#!/usr/bin/env ruby

# Reports current CPU, disk, load average, and memory use to riemann.

require File.expand_path('../../lib/riemann/tools', __FILE__)

class Riemann::Tools::Health
  include Riemann::Tools

  opt :cpu_warning, "CPU warning threshold (fraction of total jiffies)", :default => 0.9
  opt :cpu_critical, "CPU critical threshold (fraction of total jiffies)", :default => 0.95
  opt :disk_warning, "Disk warning threshold (fraction of space used)", :default => 0.9
  opt :disk_critical, "Disk critical threshold (fraction of space used)", :default => 0.95
  opt :load_warning, "Load warning threshold (load average / core)", :default => 3
  opt :load_critical, "Load critical threshold (load average / core)", :default => 8
  opt :memory_warning, "Memory warning threshold (fraction of RAM)", :default => 0.85
  opt :memory_critical, "Memory critical threshold (fraction of RAM)", :default => 0.95

  def initialize
    @limits = {
      :cpu => {:critical => opts[:cpu_critical], :warning => opts[:cpu_warning]},
      :disk => {:critical => opts[:disk_critical], :warning => opts[:disk_warning]},
      :load => {:critical => opts[:load_critical], :warning => opts[:load_warning]},
      :memory => {:critical => opts[:memory_critical], :warning => opts[:memory_warning]}
    }
    case (ostype = `uname -s`.chomp.downcase)
    when 'darwin'
      @cores = `sysctl -n hw.ncpu`.to_i
      @cpu = method :darwin_cpu
      @disk = method :disk
      @load = method :darwin_load
      @memory = method :darwin_memory
      darwin_top
    when 'freebsd'
      @cores = `sysctl -n hw.ncpu`.to_i
      @cpu = method :freebsd_cpu
      @disk = method :disk
      @load = method :freebsd_load
      @memory = method :freebsd_memory
    else
      @cores = cores
      puts "WARNING: OS '#{ostype}' not explicitly supported. Falling back to Linux" unless ostype == "linux"
      @cpu = method :linux_cpu
      @disk = method :disk
      @load = method :linux_load
      @memory = method :linux_memory
    end
  end

  def alert(service, state, metric, description)
    report(
      :service => service.to_s,
      :state => state.to_s,
      :metric => metric.to_f,
      :description => description
    )
  end

  def cores
    i = 0;
    File.read("/proc/cpuinfo").split(/\n\n/).inject({}) do |cores, p|
      physical_id = p[/physical id\s+:\s+(\d+)/, 1]
      core_id = p[/core id\s+:\s+(\d+)/, 1]
      if physical_id and core_id
        cores["#{physical_id}:#{core_id}"] = true
      elsif physical_id
        cores["#{physical_id}:"] = true
      else
        cores[i += 1] = true;
      end

      cores
    end.size
  end

  def report_pct(service, fraction, report)
    if fraction
      if fraction > @limits[service][:critical]
        alert service, :critical, fraction, "#{sprintf("%.2f", fraction * 100)}% #{report}"
      elsif fraction > @limits[service][:warning]
        alert service, :warning, fraction, "#{sprintf("%.2f", fraction * 100)}% #{report}"
      else
        alert service, :ok, fraction, "#{sprintf("%.2f", fraction * 100)}% #{report}"
      end
    end
  end

  def linux_cpu
    new = File.read('/proc/stat')
    unless new[/cpu\s+(\d+)\s+(\d+)\s+(\d+)\s+(\d+)/]
      alert 'cpu', :unknown, nil, "/proc/stat doesn't include a CPU line"
      return false
    end
    u2, n2, s2, i2 = [$1, $2, $3, $4].map { |e| e.to_i }

    if @old_cpu
      u1, n1, s1, i1 = @old_cpu

      used = (u2+n2+s2) - (u1+n1+s1)
      total = used + i2-i1
      fraction = used.to_f / total

      report_pct :cpu, fraction, "user+nice+sytem\n\n#{`ps -eo pcpu,pid,comm | sort -nrb -k1 | head -10`.chomp}"
    end

    @old_cpu = [u2, n2, s2, i2]
  end

  def linux_load
    load = File.read('/proc/loadavg').split(/\s+/)[2].to_f / @cores
    if load > @limits[:load][:critical]
      alert "load", :critical, load, "15-minute load average/core is #{load}"
    elsif load > @limits[:load][:warning]
      alert "load", :warning, load, "15-minute load average/core is #{load}"
    else
      alert "load", :ok, load, "15-minute load average/core is #{load}"
    end
  end

  def linux_memory
    m = File.read('/proc/meminfo').split(/\n/).inject({}) { |info, line|
      x = line.split(/:?\s+/)
      # Assume kB...
      info[x[0]] = x[1].to_i
      info
    }

    free = m['MemFree'].to_i + m['Buffers'].to_i + m['Cached'].to_i
    total = m['MemTotal'].to_i
    fraction = 1 - (free.to_f / total)

    report_pct :memory, fraction, "used\n\n#{`ps -eo pmem,pid,comm | sort -nrb -k1 | head -10`.chomp}"
  end

  def freebsd_cpu
    u2, n2, s2, t2, i2 = `sysctl -n kern.cp_time 2>/dev/null`.split.map{ |e| e.to_i } #FreeBSD has 5 cpu stats

    if @old_cpu
      u1, n1, s1, t1, i1 = @old_cpu

      used = (u2+n2+s2+t2) - (u1+n1+s1+t1)
      total = used + i2-i1
      fraction = used.to_f / total

      report_pct :cpu, fraction, "user+nice+sytem+interrupt\n\n#{`ps -axo pcpu,pid,comm | sort -nrb -k1 | head -10`.chomp}"
    end

    @old_cpu = [u2, n2, s2, t2, i2]
  end

  def freebsd_load
    m = `uptime`.split[-1].match(/^[0-9]*\.[0-9]*$/)
    load = m[0].to_f / @cores
    if load > @limits[:load][:critical]
      alert "load", :critical, load, "15-minute load average/core is #{load}"
    elsif load > @limits[:load][:warning]
      alert "load", :warning, load, "15-minute load average/core is #{load}"
    else
      alert "load", :ok, load, "15-minute load average/core is #{load}"
    end
  end

  def freebsd_memory
    meminfo = `sysctl -n vm.stats.vm.v_page_count vm.stats.vm.v_wire_count vm.stats.vm.v_active_count 2>/dev/null`.chomp.split
    fraction = (meminfo[1].to_f + meminfo[2].to_f) / meminfo[0].to_f

    report_pct :memory, fraction, "used\n\n#{`ps -axo pmem,pid,comm | sort -nrb -k1 | head -10`.chomp}"
  end

  def darwin_top
    raw = `top -l 1 | grep -i "^\\(cpu\\|physmem\\|load\\)"`.chomp
    @topdata = {:stamp => Time.now.to_i }
    raw.each_line do |ln|
      if ln.match(/Load Avg: [0-9.]+, [0-9.]+, ([0-9.])+/i)
        @topdata[:load] = $1.to_f
      elsif ln.match(/CPU usage: [0-9.]+% user, [0-9.]+% sys, ([0-9.]+)% idle/i)
        @topdata[:cpu] = 1  - ($1.to_f / 100)
      elsif mdat = ln.match(/PhysMem: ([0-9]+)([BKMGT]) wired, ([0-9]+)([BKMGT]) active, ([0-9]+)([BKMGT]) inactive, ([0-9]+)([BKMGT]) used, ([0-9]+)([BKMGT]) free/i)
        wired = mdat[1].to_i * (1024 ** "BKMGT".index(mdat[2]))
        active = mdat[3].to_i * (1024 ** "BKMGT".index(mdat[4]))
        inactive = mdat[5].to_i * (1024 ** "BKMGT".index(mdat[6]))
        used = mdat[7].to_i * (1024 ** "BKMGT".index(mdat[8]))
        free = mdat[9].to_i * (1024 ** "BKMGT".index(mdat[10]))
        @topdata[:memory] = (wired + active + used).to_f / (wired + active + used + inactive + free)
      end
    end
  end

  def darwin_cpu
    darwin_top unless (Time.now.to_i - @topdata[:stamp]) < opts[:interval]
    unless @topdata[:cpu]
      alert 'cpu', :unknown, nil, "unable to get CPU stats from top"
      return false
    end
    report_pct :cpu,  @topdata[:cpu], "usage\n\n#{`ps -eo pcpu,pid,comm | sort -nrb -k1 | head -10`.chomp}"
  end

  def darwin_load
    darwin_top unless (Time.now.to_i - @topdata[:stamp]) < opts[:interval]
    unless @topdata[:load]
      alert 'load', :unknown, nil, "unable to get load ave from top"
      return false
    end
    metric = @topdata[:load] / @cores
    if metric > @limits[:load][:critical]
      alert "load", :critical, metric, "15-minute load average per core is #{metric}"
    elsif metric > @limits[:load][:warning]
      alert "load", :warning, metric, "15-minute load average per core is #{metric}"
    else
      alert "load", :ok, metric, "15-minute load average per core is #{metric}"
    end
  end

  def darwin_memory
    darwin_top unless (Time.now.to_i - @topdata[:stamp]) < opts[:interval]
    unless @topdata[:memory]
      alert 'memory', :unknown, nil, "unable to get memory data from top"
      return false
    end
    report_pct :memory,  @topdata[:memory], "usage\n\n#{`ps -eo pmem,pid,comm | sort -nrb -k1 | head -10`.chomp}"
  end

  def disk
    `df -P`.split(/\n/).each do |r|
      f = r.split(/\s+/)
      next unless f[0] =~ /^\//
        next if f[0] == 'Filesystem'
      x = f[4].to_f/100

      if x > @limits[:disk][:critical]
        alert "disk #{f[5]}", :critical, x, "#{f[4]} used"
      elsif x > @limits[:disk][:warning]
        alert "disk #{f[5]}", :warning, x, "#{f[4]} used"
      else
        alert "disk #{f[5]}", :ok, x, "#{f[4]} used"
      end
    end
  end

  def tick
    @cpu.call
    @memory.call
    @disk.call
    @load.call
  end
end

Riemann::Tools::Health.run