X-Git-Url: https://git.openstreetmap.org./chef.git/blobdiff_plain/aaf1ea5c8f1f14b6b81bea7aad51743c767d310c..1874a3b948505eb5ba23d10059597d81937fa94a:/cookbooks/hardware/recipes/default.rb

diff --git a/cookbooks/hardware/recipes/default.rb b/cookbooks/hardware/recipes/default.rb
index 91b7510b9..4122a7ac8 100644
--- a/cookbooks/hardware/recipes/default.rb
+++ b/cookbooks/hardware/recipes/default.rb
@@ -17,18 +17,27 @@
 # limitations under the License.
 #
 
+include_recipe "apt"
+include_recipe "git"
+include_recipe "prometheus"
+include_recipe "sysfs"
 include_recipe "tools"
-include_recipe "munin"
 
 ohai_plugin "hardware" do
   template "ohai.rb.erb"
 end
 
-case node[:cpu][:"0"][:vendor_id]
-when "GenuineIntel"
-  package "intel-microcode"
-when "AuthenticAMD"
-  package "amd64-microcode"
+if platform?("debian")
+  package "firmware-linux"
+end
+
+if node[:cpu] && node[:cpu][:"0"] && node[:cpu][:"0"][:vendor_id]
+  case node[:cpu][:"0"][:vendor_id]
+  when "GenuineIntel"
+    package "intel-microcode"
+  when "AuthenticAMD"
+    package "amd64-microcode"
+  end
 end
 
 if node[:dmi] && node[:dmi][:system]
@@ -47,22 +56,40 @@ end
 
 units = []
 
-if node[:roles].include?("bytemark") || node[:roles].include?("exonetric")
+if node[:roles].include?("bytemark") || node[:roles].include?("exonetric") || node[:roles].include?("prgmr")
   units << "0"
 end
 
 case manufacturer
-when "HP"
+when "HP", "HPE"
+  include_recipe "apt::management-component-pack"
+
   package "hponcfg"
 
+  execute "update-ilo" do
+    action :nothing
+    command "/usr/sbin/hponcfg -f /etc/ilo-defaults.xml"
+    not_if { kitchen? }
+  end
+
+  template "/etc/ilo-defaults.xml" do
+    source "ilo-defaults.xml.erb"
+    owner "root"
+    group "root"
+    mode "644"
+    notifies :run, "execute[update-ilo]"
+  end
+
   package "hp-health" do
     action :install
     notifies :restart, "service[hp-health]"
+    only_if { platform?("ubuntu") && node[:lsb][:release].to_f < 22.04 }
   end
 
   service "hp-health" do
     action [:enable, :start]
     supports :status => true, :restart => true
+    only_if { platform?("ubuntu") && node[:lsb][:release].to_f < 22.04 }
   end
 
   if product.end_with?("Gen8", "Gen9")
@@ -75,20 +102,29 @@ when "HP"
       action [:enable, :start]
       supports :status => true, :restart => true
     end
+  elsif product.end_with?("Gen10")
+    package "amsd" do
+      action :install
+      notifies :restart, "service[amsd]"
+    end
+
+    service "amsd" do
+      action [:enable, :start]
+      supports :status => true, :restart => true
+    end
   end
 
-  units << "1"
+  units << if product.end_with?("Gen10")
+             "0"
+           else
+             "1"
+           end
 when "TYAN"
   units << "0"
 when "TYAN Computer Corporation"
   units << "0"
 when "Supermicro"
-  case product
-  when "H8DGU", "X9SCD", "X7DBU", "X7DW3", "X9DR7/E-(J)LN4F", "X9DR3-F", "X9DRW", "SYS-1028U-TN10RT+", "SYS-2028U-TN24R4T+", "SYS-1029P-WTRT", "Super Server"
-    units << "1"
-  else
-    units << "0"
-  end
+  units << "1"
 when "IBM"
   units << "0"
 when "VMware, Inc."
@@ -97,7 +133,7 @@ when "VMware, Inc."
   # Remove timeSync plugin completely
   # https://github.com/vmware/open-vm-tools/issues/302
   file "/usr/lib/open-vm-tools/plugins/vmsvc/libtimeSync.so" do
-    state :absent
+    action :delete
     notifies :restart, "service[open-vm-tools]"
   end
 
@@ -116,6 +152,7 @@ end
 units.sort.uniq.each do |unit|
   service "serial-getty@ttyS#{unit}" do
     action [:enable, :start]
+    not_if { kitchen? }
   end
 end
 
@@ -142,18 +179,21 @@ if File.exist?("/etc/default/grub")
   execute "update-grub" do
     action :nothing
     command "/usr/sbin/update-grub"
+    not_if { kitchen? }
   end
 
   template "/etc/default/grub" do
     source "grub.erb"
     owner "root"
     group "root"
-    mode 0o644
+    mode "644"
     variables :units => units, :entry => grub_entry
     notifies :run, "execute[update-grub]"
   end
 end
 
+package "initramfs-tools"
+
 execute "update-initramfs" do
   action :nothing
   command "update-initramfs -u -k all"
@@ -165,78 +205,128 @@ template "/etc/initramfs-tools/conf.d/mdadm" do
   source "initramfs-mdadm.erb"
   owner "root"
   group "root"
-  mode 0o644
+  mode "644"
   notifies :run, "execute[update-initramfs]"
 end
 
-package "haveged"
-service "haveged" do
-  action [:enable, :start]
+# haveged is only required on older kernels
+# /dev/random is not blocking anymore in 5.15+
+if Chef::Util.compare_versions(node[:kernel][:release], [5, 15]).negative?
+  package "haveged"
+  service "haveged" do
+    action [:enable, :start]
+  end
+else
+  service "haveged" do
+    action [:stop, :disable]
+  end
+  package "haveged" do
+    action :remove
+  end
 end
 
-package "ipmitool" if node[:kernel][:modules].include?("ipmi_si")
+watchdog_module = %w[hpwdt sp5100_tco].find do |module_name|
+  node[:hardware][:pci]&.any? { |_, pci| pci[:modules]&.any?(module_name) }
+end
 
-package "irqbalance"
+if node[:kernel][:modules].include?("ipmi_si")
+  package "ipmitool"
+  package "freeipmi-tools"
 
-template "/etc/default/irqbalance" do
-  source "irqbalance.erb"
-  owner "root"
-  group "root"
-  mode 0o644
+  template "/etc/prometheus/ipmi_local.yml" do
+    source "ipmi_local.yml.erb"
+    owner "root"
+    group "root"
+    mode "644"
+  end
+
+  prometheus_exporter "ipmi" do
+    port 9290
+    user "root"
+    private_devices false
+    protect_clock false
+    system_call_filter ["@system-service", "@raw-io"]
+    options "--config.file=/etc/prometheus/ipmi_local.yml"
+    subscribes :restart, "template[/etc/prometheus/ipmi_local.yml]"
+  end
+
+  watchdog_module ||= "ipmi_watchdog"
 end
 
+package "irqbalance"
+
 service "irqbalance" do
   action [:start, :enable]
   supports :status => false, :restart => true, :reload => false
-  subscribes :restart, "template[/etc/default/irqbalance]"
 end
 
-# Link Layer Discovery Protocol Daemon
 package "lldpd"
+
 service "lldpd" do
   action [:start, :enable]
   supports :status => true, :restart => true, :reload => true
 end
 
+ohai_plugin "lldp" do
+  template "lldp.rb.erb"
+end
+
+package %w[
+  rasdaemon
+  ruby-sqlite3
+]
+
+service "rasdaemon" do
+  action [:enable, :start]
+end
+
+prometheus_exporter "rasdaemon" do
+  port 9797
+  user "root"
+end
+
 tools_packages = []
 status_packages = {}
 
-node[:kernel][:modules].each_key do |modname|
-  case modname
-  when "cciss"
-    tools_packages << "ssacli"
-    status_packages["cciss-vol-status"] ||= []
-  when "hpsa"
-    tools_packages << "ssacli"
-    status_packages["cciss-vol-status"] ||= []
-  when "mptsas"
-    tools_packages << "lsiutil"
-    status_packages["mpt-status"] ||= []
-  when "mpt2sas", "mpt3sas"
-    tools_packages << "sas2ircu"
-    status_packages["sas2ircu-status"] ||= []
-  when "megaraid_mm"
-    tools_packages << "megactl"
-    status_packages["megaraid-status"] ||= []
-  when "megaraid_sas"
-    tools_packages << "megacli"
-    status_packages["megaclisas-status"] ||= []
-  when "aacraid"
-    tools_packages << "arcconf"
-    status_packages["aacraid-status"] ||= []
-  when "arcmsr"
-    tools_packages << "areca"
-  end
-end
-
-node[:block_device].each do |name, attributes|
-  next unless attributes[:vendor] == "HP" && attributes[:model] == "LOGICAL VOLUME"
-
-  if name =~ /^cciss!(c[0-9]+)d[0-9]+$/
-    status_packages["cciss-vol-status"] |= ["cciss/#{Regexp.last_match[1]}d0"]
-  else
-    Dir.glob("/sys/block/#{name}/device/scsi_generic/*").each do |sg|
-      status_packages["cciss-vol-status"] |= [File.basename(sg)]
+if node[:virtualization][:role] != "guest" ||
+   (node[:virtualization][:system] != "lxc" &&
+    node[:virtualization][:system] != "lxd" &&
+    node[:virtualization][:system] != "openvz")
+
+  node[:kernel][:modules].each_key do |modname|
+    case modname
+    when "cciss"
+      tools_packages << "ssacli"
+      status_packages["cciss-vol-status"] ||= []
+    when "hpsa"
+      tools_packages << "ssacli"
+      status_packages["cciss-vol-status"] ||= []
+    when "mptsas"
+      tools_packages << "lsiutil"
+      status_packages["mpt-status"] ||= []
+    when "mpt2sas", "mpt3sas"
+      tools_packages << "sas2ircu"
+      status_packages["sas2ircu-status"] ||= []
+    when "megaraid_sas"
+      tools_packages << "megacli"
+      status_packages["megaclisas-status"] ||= []
+    when "aacraid"
+      tools_packages << "arcconf"
+      status_packages["aacraid-status"] ||= []
+    when "arcmsr"
+      tools_packages << "areca"
+    end
+  end
+
+  node[:block_device].each do |name, attributes|
+    next unless attributes[:vendor] == "HP" && attributes[:model] == "LOGICAL VOLUME"
+
+    if name =~ /^cciss!(c[0-9]+)d[0-9]+$/
+      status_packages["cciss-vol-status"] |= ["cciss/#{Regexp.last_match[1]}d0"]
+    else
+      Dir.glob("/sys/block/#{name}/device/scsi_generic/*").each do |sg|
+        status_packages["cciss-vol-status"] |= [File.basename(sg)]
+      end
     end
   end
 end
@@ -257,8 +347,10 @@ if tools_packages.include?("areca")
   git "/opt/areca" do
     action :sync
     repository "https://git.openstreetmap.org/private/areca.git"
+    depth 1
     user "root"
     group "root"
+    not_if { kitchen? }
   end
 else
   directory "/opt/areca" do
@@ -267,92 +359,81 @@ else
   end
 end
 
-if status_packages.include?("cciss-vol-status")
-  template "/usr/local/bin/cciss-vol-statusd" do
-    source "cciss-vol-statusd.erb"
-    owner "root"
-    group "root"
-    mode 0o755
-    notifies :restart, "service[cciss-vol-statusd]"
-  end
+include_recipe "apt::hwraid" unless status_packages.empty?
 
-  systemd_service "cciss-vol-statusd" do
-    description "Check cciss_vol_status values in the background"
-    exec_start "/usr/local/bin/cciss-vol-statusd"
-    private_tmp true
-    protect_system "full"
-    protect_home true
-    no_new_privileges true
-    notifies :restart, "service[cciss-vol-statusd]"
-  end
-end
-
-%w[cciss-vol-status mpt-status sas2ircu-status megaraid-status megaclisas-status aacraid-status].each do |status_package|
+%w[cciss-vol-status mpt-status sas2ircu-status megaclisas-status aacraid-status].each do |status_package|
   if status_packages.include?(status_package)
     package status_package
 
-    template "/etc/default/#{status_package}d" do
-      source "raid.default.erb"
-      owner "root"
-      group "root"
-      mode 0o644
-      variables :devices => status_packages[status_package]
+    service "#{status_package}d" do
+      action [:stop, :disable]
     end
 
-    service "#{status_package}d" do
-      action [:start, :enable]
-      supports :status => false, :restart => true, :reload => false
-      subscribes :restart, "template[/etc/default/#{status_package}d]"
+    file "/etc/default/#{status_package}d" do
+      action :delete
     end
   else
     package status_package do
       action :purge
     end
-
-    file "/etc/default/#{status_package}d" do
-      action :delete
-    end
   end
 end
 
+systemd_service "cciss-vol-statusd" do
+  action :delete
+end
+
+template "/usr/local/bin/cciss-vol-statusd" do
+  action :delete
+end
+
 disks = if node[:hardware][:disk]
           node[:hardware][:disk][:disks]
         else
           []
         end
 
-# intel_ssds = disks.select { |d| d[:vendor] == "INTEL" && d[:model] =~ /^SSD/ }
-#
-# nvmes = if node[:hardware][:pci]
-#           node[:hardware][:pci].values.select { |pci| pci[:driver] == "nvme" }
-#         else
-#           []
-#         end
-#
-# intel_nvmes = nvmes.select { |pci| pci[:vendor_name] == "Intel Corporation" }
-#
-# if !intel_ssds.empty? || !intel_nvmes.empty?
-#   package "unzip"
-#
-#   intel_ssd_tool_version = "3.0.21"
-#
-#   remote_file "#{Chef::Config[:file_cache_path]}/Intel_SSD_Data_Center_Tool_#{intel_ssd_tool_version}_Linux.zip" do
-#     source "https://downloadmirror.intel.com/29115/eng/Intel_SSD_Data_Center_Tool_#{intel_ssd_tool_version}_Linux.zip"
-#   end
-#
-#   execute "#{Chef::Config[:file_cache_path]}/Intel_SSD_Data_Center_Tool_#{intel_ssd_tool_version}_Linux.zip" do
-#     command "unzip Intel_SSD_Data_Center_Tool_#{intel_ssd_tool_version}_Linux.zip isdct_#{intel_ssd_tool_version}-1_amd64.deb"
-#     cwd Chef::Config[:file_cache_path]
-#     user "root"
-#     group "root"
-#     not_if { File.exist?("#{Chef::Config[:file_cache_path]}/isdct_#{intel_ssd_tool_version}-1_amd64.deb") }
-#   end
-#
-#   dpkg_package "isdct" do
-#     version "#{intel_ssd_tool_version}-1"
-#     source "#{Chef::Config[:file_cache_path]}/isdct_#{intel_ssd_tool_version}-1_amd64.deb"
-#   end
-# end
+intel_ssds = disks.select { |d| d[:vendor] == "INTEL" && d[:model] =~ /^SSD/ }
+
+nvmes = if node[:hardware][:pci]
+          node[:hardware][:pci].values.select { |pci| pci[:driver] == "nvme" }
+        else
+          []
+        end
+
+unless nvmes.empty?
+  package "nvme-cli"
+end
+
+intel_nvmes = nvmes.select { |pci| pci[:vendor_name] == "Intel Corporation" }
+
+if !intel_ssds.empty? || !intel_nvmes.empty?
+  package "unzip"
+
+  sst_tool_version = "1.3"
+  sst_package_version = "#{sst_tool_version}.208-0"
+
+  # remote_file "#{Chef::Config[:file_cache_path]}/SST_CLI_Linux_#{sst_tool_version}.zip" do
+  #   source "https://downloadmirror.intel.com/743764/SST_CLI_Linux_#{sst_tool_version}.zip"
+  # end
+
+  execute "#{Chef::Config[:file_cache_path]}/SST_CLI_Linux_#{sst_tool_version}.zip" do
+    command "unzip SST_CLI_Linux_#{sst_tool_version}.zip sst_#{sst_package_version}_amd64.deb"
+    cwd Chef::Config[:file_cache_path]
+    user "root"
+    group "root"
+    not_if { ::File.exist?("#{Chef::Config[:file_cache_path]}/sst_#{sst_package_version}_amd64.deb") }
+  end
+
+  dpkg_package "sst" do
+    version "#{sst_package_version}"
+    source "#{Chef::Config[:file_cache_path]}/sst_#{sst_package_version}_amd64.deb"
+  end
+
+  dpkg_package "intelmas" do
+    action :purge
+  end
+end
 
 disks = disks.map do |disk|
   next if disk[:state] == "spun_down" || %w[unconfigured failed].any?(disk[:status])
@@ -363,31 +444,21 @@ disks = disks.map do |disk|
     if controller && controller[:device]
       device = controller[:device].sub("/dev/", "")
       smart = disk[:smart_device]
-
-      if device.start_with?("cciss/") && smart =~ /^cciss,(\d+)$/
-        array = node[:hardware][:disk][:arrays][disk[:arrays].first]
-        munin = "cciss-3#{array[:wwn]}-#{Regexp.last_match(1)}"
-      elsif smart =~ /^.*,(\d+)$/
-        munin = "#{device}-#{Regexp.last_match(1)}"
-      elsif smart =~ %r{^.*,(\d+)/(\d+)$}
-        munin = "#{device}-#{Regexp.last_match(1)}:#{Regexp.last_match(2)}"
-      end
+    elsif disk[:device]
+      device = disk[:device].sub("/dev/", "")
+      smart = disk[:smart_device]
     end
   elsif disk[:device] =~ %r{^/dev/(nvme\d+)n\d+$}
     device = Regexp.last_match(1)
-    munin = device
   elsif disk[:device]
     device = disk[:device].sub("/dev/", "")
-    munin = device
   end
 
   next if device.nil?
 
   Hash[
     :device => device,
-    :smart => smart,
-    :munin => munin,
-    :hddtemp => munin.tr("-:", "_")
+    :smart => smart
   ]
 end
 
@@ -396,18 +467,25 @@ disks = disks.compact.uniq
 if disks.count.positive?
   package "smartmontools"
 
+  template "/etc/cron.daily/update-smart-drivedb" do
+    source "update-smart-drivedb.erb"
+    owner "root"
+    group "root"
+    mode "755"
+  end
+
   template "/usr/local/bin/smartd-mailer" do
     source "smartd-mailer.erb"
     owner "root"
     group "root"
-    mode 0o755
+    mode "755"
   end
 
   template "/etc/smartd.conf" do
     source "smartd.conf.erb"
     owner "root"
     group "root"
-    mode 0o644
+    mode "644"
     variables :disks => disks
   end
 
@@ -415,27 +493,30 @@ if disks.count.positive?
     source "smartmontools.erb"
     owner "root"
     group "root"
-    mode 0o644
+    mode "644"
   end
 
-  service "smartd" do
+  service "smartmontools" do
     action [:enable, :start]
     subscribes :reload, "template[/etc/smartd.conf]"
     subscribes :restart, "template[/etc/default/smartmontools]"
   end
 
-  # Don't try and do munin monitoring of disks behind
-  # an Areca controller as they only allow one thing to
-  # talk to the controller at a time and smartd will
-  # throw errors if it clashes with munin
-  disks = disks.reject { |disk| disk[:smart]&.start_with?("areca,") }
+  template "/etc/prometheus/collectors/smart.devices" do
+    source "smart.devices.erb"
+    owner "root"
+    group "root"
+    mode "644"
+    variables :disks => disks
+  end
 
-  disks.each do |disk|
-    munin_plugin "smart_#{disk[:munin]}" do
-      target "smart_"
-      conf "munin.smart.erb"
-      conf_variables :disk => disk
-    end
+  prometheus_collector "smart" do
+    interval "15m"
+    user "root"
+    capability_bounding_set %w[CAP_DAC_OVERRIDE CAP_SYS_ADMIN CAP_SYS_RAWIO]
+    private_devices false
+    private_users false
+    protect_clock false
   end
 else
   service "smartd" do
@@ -443,28 +524,6 @@ else
   end
 end
 
-if disks.count.positive?
-  munin_plugin "hddtemp_smartctl" do
-    conf "munin.hddtemp.erb"
-    conf_variables :disks => disks
-  end
-else
-  munin_plugin "hddtemp_smartctl" do
-    action :delete
-    conf "munin.hddtemp.erb"
-  end
-end
-
-plugins = Dir.glob("/etc/munin/plugins/smart_*").map { |p| File.basename(p) } -
-          disks.map { |d| "smart_#{d[:munin]}" }
-
-plugins.each do |plugin|
-  munin_plugin plugin do
-    action :delete
-    conf "munin.smart.erb"
-  end
-end
-
 if File.exist?("/etc/mdadm/mdadm.conf")
   mdadm_conf = edit_file "/etc/mdadm/mdadm.conf" do |line|
     line.gsub!(/^MAILADDR .*$/, "MAILADDR admins@openstreetmap.org")
@@ -475,41 +534,57 @@ if File.exist?("/etc/mdadm/mdadm.conf")
   file "/etc/mdadm/mdadm.conf" do
     owner "root"
     group "root"
-    mode 0o644
+    mode "644"
     content mdadm_conf
   end
 
-  service "mdadm" do
+  service "mdmonitor" do
     action :nothing
     subscribes :restart, "file[/etc/mdadm/mdadm.conf]"
   end
 end
 
-template "/etc/modules" do
-  source "modules.erb"
-  owner "root"
-  group "root"
-  mode 0o644
+file "/etc/modules" do
+  action :delete
 end
 
-service "kmod" do
-  action :nothing
-  subscribes :start, "template[/etc/modules]"
+node[:hardware][:modules].each do |module_name|
+  kernel_module module_name do
+    action :install
+    not_if { kitchen? }
+  end
+end
+
+node[:hardware][:blacklisted_modules].each do |module_name|
+  kernel_module module_name do
+    action :blacklist
+  end
 end
 
-if node[:hardware][:watchdog]
-  package "watchdog"
+if watchdog_module
+  kernel_module watchdog_module do
+    action :install
+  end
+
+  execute "systemctl-reload" do
+    action :nothing
+    command "systemctl daemon-reload"
+    user "root"
+    group "root"
+  end
 
-  template "/etc/default/watchdog" do
-    source "watchdog.erb"
+  directory "/etc/systemd/system.conf.d" do
     owner "root"
     group "root"
-    mode 0o644
-    variables :module => node[:hardware][:watchdog]
+    mode "755"
   end
 
-  service "watchdog" do
-    action [:enable, :start]
+  template "/etc/systemd/system.conf.d/watchdog.conf" do
+    source "watchdog.conf.erb"
+    owner "root"
+    group "root"
+    mode "644"
+    notifies :run, "execute[systemctl-reload]"
   end
 end
 
@@ -551,7 +626,7 @@ unless Dir.glob("/sys/class/hwmon/hwmon*").empty?
     source "sensors.conf.erb"
     owner "root"
     group "root"
-    mode 0o644
+    mode "644"
     notifies :run, "execute[/etc/sensors.d/chef.conf]"
   end
 end
@@ -572,3 +647,14 @@ if node[:hardware][:shm_size]
     notifies :run, "execute[remount-dev-shm]"
   end
 end
+
+prometheus_collector "ohai" do
+  interval "15m"
+  user "root"
+  proc_subset "all"
+  capability_bounding_set %w[CAP_DAC_OVERRIDE CAP_SYS_ADMIN CAP_SYS_RAWIO]
+  private_devices false
+  private_users false
+  protect_clock false
+  protect_kernel_modules false
+end