nagios/check_swaping: adding check on oom_kill + change of unit

In order to standardize units, we use bytes instead of page/pagesize.

nagios/check_swaping.sh

@@ -11,11 +11,13 @@
 
 
 # Default values
-THRESHOLD_WARNING="50"
+PAGESIZE=4096
-THRESHOLD_CRITICAL="200"
+THRESHOLD_WARNING="200000"
+THRESHOLD_CRITICAL="1000000"
+THRESHOLD_OOMKILL="0"
+DISABLE_OOMKILL=""
 VMSTAT_PREVIOUS_DATA_FILE="/tmp/.monitoring_vmstat.txt"
 
-
 # Output
 OUTPUT_EXIT_STATUS=0
 OUTPUT_DETAIL_WARNING=""
@@ -28,6 +30,9 @@ REVISION="0.1"
 # Stop at the first non-catched error
 set -e
 
+# Set/guess current page size
+which getconf >/dev/null 2>&1 && PAGESIZE="$( getconf PAGESIZE )"
+
 # Include check_range()
 . $PROGPATH/utils.sh
 # If you don't have the previous file, just comment the line and uncomment the following
@@ -45,17 +50,18 @@ set -e
 usage() {
 	cat <<EOF
 Usage :
-  $0 [-w warning_threshold] [-c critical_threshold] [-f vmstat_previous_data_file ]
+  $0 [-w warning_threshold] [-c critical_threshold] [-f vmstat_previous_data_file ] [ -k oom_kill_threshold ] [ -K ]
 
 Note 1 : the script will measure the number of seconds passed since its last call and will
-         divide the measures accordingly, so write the thresholds using pages/s in mind.
+         divide the measures accordingly, so write the thresholds using bytes/second in mind.
          The script will measure against pswpin and pswpout added together.
 
-Note 2 : the thresholds use the kernel page size as unit. Use 'getconf PAGESIZE' to get it if needed.
+Note 2 : if 'oom_kill' is not available in /proc/vmstat, you can disable the check with '-K'.
 
 Default values:
   warning_threshold         : $THRESHOLD_WARNING
   critical_threshold        : $THRESHOLD_CRITICAL
+  oom_kill_threshold        : $THRESHOLD_OOMKILL
   vmstat_previous_data_file : $VMSTAT_PREVIOUS_DATA_FILE
 EOF
 }
@@ -89,7 +95,7 @@ is_int() {
 #
 # Parameters management
 #
-while getopts hw:c:f: f; do
+while getopts hw:c:f:k:K f; do
 	case "$f" in
 		'h')
 			usage
@@ -108,6 +114,14 @@ while getopts hw:c:f: f; do
 			VMSTAT_PREVIOUS_DATA_FILE="$OPTARG"
 			;;
 
+		'k')
+			THRESHOLD_OOMKILL="$OPTARG"
+			;;
+
+		'K')
+			DISABLE_OOMKILL="1"
+			;;
+
 		\?)
 			usage
 			exit 1
@@ -117,10 +131,14 @@ done
 shift $( expr $OPTIND - 1 )
 
 # Little checks
-if ! is_int "$THRESHOLD_WARNING" || ! is_int "$THRESHOLD_CRITICAL"; then
+if ! is_int "$THRESHOLD_WARNING" || ! is_int "$THRESHOLD_CRITICAL" || ! is_int "$THRESHOLD_OOMKILL" ; then
 	echo "UNKNOWN invalid parameter : one of the threshold is not an integer."
 	exit $STATE_UNKNOWN
 fi
+if [ -z "$DISABLE_OOMKILL" ] && ! grep '^oom_kill ' /proc/vmstat >/dev/null 2>&1; then
+	echo "UNKNOWN threshold set for oom_kill but the counter not available in /proc/vmstat."
+	exit $STATE_UNKNOWN
+fi
 
 # Check if the previous data file exists
 if [ ! -w "$VMSTAT_PREVIOUS_DATA_FILE" ]; then
@@ -132,8 +150,8 @@ if [ ! -w "$VMSTAT_PREVIOUS_DATA_FILE" ]; then
 fi
 
 # Read previous data
-PREVIOUS_PSWPIN="$( grep '^pswpin' "$VMSTAT_PREVIOUS_DATA_FILE" | cut -d " " -f 2 )"
+PREVIOUS_PSWPIN="$(( $( grep '^pswpin' "$VMSTAT_PREVIOUS_DATA_FILE" | cut -d " " -f 2 ) * $PAGESIZE ))"
-PREVIOUS_PSWPOUT="$( grep '^pswpout' "$VMSTAT_PREVIOUS_DATA_FILE" | cut -d " " -f 2 )"
+PREVIOUS_PSWPOUT="$(( $( grep '^pswpout' "$VMSTAT_PREVIOUS_DATA_FILE" | cut -d " " -f 2 ) * $PAGESIZE ))"
 PREVIOUS_PSWPTOTAL=$(( $PREVIOUS_PSWPIN + $PREVIOUS_PSWPOUT ))
 
 # Get time elapsed since last call
@@ -145,21 +163,28 @@ fi
 
 # Update and read current data
 update_vmstat_previous_data_file "$VMSTAT_PREVIOUS_DATA_FILE"
-CURRENT_PSWPIN="$( grep '^pswpin' "$VMSTAT_PREVIOUS_DATA_FILE" | cut -d " " -f 2 )"
+CURRENT_OOMKILL="$( grep '^oom_kill' "$VMSTAT_PREVIOUS_DATA_FILE" | cut -d " " -f 2 )"
-CURRENT_PSWPOUT="$( grep '^pswpout' "$VMSTAT_PREVIOUS_DATA_FILE" | cut -d " " -f 2 )"
+CURRENT_PSWPIN="$(( $( grep '^pswpin' "$VMSTAT_PREVIOUS_DATA_FILE" | cut -d " " -f 2 ) * $PAGESIZE ))"
+CURRENT_PSWPOUT="$(( $( grep '^pswpout' "$VMSTAT_PREVIOUS_DATA_FILE" | cut -d " " -f 2 ) * $PAGESIZE ))"
 CURRENT_PSWPTOTAL=$(( $CURRENT_PSWPIN + $CURRENT_PSWPOUT ))
 
 # Calculate the swaping rate
 PSWP_RATE=$(( ( $CURRENT_PSWPTOTAL - $PREVIOUS_PSWPTOTAL ) / $PREVIOUS_DATA_AGE ))
 
 # Generate perfdata
-OUTPUT_PERFDATA="$( printf " pswpin=%d pswpout=%d" "$CURRENT_PSWPIN" "$CURRENT_PSWPOUT" )"
+if [ -n "$CURRENT_OOMKILL" ]; then
+	OUTPUT_PERFDATA="$( printf " oomkill=%d" "$CURRENT_OOMKILL" )"
+fi
+OUTPUT_PERFDATA="$( printf "$OUTPUT_PERFDATA pswpin=%d pswpout=%d" "$CURRENT_PSWPIN" "$CURRENT_PSWPOUT" )"
 
 # Comparison
 # note: remember that numbers can be reseted to zero from time to time (64bits counter ?)
 if [ $PSWP_RATE -gt $THRESHOLD_CRITICAL ]; then
 	echo "CRITICAL swaping rate at $PSWP_RATE (limit at $THRESHOLD_CRITICAL) |$OUTPUT_PERFDATA"
 	exit $STATE_CRITICAL
+elif [ -z "$DISABLE_OOMKILL" ] && [ $CURRENT_OOMKILL -gt $THRESHOLD_OOMKILL ]; then
+	echo "CRITICAL $CURRENT_OOMKILL OOM kill(s) (limit at $THRESHOLD_OOMKILL) |$OUTPUT_PERFDATA"
+	exit $STATE_CRITICAL
 elif [ $PSWP_RATE -gt $THRESHOLD_WARNING ]; then
 	echo "WARNING swaping rate at $PSWP_RATE (limit at $THRESHOLD_WARNING) |$OUTPUT_PERFDATA"
 	exit $STATE_WARNING