LAVA
2021.01
When adding devices, there is a decision to be made about what qualifies as a distinct device type and what does not. A different type of device can mean different things, including:
However, a slight or incremental change to hardware does not necessarily mean that the updated device must be a different device type. Even if a change adds significant functionality, e.g. if USB hot-plug becomes available on revision C, that does not necessarily mean that revision C should be a different device type compared to to revision A and B.
Typically, the distinction between two device types comes down to whether the two devices can be driven in the same way at bootloader level, from initial power on.
Another example to consider is DTB support. If there is more than one DTB available for a particular family of devices, this will probably lead to multiple different device types. Also consider whether all devices of the proposed device type can boot all DTBs available for that type.
In the end, LAVA lab administrators are free to make their own choices about what qualifies as a distinct device type. Some factors to consider in this judgment call include:
Interchangeable jobs: Does a single health check work for all devices of this type? It is recommended to always test all of the supported boot methods of a device type during a single health check test job.
Interchangeable bootloaders: Some devices can change bootloader type within a single job, allowing a single device type to meet the needs of a variety of different use cases. There are issues to consider here:
Latency: Changing the bootloader on every test job may have significant costs in terms of job runtime. This is particularly noticeable in the length of time required before the actual test can start. In such cases, it may be worth considering a sub-type instead.
See also
Hardware lifetime of the device: Frequently, writing a new bootloader may cause problems on some devices where the bootloader may be stored on media which can only be written a limited number of times.
Multi-stage bootloaders: Some devices may have a first or second stage bootloader which can then load different higher level bootloaders. This is often described as a chained bootloader. Depending on the types of tests desired, some admins may choose to expose a choice of higher level bootloader or may choose to not allow interrupting the lower stage(s). For example, some test writers will want to test the firmware and some test writers will only want to interact with GRUB or later.
Equivalence: Different labs may make different decisions - if you are looking to work with an existing lab, try to follow their device type layout and ask about how a new device should be classified before committing to a decision in your own lab.
Test requirements: Talk to the test writers and establish whether an apparent hardware difference is sufficient that the device needs to be described using a different device type. Also consider whether the test writer requirements are going to change over time. Just because there is no current desire to test the experimental bootloader available on one device, this does not mean that this will remain unused in a year.
Scheduling: If two devices have the same device type, each device needs to be able to run any test job submitted for that device type. There is support within LAVA for encoding small differences between devices within a single device type - see device tag.
LAVA support: There are some considerations which are
constrained by LAVA support, For example, the V1 dispatcher had a
kvm device type but the improved device configuration design in
V2 made this unnecessary. If the only factor requiring two device
types is LAVA support, please talk to us.
Once a device type has been implemented, devices added and test jobs run, it can be awkward to change the device type. Changing the device type later will make it difficult for users to find test results across this and other devices and may cause significant issues with data consistency.
Separate device types can also complicate queries and result reporting - combining two devices which eventually end up being different device types causes issues with a loss of history when the split is finally made.
It is not a good idea to split device types arbitrarily - sooner or later there may be a requirement to look at the results of jobs across both types and having an unnecessary device type is confusing for test writers. Use a device tag to describe small differences between devices of the same device type where it makes sense.
See also
A balance needs to be drawn between test jobs which simply want to use a known working build of the firmware and/or bootloader and those test jobs where the latest build is relevant to the success or failure of the test itself. Different test writers may have different requirements here.
An example of sub-types could be juno-uboot compared to
juno-firmware. Consider the principle of test one thing at a
time. Let userspace test jobs run without needing to change the
bootloader, and let bootloader test jobs have the ability to update by
separating the device type into two sub-types.
Think about device integration here. You need to be able to interrupt the boot process at a level below whatever you are exposing to test writers. For example, to offer test writers the ability to modify and test the firmware, the platform must offer a way to replace the firmware in an automatable manner.
There are some considerations for the names of a device type in LAVA.
panda or panda-es instead of
panda1 or panda2.The name of a device type does not need to match an available template, however
the device dictionary for all devices must extend a template
file which exists on the instance. Templates are a lot more than configuration.
The format supports conditional logic, inheritance and other features of code.
On the master, device type templates are configured using jinja2 files
in the directory:
/etc/lava-server/dispatcher-config/device-types/
When creating a new device type, it is recommended to add the new template file first and use version control to track changes then copy the template file to the system location.
Warning
Adding a new device type template is the most complex
part of administering a LAVA instance and it can be a lot of work
(sometimes several months) to integrate a completely new device into
LAVA. V2 offers a different and wider range of support to what was
available in V1, but some devices may still need new support to be
written directly within the lava-dispatcher code. It may not
always be possible to automate a new device, depending on how the
device connects to LAVA, how the device is powered and whether the
software on the device allows the device to be controlled remotely.
See also
Adding new device types and creating new device type templates.
panda and panda-es device types are separate in the Cambridge
LAVA lab. When originally introduced, there was an expectation that the
hardware differences between the devices would be relevant to how the jobs
were constructed. As it turned out, no such difference was actually exploited
by the test writers.mustang device type can support both U-Boot and UEFI bootloaders but
not on the same machine at the same time. The bootloader can be changed, but
this is a custom process which is not manageable during a test job. In the
Cambridge lab, mustang implies U-Boot and a separate sub device type
called mustang-uefi is available for test jobs needing UEFI.panda devices can support operating systems like Debian as well
as supporting Android deployments using the same bootloader in both
cases (U-Boot). Therefore, only one device type was needed here.The device type exists as a django database object which can be modified using the django admin interface. The following fields are supported:
alias
where the device type name differs from the template name(s) in use by
devices of this type. (An Alias is one of the
Descriptive fields.)The device type database also includes some optional fields which may be completed by the admin to provide information for test writers:
Note
When modifying device type objects in the Django administration interface, take care with multiple selection boxes. Fields like architecture name can show in the list as being available for selection in a device type object but only the selected line or lines will actually be saved as references within the device type object. The references will show up on the device type detail page in the Information tab.