#!/usr/bin/env bash
#
# JbTecWiz Support Centre -- generated fix script
#
# Fault : Load average is high but the CPU looks idle
# Fix   : Separate CPU load from I/O wait
# Source: https://jbtecwiz.com/support/lnx-high-load
#
# Run as     : Shell
# Expect     : 30 minutes
# Risk       : low
# Reversible : yes
#
# WHEN THIS IS THE RIGHT FIX
#   Load is high and CPU is not. The load average alone cannot tell you
#   which, so start by splitting them.
#
# HOW TO UNDO IT
#   None -- diagnostic.
#
# Walks the fix one step at a time and asks before each. Steps with no
# command are yours to do -- it prints those and waits. DRYRUN=1 prints
# without executing; UNATTENDED=1 does not ask.
#
# --------------------------------------------------------------------
# NO WARRANTY - USE AT YOUR OWN RISK
#
# This script is provided by JbTecWiz as-is and with no warranty of any
# kind, express or implied. You run it entirely at your own risk.
#
# JbTecWiz accepts no liability for any loss or damage arising from its
# use, including but not limited to data loss, downtime, or configuration
# changes that turn out to be wrong for your system.
#
# You are responsible for reading this script before running it, for
# satisfying yourself that it suits the machine in front of you, and for
# having a working backup first. Some steps cannot be undone.
# --------------------------------------------------------------------

set -uo pipefail
DRYRUN="${DRYRUN:-0}"
UNATTENDED="${UNATTENDED:-0}"
failed=0

rule() { printf "\n%s\n" "$(printf '-%.0s' $(seq 1 70))"; if [ $# -gt 0 ]; then echo "$1"; fi; }
prose() { echo "$1" | fold -s -w 74 | sed "s/^/  /"; }

# Returns 0 when the caller should run the command, 1 when it should not.
# A manual step always returns 1 -- there is nothing for the caller to run.
step() {  # step <n> <do> <why> <manual|cmd> [command lines...]
  local n="$1" dotext="$2" why="$3" mode="$4"; shift 4
  rule "  Step $n of 6"
  prose "$dotext"
  if [ -n "$why" ]; then echo; prose "$why"; fi
  if [ "$mode" = "manual" ]; then
    echo; echo "  -> Do this yourself, then press Enter to carry on."
    if [ "$UNATTENDED" = "0" ] && [ "$DRYRUN" = "0" ]; then read -r _; fi
    return 1
  fi
  echo; printf "    %s\n" "$@"; echo
  if [ "$DRYRUN" = "1" ]; then echo "  (dry run -- not executed)"; return 1; fi
  if [ "$UNATTENDED" = "0" ]; then
    read -r -p "  Run this step? [Y]es / [S]kip / [Q]uit " a
    case "$a" in
      [Qq]*) echo "  Stopped at your request."; exit 0 ;;
      [Ss]*) echo "  Skipped."; return 1 ;;
    esac
  fi
  return 0
}

rule
echo "  Load average is high but the CPU looks idle"
echo "  Separate CPU load from I/O wait"
echo
echo "  Risk: low   Reversible   30 minutes"
echo
prose 'No warranty. Use at your own risk - JbTecWiz accepts no liability. Read it before you run it, and have a backup.'
rule
echo
if [ "$UNATTENDED" = "0" ] && [ "$DRYRUN" = "0" ]; then
  read -r -p "  Ready? [y/N] " go
  case "$go" in [Yy]*) ;; *) echo "  Nothing was changed."; exit 0;; esac
fi

if step 1 'Look at the CPU breakdown. The %wa column is I/O wait.' 'In vmstat, the '\''b'\'' column counts processes blocked on I/O. If b is high and '\''r'\'' is low, no amount of CPU would help.' cmd 'top -bn2 -d1 | grep -E '\''^%Cpu'\'' | tail -1' 'vmstat 1 5'; then
  top -bn2 -d1 | grep -E '^%Cpu' | tail -1
  vmstat 1 5
  if [ $? -ne 0 ]; then
    failed=$((failed+1))
    echo "  Step 1 failed. The rest of the fix may depend on it." >&2
  fi
fi

if step 2 'List the processes actually in uninterruptible sleep.' 'The wchan column names the kernel function they are stuck in -- nfs_wait, io_schedule and similar point straight at the subsystem.' cmd 'ps -eo state,pid,user,wchan:30,cmd | awk '\''$1 ~ /D/'\'''; then
  ps -eo state,pid,user,wchan:30,cmd | awk '$1 ~ /D/'
  if [ $? -ne 0 ]; then
    failed=$((failed+1))
    echo "  Step 2 failed. The rest of the fix may depend on it." >&2
  fi
fi

if step 3 'Find which device is saturated.' '%util near 100 with a high await means that device is the bottleneck. A high await with low %util usually means the storage behind it, not the disk itself.' cmd 'iostat -xz 1 5'; then
  iostat -xz 1 5
  if [ $? -ne 0 ]; then
    failed=$((failed+1))
    echo "  Step 3 failed. The rest of the fix may depend on it." >&2
  fi
fi

if step 4 'Attribute the I/O to a process.' '' cmd 'sudo iotop -oPa -n 5'; then
  sudo iotop -oPa -n 5
  if [ $? -ne 0 ]; then
    failed=$((failed+1))
    echo "  Step 4 failed. The rest of the fix may depend on it." >&2
  fi
fi

if step 5 'Check for hung NFS mounts, a very common cause of unkillable D-state processes.' '' cmd 'mount -t nfs,nfs4' 'sudo dmesg -T | grep -i '\''nfs.*not responding'\'' | tail'; then
  mount -t nfs,nfs4
  sudo dmesg -T | grep -i 'nfs.*not responding' | tail
  if [ $? -ne 0 ]; then
    failed=$((failed+1))
    echo "  Step 5 failed. The rest of the fix may depend on it." >&2
  fi
fi

if step 6 'Look for kernel hung-task warnings, which name the stuck process and its stack.' '' cmd 'sudo dmesg -T | grep -iA10 '\''hung_task\|blocked for more than'\'' | tail -40'; then
  sudo dmesg -T | grep -iA10 'hung_task\|blocked for more than' | tail -40
  if [ $? -ne 0 ]; then
    failed=$((failed+1))
    echo "  Step 6 failed. The rest of the fix may depend on it." >&2
  fi
fi

rule "  Confirm it worked"
prose 'The load figure tracks the CPU run queue again once the I/O source is dealt with.'
if [ "$DRYRUN" = "0" ]; then
  uptime; vmstat 1 3
fi

rule
if [ "$failed" -gt 0 ]; then
  echo "  Finished with $failed failed step(s)."
  echo "  Read the full write-up at https://jbtecwiz.com/support/lnx-high-load"
else
  echo "  Finished."
fi
echo
prose 'To undo: None -- diagnostic.'
rule