This patch changes the way logging is initialized in avocado app:

1. stdout/stderr/logging is replaced with in-memory handlers
2. when arg parsing is done, logging is reconfigured accordingly to
   setting and re-logs stored messages (when streams are enabled)
3. when job starts and "test" stream is enabled, "" and "avocado.test"
   streams are enabled
4. when job finishes and "test" stream enabled the additional outputs,
   the additional outputs are disabled again.

The 3 and 4 is necessarily as the "" logger contains all logs including
avocado initialization (stevedore, PIL, ...). This way the "test" stream
still outputs everything, but only during the job-execution and not
during avocado initialization and cleanup. (one can use "early" log to
see those).

Last but not least it keeps "avocado.app" output in stdout (and stderr),
but it moves all other outputs to stderr instead. This is IMO better
handling as one can distinguish between debug and "additional" output.
Signed-off-by: NLukáš Doktor <ldoktor@redhat.com>

When we abort the test and the test ignores SIGTERM, avocado hangs for
infinity. Let's wait a while and send SIGKILL.
Signed-off-by: NLukáš Doktor <ldoktor@redhat.com>

There is an issue when running SIMPLE tests with non absolute
paths, example:

 $ echo "#!/bin/bash" > /tmp/test.sh
 $ echo "exit 0" >> /tmp/test.sh
 $ chmod +x /tmp/test.sh
 $ cd /tmp
 $ avocado run test.sh

The previous work on the simple test runner, adding the filename
property, actually fixes that bug, so let's add a functional test to
avoid an regression.

Reference: https://trello.com/c/nbhyILO0Signed-off-by: NCleber Rosa <crosa@redhat.com>

The "replay_map" allows to substitute the test class with custom class.
The problem is, that the "methodName" defined in "test_params" might not
exist in this class.

We have two choices, either leave it up to the user, to substitute with
a compatible class (eg. by creating the "methodName" in "__init__"
before calling the super.__init__), or we can require them to always use
given "methodName". This patch uses the second approach as it seems less
magical to me.
Signed-off-by: NLukáš Doktor <ldoktor@redhat.com>

Using an unknown option in a given subcommand leads argparse to
show the main avocado usage instructions.

This patch makes the argparse show the subcommand help instead.

Also, argparse error exit code now matches error_codes.AVOCADO_FAIL,
which makes more sense than error_codes.AVOCADO_JOB_FAIL.

Reference:https://trello.com/c/o0TpH9aeSigned-off-by: NAmador Pahim <apahim@redhat.com>

There are a few leftover references to the old Plugin classes in
the functional tests. Let's remove them.
Signed-off-by: NCleber Rosa <crosa@redhat.com>

If a test has local imports (a fairly common use case),
we have to append the base test dir to sys.path. Otherwise,
we end up having a horrible traceback as a result:

Exception loading test
Traceback (most recent call last):
  File "/home/lmr/Code/avocado.lmr/avocado/core/runner.py", line 200, in _run_test
    instance = loader.load_test(test_factory)
  File "/home/lmr/Code/avocado.lmr/avocado/core/loader.py", line 262, in load_test
    test_module = imp.load_module(module_name, f, p, d)
  File "scylla_longevity.py", line 5, in <module>
    from sdcm.tester import ScyllaClusterTester
ImportError: No module named sdcm.tester

Which is clearly not wanted. We had this problem before
and now we have a regression. It's time for a functional
test to make sure we don't regress again.
Signed-off-by: NLucas Meneghel Rodrigues <lookkas@gmail.com>
Signed-off-by: NCleber Rosa <crosa@redhat.com>

Signed-off-by: NCleber Rosa <crosa@redhat.com>

Signed-off-by: NCleber Rosa <crosa@redhat.com>

Yes, too many plugin mentions, in this commit title, but that's what
it really is, so please bear with it.

Now that the plugin types have a finer granularity, let's show each
plugin type (based on the namespace) as a separate section.
Signed-off-by: NLucas Meneghel Rodrigues <lmr@scylladb.com>
Signed-off-by: NCleber Rosa <crosa@redhat.com>

Signed-off-by: NCleber Rosa <crosa@redhat.com>

Signed-off-by: NCleber Rosa <crosa@redhat.com>

Signed-off-by: NCleber Rosa <crosa@redhat.com>

Signed-off-by: NCleber Rosa <crosa@redhat.com>

The Avocado command line application has "subcommands". So far,
to execute these commands, a special method, named "dispatch"
was set on the command line parser result object (aka Namespace)
pointing to the `run` method of the same class.

In truthfullness, the command line parser doesn't have too much
to do with these methods. Also, the `set_defaults` method of the
command line argument parser is not exactly intended for that use.
So, let's ask the subcommands subparser to formally save the chosen
subcommand, at well, `subcommand`, and nothing more.

Since the application knows about the command plugins (the ones that
inherit from CLICmdBase), and their interface now *require* a method
called `run`, let's use the extension entry-point name to locate the
underlying object and its `run` method.
Signed-off-by: NCleber Rosa <crosa@redhat.com>

At this point, no plugin support exists in Avocado. This is in
preparation for the new plugin code to be cleanly introduced.

The plugins that play a role in the plugin architecture have
been removed, since they wouldn't be functional under the new
plugin management code. The ones that are add extra functionality
to Avocado have been kept, and will be ported to the new
architecture.

Also, most of the functional tests have been temporarily disabled.
The reason is that most of them run avocado, which depends on the
run command, which in turn, depends on the plugin archicture code.
Signed-off-by: NCleber Rosa <crosa@redhat.com>

This shows a more succint version of the help message (the so called
"usage" information). It's a certainly a matter of taste and change
of behavior, so this is marked as an RFC.

It also avoids a hackish injection of the print_help method as the
default dispatch method.
Signed-off-by: NCleber Rosa <crosa@redhat.com>

Signed-off-by: NLukáš Doktor <ldoktor@redhat.com>

Since we define the meaning of each exit code, we should be using that
ourselves.

Functional tests have been coded using literal values, and there may
be a good reason for that: proving that the API stands.  But there's a
big issue: we have not declared those codes as some kind of an
external API and guarantee stability. Also the subsequent changes in
this series actually change the meaning and values of those.
Signed-off-by: NCleber Rosa <crosa@redhat.com>

It's prone to mistakes maintain two separate ways to query for tests
(running them and displaying available ones).

This patch implements --dry-run, which follows the full "run" process,
only instead of running the test it skips them.

Additionally the --dry-run uses some tweaks to tell them apart from
normal executions, namely:

    1. No sysinfo gathering
    2. Unique id is 0000000000000000000000000000000000000000
    3. Unless user provides custom log location, fresh tmpdir is used
Signed-off-by: NLukáš Doktor <ldoktor@redhat.com>

In preparation for the large plugin refactor, let's simplify the code ASAP
and drop related (custom) functionality that is going to be replaced.

Standard setuptools entrypoints is going to provide better capability then
our custom code that supports external 'avocado_*.py'module loading.
Signed-off-by: NCleber Rosa <crosa@redhat.com>

This patch doesn't change any functionality, it only renames the
inner_runner to external_runner (the plugin is called simply external)
Signed-off-by: NLukáš Doktor <ldoktor@redhat.com>

Autopep8 style cleanup + alphabetic import + import cleanup in
avocado.core.{loader,test} and selftests.functional.test_basic.
Signed-off-by: NLukáš Doktor <ldoktor@redhat.com>

Currently we can only specify the loader plugins order via settings.
This patch adds "--loaders" argument and improves the granularity.
Now it's possible to set either @plugin_name, or TEST_TYPE, where
TEST_TYPE is the mapped name of the plugin (eg. SIMPLE, VT, ...)

There is one special name "DEFAULT", which injects all unspecified
loaders in the position of the "DEFAULT" keyword. When "DEFAULT" is
not specified, not-listed plugins/test_types are not evaluated.
(therefor it's possible to restrict some test loaders or only some
test types).

As the usecase changed slightly, the settings option was renamed
from plugins.loader_plugins_priority to plugins.loaders.
Signed-off-by: NLukáš Doktor <ldoktor@redhat.com>

Make sure we have avocado_ + __name__ prefixes to the dirs
created with tempfile, so that we have an easy way to spot
when we are not cleaning up directories properly. This mostly
concerns unittests, although the kernel_build lib and the
iso9660 lib also uses that API.
Signed-off-by: NLucas Meneghel Rodrigues <lmr@redhat.com>

Let's use a simple passtest to check for latest result,
and let's wait for avocado to finish. That should accomplish
the same result in a shorter and nicer way.
Signed-off-by: NLucas Meneghel Rodrigues <lmr@redhat.com>

The Xunit and JSON test cases inherit from PluginsTest. This test defines
some hardcoded-values tests, which are then inherited and run for each
inherited class with the exact same values. This patch creates abstract
class, which all 3 cases share and makes them inherit from it, rather
than from the actual testcase.
Signed-off-by: NLukáš Doktor <ldoktor@redhat.com>

Even though I may be the one to blame about coming up with this
"simple import magic", I believe it was a mistake and we should get
rid of them.

There are a couple of other ways to actually do development out of
of a source tree, including running unittests that do not require
this amount of boiler plate code. Examples include just setting
the PYTHONPATH environment variable to actually run (setuptools based)
`python setup.py develop`.

The little bits of what looks like the import magic that was left, is
not really import magic. It's just that functional tests need to locate
the Avocado source tree base directory to run the test runner from it.
Signed-off-by: NCleber Rosa <crosa@redhat.com>

To be honest, our test code could still be kept in the same directories and
have the same names. But I think we can improve two things here:

1) Flatten a little bit the directory structure of selftests. Two path
components are being dropped here: "all" and "avocado". So that
"selftests/all/functional/avocado" becomes simply
"selftests/functional".

2) File names match what is, by default, recognized by unittest (the
Standard Library module) based discovery of tests. That means that
doc_build_test.py becomes test_doc_build.py. Not a big deal IMHO.
Signed-off-by: NCleber Rosa <crosa@redhat.com>

Add some functional tests that ensure that we don't
get caught by mistakes made when modifying the human
test reporting code again (or at least, not by the
obvious ones).
Signed-off-by: NLucas Meneghel Rodrigues <lmr@redhat.com>

In some external test suites, what is defined as a "test" can not be
run by itself and need a specific runner script/tool.

This introduces the concept of an "inner runner", that is, a custom
test runner or tool that can deal with custom tests that do follow
the Avocado SIMPLE test definition.

This implements the Trello card:

 https://trello.com/c/TpXecE2n/486-introduce-inner-runner-to-avocado-run

More information can be found in the man page section about it.

Changes from v1:
 * removed "level" from "inner level runner" mentions
 * replicated docs from man page to "main" docs
Signed-off-by: NCleber Rosa <crosa@redhat.com>

Instead of creating at the end of the job. This allows users running
long (time wise) jobs to be monitored by "following" the
$AVOCADO_RESULT/latest/job.log file.

This implements the card described at:

 https://trello.com/c/QJ4wCD8x/447-reconsider-when-the-latest-log-symlink-is-createdSigned-off-by: NCleber Rosa <crosa@redhat.com>

On multiple places we modify the sys.path to make avocado work from
sources. Instead of `append` we should `insert` the path as when running
from avocado sources directory, we want to use the modules from here,
rather than from installed version.
Signed-off-by: NLukáš Doktor <ldoktor@redhat.com>

Previously the "file" loader was special and always matched the URL
which might be confusing and lead to problems when developing additional
plugins.

This commit tires to define the way loaders should work. They must never
fail (unless of uncaught exception which is reported and next available
plugins is used). Instead when the URL produces broken test or even a
single corrupted subtest, consider it as URL not for me and return empty
set. The loader proxy uses next plugin to resolve this URL. When no
plugin resolves the URL user is advised to run "avocado list -V" to see
the potential problems.

The order of loader plugins was also changed. Now the file-loader is
first by default and all remaining ones are loaded by the plugin
priority. The order can be changed in setting
"plugins/loader_plugins_priority". It uses "name" variable of the each
loader plugin. Not listed plugins are loaded afterwards.

Additionally the whole loader API was simplified. Now it only accepts
one parameter - list_tests - which is 3-state:

1. ALL       - all tests including broken/incorrect ones
2. AVAILABLE - available tests (for listing purposes)
2. DEFAULT   - default tests (to run when no url given)
Signed-off-by: NLukáš Doktor <ldoktor@redhat.com>

This patch renames the fail_on_error decorator to fail_on and adds
support to supply list of exceptions. This way the decorator can be used
not only to make tests fail on generic exceptions but also to specify
parts of code with expected failures of given type.
Signed-off-by: NLukáš Doktor <ldoktor@redhat.com>

Give to test writers the fail_on_error decorator, which
should turn any generic exception into an avocado
TestFail.

This is important for some tests that can raise
any exception during a complex operation, and any of those
exceptions can be considered a test failure.
Signed-off-by: NLucas Meneghel Rodrigues <lmr@redhat.com>

Make sure avocado complains loudly when people try to call
skip() from outside setUp() - this shouldn't be allowed in
avocado tests. Let the user know of the problem and ask them
to fix their tests. Also, update the docstring for skip(),
which will reduce the chances of test writers using it in
inappropriate places.
Signed-off-by: NLucas Meneghel Rodrigues <lmr@redhat.com>

On unhandeled unknown exceptions the test should finish with ERROR
rather than FAIL as the test developer didn't think it might happen.

This behavior is configurable in config when the user have a different
opinion.
Signed-off-by: NLukáš Doktor <ldoktor@redhat.com>

It was previously not possible for a test to skip itself (due to say,
missing requirements) during the setUp() phase. Any exception caught
at this point was reflected as a test failure.
Signed-off-by: NCleber Rosa <crosa@redhat.com>
Signed-off-by: NLucas Meneghel Rodrigues <lmr@redhat.com>