If there is no SchedLogfile defined and 'scontrol schedloglevel 1'
is issued from an administrator, slurmctld will segfault at the
next "sched: " log message due to NULL log file pointer. There
are obviously multiple ways to fix this issue, but in this patch
the RPC simply returns and "Operation Disabled" error immediately
if the sched log file is NULL.
Other options include opening a new logfile with a default name,
sending sched log messages to stderr, or enhancing the scontrol
interface to allow specifying a logfile name for the schedlog.
There are other cases in the schedlog code that could cause problems
for the slurmctld, but since the sched log stuff is tied in strangely
with the rest of the logging code, I didn't want to try modifying
anything in log.c, for fear of breaking the normal logging functions.
---
src/slurmctld/proc_req.c | 13 +++++++++++++
1 files changed, 13 insertions(+), 0 deletions(-)
diff --git a/src/slurmctld/proc_req.c b/src/slurmctld/proc_req.c
index 916ab43..8c302ee 100644
--- a/src/slurmctld/proc_req.c
+++ b/src/slurmctld/proc_req.c
@@ -3987,6 +3987,19 @@ inline static void
_slurm_rpc_set_schedlog_level(slurm_msg_t *msg)
return;
}
+ /*
+ * If slurmctld_conf.sched_logfile is NULL, then this operation
+ * will fail, since there is no sched logfile for which to alter
+ * the log level. (Calling sched_log_alter with a NULL filename
+ * is likely to cause a segfault at the next sched log call)
+ * So just give up and return "Operation Disabled"
+ */
+ if (slurmctld_conf.sched_logfile == NULL) {
+ error("set scheduler log level failed: no log file!");
+ slurm_send_rc_msg (msg, ESLURM_DISABLED);
+ return;
+ }
+
schedlog_level = MIN (request_msg->debug_level, (LOG_LEVEL_QUIET + 1));
schedlog_level = MAX (schedlog_level, LOG_LEVEL_QUIET);
--
1.7.1
