Writing a Lava-Test Test Definition 1.0

Note

A Lava Test Shell Definition is distinct from a test job definition, although both use YAML. Typically, the test job definition includes URLs for one or more test shell definitions. The Lava-Test Test Definition 1.0 action then executes the test shell definitions and reports results as part of the test job. See also job definition and Metadata.

A LAVA Test Definition comprises

  1. Metadata describing the test definition, used by the test writers but not read by LAVA.
  2. The actions and parameters to set up the test(s)
  3. The instructions or steps to run as part of the test(s)

For certain tests, the instructions can be included inline with the actions. For more complex tests or to share test definitions across multiple devices, environments and purposes, the test can use a repository of YAML files.

Writing a test definition YAML file

Metadata

The YAML is downloaded from the repository (or handled inline) and installed into the test image, either as a single file or as part of a git repository. (See Test definitions in version control)

Each test definition YAML file contains metadata and instructions. Metadata includes:

  1. A format string recognized by LAVA
  2. A short name of the purpose of the file
  3. A description of the instructions contained in the file.
metadata:
    format: Lava-Test Test Definition 1.0
    name: singlenode-advanced
    description: "Advanced (level 3): single node test commands for Linux Linaro ubuntu Images"

Note

the short name of the purpose of the test definition, i.e., value of field name, must not contain any non-ascii characters or special characters from the following list, including white space(s): $& "'`()<>/\|;

If the file is not under version control (i.e. not in a git repository), the version of the file must also be specified in the metadata:

metadata:
    format: Lava-Test Test Definition 1.0
    name: singlenode-advanced
    description: "Advanced (level 3): single node test commands for Linux Linaro ubuntu Images"
    version: "1.0"

Optional metadata

There are also optional metadata fields:

  1. The email address of the maintainer of this file.
  2. A list of the operating systems which this file can support.
  3. A list of devices which are expected to be able to run these instructions.
maintainer:
    - user.user@linaro.org
os:
    - ubuntu
scope:
    - functional
devices:
    - kvm
    - arndale
    - panda
    - beaglebone-black
    - beagle-xm

These fields are ignored by LAVA itself; they exist only for test writers to use for their own requirements.

Deprecated installation commands

Warning

The install element of Lava-Test Test Definition 1.0 is DEPRECATED. See Write portable test definitions. Newly written Lava-Test Test Definition 1.0 files should not use install.

The instructions within the YAML file can include installation requirements for images based on supported distributions (currently, Ubuntu or Debian):

install:
    deps:
        - curl
        - realpath
        - ntpdate
        - lsb-release
        - usbutils

Note

for an install step to work, the test must first raise a usable network interface without running any instructions from the rest of the YAML file. If this is not possible, raise a network interface manually as a run step and install or build the components directly then.

When an external PPA or package repository (specific to debian based distributions) is required for installation of packages, it can be added in the install section as follows:

install:
    keys:
        - 7C751B3F
        - 6CCD4038
    sources:
        - https://security.debian.org
        - ppa:linaro-maintainers/tools
    deps:
        - curl
        - ntpdate
        - lavacli

Debian and Ubuntu repositories must be signed for the apt package management tool to trust them as package sources. To tell the system to trust extra repositories listed here, add references to the PGP keys used in the keys list. These may be either the names of Debian keyring packages (already available in the standard Debian archive), or PGP key IDs. If using key IDs, LAVA will import them from a key server (pgp.mit.edu). PPA keys will be automatically imported using data from launchpad.net. For more information, see the documentation of apt-add-repository, man 1 apt-add-repository

See Debian apt source addition and Ubuntu PPA addition

Note

When a new source is added and there are no ‘deps’ in the ‘install’ section, then it is the test writer’s responsibility to run apt update before attempting any other apt operation elsewhere in the test definition.

Note

When keys are not added for an apt source repository listed in the sources section, packages may fail to install if the repository is not trusted. LAVA does not add the –force-yes option during apt operations which would over-ride the trust check.

The principal purpose of the test definitions in the YAML file is to run commands on the device. These are specified in the run steps:

run:
    steps:

Writing commands to run on the device

  1. All commands need to be executables available on the device. This is why the metadata section includes an “os” flag, so that commands specific to that operating system can be accessed.

  2. All tests will be run in a dedicated working directory. If a test repository is used, the local checkout of that repository will also be located within that same directory.

  3. Avoid assumptions about the base system - if a test needs a particular interpreter, executable or environment, ensure that this is available. That can be done either by using the install step in the test definition, or by building or installing the components as a series of commands in the run steps. Many images will not contain any servers or compilers and many will only have a limited range of interpreters pre-installed. Some of those may also have reduced functionality compared to versions on other systems.

  4. Keep the YAML files relatively small and clean to promote easier reuse in other tests or devices. It is often better to have many YAML files to be run in sequence than to have a large overly complex YAML file within which some tests will fail due to changed assumptions. e.g. a smoke test YAML file which checks for USB devices is not useful on devices where lsusb is not functional. It is much easier to scan through the test results if the baseline for the test is that all tests should be expected to pass on all supported platforms.

  5. Check for the existence of one of the LAVA test helper scripts, like lava-test-case, in the directories specified by the PATH environment variable to determine how the script should report results. For example, the script may want to use echo or print() when not running inside LAVA and lava-test-case only when that script exists.

  6. Avoid use of redirects and pipes inside the run steps. If the command needs to use redirection and/or pipes, use a custom script in your repository and execute that script instead. See Writing custom scripts to support tests

  7. Take care with YAML syntax. These lines will fail with wrong syntax:

    - echo "test1: pass"
    - echo test2: fail
    

    While this syntax will pass:

    - echo "test1:" "pass"
    - echo "test2:" "fail"
    

Note

Commands must not try to access files from other test definitions. If a script needs to be in multiple tests, either combine the repositories into one or copy the script into multiple repositories. The copy of the script executed will be the one below the working directory of the current test.

Using inline test definitions

Rather than refer to a separate file or VCS repository, it is also possible to create a test definition directly inside the test action of a job submission. This is called an inline test definition:

- test:
    timeout:
      minutes: 4
    definitions:
    - repository:
        metadata:
          format: Lava-Test Test Definition 1.0
          name: apache-server
          description: "server installation"
          os:
          - debian
          scope:
          - functional
        run:
          steps:
          - apt -q update
          - apt -q -y install apache2
          - lava-test-case dpkg --shell dpkg -s apache2
      # remember to use -y to allow apt to proceed without interaction
      # -q simplifies the apt output for logging.
      from: inline
      name: apache-server
      path: inline/apache-server.yaml

An inline test definition must:

  1. Use the from: inline method.
  2. Provide a path to which the definition will be written
  3. Specify the metadata, at least:
    1. format: Lava-Test Test Definition 1.0
    2. name
    3. description

Inline test definitions will be written out as single files, so if the test definition needs to call any scripts or programs, those need to be downloaded or installed before being called in the inline test definition.

Download or view the complete example: examples/test-jobs/inline-test-definition-example.yaml

Terminology reference

LAVA Test Job

The test job provides test shell definitions (and inline definitions), as well as describing the steps needed to deploy code and boot a device to a command prompt. These steps will not be portable between devices or operating system deployments.

This design is quite different from LAVA V1 because V1 used to perform magic implicit steps. In V2 test jobs need to be explicit about all steps required.

Inline definitions are often used for prototyping test definitions. They are also the recommended choice for MultiNode synchronization primitives, inserted between the other LAVA Test Shell Definitions which do the bulk of the work.

The test job definition is what is submitted to LAVA to generate a test job.

LAVA Test Shell Definition

The LAVA Test Shell Definition is a YAML file, normally stored in a git repository alongside test writer scripts. Again, this will normally not be portable between operating system deployments.

It is possible to use different scripts, with the test job selecting which scripts to use for a particular deployment as it runs.

Each line in the definition must be a single line of shell, with no redirects, functions or pipes. Ideally, the Lava-Test Test Definition 1.0 will consist of a single run step which simply calls the appropriate test writer script.

LAVA Test Helpers

The LAVA Test Helpers are scripts maintained in the LAVA codebase, like lava-test-case. These are designed to work using only the barest minimum of operating system support, to make them portable to all deployments. Where necessary they will use deployment_data to customize content.

The helpers have two main uses:

  • to embed information from LAVA into the test shell and
  • to support communication with LAVA during test jobs.

Some helpers will always be required, for example to locate and start the test shell scripts.

Helpers which are too closely tied to any one operating system are likely to be deprecated and removed after LAVA V1 is dropped, along with helpers which duplicate standard operating system support.

For example, helpers which use distribution-specific utilities to install packages or add repositories.

Supporting OS variants

Most test shells can support portable test scripts without changes to the defaults.

  • lava_test_sh_cmd specifies the location of the shell interpreter. Default: /bin/sh
  • lava_test_results_dir specifies the location of the LAVA test directory which includes lava-test-runner. If this directory does not exist, the test shell will not start. Default: '/lava-%s'
  • lava_test_shell_file specifies the file to append with any Per dispatcher environment settings. Note: this is not the same as the LAVA params support. Default: '~/.bashrc'

These values can be overridden in the job context if the test job deploys a non-standard system as long as none of the deployments specify the os.

Test Writer Scripts

Test writer scripts are scripts written by test writers, designed to be run both by LAVA and by developers. They do not need to be portable to different operating system deployments, as the choice of script to run is up to the developer or test writer. This means that the test writer has a free choice of languages, methods and tools in these scripts - whatever is available within the particular operating system deployment. This can even include building custom tools from source if so desired.

The key feature of these scripts is that they should not depend on any LAVA features or helpers for their basic functionality. That way, developers can run exactly the same scripts both inside and outside of LAVA, to help reproduce problems.

When running inside LAVA, scripts should check for the presence of lava-test-case in the PATH environment variable and behave accordingly, using lava-test-case to report results to LAVA if it is available. Otherwise, report results to the user in whatever way makes most sense.

Test writers are strongly encouraged to make their scripts verbose: add progress messages, debug statements, error handling, logging and other support to allow developers to see what is actually happening when a test is running. This will aid debugging greatly.

Finally, scripts are commonly shared amongst test writers. It is a good idea to keep them self-contained as much as possible, as this will aid reuse. Also, try to stick to the common Unix model: one script doing one task.

See also

The next section on Writing custom scripts to support tests.

Writing custom scripts to support tests

Note

Custom scripts are not available in an inline definition, unless the definition itself downloads the script, adds any dependencies and makes the script executable.

When multiple actions are necessary to get usable output, write a custom script to go alongside the YAML and execute that script as a run step:

run:
    steps:
        - $(./my-script.sh arguments)

You can choose whatever scripting language you prefer, as long as you ensure that it is available in the test image.

Take care when using cd inside custom scripts - always store the initial return value or the value of pwd before the call and change back to that directory at the end of the script.

Example of a custom script wrapping the output:

https://git.linaro.org/lava-team/refactoring.git/tree/functional/dispatcher-unittests.sh

The script is simply called directly from the test shell definition:

https://git.linaro.org/lava-team/refactoring.git/tree/functional/server-unit-tests-stretch.yaml

Example V2 job using this support:

https://git.linaro.org/lava-team/refactoring.git/tree/functional/server-jessie-stretch-debian.yaml

Note

Make sure that your custom scripts output some useful information, including some indication of progress, in all test jobs but control the total amount of output to make the logs easier to read.

Advantages of custom scripts

Detailed knowledge of the output

Custom scripts can be written to take advantage of detailed knowledge of the expected output and the test environment. They don’t have to be generic (i.e. they can be specifically targeted to one test suite). They can use a variety of tools or programming language support to parse the test output.

Increased portability

Custom scripts can also allow test writers to make the Test Shell Definition more portable, to be run outside LAVA. It is recommended to do this wherever possible and not rely on LAVA-specific helper scripts. This allows developers who do not have access to the test framework to reproduce bugs found by the test framework whilst retaining the benefits of scripts which are specific to particular test output styles.

Problem reports can be difficult for developers to debug if they cannot reproduce the bug manually, without using the complete CI system. Every effort should be made to support running the test action instructions on a DUT which has been manually deployed so that developers can add specialized debug tools and equipment which are not available within the CI.

Script interpreters

  1. shell - consider running the script with set -x to see the operation of the script in the LAVA log files. Ensure that if your script expects bash, use the bash shebang line #!/bin/bash and ensure that bash is installed in the test image. The default shell may be busybox or dash, so take care with non-POSIX constructs in your shell scripts if you use #!/bin/sh.
  2. python - ensure that python is installed in the test image. Add all the python dependencies necessary for your script. Remember that Python2 is end-of-life and python3- alternative dependencies may be required.
  3. perl - ensure that any modules required by your script are available, bearing in mind that some images may only have a basic perl installation with a limited selection of modules.

If your YAML file does not reside in a repository, the YAML run steps will need to ensure that a network interface is raised, install a tool like wget and then use that to obtain the script, setting permissions if appropriate.

Using commands as test cases

If all your test does is feed the textual output of commands to the log file, you will spend a lot of time reading log files. To make test results easier to parse, aggregate and compare, individual commands can be converted into test cases with a pass or fail result. The simplest way to do this is to use the exit value of the command. A non-zero exit value is a test case failure. This produces a simple list of passes and failures in the result bundle which can be easily tracked over time.

To use the exit value, simply precede the command with a call to lava-test-case with a test-case name (no spaces):

run:
    steps:
        - lava-test-case test-ls-command --shell ls /usr/bin/sort
        - lava-test-case test-ls-fail --shell ls /user/somewhere/else/

Use subshells instead of backticks to execute a command as an argument to another command:

- lava-test-case pointless-example --shell ls $(pwd)

For more details on the contents of the YAML file and how to construct YAML for your own tests, see the Writing Tests.

Recording test case results

lava-test-case can also be used with a parser with the extra support for checking the exit value of the call:

run:
   steps:
    - "lava-test-case fail-test --shell false"
    - "lava-test-case pass-test --shell true"

This syntax will result in extra test results:

fail-test -> fail
pass-test -> pass

Alternatively, the --result command can be used to output the result directly:

run:
   steps:
      - lava-test-case test5 --result pass
      - lava-test-case test6 --result fail

This syntax will result in the test results:

test5 -> pass
test6 -> fail

Recording test case measurements and units

Various tests require measurements and lava-test-case supports measurements and units per test at a precision of 10 digits.

--result must always be specified and only numbers can be recorded as measurements (to support charts based on measurement trends).

run:
   steps:
      - lava-test-case test5 --result pass --measurement 99 --units bottles
      - lava-test-case test6 --result fail --measurement 0 --units mugs

This syntax will result in the test results:

test5 -> pass -> 99.0000000000 bottles
test6 -> fail -> 0E-10 mugs

The simplest way to use this with real data is to use a custom script which runs lava-test-case with the relevant arguments.

Recording sets of test cases

Test Set is a way to allow test writers to subdivide individual results within a single Lava Test Shell Definition using an arbitrary label.

Some test definitions run the same test with different parameters. To distinguish between these similar tests, it can be useful to use a test set.

Recording test case references

Some test cases may relate to specific bug reports or have specific URLs associated with the result. Simple strings can be recorded separately but if you need to relate a test case result to a URL, consider using lava-test-reference:

lava-test-reference TEST_CASE_ID --result pass|fail|skip|unknown --reference URL

The TEST_CASE_ID can be the same as an existing test case or a new test case.

lava-test-reference has similar support as lava-test-case except that --measurement and --unit options are not supported.

Note

Unlike the metadata in the test shell definition itself, the reference URL, result and the test case name are stored as part of the job metadata in the test job results. See also Metadata.

run:
   steps:
      - lava-test-case checked --result pass
      - lava-test-reference checked --result pass --reference https://staging.validation.linaro.org/static/doc/v2/index.html

Note

The URL should be a simple file reference, complex query strings could fail to be parsed.

Test shell parameters

The test action in the job definition supports parameters which are passed to the test shell. These parameters can be used to allow different job definitions to use a single test shell definition in multiple ways. A common example of this is a hacking session.

The parameters themselves are inserted into the lava-test-runner and will be available to all Lava Test Shell Definitions used in that test job. The parameters are not exported. The test shell definition needs to support using the parameter and can then use that information to change how external programs behave. This may include using export, it may include changing the command line options.

Obtaining information about the device

See also

Exported parameters for details of how this support is described in the device dictionary.

Some elements of the static device configuration are exposed to the test shell, where it is safe to do so and where the admin has explicitly configured the information. The information is exposed using test shell helpers which currently include:

  • lava-target-ip - Devices with a fixed IPv4 address will populate this field. Test writers are able to use this in an LXC to connect to the device, providing that the test shell has correctly raised a network connection and suitable services are configured and running on the device:

    ping -c4 $(lava-target-ip)
    
  • lava-target-mac - An alternative to lava-target-ip, declaring the MAC address of the device. Depending on the use case, this may be useful to lookup the IP address of the device:

    echo `lava-target-mac`
    
  • lava-target-storage - Where devices have alternative storage media fitted, the id of the block device can be exported. For example, this can help provide temporary storage on the device when the test shell is running a ramdisk or NFS. Some devices may provide a USB mass storage device which could also be exported in this way.

    Note

    This provision is designed to support temporary storage on devices which typically boot over NFS or ramdisk etc. It is intended to allow test writers to run operations which would typically fail without a local filesystem or would block network traffic such that NFS would time out.

    Only a single block device is supported per method. The method itself is simply a label specified by the admin. Often it will relate to the interface used by the block device, e.g. SATA or USB but it could be any string. In the example below, UMS is the label used by the device (as an abbreviation for USB Mass Storage).

    See also

    Extra device configuration and Persistence - test writers are responsible for handling persistence issues. The recommendation is that a new filesystem is created on the block device each time it is to be used.

    The output format contains one line per device, and each line contains the method and the ID for the storage using that method, separated by a TAB character:

    $ lava-target-storage
    UMS     /dev/disk/by-id/usb-Linux_UMS_disk_0_WaRP7-0xac2400d300000054-0:0
    SATA    /dev/disk/by-id/ata-ST500DM002-1BD142_W3T79GCW
    

    Usage: lava-target-storage method

    The output format contains one line per device assigned to the specified ID, with no whitespace. The matched method is not output.:

    $ lava-target-storage UMS
    /dev/disk/by-id/usb-Linux_UMS_disk_0_WaRP7-0xac2400d300000054-0:0
    

    If there is no matching method, exit non-zero and output nothing.

Recording test case data

Simple strings

A version string or similar can be recorded as a lava-test-case name:

lava-test-case ${VERSION} --result pass

Version strings need specific handling to compare for newer, older etc. so LAVA does not support comparing or ordering of such strings beyond simple alphanumeric sorting. A custom script would be the best way to handle such results.

For example, if your test definition uses a third party code repository, then it is always useful to use whatever support exists within that repository to output details like the current version or most recent commit hash or log message. This information may be useful when debugging a failure in the tests later. If or when particular tags, branches, commits or versions fail to work, the custom script can check for the supported or unsupported versions or names and report a fail test case result.

Files

See also

In LAVA V1, data files could be published using lava-test-case-attach. In V2, there is a new way to publish directly from the DUT - the publishing API.

Measurements

lava-test-case supports recording integer or floating point measurements for a particular test case. When a measurement is supplied, a text string can also be supplied to be used as the units of that measurement, e.g. seconds or bytes. Results are used to track changes across test jobs over time, so results which cannot be compared as integers or floating point numbers cannot be used as measurements.

The lava test results

Each test job creates a set of results in a reserved test suite called lava. LAVA will reject any submission which tries to set lava as the test definition name. These results are generated directly by the LAVA actions and include useful metadata including the actual time taken for specific actions and data generated during the test operation such as the VCS commit hash of each test definition included into the overlay.

The results are available in the same ways as any other test suite. In addition to strings and measurements, the lava suite also include an element called extra.

Examples

  • The lava test suite may contain a result for the git-repo-action test case, generated during the running of the test. The extra data in this test case could look like:

    extra:
      path: lava-test-shell/smoke-tests-basic.yaml
      repository: git://git.linaro.org/lava-team/lava-functional-tests.git
      success: c50a99ebb5835501181f4e34417e38fc819a6280
    
  • The duration result for the auto-login-action records the time taken to boot the kernel and get to a login prompt. The extra data for the same result includes details of kernel messages identified during the boot including stack traces, kernel panics and other alerts, if any.

Results from any test suite can be tracked using queries, charts and / or the REST API.

Note

The results in the lava test suite are managed by the software team. The results in the other test suites are entirely down to the test writer to manage. The less often the names of the test definitions and the test cases change, the easier it will be to track those test cases over time.

Best practices for writing a LAVA test job

A test job may consist of several LAVA test definitions and multiple deployments, but this flexibility needs to be balanced against the complexity of the job and the ways to analyze the results.

As with all things in automation, the core principles of best practice can be summarized as:

  1. Start small
  2. Build slowly
  3. Change only one thing at a time
  4. Test every change

Write portable test definitions

lava-test-shell is a useful helper but that can become a limitation. Avoid relying upon the helper for anything more than the automation by putting the logic and the parsing of your test into a more competent language. Remember: as test writer, you control which languages are available inside your test.

lava-test-shell has to try and get by with not much more than busybox ash as the lowest common denominator.

Please don’t expect lava-test-shell to do everything.

Let lava-test-shell provide you with a directory layout containing your scripts, some basic information about the job and a way of reporting test case results - that’s about all it should be doing outside of the MultiNode API.

Avoid using test definitions patterns

Test definitions which can use lava-test-case should not also use test definition patterns like:

"(?P<test_case_id>.*-*):\\s+(?P<result>(pass|fail))"

Test shell definition patterns are difficult to debug and almost impossible to make portable. If you have access to lava-test-case, there is no need to also use a pattern because you already have a shell on the DUT which is capable of much better pattern matching and parsing. Start by copying the relevant part of the test output and see how parsing can be improved:

  • Is any kind of pattern needed at all? Can the process generating the output be called by a script which already understands the output?
  • If you do need a pattern, put the pattern handling inside the test shell definition scripts and use copies of different sections of output to debug the pattern matching before submitting anything to LAVA.

Note

If the DUT does not support a POSIX shell then lava-test-case will not be available either. In some cases, the test operation is executed from an LXC and this will provide the necessary shell support.

Do not lock yourself out of your tests

  1. Do not make your test code depend on the LAVA infrastructure any more than is necessary for automation. Make sure you can always run your tests by downloading the test code to a target device using a clean environment, installing its dependencies (the test code itself could do this), and running a single script. Emulation can be used in most cases where access to the device is difficult. Even if the values in the output change, the format of the output from the underlying test operation should remain the same, allowing a single script to parse the output in LAVA and in local testing.
  2. Make the LAVA-specific part as small as possible, just enough to, for example, gather any inputs that you get via LAVA, call the main test program, and translate your regular output into ways to tell lava how the test went (if needed).
  3. Standard test jobs are intended to showcase the design of the test job, not the test definitions. These test definitions tend to be very simplistic and are not intended to be examples of how to write test definitions, just how to prepare test jobs.

Rely less on install: steps

To make your test portable, the goal of the install block of any test definition should be to get the raw LAVA environment up to the point where a developer would be ready to start test-specific operations. For example, installing package dependencies. Putting the operations into the run: steps also means that the test writer can report results from these operations.

Whilst compatibility with V1 has been retained in most areas of the test shell, there can be differences in how the install steps behave between V1 and V2. Once V1 is removed, other changes are planned for the test shell to make it easier for test writers to create portable tests. It is possible that the install: behavior of the test shell could be restricted at this time.

Consider moving install: git-repos: into a run step or directly into a custom_script along with the other setup (for example, switching branches or compiling the source tree). Then, when debugging the test job, a test writer can setup a similar environment and simply call exactly the same script.

Use different test definitions for different test areas

Follow the standard UNIX model of Make each program do one thing well. Make a set of separate test definitions. Each definition should concentrate on one area of functionality and test that one area thoroughly.

Use different jobs for different test environments

While it is supported to reboot from one distribution and boot into a different one, the usefulness of this is limited. If the first environment fails, the subsequent tests might not run at all.

Use a limited number of test definitions per job

While LAVA tries to ensure that all tests are run, adding more and more test repositories to a single LAVA job increases the risk that one test will fail in a way that prevents the results from all tests being collected.

Overly long sets of test definitions also increase the complexity of the log files, which can make it hard to identify why a particular job failed.

Splitting a large job into smaller chunks also means that the device can run other jobs for other users in between the smaller jobs.

Retain at least some debug output in the final test definitions

Information about which commit or version of any third-party code is and will remain useful when debugging failures. When cloning such code, call a script in the code or use the version control tools to output information about the cloned copy. You may want to include the most recent commit message or the current commit hash or version control tag or branch name.

If an item of configuration is important to how the test operates, write a test case or a custom script which reports this information. Even if this only exists in the test job log output, it will still be useful when comparing the log files of other similar jobs.

Mock up the device output to test the scripts

Avoid waiting for a device to deploy and boot for each iteration in the development of test support scripts. Copy the output of a working device and use that as the input to the scripts which process the logs to identify results and cut out the noise.

Where possible, include such mock ups as tests which can be run in another CI process, triggered each time the scripts are modified.

Use functional tests to validate common functionality

Use the principles of Functional testing of LAVA source code to test common code used by the test jobs. For example, if a shell library is used, ensure that your smoke tests definitions are changed to use the shell library so that all health checks and functional tests provide test coverage for the shell library.

Check for specific support as a test case

If a particular package, service, script or utility must exist and / or function for the rest of your test definition to operate, test for this functionality.

Any command executed by lava-test-case <name> --shell will report a test case as pass if that command exits zero and fail if that command exited non-zero. If the command is complex or needs pipes or redirects, create a simple script which returns the exit code of the command.

Note

remember that the test shell runs under set -e, so if you need to prevent the rest of a test definition from exiting, you can report a non-zero exit code from your scripts and call the script directly instead of as a test case.

Check custom scripts for side-effects

Subtle bugs can be introduced in custom scripts, so it is important to make the scripts portable so that bugs can be reproduced outside LAVA.

When interacting directly with LAVA, for example calling lava-test-case, it is possible to introduce control flow bugs. These can cause the output of lava-test-case to be received after the end of a test run and this can generate TestError exceptions. This section covers one example when using Python, there may be others.

This example checks for lava-test-case in $PATH to determine whether to use the LAVA helpers.

import os
import subprocess


def _which_check(path, match):
    """
    Simple replacement for the `which` command found on
    Debian based systems. Allows ordinary users to query
    the PATH used at runtime.
    """
    paths = os.environ['PATH'].split(':')
    if os.getuid() != 0:
        # avoid sudo - it may ask for a password on developer systems.
        paths.extend(['/usr/local/sbin', '/usr/sbin', '/sbin'])
    for dirname in paths:
        candidate = os.path.join(dirname, path)
        if match(candidate):
            return candidate
    return None


if _which_check(path='lava-test-case', match=os.path.isfile):
    subprocess.Popen([
         'lava-test-case', 'probe-results', '--result', 'pass',
         '--measurement', str(average), '--units', 'volts'])

The error is in this line:

subprocess.Popen([

Popen calls fork but returns immediately. Unless the script also calls wait, then the output of the subprocess can occur after the above function has returned. It is easy for this to happen at the end of a test definition, leading to intermittent bugs where some tests fail.

The solution is to use the existing subprocess functions which already use wait internally. For lava-test-case, this would be check_call which waits for the process to execute and checks the return value.

The fixed example looks like:

import os
import subprocess


def _which_check(path, match):
    """
    Simple replacement for the `which` command found on
    Debian based systems. Allows ordinary users to query
    the PATH used at runtime.
    """
    paths = os.environ['PATH'].split(':')
    if os.getuid() != 0:
        # avoid sudo - it may ask for a password on developer systems.
        paths.extend(['/usr/local/sbin', '/usr/sbin', '/sbin'])
    for dirname in paths:
        candidate = os.path.join(dirname, path)
        if match(candidate):
            return candidate
    return None


if _which_check(path='lava-test-case', match=os.path.isfile):
    subprocess.check_call([
         'lava-test-case', 'probe-results', '--result', 'pass',
         '--measurement', str(average), '--units', 'volts'])

Call lava-test-raise if setup fails

Most test jobs have setup routines which ensure that dependencies are available or that the directory layout is correct etc. In most cases, these routines are called early and a failure in the setup function would undermine all subsequent test operations.

The return code of some operations can be used to trigger an early failure.

Inline

If you are using an inline definition, the syntax can be a bit awkward:

run:
   steps:
       - apt-get update -q && lava-test-case "apt-update" --result pass || lava-test-raise "apt-update"

An alternative is to put the definition into a file on a remote fileserver, use wget to download it and then execute it:

run:
  steps:
    - apt -y install wget
    - wget http://people.linaro.org/~neil.williams/setup-test.sh
    - sh -x setup-test.sh

Caution

The download step is itself a setup command and could fail, so whilst this is useful in development, using scripts from a git repository is preferable.

Using a repository

Shell library

A local shell library and a shell script can be easily used from a test shell repository:

# saved, committed and pushed as ./testdefs/lava-common

command(){
    if [ -n "$(which lava-test-case || true)" ]; then
        echo $2
        $2 && lava-test-case "$1" --result pass || lava-test-raise "$1"
    else
        echo $2
        $2
    fi
}

This snippet is also portable because if lava-test-case is not in the $PATH, the setup command is executed without needing lava-test-case or lava-test-raise. The calling script is responsible for handling the return code, typically by using set -e.

The above snippet is just an example to show the principle. The function itself continues to develop as lava-common - a small shell library which also supports a testcase function which reports a failed test case instead of lava-test-raise. Use testcase for non-fatal checks and command for fatal checks.

command(){
    # setup command - will abort the test job upon failure.
    # expects two quoted arguments
    # $1 - valid lava test case name (no spaces)
    # $2 - the full command line to execute
    # Note: avoid trying to set environment variables.
    # use an explicit export.
    CMD=""
    PREFIX=$1
    shift
    while [ "$1" != "" ]; do
      CMD="${CMD} $1" && shift;
    done;
    if [ -n "$(which lava-test-case || true)" ]; then
        echo "${CMD}"
        $CMD && lava-test-case "${PREFIX}" --result pass || lava-test-raise "${PREFIX}"
    else
        echo "${CMD}"
        $CMD
    fi
}

Calling shell script

#!/bin/sh

# saved, committed and pushed as ./testdefs/local-run.sh

. ./lava-common

command 'setup-apt' "apt-get update -q"

If the shell script is saved to a different directory, the path to the shell library will have to be updated.

See also

Custom scripts - the language used for these scripts is entirely up to the test writer to choose. Remember that some language interpreters will themselves need to be installed before scripts can be executed, requiring an initial setup shell script. That does not mean that all setup needs to be done in shell; there are key advantages to using other languages, including test writer familiarity and ease of triage.

Test shell definition

Execute using a Lava Test Shell Definition:

run:
    steps:
      ./testdefs/local-run.sh

Custom scripts

Custom scripts should check the return code of setup operations and use lava-test-raise to halt the test job immediately if a setup error occurs. This makes triage much easier as it puts the failure much closer to the actual cause within the log file.

import os
import subprocess


def _which_check(path, match):
    """
    Simple replacement for the `which` command found on
    Debian based systems. Allows ordinary users to query
    the PATH used at runtime.
    """
    paths = os.environ['PATH'].split(':')
    if os.getuid() != 0:
        # avoid sudo - it may ask for a password on developer systems.
        paths.extend(['/usr/local/sbin', '/usr/sbin', '/sbin'])
    for dirname in paths:
        candidate = os.path.join(dirname, path)
        if match(candidate):
            return candidate
    return None


values = []
# other processing populates the values list
if not values:
    if _which_check(path='lava-test-raise', match=os.path.isfile):
        subprocess.check_call(['lava-test-raise', 'setup failed'])
    else:
        print("setup failed")
    return 1

Control the amount of output from scripts and tools

Many tools available in distributions have ways to control the amount of output during operation. A balance is needed and test writers are recommended to check for available support. Wherever possible, use the available options to opt for output intended for log file output rather than your typical terminal.

When writing your own scripts, consider using set -x or wrapping certain blocks with set -x, set +x when using shell scripts. With other languages, use print() and similar functions often, especially where the script uses a conditional that can be affected by parameters from within the test job.

Specific tools

Progress bars, in general, are a particular problem. Instead of overwriting a single line of output, every iteration of the bar creates a complete new line over the serial connection and in the logs. Wherever possible, disable the progress bar behavior of all operations.

  • apt - When calling apt update or apt-get update, always use the -q option to avoid filling the log file with repeated progress output during downloads. This option still gives output but formats it in a way that is much more useful when reading log files compared to an interactive terminal.
  • wget - always use the -S --progress=dot:giga options for downloads as this reduces the total amount of progress information during the operation.
  • git clone - consider using -q on git clone operations to silence the progress bars.

Problems with output

LAVA uses pexpect to monitor the output over the serial connection for patterns which are used to pick up test cases and other test shell support. Each time a match is found, the buffer is cleared. If there is a lot of output with no pattern matches, the processing can slow down.

By default pexpect uses a buffer of 2000 bytes for the input used for pattern matches. In order to improve performance, LAVA uses a limit of 4092 bytes. This is intended to limit problems with processing slowing down but best practice remains to manage the test job output to make the logs more useful during later triage.

Large log files also have a few implications for the user interface and triage. More content makes loading links to a test job take longer and finding the right line to make that link becomes more and more difficult. Eventually, very large log files can be disabled by the admin, so that the log file can only be downloaded.

The size of the log output needs to be balanced against the need to have enough information in the logs to be able to triage the test successfully.

Although the total size of the test job log file is important, there can also be issues when a smaller log file contains large sections where none of the patterns match and this can cause the test to run more slowly.

Important

It is only the content sent over the serial connection which needs to be managed. Redirecting to files will be unaffected, subject to filesystem performance on the DUT or LXC. However, remember that at least some of the content of such files will be useful in triage or contain results directly. Therefore, it is important to manage the output of test operations to achieve the balance of sufficient information for triage and avoiding a flood of too much information causing performance issues.

Very large amounts of output can also be published for later analysis, e.g. if the original output is redirected to a file. Consider using tee here (or similar functionality) to retain some output into the logs because if the test operation fails early for any reason, the file might not be uploaded at all.

When performance is important, for example benchmarking, use a wrapper script to optimize your test shell output.

  • If a progress bar is used and cannot be turned off without losing other useful content, wrap the output of the command in a script which omits the lines generated by the progress bar. Check existing test logs for example lines and print all the other lines. Avoid the simplistic approach of redirecting to /dev/null.

    For a progress bar which outputs lines looking like: [ 98%] /data/art-test/arm64/core.oat: 95%

    Use something like this:

    #!/usr/bin/env python
    
    import fileinput
    
    def main(args):
        for line in fileinput.input('-'):
            line = line.strip()
            if line.startswith('[') and line.endswith('%'):
                continue
            print(line)
        return 0
    
    if __name__ == '__main__':
        import sys
        sys.exit(main(sys.argv))
    

    Adapted from https://git.linaro.org/lava-team/refactoring.git/tree/functional/unittests.py

    The same script can be used to drop other noise from the output.

  • Add LAVA Test Cases - avoid the habit of reporting results at the very end of a test operation or (worse) test job. This risks getting no results at all when things go wrong, as well as creating large amounts of output without any pattern matches. Most tests run many small test operations, it can be helpful to have records of which tests completed. Remember that a test set can be used to identify groups of test cases, isolating them from later test cases.

    Example: Rely less on install: steps means that after all of the output of installing dependencies, a lava-test-case should be reported that the dependencies installed correctly which also clears the buffer of the extra output.

    Example: If the test operation involves iterations over a test condition, report a lava test case every few iterations.

Control the number of test cases reported

Creating a lava-test-case involves a database operation on the master. LAVA tries to optimize these calls to allow test jobs to report several tens of thousands of test cases per test job, including supporting streaming of test cases exported through the API. However, there will always be a practical limit to the total number of test cases per test job.

Groups of test cases should be separated into test sets and then into test suites (by using separate LAVA Test Shell Definition paths) to make it easier to find the relevant test case.

When writing the test shell definition, always try to report results on-the-fly instead of waiting until the test operation has written all the data to a file. This insulates you from early failures where the file is not written or cannot be parsed after being written. Wrapper scripts can be used to report LAVA test cases during the creation of the file.