To test the new design, use the increasing number of unit tests:
$ python -m unittest discover lava_dispatcher/pipeline/
To run a single test, use the test class name as output by a failing test, without the call to discover:
$ python -m unittest lava_dispatcher.pipeline.test.test_basic.TestPipelineInit.test_pipeline_init
$ python -m unittest -v -c -f lava_dispatcher.pipeline.test.test_basic.TestPipelineInit.test_pipeline_init
Sets of tests can also be executed from the ./ci-run script of lava-dispatcher as well:
$ ./ci-run --test-suite lava_dispatcher.pipeline.test.test_basic.TestPipelineInit.test_pipeline_init
Also, install the updated lava-dispatcher package and use it to inspect the output of the pipeline using the --validate switch to lava-dispatch:
$ sudo lava-dispatch --validate --target ./devices/kvm01.yaml ./sample_jobs/kvm.yaml --output-dir=/tmp/test
Note
The refactoring has changed the behaviour of target - the value must be a path to a YAML file, not a hostname. This is because the refactored dispatcher has no local configuration, so the master sends the entire device configuration to the dispatcher as a single YAML file.
The structure of any one job will be the same each time it is run (subject to changes in the developing codebase). Each different job will have a different pipeline structure. Do not rely on any of the pipeline levels have any specific labels. When writing unit tests, only use checks based on isinstance or self.name. (The description and summary fields are subject to change to make the validation output easier to understand whereas self.name is a strict class-based label.)
(Actual output is subject to frequent change.)
!!python/object/apply:collections.OrderedDict
- - - device
- parameters:
actions:
boot:
prompts: ['linaro-test', 'root@debian:~#']
command:
amd64: {qemu_binary: qemu-system-x86_64}
methods: [qemu]
overrides: [boot_cmds, qemu_options]
parameters:
boot_cmds:
- {root: /dev/sda1}
- {console: 'ttyS0,115200'}
machine: accel=kvm:tcg
net: ['nic,model=virtio', user]
qemu_options: [-nographic]
deploy:
methods: [image]
architecture: amd64
device_type: kvm
hostname: kvm01
memory: 512
root_part: 1
- - job
- parameters: {action_timeout: 5m, device_type: kvm, job_name: kvm-pipeline, job_timeout: 15m,
output_dir: /tmp/codehelp, priority: medium, target: kvm01, yaml_line: 3}
- - '1'
- content:
description: deploy image using loopback mounts
level: '1'
name: deployimage
parameters:
deployment_data: &id001 {TESTER_PS1: 'linaro-test [rc=$(echo \$?)]# ', TESTER_PS1_INCLUDES_RC: true,
TESTER_PS1_PATTERN: 'linaro-test \[rc=(\d+)\]# ', boot_cmds: boot_cmds,
distro: debian, lava_test_dir: /lava-%s, lava_test_results_dir: /lava-%s,
lava_test_results_part_attr: root_part, lava_test_sh_cmd: /bin/bash}
summary: deploy image
valid: true
yaml_line: 12
description: deploy image using loopback mounts
summary: deploy image
- - '1.1'
- content:
description: download with retry
level: '1.1'
max_retries: 5
name: download_action
parameters:
deployment_data: *id001
sleep: 1
summary: download-retry
valid: true
description: download with retry
summary: download-retry
- - '1.2'
- content:
description: md5sum and sha256sum
level: '1.2'
name: checksum_action
parameters:
deployment_data: *id001
summary: checksum
valid: true
description: md5sum and sha256sum
summary: checksum
- - '1.3'
- content:
description: mount with offset
level: '1.3'
name: mount_action
parameters:
deployment_data: *id001
summary: mount loop
valid: true
description: mount with offset
summary: mount loop
- - 1.3.1
- content:
description: calculate offset of the image
level: 1.3.1
name: offset_action
parameters:
deployment_data: *id001
summary: offset calculation
valid: true
description: calculate offset of the image
summary: offset calculation
- - 1.3.2
- content:
description: ensure a loop back mount operation is possible
level: 1.3.2
name: loop_check
parameters:
deployment_data: *id001
summary: check available loop back support
valid: true
description: ensure a loop back mount operation is possible
summary: check available loop back support
- - 1.3.3
- content:
description: Mount using a loopback device and offset
level: 1.3.3
max_retries: 5
name: loop_mount
parameters:
deployment_data: *id001
retries: 10
sleep: 10
summary: loopback mount
valid: true
description: Mount using a loopback device and offset
summary: loopback mount
- - '1.4'
- content:
description: customise image during deployment
level: '1.4'
name: customise
parameters:
deployment_data: *id001
summary: customise image
valid: true
description: customise image during deployment
summary: customise image
- - '1.5'
- content:
description: load test definitions into image
level: '1.5'
name: test-definition
parameters:
deployment_data: *id001
summary: loading test definitions
valid: true
description: load test definitions into image
summary: loading test definitions
- - 1.5.1
- content:
description: apply git repository of tests to the test image
level: 1.5.1
max_retries: 5
name: git-repo-action
parameters:
deployment_data: *id001
sleep: 1
summary: clone git test repo
uuid: b32dd5ff-fb80-44df-90fb-5fbd5ab35fe5
valid: true
vcs_binary: /usr/bin/git
description: apply git repository of tests to the test image
summary: clone git test repo
- - 1.5.2
- content:
description: apply git repository of tests to the test image
level: 1.5.2
max_retries: 5
name: git-repo-action
parameters:
deployment_data: *id001
sleep: 1
summary: clone git test repo
uuid: 200e83ef-bb74-429e-89c1-05a64a609213
valid: true
vcs_binary: /usr/bin/git
description: apply git repository of tests to the test image
summary: clone git test repo
- - 1.5.3
- content:
description: overlay test support files onto image
level: 1.5.3
name: test-overlay
parameters:
deployment_data: *id001
summary: applying LAVA test overlay
valid: true
description: overlay test support files onto image
summary: applying LAVA test overlay
- - '1.6'
- content:
default_fixupdict: {FAIL: fail, PASS: pass, SKIP: skip, UNKNOWN: unknown}
default_pattern: (?P<test_case_id>.*-*)\s+:\s+(?P<result>(PASS|pass|FAIL|fail|SKIP|skip|UNKNOWN|unknown))
description: add lava scripts during deployment for test shell use
lava_test_dir: /usr/lib/python2.7/dist-packages/lava_dispatcher/lava_test_shell
level: '1.6'
name: lava-overlay
parameters:
deployment_data: *id001
runner_dirs: [bin, tests, results]
summary: overlay the lava support scripts
valid: true
xmod: 493
description: add lava scripts during deployment for test shell use
summary: overlay the lava support scripts
- - '1.7'
- content:
description: unmount the test image at end of deployment
level: '1.7'
max_retries: 5
name: umount
parameters:
deployment_data: *id001
sleep: 1
summary: unmount image
valid: true
description: unmount the test image at end of deployment
summary: unmount image
- - '2'
- content:
description: boot image using QEMU command line
level: '2'
name: boot_qemu_image
parameters:
parameters: {failure_retry: 2, media: tmpfs, method: kvm, yaml_line: 22}
summary: boot QEMU image
timeout: {duration: 30, name: boot_qemu_image}
valid: true
yaml_line: 22
description: boot image using QEMU command line
summary: boot QEMU image
- - '2.1'
- content:
description: Wait for a shell
level: '2.1'
name: expect-shell-connection
parameters:
parameters: {failure_retry: 2, media: tmpfs, method: kvm, yaml_line: 22}
summary: Expect a shell prompt
valid: true
description: Wait for a shell
summary: Expect a shell prompt
- - '3'
- content:
level: '3'
name: test
parameters:
parameters:
definitions:
- {from: git, name: smoke-tests, path: ubuntu/smoke-tests-basic.yaml,
repository: 'git://git.linaro.org/qa/test-definitions.git', yaml_line: 31}
- {from: git, name: singlenode-basic, path: singlenode01.yaml, repository: 'git://git.linaro.org/people/neilwilliams/multinode-yaml.git',
yaml_line: 39}
failure_retry: 3
name: kvm-basic-singlenode
yaml_line: 27
summary: test
valid: true
description: null
summary: test
- - '4'
- content:
level: '4'
description: finish the process and cleanup
name: finalize
parameters:
parameters: {}
summary: finalize the job
valid: true
description: finish the process and cleanup
summary: finalize the job
The code can be executed:
$ sudo lava-dispatch --target kvm01 lava_dispatcher/pipeline/test/sample_jobs/kvm.yaml --output-dir=/tmp/test
The hacks and workarounds in the old LavaTestShell classes may need to be marked and retained until such time as either the new model replaces the old or the bug can be fixed in both models. Whereas the submission schema, log file structure and result bundle schema have thrown away any backwards compatibility, LavaTestShell will need to at least attempt to retain compatibility whilst improving the overall design and integrating the test shell operations into the new classes.
Current possible issues include:
- If dependent test definitions use custom signal handlers, this may not work - it would depend on how the job parameters are handled by the new classes.
For a RetryAction to validate, the RetryAction subclass must be a wrapper class around a new internal_pipeline to allow the RetryAction.run() function to handle all of the retry functionality in one place.
An Action which needs to support failure_retry or which wants to use RetryAction support internally, needs a new class added which derives from RetryAction, sets a useful name, summary and description and defines a populate() function which creates the internal_pipeline. The Action with the customised run() function then gets added to the internal_pipeline of the RetryAction subclass - without changing the inheritance of the original Action.
To add Diagnostics, add subclasses of DiagnosticAction to the list of supported Diagnostic classes in the Job class. Each subclass must define a trigger classmethod which is unique across all Diagnostic subclasses. (The trigger string is used as an index in a generator hash of classes.) Trigger strings are only used inside the Diagnostic class. If an Action catches a JobError or InfrastructureError exception and wants to allow a specific Diagnostic class to run, import the relevant Diagnostic subclass and add the trigger to the current job inside the exception handling of the Action:
try:
self._run_command(cmd_list)
except JobError as exc:
self.job.triggers.append(DiagnoseNetwork.trigger())
raise JobError(exc)
return connection
Actions should only append triggers which are relevant to the JobError or InfrastructureError exception about to be raised inside an Action.run() function. Multiple triggers can be appended to a single exception. The exception itself is still raised (so that a RetryAction container will still operate).
Hint
A DownloadAction which fails to download a file could append a DiagnosticAction class which runs ifconfig or route just before raising a JobError containing the 404 message.
If the error to be diagnosed does not raise an exception, append the trigger in a conditional block and emit a JobError or InfrastructureError exception with a useful message.
Do not clear failed results of previous attempts when running a Diagnostic class - the fact that a Diagnostic was required is an indication that the job had some kind of problem.
Avoid overloading common Action classes with Diagnostics, add a new Action subclass and change specific Strategy classes (Deployment, Boot, Test) to use the new Action.
Avoid chaining Diagnostic classes - if a Diagnostic requires a command to exist, it must check that the command does exist. Raise a RuntimeError if a Strategy class leads to a Diagnostic failing to execute.
It is an error to add a Diagnostic class to any Pipeline. Pipeline Actions should be restricted to classes which have an effect on the Test itself, not simply reporting information.
Sometimes, a particular test image will support the expected command but a subsequent image would need an alternative. Generally, the expectation is that the initial command should work, therefore the fallback or helper action should not be needed. The refactoring offers support for this situation using Adjuvants.
An Adjuvant is a helper action which exists in the normal pipeline but which is normally skipped, unless the preceding Action sets a key in the PipelineContext that the adjuvant is required. A successful operation of the adjuvant clears the key in the context.
One example is the reboot command. Normal user expectation is that a reboot command as root will successfully reboot the device but LAVA needs to be sure that a reboot actually does occur, so usually uses a hard reset PDU command after a timeout. The refactoring allows LAVA to distinguish between a job where the soft reboot worked and a job where the PDU command became necessary, without causing the test itself to fail simply because the job didn’t use a hard reset.
If the ResetDevice Action determines that a reboot happened (by matching a pexpect on the bootloader initialisation), then nothing happens and the Adjuvant action (in this case, HardResetDevice) is marked in the results as skipped. If the soft reboot fails, the ResetDevice Action marks this result as failed but also sets a key in the PipelineContext so that the HardResetDevice action then executes.
Unlike Diagnostics, Adjuvants are an integral part of the pipeline and show up in the verification output and the results, whether executed or not. An Adjuvant is not a simple retry, it is a different action, typically a more aggressive or forced action. In an ideal world, the adjuvant would never be required.
A similar situation exists with firmware upgrades. In this case, the adjuvant is skipped if the firmware does not need upgrading. The preceding Action would not be set as a failure in this situation but LAVA would still be able to identify which jobs updated the firmware and which did not.
Most deployment Action classes run without needing a Connection. Once a Connection is established, the Action may need to run commands over that Connection. At this point, the Action delegates the maintenance of the run function to the Connection pexpect. i.e. the Action.run() is blocked, waiting for Connection.run_command() (or similar) to return and the Connection needs to handle timeouts, signals and other interaction over the connection. This role is taken on by the internal SignalDirector within each Connection. Unlike the old model, Connections have their own directors which takes the multinode and LMP workload out of the singlenode operations.
Devices on your desk can behave differently to those in the lab under full automation. Under automation, the hard_reset and power_off support means that the device is likely to be powered off when the first connection atttempt is made. On the desk, the device may spend more time powered on (even if the device is not running a usable system, for example the NFS location will be deleted when the previous job ends). So when writing connection classes and actions which initiate connections, check the power state of the device first.
An Action initiating a connection needs to know if it should wait for a prompt. In the run function, add:
if self.job.device.power_state not in ['on', 'off']:
self.wait(connection)
The next Action should be a ResetDevice action which understands the power state and determines whether to call the hard_reset commands or to attempt a soft reboot. In the populate function, ensure the correct ordering is in place:
self.internal_pipeline.add_action(MenuConnect())
self.internal_pipeline.add_action(ResetDevice())
Warn if the device has no automation support in the validate function:
if self.job.device.power_state in ['on', 'off']:
# to enable power to a device, either power_on or hard_reset are needed.
if self.job.device.power_command is '':
self.errors = "Unable to power on or reset the device %s" % hostname
if self.job.device.connect_command is '':
self.errors = "Unable to connect to device %s" % hostname
else:
self.logger.warning("%s may need manual intervention to reboot" % hostname)
Construct your pipeline to use Actions in the order:
Note
There may be several Retry actions necessary within these steps.
So, for a UBoot operation, this results in a pipeline like:
- UBootInterrupt - sets the Hit any key prompt in a new connection
- ResetDevice - sends the reboot command to the device
- ExpectShellSession - waits for the specified prompt to match
- UBootCommandsAction - issues the commands to UBoot
Typically, a Connection is started by an Action within the Pipeline. The call to start a Connection must not return until all operations on that Connection are complete or the Pipeline determines that the Connection needs to be terminated.
The refactored dispatcher has a different approach to logging:
Actual representation of the logs in the UI will change - these examples are the raw content of the output YAML.
- {debug: 'start: 1.4.2.3.7 test-install-overlay (max 300s)', ts: '2015-09-07T09:40:46.720450'}
- {debug: 'test-install-overlay duration: 0.02', ts: '2015-09-07T09:40:46.746036'}
- results:
test-install-overlay: !!python/object/apply:collections.OrderedDict
- - [success, a9b2300d-0864-4f9c-ba78-c2594b567fc5]
- [skipped, a9b2300d-0864-4f9c-ba78-c2594b567fc5]
- [duration, 0.024679899215698242]
- [timeout, 300.0]
- [level, 1.4.2.3.7]
- {debug: 'Received signal: <STARTTC> linux-linaro-ubuntu-pwd'}
- {target: ''}
- {target: ''}
- {target: ''}
- {target: ''}
- {debug: 'test shell timeout: 300 seconds'}
- {target: ''}
- {target: /lava-None/tests/0_smoke-tests}
- {target: <LAVA_SIGNAL_ENDTC linux-linaro-ubuntu-pwd>}
- {target: <LAVA_SIGNAL_TESTCASE TEST_CASE_ID=linux-linaro-ubuntu-pwd RESULT=pass>}
- {target: <LAVA_SIGNAL_STARTTC linux-linaro-ubuntu-uname>}
- {target: ''}
- {debug: 'Received signal: <ENDTC> linux-linaro-ubuntu-pwd'}
- {target: ''}
- {target: ''}
- {target: ''}
- {target: ''}
- {debug: 'test shell timeout: 300 seconds'}
- {debug: 'Received signal: <TESTCASE> TEST_CASE_ID=linux-linaro-ubuntu-pwd RESULT=pass'}
- {debug: 'res: {''test_case_id'': ''linux-linaro-ubuntu-pwd'', ''result'': ''pass''}
data: {''test_case_id'': ''linux-linaro-ubuntu-pwd'', ''result'': ''pass''}'}
- results: {linux-linaro-ubuntu-pwd: pass, testsuite: smoke-tests-basic}
- {info: 'ok: lava_test_shell seems to have completed'}
- debug: {curl-http: pass, direct-install: pass, direct-update: pass, linux-linaro-ubuntu-ifconfig: pass,
linux-linaro-ubuntu-ifconfig-dump: pass, linux-linaro-ubuntu-lsb_release: fail,
linux-linaro-ubuntu-lscpu: pass, linux-linaro-ubuntu-netstat: pass, linux-linaro-ubuntu-pwd: pass,
linux-linaro-ubuntu-route-dump-a: pass, linux-linaro-ubuntu-route-dump-b: pass,
linux-linaro-ubuntu-route-ifconfig-up: pass, linux-linaro-ubuntu-route-ifconfig-up-lo: pass,
linux-linaro-ubuntu-uname: pass, linux-linaro-ubuntu-vmstat: pass, ping-test: pass,
remove-tgz: pass, tar-tgz: pass}
- {debug: 'lava-test-shell duration: 26.88', ts: '2015-09-07T09:43:14.065956'}
Pipeline jobs are sent to the slave dispatcher over ZMQ as fully formatted YAML files. These files are then passed to lava-dispatch when the job starts. To reproduce issues on the slave, the original files are retained in a temporary directory after the job has completed. As long as the slave has not been rebooted since the job started, the files will be retained in /tmp/lava-dispatcher/slave/<JOB_ID>/. These can then be used to re-run the job on the command line. Also in this directory, there is an err file which tracks any exceptions caught by the slave during the job run - these are sent back to the master and appear as a failure comment. Exceptions of this kind can then generate bug reports so that the dispatcher code handles the issue instead of it falling back to the slave daemon to handle.
See also Mapping deployment actions to the python code:
The expectation is that new tasks for the dispatcher will be created by adding more specialist Actions and organising the existing Action classes into a new pipeline for the new task.
Adding new behaviour is a two step process:
A Strategy class may use conditionals to select between a number of top level Strategy Action classes, for example DeployImageAction is a top level Strategy Action class for the DeployImage strategy. If used, this conditional must only operate on job parameters and the device as the selection function is a classmethod.
A test Job will consist of multiple strategies, one for each of the listed actions in the YAML file. Typically, this may include a Deployment strategy, a Boot strategy and a Test strategy. Jobs can have multiple deployment, boot, or test actions. Strategies add top level Actions to the main pipeline in the order specified by the parser. For the parser to select the new strategy, the strategies.py module for the relevant type of action needs to import the new subclass. There should be no need to modify the parser itself.
A single top level Strategy Action implements a single strategy for the outer Pipeline. The use of Logical actions can provide sufficient complexity without adding conditionals to a single top level Strategy Action class. Image deployment actions will typically include a conditional to check if a Test action is required later so that the test definitions can be added to the overlay during deployment.
Re-use existing Action classes wherever these can be used without changes.
If two or more Action classes have very similar behaviour, re-factor to make a new base class for the common behaviour and retain the specialised classes.
Strategy selection via select() must only ever rely on the device and the job parameters. Add new parameters to the job to distinguish strategies, e.g. the boot method or deployment method.
Wherever a new class is added, that new class can be tested - if only to be sure that it is correctly initialised and added to the pipeline at the correct level. Always create a new file in the tests directory for new functionality. All unit tests need to be in a file with the test_ prefix and add a new YAML file to the sample_jobs so that the strategies to select the new code can be tested. See Basics of the YAML format.
Often the simplest way to understand the available parameters and how new statements in the device configuration or job submission show up inside the classes is to use a unit test. To run a single unit-test, for example test_function in a class called TestExtra in a file called test_extra.py, use:
$ python -m unittest -v -c -f lava_dispatcher.pipeline.test.test_extra.TestExtra.test_function
Example python code:
import os
import unittest
class TestExtra(unittest.TestCase): # pylint: disable=too-many-public-methods
def test_function(self):
print "Hello world"
When using a connection to a device, group calls over that connection to calls which are expected to return within a consistent timeout for that class. If the final command from the class starts a longer running process, e.g. boot, set the connection prompt to look for a message which will be seen on that connection within a similar timeframe to all the other calls made by that class. This allows test writers to correctly choose the timeout to extend.
Add to the documentation when adding new classes which implement new dispatcher actions, parameters or behaviour.
$ sudo apt install pylint
$ pylint -d line-too-long -d missing-docstring lava_dispatcher/pipeline/
$ sudo apt install graphviz
$ pyreverse lava_dispatcher/pipeline/
$ dot -Tpng classes_No_Name.dot > classes.png
(Actual images can be very large.)
$ sudo apt install python-meliae
Add this python snippet to a unit test or part of the code of interest:
from meliae import scanner
scanner.dump_all_objects('filename.json')
Once the test has run, the specified filename will exist. To analyse the results, start up a python interactive shell in the same directory:
$ python
>>> from meliae import loader
>>> om = loader.load('filename.json')
loaded line 64869, 64870 objs, 8.7 / 8.7 MiB read in 0.9s
checked 64869 / 64870 collapsed 5136
set parents 59733 / 59734
collapsed in 0.4s
>>> s = om.summarize(); s
Note
The python interpreter, the setup.py configuration and other tools may allocate memory as part of the test, so the figures in the output may be larger than it would seem for a small test. A basic test may give a summary of 12Mb, total size. Figures above 100Mb should prompt a check on what is using the extra memory.
Note
These provisions are under development and are likely to change substantially. e.g. it may be possible to do a lot of these tasks using secondary media and secondary connections.
There are several situations where an environment needs to be setup in a contained and tested manner and then used for one or multiple LAVA test operations.
One solution is to use MultiNode and this works well when the device under test supports a secondary connection, e.g. ethernet.
MultiNode has requirements on a POSIX-type command line shell to be able to pass messages, e.g. busybox.
QEMU tests involve downloading a pre-built chroot based on a stable distribution release of a foreign architecture and running tests inside that chroot.
Android tests may involve setting up a VM or a configured chroot to expose USB devices whilst retaining the ability to use different versions of tools for different tests.