Add smart monitoring using a textcollector

Collects the smart data using smartctl and outputs them in the
textcollector dir. This expects smartd to be configured to regularly
self tests on a regular interval to detect if a disk is broken.

docs/monitoring.md

@@ -73,3 +73,7 @@ The Nginx access logs/systemd logs are indexed by loki. For non webserver hosts
 ### AUR monitoring
 
 Some fun statistics are scraped from aur.archlinux.org using `curl` and `hq` as there is no proper AUR prometheus endpoint as of yet. The statistics are the AUR packages and users and is retrieved every 5 minutes.
+
+### Smart
+
+TODO:

hosts

@@ -182,3 +182,10 @@ america.mirror.pkgbuild.com
 europe.mirror.pkgbuild.com
 repro2.pkgbuild.com
 runner1.archlinux.org
+
+[dedicated_servers]
+gemini.archlinux.org
+build.archlinux.org
+runner1.archlinux.org
+runner2.archlinux.org
+secure-runner1.archlinux.org

roles/prometheus_exporters/files/smart-textcollector.sh

+#!/bin/bash
+
+set -o errexit
+set -o nounset
+
+if (( $# != 1 )); then
+  echo "Missing textcollector directory argument"
+  exit 1
+fi
+
+TEXTFILE_COLLECTOR_DIR=${1}
+PROM_FILE=$TEXTFILE_COLLECTOR_DIR/smart.prom
+
+TMP_FILE=$PROM_FILE.$$
+[ -e $TMP_FILE ] && rm -f $TMP_FILE
+
+trap "rm -f $TMP_FILE" EXIT
+
+# Metric types
+echo "# HELP smart_device_smart_healthy SMART metric device_smart_healthy" >> $TMP_FILE
+echo "# TYPE smart_device_smart_healthy gauge" >> $TMP_FILE
+
+echo "# HELP smart_temperature_celsius SMART metric temperature_celsius" >> $TMP_FILE
+echo "# TYPE smart_temperature_celsius gauge" >> $TMP_FILE
+
+echo "# HELP smart_device_info Device information, family/model name" >> $TMP_FILE
+echo "# TYPE smart_device_info gauge" >> $TMP_FILE
+
+echo "# HELP smart_device_self_test Self test status" >> $TMP_FILE
+echo "# TYPE smart_device_self_test gauge" >> $TMP_FILE
+
+devices="$(smartctl --scan-open --json)"
+devices_total="$(echo $devices | jq '.devices | length')"
+
+for ((i=0; i < $devices_total; i++)); do
+	disk=$(echo $devices | jq -r ".devices[${i}].name")
+	type=$(echo $devices | jq -r ".devices[${i}].type")
+	info=$(smartctl -a --json $disk)
+
+	status=$(echo $info | jq '.smart_status.passed')
+	if [[ "$status" == "true" ]]; then
+		echo "smart_device_smart_healthy{disk=\"${disk}\"} 1" >> $TMP_FILE
+	else
+		echo "smart_device_smart_healthy{disk=\"${disk}\"} 0" >> $TMP_FILE
+	fi
+
+	status=$(echo $info | jq '.ata_smart_data.self_test.status.passed')
+	if [[ "$status" == "true" ]]; then
+		echo "smart_device_self_test{disk=\"${disk}\"} 1" >> $TMP_FILE
+	else
+		echo "smart_device_self_test{disk=\"${disk}\"} 0" >> $TMP_FILE
+	fi
+
+	echo "smart_temperature_celsius{disk=\"${disk}\"} $(echo $info | jq '.temperature.current')" >> $TMP_FILE
+
+	# disk information
+        model_family=$(echo $info | jq -r '.model_family')
+        model_name=$(echo $info | jq -r '.model_name')
+        serial_number=$(echo $info | jq -r '.serial_number')
+	echo "smart_device_info{disk=\"$disk\",type=\"$type\",model_family=\"$model_family\",model_name=\"$model_name\",serial_number=\"$serial_number\"} 1" >> $TMP_FILE
+done
+
+mv -f $TMP_FILE $PROM_FILE

roles/prometheus_exporters/tasks/main.yml

@@ -7,6 +7,10 @@
   pacman: name=prometheus-blackbox-exporter state=present
   when: "'prometheus' in group_names"
 
+- name: install smartmontools for dedicated servers
+  pacman: name=smartmontools state=present
+  when: "'dedicated_servers' in group_names"
+
 - name: install prometheus-memcached-exporter
   pacman: name=prometheus-memcached-exporter state=present
   when: "'memcached' in group_names"
@@ -61,6 +65,7 @@
     - btrfs-textcollector.sh
     - aur-textcollector.sh
     - fail2ban-textcollector.sh
+    - smart-textcollector.sh
 
 - name: install arch textcollector service
   template: src=prometheus-arch-textcollector.service.j2 dest=/etc/systemd/system/prometheus-arch-textcollector.service owner=root group=root mode=644
@@ -85,6 +90,18 @@
     - { name: borg-offsite, service: borg-backup-offsite }
   when: "'borg_clients' in group_names"
 
+- name: install smart textcollector service
+  template: src=prometheus-smart-textcollector.service.j2 dest=/etc/systemd/system/prometheus-smart-textcollector.service owner=root group=root mode=644
+  when: "'dedicated_servers' in group_names"
+
+- name: install smart textcollector timer
+  template: src=prometheus-smart-textcollector.timer.j2 dest=/etc/systemd/system/prometheus-smart-textcollector.timer owner=root group=root mode=644
+  when: "'dedicated_servers' in group_names"
+
+- name: enable and start prometheus smart textcollector timer
+  systemd: name=prometheus-smart-textcollector.timer enabled=yes daemon_reload=yes state=started
+  when: "'dedicated_servers' in group_names"
+
 - name: install hetzner textcollector service
   template: src=prometheus-hetzner-textcollector.service.j2 dest=/etc/systemd/system/prometheus-hetzner-textcollector.service owner=root group=root mode=644
   when: "inventory_hostname == 'monitoring.archlinux.org'"

roles/prometheus_exporters/templates/prometheus-smart-textcollector.service.j2

+[Unit]
+Description=Prometheus Smart Exporter TextCollector
+
+[Service]
+Type=oneshot
+ExecStart=/usr/local/bin/smart-textcollector.sh {{ prometheus_textfile_dir }}
+
+NoNewPrivileges=true
+LockPersonality=true
+
+PrivateTmp=true
+ProtectSystem=strict
+ProtectHome=read-only
+ReadWritePaths={{ prometheus_textfile_dir }}
+
+MemoryDenyWriteExecute=true
+RemoveIPC=true
+RestrictRealtime=true
+RestrictNamespaces=true
+RestrictSUIDSGID=true
+
+PrivateNetwork=true
+ProtectHostname=true
+ProtectControlGroups=true
+ProtectKernelLogs=true
+ProtectKernelTunables=true
+ProtectKernelModules=true
+#ProtectClock=true
+
+SystemCallArchitectures=native

roles/prometheus_exporters/templates/prometheus-smart-textcollector.timer.j2

+[Unit]
+Description=Prometheus Smart Exporter TextCollector Timer
+
+[Timer]
+OnUnitActiveSec=1h
+OnBootSec=15min
+RandomizedDelaySec=1min
+
+[Install]
+WantedBy=timers.target