static void _signal_while_allocating(int signo)
{
allocation_interrupted = true;
if (pending_job_id != 0) {
slurm_complete_job(pending_job_id, 0);
}
}
static void _signal_while_allocating(int signo)
{
destroy_job = 1;
if (pending_job_id != 0) {
slurm_complete_job(pending_job_id, NO_VAL);
info("Job allocation %u has been revoked.", pending_job_id);
}
}
static void _signal_while_allocating(int signo)
{
debug("Got signal %d", signo);
if (signo == SIGCONT)
return;
destroy_job = 1;
if (pending_job_id != 0) {
slurm_complete_job(pending_job_id, NO_VAL);
info("Job allocation %u has been revoked.", pending_job_id);
}
}
static void *_safe_signal_while_allocating(void *in_data)
{
int signo = *(int *)in_data;
debug("Got signal %d", signo);
if (signo == SIGCONT)
return NULL;
destroy_job = 1;
if (pending_job_id != 0) {
slurm_complete_job(pending_job_id, NO_VAL);
info("Job allocation %u has been revoked.", pending_job_id);
}
return NULL;
}
/*
* slurm_allocate_resources_blocking
* allocate resources for a job request. This call will block until
* the allocation is granted, or the specified timeout limit is reached.
* IN req - description of resource allocation request
* IN timeout - amount of time, in seconds, to wait for a response before
* giving up.
* A timeout of zero will wait indefinitely.
* IN pending_callback - If the allocation cannot be granted immediately,
* the controller will put the job in the PENDING state. If
* pending callback is not NULL, it will be called with the job_id
* of the pending job as the sole parameter.
*
* RET allocation structure on success, NULL on error set errno to
* indicate the error (errno will be ETIMEDOUT if the timeout is reached
* with no allocation granted)
* NOTE: free the response using slurm_free_resource_allocation_response_msg()
*/
resource_allocation_response_msg_t *
slurm_allocate_resources_blocking (const job_desc_msg_t *user_req,
time_t timeout,
void(*pending_callback)(uint32_t job_id))
{
int rc;
slurm_msg_t req_msg;
slurm_msg_t resp_msg;
resource_allocation_response_msg_t *resp = NULL;
char *hostname = NULL;
uint32_t job_id;
job_desc_msg_t *req;
listen_t *listen = NULL;
int errnum = SLURM_SUCCESS;
slurm_msg_t_init(&req_msg);
slurm_msg_t_init(&resp_msg);
/* make a copy of the user's job description struct so that we
* can make changes before contacting the controller */
req = (job_desc_msg_t *)xmalloc(sizeof(job_desc_msg_t));
if (req == NULL)
return NULL;
memcpy(req, user_req, sizeof(job_desc_msg_t));
/*
* set Node and session id for this request
*/
if (req->alloc_sid == NO_VAL)
req->alloc_sid = getsid(0);
if (user_req->alloc_node != NULL) {
req->alloc_node = xstrdup(user_req->alloc_node);
} else if ((hostname = xshort_hostname()) != NULL) {
req->alloc_node = hostname;
} else {
error("Could not get local hostname,"
" forcing immediate allocation mode.");
req->immediate = 1;
}
if (!req->immediate) {
listen = _create_allocation_response_socket(hostname);
if (listen == NULL) {
xfree(req);
return NULL;
}
req->alloc_resp_port = listen->port;
}
req_msg.msg_type = REQUEST_RESOURCE_ALLOCATION;
req_msg.data = req;
rc = slurm_send_recv_controller_msg(&req_msg, &resp_msg);
if (rc == SLURM_SOCKET_ERROR) {
int errnum = errno;
destroy_forward(&req_msg.forward);
destroy_forward(&resp_msg.forward);
if (!req->immediate)
_destroy_allocation_response_socket(listen);
xfree(req);
errno = errnum;
return NULL;
}
switch (resp_msg.msg_type) {
case RESPONSE_SLURM_RC:
if (_handle_rc_msg(&resp_msg) < 0) {
/* will reach this when the allocation fails */
errnum = errno;
} else {
/* shouldn't get here */
errnum = -1;
}
break;
case RESPONSE_RESOURCE_ALLOCATION:
/* Yay, the controller has acknowledged our request! But did
we really get an allocation yet? */
resp = (resource_allocation_response_msg_t *) resp_msg.data;
if (resp->node_cnt > 0) {
/* yes, allocation has been granted */
errno = SLURM_PROTOCOL_SUCCESS;
} else if (!req->immediate) {
if (resp->error_code != SLURM_SUCCESS)
info("%s", slurm_strerror(resp->error_code));
/* no, we need to wait for a response */
job_id = resp->job_id;
slurm_free_resource_allocation_response_msg(resp);
if (pending_callback != NULL)
pending_callback(job_id);
resp = _wait_for_allocation_response(job_id, listen,
timeout);
/* If NULL, we didn't get the allocation in
the time desired, so just free the job id */
if ((resp == NULL) && (errno != ESLURM_ALREADY_DONE)) {
errnum = errno;
slurm_complete_job(job_id, -1);
}
}
break;
default:
errnum = SLURM_UNEXPECTED_MSG_ERROR;
resp = NULL;
}
destroy_forward(&req_msg.forward);
destroy_forward(&resp_msg.forward);
if (!req->immediate)
_destroy_allocation_response_socket(listen);
xfree(req);
errno = errnum;
return resp;
}
resource_allocation_response_msg_t *
allocate_nodes(void)
{
resource_allocation_response_msg_t *resp = NULL;
job_desc_msg_t *j = job_desc_msg_create_from_opts();
slurm_allocation_callbacks_t callbacks;
int i;
if (!j)
return NULL;
/* Do not re-use existing job id when submitting new job
* from within a running job */
if ((j->job_id != NO_VAL) && !opt.jobid_set) {
info("WARNING: Creating SLURM job allocation from within "
"another allocation");
info("WARNING: You are attempting to initiate a second job");
if (!opt.jobid_set) /* Let slurmctld set jobid */
j->job_id = NO_VAL;
}
callbacks.ping = _ping_handler;
callbacks.timeout = _timeout_handler;
callbacks.job_complete = _job_complete_handler;
callbacks.user_msg = _user_msg_handler;
callbacks.node_fail = _node_fail_handler;
/* create message thread to handle pings and such from slurmctld */
msg_thr = slurm_allocation_msg_thr_create(&j->other_port, &callbacks);
/* NOTE: Do not process signals in separate pthread. The signal will
* cause slurm_allocate_resources_blocking() to exit immediately. */
xsignal_unblock(sig_array);
for (i = 0; sig_array[i]; i++)
xsignal(sig_array[i], _signal_while_allocating);
while (!resp) {
resp = slurm_allocate_resources_blocking(j, opt.immediate,
_set_pending_job_id);
if (destroy_job) {
/* cancelled by signal */
break;
} else if (!resp && !_retry()) {
break;
}
}
if (resp && !destroy_job) {
/*
* Allocation granted!
*/
pending_job_id = resp->job_id;
#ifdef HAVE_BG
if (!_wait_bluegene_block_ready(resp)) {
if(!destroy_job)
error("Something is wrong with the "
"boot of the block.");
goto relinquish;
}
#else
if (!_wait_nodes_ready(resp)) {
if(!destroy_job)
error("Something is wrong with the "
"boot of the nodes.");
goto relinquish;
}
#endif
} else if (destroy_job) {
goto relinquish;
}
xsignal_block(sig_array);
job_desc_msg_destroy(j);
return resp;
relinquish:
slurm_free_resource_allocation_response_msg(resp);
if (!destroy_job)
slurm_complete_job(resp->job_id, 1);
exit(error_exit);
return NULL;
}
resource_allocation_response_msg_t *
allocate_nodes(bool handle_signals)
{
resource_allocation_response_msg_t *resp = NULL;
job_desc_msg_t *j = job_desc_msg_create_from_opts();
slurm_allocation_callbacks_t callbacks;
int i;
if (!j)
return NULL;
/* Do not re-use existing job id when submitting new job
* from within a running job */
if ((j->job_id != NO_VAL) && !opt.jobid_set) {
info("WARNING: Creating SLURM job allocation from within "
"another allocation");
info("WARNING: You are attempting to initiate a second job");
if (!opt.jobid_set) /* Let slurmctld set jobid */
j->job_id = NO_VAL;
}
callbacks.ping = _ping_handler;
callbacks.timeout = _timeout_handler;
callbacks.job_complete = _job_complete_handler;
callbacks.job_suspend = NULL;
callbacks.user_msg = _user_msg_handler;
callbacks.node_fail = _node_fail_handler;
/* create message thread to handle pings and such from slurmctld */
msg_thr = slurm_allocation_msg_thr_create(&j->other_port, &callbacks);
/* NOTE: Do not process signals in separate pthread. The signal will
* cause slurm_allocate_resources_blocking() to exit immediately. */
if (handle_signals) {
xsignal_unblock(sig_array);
for (i = 0; sig_array[i]; i++)
xsignal(sig_array[i], _signal_while_allocating);
}
while (!resp) {
resp = slurm_allocate_resources_blocking(j, opt.immediate,
_set_pending_job_id);
if (destroy_job) {
/* cancelled by signal */
break;
} else if (!resp && !_retry()) {
break;
}
}
if (resp && !destroy_job) {
/*
* Allocation granted!
*/
pending_job_id = resp->job_id;
/*
* These values could be changed while the job was
* pending so overwrite the request with what was
* allocated so we don't have issues when we use them
* in the step creation.
*/
if (opt.pn_min_memory != NO_VAL)
opt.pn_min_memory = (resp->pn_min_memory &
(~MEM_PER_CPU));
else if (opt.mem_per_cpu != NO_VAL)
opt.mem_per_cpu = (resp->pn_min_memory &
(~MEM_PER_CPU));
/*
* FIXME: timelimit should probably also be updated
* here since it could also change.
*/
#ifdef HAVE_BG
uint32_t node_cnt = 0;
select_g_select_jobinfo_get(resp->select_jobinfo,
SELECT_JOBDATA_NODE_CNT,
&node_cnt);
if ((node_cnt == 0) || (node_cnt == NO_VAL)) {
opt.min_nodes = node_cnt;
opt.max_nodes = node_cnt;
} /* else we just use the original request */
if (!_wait_bluegene_block_ready(resp)) {
if (!destroy_job)
error("Something is wrong with the "
"boot of the block.");
goto relinquish;
}
#else
opt.min_nodes = resp->node_cnt;
opt.max_nodes = resp->node_cnt;
if (!_wait_nodes_ready(resp)) {
if (!destroy_job)
error("Something is wrong with the "
"boot of the nodes.");
goto relinquish;
}
#endif
} else if (destroy_job) {
goto relinquish;
}
if (handle_signals)
xsignal_block(sig_array);
job_desc_msg_destroy(j);
return resp;
relinquish:
if (resp) {
if (!destroy_job)
slurm_complete_job(resp->job_id, 1);
slurm_free_resource_allocation_response_msg(resp);
}
exit(error_exit);
return NULL;
}
/*
* slurm_allocate_pack_job_blocking
* allocate resources for a list of job requests. This call will block
* until the entire allocation is granted, or the specified timeout limit
* is reached.
* IN req - List of resource allocation requests
* IN timeout - amount of time, in seconds, to wait for a response before
* giving up.
* A timeout of zero will wait indefinitely.
* IN pending_callback - If the allocation cannot be granted immediately,
* the controller will put the job in the PENDING state. If
* pending callback is not NULL, it will be called with the job_id
* of the pending job as the sole parameter.
*
* RET List of allocation structures on success, NULL on error set errno to
* indicate the error (errno will be ETIMEDOUT if the timeout is reached
* with no allocation granted)
* NOTE: free the response using list_destroy()
*/
List slurm_allocate_pack_job_blocking(List job_req_list, time_t timeout,
void(*pending_callback)(uint32_t job_id))
{
int rc;
slurm_msg_t req_msg;
slurm_msg_t resp_msg;
List resp = NULL;
char *local_hostname = NULL;
job_desc_msg_t *req;
listen_t *listen = NULL;
int errnum = SLURM_SUCCESS;
ListIterator iter;
bool immediate_flag = false;
bool immediate_logged = false;
uint32_t node_cnt = 0, job_id = 0;
slurm_msg_t_init(&req_msg);
slurm_msg_t_init(&resp_msg);
/*
* set node name and session ID for this request
*/
if (!immediate_flag) {
listen = _create_allocation_response_socket(local_hostname);
if (listen == NULL)
return NULL;
}
local_hostname = xshort_hostname();
iter = list_iterator_create(job_req_list);
while ((req = (job_desc_msg_t *) list_next(iter))) {
if (req->alloc_sid == NO_VAL)
req->alloc_sid = getsid(0);
if (listen)
req->alloc_resp_port = listen->port;
if (!req->alloc_node) {
if (local_hostname) {
req->alloc_node = local_hostname;
} else if (immediate_logged) {
req->immediate = 1;
} else {
req->immediate = 1;
error("Could not get local hostname, forcing "
"immediate allocation mode");
immediate_logged = true;
}
}
if (req->immediate)
immediate_flag = true;
}
list_iterator_destroy(iter);
req_msg.msg_type = REQUEST_JOB_PACK_ALLOCATION;
req_msg.data = job_req_list;
rc = slurm_send_recv_controller_msg(&req_msg, &resp_msg,
working_cluster_rec);
if (rc == SLURM_SOCKET_ERROR) {
int errnum = errno;
destroy_forward(&req_msg.forward);
destroy_forward(&resp_msg.forward);
if (listen)
_destroy_allocation_response_socket(listen);
iter = list_iterator_create(job_req_list);
while ((req = (job_desc_msg_t *)list_next(iter))) {
if (req->alloc_node == local_hostname)
req->alloc_node = NULL;
}
list_iterator_destroy(iter);
xfree(local_hostname);
errno = errnum;
return NULL;
}
switch (resp_msg.msg_type) {
case RESPONSE_SLURM_RC:
if (_handle_rc_msg(&resp_msg) < 0) {
/* will reach this when the allocation fails */
errnum = errno;
} else {
/* shouldn't get here */
errnum = SLURM_ERROR;
}
break;
case RESPONSE_JOB_PACK_ALLOCATION:
/* Yay, the controller has acknowledged our request!
* Test if we have an allocation yet? */
resp = (List) resp_msg.data;
_pack_alloc_test(resp, &node_cnt, &job_id);
if (node_cnt > 0) {
/* yes, allocation has been granted */
errno = SLURM_PROTOCOL_SUCCESS;
} else if (immediate_flag) {
debug("Immediate allocation not granted");
} else {
/* no, we need to wait for a response */
FREE_NULL_LIST(resp);
if (pending_callback != NULL)
pending_callback(job_id);
_wait_for_allocation_response(job_id, listen,
RESPONSE_JOB_PACK_ALLOCATION,
timeout, (void **) &resp);
/* If NULL, we didn't get the allocation in
* the time desired, so just free the job id */
if ((resp == NULL) && (errno != ESLURM_ALREADY_DONE)) {
errnum = errno;
slurm_complete_job(job_id, -1);
}
}
break;
default:
errnum = SLURM_UNEXPECTED_MSG_ERROR;
}
destroy_forward(&req_msg.forward);
destroy_forward(&resp_msg.forward);
if (listen)
_destroy_allocation_response_socket(listen);
iter = list_iterator_create(job_req_list);
while ((req = (job_desc_msg_t *)list_next(iter))) {
if (req->alloc_node == local_hostname)
req->alloc_node = NULL;
}
list_iterator_destroy(iter);
xfree(local_hostname);
errno = errnum;
return resp;
}
/*
* slurm_allocate_resources_blocking
* allocate resources for a job request. This call will block until
* the allocation is granted, or the specified timeout limit is reached.
* IN req - description of resource allocation request
* IN timeout - amount of time, in seconds, to wait for a response before
* giving up.
* A timeout of zero will wait indefinitely.
* IN pending_callback - If the allocation cannot be granted immediately,
* the controller will put the job in the PENDING state. If
* pending callback is not NULL, it will be called with the job_id
* of the pending job as the sole parameter.
*
* RET allocation structure on success, NULL on error set errno to
* indicate the error (errno will be ETIMEDOUT if the timeout is reached
* with no allocation granted)
* NOTE: free the response using slurm_free_resource_allocation_response_msg()
*/
resource_allocation_response_msg_t *
slurm_allocate_resources_blocking (const job_desc_msg_t *user_req,
time_t timeout,
void(*pending_callback)(uint32_t job_id))
{
int rc;
slurm_msg_t req_msg;
slurm_msg_t resp_msg;
resource_allocation_response_msg_t *resp = NULL;
uint32_t job_id;
job_desc_msg_t *req;
listen_t *listen = NULL;
int errnum = SLURM_SUCCESS;
bool already_done = false;
slurm_msg_t_init(&req_msg);
slurm_msg_t_init(&resp_msg);
/* make a copy of the user's job description struct so that we
* can make changes before contacting the controller */
req = (job_desc_msg_t *)xmalloc(sizeof(job_desc_msg_t));
if (req == NULL)
return NULL;
memcpy(req, user_req, sizeof(job_desc_msg_t));
/*
* set Node and session id for this request
*/
if (req->alloc_sid == NO_VAL)
req->alloc_sid = getsid(0);
if (!req->immediate) {
listen = _create_allocation_response_socket();
if (listen == NULL) {
xfree(req);
return NULL;
}
req->alloc_resp_port = listen->port;
}
req_msg.msg_type = REQUEST_RESOURCE_ALLOCATION;
req_msg.data = req;
rc = slurm_send_recv_controller_msg(&req_msg, &resp_msg,
working_cluster_rec);
if (rc == SLURM_ERROR) {
int errnum = errno;
destroy_forward(&req_msg.forward);
destroy_forward(&resp_msg.forward);
if (!req->immediate)
_destroy_allocation_response_socket(listen);
xfree(req);
errno = errnum;
return NULL;
}
switch (resp_msg.msg_type) {
case RESPONSE_SLURM_RC:
if (_handle_rc_msg(&resp_msg) < 0) {
/* will reach this when the allocation fails */
errnum = errno;
} else {
/* shouldn't get here */
errnum = SLURM_ERROR;
}
break;
case RESPONSE_RESOURCE_ALLOCATION:
/* Yay, the controller has acknowledged our request!
* Test if we have an allocation yet? */
resp = (resource_allocation_response_msg_t *) resp_msg.data;
if (resp->node_cnt > 0) {
/* yes, allocation has been granted */
errno = SLURM_SUCCESS;
} else if (!req->immediate) {
if (resp->error_code != SLURM_SUCCESS)
info("%s", slurm_strerror(resp->error_code));
/* no, we need to wait for a response */
/* print out any user messages before we wait. */
print_multi_line_string(resp->job_submit_user_msg,
-1, LOG_LEVEL_INFO);
job_id = resp->job_id;
slurm_free_resource_allocation_response_msg(resp);
if (pending_callback != NULL)
pending_callback(job_id);
_wait_for_allocation_response(job_id, listen,
RESPONSE_RESOURCE_ALLOCATION,
timeout, (void **) &resp);
/* If NULL, we didn't get the allocation in
the time desired, so just free the job id */
if ((resp == NULL) && (errno != ESLURM_ALREADY_DONE)) {
errnum = errno;
slurm_complete_job(job_id, -1);
}
if ((resp == NULL) && (errno == ESLURM_ALREADY_DONE))
already_done = true;
}
break;
default:
errnum = SLURM_UNEXPECTED_MSG_ERROR;
resp = NULL;
}
destroy_forward(&req_msg.forward);
destroy_forward(&resp_msg.forward);
if (!req->immediate)
_destroy_allocation_response_socket(listen);
xfree(req);
if (!resp && already_done && (errnum == SLURM_SUCCESS))
errnum = ESLURM_ALREADY_DONE;
errno = errnum;
return resp;
}
int main(int argc, char *argv[])
{
log_options_t logopt = LOG_OPTS_STDERR_ONLY;
job_desc_msg_t desc;
resource_allocation_response_msg_t *alloc;
time_t before, after;
allocation_msg_thread_t *msg_thr;
char **env = NULL;
int status = 0;
int retries = 0;
pid_t pid = getpid();
pid_t tpgid = 0;
pid_t rc_pid = 0;
int i, rc = 0;
static char *msg = "Slurm job queue full, sleeping and retrying.";
slurm_allocation_callbacks_t callbacks;
log_init(xbasename(argv[0]), logopt, 0, NULL);
_set_exit_code();
if (spank_init_allocator() < 0) {
error("Failed to initialize plugin stack");
exit(error_exit);
}
/* Be sure to call spank_fini when salloc exits
*/
if (atexit((void (*) (void)) spank_fini) < 0)
error("Failed to register atexit handler for plugins: %m");
if (initialize_and_process_args(argc, argv) < 0) {
error("salloc parameter parsing");
exit(error_exit);
}
/* reinit log with new verbosity (if changed by command line) */
if (opt.verbose || opt.quiet) {
logopt.stderr_level += opt.verbose;
logopt.stderr_level -= opt.quiet;
logopt.prefix_level = 1;
log_alter(logopt, 0, NULL);
}
if (spank_init_post_opt() < 0) {
error("Plugin stack post-option processing failed");
exit(error_exit);
}
_set_spank_env();
_set_submit_dir_env();
if (opt.cwd && chdir(opt.cwd)) {
error("chdir(%s): %m", opt.cwd);
exit(error_exit);
}
if (opt.get_user_env_time >= 0) {
char *user = uid_to_string(opt.uid);
if (strcmp(user, "nobody") == 0) {
error("Invalid user id %u: %m", (uint32_t)opt.uid);
exit(error_exit);
}
env = env_array_user_default(user,
opt.get_user_env_time,
opt.get_user_env_mode);
xfree(user);
if (env == NULL)
exit(error_exit); /* error already logged */
_set_rlimits(env);
}
/*
* Job control for interactive salloc sessions: only if ...
*
* a) input is from a terminal (stdin has valid termios attributes),
* b) controlling terminal exists (non-negative tpgid),
* c) salloc is not run in allocation-only (--no-shell) mode,
* d) salloc runs in its own process group (true in interactive
* shells that support job control),
* e) salloc has been configured at compile-time to support background
* execution and is not currently in the background process group.
*/
if (tcgetattr(STDIN_FILENO, &saved_tty_attributes) < 0) {
/*
* Test existence of controlling terminal (tpgid > 0)
* after first making sure stdin is not redirected.
*/
} else if ((tpgid = tcgetpgrp(STDIN_FILENO)) < 0) {
if (!opt.no_shell) {
error("no controlling terminal: please set --no-shell");
exit(error_exit);
}
} else if ((!opt.no_shell) && (pid == getpgrp())) {
if (tpgid == pid)
is_interactive = true;
#ifdef SALLOC_RUN_FOREGROUND
while (tcgetpgrp(STDIN_FILENO) != pid) {
if (!is_interactive) {
error("Waiting for program to be placed in "
"the foreground");
is_interactive = true;
}
killpg(pid, SIGTTIN);
}
#endif
}
/*
* Reset saved tty attributes at exit, in case a child
* process died before properly resetting terminal.
*/
if (is_interactive)
atexit (_reset_input_mode);
/*
* Request a job allocation
*/
slurm_init_job_desc_msg(&desc);
if (_fill_job_desc_from_opts(&desc) == -1) {
exit(error_exit);
}
if (opt.gid != (gid_t) -1) {
if (setgid(opt.gid) < 0) {
error("setgid: %m");
exit(error_exit);
}
}
callbacks.ping = _ping_handler;
callbacks.timeout = _timeout_handler;
callbacks.job_complete = _job_complete_handler;
callbacks.user_msg = _user_msg_handler;
callbacks.node_fail = _node_fail_handler;
/* create message thread to handle pings and such from slurmctld */
msg_thr = slurm_allocation_msg_thr_create(&desc.other_port,
&callbacks);
/* NOTE: Do not process signals in separate pthread. The signal will
* cause slurm_allocate_resources_blocking() to exit immediately. */
for (i = 0; sig_array[i]; i++)
xsignal(sig_array[i], _signal_while_allocating);
before = time(NULL);
while ((alloc = slurm_allocate_resources_blocking(&desc, opt.immediate,
_pending_callback)) == NULL) {
if ((errno != ESLURM_ERROR_ON_DESC_TO_RECORD_COPY) ||
(retries >= MAX_RETRIES))
break;
if (retries == 0)
error("%s", msg);
else
debug("%s", msg);
sleep (++retries);
}
/* become the user after the allocation has been requested. */
if (opt.uid != (uid_t) -1) {
if (setuid(opt.uid) < 0) {
error("setuid: %m");
exit(error_exit);
}
}
if (alloc == NULL) {
if (allocation_interrupted) {
/* cancelled by signal */
info("Job aborted due to signal");
} else if (errno == EINTR) {
error("Interrupted by signal."
" Allocation request rescinded.");
} else if (opt.immediate &&
((errno == ETIMEDOUT) ||
(errno == ESLURM_NOT_TOP_PRIORITY) ||
(errno == ESLURM_NODES_BUSY))) {
error("Unable to allocate resources: %m");
error_exit = immediate_exit;
} else {
error("Failed to allocate resources: %m");
}
slurm_allocation_msg_thr_destroy(msg_thr);
exit(error_exit);
} else if (!allocation_interrupted) {
/*
* Allocation granted!
*/
info("Granted job allocation %u", alloc->job_id);
pending_job_id = alloc->job_id;
#ifdef HAVE_BG
if (!_wait_bluegene_block_ready(alloc)) {
if(!allocation_interrupted)
error("Something is wrong with the "
"boot of the block.");
goto relinquish;
}
#else
if (!_wait_nodes_ready(alloc)) {
if(!allocation_interrupted)
error("Something is wrong with the "
"boot of the nodes.");
goto relinquish;
}
#endif
}
after = time(NULL);
if (opt.bell == BELL_ALWAYS
|| (opt.bell == BELL_AFTER_DELAY
&& ((after - before) > DEFAULT_BELL_DELAY))) {
_ring_terminal_bell();
}
if (opt.no_shell)
exit(0);
if (allocation_interrupted) {
/* salloc process received a signal after
* slurm_allocate_resources_blocking returned with the
* allocation, but before the new signal handlers were
* registered.
*/
goto relinquish;
}
/*
* Run the user's command.
*/
if (env_array_for_job(&env, alloc, &desc) != SLURM_SUCCESS)
goto relinquish;
/* Add default task count for srun, if not already set */
if (opt.ntasks_set) {
env_array_append_fmt(&env, "SLURM_NTASKS", "%d", opt.ntasks);
/* keep around for old scripts */
env_array_append_fmt(&env, "SLURM_NPROCS", "%d", opt.ntasks);
}
if (opt.cpus_per_task > 1) {
env_array_append_fmt(&env, "SLURM_CPUS_PER_TASK", "%d",
opt.cpus_per_task);
}
if (opt.overcommit) {
env_array_append_fmt(&env, "SLURM_OVERCOMMIT", "%d",
opt.overcommit);
}
if (opt.acctg_freq >= 0) {
env_array_append_fmt(&env, "SLURM_ACCTG_FREQ", "%d",
opt.acctg_freq);
}
if (opt.network)
env_array_append_fmt(&env, "SLURM_NETWORK", "%s", opt.network);
env_array_set_environment(env);
env_array_free(env);
pthread_mutex_lock(&allocation_state_lock);
if (allocation_state == REVOKED) {
error("Allocation was revoked for job %u before command could "
"be run", alloc->job_id);
pthread_mutex_unlock(&allocation_state_lock);
if (slurm_complete_job(alloc->job_id, status) != 0) {
error("Unable to clean up allocation for job %u: %m",
alloc->job_id);
}
return 1;
}
allocation_state = GRANTED;
pthread_mutex_unlock(&allocation_state_lock);
/* Ensure that salloc has initial terminal foreground control. */
if (is_interactive) {
/*
* Ignore remaining job-control signals (other than those in
* sig_array, which at this state act like SIG_IGN).
*/
xsignal(SIGTSTP, SIG_IGN);
xsignal(SIGTTIN, SIG_IGN);
xsignal(SIGTTOU, SIG_IGN);
pid = getpid();
setpgid(pid, pid);
tcsetpgrp(STDIN_FILENO, pid);
}
command_pid = _fork_command(command_argv);
/*
* Wait for command to exit, OR for waitpid to be interrupted by a
* signal. Either way, we are going to release the allocation next.
*/
if (command_pid > 0) {
setpgid(command_pid, command_pid);
if (is_interactive)
tcsetpgrp(STDIN_FILENO, command_pid);
/* NOTE: Do not process signals in separate pthread.
* The signal will cause waitpid() to exit immediately. */
xsignal(SIGHUP, _exit_on_signal);
/* Use WUNTRACED to treat stopped children like terminated ones */
do {
rc_pid = waitpid(command_pid, &status, WUNTRACED);
} while ((rc_pid == -1) && (!exit_flag));
if ((rc_pid == -1) && (errno != EINTR))
error("waitpid for %s failed: %m", command_argv[0]);
}
if (is_interactive)
tcsetpgrp(STDIN_FILENO, pid);
/*
* Relinquish the job allocation (if not already revoked).
*/
relinquish:
pthread_mutex_lock(&allocation_state_lock);
if (allocation_state != REVOKED) {
pthread_mutex_unlock(&allocation_state_lock);
info("Relinquishing job allocation %d", alloc->job_id);
if ((slurm_complete_job(alloc->job_id, status) != 0) &&
(slurm_get_errno() != ESLURM_ALREADY_DONE))
error("Unable to clean up job allocation %d: %m",
alloc->job_id);
pthread_mutex_lock(&allocation_state_lock);
allocation_state = REVOKED;
}
pthread_mutex_unlock(&allocation_state_lock);
slurm_free_resource_allocation_response_msg(alloc);
slurm_allocation_msg_thr_destroy(msg_thr);
/*
* Figure out what return code we should use. If the user's command
* exited normally, return the user's return code.
*/
rc = 1;
if (rc_pid != -1) {
if (WIFEXITED(status)) {
rc = WEXITSTATUS(status);
} else if (WIFSIGNALED(status)) {
verbose("Command \"%s\" was terminated by signal %d",
command_argv[0], WTERMSIG(status));
/* if we get these signals we return a normal
* exit since this was most likely sent from the
* user */
switch(WTERMSIG(status)) {
case SIGHUP:
case SIGINT:
case SIGQUIT:
case SIGKILL:
rc = 0;
break;
default:
break;
}
}
}
return rc;
}
/*
* Allocate nodes for heterogeneous/pack job from the slurm controller --
* retrying the attempt if the controller appears to be down, and optionally
* waiting for resources if none are currently available (see opt.immediate)
*
* Returns a pointer to a resource_allocation_response_msg which must
* be freed with slurm_free_resource_allocation_response_msg()
*/
List allocate_pack_nodes(bool handle_signals)
{
resource_allocation_response_msg_t *resp = NULL;
bool jobid_log = true;
job_desc_msg_t *j, *first_job = NULL;
slurm_allocation_callbacks_t callbacks;
ListIterator opt_iter, resp_iter;
slurm_opt_t *opt_local, *first_opt = NULL;
List job_req_list = NULL, job_resp_list = NULL;
uint32_t my_job_id = 0;
int i, k;
job_req_list = list_create(NULL);
opt_iter = list_iterator_create(opt_list);
while ((opt_local = list_next(opt_iter))) {
srun_opt_t *srun_opt = opt_local->srun_opt;
xassert(srun_opt);
if (!first_opt)
first_opt = opt_local;
if (srun_opt->relative_set && srun_opt->relative)
fatal("--relative option invalid for job allocation request");
if ((j = _job_desc_msg_create_from_opts(opt_local)) == NULL)
return NULL;
if (!first_job)
first_job = j;
j->origin_cluster = xstrdup(slurmctld_conf.cluster_name);
/* Do not re-use existing job id when submitting new job
* from within a running job */
if ((j->job_id != NO_VAL) && !opt_local->jobid_set) {
if (jobid_log) {
jobid_log = false; /* log once */
info("WARNING: Creating SLURM job allocation from within "
"another allocation");
info("WARNING: You are attempting to initiate a second job");
}
if (!opt_local->jobid_set) /* Let slurmctld set jobid */
j->job_id = NO_VAL;
}
list_append(job_req_list, j);
}
list_iterator_destroy(opt_iter);
if (!first_job) {
error("%s: No job requests found", __func__);
return NULL;
}
if (first_opt && first_opt->clusters &&
(slurmdb_get_first_pack_cluster(job_req_list, first_opt->clusters,
&working_cluster_rec)
!= SLURM_SUCCESS)) {
print_db_notok(first_opt->clusters, 0);
return NULL;
}
callbacks.ping = _ping_handler;
callbacks.timeout = _timeout_handler;
callbacks.job_complete = _job_complete_handler;
callbacks.job_suspend = NULL;
callbacks.user_msg = _user_msg_handler;
callbacks.node_fail = _node_fail_handler;
/* create message thread to handle pings and such from slurmctld */
msg_thr = slurm_allocation_msg_thr_create(&first_job->other_port,
&callbacks);
/* NOTE: Do not process signals in separate pthread. The signal will
* cause slurm_allocate_resources_blocking() to exit immediately. */
if (handle_signals) {
xsignal_unblock(sig_array);
for (i = 0; sig_array[i]; i++)
xsignal(sig_array[i], _signal_while_allocating);
}
while (first_opt && !job_resp_list) {
job_resp_list = slurm_allocate_pack_job_blocking(job_req_list,
first_opt->immediate, _set_pending_job_id);
if (destroy_job) {
/* cancelled by signal */
break;
} else if (!job_resp_list && !_retry()) {
break;
}
}
if (job_resp_list && !destroy_job) {
/*
* Allocation granted!
*/
opt_iter = list_iterator_create(opt_list);
resp_iter = list_iterator_create(job_resp_list);
while ((opt_local = list_next(opt_iter))) {
resp = (resource_allocation_response_msg_t *)
list_next(resp_iter);
if (!resp)
break;
if (pending_job_id == 0)
pending_job_id = resp->job_id;
if (my_job_id == 0) {
my_job_id = resp->job_id;
i = list_count(opt_list);
k = list_count(job_resp_list);
if (i != k) {
error("%s: request count != response count (%d != %d)",
__func__, i, k);
goto relinquish;
}
}
/*
* These values could be changed while the job was
* pending so overwrite the request with what was
* allocated so we don't have issues when we use them
* in the step creation.
*
* NOTE: pn_min_memory here is an int64, not uint64.
* These operations may have some bizarre side effects
*/
if (opt_local->pn_min_memory != NO_VAL64)
opt_local->pn_min_memory =
(resp->pn_min_memory & (~MEM_PER_CPU));
else if (opt_local->mem_per_cpu != NO_VAL64)
opt_local->mem_per_cpu =
(resp->pn_min_memory & (~MEM_PER_CPU));
#ifdef HAVE_BG
uint32_t node_cnt = 0;
select_g_select_jobinfo_get(resp->select_jobinfo,
SELECT_JOBDATA_NODE_CNT,
&node_cnt);
if ((node_cnt == 0) || (node_cnt == NO_VAL)) {
opt_local->min_nodes = node_cnt;
opt_local->max_nodes = node_cnt;
} /* else we just use the original request */
if (!_wait_bluegene_block_ready(resp)) {
if (!destroy_job)
error("Something is wrong with the "
"boot of the block.");
goto relinquish;
}
#else
opt_local->min_nodes = resp->node_cnt;
opt_local->max_nodes = resp->node_cnt;
if (resp->working_cluster_rec)
slurm_setup_remote_working_cluster(resp);
if (!_wait_nodes_ready(resp)) {
if (!destroy_job)
error("Something is wrong with the "
"boot of the nodes.");
goto relinquish;
}
#endif
}
list_iterator_destroy(resp_iter);
list_iterator_destroy(opt_iter);
} else if (destroy_job) {
goto relinquish;
}
if (handle_signals)
xsignal_block(sig_array);
return job_resp_list;
relinquish:
if (job_resp_list) {
if (!destroy_job && my_job_id)
slurm_complete_job(my_job_id, 1);
list_destroy(job_resp_list);
}
exit(error_exit);
return NULL;
}
/*
* Allocate nodes from the slurm controller -- retrying the attempt
* if the controller appears to be down, and optionally waiting for
* resources if none are currently available (see opt.immediate)
*
* Returns a pointer to a resource_allocation_response_msg which must
* be freed with slurm_free_resource_allocation_response_msg()
*/
extern resource_allocation_response_msg_t *
allocate_nodes(bool handle_signals, slurm_opt_t *opt_local)
{
srun_opt_t *srun_opt = opt_local->srun_opt;
resource_allocation_response_msg_t *resp = NULL;
job_desc_msg_t *j;
slurm_allocation_callbacks_t callbacks;
int i;
xassert(srun_opt);
if (srun_opt->relative_set && srun_opt->relative)
fatal("--relative option invalid for job allocation request");
if ((j = _job_desc_msg_create_from_opts(&opt)) == NULL)
return NULL;
if (opt_local->clusters &&
(slurmdb_get_first_avail_cluster(j, opt_local->clusters,
&working_cluster_rec)
!= SLURM_SUCCESS)) {
print_db_notok(opt_local->clusters, 0);
return NULL;
}
j->origin_cluster = xstrdup(slurmctld_conf.cluster_name);
/* Do not re-use existing job id when submitting new job
* from within a running job */
if ((j->job_id != NO_VAL) && !opt_local->jobid_set) {
info("WARNING: Creating SLURM job allocation from within "
"another allocation");
info("WARNING: You are attempting to initiate a second job");
if (!opt_local->jobid_set) /* Let slurmctld set jobid */
j->job_id = NO_VAL;
}
callbacks.ping = _ping_handler;
callbacks.timeout = _timeout_handler;
callbacks.job_complete = _job_complete_handler;
callbacks.job_suspend = NULL;
callbacks.user_msg = _user_msg_handler;
callbacks.node_fail = _node_fail_handler;
/* create message thread to handle pings and such from slurmctld */
msg_thr = slurm_allocation_msg_thr_create(&j->other_port, &callbacks);
/* NOTE: Do not process signals in separate pthread. The signal will
* cause slurm_allocate_resources_blocking() to exit immediately. */
if (handle_signals) {
xsignal_unblock(sig_array);
for (i = 0; sig_array[i]; i++)
xsignal(sig_array[i], _signal_while_allocating);
}
while (!resp) {
resp = slurm_allocate_resources_blocking(j,
opt_local->immediate,
_set_pending_job_id);
if (destroy_job) {
/* cancelled by signal */
break;
} else if (!resp && !_retry()) {
break;
}
}
if (resp)
print_multi_line_string(resp->job_submit_user_msg, -1);
if (resp && !destroy_job) {
/*
* Allocation granted!
*/
pending_job_id = resp->job_id;
/*
* These values could be changed while the job was
* pending so overwrite the request with what was
* allocated so we don't have issues when we use them
* in the step creation.
*/
opt_local->pn_min_memory = NO_VAL64;
opt_local->mem_per_cpu = NO_VAL64;
if (resp->pn_min_memory != NO_VAL64) {
if (resp->pn_min_memory & MEM_PER_CPU) {
opt_local->mem_per_cpu = (resp->pn_min_memory &
(~MEM_PER_CPU));
} else {
opt_local->pn_min_memory = resp->pn_min_memory;
}
}
#ifdef HAVE_BG
uint32_t node_cnt = 0;
select_g_select_jobinfo_get(resp->select_jobinfo,
SELECT_JOBDATA_NODE_CNT,
&node_cnt);
if ((node_cnt == 0) || (node_cnt == NO_VAL)) {
opt_local->min_nodes = node_cnt;
opt_local->max_nodes = node_cnt;
} /* else we just use the original request */
if (!_wait_bluegene_block_ready(resp)) {
if (!destroy_job)
error("Something is wrong with the "
"boot of the block.");
goto relinquish;
}
#else
opt_local->min_nodes = resp->node_cnt;
opt_local->max_nodes = resp->node_cnt;
if (resp->working_cluster_rec)
slurm_setup_remote_working_cluster(resp);
if (!_wait_nodes_ready(resp)) {
if (!destroy_job)
error("Something is wrong with the boot of the nodes.");
goto relinquish;
}
#endif
} else if (destroy_job) {
goto relinquish;
}
if (handle_signals)
xsignal_block(sig_array);
job_desc_msg_destroy(j);
return resp;
relinquish:
if (resp) {
if (!destroy_job)
slurm_complete_job(resp->job_id, 1);
slurm_free_resource_allocation_response_msg(resp);
}
exit(error_exit);
return NULL;
}
