There are many potential reasons why tests running in LAVA might fail, or produce unexpected behavior. Some of them can be easy to track down, but others may be more difficult. The devices, software and test suites can vary massively from one test job to the next, but nonetheless a few common ideas may help you to work out what’s going wrong.
This may seem obvious, but it is all too easy to miss real problems in the test logs! For people not used to diagnosing failures, it is worth reading all the way from deployment through test device boot to the end of the logfile. If a test job fails to complete successfully, it can often be caused by a problem much earlier in the test - don’t assume that the final few lines of the logfile will tell the whole story:
When writing tests, make things verbose to give yourself more useful logs in case they fail.
Certain operations will cause a failure comment to be automatically added to the testjob.
See also
If the test system does not (seem to) boot at all, there are a few things worth checking:
See also
Did the kernel boot OK but then fail to find the root filesystem? This is a common failure mode, and there are quite a few possible causes. Here are some of the more common failure cases.
sd_mod
), filesystems (e.g. ext4
) or network interfaces (e.g.
e1000e
) if you’re using NFS for the rootfs. You should be able to see
what devices are found by the kernel by reading the boot messages; check that
the device you are expecting to use does show up there.root=
parameter.init
program, in the correct
location. In an initramfs, the default location is /init
; this can be
over-ridden on the kernel command line using the init=
parameter.This is a common theme throughout the suggested workflow for developing tests in LAVA. Start with simple test jobs and verify they work as expected. Add complexity one step at a time, ensuring that each new option or test suite added behaves as expected. It’s much easier to work out what has broken in a test job if you’ve made just one small change to a previous test job that worked fine.
Similarly, if you have a complex test job that’s not working correctly then often the easiest way to find the problem is to simplify the job - remove some of the complexity and re-test. By removing the complex setup in the test, it should be possible to identify the cause of the failure.
If there are standard test jobs available for the device type in question, it might be useful to compare your test job to one of those standard jobs, or even start with one and append your test definitions.
When developing a test, resist the urge to make too many changes at once - test one element at a time. Avoid changing the deployed files and the test definition in the same job. When the deployed files change, use an older test definition and an inline definition to explicitly check for any new support your test will want to use from those new files. If you change too many variables at once, it may become impossible to work out what change caused things to break.
Especially when developing a new test, add plenty of output to explain what is going on. If you are starting with a new test device or new boot files, make it easy to diagnose problems later by adding diagnostics early in the process. In general, it is much easier to debug a failed test when it is clear about what it expects to be happening than one which just stops or says “error” in the middle of a test. The presence of debug information in a known working test job can be invaluable when checking why a different test job or test case failed.
If your test configures one or more network interfaces, add the output of
ifconfig
or ip a show
afterwards to show that it worked. Consider
adding calls to route
or running cat /etc/resolv.conf
as well.
If your test uses a specific block device or filesystem, add the
output of df
or mount
to show what devices and filesystems are
available.
Check the kernel support available inside the test image by running
commands to output details into the test job log file. Once you know which
parts of /dev/
, /proc/
and /sys
are relevant to the commands used
in your test definition, use grep
and cat
to ensure that details
about the available support are available when you come to debug the test
job.
Check the available kernel modules using lsmod
or by outputting the
contents of modules.dep
, depending on the configuration of the kernel
used in the test job.
Use the metadata to reference the build log and configuration of files used in the test job, especially the kernel, initramfs and / or NFS.
See also
If you are writing shell scripts to wrap tests, try using set -x
- this
will tell the shell to log all lines of your script as it runs them. For
example:
#!/bin/sh
set -e
set -x
echo "foo"
a=1
if [ $a -eq 1 ]; then
echo "yes"
fi
will give the following output:
+ echo foo
foo
+ a=1
+ [ 1 -eq 1 ]
+ echo yes
yes
The debug statements used when the test definitions are being developed can be retained in the final test definitions for later reference. It is much better to have the debug information available in every test than to have to resubmit the test job only to find that the problem is intermittent or can only be reproduced in particular operations.
Debug checks which become common across a range of test job definitions or
which are particularly important for quick triage can also be run as test cases
so that the presence or absence of a critical element of the test shows up as a
pass or fail. Many such checks will need to use scripts to isolate the relevant
information from the available data in proc
or dmesg
etc.
There are some common mistakes using LAVA which can cause issues. If you are experiencing weird problems with your test job, maybe considering these will help.
Triage will be a lot easier if you follow these guidelines when using files you have built or modified yourself in LAVA test jobs:
Use a checksums on all downloaded copies of locally rebuilt files. Frequent rebuilds lead to confusion about whether the file you have just built is the same file as the test job uses. Even when you are sure you have updated the file correctly, there may be caches between the upload location and the worker.
Always update the metadata every time a local file is rebuilt for use in a testjob. Include details of what was changed to require the file to be rebuilt and when that change was made.
Always include and update files describing the configuration of the locally
built file. If building a kernel, enabling /proc/config.gz
can save large
amounts of time in triage. Upload the full configuration and build log of all
files and include the URL to those files in the metadata. It can be
very difficult for anyone to help you debug your test jobs if the details of
how the test job files were built is not available. Consider using version
control software for the test job definitions, configuration files, build
logs or changelogs to make it easier to track what has changed. When
rebuilding local files for your test jobs, please remember:
Change one thing at a time.
Retain old copies of locally built files, especially if test jobs using those files ran successfully.
Compare your configuration with known working test jobs.
See also
Pipes, redirects and nested sub shells will not work reliably when put directly into the YAML. Use a wrapper script (with set -x) instead for safety:
#!/bin/sh
set -e
set -x
ifconfig|grep "inet addr"|grep -v "127.0.0.1"|cut -d: -f2|cut -d' ' -f1
Un-nested sub-shells do work, though:
- lava-test-case multinode-send-network --shell lava-send network hostname=$(hostname) fqdn=$(hostname -f)
If you use a custom result parser, configure one of your YAML files to output the entire test result output to stdout so that you can reliably capture a representative block of output. Test your proposed result parser against the block using your favorite language.
Comment out the parser from the YAML if there are particular problems, just to see what the default LAVA parsers can provide.
Note
Parsers can be difficult to debug after being parsed from YAML into shell. LAVA developers used to recommend the use of custom parsers, but experience has shown this to be a mistake. Instead, it is suggested that new test definitions should use custom scripts. This allows the parsing to be debugged outside LAVA, as well as making the test itself more portable.
cd
in your YAML, always store where you were and where you end
up using pwd
.realpath
and use that to debug your directory
structure.A first step in triage of a test job failure can be to replicate the steps manually. If this works, then consider the differences between running a test manually and through automation:
Some devices have substantial requirements for infrastructure to support the automation: switchable USB hubs, relays, remote power control, multiple serial connections, VLANd support, etc.
Triaging of test job failures in one automated system typically needs to be done on the same instance or, if using another instance, using infrastructure which is as close as possible to the original instance. It will still be difficult to identify the problem, especially with intermittent failures, unless key elements of the test instance can be disabled, replaced or otherwise eliminated from the test process without generating new failures.
LAVA tries to identify the likely cause of the error and raise the
correct exception. (This can be tracked in the job
test case
created by every test job in the lava
test suite of the results.)
It can be particularly hard to identify the cause of timeouts. Pay close attention to all devices across the instance to see if a third party element (like a distribution mirror) is the cause. Look for common factors - both those which trigger a failure and those which do not.
When investigating intermittent errors, see if the error can be provoked in a health check and then use looping mode to generate data on how often the error occurs whilst keeping the test job identical.
If the health check does generate the error, the device will go offline. Infrastructure problems can be debugged whilst keeping the device(s) offline by Running lava-run directly.
See also
Manual keyboard entry has noticeable gaps between every key press. Even the fastest typist will not approach the speed at which a computer can transmit the same string over a serial connection.
In automation, strings will be sent as quickly as the connection allows. Some devices may then fail to process the characters correctly. This might manifest in several ways, including:
rot
instead of root
or erverip
instead of set serverip
. Often at the start of a line, although
also includes loss of the newline itself, causing lines to join
together. set foo ; set bar
can be changed to set foo set
bar
, causing a failure to process bar
.orot
instead of root
. This is less
common than missing characters and can sometimes indicate a hardware
problem on the device. However, replicating an input speed which is
closer to human typing can still alleviate the problem.Note
This is not the same as the replacement of characters by invalid characters which is a different type of serial corruption. If you see ASCII strings being output to the device but unprintable or otherwise incorrect characters being received, then this could be a hardware problem with the DUT or the connections to it. Intermittent single bit flips in the serial data stream are all too common.
LAVA supports specifying character delays in the boot and test actions to help alleviate these problems. These are device-specific features, so best controlled in the device configuration.
The boot
action suffers from this problem more frequently than the
test
action, typically because boot
has to interact with
processes executed by firmware or a bootloader where processing can be
more limited than in a POSIX-type test environment.
In the device-type template, set the number of milliseconds to add
between each character of every string sent to the DUT during the
boot
action:
{% set boot_character_delay = 10 %}
base.jinja2
will then handle this variable to set the boot
character delay to 10 milliseconds. Some devices may need more, up to
100 or 500 milliseconds. In the case of such long delays, it is also
necessary to consider the overall boot timeout and specify a minimum
for the relevant boot action in the device-type template.
In the device-type template, set the number of milliseconds to add
between each character of every string sent to the DUT during the
test
action:
{% set test_character_delay = 10 %}
base.jinja2
will then handle this variable to set the test
character delay to 10 milliseconds.
MultiNode tests are necessarily more complex than jobs running on single test devices, and so there are extra places where errors can creep in and cause unexpected failures.
This may seem obvious, but one of the most common causes of MultiNode test failure is nothing to do with MultiNode. If your MultiNode tests are failing to boot correctly, check that the basics of each of the desired roles works independently. Remove the MultiNode pieces and just check that the specified deploy and boot actions work alone in a single-node test with the right device-type. Then add back the MultiNode configuration, changing one thing at a time and ensuring that things still work as you build up complexity.
A lava-wait must be preceded by a lava-send from at least one other device in the group, or the waiting device will timeout
This can be a particular problem if you remove test definitions or edit a YAML file without checking other uses of the same file. The simplest (and hence recommended) way to use the MultiNode synchronization calls is using inline definitions.
Always check whether the test result came back as a failure due to some cause other than the test definition itself. Particularly with MultiNode test jobs, a test can fail for other reasons like an unrelated failure on a different board within the group.