Lava Dispatcher Design¶

This is the developer documentation for the new V2 dispatcher design. See Advanced Use Cases for information for lab administrators and users of the new design.

The refactoring takes place alongside the V1 dispatcher and existing JSON jobs are unaffected. A migration will take place where individual devices are configured for pipeline support and individual jobs are then re-written using the pipeline_schema. The administrator of each instance will be able to manage their own migration and at some point after validation.linaro.org has completed the migration of all devices to pipeline support, the support for the current dispatcher will be removed.

validation.linaro.org supports LAVA V2 pipeline submissions as of the 2016.2 release and the V2 support will continue to expand in subsequent releases.

The LAVA developers use a playground instance for testing of selected changes prior to merging into master. There is also a staging instance for testing of the current master branch and release candidates for the next production release.

Devices indicate their support for pipeline jobs in the detailed device information for each device and device type.

Pipeline Architecture¶

Principal changes¶

Database isolation - Only the master daemon has a connection to the database. This simplifies the architecture and avoids the use of fault-intolerant database connections to remote workers.
Drop use of SSHFS between workers and master - this was awkward to configure and problematic over external connections.
Move configuration onto the master - The worker becomes a simple slave which receives all configuration and tasks from the master.

Objectives¶

The new dispatcher design is intended to make it easier to adapt the dispatcher flow to new boards, new mechanisms and new deployments. It also shifts support to do less work on the dispatcher, make fewer assumptions about the test in the dispatcher configuration and put more flexibility into the hands of the test writer.

Note

The new code is still developing, some areas are absent, some areas will change substantially before the migration completes. There may be changes to the submission formats but these will be announced on the lava-announce mailing list.

From 2015.8 onwards the sample jobs supporting the unit tests conform to the LAVA schema.

Design¶

Start with a Job which is broken up into a Deployment, a Boot and a Test class. Results are transmitted live during any part of the job.


Deployment
	DeployAction
		DownloadAction
		ChecksumAction
		MountAction
		CustomiseAction
		TestDefAction
		UnmountAction
	BootAction
	TestAction

The Job manages the Actions using a Pipeline structure. Actions can specialise actions by using internal pipelines and an Action can include support for retries and other logical functions:

DownloadAction
	HttpDownloadAction
	FileDownloadAction

If a Job includes one or more Test definitions, the Deployment can then extend the Deployment to overlay the LAVA test scripts without needing to mount the image twice:

DeployAction
	OverlayAction
		MultinodeOverlayAction
		LMPOverlayAction

The TestDefinitionAction has a similar structure with specialist tasks being handed off to cope with particular tools:

TestDefinitionAction
	RepoAction
		GitRepoAction
		BzrRepoAction
		TarRepoAction
		UrlRepoAction

Following the code flow¶

Filename	Role
lava/dispatcher/commands.py	Command line arguments, call to YAML parser
lava_dispatcher/device.py	YAML Parser to create the Device object
lava_dispatcher/parser.py	YAML Parser to create the Job object
....actions/deploy/	Handlers for different deployment strategies
....actions/boot/	Handlers for different boot strategies
....actions/test/	Handlers for different LavaTestShell strategies
....actions/deploy/image.py	DeployImages strategy creates DeployImagesAction
....actions/deploy/image.py	DeployImagesAction.populate adds deployment actions to the Job pipeline
*repeat for each strategy*	each `populate` function adds more Actions
....action.py	`Pipeline.run_actions()` to start

The deployment is determined from the device_type specified in the Job (or the device_type of the specified target) by reading the list of support methods from the device_types YAML configuration.

Each Action can define an internal pipeline and add sub-actions in the Action.populate function.

Particular Logic Actions (like RetryAction) require an internal pipeline so that all actions added to that pipeline can be retried in the same order. (Remember that actions must be idempotent.) Actions which fail with a JobError or InfrastructureError can trigger Diagnostic actions. See Logical actions.

actions:
  deploy:
    allow:
      - image
  boot:
    allow:
      - image

This then matches the python class structure:

actions/
   deploy/
       image.py

The class defines the list of Action classes needed to implement this deployment. See also Dispatcher Action Reference.

Pipeline construction and flow¶

The pipeline is a FIFO and has branches which are handled as a tree walk. The top level object is the job, based on the YAML definition supplied by the lava-master. The definition is processed by the scheduler and the submission interface with information specific to the actual device. The processed definition is parsed to generate the top level pipeline and strategy classes. Each strategy class adds a top level action to the top level pipeline. The top level action then populates branches containing more actions.

Actions are populated, validated and executed in strict order. The next action in any branch waits until all branches of the preceding action have completed. Populating an action in a pipeline creates a level string, e.g. all actions in level 1.2.1, including all actions in sublevel 1.2.1.2 are executed before the pipeline moves on to processing level 1.3 or 2:

Deploy (1)
   |
   \___ 1.1
   |
   \ __ 1.2
   |     |
   |     \_ 1.2.1
   |     |   |
   |     |   \_ 1.2.1.1
   |     |   |
   |     |   \_ 1.2.1.2
   |     |         |
   |     |         \__ 1.2.1.2.1
   |     |
   |     \__1.2.2
   |
   \____1.3
   |
  Boot (2)
   |
   \_ 2.1
   |
   \_ 2.2

One device per job. One top level pipeline per job
- loads only the configuration required for this one job.
A NewDevice is built from the target specified (commands.py)
A Job is generated from the YAML by the parser.
The top level Pipeline is constructed by the parser.
Strategy classes are initialised by the parser
1. Strategy classes add the top level Action for that strategy to the top level pipeline.
2. Top level pipeline calls populate() on each top level Action added.
  1. Each Action.populate() function may construct one internal pipeline, based on parameters.
  2. internal pipelines call populate() on each Action added.
  3. A sublevel is set for each action in the internal pipeline. Level 1 creates 1.1 and level 2.3.2 creates 2.3.2.1.
Parser waits while each Strategy completes branch population.
Parser adds the FinalizeAction to the top-level pipeline
Loghandlers are set up
Job validates the completed pipeline
1. Dynamic data can be added to the context
If --validate not specified, the job runs.
1. Each run() function can add dynamic data to the context and/or results to the pipeline.
2. Pipeline walks along the branches, executing actions.
Job ends, check for errors
Completed pipeline is available.

Using strategy classes¶

Strategies are ways of meeting the requirements of the submitted job within the limits of available devices and code support.

If an internal pipeline would need to allow for optional actions, those actions still need to be idempotent. Therefore, the pipeline can include all actions, with each action being responsible for checking whether anything actually needs to be done. The populate function should avoid using conditionals. An explicit select function can be used instead.

Whenever there is a need for a particular job to use a different Action based solely on job parameters or device configuration, that decision should occur in the Strategy selection using classmethod support.

Where a class is used in lots of different strategies, identify whether there is a match between particular strategies always needing particular options within the class. At this point, the class can be split and particular strategies use a specialised class implementing the optional behaviour and calling down to the base class for the rest.

If there is no clear match, for example in testdef.py where any particular job could use a different VCS or URL without actually being a different strategy, a select function is preferable. A select handler allows the pipeline to contain only classes supporting git repositories when only git repositories are in use for that job.

The list of available strategies can be determined in the codebase from the module imports in the strategies.py file for each action type.

This results in more classes but a cleaner (and more predictable) pipeline construction.

Lava test shell scripts¶

Note

See LAVA review criteria - it is a mistake to think of the LAVA test support scripts as an overlay - the scripts are an extension to the test. Wherever possible, current deployments are being changed to supply the extensions alongside the deployment instead of overlaying, and thereby altering, the deployment.

The LAVA scripts are a standard addition to a LAVA test and are handled as a single unit. Using idempotent actions, the test script extension can support LMP or MultiNode or other custom requirements without requiring this support to be added to all tests. The extensions are created during the deploy strategy and specific deployments can override the ApplyExtensionAction to unpack the extension tarball alongside the test during the deployment phase and then mount the extension inside the image. The tarball itself remains in the output directory and becomes part of the test records. The checksum of the overlay is added to the test job log.

Pipeline error handling¶

RuntimeError Exception¶

Runtime errors include:

Parser fails to handle device configuration
Parser fails to handle submission YAML
Parser fails to locate a Strategy class for the Job.
Code errors in Action classes cause Pipeline to fail.
Errors in YAML cause errors upon pipeline validation.

Each runtime error is a bug in the code - wherever possible, implement a unit test to prevent regressions.

InfrastructureError Exception¶

Infrastructure errors include:

Missing dependencies on the dispatcher
Device configuration errors

JobError Exception¶

Job errors include:

Failed to find the specified URL.
Failed in an operation to create the necessary extensions.

TestError Exception¶

Test errors include:

Failed to handle a signal generated by the device
Failed to parse a test case

Result bundle identifiers¶

Old style result bundles are assigned a text based UUID during submission. This has several issues:

The UUID is not sequential or predictable, so finding this one, the next one or the previous one requires a database lookup for each. The new dispatcher model will not have a persistent database connection.
The UUID is not available to the dispatcher while running the job, so cannot be cross-referenced to logs inside the job.
The UUID makes the final URL of individual test results overly long, unmemorable and complex, especially as the test run is also given a separate UUID in the old dispatcher model.

The new dispatcher creates a pipeline where every action within the pipeline is guaranteed to have a unique level string which is strictly sequential, related directly to the type of action and shorter than a UUID. To make a pipeline result unique on a per instance basis, the only requirement is that the result includes the JobID which is a sequential number, passed to the job in the submission YAML. This could also have been a UUID but the JobID is already a unique ID for this instance.

When bundles are downloaded, the database query will need to assign a UUID to that downloaded file but the file will also include the job number and the query can also insert the source of the bundle in a comment in the YAML. This will allow bundles to be uploaded to a different instance using lava-tool without the risk of collisions. It is also possible that the results could provide a link back to the original job log file and other data - if the original server is visible to users of the server to which the bundle was later uploaded.

LAVA review criteria¶

These criteria are to help developers control the development of new code. Any of these criteria can be cited in a code review as reasons for a review to be improved.

Keep the dispatcher dumb¶

There is a temptation to make the dispatcher clever but this only restricts the test writer from doing their own clever tests by hard coding commands into the dispatcher codebase. If the dispatcher needs some information about the test image, that information must be retrieved from the job submission parameters, not by calculating in the dispatcher or running commands inside the test image. Exceptions to this are the metrics already calculated during download, like file size and checksums. Any information about the test image which is permanent within that image, e.g. the partition UUID strings or the network interface list, can be identified by the process creating that image or by a script which is run before the image is compressed and made available for testing. If a test uses a tarball instead of an image, the test must be explicit about the filesystem to use when unpacking that tarball for use in the test as well as the size and location of the partition to use.

LAVA will need to implement some safeguards for tests which still need to deploy any test data to the media hosting the bootloader (e.g. fastboot, SD card or UEFI) in order to avoid overwriting the bootloader itself. Therefore, although SD card partitions remain available for LAVA tests where no other media are supportable by the device, those tests can only use tarballs and pre-defined partitions on the SD card. The filesystem to use on those partitions needs to be specified by the test writer.

Avoid defaults in dispatcher code¶

Constants and defaults are going to need an override somewhere for some device or test, eventually. Code defensively and put constants into the utilities module to support modification. Put defaults into the YAML, not the python code. It is better to have an extra line in the device_type than a string in the python code as this can later be extended to a device or a job submission.

Let the test fail and diagnose later¶

Avoid guessing in LAVA code. If any operation in the dispatcher could go in multiple paths, those paths must be made explicit to the test writer. Report the available data, proceed according to the job definition and diagnose the state of the device afterwards, where appropriate.

Avoid trying to be helpful in the test image. Anticipating an error and trying to code around it is a mistake. Possible solutions include but are not limited to:

Provide an optional, idempotent, class which only acts if a specific option is passed in the job definition. e.g. AutoLoginAction.
Provide a diagnostic class which triggers if the expected problem arises. Report on the actual device state and document how to improve the job submission to avoid the problem in future.
Split the deployment strategy to explicitly code for each possible path.

AutoLogin is a good example of the problem here. For too long, LAVA has made assumptions about the incoming image, requiring hacks like linaro-overlay packages to be added to basic bootstrap images or disabling passwords for the root user. These helpful steps act to make it harder to use unchanged third party images in LAVA tests. AutoLogin is the de facto default for non-Linaro images.

Another example is the assumption in various parts of LAVA that the test image will raise a network interface and repeatedly calling ping on the assumption that the interface will appear, somehow, eventually.

Treat the deployment as a black box¶

LAVA has claimed to do this for a long time but the refactored dispatcher is pushing this further. Do not think of the LAVA scripts as an overlay, the LAVA scripts are extensions. When a test wants an image deployed, the LAVA extensions should be deployed alongside the image and then mounted to create a /lava-$hostname/ directory. Images for testing within LAVA are no longer broken up or redeployed but must be deployed intact. This avoids LAVA needing to know anything about issues like SELinux or specific filesystems but may involve multiple images for systems like Android where data may exist on different physical devices.

Only protect the essential components¶

LAVA has had a tendency to hardcode commands and operations and there are critical areas which must still be protected from changes in the test but these critical areas are restricted to:

The dispatcher.
Unbricking devices.

Any process which has to run on the dispatcher itself must be fully protected from mistakes within tests. This means that all commands to be executed by the dispatcher are hardcoded into the dispatcher python code with only limited support for overriding parameters or specifying tainted user data.

Tests are prevented from requiring new software to be installed on any dispatcher which is not already a dependency of lava-dispatcher. Issues arising from this need to be resolved using MultiNode.

Until such time as there is a general and reliable method of deploying and testing new bootloaders within LAVA tests, the bootloader / firmware installed by the lab admin is deemed sacrosanct and must not be altered or replaced in a test job. However, bootloaders are generally resilient to errors in the commands, so the commands given to the bootloader remain accessible to test writers.

It is not practical to scan all test definitions for potentially harmful commands. If a test inadvertently corrupts the SD card in such a way that the bootloader is corrupted, that is an issue for the lab admins to take up with the test submitter.

Give the test writer enough rope¶

Within the provisos of Only protect the essential components, the test writer needs to be given enough rope and then let LAVA diagnose issues after the event.

There is no reason to restrict the test writer to using LAVA commands inside the test image - as long as the essential components remain protected.

Examples:

KVM devices need to protect the QEMU command line because these commands run on the dispatcher
VM devices running on an arndale do not need the command line to be coded within LAVA. There have already been bug reports on this issue.

Diagnostic subclasses report on the state of the device after some kind of error. This reporting can include:

The presence or absence of expected files (like /dev/disk/by-id/ or /proc/net/pnp).
Data about running processes or interfaces, e.g. ifconfig

It is a mistake to attempt to calculate data about a test image - instead, require that the information is provided and diagnose the actual information if the attempt to use the specified information fails.

Guidance¶

If the command is to run inside a deployment, require that the full command line can be specified by the test writer. Remember: Avoid defaults in dispatcher code. It is recommended to have default commands where appropriate but these defaults need to support overrides in the job submission. This includes using a locally built binary instead of an executable installed in /usr/bin or similar.
If the command is run on a dispatcher, require that the binary to be run on the dispatcher is actually installed on the dispatcher. If /usr/bin/git does not exist, this is a validation error. There should be no circumstances where a tool required on the dispatcher cannot be identified during validation of the pipeline.
An error from running the command on the dispatcher with user-specified parameters is a JobError.
Where it is safe to do so, offer overrides for supportable commandline options.

The codebase itself will help identify how much control is handed over to the test writer. self.run_command() is a dispatcher call and needs to be protected. connection.sendline() is a deployment call and does not need to be protected.

Providing gold standard images¶

Test writers are strongly recommended to only use a known working setup for their job. A set of gold standard jobs has been defined in association with the QA team. These jobs will provide a known baseline for test definition writers, in a similar manner as the existing QA test definitions provide a base for more elaborate testing.

Creating a gold standard image¶

Part of the benefit of a standard image is that the methods for building the image - and therefore the methods for updating it, modifying it and preparing custom images based upon it - must be documented clearly.

Where possible, standard tools familiar to developers of the OS concerned should be used, e.g. debootstrap for Debian based images. The image can also be a standard OS installation. Gold standard images are not “Linaro” images and should not require Linaro tools. Use AutoLogin support where required instead of modifying existing images to add Linaro-specific tools.

All gold standard images need to be kept up to date with the base OS as many tests will want to install extra software on top and it will waste time during the test if a lot of other packages need to be updated at the same time. An update of a gold standard image still needs to be tested for equivalent or improved performance compared to the current image before replacing it.

The documentation for building and updating the image needs to be provided alongside the image itself as a README. This text file should also be reproduced on a wiki page and contain a link to that page. Any wiki can be used - if a suitable page does not already exist elsewhere, use wiki.linaro.org.

Other gold standard components¶

The standard does not have to be a complete OS image - a kernel with a DTB (and possibly an initrd) can also count as a standard ramdisk image. Similarly, a combination of kernel and rootfs can count as a standard NFS configuration.

The same requirement exists for documenting how to build, modify and update all components of the “image” and the set of components need to be tested as a whole to represent a test using the standard.

In addition, information about the prompts within the image needs to be exposed. LAVA no longer has a list of potential prompts and each job must specify a list of prompts to use for the job.

Other information should also be provided, for example, memory requirements or CPU core requirements for images to be used with QEMU or dependencies on other components (like firmware or kernel support).

Test writers need to have enough information to submit a job without needing to resubmit after identifying and providing missing data.

One or more sample test jobs is one way of providing this information but it is still recommended to provide the prompts and other information explicitly.

Secondary media¶

With the migration from master images on an SD card to dynamic master images over NFS, other possibilities arise from the refactoring.

Deploy a ramdisk, boot and deploy an entire image to a USB key, boot and direct bootloader at USB filesystem, including kernel and initrd.
Deploy an NFS system, boot and bootstrap an image to SATA, boot and direct bootloader at SATA filesystem, including kernel and initrd.
Deploy using a script written by the test author (e.g. debootstrap) which is installed in the initial deployment. Parameters for the script need to be contained within the test image.

Device configuration design¶

Device configuration, as received by lava_dispatch has moved to YAML and the database device configuration has moved to Jinja2 templates. This method has a much larger scope of possible methods, related to the pipeline strategies as well as allowing simple overrides and reuse of common device configuration stanzas.

There is no need for the device configuration to include the hostname in the YAML as there is nothing on the dispatcher to check against - the dispatcher uses the command line arguments and the supplied device configuration. The configuration includes all the data the dispatcher needs to be able to run the job on the device attached to the specified ports.

The device type configuration on the dispatcher is replaced by a device type template on the server which is used to generate the YAML device configuration sent to the dispatcher.

Device Dictionary¶

The normal admin flow for individual devices will be to make changes to the device dictionary of that device. In time, an editable interface will exist within the admin interface. Initially, changes to the dictionary are made from the command line with details being available in a read-only view in the admin interface.

The device dictionary acts as a set of variables inside the template, in a very similar manner to how Django handles HTML templates. In turn, a device type template will extend a base template.

It is a bug in the template if a missing value causes a broken device configuration to be generated. Values which are not included in the specified template will be ignored.

Once the device dictionary has been populated, the scheduler can be told that the device is a pipeline device in the admin interface.

Note

Several parts of this process still need helpers and tools or may give unexpected errors - there is a lot of ongoing work in this area.

Exporting an existing device dictionary¶

If the local instance has a working pipeline device called mypanda, the device dictionary can be exported as a Jinja2 child template which extends a device type jinja template:

$ sudo lava-server manage device-dictionary --hostname mypanda --export
{% extends 'panda.jinja2' %}
{% set power_off_command = '/usr/bin/pduclient --daemon tweetypie --hostname pdu --command off --port 08' %}
{% set hard_reset_command = '/usr/bin/pduclient --daemon tweetypie --hostname pdu --command reboot --port 08' %}
{% set connection_list = [‘uart0’] %}
{% set connection_commands = {‘uart0’: ‘telnet dispatcher01 7001’} %}
{% set connection_tags = {‘uart0’: [‘primary’, 'telnet']} %}
{% set power_on_command = '/usr/bin/pduclient --daemon tweetypie --hostname pdu --command on --port 08' %}

This dictionary declares that the device inherits the rest of the device configuration from the panda device type. Settings specific to this one device are then specified.

Reviewing an existing device dictionary¶

To populate the full configuration using the device dictionary and the associated templates, use the review option:

$ sudo lava-server manage device-dictionary --hostname mypanda --review

Example device configuration review¶

device_type: beaglebone-black
commands:
  connect: telnet localhost 6000
  hard_reset: /usr/bin/pduclient --daemon localhost --hostname pdu --command reboot --port 08
  power_off: /usr/bin/pduclient --daemon localhost --hostname pdu --command off --port 08
  power_on: /usr/bin/pduclient --daemon localhost --hostname pdu --command on --port 08

parameters:
 bootm:
  kernel: '0x80200000'
  ramdisk: '0x81600000'
  dtb: '0x815f0000'
 bootz:
  kernel: '0x81000000'
  ramdisk: '0x82000000'
  dtb: '0x81f00000'

actions:
 deploy:
   # list of deployment methods which this device supports
   methods:
     # - image # not ready yet
     - tftp

 boot:
   # list of boot methods which this device supports.
   methods:
     - u-boot:
         parameters:
           bootloader_prompt: U-Boot
           boot_message: Booting Linux
           send_char: False
           # interrupt: # character needed to interrupt u-boot, single whitespace by default
         # method specific stanza
         oe:
           commands:
           - setenv initrd_high '0xffffffff'
           - setenv fdt_high '0xffffffff'
           - setenv bootcmd 'fatload mmc 0:3 0x80200000 uImage; fatload mmc 0:3 0x815f0000 board.dtb;
             bootm 0x80200000 - 0x815f0000'
           - setenv bootargs 'console=ttyO0,115200n8 root=/dev/mmcblk0p5 rootwait ro'
           - boot
         nfs:
           commands:
           - setenv autoload no
           - setenv initrd_high '0xffffffff'
           - setenv fdt_high '0xffffffff'
           - setenv kernel_addr_r '{KERNEL_ADDR}'
           - setenv initrd_addr_r '{RAMDISK_ADDR}'
           - setenv fdt_addr_r '{DTB_ADDR}'
           - setenv loadkernel 'tftp ${kernel_addr_r} {KERNEL}'
           - setenv loadinitrd 'tftp ${initrd_addr_r} {RAMDISK}; setenv initrd_size ${filesize}'
           - setenv loadfdt 'tftp ${fdt_addr_r} {DTB}'
           # this could be a pycharm bug or a YAML problem with colons. Use &#58; for now.
           # alternatively, construct the nfsroot argument from values.
           - setenv nfsargs 'setenv bootargs console=ttyO0,115200n8 root=/dev/nfs rw nfsroot={SERVER_IP}&#58;{NFSROOTFS},tcp,hard,intr ip=dhcp'
           - setenv bootcmd 'dhcp; setenv serverip {SERVER_IP}; run loadkernel; run loadinitrd; run loadfdt; run nfsargs; {BOOTX}'
           - boot
         ramdisk:
           commands:
           - setenv autoload no
           - setenv initrd_high '0xffffffff'
           - setenv fdt_high '0xffffffff'
           - setenv kernel_addr_r '{KERNEL_ADDR}'
           - setenv initrd_addr_r '{RAMDISK_ADDR}'
           - setenv fdt_addr_r '{DTB_ADDR}'
           - setenv loadkernel 'tftp ${kernel_addr_r} {KERNEL}'
           - setenv loadinitrd 'tftp ${initrd_addr_r} {RAMDISK}; setenv initrd_size ${filesize}'
           - setenv loadfdt 'tftp ${fdt_addr_r} {DTB}'
           - setenv bootargs 'console=ttyO0,115200n8 root=/dev/ram0 ip=dhcp'
           - setenv bootcmd 'dhcp; setenv serverip {SERVER_IP}; run loadkernel; run loadinitrd; run loadfdt; {BOOTX}'
           - boot

Importing configuration using a known template¶

To add or update the device dictionary, a file using the same syntax as the export content can be imported into the database:

$ sudo lava-server manage device-dictionary --hostname mypanda --import mypanda.yaml

(The file extension is unnecessary and the content is not actually YAML but will be rendered as YAML when the templates are used.)

Creating a new template¶

Start with the base.yaml template and use the structure of that template to ensure that your template remains valid YAML.

Start with a complete device configuration (in YAML) which works on the lava-dispatch command line, then iterate over changes in the template to produce the same output.

Note

A helper is being planned for this step.

Running lava-dispatch directly¶

lava-dispatch only accepts a YAML file for pipeline jobs - the old behaviour of looking up the file based on the device hostname has been dropped. The absolute or relative path to the YAML file must be specified to the --target option. --output-dir must also be specified:

sudo lava-dispatch --target devices/fred.conf panda-ramdisk.yaml --output-dir=/tmp/test