Previously, includes had to receive variables via a special 'vars'
field. With this patch, the include syntax becomes a more natural
datastructure without special fields and is more akin to the way
role includes/dependencies work.
Tested with the following playbook:
---
- hosts: localhost
connection: local
tasks:
- { include: inc1.yml, a: 1 }
- include: inc2.yml
b: 2
- include: inc3.yml
with_items:
- x
- y
- z
Fixes#3481
Still compatible with user: but deprecating it so we can have
a matching remote_user: in tasks, cannot be user: because of the
module of the same name. #3932
Signed-off-by: Brian Coca <briancoca+dev@gmail.com>
Reported by Rumen:
TASK: [fail FAIL] *************************************************************
skipping: [hostname.com]
failed: [hostname.com] => {"failed": true}
msg: Failed as requested from task
Traceback (most recent call last):
File "/usr/local/bin/ansible-playbook", line 268, in <module>
sys.exit(main(sys.argv[1:]))
File "/usr/local/bin/ansible-playbook", line 208, in main
pb.run()
File "/Library/Python/2.7/site-packages/ansible/playbook/__init__.py", line 262, in run
if not self._run_play(play):
File "/Library/Python/2.7/site-packages/ansible/playbook/__init__.py", line 580, in _run_play
if (hosts_count - len(host_list)) > int((play.max_fail_pct)/100.0 * hosts_count):
TypeError: object of type 'NoneType' has no len()
The play was just checking for the presence of the keyword in the
YAML datastructure, and not the value of the field, so doing something
like variable substitution was always causing the play to be accelerated
* Default variables are now fed directly into roles, just like the
other variables, so that roles see their unique values rather
than those set at the global level.
* Role dependency duplicates are now determined by checking the params used
when specifying them as dependencies rather than just on the name of the
role. For example, the following would be included twice without having
to specify "allow_duplicates: true":
dependencies:
- { role: foo, x: 1 }
- { role: foo, x: 2 }
The 'always_run' task clause allows one to execute a task even in
check mode.
While here implement Runner.noop_on_check() to check if a runner
really should execute its task, with respect to check mode option
and 'always_run' clause.
Also add the optional 'jinja2' argument to check_conditional() :
it allows to give this function a jinja2 expression without exposing
the 'jinja2_compare' implementation mechanism.
Tests `test_playbook_undefined_varsX_fail` check if ansible detects
undefined variables when `error_on_undefined_vars` is enabled. These
tests fail without "Improve behavior with error_on_undefined_vars
enabled" patch.
Tests `test_playbook_undefined_varsX_ignore` check if ansible ignores
undefined variables when `error_on_undefined_vars` is disabled.
Also modify PlayBook._run_task_internal() so error_on_undefined_vars is
testable.
The block that added the original list of roles was indented too far,
and was only being reached if a role had dependencies. This resulted
in roles without dependencies from being added to the list of roles.
Credit goes to looped for reporting and diagnosing the issue.
Fixes#3686
Dependencies are enabled by adding a new directory/file named
meta/main.yml to the role. The format of the dependencies are:
dependencies:
- { role: foo, x: 1, y: 2 }
- { role: bar, x: 3, y: 4 }
...
Dependencies inherit variables as they are seen at the time of the
dependency inclusion. For example, if foo(x=1, y=2) has a dependency
on bar(x=3,z=4), then bar will have variables (x=3,y=2,z=4).
Different roles can have dependencies on the same role, and this
variable inheritence allows for the reuse of generic roles quite easily.
For example:
Role 'car' has the following dependencies:
dependencies:
- { role: wheel, n: 1 }
- { role: wheel, n: 2 }
- { role: wheel, n: 3 }
- { role: wheel, n: 4 }
Role 'wheel' has the following dependencies:
dependencies:
- { role: tire }
- { role: brake }
The role 'car' is then used as follows:
- { role: car, type: honda }
And tasks/main.yml in each role simply contains the following:
- name: {{ type }} whatever {{ n }}
command: echo ''
TASK: [honda tire 1]
TASK: [honda brake 1]
TASK: [honda wheel 1]
TASK: [honda tire 2]
TASK: [honda brake 2]
TASK: [honda wheel 2]
TASK: [honda tire 3]
TASK: [honda brake 3]
TASK: [honda wheel 3]
TASK: [honda tire 4]
TASK: [honda brake 4]
TASK: [honda wheel 4]
TASK: [I'm a honda] <- (this is in roles/car/tasks/main.yml)
Previous patch was reverted due to the fact that there was an issue
with the results not always being a dictionary (they're sometimes
a unicode string, ie. when the with_items is used with yum). This
minor change corrects that by checking for a dict object.
Previous commit c3659741 expanded sudo_user during task construction,
but this is too early as it does not pick up variables set during
the play.
This commit moves sudo_user expansion to the runner after variables
have been merged.
If a variable was provided for an include, in either of these ways:
---
- hosts: all
tasks:
- include: included.yml param=www-data
- include: included.yml
vars:
param: www-data
and then that param was used as the value of sudo_user in the included
tasks:
---
- name: do something as a parameterized sudo_user
command: whoami
sudo: yes
sudo_user: $param
you would receive a "failed to parse: usage: sudo" error back and the
command would not execute.
This seemed to be due to a missing call to template.template somewhere,
because the final value being passed through ssh was still `$param`.
After some digging, the issue seems to instead have been a problem with
providing the wrong context to the template for expansion. Inside the
`Task` logic, it was passing `play.vars` as the context, where
`module_vars` seemed more appropriate. After replacing it, my test case
above ran without issue. There was a comment above suggesting that the
template call might be unnecessary, but removing it made the original
error return, since it is not getting escaped later down the line. I
removed the comment since it was inaccurate.
I tried to actually incorporate my test case above into the test suite
as a regression test, but was unable to figure out how to structure it.
The existing test infrastructure seemed to only be testing for correct
number of counts in things (ok vs. changed, etc.), without regard for
whether the content generated by the command is correct. If there is an
example of a test similar to this one (where I would want to check the
JSON generated to make sure sudo_user had been converted), please let me
know and I will be happy to submit an additional patch.
As documented in #2623, early variable substitution causes when_
tests to fail and possibly other side effects.
I can see the reason for this early substitution, likely introduced
in 1dfe60a6, to allow many playbook parameters to be templated.
This is a valid goal, but the recursive nature of the utils.template
function means that it goes too far.
At this point removing tasks from the list of parameters to be
substituted seems sufficient to make my tests pass. It may be the
case that other parameters should be excluded, but I suspect not.
Adding a test case. I would prefer to analyse not just the aggregate
statistics but also whether the results are as expected - I can't
see an easy way to do that with the available callbacks at present.
evaluate and replace '$item' with ''. Really it doesn't make sense to include multiple playbooks
via a loop variable, as you can do this with task + with_items already (and it's a simpler code
path). Given this is undocumented, this removes that feature, and we'll consider next how to
also add 'with_items' support directly to roles.
When operating on a unicode string in python 2.6, shlex.split returns
a result that does not work with the file constructor.
To reproduce this requires a task include that is templated (this is
because the templated string is a unicode result, whereas a non-
templated string is a non-unicode string)
[will@centos6.3] $ python
Python 2.6.6 (r266:84292, Sep 11 2012, 08:34:23)
[GCC 4.4.6 20120305 (Red Hat 4.4.6-4)] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> import shlex
>>> shlex.split(u'abc')
['a\x00\x00\x00b\x00\x00\x00c\x00\x00\x00']
[will@fedora17] $ python
Python 2.7.3 (default, Jul 24 2012, 10:05:38)
[GCC 4.7.0 20120507 (Red Hat 4.7.0-5)] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> import shlex
>>> shlex.split(u'abc')
['abc']
The proposed fix (coercing the include parameters to string before the
shlex.split) may not be ideal but it does fix the bug for my test case.
Instead of having to remember when to use which one, rename template_ds
to template and move the last bit of code from template to varReplace
(which gets used for all string replacements, in the end).
This means that you can template any data type without worrying about
whether it's a string or not, and the right thing will happen.
global_vars has higher precedence than inventory. Putting the all
group's variables into it overrides all other groups and hosts.
Partially fixes#1647.
Executive summary: skipping a host corrupts a variable (when it is registered)
We have a play existing out of multiple tasks that check a condition, if one of these tasks fails we want to skip all next tasks in the playbook. I noticed that if we skip a task because a certain condition is met, and this task has a register-attribute, I loose the value in the variable. Which means we cannot use that variable in subsequent tasks to evaluate because it was skipped:
```
- action: command test -d /some/directory
register: task
- action: command test -f /some/directory/file
register: task
only_if: '${task.rc} == 0'
- action: do something else
only_if: '${task.rc} == 0'
```
In the above example, if the second task is skipped (because the first failed), the third action will end with a "SyntaxError: invalid syntax" complaining about the unsubstituted ${task.rc} (even though it was set by the first task and used for skipping the second).
The following play demonstrates the problem:
```
- name: Test register on ignored tasks
hosts: all
gather_facts: no
vars:
skip: true
task: { 'rc': 666 }
tasks:
- action: debug msg='skip = ${skip}, task.rc = ${task.rc}'
- name: Skip this task, just to test if task has changed
action: command ls
register: task
only_if: '${skip} != True'
- action: debug msg='skip = ${skip}, task.rc = ${task.rc}'
- name: Now use task value
action: command echo 'Works !'
only_if: '${task.rc} == 0'
```
And the enclosed fix, fixes the above problem.
After spending 10 minutes to find which playbook had an action/local_action missing, I changed the error to include the task name (if set). The error eventually was caused because I added a name to a task, but the dash before the existing action was not removed.
We now check explicitely for 'module_setup' in the SETUP_CACHE in order to avoid skipping setup because SETUP_CACHE was populated some other way. Other modules can implement the same mechanism to test if they've already run.
This closes#1206.
This change makes a distinction between no_hosts_matched and no_hosts_remaining.
In both cases we do not start facts-gathering, or run any tasks.
In the case that there are no more hosts remaining, we abort running tasks and abort the playbook.
I also cleaned up the leftovers from the previous patchsets, as these are no longer required.
This closes#1187.
Example playbook:
```yaml
---
- hosts: emptygroup
tasks:
- action: command date
- action: command false
- hosts: all
gather_facts: False
tasks:
- action: command ls
- action: command false
- action: command true
- hosts: all
tasks:
- action: command true
- action: command false
- hosts: all
tasks:
- action: command pwd
```
Sometimes you may want to allow variables through host_vars or inventory, but prompt for a value if it is not set or if the value does not conform to something specific. This option allows you to specify when you want to offer a prompt.
This patch also moves check_conditional to utils, and adds an is_unset() function which is nicer to read:
only_if: "not is_set('${var}')"
vs
only_if: "is_unset('${var}')"
James Cammarata
Michael DeHaan
bennojoy
Thomas Omans
Hiroaki Nakamura
Brian Coca
Kavin Kankeshwar
kavink
smoothify
Wincent Colaiuta
Jocelyn Jaubert
Stoned Elipot
Kavin Kankeshwar
Petr Svoboda
René Moser
Serge van Ginderachter
Michael DeHaan
Chris Church
Ralph Tice
Rodney Quillo
jeromew
Tin Tvrtkovic
Alan Descoins
Ninety Thirty
George Miroshnykov
David Golden
Kent Frazier
Tim Habermaas
Matt Coddington
willthames
Junegunn Choi
David Hughes
milan
Seth Vidal
Daniel Hokka Zakrisson
Rodney Quillo
Nandor Sivok
Dag Wieers
Aleksej Romanov
Aurélien Bondis
Jonathan Palley