#!/bin/bash
# Ref : OPSTEAM-15119 - find old drives with reallocated sectors and ata errors
# This helps prevent crashes, grub issues, fscks etc on older shared servers
# SYSENG-25357 - refactored to run smartctl only once and added OK status 
#

# If not physical server, exit
if ! [[ $(systemd-detect-virt) == "none" ]]; then
  exit 0
fi

# SSD drive variables - default values
drive_list=$(find /sys/block/ | grep -E 'sd|nvm' | awk -F 'block/' '{print $2}')
drive_age_min=39420 # in hours, 4.5 years in age minimum
drive_realloc_min=100 # reallocated sectors minimum
drive_ata_errors_min=20 # minimum ata errors

# NVMe drive variables - default values
nvme_age_min=43800 # in hours, at least aged 5 years, NVMes are built to last longer
nvme_percentage_used=90 # 'Percentage Used' attribute for NMVe
nvme_media_errors=500 # 'Media and Data Integrity' Errors attribute for NMVe

for drive in ${drive_list}; do
  drive_data=$(smartctl -a /dev/"${drive}")
  if [[ "${drive}" =~ "sd" ]]; then
    realloc_sectors=$(awk '/Reallocated_Sector/ {print $NF}' <<< "$drive_data")
    power_on_hours=$(awk '/Power_On_Hours/ {print $NF}' <<< "$drive_data")
    ata_errors=$(awk '/ATA Error/ {print $NF}' <<< "$drive_data")

    if [[ ${realloc_sectors} -gt ${drive_realloc_min} ]] && [[ ${power_on_hours} -gt ${drive_age_min} ]] && [[ ${ata_errors} -gt ${drive_ata_errors_min} ]]; then
      printf '%s\n' "check_old_drive_health_$drive - /dev/$drive age > 4.5y|reallocated sectors: ${realloc_sectors}|ata errors: ${ata_errors}"
      exit 2
    else
      printf '%s\n' "check_old_drive_health_$drive - status OK"
      exit 0
    fi
  else
    media_errors=$(awk '/Media and Data Integrity Errors/ {print $NF}' <<< "$drive_data")
    nmve_power_on_hours=$(awk '/Power On Hours/ {print $NF}' <<< "$drive_data")
    percentage_used=$(awk '/Percentage Used:/ {print $NF}' <<< "$drive_data")

    if [[ ${media_errors} -gt ${nvme_media_errors} ]] && [[ ${nmve_power_on_hours} -gt ${nvme_age_min} ]] && [[ ${percentage_used} -gt ${nvme_percentage_used} ]]; then
      printf '%s\n' "check_old_drive_health_$drive - /dev/$drive age > 5y|media errors: ${media_errors}|life used: ${percentage_used}"
      exit 2
    else
      printf '%s\n' "check_old_drive_health_$drive - status OK"
      exit 0
    fi
  fi
done