From 917f8bb298c409e380fb90a621e3059d35a08827 Mon Sep 17 00:00:00 2001
From: Moe Jette <jette1@llnl.gov>
Date: Tue, 3 Jan 2006 17:01:03 +0000
Subject: [PATCH] Send REQUEST_SUSPEND RPC to slurmd rather than signal RPCs.
---
src/slurmctld/agent.c | 5 ++-
src/slurmctld/job_mgr.c | 55 ++++++++++++++++++++---
src/slurmd/slurmd/req.c | 96 ++++++++++++++++++++++++++++++++++++++++-
3 files changed, 148 insertions(+), 8 deletions(-)
diff --git a/src/slurmctld/agent.c b/src/slurmctld/agent.c
index 8fc618d77f7..12ca88d199d 100644
--- a/src/slurmctld/agent.c
+++ b/src/slurmctld/agent.c
@@ -4,9 +4,9 @@
*
* $Id$
*****************************************************************************
- * Copyright (C) 2002 The Regents of the University of California.
+ * Copyright (C) 2002-5 The Regents of the University of California.
* Produced at Lawrence Livermore National Laboratory (cf, DISCLAIMER).
- * Written by Morris Jette <jette@llnl.gov>, et. al.
+ * Written by Morris Jette <jette1@llnl.gov>, et. al.
* Derived from pdsh written by Jim Garlick <garlick1@llnl.gov>
* UCRL-CODE-2002-040.
*
@@ -313,6 +313,7 @@ static int _valid_agent_arg(agent_arg_t *agent_arg_ptr)
(agent_arg_ptr->msg_type == REQUEST_PING) ||
(agent_arg_ptr->msg_type == REQUEST_BATCH_JOB_LAUNCH) ||
(agent_arg_ptr->msg_type == REQUEST_SHUTDOWN) ||
+ (agent_arg_ptr->msg_type == REQUEST_SUSPEND) ||
(agent_arg_ptr->msg_type == REQUEST_RECONFIGURE) ||
(agent_arg_ptr->msg_type == RESPONSE_RESOURCE_ALLOCATION) ||
(agent_arg_ptr->msg_type == REQUEST_NODE_REGISTRATION_STATUS));
diff --git a/src/slurmctld/job_mgr.c b/src/slurmctld/job_mgr.c
index 0b1f9be6ab7..be9b1e9d305 100644
--- a/src/slurmctld/job_mgr.c
+++ b/src/slurmctld/job_mgr.c
@@ -5,7 +5,7 @@
*
* $Id$
*****************************************************************************
- * Copyright (C) 2002 The Regents of the University of California.
+ * Copyright (C) 2002-5 The Regents of the University of California.
* Produced at Lawrence Livermore National Laboratory (cf, DISCLAIMER).
* Written by Morris Jette <jette1@llnl.gov>
* UCRL-CODE-2002-040.
@@ -129,6 +129,7 @@ static void _set_job_id(struct job_record *job_ptr);
static void _set_job_prio(struct job_record *job_ptr);
static void _signal_batch_job(struct job_record *job_ptr, uint16_t signal);
static void _signal_job(struct job_record *job_ptr, int signal);
+static void _suspend_job(struct job_record *job_ptr, uint16_t op);
static int _suspend_job_nodes(struct job_record *job_ptr);
static bool _top_priority(struct job_record *job_ptr);
static int _validate_job_create_req(job_desc_msg_t * job_desc);
@@ -3934,6 +3935,52 @@ static void _signal_job(struct job_record *job_ptr, int signal)
return;
}
+/* Send suspend request to slumrd of all nodes associated with a job */
+static void _suspend_job(struct job_record *job_ptr, uint16_t op)
+{
+ agent_arg_t *agent_args;
+ suspend_msg_t *sus_ptr;
+ int i, buf_rec_size = 0;
+
+ agent_args = xmalloc(sizeof(agent_arg_t));
+ agent_args->msg_type = REQUEST_SUSPEND;
+ agent_args->retry = 1;
+ sus_ptr = xmalloc(sizeof(suspend_msg_t));
+ sus_ptr->job_id = job_ptr->job_id;
+ sus_ptr->op = op;
+
+ for (i = 0; i < node_record_count; i++) {
+ if (bit_test(job_ptr->node_bitmap, i) == 0)
+ continue;
+ if ((agent_args->node_count + 1) > buf_rec_size) {
+ buf_rec_size += 128;
+ xrealloc((agent_args->slurm_addr),
+ (sizeof(struct sockaddr_in) *
+ buf_rec_size));
+ xrealloc((agent_args->node_names),
+ (MAX_NAME_LEN * buf_rec_size));
+ }
+ agent_args->slurm_addr[agent_args->node_count] =
+ node_record_table_ptr[i].slurm_addr;
+ strncpy(&agent_args->
+ node_names[MAX_NAME_LEN * agent_args->node_count],
+ node_record_table_ptr[i].name, MAX_NAME_LEN);
+ agent_args->node_count++;
+#ifdef HAVE_FRONT_END /* Operate only on front-end */
+ break;
+#endif
+ }
+
+ if (agent_args->node_count == 0) {
+ xfree(sus_ptr);
+ xfree(agent_args);
+ return;
+ }
+
+ agent_args->msg_args = sus_ptr;
+ agent_queue_request(agent_args);
+ return;
+}
/* Specified job is being suspended, release allocated nodes */
static int _suspend_job_nodes(struct job_record *job_ptr)
{
@@ -4085,8 +4132,7 @@ extern int job_suspend(suspend_msg_t *sus_ptr, uid_t uid,
rc = _suspend_job_nodes(job_ptr);
if (rc != SLURM_SUCCESS)
goto reply;
- _signal_batch_job(job_ptr, SIGSTOP);
- _signal_job(job_ptr, SIGSTOP);
+ _suspend_job(job_ptr, sus_ptr->op);
job_ptr->job_state = JOB_SUSPENDED;
if (job_ptr->suspend_time) {
job_ptr->pre_sus_time +=
@@ -4105,8 +4151,7 @@ extern int job_suspend(suspend_msg_t *sus_ptr, uid_t uid,
rc = _resume_job_nodes(job_ptr);
if (rc != SLURM_SUCCESS)
goto reply;
- _signal_job(job_ptr, SIGCONT);
- _signal_batch_job(job_ptr, SIGCONT);
+ _suspend_job(job_ptr, sus_ptr->op);
job_ptr->job_state = JOB_RUNNING;
if (job_ptr->time_limit != INFINITE) {
/* adjust effective time_limit */
diff --git a/src/slurmd/slurmd/req.c b/src/slurmd/slurmd/req.c
index 420c0ca58a9..e6262afa6ae 100644
--- a/src/slurmd/slurmd/req.c
+++ b/src/slurmd/slurmd/req.c
@@ -2,7 +2,7 @@
* src/slurmd/slurmd/req.c - slurmd request handling
* $Id$
*****************************************************************************
- * Copyright (C) 2002 The Regents of the University of California.
+ * Copyright (C) 2002-5 The Regents of the University of California.
* Produced at Lawrence Livermore National Laboratory (cf, DISCLAIMER).
* Written by Mark Grondona <mgrondona@llnl.gov>.
* UCRL-CODE-2002-040.
@@ -79,6 +79,7 @@ static void _rpc_terminate_tasks(slurm_msg_t *, slurm_addr *);
static void _rpc_timelimit(slurm_msg_t *, slurm_addr *);
static void _rpc_reattach_tasks(slurm_msg_t *, slurm_addr *);
static void _rpc_signal_job(slurm_msg_t *, slurm_addr *);
+static void _rpc_suspend_job(slurm_msg_t *, slurm_addr *);
static void _rpc_terminate_job(slurm_msg_t *, slurm_addr *);
static void _rpc_update_time(slurm_msg_t *, slurm_addr *);
static void _rpc_shutdown(slurm_msg_t *msg, slurm_addr *cli_addr);
@@ -152,6 +153,10 @@ slurmd_req(slurm_msg_t *msg, slurm_addr *cli)
_rpc_signal_job(msg, cli);
slurm_free_signal_job_msg(msg->data);
break;
+ case REQUEST_SUSPEND:
+ _rpc_suspend_job(msg, cli);
+ slurm_free_suspend_msg(msg->data);
+ break;
case REQUEST_TERMINATE_JOB:
debug2("Processing RPC: REQUEST_TERMINATE_JOB");
_rpc_terminate_job(msg, cli);
@@ -1425,6 +1430,95 @@ _rpc_signal_job(slurm_msg_t *msg, slurm_addr *cli)
}
}
+/*
+ * Send a job suspend/resume request through the appropriate slurmstepds for
+ * each job step belonging to a given job allocation.
+ */
+static void
+_rpc_suspend_job(slurm_msg_t *msg, slurm_addr *cli)
+{
+ suspend_msg_t *req = msg->data;
+ uid_t req_uid = g_slurm_auth_get_uid(msg->cred);
+ uid_t job_uid;
+ List steps;
+ ListIterator i;
+ step_loc_t *stepd;
+ int step_cnt = 0;
+ int fd, rc = SLURM_SUCCESS, sig_num;
+ char *sig_name;
+
+ if (req->op == SUSPEND_JOB) {
+ sig_name = "STOP";
+ sig_num = SIGSTOP;
+ } else if (req->op == RESUME_JOB) {
+ sig_name = "CONT";
+ sig_num = SIGCONT;
+ } else {
+ error("REQUEST_SUSPEND: bad op code %u",
+ req->op);
+ rc = ESLURM_NOT_SUPPORTED;
+ goto fini;
+ }
+ debug("_rpc_suspend_job jobid=%u uid=%d signal=%s",
+ req->job_id, req_uid, sig_name);
+ job_uid = _get_job_uid(req->job_id);
+ /*
+ * check that requesting user ID is the SLURM UID
+ */
+ if (!_slurm_authorized_user(req_uid)) {
+ error("Security violation: signal_job(%u) from uid %ld",
+ req->job_id, (long) req_uid);
+ rc = ESLURM_USER_ID_MISSING;
+ goto fini;
+ }
+
+ /*
+ * Loop through all job steps for this job and signal the
+ * step's process group through the slurmstepd.
+ */
+ steps = stepd_available(conf->spooldir, conf->node_name);
+ i = list_iterator_create(steps);
+ while (stepd = list_next(i)) {
+ if (stepd->jobid != req->job_id) {
+ /* multiple jobs expected on shared nodes */
+ debug3("Step from other job: jobid=%u (this jobid=%u)",
+ stepd->jobid, req->job_id);
+ continue;
+ }
+ step_cnt++;
+
+ fd = stepd_connect(stepd->directory, stepd->nodename,
+ stepd->jobid, stepd->stepid);
+ if (fd == -1) {
+ debug3("Unable to connect to step %u.%u",
+ stepd->jobid, stepd->stepid);
+ continue;
+ }
+
+ debug2(" signal %d to job %u.%u",
+ sig_num, stepd->jobid, stepd->stepid);
+ if (stepd_signal(fd, sig_num) < 0)
+ debug("signal jobid=%u failed: %m", stepd->jobid);
+ close(fd);
+ }
+ list_iterator_destroy(i);
+ list_destroy(steps);
+ if (step_cnt == 0)
+ debug2("No steps in jobid %u to send signal %d",
+ req->job_id, sig_num);
+
+ /*
+ * At this point, if connection still open, we send controller
+ * a reply.
+ */
+ fini: if (msg->conn_fd >= 0) {
+ slurm_send_rc_msg(msg, rc);
+ if (slurm_close_accepted_conn(msg->conn_fd) < 0)
+ error ("_rpc_signal_job: close(%d): %m", msg->conn_fd);
+ msg->conn_fd = -1;
+ }
+}
+
static void
_rpc_terminate_job(slurm_msg_t *msg, slurm_addr *cli)
{
--
GitLab