static void send_self_exe(int fd, short args, void* data)
{
orte_self_send_xfer_t *xfer = (orte_self_send_xfer_t*)data;
ORTE_ACQUIRE_OBJECT(xfer);
OPAL_OUTPUT_VERBOSE((1, orte_rml_base_framework.framework_output,
"%s rml_send_to_self callback executing for tag %d",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), xfer->tag));
/* execute the send callback function - note that
* send-to-self always returns a SUCCESS status
*/
if (NULL != xfer->iov) {
if (NULL != xfer->cbfunc.iov) {
/* non-blocking iovec send */
xfer->cbfunc.iov(ORTE_SUCCESS, ORTE_PROC_MY_NAME, xfer->iov, xfer->count,
xfer->tag, xfer->cbdata);
}
} else if (NULL != xfer->buffer) {
if (NULL != xfer->cbfunc.buffer) {
/* non-blocking buffer send */
xfer->cbfunc.buffer(ORTE_SUCCESS, ORTE_PROC_MY_NAME, xfer->buffer,
xfer->tag, xfer->cbdata);
}
} else {
/* should never happen */
abort();
}
/* cleanup the memory */
OBJ_RELEASE(xfer);
}
/* 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);
}
/*
* Event callback when there is data available on the registered
* socket to recv. This is called for the listen sockets to accept an
* incoming connection, on new sockets trying to complete the software
* connection process, and for probes. Data on an established
* connection is handled elsewhere.
*/
static void recv_handler(int sd, short flg, void *cbdata)
{
mca_oob_tcp_conn_op_t *op = (mca_oob_tcp_conn_op_t*)cbdata;
int flags;
mca_oob_tcp_hdr_t hdr;
mca_oob_tcp_peer_t *peer;
ORTE_ACQUIRE_OBJECT(op);
opal_output_verbose(OOB_TCP_DEBUG_CONNECT, orte_oob_base_framework.framework_output,
"%s:tcp:recv:handler called",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
/* get the handshake */
if (ORTE_SUCCESS != mca_oob_tcp_peer_recv_connect_ack(NULL, sd, &hdr)) {
goto cleanup;
}
/* finish processing ident */
if (MCA_OOB_TCP_IDENT == hdr.type) {
if (NULL == (peer = mca_oob_tcp_peer_lookup(&hdr.origin))) {
/* should never happen */
mca_oob_tcp_peer_close(peer);
goto cleanup;
}
/* set socket up to be non-blocking */
if ((flags = fcntl(sd, F_GETFL, 0)) < 0) {
opal_output(0, "%s mca_oob_tcp_recv_connect: fcntl(F_GETFL) failed: %s (%d)",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), strerror(opal_socket_errno), opal_socket_errno);
} else {
flags |= O_NONBLOCK;
if (fcntl(sd, F_SETFL, flags) < 0) {
opal_output(0, "%s mca_oob_tcp_recv_connect: fcntl(F_SETFL) failed: %s (%d)",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), strerror(opal_socket_errno), opal_socket_errno);
}
}
/* is the peer instance willing to accept this connection */
peer->sd = sd;
if (mca_oob_tcp_peer_accept(peer) == false) {
if (OOB_TCP_DEBUG_CONNECT <= opal_output_get_verbosity(orte_oob_base_framework.framework_output)) {
opal_output(0, "%s-%s mca_oob_tcp_recv_connect: "
"rejected connection from %s connection state %d",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer->name)),
ORTE_NAME_PRINT(&(hdr.origin)),
peer->state);
}
CLOSE_THE_SOCKET(sd);
}
}
cleanup:
OBJ_RELEASE(op);
}
/* after we map, we are ready to launch the daemons */
static void map_complete(int fd, short args, void *cbdata)
{
orte_state_caddy_t *state = (orte_state_caddy_t*)cbdata;
orte_job_t *jdata;
ORTE_ACQUIRE_OBJECT(caddy);
jdata = state->jdata;
jdata->state = ORTE_JOB_STATE_MAP_COMPLETE;
/* move to the map stage */
ORTE_ACTIVATE_JOB_STATE(jdata, ORTE_JOB_STATE_LAUNCH_DAEMONS);
/* cleanup */
OBJ_RELEASE(state);
}
static void vm_ready(int fd, short args, void *cbdata)
{
orte_state_caddy_t *state = (orte_state_caddy_t*)cbdata;
orte_job_t *jdata;
ORTE_ACQUIRE_OBJECT(caddy);
jdata = state->jdata;
/* now that the daemons are launched, we are ready
* to roll
*/
jdata->state = ORTE_JOB_STATE_VM_READY;
ORTE_ACTIVATE_JOB_STATE(jdata, ORTE_JOB_STATE_SYSTEM_PREP);
OBJ_RELEASE(state);
}
static void pmix_server_release(int status, opal_buffer_t *buf, void *cbdata)
{
orte_pmix_mdx_caddy_t *cd=(orte_pmix_mdx_caddy_t*)cbdata;
char *data = NULL;
int32_t ndata = 0;
int rc = OPAL_SUCCESS;
ORTE_ACQUIRE_OBJECT(cd);
/* unload the buffer */
if (NULL != buf) {
rc = opal_dss.unload(buf, (void**)&data, &ndata);
}
if (OPAL_SUCCESS == rc) {
rc = status;
}
cd->cbfunc(rc, data, ndata, cd->cbdata, relcb, data);
OBJ_RELEASE(cd);
}
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;
char *rtmod;
ORTE_ACQUIRE_OBJECT(caddy);
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);
rtmod = orte_rml.get_routed(orte_mgmt_conduit);
/* 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(rtmod, proc);
/* if all my routes and local children are gone, then terminate ourselves */
if (0 == orte_routed.num_routes(rtmod)) {
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(rtmod)));
}
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)) {
/* 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);
if (!orte_enable_recovery) {
/* output an error message so the user knows what happened */
orte_show_help("help-errmgr-base.txt", "node-died", true,
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
orte_process_info.nodename,
ORTE_NAME_PRINT(proc),
pptr->node->name);
/* 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);
}
}
/* if recovery is enabled, then we are done - otherwise,
* abort the system */
if (!orte_enable_recovery) {
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(rtmod)) {
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:
/* if this is a continuously operating job, then there is nothing more
* to do - we let the job continue to run */
if (orte_get_attribute(&jdata->attributes, ORTE_JOB_CONTINUOUS_OP, NULL, OPAL_BOOL) ||
ORTE_FLAG_TEST(jdata, ORTE_JOB_FLAG_RECOVERABLE)) {
/* always mark the waitpid as having fired */
ORTE_ACTIVATE_PROC_STATE(&pptr->name, ORTE_PROC_STATE_WAITPID_FIRED);
/* if this is a remote proc, we won't hear anything more about it
* as the default behavior would be to terminate the job. So be sure to
* mark the IOF as having completed too so we correctly mark this proc
* as dead and notify everyone as required */
if (!ORTE_FLAG_TEST(pptr, ORTE_PROC_FLAG_LOCAL)) {
ORTE_ACTIVATE_PROC_STATE(&pptr->name, ORTE_PROC_STATE_IOF_COMPLETE);
}
goto cleanup;
}
/* 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(rtmod, 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;
case ORTE_PROC_STATE_NO_PATH_TO_TARGET:
OPAL_OUTPUT_VERBOSE((5, orte_errmgr_base_framework.framework_output,
"%s errmgr:hnp: no message path to proc %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(proc)));
orte_show_help("help-errmgr-base.txt", "no-path", true,
orte_process_info.nodename, pptr->node->name);
/* 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;
case ORTE_PROC_STATE_FAILED_TO_CONNECT:
OPAL_OUTPUT_VERBOSE((5, orte_errmgr_base_framework.framework_output,
"%s errmgr:hnp: cannot connect to proc %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(proc)));
orte_show_help("help-errmgr-base.txt", "no-connect", true,
orte_process_info.nodename, pptr->node->name);
/* 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 job_errors(int fd, short args, void *cbdata)
{
orte_state_caddy_t *caddy = (orte_state_caddy_t*)cbdata;
orte_job_t *jdata;
orte_job_state_t jobstate;
orte_exit_code_t sts;
orte_proc_t *aborted_proc;
opal_buffer_t *answer;
int32_t rc, ret;
int room, *rmptr;
ORTE_ACQUIRE_OBJECT(caddy);
/*
* if orte is trying to shutdown, just let it
*/
if (orte_finalizing) {
return;
}
/* ensure we have an error exit status */
ORTE_UPDATE_EXIT_STATUS(ORTE_ERROR_DEFAULT_EXIT_CODE);
/* if the jdata is NULL, then we abort as this
* is reporting an unrecoverable error
*/
if (NULL == caddy->jdata) {
ORTE_ACTIVATE_JOB_STATE(NULL, ORTE_JOB_STATE_FORCED_EXIT);
OBJ_RELEASE(caddy);
return;
}
/* update the state */
jdata = caddy->jdata;
jobstate = caddy->job_state;
jdata->state = jobstate;
OPAL_OUTPUT_VERBOSE((1, orte_errmgr_base_framework.framework_output,
"%s errmgr:default_hnp: job %s reported state %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_JOBID_PRINT(jdata->jobid),
orte_job_state_to_str(jobstate)));
if (ORTE_JOB_STATE_NEVER_LAUNCHED == jobstate ||
ORTE_JOB_STATE_ALLOC_FAILED == jobstate ||
ORTE_JOB_STATE_MAP_FAILED == jobstate ||
ORTE_JOB_STATE_CANNOT_LAUNCH == jobstate) {
if (1 == ORTE_LOCAL_JOBID(jdata->jobid)) {
/* this is the primary job */
orte_never_launched = true;
}
/* disable routing as we may not have performed the daemon
* wireup - e.g., in a managed environment, all the daemons
* "phone home", but don't actually wireup into the routed
* network until they receive the launch message
*/
orte_routing_is_enabled = false;
jdata->num_terminated = jdata->num_procs;
/* activate the terminated state so we can exit */
ORTE_ACTIVATE_JOB_STATE(caddy->jdata, ORTE_JOB_STATE_TERMINATED);
/* if it was a dynamic spawn, then we better tell them this didn't work */
if (ORTE_JOBID_INVALID != jdata->originator.jobid) {
rc = jobstate;
answer = OBJ_NEW(opal_buffer_t);
if (ORTE_SUCCESS != (ret = opal_dss.pack(answer, &rc, 1, OPAL_INT32))) {
ORTE_ERROR_LOG(ret);
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
OBJ_RELEASE(caddy);
return;
}
if (ORTE_SUCCESS != (ret = opal_dss.pack(answer, &jdata->jobid, 1, ORTE_JOBID))) {
ORTE_ERROR_LOG(ret);
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
OBJ_RELEASE(caddy);
return;
}
/* pack the room number */
rmptr = &room;
if (orte_get_attribute(&jdata->attributes, ORTE_JOB_ROOM_NUM, (void**)&rmptr, OPAL_INT)) {
if (ORTE_SUCCESS != (ret = opal_dss.pack(answer, &room, 1, OPAL_INT))) {
ORTE_ERROR_LOG(ret);
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
OBJ_RELEASE(caddy);
return;
}
}
OPAL_OUTPUT_VERBOSE((5, orte_errmgr_base_framework.framework_output,
"%s errmgr:hnp sending dyn error release of job %s to %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_JOBID_PRINT(jdata->jobid),
ORTE_NAME_PRINT(&jdata->originator)));
if (0 > (ret = orte_rml.send_buffer_nb(orte_mgmt_conduit,
&jdata->originator, answer,
ORTE_RML_TAG_LAUNCH_RESP,
orte_rml_send_callback, NULL))) {
ORTE_ERROR_LOG(ret);
OBJ_RELEASE(answer);
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
}
}
OBJ_RELEASE(caddy);
return;
}
if (ORTE_JOB_STATE_FAILED_TO_START == jobstate ||
ORTE_JOB_STATE_FAILED_TO_LAUNCH == jobstate) {
/* the job object for this job will have been NULL'd
* in the array if the job was solely local. If it isn't
* NULL, then we need to tell everyone else to die
*/
aborted_proc = NULL;
if (orte_get_attribute(&jdata->attributes, ORTE_JOB_ABORTED_PROC, (void**)&aborted_proc, OPAL_PTR)) {
sts = aborted_proc->exit_code;
if (ORTE_PROC_MY_NAME->jobid == jdata->jobid) {
if (WIFSIGNALED(sts)) { /* died on signal */
#ifdef WCOREDUMP
if (WCOREDUMP(sts)) {
orte_show_help("help-plm-base.txt", "daemon-died-signal-core", true,
WTERMSIG(sts));
sts = WTERMSIG(sts);
} else {
orte_show_help("help-plm-base.txt", "daemon-died-signal", true,
WTERMSIG(sts));
sts = WTERMSIG(sts);
}
#else
orte_show_help("help-plm-base.txt", "daemon-died-signal", true,
WTERMSIG(sts));
sts = WTERMSIG(sts);
#endif /* WCOREDUMP */
} else {
orte_show_help("help-plm-base.txt", "daemon-died-no-signal", true,
WEXITSTATUS(sts));
sts = WEXITSTATUS(sts);
}
}
}
/* if this is the daemon job, then we need to ensure we
* output an error message indicating we couldn't launch the
* daemons */
if (jdata->jobid == ORTE_PROC_MY_NAME->jobid) {
orte_show_help("help-errmgr-base.txt", "failed-daemon-launch", true);
}
}
/* if the daemon job aborted and we haven't heard from everyone yet,
* then this could well have been caused by a daemon not finding
* a way back to us. In this case, output a message indicating a daemon
* died without reporting. Otherwise, say nothing as we
* likely already output an error message */
if (ORTE_JOB_STATE_ABORTED == jobstate &&
jdata->jobid == ORTE_PROC_MY_NAME->jobid &&
jdata->num_procs != jdata->num_reported) {
orte_show_help("help-errmgr-base.txt", "failed-daemon", true);
}
/* abort the job */
ORTE_ACTIVATE_JOB_STATE(caddy->jdata, ORTE_JOB_STATE_FORCED_EXIT);
/* set the global abnormal exit flag */
orte_abnormal_term_ordered = true;
OBJ_RELEASE(caddy);
}
static void wakeup(int sd, short args, void *cbdata)
{
/* nothing more we can do */
ORTE_ACQUIRE_OBJECT(cbdata);
orte_quit(0, 0, NULL);
}
int main(int argc, char *argv[])
{
int ret = 0;
opal_cmd_line_t *cmd_line = NULL;
char *rml_uri;
#if OPAL_ENABLE_FT_CR == 1
char * tmp_env_var = NULL;
#endif
/* init enough of opal to process cmd lines */
if (OPAL_SUCCESS != opal_init_util(&argc, &argv)) {
fprintf(stderr, "OPAL failed to initialize -- orted aborting\n");
exit(1);
}
/* setup to check common command line options that just report and die */
cmd_line = OBJ_NEW(opal_cmd_line_t);
opal_cmd_line_create(cmd_line, orte_server_cmd_line_opts);
mca_base_cmd_line_setup(cmd_line);
if (OPAL_SUCCESS != (ret = opal_cmd_line_parse(cmd_line, false, false,
argc, argv))) {
if (OPAL_ERR_SILENT != ret) {
fprintf(stderr, "%s: command line error (%s)\n", argv[0],
opal_strerror(ret));
}
return 1;
}
/* check for help request */
if (help) {
char *str, *args = NULL;
args = opal_cmd_line_get_usage_msg(cmd_line);
str = opal_show_help_string("help-orte-server.txt",
"orteserver:usage", false,
argv[0], args);
if (NULL != str) {
printf("%s", str);
free(str);
}
free(args);
/* If we show the help message, that should be all we do */
return 0;
}
/*
* Since this process can now handle MCA/GMCA parameters, make sure to
* process them.
*/
mca_base_cmd_line_process_args(cmd_line, &environ, &environ);
/* if debug is set, then set orte_debug_flag so that the data server
* code will output
*/
if (debug) {
putenv(OPAL_MCA_PREFIX"orte_debug=1");
}
/* detach from controlling terminal
* otherwise, remain attached so output can get to us
*/
if(debug == false &&
no_daemonize == false) {
opal_daemon_init(NULL);
}
#if OPAL_ENABLE_FT_CR == 1
/* Disable the checkpoint notification routine for this
* tool. As we will never need to checkpoint this tool.
* Note: This must happen before opal_init().
*/
opal_cr_set_enabled(false);
/* Select the none component, since we don't actually use a checkpointer */
(void) mca_base_var_env_name("crs", &tmp_env_var);
opal_setenv(tmp_env_var,
"none",
true, &environ);
free(tmp_env_var);
tmp_env_var = NULL;
/* Mark as a tool program */
(void) mca_base_var_env_name("opal_cr_is_tool", &tmp_env_var);
opal_setenv(tmp_env_var,
"1",
true, &environ);
free(tmp_env_var);
#endif
/* don't want session directories */
orte_create_session_dirs = false;
/* Perform the standard init, but flag that we are an HNP */
if (ORTE_SUCCESS != (ret = orte_init(&argc, &argv, ORTE_PROC_HNP))) {
fprintf(stderr, "orte-server: failed to initialize -- aborting\n");
exit(1);
}
/* report out our URI, if we were requested to do so, using syntax
* proposed in an email thread by Jeff Squyres
*/
if (NULL != report_uri) {
orte_oob_base_get_addr(&rml_uri);
if (0 == strcmp(report_uri, "-")) {
/* if '-', then output to stdout */
printf("%s\n", rml_uri);
} else if (0 == strcmp(report_uri, "+")) {
/* if '+', output to stderr */
fprintf(stderr, "%s\n", rml_uri);
} else {
/* treat it as a filename and output into it */
FILE *fp;
fp = fopen(report_uri, "w");
if (NULL == fp) {
fprintf(stderr, "orte-server: failed to open designated file %s -- aborting\n", report_uri);
orte_finalize();
exit(1);
}
fprintf(fp, "%s\n", rml_uri);
fclose(fp);
}
free(rml_uri);
}
/* setup the data server to listen for commands */
if (ORTE_SUCCESS != (ret = orte_data_server_init())) {
fprintf(stderr, "orte-server: failed to start data server -- aborting\n");
orte_finalize();
exit(1);
}
/* setup to listen for commands sent specifically to me */
orte_rml.recv_buffer_nb(ORTE_NAME_WILDCARD, ORTE_RML_TAG_DAEMON,
ORTE_RML_NON_PERSISTENT, orte_daemon_recv, NULL);
/* Set signal handlers to catch kill signals so we can properly clean up
* after ourselves.
*/
opal_event_set(orte_event_base, &term_handler, SIGTERM, OPAL_EV_SIGNAL,
shutdown_callback, NULL);
opal_event_add(&term_handler, NULL);
opal_event_set(orte_event_base, &int_handler, SIGINT, OPAL_EV_SIGNAL,
shutdown_callback, NULL);
opal_event_add(&int_handler, NULL);
/* We actually do *not* want the server to voluntarily yield() the
processor more than necessary. The server 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 the server, we want the
OS to wake up the server in a timely fashion (which most OS's
seem good about doing) and then we want the server to process
the message as fast as possible. If the server yields and lets
aggressive MPI applications get the processor back, it may be a
long time before the OS schedules the server to run again
(particularly if there is no IO event to wake it up). Hence,
publish and lookup (for example) may be significantly delayed
before being delivered to MPI processes, which can be
problematic in some scenarios (e.g., COMM_SPAWN). */
opal_progress_set_yield_when_idle(false);
/* Change the default behavior of libevent such that we want to
continually block rather than blocking for the default timeout
and then looping around the progress engine again. There
should be nothing in the server that cannot block in libevent
until "something" happens (i.e., there's no need to keep
cycling through progress because the only things that should
happen will happen in libevent). This is a minor optimization,
but what the heck... :-) */
opal_progress_set_event_flag(OPAL_EVLOOP_ONCE);
if (debug) {
opal_output(0, "%s orte-server: up and running!", ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
}
/* wait to hear we are done */
while (orte_event_base_active) {
opal_event_loop(orte_event_base, OPAL_EVLOOP_ONCE);
}
ORTE_ACQUIRE_OBJECT(orte_event_base_active);
/* should never get here, but if we do... */
/* Finalize and clean up ourselves */
orte_finalize();
return orte_exit_status;
}
static void job_errors(int fd, short args, void *cbdata)
{
orte_state_caddy_t *caddy = (orte_state_caddy_t*)cbdata;
orte_job_t *jdata;
orte_job_state_t jobstate;
opal_buffer_t *answer;
int32_t rc, ret;
int room, *rmptr;
ORTE_ACQUIRE_OBJECT(caddy);
/*
* if orte is trying to shutdown, just let it
*/
if (orte_finalizing) {
return;
}
/* if the jdata is NULL, then we ignore it as this
* is reporting an unrecoverable error
*/
if (NULL == caddy->jdata) {
ORTE_ERROR_LOG(ORTE_ERR_BAD_PARAM);
OBJ_RELEASE(caddy);
return;
}
/* update the state */
jdata = caddy->jdata;
jobstate = caddy->job_state;
jdata->state = jobstate;
OPAL_OUTPUT_VERBOSE((1, orte_errmgr_base_framework.framework_output,
"%s errmgr:dvm: job %s reported state %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_JOBID_PRINT(jdata->jobid),
orte_job_state_to_str(jobstate)));
if (jdata->jobid == ORTE_PROC_MY_NAME->jobid) {
/* if the daemon job aborted and we haven't heard from everyone yet,
* then this could well have been caused by a daemon not finding
* a way back to us. In this case, output a message indicating a daemon
* died without reporting. Otherwise, say nothing as we
* likely already output an error message */
if (ORTE_JOB_STATE_ABORTED == jobstate &&
jdata->num_procs != jdata->num_reported) {
orte_routing_is_enabled = false;
orte_show_help("help-errmgr-base.txt", "failed-daemon", true);
}
/* there really isn't much else we can do since the problem
* is in the DVM itself, so best just to terminate */
jdata->num_terminated = jdata->num_procs;
/* activate the terminated state so we can exit */
ORTE_ACTIVATE_JOB_STATE(jdata, ORTE_JOB_STATE_TERMINATED);
OBJ_RELEASE(caddy);
return;
}
/* all other cases involve jobs submitted to the DVM - therefore,
* we only inform the submitter of the problem, but do NOT terminate
* the DVM itself */
rc = jobstate;
answer = OBJ_NEW(opal_buffer_t);
if (ORTE_SUCCESS != (ret = opal_dss.pack(answer, &rc, 1, OPAL_INT32))) {
ORTE_ERROR_LOG(ret);
OBJ_RELEASE(caddy);
return;
}
if (ORTE_SUCCESS != (ret = opal_dss.pack(answer, &jdata->jobid, 1, ORTE_JOBID))) {
ORTE_ERROR_LOG(ret);
OBJ_RELEASE(caddy);
return;
}
/* pack the room number */
rmptr = &room;
if (orte_get_attribute(&jdata->attributes, ORTE_JOB_ROOM_NUM, (void**)&rmptr, OPAL_INT)) {
if (ORTE_SUCCESS != (ret = opal_dss.pack(answer, &room, 1, OPAL_INT))) {
ORTE_ERROR_LOG(ret);
OBJ_RELEASE(caddy);
return;
}
}
OPAL_OUTPUT_VERBOSE((5, orte_errmgr_base_framework.framework_output,
"%s errmgr:dvm sending notification of job %s failure to %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_JOBID_PRINT(jdata->jobid),
ORTE_NAME_PRINT(&jdata->originator)));
if (0 > (ret = orte_rml.send_buffer_nb(orte_mgmt_conduit,
&jdata->originator, answer,
ORTE_RML_TAG_LAUNCH_RESP,
orte_rml_send_callback, NULL))) {
ORTE_ERROR_LOG(ret);
OBJ_RELEASE(answer);
}
/* ensure we terminate any processes left running in the DVM */
_terminate_job(jdata->jobid);
/* cleanup */
OBJ_RELEASE(caddy);
}
int orte_daemon(int argc, char *argv[])
{
int ret = 0;
opal_cmd_line_t *cmd_line = NULL;
int i;
opal_buffer_t *buffer;
char hostname[OPAL_MAXHOSTNAMELEN];
pmix_value_t val;
pmix_proc_t proc;
pmix_status_t prc;
myxfer_t xfer;
pmix_data_buffer_t pbuf;
pmix_byte_object_t pbo;
opal_byte_object_t bo, *boptr;
/* initialize the globals */
memset(&orted_globals, 0, sizeof(orted_globals));
/* initialize the singleton died pipe to an illegal value so we can detect it was set */
orted_globals.singleton_died_pipe = -1;
bucket = OBJ_NEW(opal_buffer_t);
/* setup to check common command line options that just report and die */
cmd_line = OBJ_NEW(opal_cmd_line_t);
if (OPAL_SUCCESS != opal_cmd_line_create(cmd_line, orte_cmd_line_opts)) {
OBJ_RELEASE(cmd_line);
exit(1);
}
mca_base_cmd_line_setup(cmd_line);
if (ORTE_SUCCESS != (ret = opal_cmd_line_parse(cmd_line, false, false,
argc, argv))) {
char *args = NULL;
args = opal_cmd_line_get_usage_msg(cmd_line);
fprintf(stderr, "Usage: %s [OPTION]...\n%s\n", argv[0], args);
free(args);
OBJ_RELEASE(cmd_line);
return ret;
}
/*
* Since this process can now handle MCA/GMCA parameters, make sure to
* process them.
*/
mca_base_cmd_line_process_args(cmd_line, &environ, &environ);
/* Ensure that enough of OPAL is setup for us to be able to run */
/*
* NOTE: (JJH)
* We need to allow 'mca_base_cmd_line_process_args()' to process command
* line arguments *before* calling opal_init_util() since the command
* line could contain MCA parameters that affect the way opal_init_util()
* functions. AMCA parameters are one such option normally received on the
* command line that affect the way opal_init_util() behaves.
* 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 != opal_init_util(&argc, &argv)) {
fprintf(stderr, "OPAL failed to initialize -- orted aborting\n");
exit(1);
}
/* save the environment for launch purposes. This MUST be
* done so that we can pass it to any local procs we
* spawn - otherwise, those local procs won't see any
* non-MCA envars that were set in the enviro when the
* orted was executed - e.g., by .csh
*/
orte_launch_environ = opal_argv_copy(environ);
/* purge any ess/pmix flags set in the environ when we were launched */
opal_unsetenv(OPAL_MCA_PREFIX"ess", &orte_launch_environ);
opal_unsetenv(OPAL_MCA_PREFIX"pmix", &orte_launch_environ);
/* if orte_daemon_debug is set, let someone know we are alive right
* away just in case we have a problem along the way
*/
if (orted_globals.debug) {
gethostname(hostname, sizeof(hostname));
fprintf(stderr, "Daemon was launched on %s - beginning to initialize\n", hostname);
}
/* check for help request */
if (orted_globals.help) {
char *args = NULL;
args = opal_cmd_line_get_usage_msg(cmd_line);
orte_show_help("help-orted.txt", "orted:usage", false,
argv[0], args);
free(args);
return 1;
}
#if defined(HAVE_SETSID)
/* see if we were directed to separate from current session */
if (orted_globals.set_sid) {
setsid();
}
#endif
/* see if they want us to spin until they can connect a debugger to us */
i=0;
while (orted_spin_flag) {
i++;
if (1000 < i) i=0;
}
/* detach from controlling terminal
* otherwise, remain attached so output can get to us
*/
if(!orte_debug_flag &&
!orte_debug_daemons_flag &&
orted_globals.daemonize) {
opal_daemon_init(NULL);
}
/* Set the flag telling OpenRTE that I am NOT a
* singleton, but am "infrastructure" - prevents setting
* up incorrect infrastructure that only a singleton would
* require.
*/
if (orted_globals.hnp) {
if (ORTE_SUCCESS != (ret = orte_init(&argc, &argv, ORTE_PROC_HNP))) {
ORTE_ERROR_LOG(ret);
return ret;
}
} else {
if (ORTE_SUCCESS != (ret = orte_init(&argc, &argv, ORTE_PROC_DAEMON))) {
ORTE_ERROR_LOG(ret);
return ret;
}
}
/* finalize the OPAL utils. As they are opened again from orte_init->opal_init
* we continue to have a reference count on them. So we have to finalize them twice...
*/
opal_finalize_util();
/* bind ourselves if so directed */
if (NULL != orte_daemon_cores) {
char **cores=NULL, tmp[128];
hwloc_obj_t pu;
hwloc_cpuset_t ours, res;
int core;
/* could be a collection of comma-delimited ranges, so
* use our handy utility to parse it
*/
orte_util_parse_range_options(orte_daemon_cores, &cores);
if (NULL != cores) {
ours = hwloc_bitmap_alloc();
hwloc_bitmap_zero(ours);
res = hwloc_bitmap_alloc();
for (i=0; NULL != cores[i]; i++) {
core = strtoul(cores[i], NULL, 10);
if (NULL == (pu = opal_hwloc_base_get_pu(opal_hwloc_topology, core, OPAL_HWLOC_LOGICAL))) {
/* turn off the show help forwarding as we won't
* be able to cycle the event library to send
*/
orte_show_help_finalize();
/* the message will now come out locally */
orte_show_help("help-orted.txt", "orted:cannot-bind",
true, orte_process_info.nodename,
orte_daemon_cores);
ret = ORTE_ERR_NOT_SUPPORTED;
hwloc_bitmap_free(ours);
hwloc_bitmap_free(res);
goto DONE;
}
hwloc_bitmap_or(res, ours, pu->cpuset);
hwloc_bitmap_copy(ours, res);
}
/* if the result is all zeros, then don't bind */
if (!hwloc_bitmap_iszero(ours)) {
(void)hwloc_set_cpubind(opal_hwloc_topology, ours, 0);
if (opal_hwloc_report_bindings) {
opal_hwloc_base_cset2mapstr(tmp, sizeof(tmp), opal_hwloc_topology, ours);
opal_output(0, "Daemon %s is bound to cores %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), tmp);
}
}
/* cleanup */
hwloc_bitmap_free(ours);
hwloc_bitmap_free(res);
opal_argv_free(cores);
}
}
if ((int)ORTE_VPID_INVALID != orted_debug_failure) {
orted_globals.abort=false;
/* some vpid was ordered to fail. The value can be positive
* or negative, depending upon the desired method for failure,
* so need to check both here
*/
if (0 > orted_debug_failure) {
orted_debug_failure = -1*orted_debug_failure;
orted_globals.abort = true;
}
/* are we the specified vpid? */
if ((int)ORTE_PROC_MY_NAME->vpid == orted_debug_failure) {
/* if the user specified we delay, then setup a timer
* and have it kill us
*/
if (0 < orted_debug_failure_delay) {
ORTE_TIMER_EVENT(orted_debug_failure_delay, 0, shutdown_callback, ORTE_SYS_PRI);
} else {
opal_output(0, "%s is executing clean %s", ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
orted_globals.abort ? "abort" : "abnormal termination");
/* do -not- call finalize as this will send a message to the HNP
* indicating clean termination! Instead, just forcibly cleanup
* the local session_dir tree and exit
*/
orte_session_dir_cleanup(ORTE_JOBID_WILDCARD);
/* if we were ordered to abort, do so */
if (orted_globals.abort) {
abort();
}
/* otherwise, return with non-zero status */
ret = ORTE_ERROR_DEFAULT_EXIT_CODE;
goto DONE;
}
}
}
/* insert our contact info into our process_info struct so we
* have it for later use and set the local daemon field to our name
*/
orte_oob_base_get_addr(&orte_process_info.my_daemon_uri);
if (NULL == orte_process_info.my_daemon_uri) {
/* no way to communicate */
ret = ORTE_ERROR;
goto DONE;
}
ORTE_PROC_MY_DAEMON->jobid = ORTE_PROC_MY_NAME->jobid;
ORTE_PROC_MY_DAEMON->vpid = ORTE_PROC_MY_NAME->vpid;
PMIX_VALUE_LOAD(&val, orte_process_info.my_daemon_uri, OPAL_STRING);
(void)opal_snprintf_jobid(proc.nspace, PMIX_MAX_NSLEN, ORTE_PROC_MY_NAME->jobid);
proc.rank = ORTE_PROC_MY_NAME->vpid;
if (PMIX_SUCCESS != (prc = PMIx_Store_internal(&proc, PMIX_PROC_URI, &val))) {
PMIX_ERROR_LOG(prc);
PMIX_VALUE_DESTRUCT(&val);
ret = ORTE_ERROR;
goto DONE;
}
PMIX_VALUE_DESTRUCT(&val);
/* if I am also the hnp, then update that contact info field too */
if (ORTE_PROC_IS_HNP) {
orte_process_info.my_hnp_uri = strdup(orte_process_info.my_daemon_uri);
ORTE_PROC_MY_HNP->jobid = ORTE_PROC_MY_NAME->jobid;
ORTE_PROC_MY_HNP->vpid = ORTE_PROC_MY_NAME->vpid;
}
/* setup the primary daemon command receive function */
orte_rml.recv_buffer_nb(ORTE_NAME_WILDCARD, ORTE_RML_TAG_DAEMON,
ORTE_RML_PERSISTENT, orte_daemon_recv, NULL);
/* output a message indicating we are alive, our name, and our pid
* for debugging purposes
*/
if (orte_debug_daemons_flag) {
fprintf(stderr, "Daemon %s checking in as pid %ld on host %s\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), (long)orte_process_info.pid,
orte_process_info.nodename);
}
/* We actually do *not* want the orted to voluntarily yield() the
processor more than necessary. The orted 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 the orted, we want the
OS to wake up the orted in a timely fashion (which most OS's
seem good about doing) and then we want the orted to process
the message as fast as possible. If the orted yields and lets
aggressive MPI applications get the processor back, it may be a
long time before the OS schedules the orted 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);
/* Change the default behavior of libevent such that we want to
continually block rather than blocking for the default timeout
and then looping around the progress engine again. There
should be nothing in the orted that cannot block in libevent
until "something" happens (i.e., there's no need to keep
cycling through progress because the only things that should
happen will happen in libevent). This is a minor optimization,
but what the heck... :-) */
opal_progress_set_event_flag(OPAL_EVLOOP_ONCE);
/* if requested, report my uri to the indicated pipe */
if (orted_globals.uri_pipe > 0) {
orte_job_t *jdata;
orte_proc_t *p;
orte_node_t *node;
orte_app_context_t *app;
char *tmp, *nptr, *sysinfo;
char **singenv=NULL, *string_key, *env_str;
/* setup the singleton's job */
jdata = OBJ_NEW(orte_job_t);
/* default to ompi for now */
opal_argv_append_nosize(&jdata->personality, "ompi");
orte_plm_base_create_jobid(jdata);
opal_hash_table_set_value_uint32(orte_job_data, jdata->jobid, jdata);
/* must create a map for it (even though it has no
* info in it) so that the job info will be picked
* up in subsequent pidmaps or other daemons won't
* know how to route
*/
jdata->map = OBJ_NEW(orte_job_map_t);
/* setup an app_context for the singleton */
app = OBJ_NEW(orte_app_context_t);
app->app = strdup("singleton");
app->num_procs = 1;
opal_pointer_array_add(jdata->apps, app);
jdata->num_apps = 1;
/* setup a proc object for the singleton - since we
* -must- be the HNP, and therefore we stored our
* node on the global node pool, and since the singleton
* -must- be on the same node as us, indicate that
*/
p = OBJ_NEW(orte_proc_t);
p->name.jobid = jdata->jobid;
p->name.vpid = 0;
p->parent = 0;
ORTE_FLAG_SET(p, ORTE_PROC_FLAG_ALIVE);
p->state = ORTE_PROC_STATE_RUNNING;
p->app_idx = 0;
/* obviously, it is on my node */
node = (orte_node_t*)opal_pointer_array_get_item(orte_node_pool, 0);
p->node = node;
OBJ_RETAIN(node); /* keep accounting straight */
opal_pointer_array_add(jdata->procs, p);
jdata->num_procs = 1;
/* add the node to the job map */
OBJ_RETAIN(node);
opal_pointer_array_add(jdata->map->nodes, node);
jdata->map->num_nodes++;
/* and it obviously is on the node */
OBJ_RETAIN(p);
opal_pointer_array_add(node->procs, p);
node->num_procs++;
/* and obviously it is one of my local procs */
OBJ_RETAIN(p);
opal_pointer_array_add(orte_local_children, p);
jdata->num_local_procs = 1;
/* set the trivial */
p->local_rank = 0;
p->node_rank = 0;
p->app_rank = 0;
p->state = ORTE_PROC_STATE_RUNNING;
p->app_idx = 0;
ORTE_FLAG_SET(p, ORTE_PROC_FLAG_LOCAL);
/* register the singleton's nspace with our PMIx server */
if (ORTE_SUCCESS != (ret = orte_pmix_server_register_nspace(jdata))) {
ORTE_ERROR_LOG(ret);
goto DONE;
}
/* use setup fork to create the envars needed by the singleton */
(void)opal_snprintf_jobid(proc.nspace, PMIX_MAX_NSLEN, p->name.jobid);
proc.rank = p->name.vpid;
if (PMIX_SUCCESS != (prc = PMIx_server_setup_fork(&proc, &singenv))) {
PMIX_ERROR_LOG(prc);
ret = ORTE_ERROR;
goto DONE;
}
/* append the transport key to the envars needed by the singleton */
if (!orte_get_attribute(&jdata->attributes, ORTE_JOB_TRANSPORT_KEY, (void**)&string_key, OPAL_STRING) || NULL == string_key) {
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
goto DONE;
}
opal_asprintf(&env_str, OPAL_MCA_PREFIX"orte_precondition_transports=%s", string_key);
opal_argv_append_nosize(&singenv, env_str);
free(env_str);
nptr = opal_argv_join(singenv, '*');
opal_argv_free(singenv);
/* create a string that contains our uri + sysinfo + PMIx server URI envars */
orte_util_convert_sysinfo_to_string(&sysinfo, orte_local_cpu_type, orte_local_cpu_model);
opal_asprintf(&tmp, "%s[%s]%s", orte_process_info.my_daemon_uri, sysinfo, nptr);
free(sysinfo);
free(nptr);
/* pass that info to the singleton */
if (OPAL_SUCCESS != (ret = opal_fd_write(orted_globals.uri_pipe, strlen(tmp)+1, tmp))) { ; /* need to add 1 to get the NULL */
ORTE_ERROR_LOG(ret);
goto DONE;
}
/* cleanup */
free(tmp);
close(orted_globals.uri_pipe);
/* since a singleton spawned us, we need to harvest
* any MCA params from the local environment so
* we can pass them along to any subsequent daemons
* we may start as the result of a comm_spawn
*/
for (i=0; NULL != environ[i]; i++) {
if (0 == strncmp(environ[i], OPAL_MCA_PREFIX, 9)) {
/* make a copy to manipulate */
tmp = strdup(environ[i]);
/* find the equal sign */
nptr = strchr(tmp, '=');
*nptr = '\0';
nptr++;
/* add the mca param to the orted cmd line */
opal_argv_append_nosize(&orted_cmd_line, "-"OPAL_MCA_CMD_LINE_ID);
opal_argv_append_nosize(&orted_cmd_line, &tmp[9]);
opal_argv_append_nosize(&orted_cmd_line, nptr);
free(tmp);
}
}
}
/* if we were given a pipe to monitor for singleton termination, set that up */
if (orted_globals.singleton_died_pipe > 0) {
/* register shutdown handler */
pipe_handler = (opal_event_t*)malloc(sizeof(opal_event_t));
opal_event_set(orte_event_base, pipe_handler,
orted_globals.singleton_died_pipe,
OPAL_EV_READ,
pipe_closed,
pipe_handler);
opal_event_add(pipe_handler, NULL);
}
/* If I have a parent, then save his contact info so
* any messages we send can flow thru him.
*/
orte_parent_uri = NULL;
(void) mca_base_var_register ("orte", "orte", NULL, "parent_uri",
"URI for the parent if tree launch is enabled.",
MCA_BASE_VAR_TYPE_STRING, NULL, 0,
MCA_BASE_VAR_FLAG_INTERNAL,
OPAL_INFO_LVL_9,
MCA_BASE_VAR_SCOPE_CONSTANT,
&orte_parent_uri);
if (NULL != orte_parent_uri) {
/* set the contact info into our local database */
ret = orte_rml_base_parse_uris(orte_parent_uri, ORTE_PROC_MY_PARENT, NULL);
if (ORTE_SUCCESS != ret) {
ORTE_ERROR_LOG(ret);
goto DONE;
}
PMIX_VALUE_LOAD(&val, orte_parent_uri, OPAL_STRING);
OPAL_PMIX_CONVERT_NAME(&proc, ORTE_PROC_MY_PARENT);
if (PMIX_SUCCESS != (prc = PMIx_Store_internal(&proc, PMIX_PROC_URI, &val))) {
PMIX_ERROR_LOG(prc);
PMIX_VALUE_DESTRUCT(&val);
ret = ORTE_ERROR;
goto DONE;
}
PMIX_VALUE_DESTRUCT(&val);
/* tell the routed module that we have a path
* back to the HNP
*/
if (ORTE_SUCCESS != (ret = orte_routed.update_route(ORTE_PROC_MY_HNP, ORTE_PROC_MY_PARENT))) {
ORTE_ERROR_LOG(ret);
goto DONE;
}
/* and a path to our parent */
if (ORTE_SUCCESS != (ret = orte_routed.update_route(ORTE_PROC_MY_PARENT, ORTE_PROC_MY_PARENT))) {
ORTE_ERROR_LOG(ret);
goto DONE;
}
/* set the lifeline to point to our parent so that we
* can handle the situation if that lifeline goes away
*/
if (ORTE_SUCCESS != (ret = orte_routed.set_lifeline(ORTE_PROC_MY_PARENT))) {
ORTE_ERROR_LOG(ret);
goto DONE;
}
}
/* if we are not the HNP...the only time we will be an HNP
* is if we are launched by a singleton to provide support
* for it
*/
if (!ORTE_PROC_IS_HNP) {
orte_process_name_t target;
/* setup the rollup callback */
orte_rml.recv_buffer_nb(ORTE_NAME_WILDCARD, ORTE_RML_TAG_ORTED_CALLBACK,
ORTE_RML_PERSISTENT, rollup, NULL);
/* define the target jobid */
target.jobid = ORTE_PROC_MY_NAME->jobid;
if (orte_fwd_mpirun_port || orte_static_ports || NULL != orte_parent_uri) {
/* we start by sending to ourselves */
target.vpid = ORTE_PROC_MY_NAME->vpid;
/* since we will be waiting for any children to send us
* their rollup info before sending to our parent, save
* a little time in the launch phase by "warming up" the
* connection to our parent while we wait for our children */
buffer = OBJ_NEW(opal_buffer_t); // zero-byte message
orte_rml.recv_buffer_nb(ORTE_PROC_MY_PARENT, ORTE_RML_TAG_NODE_REGEX_REPORT,
ORTE_RML_PERSISTENT, node_regex_report, &node_regex_waiting);
node_regex_waiting = true;
if (0 > (ret = orte_rml.send_buffer_nb(ORTE_PROC_MY_PARENT, buffer,
ORTE_RML_TAG_WARMUP_CONNECTION,
orte_rml_send_callback, NULL))) {
ORTE_ERROR_LOG(ret);
OBJ_RELEASE(buffer);
goto DONE;
}
} else {
target.vpid = 0;
}
/* send the information to the orted report-back point - this function
* will process the data, but also counts the number of
* orteds that reported back so the launch procedure can continue.
* We need to do this at the last possible second as the HNP
* can turn right around and begin issuing orders to us
*/
buffer = OBJ_NEW(opal_buffer_t);
/* insert our name for rollup purposes */
if (ORTE_SUCCESS != (ret = opal_dss.pack(buffer, ORTE_PROC_MY_NAME, 1, ORTE_NAME))) {
ORTE_ERROR_LOG(ret);
OBJ_RELEASE(buffer);
goto DONE;
}
/* get any connection info we may have pushed */
{
pmix_info_t *info;
size_t ninfo;
pmix_value_t *vptr;
OPAL_PMIX_CONVERT_NAME(&proc, ORTE_PROC_MY_NAME);
boptr = &bo;
bo.bytes = NULL;
bo.size = 0;
if (PMIX_SUCCESS == PMIx_Get(&proc, NULL, NULL, 0, &vptr) && NULL != vptr) {
/* the data is returned as a pmix_data_array_t */
if (PMIX_DATA_ARRAY != vptr->type || NULL == vptr->data.darray ||
PMIX_INFO != vptr->data.darray->type || NULL == vptr->data.darray->array) {
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
OBJ_RELEASE(buffer);
goto DONE;
}
/* use the PMIx data support to pack it */
info = (pmix_info_t*)vptr->data.darray->array;
ninfo = vptr->data.darray->size;
PMIX_DATA_BUFFER_CONSTRUCT(&pbuf);
if (PMIX_SUCCESS != (prc = PMIx_Data_pack(&proc, &pbuf, &ninfo, 1, PMIX_SIZE))) {
PMIX_ERROR_LOG(prc);
ret = ORTE_ERROR;
OBJ_RELEASE(buffer);
goto DONE;
}
if (PMIX_SUCCESS != (prc = PMIx_Data_pack(&proc, &pbuf, info, ninfo, PMIX_INFO))) {
PMIX_ERROR_LOG(prc);
ret = ORTE_ERROR;
OBJ_RELEASE(buffer);
goto DONE;
}
PMIX_DATA_BUFFER_UNLOAD(&pbuf, pbo.bytes, pbo.size);
bo.bytes = (uint8_t*)pbo.bytes;
bo.size = pbo.size;
PMIX_VALUE_RELEASE(vptr);
}
if (ORTE_SUCCESS != (ret = opal_dss.pack(buffer, &boptr, 1, OPAL_BYTE_OBJECT))) {
ORTE_ERROR_LOG(ret);
OBJ_RELEASE(buffer);
goto DONE;
}
}
/* include our node name */
opal_dss.pack(buffer, &orte_process_info.nodename, 1, OPAL_STRING);
/* if requested, include any non-loopback aliases for this node */
if (orte_retain_aliases) {
char **aliases=NULL;
uint8_t naliases, ni;
char hostname[OPAL_MAXHOSTNAMELEN];
/* if we stripped the prefix or removed the fqdn,
* include full hostname as an alias
*/
gethostname(hostname, sizeof(hostname));
if (strlen(orte_process_info.nodename) < strlen(hostname)) {
opal_argv_append_nosize(&aliases, hostname);
}
opal_ifgetaliases(&aliases);
naliases = opal_argv_count(aliases);
if (ORTE_SUCCESS != (ret = opal_dss.pack(buffer, &naliases, 1, OPAL_UINT8))) {
ORTE_ERROR_LOG(ret);
OBJ_RELEASE(buffer);
opal_argv_free(aliases);
goto DONE;
}
for (ni=0; ni < naliases; ni++) {
if (ORTE_SUCCESS != (ret = opal_dss.pack(buffer, &aliases[ni], 1, OPAL_STRING))) {
ORTE_ERROR_LOG(ret);
OBJ_RELEASE(buffer);
opal_argv_free(aliases);
goto DONE;
}
}
opal_argv_free(aliases);
}
/* always send back our topology signature - this is a small string
* and won't hurt anything */
if (ORTE_SUCCESS != (ret = opal_dss.pack(buffer, &orte_topo_signature, 1, OPAL_STRING))) {
ORTE_ERROR_LOG(ret);
}
/* if we are rank=1, then send our topology back - otherwise, mpirun
* will request it if necessary */
if (1 == ORTE_PROC_MY_NAME->vpid) {
opal_buffer_t data;
int8_t flag;
uint8_t *cmpdata;
size_t cmplen;
/* setup an intermediate buffer */
OBJ_CONSTRUCT(&data, opal_buffer_t);
if (ORTE_SUCCESS != (ret = opal_dss.pack(&data, &opal_hwloc_topology, 1, OPAL_HWLOC_TOPO))) {
ORTE_ERROR_LOG(ret);
}
if (opal_compress.compress_block((uint8_t*)data.base_ptr, data.bytes_used,
&cmpdata, &cmplen)) {
/* the data was compressed - mark that we compressed it */
flag = 1;
if (ORTE_SUCCESS != (ret = opal_dss.pack(buffer, &flag, 1, OPAL_INT8))) {
ORTE_ERROR_LOG(ret);
free(cmpdata);
OBJ_DESTRUCT(&data);
}
/* pack the compressed length */
if (ORTE_SUCCESS != (ret = opal_dss.pack(buffer, &cmplen, 1, OPAL_SIZE))) {
ORTE_ERROR_LOG(ret);
free(cmpdata);
OBJ_DESTRUCT(&data);
}
/* pack the uncompressed length */
if (ORTE_SUCCESS != (ret = opal_dss.pack(buffer, &data.bytes_used, 1, OPAL_SIZE))) {
ORTE_ERROR_LOG(ret);
free(cmpdata);
OBJ_DESTRUCT(&data);
}
/* pack the compressed info */
if (ORTE_SUCCESS != (ret = opal_dss.pack(buffer, cmpdata, cmplen, OPAL_UINT8))) {
ORTE_ERROR_LOG(ret);
free(cmpdata);
OBJ_DESTRUCT(&data);
}
OBJ_DESTRUCT(&data);
free(cmpdata);
} else {
/* mark that it was not compressed */
flag = 0;
if (ORTE_SUCCESS != (ret = opal_dss.pack(buffer, &flag, 1, OPAL_INT8))) {
ORTE_ERROR_LOG(ret);
OBJ_DESTRUCT(&data);
free(cmpdata);
}
/* transfer the payload across */
opal_dss.copy_payload(buffer, &data);
OBJ_DESTRUCT(&data);
}
}
/* collect our network inventory */
memset(&xfer, 0, sizeof(myxfer_t));
OPAL_PMIX_CONSTRUCT_LOCK(&xfer.lock);
if (PMIX_SUCCESS != (prc = PMIx_server_collect_inventory(NULL, 0, infocbfunc, &xfer))) {
PMIX_ERROR_LOG(prc);
ret = ORTE_ERR_NOT_SUPPORTED;
goto DONE;
}
OPAL_PMIX_WAIT_THREAD(&xfer.lock);
if (NULL != xfer.info) {
PMIX_DATA_BUFFER_CONSTRUCT(&pbuf);
if (PMIX_SUCCESS != (prc = PMIx_Data_pack(NULL, &pbuf, &xfer.ninfo, 1, PMIX_SIZE))) {
PMIX_ERROR_LOG(prc);
ret = ORTE_ERROR;
OBJ_RELEASE(buffer);
goto DONE;
}
if (PMIX_SUCCESS != (prc = PMIx_Data_pack(NULL, &pbuf, xfer.info, xfer.ninfo, PMIX_INFO))) {
PMIX_ERROR_LOG(prc);
ret = ORTE_ERROR;
OBJ_RELEASE(buffer);
goto DONE;
}
PMIX_DATA_BUFFER_UNLOAD(&pbuf, pbo.bytes, pbo.size);
bo.bytes = (uint8_t*)pbo.bytes;
bo.size = pbo.size;
if (ORTE_SUCCESS != (ret = opal_dss.pack(buffer, &boptr, 1, OPAL_BYTE_OBJECT))) {
ORTE_ERROR_LOG(ret);
OBJ_RELEASE(buffer);
goto DONE;
}
}
/* send it to the designated target */
if (0 > (ret = orte_rml.send_buffer_nb(&target, buffer,
ORTE_RML_TAG_ORTED_CALLBACK,
orte_rml_send_callback, NULL))) {
ORTE_ERROR_LOG(ret);
OBJ_RELEASE(buffer);
goto DONE;
}
}
/* if we are tree-spawning, then we need to capture the MCA params
* from our cmd line so we can pass them along to the daemons we spawn -
* otherwise, only the first layer of daemons will ever see them
*/
if (orted_globals.tree_spawn) {
int j, k;
bool ignore;
char *no_keep[] = {
"orte_hnp_uri",
"orte_ess_jobid",
"orte_ess_vpid",
"orte_ess_num_procs",
"orte_parent_uri",
"mca_base_env_list",
NULL
};
for (i=0; i < argc; i++) {
if (0 == strcmp("-"OPAL_MCA_CMD_LINE_ID, argv[i]) ||
0 == strcmp("--"OPAL_MCA_CMD_LINE_ID, argv[i]) ) {
ignore = false;
/* see if this is something we cannot pass along */
for (k=0; NULL != no_keep[k]; k++) {
if (0 == strcmp(no_keep[k], argv[i+1])) {
ignore = true;
break;
}
}
if (!ignore) {
/* see if this is already present so we at least can
* avoid growing the cmd line with duplicates
*/
if (NULL != orted_cmd_line) {
for (j=0; NULL != orted_cmd_line[j]; j++) {
if (0 == strcmp(argv[i+1], orted_cmd_line[j])) {
/* already here - ignore it */
ignore = true;
break;
}
}
}
if (!ignore) {
opal_argv_append_nosize(&orted_cmd_line, argv[i]);
opal_argv_append_nosize(&orted_cmd_line, argv[i+1]);
opal_argv_append_nosize(&orted_cmd_line, argv[i+2]);
}
}
i += 2;
}
}
}
if (orte_debug_daemons_flag) {
opal_output(0, "%s orted: up and running - waiting for commands!", ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
}
ret = ORTE_SUCCESS;
/* loop the event lib until an exit event is detected */
while (orte_event_base_active) {
opal_event_loop(orte_event_base, OPAL_EVLOOP_ONCE);
}
ORTE_ACQUIRE_OBJECT(orte_event_base_active);
/* ensure all local procs are dead */
orte_odls.kill_local_procs(NULL);
DONE:
/* update the exit status, in case it wasn't done */
ORTE_UPDATE_EXIT_STATUS(ret);
/* cleanup and leave */
orte_finalize();
opal_finalize_util();
orte_session_dir_cleanup(ORTE_JOBID_WILDCARD);
/* cleanup the process info */
orte_proc_info_finalize();
if (orte_debug_flag) {
fprintf(stderr, "exiting with status %d\n", orte_exit_status);
}
exit(orte_exit_status);
}
static void launch_daemons(int fd, short args, void *cbdata)
{
orte_app_context_t *app;
orte_node_t *node;
orte_std_cntr_t n;
orte_job_map_t *map;
char *jobid_string = NULL;
char *param;
char **argv = NULL;
int argc;
int rc;
char *tmp;
char** env = NULL;
char *nodelist_flat;
char **nodelist_argv;
char *name_string;
char **custom_strings;
int num_args, i;
char *cur_prefix;
int proc_vpid_index;
bool failed_launch=true;
orte_job_t *daemons;
orte_state_caddy_t *state = (orte_state_caddy_t*)cbdata;
ORTE_ACQUIRE_OBJECT(state);
OPAL_OUTPUT_VERBOSE((1, orte_plm_base_framework.framework_output,
"%s plm:slurm: LAUNCH DAEMONS CALLED",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
/* if we are launching debugger daemons, then just go
* do it - no new daemons will be launched
*/
if (ORTE_FLAG_TEST(state->jdata, ORTE_JOB_FLAG_DEBUGGER_DAEMON)) {
state->jdata->state = ORTE_JOB_STATE_DAEMONS_LAUNCHED;
ORTE_ACTIVATE_JOB_STATE(state->jdata, ORTE_JOB_STATE_DAEMONS_REPORTED);
OBJ_RELEASE(state);
return;
}
/* start by setting up the virtual machine */
daemons = orte_get_job_data_object(ORTE_PROC_MY_NAME->jobid);
if (ORTE_SUCCESS != (rc = orte_plm_base_setup_virtual_machine(state->jdata))) {
ORTE_ERROR_LOG(rc);
goto cleanup;
}
/* if we don't want to launch, then don't attempt to
* launch the daemons - the user really wants to just
* look at the proposed process map
*/
if (orte_do_not_launch) {
/* set the state to indicate the daemons reported - this
* will trigger the daemons_reported event and cause the
* job to move to the following step
*/
state->jdata->state = ORTE_JOB_STATE_DAEMONS_LAUNCHED;
ORTE_ACTIVATE_JOB_STATE(state->jdata, ORTE_JOB_STATE_DAEMONS_REPORTED);
OBJ_RELEASE(state);
return;
}
/* Get the map for this job */
if (NULL == (map = daemons->map)) {
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
rc = ORTE_ERR_NOT_FOUND;
goto cleanup;
}
if (0 == map->num_new_daemons) {
/* set the state to indicate the daemons reported - this
* will trigger the daemons_reported event and cause the
* job to move to the following step
*/
OPAL_OUTPUT_VERBOSE((1, orte_plm_base_framework.framework_output,
"%s plm:slurm: no new daemons to launch",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
state->jdata->state = ORTE_JOB_STATE_DAEMONS_LAUNCHED;
ORTE_ACTIVATE_JOB_STATE(state->jdata, ORTE_JOB_STATE_DAEMONS_REPORTED);
OBJ_RELEASE(state);
return;
}
/* need integer value for command line parameter */
asprintf(&jobid_string, "%lu", (unsigned long) daemons->jobid);
/*
* start building argv array
*/
argv = NULL;
argc = 0;
/*
* SLURM srun OPTIONS
*/
/* add the srun command */
opal_argv_append(&argc, &argv, "srun");
/* start one orted on each node */
opal_argv_append(&argc, &argv, "--ntasks-per-node=1");
if (!orte_enable_recovery) {
/* kill the job if any orteds die */
opal_argv_append(&argc, &argv, "--kill-on-bad-exit");
}
/* ensure the orteds are not bound to a single processor,
* just in case the TaskAffinity option is set by default.
* This will *not* release the orteds from any cpu-set
* constraint, but will ensure it doesn't get
* bound to only one processor
*/
opal_argv_append(&argc, &argv, "--cpu_bind=none");
#if SLURM_CRAY_ENV
/*
* If in a SLURM/Cray env. make sure that Cray PMI is not pulled in,
* neither as a constructor run when orteds start, nor selected
* when pmix components are registered
*/
opal_setenv("PMI_NO_PREINITIALIZE", "1", false, &orte_launch_environ);
opal_setenv("PMI_NO_FORK", "1", false, &orte_launch_environ);
opal_setenv("OMPI_NO_USE_CRAY_PMI", "1", false, &orte_launch_environ);
#endif
/* Append user defined arguments to srun */
if ( NULL != mca_plm_slurm_component.custom_args ) {
custom_strings = opal_argv_split(mca_plm_slurm_component.custom_args, ' ');
num_args = opal_argv_count(custom_strings);
for (i = 0; i < num_args; ++i) {
opal_argv_append(&argc, &argv, custom_strings[i]);
}
opal_argv_free(custom_strings);
}
/* create nodelist */
nodelist_argv = NULL;
for (n=0; n < map->nodes->size; n++ ) {
if (NULL == (node = (orte_node_t*)opal_pointer_array_get_item(map->nodes, n))) {
continue;
}
/* if the daemon already exists on this node, then
* don't include it
*/
if (ORTE_FLAG_TEST(node, ORTE_NODE_FLAG_DAEMON_LAUNCHED)) {
continue;
}
/* otherwise, add it to the list of nodes upon which
* we need to launch a daemon
*/
opal_argv_append_nosize(&nodelist_argv, node->name);
}
if (0 == opal_argv_count(nodelist_argv)) {
orte_show_help("help-plm-slurm.txt", "no-hosts-in-list", true);
rc = ORTE_ERR_FAILED_TO_START;
goto cleanup;
}
nodelist_flat = opal_argv_join(nodelist_argv, ',');
opal_argv_free(nodelist_argv);
/* if we are using all allocated nodes, then srun doesn't
* require any further arguments
*/
if (map->num_new_daemons < orte_num_allocated_nodes) {
asprintf(&tmp, "--nodes=%lu", (unsigned long)map->num_new_daemons);
opal_argv_append(&argc, &argv, tmp);
free(tmp);
asprintf(&tmp, "--nodelist=%s", nodelist_flat);
opal_argv_append(&argc, &argv, tmp);
free(tmp);
}
/* tell srun how many tasks to run */
asprintf(&tmp, "--ntasks=%lu", (unsigned long)map->num_new_daemons);
opal_argv_append(&argc, &argv, tmp);
free(tmp);
OPAL_OUTPUT_VERBOSE((2, orte_plm_base_framework.framework_output,
"%s plm:slurm: launching on nodes %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), nodelist_flat));
free(nodelist_flat);
/*
* ORTED OPTIONS
*/
/* add the daemon command (as specified by user) */
orte_plm_base_setup_orted_cmd(&argc, &argv);
/* Add basic orted command line options, including debug flags */
orte_plm_base_orted_append_basic_args(&argc, &argv,
"slurm", &proc_vpid_index);
/* tell the new daemons the base of the name list so they can compute
* their own name on the other end
*/
rc = orte_util_convert_vpid_to_string(&name_string, map->daemon_vpid_start);
if (ORTE_SUCCESS != rc) {
opal_output(0, "plm_slurm: unable to get daemon vpid as string");
goto cleanup;
}
free(argv[proc_vpid_index]);
argv[proc_vpid_index] = strdup(name_string);
free(name_string);
/* Copy the prefix-directory specified in the
corresponding app_context. If there are multiple,
different prefix's in the app context, complain (i.e., only
allow one --prefix option for the entire slurm run -- we
don't support different --prefix'es for different nodes in
the SLURM plm) */
cur_prefix = NULL;
for (n=0; n < state->jdata->apps->size; n++) {
char * app_prefix_dir;
if (NULL == (app = (orte_app_context_t*)opal_pointer_array_get_item(state->jdata->apps, n))) {
continue;
}
app_prefix_dir = NULL;
orte_get_attribute(&app->attributes, ORTE_APP_PREFIX_DIR, (void**)&app_prefix_dir, OPAL_STRING);
/* Check for already set cur_prefix_dir -- if different,
complain */
if (NULL != app_prefix_dir) {
if (NULL != cur_prefix &&
0 != strcmp (cur_prefix, app_prefix_dir)) {
orte_show_help("help-plm-slurm.txt", "multiple-prefixes",
true, cur_prefix, app_prefix_dir);
goto cleanup;
}
/* If not yet set, copy it; iff set, then it's the
* same anyway
*/
if (NULL == cur_prefix) {
cur_prefix = strdup(app_prefix_dir);
OPAL_OUTPUT_VERBOSE((1, orte_plm_base_framework.framework_output,
"%s plm:slurm: Set prefix:%s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
cur_prefix));
}
free(app_prefix_dir);
}
}
/* protect the args in case someone has a script wrapper around srun */
mca_base_cmd_line_wrap_args(argv);
/* setup environment */
env = opal_argv_copy(orte_launch_environ);
if (0 < opal_output_get_verbosity(orte_plm_base_framework.framework_output)) {
param = opal_argv_join(argv, ' ');
OPAL_OUTPUT_VERBOSE((1, orte_plm_base_framework.framework_output,
"%s plm:slurm: final top-level argv:\n\t%s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
(NULL == param) ? "NULL" : param));
if (NULL != param) free(param);
}
/* exec the daemon(s) */
if (ORTE_SUCCESS != (rc = plm_slurm_start_proc(argc, argv, env, cur_prefix))) {
ORTE_ERROR_LOG(rc);
goto cleanup;
}
/* indicate that the daemons for this job were launched */
state->jdata->state = ORTE_JOB_STATE_DAEMONS_LAUNCHED;
daemons->state = ORTE_JOB_STATE_DAEMONS_LAUNCHED;
/* flag that launch was successful, so far as we currently know */
failed_launch = false;
cleanup:
if (NULL != argv) {
opal_argv_free(argv);
}
if (NULL != env) {
opal_argv_free(env);
}
if(NULL != jobid_string) {
free(jobid_string);
}
/* cleanup the caddy */
OBJ_RELEASE(state);
/* check for failed launch - if so, force terminate */
if (failed_launch) {
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
}
}
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;
orte_proc_state_t state;
orte_job_t *jdata;
orte_proc_t *pdata, *pptr;
opal_buffer_t *alert;
int rc, i;
orte_plm_cmd_flag_t cmd;
orte_std_cntr_t index;
orte_job_map_t *map;
orte_node_t *node;
orte_process_name_t target;
ORTE_ACQUIRE_OBJECT(caddy);
proc = &caddy->name;
state = caddy->proc_state;
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++;
if (jdata->num_launched == jdata->num_local_procs) {
/* tell the state machine that all local procs for this job
* were launched so that it can do whatever it needs to do,
* like send a state update message for all procs to the HNP
*/
ORTE_ACTIVATE_JOB_STATE(jdata, ORTE_JOB_STATE_LOCAL_LAUNCH_COMPLETE);
}
/* 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_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)
*/
if (orte_orteds_term_ordered &&
0 == orte_routed.num_routes()) {
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_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;
}
if (!ORTE_FLAG_TEST(pptr, ORTE_PROC_FLAG_TOOL)) {
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;
}
/* if requested, check fd status for leaks */
if (orte_state_base_run_fdcheck) {
orte_state_base_check_fds(jdata);
}
/* if ompi-server is around, then notify it to purge
* any session-related info */
if (NULL != orte_data_server_uri) {
target.jobid = jdata->jobid;
target.vpid = ORTE_VPID_WILDCARD;
orte_state_base_notify_data_server(&target);
}
/* cleanup the job info */
opal_hash_table_set_value_uint32(orte_job_data, jdata->jobid, NULL);
OBJ_RELEASE(jdata);
}
}
cleanup:
OBJ_RELEASE(caddy);
}
static void track_jobs(int fd, short argc, void *cbdata)
{
orte_state_caddy_t *caddy = (orte_state_caddy_t*)cbdata;
opal_buffer_t *alert;
orte_plm_cmd_flag_t cmd;
int rc, i;
orte_proc_state_t running = ORTE_PROC_STATE_RUNNING;
orte_proc_t *child;
orte_vpid_t null=ORTE_VPID_INVALID;
ORTE_ACQUIRE_OBJECT(caddy);
if (ORTE_JOB_STATE_LOCAL_LAUNCH_COMPLETE == caddy->job_state) {
OPAL_OUTPUT_VERBOSE((5, orte_state_base_framework.framework_output,
"%s state:orted:track_jobs sending local launch complete for job %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_JOBID_PRINT(caddy->jdata->jobid)));
/* update the HNP with all proc states for this job */
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);
OBJ_RELEASE(alert);
goto cleanup;
}
/* pack the jobid */
if (ORTE_SUCCESS != (rc = opal_dss.pack(alert, &caddy->jdata->jobid, 1, ORTE_JOBID))) {
ORTE_ERROR_LOG(rc);
OBJ_RELEASE(alert);
goto cleanup;
}
for (i=0; i < orte_local_children->size; i++) {
if (NULL == (child = (orte_proc_t*)opal_pointer_array_get_item(orte_local_children, i))) {
continue;
}
/* if this child is part of the job... */
if (child->name.jobid == caddy->jdata->jobid) {
/* pack the child's vpid */
if (ORTE_SUCCESS != (rc = opal_dss.pack(alert, &(child->name.vpid), 1, ORTE_VPID))) {
ORTE_ERROR_LOG(rc);
OBJ_RELEASE(alert);
goto cleanup;
}
/* pack the pid */
if (ORTE_SUCCESS != (rc = opal_dss.pack(alert, &child->pid, 1, OPAL_PID))) {
ORTE_ERROR_LOG(rc);
OBJ_RELEASE(alert);
goto cleanup;
}
/* if this proc failed to start, then send that info */
if (ORTE_PROC_STATE_UNTERMINATED < child->state) {
if (ORTE_SUCCESS != (rc = opal_dss.pack(alert, &child->state, 1, ORTE_PROC_STATE))) {
ORTE_ERROR_LOG(rc);
OBJ_RELEASE(alert);
goto cleanup;
}
} else {
/* pack the RUNNING state to avoid any race conditions */
if (ORTE_SUCCESS != (rc = opal_dss.pack(alert, &running, 1, ORTE_PROC_STATE))) {
ORTE_ERROR_LOG(rc);
OBJ_RELEASE(alert);
goto cleanup;
}
}
/* pack its exit code */
if (ORTE_SUCCESS != (rc = opal_dss.pack(alert, &child->exit_code, 1, ORTE_EXIT_CODE))) {
ORTE_ERROR_LOG(rc);
OBJ_RELEASE(alert);
goto cleanup;
}
}
}
/* flag that this job is complete so the receiver can know */
if (ORTE_SUCCESS != (rc = opal_dss.pack(alert, &null, 1, ORTE_VPID))) {
ORTE_ERROR_LOG(rc);
OBJ_RELEASE(alert);
goto cleanup;
}
/* 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);
OBJ_RELEASE(alert);
}
}
cleanup:
OBJ_RELEASE(caddy);
}
void orte_iof_orted_read_handler(int fd, short event, void *cbdata)
{
orte_iof_read_event_t *rev = (orte_iof_read_event_t*)cbdata;
unsigned char data[ORTE_IOF_BASE_MSG_MAX];
opal_buffer_t *buf=NULL;
int rc;
int32_t numbytes;
orte_iof_proc_t *proct = (orte_iof_proc_t*)rev->proc;
ORTE_ACQUIRE_OBJECT(rev);
/* As we may use timer events, fd can be bogus (-1)
* use the right one here
*/
fd = rev->fd;
/* read up to the fragment size */
#if !defined(__WINDOWS__)
numbytes = read(fd, data, sizeof(data));
#else
{
DWORD readed;
HANDLE handle = (HANDLE)_get_osfhandle(fd);
ReadFile(handle, data, sizeof(data), &readed, NULL);
numbytes = (int)readed;
}
#endif /* !defined(__WINDOWS__) */
if (NULL == proct) {
/* nothing we can do */
ORTE_ERROR_LOG(ORTE_ERR_ADDRESSEE_UNKNOWN);
return;
}
OPAL_OUTPUT_VERBOSE((1, orte_iof_base_framework.framework_output,
"%s iof:orted:read handler read %d bytes from %s, fd %d",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
numbytes, ORTE_NAME_PRINT(&proct->name), fd));
if (numbytes <= 0) {
if (0 > numbytes) {
/* either we have a connection error or it was a non-blocking read */
if (EAGAIN == errno || EINTR == errno) {
/* non-blocking, retry */
ORTE_IOF_READ_ACTIVATE(rev);
return;
}
OPAL_OUTPUT_VERBOSE((1, orte_iof_base_framework.framework_output,
"%s iof:orted:read handler %s Error on connection:%d",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&proct->name), fd));
}
/* numbytes must have been zero, so go down and close the fd etc */
goto CLEAN_RETURN;
}
/* see if the user wanted the output directed to files */
if (NULL != rev->sink) {
/* output to the corresponding file */
orte_iof_base_write_output(&proct->name, rev->tag, data, numbytes, rev->sink->wev);
}
if (!proct->copy) {
/* re-add the event */
ORTE_IOF_READ_ACTIVATE(rev);
return;
}
/* prep the buffer */
buf = OBJ_NEW(opal_buffer_t);
/* pack the stream first - we do this so that flow control messages can
* consist solely of the tag
*/
if (ORTE_SUCCESS != (rc = opal_dss.pack(buf, &rev->tag, 1, ORTE_IOF_TAG))) {
ORTE_ERROR_LOG(rc);
goto CLEAN_RETURN;
}
/* pack name of process that gave us this data */
if (ORTE_SUCCESS != (rc = opal_dss.pack(buf, &proct->name, 1, ORTE_NAME))) {
ORTE_ERROR_LOG(rc);
goto CLEAN_RETURN;
}
/* pack the data - only pack the #bytes we read! */
if (ORTE_SUCCESS != (rc = opal_dss.pack(buf, &data, numbytes, OPAL_BYTE))) {
ORTE_ERROR_LOG(rc);
goto CLEAN_RETURN;
}
/* start non-blocking RML call to forward received data */
OPAL_OUTPUT_VERBOSE((1, orte_iof_base_framework.framework_output,
"%s iof:orted:read handler sending %d bytes to HNP",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), numbytes));
orte_rml.send_buffer_nb(orte_mgmt_conduit,
ORTE_PROC_MY_HNP, buf, ORTE_RML_TAG_IOF_HNP,
orte_rml_send_callback, NULL);
/* re-add the event */
ORTE_IOF_READ_ACTIVATE(rev);
return;
CLEAN_RETURN:
/* must be an error, or zero bytes were read indicating that the
* proc terminated this IOF channel - either way, release the
* corresponding event. This deletes the read event and closes
* the file descriptor */
if (rev->tag & ORTE_IOF_STDOUT) {
if( NULL != proct->revstdout ) {
orte_iof_base_static_dump_output(proct->revstdout);
OBJ_RELEASE(proct->revstdout);
}
} else if (rev->tag & ORTE_IOF_STDERR) {
if( NULL != proct->revstderr ) {
orte_iof_base_static_dump_output(proct->revstderr);
OBJ_RELEASE(proct->revstderr);
}
} else if (rev->tag & ORTE_IOF_STDDIAG) {
if( NULL != proct->revstddiag ) {
orte_iof_base_static_dump_output(proct->revstddiag);
OBJ_RELEASE(proct->revstddiag);
}
}
/* check to see if they are all done */
if (NULL == proct->revstdout &&
NULL == proct->revstderr &&
NULL == proct->revstddiag) {
/* this proc's iof is complete */
ORTE_ACTIVATE_PROC_STATE(&proct->name, ORTE_PROC_STATE_IOF_COMPLETE);
}
if (NULL != buf) {
OBJ_RELEASE(buf);
}
return;
}
