_idx_t)jdata->apps->size; idx++) {
if (NULL == (app = (orte_app_context_t*)opal_pointer_array_get_item(jdata->apps, idx))) {
continue;
}
/* if the number of total procs was given, set that
* limit - otherwise, set to max so we simply fill
* all the nodes with the pattern
*/
if (0 < app->num_procs) {
total_procs = app->num_procs;
} else {
total_procs = ORTE_VPID_MAX;
}
/* get the available nodes */
OBJ_CONSTRUCT(&node_list, opal_list_t);
if(ORTE_SUCCESS != (rc = orte_rmaps_base_get_target_nodes(&node_list, &num_slots, app,
jdata->map->mapping, initial_map, false))) {
ORTE_ERROR_LOG(rc);
goto error;
}
/* flag that all subsequent requests should not reset the node->mapped flag */
initial_map = false;
/* if a bookmark exists from some prior mapping, set us to start there */
jdata->bookmark = orte_rmaps_base_get_starting_point(&node_list, jdata);
/* cycle across the nodes */
nprocs_mapped = 0;
for (item = opal_list_get_first(&node_list);
item != opal_list_get_end(&node_list);
item = opal_list_get_next(item)) {
node = (orte_node_t*)item;
#if OPAL_HAVE_HWLOC
/* bozo check */
if (NULL == node->topology) {
orte_show_help("help-orte-rmaps-ppr.txt", "ppr-topo-missing",
true, node->name);
rc = ORTE_ERR_SILENT;
goto error;
}
#endif
/* add the node to the map, if needed */
if (!ORTE_FLAG_TEST(node, ORTE_NODE_FLAG_MAPPED)) {
if (ORTE_SUCCESS > (rc = opal_pointer_array_add(jdata->map->nodes, (void*)node))) {
ORTE_ERROR_LOG(rc);
goto error;
}
ORTE_FLAG_SET(node, ORTE_NODE_FLAG_MAPPED);
OBJ_RETAIN(node); /* maintain accounting on object */
jdata->map->num_nodes++;
}
/* if we are mapping solely at the node level, just put
* that many procs on this node
*/
if (OPAL_HWLOC_NODE_LEVEL == start) {
#if OPAL_HAVE_HWLOC
obj = hwloc_get_root_obj(node->topology);
#endif
for (j=0; j < ppr[start] && nprocs_mapped < total_procs; j++) {
if (NULL == (proc = orte_rmaps_base_setup_proc(jdata, node, idx))) {
rc = ORTE_ERR_OUT_OF_RESOURCE;
goto error;
}
nprocs_mapped++;
#if OPAL_HAVE_HWLOC
orte_set_attribute(&proc->attributes, ORTE_PROC_HWLOC_LOCALE, ORTE_ATTR_LOCAL, obj, OPAL_PTR);
#endif
}
#if OPAL_HAVE_HWLOC
} else {
/* get the number of lowest resources on this node */
nobjs = opal_hwloc_base_get_nbobjs_by_type(node->topology,
lowest, cache_level,
OPAL_HWLOC_AVAILABLE);
/* map the specified number of procs to each such resource on this node,
* recording the locale of each proc so we know its cpuset
*/
for (i=0; i < nobjs; i++) {
obj = opal_hwloc_base_get_obj_by_type(node->topology,
lowest, cache_level,
i, OPAL_HWLOC_AVAILABLE);
for (j=0; j < ppr[start] && nprocs_mapped < total_procs; j++) {
if (NULL == (proc = orte_rmaps_base_setup_proc(jdata, node, idx))) {
rc = ORTE_ERR_OUT_OF_RESOURCE;
goto error;
}
nprocs_mapped++;
orte_set_attribute(&proc->attributes, ORTE_PROC_HWLOC_LOCALE, ORTE_ATTR_LOCAL, obj, OPAL_PTR);
}
}
if (pruning_reqd) {
/* go up the ladder and prune the procs according to
* the specification, adjusting the count of procs on the
* node as we go
*/
level--;
prune(jdata->jobid, idx, node, &level, &nprocs_mapped);
}
#endif
}
/* set the total slots used */
if ((int)node->num_procs <= node->slots) {
node->slots_inuse = (int)node->num_procs;
} else {
node->slots_inuse = node->slots;
}
/* if no-oversubscribe was specified, check to see if
* we have violated the total slot specification - regardless,
* if slots_max was given, we are not allowed to violate it!
*/
if ((node->slots < (int)node->num_procs) ||
(0 < node->slots_max && node->slots_max < (int)node->num_procs)) {
if (ORTE_MAPPING_NO_OVERSUBSCRIBE & ORTE_GET_MAPPING_DIRECTIVE(jdata->map->mapping)) {
orte_show_help("help-orte-rmaps-base.txt", "orte-rmaps-base:alloc-error",
true, node->num_procs, app->app);
rc = ORTE_ERR_SILENT;
goto error;
}
/* flag the node as oversubscribed so that sched-yield gets
* properly set
*/
ORTE_FLAG_SET(node, ORTE_NODE_FLAG_OVERSUBSCRIBED);
}
/* if we haven't mapped all the procs, continue on to the
* next node
*/
if (total_procs == nprocs_mapped) {
break;
}
}
if (0 == app->num_procs) {
app->num_procs = nprocs_mapped;
}
if (ORTE_VPID_MAX != total_procs && nprocs_mapped < total_procs) {
/* couldn't map them all */
orte_show_help("help-orte-rmaps-ppr.txt", "ppr-too-many-procs",
true, app->app, app->num_procs, jdata->map->ppr);
rc = ORTE_ERR_SILENT;
goto error;
}
/* compute vpids and add proc objects to the job */
if (ORTE_SUCCESS != (rc = orte_rmaps_base_compute_vpids(jdata, app, &node_list))) {
ORTE_ERROR_LOG(rc);
goto error;
}
/* track the total number of processes we mapped - must update
* this AFTER we compute vpids so that computation is done
* correctly
*/
jdata->num_procs += app->num_procs;
while (NULL != (item = opal_list_remove_first(&node_list))) {
OBJ_RELEASE(item);
}
OBJ_DESTRUCT(&node_list);
}
return ORTE_SUCCESS;
error:
while (NULL != (item = opal_list_remove_first(&node_list))) {
OBJ_RELEASE(item);
}
OBJ_DESTRUCT(&node_list);
return rc;
}
static void proc_errors(int fd, short args, void *cbdata)
{
orte_state_caddy_t *caddy = (orte_state_caddy_t*)cbdata;
orte_job_t *jdata;
orte_proc_t *pptr, *proct;
orte_process_name_t *proc = &caddy->name;
orte_proc_state_t state = caddy->proc_state;
int i;
int32_t i32, *i32ptr;
OPAL_OUTPUT_VERBOSE((1, orte_errmgr_base_framework.framework_output,
"%s errmgr:default_hnp: for proc %s state %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(proc),
orte_proc_state_to_str(state)));
/*
* if orte is trying to shutdown, just let it
*/
if (orte_finalizing) {
goto cleanup;
}
/* get the job object */
if (NULL == (jdata = orte_get_job_data_object(proc->jobid))) {
/* could be a race condition */
goto cleanup;
}
pptr = (orte_proc_t*)opal_pointer_array_get_item(jdata->procs, proc->vpid);
/* we MUST handle a communication failure before doing anything else
* as it requires some special care to avoid normal termination issues
* for local application procs
*/
if (ORTE_PROC_STATE_COMM_FAILED == state) {
/* is this to a daemon? */
if (ORTE_PROC_MY_NAME->jobid != proc->jobid) {
/* nope - ignore it */
OPAL_OUTPUT_VERBOSE((5, orte_errmgr_base_framework.framework_output,
"%s Comm failure to non-daemon proc - ignoring it",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
goto cleanup;
}
/* if this is my own connection, ignore it */
if (ORTE_PROC_MY_NAME->vpid == proc->vpid) {
OPAL_OUTPUT_VERBOSE((5, orte_errmgr_base_framework.framework_output,
"%s Comm failure on my own connection - ignoring it",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
goto cleanup;
}
/* mark the daemon as gone */
ORTE_FLAG_UNSET(pptr, ORTE_PROC_FLAG_ALIVE);
/* if we have ordered orteds to terminate or abort
* is in progress, record it */
if (orte_orteds_term_ordered || orte_abnormal_term_ordered) {
OPAL_OUTPUT_VERBOSE((5, orte_errmgr_base_framework.framework_output,
"%s Comm failure: daemons terminating - recording daemon %s as gone",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), ORTE_NAME_PRINT(proc)));
/* remove from dependent routes, if it is one */
orte_routed.route_lost(proc);
/* if all my routes and local children are gone, then terminate ourselves */
if (0 == orte_routed.num_routes()) {
for (i=0; i < orte_local_children->size; i++) {
if (NULL != (proct = (orte_proc_t*)opal_pointer_array_get_item(orte_local_children, i)) &&
ORTE_FLAG_TEST(pptr, ORTE_PROC_FLAG_ALIVE) && proct->state < ORTE_PROC_STATE_UNTERMINATED) {
/* at least one is still alive */
OPAL_OUTPUT_VERBOSE((5, orte_errmgr_base_framework.framework_output,
"%s Comm failure: at least one proc (%s) still alive",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&proct->name)));
goto cleanup;
}
}
/* call our appropriate exit procedure */
OPAL_OUTPUT_VERBOSE((5, orte_errmgr_base_framework.framework_output,
"%s errmgr_hnp: all routes and children gone - ordering exit",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
ORTE_ACTIVATE_JOB_STATE(NULL, ORTE_JOB_STATE_DAEMONS_TERMINATED);
} else {
OPAL_OUTPUT_VERBOSE((5, orte_errmgr_base_framework.framework_output,
"%s Comm failure: %d routes remain alive",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
(int)orte_routed.num_routes()));
}
goto cleanup;
}
OPAL_OUTPUT_VERBOSE((5, orte_errmgr_base_framework.framework_output,
"%s Comm failure: daemon %s - aborting",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), ORTE_NAME_PRINT(proc)));
/* record the first one to fail */
if (!ORTE_FLAG_TEST(jdata, ORTE_JOB_FLAG_ABORTED)) {
/* output an error message so the user knows what happened */
orte_show_help("help-errmgr-base.txt", "node-died", true, pptr->node->name);
/* mark the daemon job as failed */
jdata->state = ORTE_JOB_STATE_COMM_FAILED;
/* point to the lowest rank to cause the problem */
orte_set_attribute(&jdata->attributes, ORTE_JOB_ABORTED_PROC, ORTE_ATTR_LOCAL, pptr, OPAL_PTR);
/* retain the object so it doesn't get free'd */
OBJ_RETAIN(pptr);
ORTE_FLAG_SET(jdata, ORTE_JOB_FLAG_ABORTED);
/* update our exit code */
ORTE_UPDATE_EXIT_STATUS(pptr->exit_code);
/* just in case the exit code hadn't been set, do it here - this
* won't override any reported exit code */
ORTE_UPDATE_EXIT_STATUS(ORTE_ERR_COMM_FAILURE);
}
/* abort the system */
default_hnp_abort(jdata);
goto cleanup;
}
/* update the proc state - can get multiple reports on a proc
* depending on circumstances, so ensure we only do this once
*/
if (pptr->state < ORTE_PROC_STATE_TERMINATED) {
pptr->state = state;
}
/* if we were ordered to terminate, mark this proc as dead and see if
* any of our routes or local children remain alive - if not, then
* terminate ourselves. */
if (orte_orteds_term_ordered) {
for (i=0; i < orte_local_children->size; i++) {
if (NULL != (proct = (orte_proc_t*)opal_pointer_array_get_item(orte_local_children, i))) {
if (ORTE_FLAG_TEST(proct, ORTE_PROC_FLAG_ALIVE)) {
goto keep_going;
}
}
}
/* if all my routes and children are gone, then terminate
ourselves nicely (i.e., this is a normal termination) */
if (0 == orte_routed.num_routes()) {
OPAL_OUTPUT_VERBOSE((2, orte_errmgr_base_framework.framework_output,
"%s errmgr:default:hnp all routes gone - exiting",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
ORTE_ACTIVATE_JOB_STATE(NULL, ORTE_JOB_STATE_DAEMONS_TERMINATED);
}
}
keep_going:
/* ensure we record the failed proc properly so we can report
* the error once we terminate
*/
switch (state) {
case ORTE_PROC_STATE_KILLED_BY_CMD:
OPAL_OUTPUT_VERBOSE((5, orte_errmgr_base_framework.framework_output,
"%s errmgr:hnp: proc %s killed by cmd",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(proc)));
/* we ordered this proc to die, so it isn't an abnormal termination
* and we don't flag it as such
*/
if (jdata->num_terminated >= jdata->num_procs) {
/* this job has terminated */
ORTE_ACTIVATE_JOB_STATE(jdata, ORTE_JOB_STATE_TERMINATED);
}
/* don't abort the job as this isn't an abnormal termination */
break;
case ORTE_PROC_STATE_ABORTED:
OPAL_OUTPUT_VERBOSE((5, orte_errmgr_base_framework.framework_output,
"%s errmgr:hnp: proc %s aborted",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(proc)));
if (!ORTE_FLAG_TEST(jdata, ORTE_JOB_FLAG_ABORTED)) {
jdata->state = ORTE_JOB_STATE_ABORTED;
/* point to the first rank to cause the problem */
orte_set_attribute(&jdata->attributes, ORTE_JOB_ABORTED_PROC, ORTE_ATTR_LOCAL, pptr, OPAL_PTR);
/* retain the object so it doesn't get free'd */
OBJ_RETAIN(pptr);
ORTE_FLAG_SET(jdata, ORTE_JOB_FLAG_ABORTED);
ORTE_UPDATE_EXIT_STATUS(pptr->exit_code);
/* abnormal termination - abort, but only do it once
* to avoid creating a lot of confusion */
default_hnp_abort(jdata);
}
break;
case ORTE_PROC_STATE_ABORTED_BY_SIG:
OPAL_OUTPUT_VERBOSE((5, orte_errmgr_base_framework.framework_output,
"%s errmgr:hnp: proc %s aborted by signal",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(proc)));
ORTE_UPDATE_EXIT_STATUS(pptr->exit_code);
/* track the number of non-zero exits */
i32 = 0;
i32ptr = &i32;
orte_get_attribute(&jdata->attributes, ORTE_JOB_NUM_NONZERO_EXIT, (void**)&i32ptr, OPAL_INT32);
++i32;
orte_set_attribute(&jdata->attributes, ORTE_JOB_NUM_NONZERO_EXIT, ORTE_ATTR_LOCAL, i32ptr, OPAL_INT32);
if (orte_abort_non_zero_exit) {
if (!ORTE_FLAG_TEST(jdata, ORTE_JOB_FLAG_ABORTED)) {
jdata->state = ORTE_JOB_STATE_ABORTED_BY_SIG;
/* point to the first rank to cause the problem */
orte_set_attribute(&jdata->attributes, ORTE_JOB_ABORTED_PROC, ORTE_ATTR_LOCAL, pptr, OPAL_PTR);
/* retain the object so it doesn't get free'd */
OBJ_RETAIN(pptr);
ORTE_FLAG_SET(jdata, ORTE_JOB_FLAG_ABORTED);
ORTE_UPDATE_EXIT_STATUS(pptr->exit_code);
/* abnormal termination - abort, but only do it once
* to avoid creating a lot of confusion */
default_hnp_abort(jdata);
}
} else {
/* user requested we consider this normal termination */
if (jdata->num_terminated >= jdata->num_procs) {
/* this job has terminated */
ORTE_ACTIVATE_JOB_STATE(jdata, ORTE_JOB_STATE_TERMINATED);
}
}
break;
case ORTE_PROC_STATE_TERM_WO_SYNC:
OPAL_OUTPUT_VERBOSE((5, orte_errmgr_base_framework.framework_output,
"%s errmgr:hnp: proc %s terminated without sync",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(proc)));
if (!ORTE_FLAG_TEST(jdata, ORTE_JOB_FLAG_ABORTED)) {
jdata->state = ORTE_JOB_STATE_ABORTED_WO_SYNC;
/* point to the first rank to cause the problem */
orte_set_attribute(&jdata->attributes, ORTE_JOB_ABORTED_PROC, ORTE_ATTR_LOCAL, pptr, OPAL_PTR);
/* retain the object so it doesn't get free'd */
OBJ_RETAIN(pptr);
ORTE_FLAG_SET(jdata, ORTE_JOB_FLAG_ABORTED);
ORTE_UPDATE_EXIT_STATUS(pptr->exit_code);
/* now treat a special case - if the proc exit'd without a required
* sync, it may have done so with a zero exit code. We want to ensure
* that the user realizes there was an error, so in this -one- case,
* we overwrite the process' exit code with the default error code
*/
ORTE_UPDATE_EXIT_STATUS(ORTE_ERROR_DEFAULT_EXIT_CODE);
/* abnormal termination - abort, but only do it once
* to avoid creating a lot of confusion */
default_hnp_abort(jdata);
}
break;
case ORTE_PROC_STATE_FAILED_TO_START:
case ORTE_PROC_STATE_FAILED_TO_LAUNCH:
OPAL_OUTPUT_VERBOSE((5, orte_errmgr_base_framework.framework_output,
"%s errmgr:hnp: proc %s %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(proc),
orte_proc_state_to_str(state)));
if (!ORTE_FLAG_TEST(jdata, ORTE_JOB_FLAG_ABORTED)) {
if (ORTE_PROC_STATE_FAILED_TO_START) {
jdata->state = ORTE_JOB_STATE_FAILED_TO_START;
} else {
jdata->state = ORTE_JOB_STATE_FAILED_TO_LAUNCH;
}
/* point to the first rank to cause the problem */
orte_set_attribute(&jdata->attributes, ORTE_JOB_ABORTED_PROC, ORTE_ATTR_LOCAL, pptr, OPAL_PTR);
/* retain the object so it doesn't get free'd */
OBJ_RETAIN(pptr);
ORTE_FLAG_SET(jdata, ORTE_JOB_FLAG_ABORTED);
ORTE_UPDATE_EXIT_STATUS(pptr->exit_code);
/* abnormal termination - abort, but only do it once
* to avoid creating a lot of confusion */
default_hnp_abort(jdata);
}
/* if this was a daemon, report it */
if (jdata->jobid == ORTE_PROC_MY_NAME->jobid) {
/* output a message indicating we failed to launch a daemon */
orte_show_help("help-errmgr-base.txt", "failed-daemon-launch", true);
}
break;
case ORTE_PROC_STATE_CALLED_ABORT:
OPAL_OUTPUT_VERBOSE((5, orte_errmgr_base_framework.framework_output,
"%s errmgr:hnp: proc %s called abort with exit code %d",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(proc), pptr->exit_code));
if (!ORTE_FLAG_TEST(jdata, ORTE_JOB_FLAG_ABORTED)) {
jdata->state = ORTE_JOB_STATE_CALLED_ABORT;
/* point to the first proc to cause the problem */
orte_set_attribute(&jdata->attributes, ORTE_JOB_ABORTED_PROC, ORTE_ATTR_LOCAL, pptr, OPAL_PTR);
/* retain the object so it doesn't get free'd */
OBJ_RETAIN(pptr);
ORTE_FLAG_SET(jdata, ORTE_JOB_FLAG_ABORTED);
ORTE_UPDATE_EXIT_STATUS(pptr->exit_code);
/* abnormal termination - abort, but only do it once
* to avoid creating a lot of confusion */
default_hnp_abort(jdata);
}
break;
case ORTE_PROC_STATE_TERM_NON_ZERO:
OPAL_OUTPUT_VERBOSE((5, orte_errmgr_base_framework.framework_output,
"%s errmgr:hnp: proc %s exited with non-zero status %d",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(proc),
pptr->exit_code));
ORTE_UPDATE_EXIT_STATUS(pptr->exit_code);
/* track the number of non-zero exits */
i32 = 0;
i32ptr = &i32;
orte_get_attribute(&jdata->attributes, ORTE_JOB_NUM_NONZERO_EXIT, (void**)&i32ptr, OPAL_INT32);
++i32;
orte_set_attribute(&jdata->attributes, ORTE_JOB_NUM_NONZERO_EXIT, ORTE_ATTR_LOCAL, i32ptr, OPAL_INT32);
if (orte_abort_non_zero_exit) {
if (!ORTE_FLAG_TEST(jdata, ORTE_JOB_FLAG_ABORTED)) {
jdata->state = ORTE_JOB_STATE_NON_ZERO_TERM;
/* point to the first rank to cause the problem */
orte_set_attribute(&jdata->attributes, ORTE_JOB_ABORTED_PROC, ORTE_ATTR_LOCAL, pptr, OPAL_PTR);
/* retain the object so it doesn't get free'd */
OBJ_RETAIN(pptr);
ORTE_FLAG_SET(jdata, ORTE_JOB_FLAG_ABORTED);
/* abnormal termination - abort, but only do it once
* to avoid creating a lot of confusion */
default_hnp_abort(jdata);
}
} else {
/* user requested we consider this normal termination */
if (jdata->num_terminated >= jdata->num_procs) {
/* this job has terminated */
ORTE_ACTIVATE_JOB_STATE(jdata, ORTE_JOB_STATE_TERMINATED);
}
}
break;
case ORTE_PROC_STATE_HEARTBEAT_FAILED:
OPAL_OUTPUT_VERBOSE((5, orte_errmgr_base_framework.framework_output,
"%s errmgr:hnp: proc %s heartbeat failed",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(proc)));
if (!ORTE_FLAG_TEST(jdata, ORTE_JOB_FLAG_ABORTED)) {
jdata->state = ORTE_JOB_STATE_HEARTBEAT_FAILED;
/* point to the first rank to cause the problem */
orte_set_attribute(&jdata->attributes, ORTE_JOB_ABORTED_PROC, ORTE_ATTR_LOCAL, pptr, OPAL_PTR);
/* retain the object so it doesn't get free'd */
OBJ_RETAIN(pptr);
ORTE_FLAG_SET(jdata, ORTE_JOB_FLAG_ABORTED);
ORTE_UPDATE_EXIT_STATUS(pptr->exit_code);
/* abnormal termination - abort, but only do it once
* to avoid creating a lot of confusion */
default_hnp_abort(jdata);
}
/* remove from dependent routes, if it is one */
orte_routed.route_lost(proc);
break;
case ORTE_PROC_STATE_UNABLE_TO_SEND_MSG:
OPAL_OUTPUT_VERBOSE((5, orte_errmgr_base_framework.framework_output,
"%s errmgr:hnp: unable to send message to proc %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(proc)));
/* if this proc is one of my daemons, then we are truly
* hosed - so just exit out
*/
if (ORTE_PROC_MY_NAME->jobid == proc->jobid) {
ORTE_ACTIVATE_JOB_STATE(NULL, ORTE_JOB_STATE_DAEMONS_TERMINATED);
break;
}
if (!ORTE_FLAG_TEST(jdata, ORTE_JOB_FLAG_ABORTED)) {
/* abnormal termination - abort, but only do it once
* to avoid creating a lot of confusion */
default_hnp_abort(jdata);
}
break;
default:
/* shouldn't get this, but terminate job if required */
OPAL_OUTPUT_VERBOSE((5, orte_errmgr_base_framework.framework_output,
"%s errmgr:hnp: proc %s default error %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(proc),
orte_proc_state_to_str(state)));
if (jdata->num_terminated == jdata->num_procs) {
ORTE_ACTIVATE_JOB_STATE(jdata, ORTE_JOB_STATE_TERMINATED);
}
break;
}
/* if the waitpid fired, be sure to let the state machine know */
if (ORTE_FLAG_TEST(pptr, ORTE_PROC_FLAG_WAITPID)) {
ORTE_ACTIVATE_PROC_STATE(&pptr->name, ORTE_PROC_STATE_WAITPID_FIRED);
}
cleanup:
OBJ_RELEASE(caddy);
}
static void track_procs(int fd, short argc, void *cbdata)
{
orte_state_caddy_t *caddy = (orte_state_caddy_t*)cbdata;
orte_process_name_t *proc = &caddy->name;
orte_proc_state_t state = caddy->proc_state;
orte_job_t *jdata;
orte_proc_t *pdata, *pptr;
opal_buffer_t *alert;
int rc, i;
orte_plm_cmd_flag_t cmd;
char *rtmod;
orte_std_cntr_t index;
orte_job_map_t *map;
orte_node_t *node;
OPAL_OUTPUT_VERBOSE((5, orte_state_base_framework.framework_output,
"%s state:orted:track_procs called for proc %s state %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(proc),
orte_proc_state_to_str(state)));
/* get the job object for this proc */
if (NULL == (jdata = orte_get_job_data_object(proc->jobid))) {
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
goto cleanup;
}
pdata = (orte_proc_t*)opal_pointer_array_get_item(jdata->procs, proc->vpid);
if (ORTE_PROC_STATE_RUNNING == state) {
/* update the proc state */
pdata->state = state;
jdata->num_launched++;
/* don't update until we are told that all are done */
} else if (ORTE_PROC_STATE_REGISTERED == state) {
/* update the proc state */
pdata->state = state;
jdata->num_reported++;
if (jdata->num_reported == jdata->num_local_procs) {
/* once everyone registers, notify the HNP */
OPAL_OUTPUT_VERBOSE((5, orte_state_base_framework.framework_output,
"%s state:orted: notifying HNP all local registered",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
alert = OBJ_NEW(opal_buffer_t);
/* pack registered command */
cmd = ORTE_PLM_REGISTERED_CMD;
if (ORTE_SUCCESS != (rc = opal_dss.pack(alert, &cmd, 1, ORTE_PLM_CMD))) {
ORTE_ERROR_LOG(rc);
goto cleanup;
}
/* pack the jobid */
if (ORTE_SUCCESS != (rc = opal_dss.pack(alert, &proc->jobid, 1, ORTE_JOBID))) {
ORTE_ERROR_LOG(rc);
goto cleanup;
}
/* pack all the local child vpids */
for (i=0; i < orte_local_children->size; i++) {
if (NULL == (pptr = (orte_proc_t*)opal_pointer_array_get_item(orte_local_children, i))) {
continue;
}
if (pptr->name.jobid == proc->jobid) {
if (ORTE_SUCCESS != (rc = opal_dss.pack(alert, &pptr->name.vpid, 1, ORTE_VPID))) {
ORTE_ERROR_LOG(rc);
goto cleanup;
}
}
}
/* send it */
if (0 > (rc = orte_rml.send_buffer_nb(orte_mgmt_conduit,
ORTE_PROC_MY_HNP, alert,
ORTE_RML_TAG_PLM,
orte_rml_send_callback, NULL))) {
ORTE_ERROR_LOG(rc);
} else {
rc = ORTE_SUCCESS;
}
}
} else if (ORTE_PROC_STATE_IOF_COMPLETE == state) {
/* do NOT update the proc state as this can hit
* while we are still trying to notify the HNP of
* successful launch for short-lived procs
*/
ORTE_FLAG_SET(pdata, ORTE_PROC_FLAG_IOF_COMPLETE);
/* Release the stdin IOF file descriptor for this child, if one
* was defined. File descriptors for the other IOF channels - stdout,
* stderr, and stddiag - were released when their associated pipes
* were cleared and closed due to termination of the process
* Do this after we handle termination in case the IOF needs
* to check to see if all procs from the job are actually terminated
*/
if (NULL != orte_iof.close) {
orte_iof.close(proc, ORTE_IOF_STDALL);
}
if (ORTE_FLAG_TEST(pdata, ORTE_PROC_FLAG_WAITPID) &&
!ORTE_FLAG_TEST(pdata, ORTE_PROC_FLAG_RECORDED)) {
ORTE_ACTIVATE_PROC_STATE(proc, ORTE_PROC_STATE_TERMINATED);
}
} else if (ORTE_PROC_STATE_WAITPID_FIRED == state) {
/* do NOT update the proc state as this can hit
* while we are still trying to notify the HNP of
* successful launch for short-lived procs
*/
ORTE_FLAG_SET(pdata, ORTE_PROC_FLAG_WAITPID);
if (ORTE_FLAG_TEST(pdata, ORTE_PROC_FLAG_IOF_COMPLETE) &&
!ORTE_FLAG_TEST(pdata, ORTE_PROC_FLAG_RECORDED)) {
ORTE_ACTIVATE_PROC_STATE(proc, ORTE_PROC_STATE_TERMINATED);
}
} else if (ORTE_PROC_STATE_TERMINATED == state) {
/* if this proc has not already recorded as terminated, then
* update the accounting here */
if (!ORTE_FLAG_TEST(pdata, ORTE_PROC_FLAG_RECORDED)) {
jdata->num_terminated++;
}
/* update the proc state */
ORTE_FLAG_SET(pdata, ORTE_PROC_FLAG_RECORDED);
ORTE_FLAG_UNSET(pdata, ORTE_PROC_FLAG_ALIVE);
pdata->state = state;
/* Clean up the session directory as if we were the process
* itself. This covers the case where the process died abnormally
* and didn't cleanup its own session directory.
*/
orte_session_dir_finalize(proc);
/* if we are trying to terminate and our routes are
* gone, then terminate ourselves IF no local procs
* remain (might be some from another job)
*/
rtmod = orte_rml.get_routed(orte_mgmt_conduit);
if (orte_orteds_term_ordered &&
0 == orte_routed.num_routes(rtmod)) {
for (i=0; i < orte_local_children->size; i++) {
if (NULL != (pdata = (orte_proc_t*)opal_pointer_array_get_item(orte_local_children, i)) &&
ORTE_FLAG_TEST(pdata, ORTE_PROC_FLAG_ALIVE)) {
/* at least one is still alive */
OPAL_OUTPUT_VERBOSE((5, orte_state_base_framework.framework_output,
"%s state:orted all routes gone but proc %s still alive",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&pdata->name)));
goto cleanup;
}
}
/* call our appropriate exit procedure */
OPAL_OUTPUT_VERBOSE((5, orte_state_base_framework.framework_output,
"%s state:orted all routes and children gone - exiting",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
ORTE_ACTIVATE_JOB_STATE(NULL, ORTE_JOB_STATE_DAEMONS_TERMINATED);
goto cleanup;
}
/* track job status */
if (jdata->num_terminated == jdata->num_local_procs &&
!orte_get_attribute(&jdata->attributes, ORTE_JOB_TERM_NOTIFIED, NULL, OPAL_BOOL)) {
/* pack update state command */
cmd = ORTE_PLM_UPDATE_PROC_STATE;
alert = OBJ_NEW(opal_buffer_t);
if (ORTE_SUCCESS != (rc = opal_dss.pack(alert, &cmd, 1, ORTE_PLM_CMD))) {
ORTE_ERROR_LOG(rc);
goto cleanup;
}
/* pack the job info */
if (ORTE_SUCCESS != (rc = pack_state_update(alert, jdata))) {
ORTE_ERROR_LOG(rc);
}
/* send it */
OPAL_OUTPUT_VERBOSE((5, orte_state_base_framework.framework_output,
"%s state:orted: SENDING JOB LOCAL TERMINATION UPDATE FOR JOB %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_JOBID_PRINT(jdata->jobid)));
if (0 > (rc = orte_rml.send_buffer_nb(orte_mgmt_conduit,
ORTE_PROC_MY_HNP, alert,
ORTE_RML_TAG_PLM,
orte_rml_send_callback, NULL))) {
ORTE_ERROR_LOG(rc);
}
/* mark that we sent it so we ensure we don't do it again */
orte_set_attribute(&jdata->attributes, ORTE_JOB_TERM_NOTIFIED, ORTE_ATTR_LOCAL, NULL, OPAL_BOOL);
/* cleanup the procs as these are gone */
for (i=0; i < orte_local_children->size; i++) {
if (NULL == (pptr = (orte_proc_t*)opal_pointer_array_get_item(orte_local_children, i))) {
continue;
}
/* if this child is part of the job... */
if (pptr->name.jobid == jdata->jobid) {
/* clear the entry in the local children */
opal_pointer_array_set_item(orte_local_children, i, NULL);
OBJ_RELEASE(pptr); // maintain accounting
}
}
/* tell the IOF that the job is complete */
if (NULL != orte_iof.complete) {
orte_iof.complete(jdata);
}
/* tell the PMIx subsystem the job is complete */
if (NULL != opal_pmix.server_deregister_nspace) {
opal_pmix.server_deregister_nspace(jdata->jobid, NULL, NULL);
}
/* release the resources */
if (NULL != jdata->map) {
map = jdata->map;
for (index = 0; index < map->nodes->size; index++) {
if (NULL == (node = (orte_node_t*)opal_pointer_array_get_item(map->nodes, index))) {
continue;
}
OPAL_OUTPUT_VERBOSE((2, orte_state_base_framework.framework_output,
"%s state:orted releasing procs from node %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
node->name));
for (i = 0; i < node->procs->size; i++) {
if (NULL == (pptr = (orte_proc_t*)opal_pointer_array_get_item(node->procs, i))) {
continue;
}
if (pptr->name.jobid != jdata->jobid) {
/* skip procs from another job */
continue;
}
node->slots_inuse--;
node->num_procs--;
OPAL_OUTPUT_VERBOSE((2, orte_state_base_framework.framework_output,
"%s state:orted releasing proc %s from node %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&pptr->name), node->name));
/* set the entry in the node array to NULL */
opal_pointer_array_set_item(node->procs, i, NULL);
/* release the proc once for the map entry */
OBJ_RELEASE(pptr);
}
/* set the node location to NULL */
opal_pointer_array_set_item(map->nodes, index, NULL);
/* maintain accounting */
OBJ_RELEASE(node);
/* flag that the node is no longer in a map */
ORTE_FLAG_UNSET(node, ORTE_NODE_FLAG_MAPPED);
}
OBJ_RELEASE(map);
jdata->map = NULL;
}
/* cleanup the job info */
opal_hash_table_set_value_uint32(orte_job_data, jdata->jobid, NULL);
OBJ_RELEASE(jdata);
}
}
cleanup:
OBJ_RELEASE(caddy);
}
int orte_ess_base_app_setup(bool db_restrict_local)
{
int ret;
char *error = NULL;
opal_list_t transports;
OPAL_TIMING_ENV_INIT(ess_base_setup);
/*
* stdout/stderr buffering
* If the user requested to override the default setting then do
* as they wish.
*/
if( orte_ess_base_std_buffering > -1 ) {
if( 0 == orte_ess_base_std_buffering ) {
setvbuf(stdout, NULL, _IONBF, 0);
setvbuf(stderr, NULL, _IONBF, 0);
}
else if( 1 == orte_ess_base_std_buffering ) {
setvbuf(stdout, NULL, _IOLBF, 0);
setvbuf(stderr, NULL, _IOLBF, 0);
}
else if( 2 == orte_ess_base_std_buffering ) {
setvbuf(stdout, NULL, _IOFBF, 0);
setvbuf(stderr, NULL, _IOFBF, 0);
}
}
/* if I am an MPI app, we will let the MPI layer define and
* control the opal_proc_t structure. Otherwise, we need to
* do so here */
if (ORTE_PROC_NON_MPI) {
orte_process_info.super.proc_name = *(opal_process_name_t*)ORTE_PROC_MY_NAME;
orte_process_info.super.proc_hostname = orte_process_info.nodename;
orte_process_info.super.proc_flags = OPAL_PROC_ALL_LOCAL;
orte_process_info.super.proc_arch = opal_local_arch;
opal_proc_local_set(&orte_process_info.super);
}
/* open and setup the state machine */
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_state_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_state_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_state_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_state_base_select";
goto error;
}
OPAL_TIMING_ENV_NEXT(ess_base_setup, "state_framework_open");
/* open the errmgr */
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_errmgr_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_errmgr_base_open";
goto error;
}
OPAL_TIMING_ENV_NEXT(ess_base_setup, "errmgr_framework_open");
/* setup my session directory */
if (orte_create_session_dirs) {
OPAL_OUTPUT_VERBOSE((2, orte_ess_base_framework.framework_output,
"%s setting up session dir with\n\ttmpdir: %s\n\thost %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
(NULL == orte_process_info.tmpdir_base) ? "UNDEF" : orte_process_info.tmpdir_base,
orte_process_info.nodename));
if (ORTE_SUCCESS != (ret = orte_session_dir(true, ORTE_PROC_MY_NAME))) {
ORTE_ERROR_LOG(ret);
error = "orte_session_dir";
goto error;
}
/* Once the session directory location has been established, set
the opal_output env file location to be in the
proc-specific session directory. */
opal_output_set_output_file_info(orte_process_info.proc_session_dir,
"output-", NULL, NULL);
/* register the directory for cleanup */
if (NULL != opal_pmix.register_cleanup) {
if (orte_standalone_operation) {
if (OPAL_SUCCESS != (ret = opal_pmix.register_cleanup(orte_process_info.top_session_dir, true, false, true))) {
ORTE_ERROR_LOG(ret);
error = "register cleanup";
goto error;
}
} else {
if (OPAL_SUCCESS != (ret = opal_pmix.register_cleanup(orte_process_info.job_session_dir, true, false, false))) {
ORTE_ERROR_LOG(ret);
error = "register cleanup";
goto error;
}
}
}
}
OPAL_TIMING_ENV_NEXT(ess_base_setup, "create_session_dirs");
/* Setup the communication infrastructure */
/* Routed system */
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_routed_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_routed_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_routed_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_routed_base_select";
goto error;
}
OPAL_TIMING_ENV_NEXT(ess_base_setup, "routed_framework_open");
/*
* OOB Layer
*/
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_oob_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_oob_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_oob_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_oob_base_select";
goto error;
}
OPAL_TIMING_ENV_NEXT(ess_base_setup, "oob_framework_open");
/* Runtime Messaging Layer */
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_rml_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_rml_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_rml_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_rml_base_select";
goto error;
}
OPAL_TIMING_ENV_NEXT(ess_base_setup, "rml_framework_open");
/* if we have info on the HNP and local daemon, process it */
if (NULL != orte_process_info.my_hnp_uri) {
/* we have to set the HNP's name, even though we won't route messages directly
* to it. This is required to ensure that we -do- send messages to the correct
* HNP name
*/
if (ORTE_SUCCESS != (ret = orte_rml_base_parse_uris(orte_process_info.my_hnp_uri,
ORTE_PROC_MY_HNP, NULL))) {
ORTE_ERROR_LOG(ret);
error = "orte_rml_parse_HNP";
goto error;
}
}
if (NULL != orte_process_info.my_daemon_uri) {
opal_value_t val;
/* extract the daemon's name so we can update the routing table */
if (ORTE_SUCCESS != (ret = orte_rml_base_parse_uris(orte_process_info.my_daemon_uri,
ORTE_PROC_MY_DAEMON, NULL))) {
ORTE_ERROR_LOG(ret);
error = "orte_rml_parse_daemon";
goto error;
}
/* Set the contact info in the database - this won't actually establish
* the connection, but just tells us how to reach the daemon
* if/when we attempt to send to it
*/
OBJ_CONSTRUCT(&val, opal_value_t);
val.key = OPAL_PMIX_PROC_URI;
val.type = OPAL_STRING;
val.data.string = orte_process_info.my_daemon_uri;
if (OPAL_SUCCESS != (ret = opal_pmix.store_local(ORTE_PROC_MY_DAEMON, &val))) {
ORTE_ERROR_LOG(ret);
val.key = NULL;
val.data.string = NULL;
OBJ_DESTRUCT(&val);
error = "store DAEMON URI";
goto error;
}
val.key = NULL;
val.data.string = NULL;
OBJ_DESTRUCT(&val);
}
/* setup the errmgr */
if (ORTE_SUCCESS != (ret = orte_errmgr_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_errmgr_base_select";
goto error;
}
OPAL_TIMING_ENV_NEXT(ess_base_setup, "errmgr_select");
/* get a conduit for our use - we never route IO over fabric */
OBJ_CONSTRUCT(&transports, opal_list_t);
orte_set_attribute(&transports, ORTE_RML_TRANSPORT_TYPE,
ORTE_ATTR_LOCAL, orte_mgmt_transport, OPAL_STRING);
if (ORTE_RML_CONDUIT_INVALID == (orte_mgmt_conduit = orte_rml.open_conduit(&transports))) {
ret = ORTE_ERR_OPEN_CONDUIT_FAIL;
error = "orte_rml_open_mgmt_conduit";
goto error;
}
OPAL_LIST_DESTRUCT(&transports);
OBJ_CONSTRUCT(&transports, opal_list_t);
orte_set_attribute(&transports, ORTE_RML_TRANSPORT_TYPE,
ORTE_ATTR_LOCAL, orte_coll_transport, OPAL_STRING);
if (ORTE_RML_CONDUIT_INVALID == (orte_coll_conduit = orte_rml.open_conduit(&transports))) {
ret = ORTE_ERR_OPEN_CONDUIT_FAIL;
error = "orte_rml_open_coll_conduit";
goto error;
}
OPAL_LIST_DESTRUCT(&transports);
OPAL_TIMING_ENV_NEXT(ess_base_setup, "rml_open_conduit");
/*
* Group communications
*/
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_grpcomm_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_grpcomm_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_grpcomm_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_grpcomm_base_select";
goto error;
}
OPAL_TIMING_ENV_NEXT(ess_base_setup, "grpcomm_framework_open");
/* open the distributed file system */
if (ORTE_SUCCESS != (ret = mca_base_framework_open(&orte_dfs_base_framework, 0))) {
ORTE_ERROR_LOG(ret);
error = "orte_dfs_base_open";
goto error;
}
if (ORTE_SUCCESS != (ret = orte_dfs_base_select())) {
ORTE_ERROR_LOG(ret);
error = "orte_dfs_base_select";
goto error;
}
OPAL_TIMING_ENV_NEXT(ess_base_setup, "dfs_framework_open");
return ORTE_SUCCESS;
error:
orte_show_help("help-orte-runtime.txt",
"orte_init:startup:internal-failure",
true, error, ORTE_ERROR_NAME(ret), ret);
return ret;
}
/*
* Sequentially map the ranks according to the placement in the
* specified hostfile
*/
static int orte_rmaps_seq_map(orte_job_t *jdata)
{
orte_job_map_t *map;
orte_app_context_t *app;
int i, n;
orte_std_cntr_t j;
opal_list_item_t *item;
orte_node_t *node, *nd;
seq_node_t *sq, *save=NULL, *seq;;
orte_vpid_t vpid;
orte_std_cntr_t num_nodes;
int rc;
opal_list_t default_seq_list;
opal_list_t node_list, *seq_list, sq_list;
orte_proc_t *proc;
mca_base_component_t *c = &mca_rmaps_seq_component.base_version;
char *hosts = NULL, *sep, *eptr;
FILE *fp;
opal_hwloc_resource_type_t rtype;
OPAL_OUTPUT_VERBOSE((1, orte_rmaps_base_framework.framework_output,
"%s rmaps:seq called on job %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_JOBID_PRINT(jdata->jobid)));
/* this mapper can only handle initial launch
* when seq mapping is desired - allow
* restarting of failed apps
*/
if (ORTE_FLAG_TEST(jdata, ORTE_JOB_FLAG_RESTART)) {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:seq: job %s is being restarted - seq cannot map",
ORTE_JOBID_PRINT(jdata->jobid));
return ORTE_ERR_TAKE_NEXT_OPTION;
}
if (NULL != jdata->map->req_mapper) {
if (0 != strcasecmp(jdata->map->req_mapper, c->mca_component_name)) {
/* a mapper has been specified, and it isn't me */
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:seq: job %s not using sequential mapper",
ORTE_JOBID_PRINT(jdata->jobid));
return ORTE_ERR_TAKE_NEXT_OPTION;
}
/* we need to process it */
goto process;
}
if (ORTE_MAPPING_SEQ != ORTE_GET_MAPPING_POLICY(jdata->map->mapping)) {
/* I don't know how to do these - defer */
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:seq: job %s not using seq mapper",
ORTE_JOBID_PRINT(jdata->jobid));
return ORTE_ERR_TAKE_NEXT_OPTION;
}
process:
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:seq: mapping job %s",
ORTE_JOBID_PRINT(jdata->jobid));
/* flag that I did the mapping */
if (NULL != jdata->map->last_mapper) {
free(jdata->map->last_mapper);
}
jdata->map->last_mapper = strdup(c->mca_component_name);
/* convenience def */
map = jdata->map;
/* if there is a default hostfile, go and get its ordered list of nodes */
OBJ_CONSTRUCT(&default_seq_list, opal_list_t);
if (NULL != orte_default_hostfile) {
char *hstname = NULL;
/* open the file */
fp = fopen(orte_default_hostfile, "r");
if (NULL == fp) {
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
rc = ORTE_ERR_NOT_FOUND;
goto error;
}
while (NULL != (hstname = orte_getline(fp))) {
if (0 == strlen(hstname)) {
free(hstname);
/* blank line - ignore */
continue;
}
if( '#' == hstname[0] ) {
free(hstname);
/* Comment line - ignore */
continue;
}
sq = OBJ_NEW(seq_node_t);
if (NULL != (sep = strchr(hstname, ' '))) {
*sep = '\0';
sep++;
/* remove any trailing space */
eptr = sep + strlen(sep) - 1;
while (eptr > sep && isspace(*eptr)) {
eptr--;
}
*(eptr+1) = 0;
sq->cpuset = strdup(sep);
}
// Strip off the FQDN if present, ignore IP addresses
if( !orte_keep_fqdn_hostnames && !opal_net_isaddr(hstname) ) {
char *ptr;
if (NULL != (ptr = strchr(hstname, '.'))) {
*ptr = '\0';
}
}
sq->hostname = hstname;
opal_list_append(&default_seq_list, &sq->super);
}
fclose(fp);
}
/* start at the beginning... */
vpid = 0;
jdata->num_procs = 0;
if (0 < opal_list_get_size(&default_seq_list)) {
save = (seq_node_t*)opal_list_get_first(&default_seq_list);
}
/* default to LOGICAL processors */
if (orte_get_attribute(&jdata->attributes, ORTE_JOB_PHYSICAL_CPUIDS, NULL, OPAL_BOOL)) {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:seq: using PHYSICAL processors");
rtype = OPAL_HWLOC_PHYSICAL;
} else {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:seq: using LOGICAL processors");
rtype = OPAL_HWLOC_LOGICAL;
}
/* initialize all the nodes as not included in this job map */
for (j=0; j < orte_node_pool->size; j++) {
if (NULL != (node = (orte_node_t*)opal_pointer_array_get_item(orte_node_pool, j))) {
ORTE_FLAG_UNSET(node, ORTE_NODE_FLAG_MAPPED);
}
}
/* cycle through the app_contexts, mapping them sequentially */
for(i=0; i < jdata->apps->size; i++) {
if (NULL == (app = (orte_app_context_t*)opal_pointer_array_get_item(jdata->apps, i))) {
continue;
}
/* dash-host trumps hostfile */
if (orte_get_attribute(&app->attributes, ORTE_APP_DASH_HOST, (void**)&hosts, OPAL_STRING)) {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:seq: using dash-host nodes on app %s", app->app);
OBJ_CONSTRUCT(&node_list, opal_list_t);
/* dash host entries cannot specify cpusets, so used the std function to retrieve the list */
if (ORTE_SUCCESS != (rc = orte_util_get_ordered_dash_host_list(&node_list, hosts))) {
ORTE_ERROR_LOG(rc);
free(hosts);
goto error;
}
free(hosts);
/* transfer the list to a seq_node_t list */
OBJ_CONSTRUCT(&sq_list, opal_list_t);
while (NULL != (nd = (orte_node_t*)opal_list_remove_first(&node_list))) {
sq = OBJ_NEW(seq_node_t);
sq->hostname = strdup(nd->name);
opal_list_append(&sq_list, &sq->super);
OBJ_RELEASE(nd);
}
OBJ_DESTRUCT(&node_list);
seq_list = &sq_list;
} else if (orte_get_attribute(&app->attributes, ORTE_APP_HOSTFILE, (void**)&hosts, OPAL_STRING)) {
char *hstname;
if (NULL == hosts) {
rc = ORTE_ERR_NOT_FOUND;
goto error;
}
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:seq: using hostfile %s nodes on app %s", hosts, app->app);
OBJ_CONSTRUCT(&sq_list, opal_list_t);
/* open the file */
fp = fopen(hosts, "r");
if (NULL == fp) {
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
rc = ORTE_ERR_NOT_FOUND;
OBJ_DESTRUCT(&sq_list);
goto error;
}
while (NULL != (hstname = orte_getline(fp))) {
if (0 == strlen(hstname)) {
free(hstname);
/* blank line - ignore */
continue;
}
if( '#' == hstname[0] ) {
free(hstname);
/* Comment line - ignore */
continue;
}
sq = OBJ_NEW(seq_node_t);
if (NULL != (sep = strchr(hstname, ' '))) {
*sep = '\0';
sep++;
/* remove any trailing space */
eptr = sep + strlen(sep) - 1;
while (eptr > sep && isspace(*eptr)) {
eptr--;
}
*(eptr+1) = 0;
sq->cpuset = strdup(sep);
}
// Strip off the FQDN if present, ignore IP addresses
if( !orte_keep_fqdn_hostnames && !opal_net_isaddr(hstname) ) {
char *ptr;
if (NULL != (ptr = strchr(hstname, '.'))) {
(*ptr) = '\0';
}
}
sq->hostname = hstname;
opal_list_append(&sq_list, &sq->super);
}
fclose(fp);
free(hosts);
seq_list = &sq_list;
} else if (0 < opal_list_get_size(&default_seq_list)) {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:seq: using default hostfile nodes on app %s", app->app);
seq_list = &default_seq_list;
} else {
/* can't do anything - no nodes available! */
orte_show_help("help-orte-rmaps-base.txt",
"orte-rmaps-base:no-available-resources",
true);
return ORTE_ERR_SILENT;
}
/* check for nolocal and remove the head node, if required */
if (map->mapping & ORTE_MAPPING_NO_USE_LOCAL) {
for (item = opal_list_get_first(seq_list);
item != opal_list_get_end(seq_list);
item = opal_list_get_next(item) ) {
seq = (seq_node_t*)item;
/* need to check ifislocal because the name in the
* hostfile may not have been FQDN, while name returned
* by gethostname may have been (or vice versa)
*/
if (orte_ifislocal(seq->hostname)) {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:seq: removing head node %s", seq->hostname);
opal_list_remove_item(seq_list, item);
OBJ_RELEASE(item); /* "un-retain" it */
}
}
}
if (NULL == seq_list || 0 == (num_nodes = (orte_std_cntr_t)opal_list_get_size(seq_list))) {
orte_show_help("help-orte-rmaps-base.txt",
"orte-rmaps-base:no-available-resources",
true);
return ORTE_ERR_SILENT;
}
/* if num_procs wasn't specified, set it now */
if (0 == app->num_procs) {
app->num_procs = num_nodes;
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:seq: setting num procs to %s for app %s",
ORTE_VPID_PRINT(app->num_procs), app->app);
} else if (num_nodes < app->num_procs) {
orte_show_help("help-orte-rmaps-base.txt", "seq:not-enough-resources", true,
app->num_procs, num_nodes);
return ORTE_ERR_SILENT;
}
if (seq_list == &default_seq_list) {
sq = save;
} else {
sq = (seq_node_t*)opal_list_get_first(seq_list);
}
for (n=0; n < app->num_procs; n++) {
/* find this node on the global array - this is necessary so
* that our mapping gets saved on that array as the objects
* returned by the hostfile function are -not- on the array
*/
node = NULL;
for (j=0; j < orte_node_pool->size; j++) {
if (NULL == (node = (orte_node_t*)opal_pointer_array_get_item(orte_node_pool, j))) {
continue;
}
if (0 == strcmp(sq->hostname, node->name)) {
break;
}
}
if (NULL == node) {
/* wasn't found - that is an error */
orte_show_help("help-orte-rmaps-seq.txt",
"orte-rmaps-seq:resource-not-found",
true, sq->hostname);
rc = ORTE_ERR_SILENT;
goto error;
}
/* ensure the node is in the map */
if (!ORTE_FLAG_TEST(node, ORTE_NODE_FLAG_MAPPED)) {
OBJ_RETAIN(node);
opal_pointer_array_add(map->nodes, node);
jdata->map->num_nodes++;
ORTE_FLAG_SET(node, ORTE_NODE_FLAG_MAPPED);
}
proc = orte_rmaps_base_setup_proc(jdata, node, i);
if ((node->slots < (int)node->num_procs) ||
(0 < node->slots_max && node->slots_max < (int)node->num_procs)) {
if (ORTE_MAPPING_NO_OVERSUBSCRIBE & ORTE_GET_MAPPING_DIRECTIVE(jdata->map->mapping)) {
orte_show_help("help-orte-rmaps-base.txt", "orte-rmaps-base:alloc-error",
true, node->num_procs, app->app);
ORTE_UPDATE_EXIT_STATUS(ORTE_ERROR_DEFAULT_EXIT_CODE);
rc = ORTE_ERR_SILENT;
goto error;
}
/* flag the node as oversubscribed so that sched-yield gets
* properly set
*/
ORTE_FLAG_SET(node, ORTE_NODE_FLAG_OVERSUBSCRIBED);
ORTE_FLAG_SET(jdata, ORTE_JOB_FLAG_OVERSUBSCRIBED);
/* check for permission */
if (ORTE_FLAG_TEST(node, ORTE_NODE_FLAG_SLOTS_GIVEN)) {
/* if we weren't given a directive either way, then we will error out
* as the #slots were specifically given, either by the host RM or
* via hostfile/dash-host */
if (!(ORTE_MAPPING_SUBSCRIBE_GIVEN & ORTE_GET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping))) {
orte_show_help("help-orte-rmaps-base.txt", "orte-rmaps-base:alloc-error",
true, app->num_procs, app->app);
ORTE_UPDATE_EXIT_STATUS(ORTE_ERROR_DEFAULT_EXIT_CODE);
return ORTE_ERR_SILENT;
} else if (ORTE_MAPPING_NO_OVERSUBSCRIBE & ORTE_GET_MAPPING_DIRECTIVE(jdata->map->mapping)) {
/* if we were explicitly told not to oversubscribe, then don't */
orte_show_help("help-orte-rmaps-base.txt", "orte-rmaps-base:alloc-error",
true, app->num_procs, app->app);
ORTE_UPDATE_EXIT_STATUS(ORTE_ERROR_DEFAULT_EXIT_CODE);
return ORTE_ERR_SILENT;
}
}
}
/* assign the vpid */
proc->name.vpid = vpid++;
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:seq: assign proc %s to node %s for app %s",
ORTE_VPID_PRINT(proc->name.vpid), sq->hostname, app->app);
/* record the cpuset, if given */
if (NULL != sq->cpuset) {
hwloc_cpuset_t bitmap;
char *cpu_bitmap;
if (NULL == node->topology) {
/* not allowed - for sequential cpusets, we must have
* the topology info
*/
orte_show_help("help-orte-rmaps-base.txt", "rmaps:no-topology", true, node->name);
rc = ORTE_ERR_SILENT;
goto error;
}
/* if we are using hwthreads as cpus and binding to hwthreads, then
* we can just copy the cpuset across as it already specifies things
* at that level */
if (opal_hwloc_use_hwthreads_as_cpus &&
OPAL_BIND_TO_HWTHREAD == OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy)) {
cpu_bitmap = strdup(sq->cpuset);
} else {
/* setup the bitmap */
bitmap = hwloc_bitmap_alloc();
/* parse the slot_list to find the socket and core */
if (ORTE_SUCCESS != (rc = opal_hwloc_base_slot_list_parse(sq->cpuset, node->topology, rtype, bitmap))) {
ORTE_ERROR_LOG(rc);
hwloc_bitmap_free(bitmap);
goto error;
}
/* note that we cannot set the proc locale to any specific object
* as the slot list may have assigned it to more than one - so
* leave that field NULL
*/
/* set the proc to the specified map */
hwloc_bitmap_list_asprintf(&cpu_bitmap, bitmap);
hwloc_bitmap_free(bitmap);
}
orte_set_attribute(&proc->attributes, ORTE_PROC_CPU_BITMAP, ORTE_ATTR_GLOBAL, cpu_bitmap, OPAL_STRING);
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:seq: binding proc %s to cpuset %s bitmap %s",
ORTE_VPID_PRINT(proc->name.vpid), sq->cpuset, cpu_bitmap);
/* we are going to bind to cpuset since the user is specifying the cpus */
OPAL_SET_BINDING_POLICY(jdata->map->binding, OPAL_BIND_TO_CPUSET);
/* note that the user specified the mapping */
ORTE_SET_MAPPING_POLICY(jdata->map->mapping, ORTE_MAPPING_BYUSER);
ORTE_SET_MAPPING_DIRECTIVE(jdata->map->mapping, ORTE_MAPPING_GIVEN);
/* cleanup */
free(cpu_bitmap);
} else {
hwloc_obj_t locale;
/* assign the locale - okay for the topo to be null as
* it just means it wasn't returned
*/
if (NULL != node->topology) {
locale = hwloc_get_root_obj(node->topology);
orte_set_attribute(&proc->attributes, ORTE_PROC_HWLOC_LOCALE,
ORTE_ATTR_LOCAL, locale, OPAL_PTR);
}
}
/* add to the jdata proc array */
if (ORTE_SUCCESS != (rc = opal_pointer_array_set_item(jdata->procs, proc->name.vpid, proc))) {
ORTE_ERROR_LOG(rc);
goto error;
}
/* move to next node */
sq = (seq_node_t*)opal_list_get_next(&sq->super);
}
/** track the total number of processes we mapped */
jdata->num_procs += app->num_procs;
/* cleanup the node list if it came from this app_context */
if (seq_list != &default_seq_list) {
OPAL_LIST_DESTRUCT(seq_list);
} else {
save = sq;
}
}
return ORTE_SUCCESS;
error:
OPAL_LIST_DESTRUCT(&default_seq_list);
return rc;
}
static void recv_data(int fd, short args, void *cbdata)
{
bool found;
int i, rc;
orte_node_t *nd, *nd2;
opal_list_t nds, ndtmp;
opal_list_item_t *item, *itm;
char recv_msg[8192];
int nbytes, idx, sjob;
char **alloc, *nodelist, *tpn;
local_jobtracker_t *ptr, *jtrk;
local_apptracker_t *aptrk;
orte_app_context_t *app;
orte_jobid_t jobid;
orte_job_t *jdata;
char **dash_host = NULL;
opal_output_verbose(2, orte_ras_base_framework.framework_output,
"%s ras:slurm: dynamic allocation - data recvd",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
/* read the data from the socket and put it in the
* nodes field of op
*/
memset(recv_msg, 0, sizeof(recv_msg));
nbytes = read(fd, recv_msg, sizeof(recv_msg) - 1);
opal_output_verbose(2, orte_ras_base_framework.framework_output,
"%s ras:slurm: dynamic allocation msg: %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), recv_msg);
/* check if we got something */
if (0 == nbytes || 0 == strlen(recv_msg) || strstr(recv_msg, "failure") != NULL) {
/* show an error here - basically, a "nothing was available"
* message
*/
orte_show_help("help-ras-slurm.txt", "slurm-dyn-alloc-failed", true,
(0 == strlen(recv_msg)) ? "NO MSG" : recv_msg);
ORTE_ACTIVATE_JOB_STATE(NULL, ORTE_JOB_STATE_ALLOC_FAILED);
return;
}
/* break the message into its component parts, separated by colons */
alloc = opal_argv_split(recv_msg, ':');
/* the first section contains the ORTE jobid for this allocation */
tpn = strchr(alloc[0], '=');
orte_util_convert_string_to_jobid(&jobid, tpn+1);
/* get the corresponding job object */
jdata = orte_get_job_data_object(jobid);
jtrk = NULL;
/* find the associated tracking object */
for (item = opal_list_get_first(&jobs);
item != opal_list_get_end(&jobs);
item = opal_list_get_next(item)) {
ptr = (local_jobtracker_t*)item;
if (ptr->jobid == jobid) {
jtrk = ptr;
break;
}
}
if (NULL == jtrk) {
orte_show_help("help-ras-slurm.txt", "slurm-dyn-alloc-failed", true, "NO JOB TRACKER");
ORTE_ACTIVATE_JOB_STATE(NULL, ORTE_JOB_STATE_ALLOC_FAILED);
opal_argv_free(alloc);
return;
}
/* stop the timeout event */
opal_event_del(&jtrk->timeout_ev);
/* cycle across all the remaining parts - each is the allocation for
* an app in this job
*/
OBJ_CONSTRUCT(&nds, opal_list_t);
OBJ_CONSTRUCT(&ndtmp, opal_list_t);
idx = -1;
sjob = -1;
nodelist = NULL;
tpn = NULL;
for (i=1; NULL != alloc[i]; i++) {
if (ORTE_SUCCESS != parse_alloc_msg(alloc[i], &idx, &sjob, &nodelist, &tpn)) {
orte_show_help("help-ras-slurm.txt", "slurm-dyn-alloc-failed", true, jtrk->cmd);
ORTE_ACTIVATE_JOB_STATE(jdata, ORTE_JOB_STATE_ALLOC_FAILED);
opal_argv_free(alloc);
if (NULL != nodelist) {
free(nodelist);
}
if (NULL != tpn) {
free(tpn);
}
return;
}
if (idx < 0) {
orte_show_help("help-ras-slurm.txt", "slurm-dyn-alloc-failed", true, jtrk->cmd);
ORTE_ACTIVATE_JOB_STATE(jdata, ORTE_JOB_STATE_ALLOC_FAILED);
opal_argv_free(alloc);
free(nodelist);
free(tpn);
return;
}
if (NULL == (app = (orte_app_context_t*)opal_pointer_array_get_item(jdata->apps, idx))) {
orte_show_help("help-ras-slurm.txt", "slurm-dyn-alloc-failed", true, jtrk->cmd);
ORTE_ACTIVATE_JOB_STATE(jdata, ORTE_JOB_STATE_ALLOC_FAILED);
opal_argv_free(alloc);
free(nodelist);
free(tpn);
return;
}
/* release the current dash_host as that contained the *desired* allocation */
orte_remove_attribute(&app->attributes, ORTE_APP_DASH_HOST);
/* track the Slurm jobid */
if (NULL == (aptrk = (local_apptracker_t*)opal_pointer_array_get_item(&jtrk->apps, idx))) {
aptrk = OBJ_NEW(local_apptracker_t);
opal_pointer_array_set_item(&jtrk->apps, idx, aptrk);
}
aptrk->sjob = sjob;
/* since the nodelist/tpn may contain regular expressions, parse them */
if (ORTE_SUCCESS != (rc = orte_ras_slurm_discover(nodelist, tpn, &ndtmp))) {
ORTE_ERROR_LOG(rc);
ORTE_ACTIVATE_JOB_STATE(jdata, ORTE_JOB_STATE_ALLOC_FAILED);
opal_argv_free(alloc);
free(nodelist);
free(tpn);
return;
}
/* transfer the discovered nodes to our node list, and construct
* the new dash_host entry to match what was allocated
*/
while (NULL != (item = opal_list_remove_first(&ndtmp))) {
nd = (orte_node_t*)item;
opal_argv_append_nosize(&dash_host, nd->name);
/* check for duplicates */
found = false;
for (itm = opal_list_get_first(&nds);
itm != opal_list_get_end(&nds);
itm = opal_list_get_next(itm)) {
nd2 = (orte_node_t*)itm;
if (0 == strcmp(nd->name, nd2->name)) {
found = true;
nd2->slots += nd->slots;
OBJ_RELEASE(item);
break;
}
}
if (!found) {
/* append the new node to our list */
opal_list_append(&nds, item);
}
}
/* cleanup */
free(nodelist);
free(tpn);
}
/* cleanup */
opal_argv_free(alloc);
OBJ_DESTRUCT(&ndtmp);
if (NULL != dash_host) {
tpn = opal_argv_join(dash_host, ',');
orte_set_attribute(&app->attributes, ORTE_APP_DASH_HOST, ORTE_ATTR_LOCAL, (void*)tpn, OPAL_STRING);
opal_argv_free(dash_host);
free(tpn);
}
if (opal_list_is_empty(&nds)) {
/* if we get here, then we were able to contact slurm,
* which means we are in an actively managed cluster.
* However, slurm indicated that nothing is currently
* available that meets our requirements. This is a fatal
* situation - we do NOT have the option of running on
* user-specified hosts as the cluster is managed.
*/
OBJ_DESTRUCT(&nds);
orte_show_help("help-ras-base.txt", "ras-base:no-allocation", true);
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
}
/* store the found nodes */
if (ORTE_SUCCESS != (rc = orte_ras_base_node_insert(&nds, jdata))) {
ORTE_ERROR_LOG(rc);
OBJ_DESTRUCT(&nds);
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
return;
}
OBJ_DESTRUCT(&nds);
/* default to no-oversubscribe-allowed for managed systems */
if (!(ORTE_MAPPING_SUBSCRIBE_GIVEN & ORTE_GET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping))) {
ORTE_SET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping, ORTE_MAPPING_NO_OVERSUBSCRIBE);
}
/* flag that the allocation is managed */
orte_managed_allocation = true;
/* move the job along */
ORTE_ACTIVATE_JOB_STATE(jdata, ORTE_JOB_STATE_ALLOCATION_COMPLETE);
/* all done */
return;
}
/* process incoming messages in order of receipt */
void orte_plm_base_recv(int status, orte_process_name_t* sender,
opal_buffer_t* buffer, orte_rml_tag_t tag,
void* cbdata)
{
orte_plm_cmd_flag_t command;
orte_std_cntr_t count;
orte_jobid_t job;
orte_job_t *jdata, *parent;
opal_buffer_t *answer;
orte_vpid_t vpid;
orte_proc_t *proc;
orte_proc_state_t state;
orte_exit_code_t exit_code;
int32_t rc=ORTE_SUCCESS, ret;
orte_app_context_t *app, *child_app;
orte_process_name_t name;
pid_t pid;
bool running;
int i;
char **env;
char *prefix_dir;
OPAL_OUTPUT_VERBOSE((5, orte_plm_base_framework.framework_output,
"%s plm:base:receive processing msg",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
count = 1;
if (ORTE_SUCCESS != (rc = opal_dss.unpack(buffer, &command, &count, ORTE_PLM_CMD))) {
ORTE_ERROR_LOG(rc);
goto CLEANUP;
}
switch (command) {
case ORTE_PLM_LAUNCH_JOB_CMD:
OPAL_OUTPUT_VERBOSE((5, orte_plm_base_framework.framework_output,
"%s plm:base:receive job launch command from %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(sender)));
/* unpack the job object */
count = 1;
if (ORTE_SUCCESS != (rc = opal_dss.unpack(buffer, &jdata, &count, ORTE_JOB))) {
ORTE_ERROR_LOG(rc);
goto ANSWER_LAUNCH;
}
/* record the sender so we know who to respond to */
jdata->originator.jobid = sender->jobid;
jdata->originator.vpid = sender->vpid;
/* get the parent's job object */
if (NULL != (parent = orte_get_job_data_object(sender->jobid))) {
/* if the prefix was set in the parent's job, we need to transfer
* that prefix to the child's app_context so any further launch of
* orteds can find the correct binary. There always has to be at
* least one app_context in both parent and child, so we don't
* need to check that here. However, be sure not to overwrite
* the prefix if the user already provided it!
*/
app = (orte_app_context_t*)opal_pointer_array_get_item(parent->apps, 0);
child_app = (orte_app_context_t*)opal_pointer_array_get_item(jdata->apps, 0);
prefix_dir = NULL;
if (orte_get_attribute(&app->attributes, ORTE_APP_PREFIX_DIR, (void**)&prefix_dir, OPAL_STRING) &&
!orte_get_attribute(&child_app->attributes, ORTE_APP_PREFIX_DIR, NULL, OPAL_STRING)) {
orte_set_attribute(&child_app->attributes, ORTE_APP_PREFIX_DIR, ORTE_ATTR_GLOBAL, prefix_dir, OPAL_STRING);
}
if (NULL != prefix_dir) {
free(prefix_dir);
}
}
/* if the user asked to forward any envars, cycle through the app contexts
* in the comm_spawn request and add them
*/
if (NULL != orte_forwarded_envars) {
for (i=0; i < jdata->apps->size; i++) {
if (NULL == (app = (orte_app_context_t*)opal_pointer_array_get_item(jdata->apps, i))) {
continue;
}
env = opal_environ_merge(orte_forwarded_envars, app->env);
opal_argv_free(app->env);
app->env = env;
}
}
OPAL_OUTPUT_VERBOSE((5, orte_plm_base_framework.framework_output,
"%s plm:base:receive adding hosts",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
/* process any add-hostfile and add-host options that were provided */
if (ORTE_SUCCESS != (rc = orte_ras_base_add_hosts(jdata))) {
ORTE_ERROR_LOG(rc);
goto ANSWER_LAUNCH;
}
if (NULL != parent) {
if (NULL == parent->bookmark) {
/* find the sender's node in the job map */
if (NULL != (proc = (orte_proc_t*)opal_pointer_array_get_item(parent->procs, sender->vpid))) {
/* set the bookmark so the child starts from that place - this means
* that the first child process could be co-located with the proc
* that called comm_spawn, assuming slots remain on that node. Otherwise,
* the procs will start on the next available node
*/
jdata->bookmark = proc->node;
}
} else {
jdata->bookmark = parent->bookmark;
}
}
/* launch it */
OPAL_OUTPUT_VERBOSE((5, orte_plm_base_framework.framework_output,
"%s plm:base:receive calling spawn",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
if (ORTE_SUCCESS != (rc = orte_plm.spawn(jdata))) {
ORTE_ERROR_LOG(rc);
goto ANSWER_LAUNCH;
}
break;
ANSWER_LAUNCH:
OPAL_OUTPUT_VERBOSE((5, orte_plm_base_framework.framework_output,
"%s plm:base:receive - error on launch: %d",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), rc));
/* setup the response */
answer = OBJ_NEW(opal_buffer_t);
/* pack the error code to be returned */
if (ORTE_SUCCESS != (ret = opal_dss.pack(answer, &rc, 1, OPAL_INT32))) {
ORTE_ERROR_LOG(ret);
}
/* send the response back to the sender */
if (0 > (ret = orte_rml.send_buffer_nb(sender, answer, ORTE_RML_TAG_PLM_PROXY,
orte_rml_send_callback, NULL))) {
ORTE_ERROR_LOG(ret);
OBJ_RELEASE(answer);
}
break;
case ORTE_PLM_UPDATE_PROC_STATE:
opal_output_verbose(5, orte_plm_base_framework.framework_output,
"%s plm:base:receive update proc state command from %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(sender));
count = 1;
while (ORTE_SUCCESS == (rc = opal_dss.unpack(buffer, &job, &count, ORTE_JOBID))) {
opal_output_verbose(5, orte_plm_base_framework.framework_output,
"%s plm:base:receive got update_proc_state for job %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_JOBID_PRINT(job));
name.jobid = job;
running = false;
/* get the job object */
jdata = orte_get_job_data_object(job);
count = 1;
while (ORTE_SUCCESS == (rc = opal_dss.unpack(buffer, &vpid, &count, ORTE_VPID))) {
if (ORTE_VPID_INVALID == vpid) {
/* flag indicates that this job is complete - move on */
break;
}
name.vpid = vpid;
/* unpack the pid */
count = 1;
if (ORTE_SUCCESS != (rc = opal_dss.unpack(buffer, &pid, &count, OPAL_PID))) {
ORTE_ERROR_LOG(rc);
goto CLEANUP;
}
/* unpack the state */
count = 1;
if (ORTE_SUCCESS != (rc = opal_dss.unpack(buffer, &state, &count, ORTE_PROC_STATE))) {
ORTE_ERROR_LOG(rc);
goto CLEANUP;
}
if (ORTE_PROC_STATE_RUNNING == state) {
running = true;
}
/* unpack the exit code */
count = 1;
if (ORTE_SUCCESS != (rc = opal_dss.unpack(buffer, &exit_code, &count, ORTE_EXIT_CODE))) {
ORTE_ERROR_LOG(rc);
goto CLEANUP;
}
OPAL_OUTPUT_VERBOSE((5, orte_plm_base_framework.framework_output,
"%s plm:base:receive got update_proc_state for vpid %lu state %s exit_code %d",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
(unsigned long)vpid, orte_proc_state_to_str(state), (int)exit_code));
if (NULL != jdata) {
/* get the proc data object */
if (NULL == (proc = (orte_proc_t*)opal_pointer_array_get_item(jdata->procs, vpid))) {
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
}
/* NEVER update the proc state before activating the state machine - let
* the state cbfunc update it as it may need to compare this
* state against the prior proc state */
proc->pid = pid;
proc->exit_code = exit_code;
ORTE_ACTIVATE_PROC_STATE(&name, state);
}
}
/* record that we heard back from a daemon during app launch */
if (running && NULL != jdata) {
jdata->num_daemons_reported++;
if (orte_report_launch_progress) {
if (0 == jdata->num_daemons_reported % 100 ||
jdata->num_daemons_reported == orte_process_info.num_procs) {
ORTE_ACTIVATE_JOB_STATE(jdata, ORTE_JOB_STATE_REPORT_PROGRESS);
}
}
}
/* prepare for next job */
count = 1;
}
if (ORTE_ERR_UNPACK_READ_PAST_END_OF_BUFFER != rc) {
ORTE_ERROR_LOG(rc);
} else {
rc = ORTE_SUCCESS;
}
break;
case ORTE_PLM_REGISTERED_CMD:
count=1;
if (ORTE_SUCCESS != (rc = opal_dss.unpack(buffer, &job, &count, ORTE_JOBID))) {
ORTE_ERROR_LOG(rc);
goto DEPART;
}
name.jobid = job;
/* get the job object */
if (NULL == (jdata = orte_get_job_data_object(job))) {
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
rc = ORTE_ERR_NOT_FOUND;
goto DEPART;
}
count=1;
while (ORTE_SUCCESS == opal_dss.unpack(buffer, &vpid, &count, ORTE_VPID)) {
name.vpid = vpid;
ORTE_ACTIVATE_PROC_STATE(&name, ORTE_PROC_STATE_REGISTERED);
count=1;
}
break;
default:
ORTE_ERROR_LOG(ORTE_ERR_VALUE_OUT_OF_BOUNDS);
rc = ORTE_ERR_VALUE_OUT_OF_BOUNDS;
break;
}
CLEANUP:
if (ORTE_SUCCESS != rc) {
goto DEPART;
}
DEPART:
/* see if an error occurred - if so, wakeup the HNP so we can exit */
if (ORTE_PROC_IS_HNP && ORTE_SUCCESS != rc) {
jdata = NULL;
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
}
OPAL_OUTPUT_VERBOSE((5, orte_plm_base_framework.framework_output,
"%s plm:base:receive done processing commands",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
}
static void proc_errors(int fd, short args, void *cbdata)
{
orte_state_caddy_t *caddy = (orte_state_caddy_t*)cbdata;
orte_job_t *jdata;
orte_process_name_t *proc = &caddy->name;
orte_proc_state_t state = caddy->proc_state;
orte_proc_t *child, *ptr;
opal_buffer_t *alert;
orte_plm_cmd_flag_t cmd;
int rc=ORTE_SUCCESS;
int i;
OPAL_OUTPUT_VERBOSE((2, orte_errmgr_base_framework.framework_output,
"%s errmgr:default_orted:proc_errors process %s error state %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(proc),
orte_proc_state_to_str(state)));
/*
* if orte is trying to shutdown, just let it
*/
if (orte_finalizing) {
OPAL_OUTPUT_VERBOSE((2, orte_errmgr_base_framework.framework_output,
"%s errmgr:default_orted:proc_errors finalizing - ignoring error",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
goto cleanup;
}
/* if this is a heartbeat failure, let the HNP handle it */
if (ORTE_PROC_STATE_HEARTBEAT_FAILED == state) {
OPAL_OUTPUT_VERBOSE((2, orte_errmgr_base_framework.framework_output,
"%s errmgr:default_orted:proc_errors heartbeat failed - ignoring error",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
goto cleanup;
}
/* if this was a failed comm, then see if it was to our
* lifeline
*/
if (ORTE_PROC_STATE_LIFELINE_LOST == state ||
ORTE_PROC_STATE_UNABLE_TO_SEND_MSG == state) {
OPAL_OUTPUT_VERBOSE((2, orte_errmgr_base_framework.framework_output,
"%s errmgr:orted lifeline lost - exiting",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
/* set our exit status */
ORTE_UPDATE_EXIT_STATUS(ORTE_ERROR_DEFAULT_EXIT_CODE);
/* kill our children */
killprocs(ORTE_JOBID_WILDCARD, ORTE_VPID_WILDCARD);
/* terminate - our routed children will see
* us leave and automatically die
*/
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
goto cleanup;
}
/* get the job object */
if (NULL == (jdata = orte_get_job_data_object(proc->jobid))) {
/* must already be complete */
OPAL_OUTPUT_VERBOSE((2, orte_errmgr_base_framework.framework_output,
"%s errmgr:default_orted:proc_errors NULL jdata - ignoring error",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
goto cleanup;
}
if (ORTE_PROC_STATE_COMM_FAILED == state) {
/* if it is our own connection, ignore it */
if (OPAL_EQUAL == orte_util_compare_name_fields(ORTE_NS_CMP_ALL, ORTE_PROC_MY_NAME, proc)) {
OPAL_OUTPUT_VERBOSE((2, orte_errmgr_base_framework.framework_output,
"%s errmgr:default_orted:proc_errors comm_failed to self - ignoring error",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
goto cleanup;
}
/* was it a daemon? */
if (proc->jobid != ORTE_PROC_MY_NAME->jobid) {
/* nope - we can't seem to trust that we will catch the waitpid
* in this situation, so push this over to be handled as if
* it were a waitpid trigger so we don't create a bunch of
* duplicate code */
OPAL_OUTPUT_VERBOSE((2, orte_errmgr_base_framework.framework_output,
"%s errmgr:default_orted:proc_errors comm_failed to non-daemon - handling as waitpid",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
/* get the proc_t */
if (NULL == (child = (orte_proc_t*)opal_pointer_array_get_item(jdata->procs, proc->vpid))) {
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
goto cleanup;
}
/* leave the exit code alone - process this as a waitpid */
odls_base_default_wait_local_proc(child, NULL);
goto cleanup;
}
OPAL_OUTPUT_VERBOSE((2, orte_errmgr_base_framework.framework_output,
"%s errmgr:default:orted daemon %s exited",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(proc)));
/* if we are using static ports, then it is possible that the HNP
* will not see this termination. So if the HNP didn't order us
* to terminate, then we should ensure it knows */
if (orte_static_ports && !orte_orteds_term_ordered) {
/* send an alert to the HNP */
alert = OBJ_NEW(opal_buffer_t);
/* pack update state command */
cmd = ORTE_PLM_UPDATE_PROC_STATE;
if (ORTE_SUCCESS != (rc = opal_dss.pack(alert, &cmd, 1, ORTE_PLM_CMD))) {
ORTE_ERROR_LOG(rc);
return;
}
/* get the proc_t */
if (NULL == (child = (orte_proc_t*)opal_pointer_array_get_item(jdata->procs, proc->vpid))) {
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
goto cleanup;
}
/* set the exit code to reflect the problem */
child->exit_code = ORTE_ERR_COMM_FAILURE;
/* pack only the data for this daemon - have to start with the jobid
* so the receiver can unpack it correctly
*/
if (ORTE_SUCCESS != (rc = opal_dss.pack(alert, &proc->jobid, 1, ORTE_JOBID))) {
ORTE_ERROR_LOG(rc);
return;
}
/* now pack the daemon's info */
if (ORTE_SUCCESS != (rc = pack_state_for_proc(alert, child))) {
ORTE_ERROR_LOG(rc);
return;
}
/* send it */
OPAL_OUTPUT_VERBOSE((5, orte_errmgr_base_framework.framework_output,
"%s errmgr:default_orted reporting lost connection to daemon %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(proc)));
if (0 > (rc = orte_rml.send_buffer_nb(ORTE_PROC_MY_HNP, alert,
ORTE_RML_TAG_PLM,
orte_rml_send_callback, NULL))) {
ORTE_ERROR_LOG(rc);
OBJ_RELEASE(alert);
}
/* mark that we notified the HNP for this job so we don't do it again */
orte_set_attribute(&jdata->attributes, ORTE_JOB_FAIL_NOTIFIED, ORTE_ATTR_LOCAL, NULL, OPAL_BOOL);
/* continue on */
goto cleanup;
}
if (orte_orteds_term_ordered) {
/* are any of my children still alive */
for (i=0; i < orte_local_children->size; i++) {
if (NULL != (child = (orte_proc_t*)opal_pointer_array_get_item(orte_local_children, i))) {
if (ORTE_FLAG_TEST(child, ORTE_PROC_FLAG_ALIVE)) {
OPAL_OUTPUT_VERBOSE((5, orte_state_base_framework.framework_output,
"%s errmgr:default:orted[%s(%d)] proc %s is alive",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
__FILE__, __LINE__,
ORTE_NAME_PRINT(&child->name)));
goto cleanup;
}
}
}
/* if all my routes and children are gone, then terminate
ourselves nicely (i.e., this is a normal termination) */
if (0 == orte_routed.num_routes()) {
OPAL_OUTPUT_VERBOSE((2, orte_errmgr_base_framework.framework_output,
"%s errmgr:default:orted all routes gone - exiting",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
ORTE_ACTIVATE_JOB_STATE(NULL, ORTE_JOB_STATE_DAEMONS_TERMINATED);
} else {
OPAL_OUTPUT_VERBOSE((2, orte_errmgr_base_framework.framework_output,
"%s errmgr:default:orted not exiting, num_routes() == %d",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
(int)orte_routed.num_routes()));
}
}
/* if not, then we can continue */
goto cleanup;
}
if (NULL == (child = (orte_proc_t*)opal_pointer_array_get_item(jdata->procs, proc->vpid))) {
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
goto cleanup;
}
/* if this is not a local proc for this job, we can
* ignore this call
*/
if (!ORTE_FLAG_TEST(child, ORTE_PROC_FLAG_LOCAL)) {
OPAL_OUTPUT_VERBOSE((2, orte_errmgr_base_framework.framework_output,
"%s errmgr:default_orted:proc_errors proc is not local - ignoring error",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
goto cleanup;
}
OPAL_OUTPUT_VERBOSE((2, orte_errmgr_base_framework.framework_output,
"%s errmgr:default_orted got state %s for proc %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
orte_proc_state_to_str(state),
ORTE_NAME_PRINT(proc)));
if (ORTE_PROC_STATE_TERM_NON_ZERO == state) {
/* update the state */
child->state = state;
/* report this as abnormal termination to the HNP, unless we already have
* done so for this job */
if (!orte_get_attribute(&jdata->attributes, ORTE_JOB_FAIL_NOTIFIED, NULL, OPAL_BOOL)) {
alert = OBJ_NEW(opal_buffer_t);
/* pack update state command */
cmd = ORTE_PLM_UPDATE_PROC_STATE;
if (ORTE_SUCCESS != (rc = opal_dss.pack(alert, &cmd, 1, ORTE_PLM_CMD))) {
ORTE_ERROR_LOG(rc);
return;
}
/* pack only the data for this proc - have to start with the jobid
* so the receiver can unpack it correctly
*/
if (ORTE_SUCCESS != (rc = opal_dss.pack(alert, &proc->jobid, 1, ORTE_JOBID))) {
ORTE_ERROR_LOG(rc);
return;
}
/* now pack the child's info */
if (ORTE_SUCCESS != (rc = pack_state_for_proc(alert, child))) {
ORTE_ERROR_LOG(rc);
return;
}
/* send it */
OPAL_OUTPUT_VERBOSE((5, orte_errmgr_base_framework.framework_output,
"%s errmgr:default_orted reporting proc %s abnormally terminated with non-zero status (local procs = %d)",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&child->name),
jdata->num_local_procs));
if (0 > (rc = orte_rml.send_buffer_nb(ORTE_PROC_MY_HNP, alert,
ORTE_RML_TAG_PLM,
orte_rml_send_callback, NULL))) {
ORTE_ERROR_LOG(rc);
OBJ_RELEASE(alert);
}
/* mark that we notified the HNP for this job so we don't do it again */
orte_set_attribute(&jdata->attributes, ORTE_JOB_FAIL_NOTIFIED, ORTE_ATTR_LOCAL, NULL, OPAL_BOOL);
}
/* if the proc has terminated, notify the state machine */
if (ORTE_FLAG_TEST(child, ORTE_PROC_FLAG_IOF_COMPLETE) &&
ORTE_FLAG_TEST(child, ORTE_PROC_FLAG_WAITPID) &&
!ORTE_FLAG_TEST(child, ORTE_PROC_FLAG_RECORDED)) {
ORTE_ACTIVATE_PROC_STATE(proc, ORTE_PROC_STATE_TERMINATED);
}
goto cleanup;
}
if (ORTE_PROC_STATE_FAILED_TO_START == state ||
ORTE_PROC_STATE_FAILED_TO_LAUNCH == state) {
/* update the proc state */
child->state = state;
/* count the proc as having "terminated" */
jdata->num_terminated++;
/* leave the error report in this case to the
* state machine, which will receive notice
* when all local procs have attempted to start
* so that we send a consolidated error report
* back to the HNP
*/
goto cleanup;
}
if (ORTE_PROC_STATE_TERMINATED < state) {
/* if we were ordered to terminate, see if
* any of our routes or local children remain alive - if not, then
* terminate ourselves. */
if (orte_orteds_term_ordered) {
/* mark the child as no longer alive and update the counters, if necessary.
* we have to do this here as we aren't going to send this to the state
* machine, and we want to keep the bookkeeping accurate just in case */
if (ORTE_FLAG_TEST(child, ORTE_PROC_FLAG_ALIVE)) {
ORTE_FLAG_UNSET(child, ORTE_PROC_FLAG_ALIVE);
}
if (!ORTE_FLAG_TEST(child, ORTE_PROC_FLAG_RECORDED)) {
ORTE_FLAG_SET(child, ORTE_PROC_FLAG_RECORDED);
jdata->num_terminated++;
}
for (i=0; i < orte_local_children->size; i++) {
if (NULL != (child = (orte_proc_t*)opal_pointer_array_get_item(orte_local_children, i))) {
if (ORTE_FLAG_TEST(child, ORTE_PROC_FLAG_ALIVE)) {
goto keep_going;
}
}
}
/* if all my routes and children are gone, then terminate
ourselves nicely (i.e., this is a normal termination) */
if (0 == orte_routed.num_routes()) {
OPAL_OUTPUT_VERBOSE((2, orte_errmgr_base_framework.framework_output,
"%s errmgr:default:orted all routes gone - exiting",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
ORTE_ACTIVATE_JOB_STATE(NULL, ORTE_JOB_STATE_DAEMONS_TERMINATED);
}
/* no need to alert the HNP - we are already on our way out */
goto cleanup;
}
keep_going:
/* if the job hasn't completed and the state is abnormally
* terminated, then we need to alert the HNP right away - but
* only do this once!
*/
if (!orte_get_attribute(&jdata->attributes, ORTE_JOB_FAIL_NOTIFIED, NULL, OPAL_BOOL)) {
alert = OBJ_NEW(opal_buffer_t);
/* pack update state command */
cmd = ORTE_PLM_UPDATE_PROC_STATE;
if (ORTE_SUCCESS != (rc = opal_dss.pack(alert, &cmd, 1, ORTE_PLM_CMD))) {
ORTE_ERROR_LOG(rc);
return;
}
/* pack only the data for this proc - have to start with the jobid
* so the receiver can unpack it correctly
*/
if (ORTE_SUCCESS != (rc = opal_dss.pack(alert, &proc->jobid, 1, ORTE_JOBID))) {
ORTE_ERROR_LOG(rc);
return;
}
child->state = state;
/* now pack the child's info */
if (ORTE_SUCCESS != (rc = pack_state_for_proc(alert, child))) {
ORTE_ERROR_LOG(rc);
return;
}
OPAL_OUTPUT_VERBOSE((5, orte_errmgr_base_framework.framework_output,
"%s errmgr:default_orted reporting proc %s aborted to HNP (local procs = %d)",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&child->name),
jdata->num_local_procs));
/* send it */
if (0 > (rc = orte_rml.send_buffer_nb(ORTE_PROC_MY_HNP, alert,
ORTE_RML_TAG_PLM,
orte_rml_send_callback, NULL))) {
ORTE_ERROR_LOG(rc);
}
/* mark that we notified the HNP for this job so we don't do it again */
orte_set_attribute(&jdata->attributes, ORTE_JOB_FAIL_NOTIFIED, ORTE_ATTR_LOCAL, NULL, OPAL_BOOL);
}
/* if the proc has terminated, notify the state machine */
if (ORTE_FLAG_TEST(child, ORTE_PROC_FLAG_IOF_COMPLETE) &&
ORTE_FLAG_TEST(child, ORTE_PROC_FLAG_WAITPID) &&
!ORTE_FLAG_TEST(child, ORTE_PROC_FLAG_RECORDED)) {
ORTE_ACTIVATE_PROC_STATE(proc, ORTE_PROC_STATE_TERMINATED);
}
goto cleanup;
}
/* only other state is terminated - see if anyone is left alive */
if (!any_live_children(proc->jobid)) {
alert = OBJ_NEW(opal_buffer_t);
/* pack update state command */
cmd = ORTE_PLM_UPDATE_PROC_STATE;
if (ORTE_SUCCESS != (rc = opal_dss.pack(alert, &cmd, 1, ORTE_PLM_CMD))) {
ORTE_ERROR_LOG(rc);
return;
}
/* pack the data for the job */
if (ORTE_SUCCESS != (rc = pack_state_update(alert, jdata))) {
ORTE_ERROR_LOG(rc);
return;
}
OPAL_OUTPUT_VERBOSE((5, orte_errmgr_base_framework.framework_output,
"%s errmgr:default_orted reporting all procs in %s terminated",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_JOBID_PRINT(jdata->jobid)));
/* remove all of this job's children from the global list */
for (i=0; i < orte_local_children->size; i++) {
if (NULL == (ptr = (orte_proc_t*)opal_pointer_array_get_item(orte_local_children, i))) {
continue;
}
if (jdata->jobid == ptr->name.jobid) {
opal_pointer_array_set_item(orte_local_children, i, NULL);
OBJ_RELEASE(ptr);
}
}
/* ensure the job's local session directory tree is removed */
orte_session_dir_cleanup(jdata->jobid);
/* remove this job from our local job data since it is complete */
opal_pointer_array_set_item(orte_job_data, ORTE_LOCAL_JOBID(jdata->jobid), NULL);
OBJ_RELEASE(jdata);
/* send it */
if (0 > (rc = orte_rml.send_buffer_nb(ORTE_PROC_MY_HNP, alert,
ORTE_RML_TAG_PLM,
orte_rml_send_callback, NULL))) {
ORTE_ERROR_LOG(rc);
}
return;
}
cleanup:
OBJ_RELEASE(caddy);
}