Add configure-swap role

This is a role to configure swap and /opt on an infra host.

Originally we wrote this for devstack-gate with
Icd4e6d8ab84471ad06e69c3e0f9bac92776efc78.  It was ported into the
native devstack job with Iffe54fbccbccd68db08f79a1b51dd7f76dbff408.

This is really generic, and it's quite possible non-devstack jobs
would like to setup swap for jobs too, so o-z-j is a better home for
the role.

Change-Id: I0e9c846ace7fac8a1340746c6818fba6ec963018

.gitignore

@@ -1,3 +1,7 @@
+# Backup files
+*~
+\#*\#
+
 # Unit test / coverage reports
 .coverage
 .tox

roles/configure-swap/README.rst

@@ -0,0 +1,11 @@
+Configure a swap partition
+
+Creates a swap partition on the ephemeral block device (the rest of which
+will be mounted on /opt).
+
+**Role Variables**
+
+.. zuul:rolevar:: configure_swap_size
+   :default: 8192
+
+   The size of the swap partition, in MiB.

roles/configure-swap/defaults/main.yaml

@@ -0,0 +1,2 @@
+# Default swap partition/file size, in MiB
+configure_swap_size: 8192

roles/configure-swap/tasks/ephemeral.yaml

@@ -0,0 +1,118 @@
+---
+
+# Configure attached ephemeral devices for storage and swap
+
+- assert:
+    that:
+      - "ephemeral_device is defined"
+
+- name: Set partition names
+  set_fact:
+    swap_partition: "{{ ephemeral_device}}1"
+    opt_partition: "{{ ephemeral_device}}2"
+
+- name: Ensure ephemeral device is unmounted
+  become: yes
+  mount:
+    name: "{{ ephemeral_device }}"
+    state: unmounted
+
+- name: Get existing partitions
+  become: yes
+  parted:
+    device: "{{ ephemeral_device }}"
+    unit: MiB
+  register: ephemeral_partitions
+
+- name: Remove any existing partitions
+  become: yes
+  parted:
+    device: "{{ ephemeral_device }}"
+    number: "{{ item.num }}"
+    state: absent
+  with_items:
+    - "{{ ephemeral_partitions.partitions }}"
+
+- name: Create new disk label
+  become: yes
+  parted:
+    label: msdos
+    device: "{{ ephemeral_device }}"
+
+- name: Create swap partition
+  become: yes
+  parted:
+    device: "{{ ephemeral_device }}"
+    number: 1
+    state: present
+    part_start: '0%'
+    part_end: "{{ configure_swap_size }}MiB"
+
+- name: Create opt partition
+  become: yes
+  parted:
+    device: "{{ ephemeral_device }}"
+    number: 2
+    state: present
+    part_start: "{{ configure_swap_size }}MiB"
+    part_end: "100%"
+
+- name: Make swap on partition
+  become: yes
+  command: "mkswap {{ swap_partition }}"
+
+- name: Write swap to fstab
+  become: yes
+  mount:
+    path: none
+    src: "{{ swap_partition }}"
+    fstype: swap
+    opts: sw
+    passno: 0
+    dump: 0
+    state: present
+
+# XXX: does "parted" plugin ensure the partition is available
+# before moving on?  No udev settles here ...
+
+- name: Add all swap
+  become: yes
+  command: swapon -a
+
+- name: Create /opt filesystem
+  become: yes
+  filesystem:
+    fstype: ext4
+    # The default ratio is 16384 bytes per inode or so. Reduce that to 8192
+    # bytes per inode so that we get roughly twice the number of inodes as
+    # by default. This should still be well above the block size of 4096.
+    # We do this because we have found in at least a couple locations that
+    # more inodes is useful and is painful to fix after the fact.
+    opts: -i 8192
+    dev: "{{ opt_partition }}"
+
+# Rackspace at least does not have enough room for two devstack
+# installs on the primary partition.  We copy in the existing /opt to
+# the new partition on the ephemeral device, and then overmount /opt
+# to there for the test runs.
+#
+# NOTE(ianw): the existing "mount" touches fstab.  There is currently (Sep2017)
+# work in [1] to split mount & fstab into separate parts, but for now we bundle
+# it into an atomic shell command
+# [1] https://github.com/ansible/ansible/pull/27174
+- name: Copy old /opt
+  become: yes
+  shell: |
+    mount {{ opt_partition }} /mnt
+    find /opt/ -mindepth 1 -maxdepth 1 -exec mv {} /mnt/ \;
+    umount /mnt
+
+# This overmounts any existing /opt
+- name: Add opt to fstab and mount
+  become: yes
+  mount:
+    path: /opt
+    src: "{{ opt_partition }}"
+    fstype: ext4
+    opts: noatime
+    state: mounted

roles/configure-swap/tasks/main.yaml

@@ -0,0 +1,65 @@
+---
+
+# On RAX hosts, we have a small root partition and a large,
+# unallocated ephemeral device attached at /dev/xvde
+- name: Set ephemeral device if /dev/xvde exists
+  when: ansible_devices["xvde"] is defined
+  set_fact:
+    ephemeral_device: "/dev/xvde"
+
+# On other providers, we have a device called "ephemeral0".
+#
+# NOTE(ianw): Once [1] is in our ansible (2.4 era?), we can figure
+# this out more directly by walking the device labels in the facts
+#
+# [1] https://github.com/ansible/ansible/commit/d46dd99f47c0ee5081d15bc5b741e9096d8bfd3e
+- name: Set ephemeral device by label
+  when: ephemeral_device is undefined
+  block:
+    - name: Get ephemeral0 device node
+      command: /sbin/blkid -L ephemeral0
+      register: ephemeral0
+      # If this doesn't exist, returns !0
+      ignore_errors: yes
+      changed_when: False
+
+    - name: Set ephemeral device if LABEL exists
+      when: "ephemeral0.rc == 0"
+      set_fact:
+        ephemeral_device: "{{ ephemeral0.stdout }}"
+
+# If we have ephemeral storage and we don't appear to have setup swap,
+# we will create a swap and move /opt to a large data partition there.
+- include: ephemeral.yaml
+  static: no
+  when:
+    - ephemeral_device is defined
+    - ansible_memory_mb['swap']['total'] | int + 10 <= configure_swap_size
+
+# If no ephemeral device and no swap, then we will setup some swap
+# space on the root device to ensure all hosts a consistent memory
+# environment.
+- include: root.yaml
+  static: no
+  when:
+    - ephemeral_device is undefined
+    - ansible_memory_mb['swap']['total'] | int + 10 <= configure_swap_size
+
+# ensure a standard level of swappiness.  Some platforms
+# (rax+centos7) come with swappiness of 0 (presumably because the
+# vm doesn't come with swap setup ... but we just did that above),
+# which depending on the kernel version can lead to the OOM killer
+# kicking in on some processes despite swap being available;
+# particularly things like mysql which have very high ratio of
+# anonymous-memory to file-backed mappings.
+#
+# This sets swappiness low; we really don't want to be relying on
+# cloud I/O based swap during our runs if we can help it
+- name: Set swappiness
+  become: yes
+  sysctl:
+    name: vm.swappiness
+    value: 30
+    state: present
+
+- debug:  var=ephemeral_device

roles/configure-swap/tasks/root.yaml

@@ -0,0 +1,65 @@
+---
+
+# If no ephemeral devices are available, use root filesystem
+
+- name: Calculate required swap
+  set_fact:
+    swap_required: "{{ configure_swap_size - ansible_memory_mb['swap']['total'] | int }}"
+
+- block:
+    - name: Get root filesystem
+      shell: df --output='fstype' /root | tail -1
+      register: root_fs
+
+    - name: Save root filesystem
+      set_fact:
+        root_filesystem: "{{ root_fs.stdout }}"
+
+    - debug: var=root_filesystem
+
+# Note, we don't use a sparse device to avoid wedging when disk space
+# and memory are both unavailable.
+
+# Cannot fallocate on filesystems like XFS, so use slower dd
+- name: Create swap backing file for non-EXT fs
+  when: '"ext" not in root_filesystem'
+  become: yes
+  command: dd if=/dev/zero of=/root/swapfile bs=1M count={{ swap_required }}
+  args:
+    creates: /root/swapfile
+
+- name: Create sparse swap backing file for EXT fs
+  when: '"ext" in root_filesystem'
+  become: yes
+  command: fallocate -l {{ swap_required }}M /root/swapfile
+  args:
+    creates: /root/swapfile
+
+- name: Ensure swapfile perms
+  become: yes
+  file:
+    path: /root/swapfile
+    owner: root
+    group: root
+    mode: 0600
+
+- name: Make swapfile
+  become: yes
+  command: mkswap /root/swapfile
+
+- name: Write swap to fstab
+  become: yes
+  mount:
+    path: none
+    src: /root/swapfile
+    fstype: swap
+    opts: sw
+    passno: 0
+    dump: 0
+    state: present
+
+- name: Add all swap
+  become: yes
+  command: swapon -a
+
+- debug: var=swap_required