Commit 27c9272c authored by Emma Anholt's avatar Emma Anholt Committed by Marge Bot
Browse files

docs: Move the gitlab-ci docs to RST.



I tried not to edit too much meaning in the process, but I did shuffle
some stuff around to work as structured documentation.
Reviewed-by: Eric Engestrom's avatarEric Engestrom <eric@engestrom.ch>
Reviewed-by: Erik Faye-Lund 's avatarErik Faye-Lund <erik.faye-lund@collabora.com>
Part-of: <!5510>
parent a2ca7e09
# Mesa testing
LAVA CI
=======
The goal of the "test" stage of the .gitlab-ci.yml is to do pre-merge
testing of Mesa drivers on various platforms, so that we can ensure no
regressions are merged, as long as developers are merging code using
marge-bot.
There are currently 4 automated testing systems deployed for Mesa.
LAVA and gitlab-runner on the DUTs are used in pre-merge testing and
are described in this document. Managing bare metal using
gitlab-runner is described under [bare-metal/README.md]. Intel also
has a jenkins-based CI system with restricted access that isn't
connected to gitlab.
## Mesa testing using LAVA
[LAVA](https://lavasoftware.org/) is a system for functional testing
`LAVA <https://lavasoftware.org/>`_ is a system for functional testing
of boards including deploying custom bootloaders and kernels. This is
particularly relevant to testing Mesa because we often need to change
kernels for UAPI changes (and this lets us do full testing of a new
......@@ -22,7 +9,8 @@ kernel during development), and our workloads can easily take down
boards when mistakes are made (kernel oopses, OOMs that take out
critical system services).
### Mesa-LAVA software architecture
Mesa-LAVA software architecture
-------------------------------
The gitlab-runner will run on some host that has access to the LAVA
lab, with tags like "lava-mesa-boardname" to control only taking in
......@@ -35,7 +23,8 @@ The LAVA instance manages scheduling those jobs to the boards present.
For a job, it will deploy the kernel, device tree, and the ramdisk
containing the CTS.
### Deploying a new Mesa-LAVA lab
Deploying a new Mesa-LAVA lab
-----------------------------
You'll want to start with setting up your LAVA instance and getting
some boards booting using test jobs. Start with the stock QEMU
......@@ -44,12 +33,12 @@ to define your actual boards.
The device type in lava-gitlab-ci.yml is the device type you create in
your LAVA instance, which doesn't have to match the board's name in
`/etc/lava-dispatcher/device-types`. You create your boards under
``/etc/lava-dispatcher/device-types``. You create your boards under
that device type and the Mesa jobs will be scheduled to any of them.
Instantiate your boards by creating them in the UI or at the command
line attached to that device type, then populate their dictionary
(using an "extends" line probably referencing the board's template in
`/etc/lava-dispatcher/device-types`). Now, go find a relevant
``/etc/lava-dispatcher/device-types``). Now, go find a relevant
healthcheck job for your board as a test job definition, or cobble
something together from a board that boots using the same boot_method
and some public images, and figure out how to get your boards booting.
......@@ -68,26 +57,24 @@ the binfmt support.
The docker image will need access to the lava instance. If it's on a
public network it should be fine. If you're running the LAVA instance
on localhost, you'll need to set `network_mode="host"` in
`/etc/gitlab-runner/config.toml` so it can access localhost. Create a
on localhost, you'll need to set ``network_mode="host"`` in
``/etc/gitlab-runner/config.toml`` so it can access localhost. Create a
gitlab-runner user in your LAVA instance, log in under that user on
the web interface, and create an API token. Copy that into a
`lavacli.yaml`:
``lavacli.yaml``:
.. code-block:: yaml
```
default:
token: <token contents>
uri: <url to the instance>
username: gitlab-runner
```
default:
token: <token contents>
uri: <url to the instance>
username: gitlab-runner
Add a volume mount of that `lavacli.yaml` to
`/etc/gitlab-runner/config.toml` so that the docker container can
access it. You probably have a `volumes = ["/cache"]` already, so now it would be
Add a volume mount of that ``lavacli.yaml`` to
``/etc/gitlab-runner/config.toml`` so that the docker container can
access it. You probably have a ``volumes = ["/cache"]`` already, so now it would be::
```
volumes = ["/home/anholt/lava-config/lavacli.yaml:/root/.config/lavacli.yaml", "/cache"]
```
volumes = ["/home/anholt/lava-config/lavacli.yaml:/root/.config/lavacli.yaml", "/cache"]
Note that this token is visible to anybody that can submit MRs to
Mesa! It is not an actual secret. We could just bake it into the
......@@ -96,117 +83,4 @@ LAVA instance is separated from the Mesa branches (particularly
relevant as we have many stable branches all using CI).
Now it's time to define your test runner in
`.gitlab-ci/lava-gitlab-ci.yml`.
## Mesa testing using gitlab-runner on DUTs
### Software architecture
For freedreno and llvmpipe CI, we're using gitlab-runner on the test
devices (DUTs), cached docker containers with VK-GL-CTS, and the
normal shared x86_64 runners to build the Mesa drivers to be run
inside of those containers on the DUTs.
The docker containers are rebuilt from the debian-install.sh script
when DEBIAN\_TAG is changed in .gitlab-ci.yml, and
debian-test-install.sh when DEBIAN\_ARM64\_TAG is changed in
.gitlab-ci.yml. The resulting images are around 500MB, and are
expected to change approximately weekly (though an individual
developer working on them may produce many more images while trying to
come up with a working MR!).
gitlab-runner is a client that polls gitlab.freedesktop.org for
available jobs, with no inbound networking requirements. Jobs can
have tags, so we can have DUT-specific jobs that only run on runners
with that tag marked in the gitlab UI.
Since dEQP takes a long time to run, we mark the job as "parallel" at
some level, which spawns multiple jobs from one definition, and then
deqp-runner.sh takes the corresponding fraction of the test list for
that job.
To reduce dEQP runtime (or avoid tests with unreliable results), a
deqp-runner.sh invocation can provide a list of tests to skip. If
your driver is not yet conformant, you can pass a list of expected
failures, and the job will only fail on tests that aren't listed (look
at the job's log for which specific tests failed).
### DUT requirements
#### DUTs must have a stable kernel and GPU reset.
If the system goes down during a test run, that job will eventually
time out and fail (default 1 hour). However, if the kernel can't
reliably reset the GPU on failure, bugs in one MR may leak into
spurious failures in another MR. This would be an unacceptable impact
on Mesa developers working on other drivers.
#### DUTs must be able to run docker
The Mesa gitlab-runner based test architecture is built around docker,
so that we can cache the debian package installation and CTS build
step across multiple test runs. Since the images are large and change
approximately weekly, the DUTs also need to be running some script to
prune stale docker images periodically in order to not run out of disk
space as we rev those containers (perhaps [this
script](https://gitlab.com/gitlab-org/gitlab-runner/issues/2980#note_169233611)).
Note that docker doesn't allow containers to be stored on NFS, and
doesn't allow multiple docker daemons to interact with the same
network block device, so you will probably need some sort of physical
storage on your DUTs.
#### DUTs must be public
By including your device in .gitlab-ci.yml, you're effectively letting
anyone on the internet run code on your device. docker containers may
provide some limited protection, but how much you trust that and what
you do to mitigate hostile access is up to you.
#### DUTs must expose the dri device nodes to the containers.
Obviously, to get access to the HW, we need to pass the render node
through. This is done by adding `devices = ["/dev/dri"]` to the
`runners.docker` section of /etc/gitlab-runner/config.toml.
### HW CI farm expectations
To make sure that testing of one vendor's drivers doesn't block
unrelated work by other vendors, we require that a given driver's test
farm produces a spurious failure no more than once a week. If every
driver had CI and failed once a week, we would be seeing someone's
code getting blocked on a spurious failure daily, which is an
unacceptable cost to the project.
Additionally, the test farm needs to be able to provide a short enough
turnaround time that people can regularly use the "Merge when pipeline
succeeds" button successfully (until we get
[marge-bot](https://github.com/smarkets/marge-bot) in place on
freedesktop.org). As a result, we require that the test farm be able
to handle a whole pipeline's worth of jobs in less than 5 minutes (to
compare, the build stage is about 10 minutes, if you could get all
your jobs scheduled on the shared runners in time.).
If a test farm is short the HW to provide these guarantees, consider
dropping tests to reduce runtime.
`VK-GL-CTS/scripts/log/bottleneck_report.py` can help you find what
tests were slow in a `results.qpa` file. Or, you can have a job with
no `parallel` field set and:
```
variables:
CI_NODE_INDEX: 1
CI_NODE_TOTAL: 10
```
to just run 1/10th of the test list.
If a HW CI farm goes offline (network dies and all CI pipelines end up
stalled) or its runners are consistenly spuriously failing (disk
full?), and the maintainer is not immediately available to fix the
issue, please push through an MR disabling that farm's jobs by adding
'.' to the front of the jobs names until the maintainer can bring
things back up. If this happens, the farm maintainer should provide a
report to mesa-dev@lists.freedesktop.org after the fact explaining
what happened and what the mitigation plan is for that failure next
time.
``.gitlab-ci/lava-gitlab-ci.yml``.
# bare-metal Mesa testing
Bare-metal CI
=============
Testing Mesa with gitlab-runner running on the devices being tested
(DUTs) proved to be too unstable, so this set of scripts is for
running Mesa testing on bare-metal boards connected to a separate
system using gitlab-runner. Currently only "fastboot" and "ChromeOS
Servo" devices are supported.
The bare-metal scripts run on a system with gitlab-runner and docker,
connected to potentially multiple bare-metal boards that run tests of
Mesa. Currently only "fastboot" and "ChromeOS Servo" devices are
supported.
In comparison with LAVA, this doesn't involve maintaining a separate
webservice with its own job scheduler and replicating jobs between the
......@@ -12,7 +12,8 @@ two. It also places more of the board support in git, instead of
webservice configuration. On the other hand, the serial interactions
and bootloader support are more primitive.
## Requirements (fastboot)
Requirements (fastboot)
-----------------------
This testing requires power control of the DUTs by the gitlab-runner
machine, since this is what we use to reset the system and get back to
......@@ -31,7 +32,8 @@ The boards should have networking, so that (in a future iteration of
this code) we can extract the dEQP .xml results to artifacts on
gitlab.
## Requirements (servo)
Requirements (servo)
--------------------
For servo-connected boards, we can use the EC connection for power
control to reboot the board. However, loading a kernel is not as easy
......@@ -47,46 +49,45 @@ at the cost of needing more storage on the runner.
Telling the board about where its TFTP and NFS should come from is
done using dnsmasq on the runner host. For example, this snippet in
the dnsmasq.conf.d in the google farm, with the gitlab-runner host we
call "servo".
call "servo"::
```
dhcp-host=1c:69:7a:0d:a3:d3,10.42.0.10,set:servo
dhcp-host=1c:69:7a:0d:a3:d3,10.42.0.10,set:servo
# Fixed dhcp addresses for my sanity, and setting a tag for
# specializing other DHCP options
dhcp-host=a0:ce:c8:c8:d9:5d,10.42.0.11,set:cheza1
dhcp-host=a0:ce:c8:c8:d8:81,10.42.0.12,set:cheza2
# Fixed dhcp addresses for my sanity, and setting a tag for
# specializing other DHCP options
dhcp-host=a0:ce:c8:c8:d9:5d,10.42.0.11,set:cheza1
dhcp-host=a0:ce:c8:c8:d8:81,10.42.0.12,set:cheza2
# Specify the next server, watch out for the double ',,'. The
# filename didn't seem to get picked up by the bootloader, so we use
# tftp-unique-root and mount directories like
# /srv/tftp/10.42.0.11/jwerner/cheza as /tftp in the job containers.
tftp-unique-root
dhcp-boot=tag:cheza1,cheza1/vmlinuz,,10.42.0.10
dhcp-boot=tag:cheza2,cheza2/vmlinuz,,10.42.0.10
# Specify the next server, watch out for the double ',,'. The
# filename didn't seem to get picked up by the bootloader, so we use
# tftp-unique-root and mount directories like
# /srv/tftp/10.42.0.11/jwerner/cheza as /tftp in the job containers.
tftp-unique-root
dhcp-boot=tag:cheza1,cheza1/vmlinuz,,10.42.0.10
dhcp-boot=tag:cheza2,cheza2/vmlinuz,,10.42.0.10
dhcp-option=tag:cheza1,option:root-path,/srv/nfs/cheza1
dhcp-option=tag:cheza2,option:root-path,/srv/nfs/cheza2
```
dhcp-option=tag:cheza1,option:root-path,/srv/nfs/cheza1
dhcp-option=tag:cheza2,option:root-path,/srv/nfs/cheza2
## Setup
Setup
-----
Each board will be registered in fd.o gitlab. You'll want something
like this to register a fastboot board:
```
sudo gitlab-runner register \
--url https://gitlab.freedesktop.org \
--registration-token $1 \
--name MY_BOARD_NAME \
--tag-list MY_BOARD_TAG \
--executor docker \
--docker-image "alpine:latest" \
--docker-volumes "/dev:/dev" \
--docker-network-mode "host" \
--docker-privileged \
--non-interactive
```
.. code-block:: console
sudo gitlab-runner register \
--url https://gitlab.freedesktop.org \
--registration-token $1 \
--name MY_BOARD_NAME \
--tag-list MY_BOARD_TAG \
--executor docker \
--docker-image "alpine:latest" \
--docker-volumes "/dev:/dev" \
--docker-network-mode "host" \
--docker-privileged \
--non-interactive
For a servo board, you'll need to also volume mount the board's NFS
root dir at /nfs and TFTP kernel directory at /tftp.
......@@ -107,18 +108,16 @@ with fastboot, and the servo serial devices are acctually links to
spin up a server to collect XML results for fastboot.
Once you've added your boards, you're going to need to add a little
more customization in `/etc/gitlab-runner/config.toml`. First, add
`concurrent = <number of boards>` at the top ("we should have up to
more customization in ``/etc/gitlab-runner/config.toml``. First, add
``concurrent = <number of boards>`` at the top ("we should have up to
this many jobs running managed by this gitlab-runner"). Then for each
board's runner, set `limit = 1` ("only 1 job served by this board at a
board's runner, set ``limit = 1`` ("only 1 job served by this board at a
time"). Finally, add the board-specific environment variables
required by your bare-metal script, something like:
required by your bare-metal script, something like::
```
[[runners]]
name = "google-freedreno-db410c-1"
environment = ["BM_SERIAL=/dev/ttyDB410c8", "BM_POWERUP=google-power-up.sh 8", "BM_FASTBOOT_SERIAL=15e9e390"]
```
[[runners]]
name = "google-freedreno-db410c-1"
environment = ["BM_SERIAL=/dev/ttyDB410c8", "BM_POWERUP=google-power-up.sh 8", "BM_FASTBOOT_SERIAL=15e9e390"]
Once you've updated your runners' configs, restart with `sudo service
gitlab-runner restart`
Once you've updated your runners' configs, restart with ``sudo service
gitlab-runner restart``
Docker CI
=========
For llvmpipe and swrast CI, we run tests in a container containing
VK-GL-CTS, on the shared gitlab runners provided by `freedesktop
<http://freedesktop.org>`_
Software architecture
---------------------
The docker containers are rebuilt from the debian-install.sh script
when DEBIAN\_TAG is changed in .gitlab-ci.yml, and
debian-test-install.sh when DEBIAN\_ARM64\_TAG is changed in
.gitlab-ci.yml. The resulting images are around 500MB, and are
expected to change approximately weekly (though an individual
developer working on them may produce many more images while trying to
come up with a working MR!).
gitlab-runner is a client that polls gitlab.freedesktop.org for
available jobs, with no inbound networking requirements. Jobs can
have tags, so we can have DUT-specific jobs that only run on runners
with that tag marked in the gitlab UI.
Since dEQP takes a long time to run, we mark the job as "parallel" at
some level, which spawns multiple jobs from one definition, and then
deqp-runner.sh takes the corresponding fraction of the test list for
that job.
To reduce dEQP runtime (or avoid tests with unreliable results), a
deqp-runner.sh invocation can provide a list of tests to skip. If
your driver is not yet conformant, you can pass a list of expected
failures, and the job will only fail on tests that aren't listed (look
at the job's log for which specific tests failed).
DUT requirements
----------------
In addition to the general :ref:`CI-farm-expectations`, using
docker requiers:
* DUTs must have a stable kernel and GPU reset (if applicable).
If the system goes down during a test run, that job will eventually
time out and fail (default 1 hour). However, if the kernel can't
reliably reset the GPU on failure, bugs in one MR may leak into
spurious failures in another MR. This would be an unacceptable impact
on Mesa developers working on other drivers.
* DUTs must be able to run docker
The Mesa gitlab-runner based test architecture is built around docker,
so that we can cache the debian package installation and CTS build
step across multiple test runs. Since the images are large and change
approximately weekly, the DUTs also need to be running some script to
prune stale docker images periodically in order to not run out of disk
space as we rev those containers (perhaps `this script
<https://gitlab.com/gitlab-org/gitlab-runner/issues/2980#note_169233611>`_).
Note that docker doesn't allow containers to be stored on NFS, and
doesn't allow multiple docker daemons to interact with the same
network block device, so you will probably need some sort of physical
storage on your DUTs.
* DUTs must be public
By including your device in .gitlab-ci.yml, you're effectively letting
anyone on the internet run code on your device. docker containers may
provide some limited protection, but how much you trust that and what
you do to mitigate hostile access is up to you.
* DUTs must expose the dri device nodes to the containers.
Obviously, to get access to the HW, we need to pass the render node
through. This is done by adding ``devices = ["/dev/dri"]`` to the
``runners.docker`` section of /etc/gitlab-runner/config.toml.
Continuous Integration
======================
GitLab CI
---------
......@@ -18,6 +17,7 @@ The CI runs a number of tests, from trivial build-testing to complex GPU renderi
- Sanity checks (``meson test`` & ``scons check``)
- Some drivers (softpipe, llvmpipe, freedreno and panfrost) are also tested
using `VK-GL-CTS <https://github.com/KhronosGroup/VK-GL-CTS>`__
- Replay of application traces
A typical run takes between 20 and 30 minutes, although it can go up very quickly
if the GitLab runners are overwhelmed, which happens sometimes. When it does happen,
......@@ -42,6 +42,15 @@ about it on ``#freedesktop`` on Freenode and tag `Daniel Stone
`Eric Anholt <https://gitlab.freedesktop.org/anholt>`__ (``anholt`` on
IRC).
The three gitlab CI systems currently integrated are:
.. toctree::
:maxdepth: 1
bare-metal
LAVA
docker
Intel CI
--------
......@@ -74,3 +83,46 @@ it on ``#dri-devel`` on Freenode and tag `Clayton Craft
<https://gitlab.freedesktop.org/craftyguy>`__ (``craftyguy`` on IRC) or
`Nico Cortes <https://gitlab.freedesktop.org/ngcortes>`__ (``ngcortes``
on IRC).
.. _CI-farm-expectations:
CI farm expectations
--------------------
To make sure that testing of one vendor's drivers doesn't block
unrelated work by other vendors, we require that a given driver's test
farm produces a spurious failure no more than once a week. If every
driver had CI and failed once a week, we would be seeing someone's
code getting blocked on a spurious failure daily, which is an
unacceptable cost to the project.
Additionally, the test farm needs to be able to provide a short enough
turnaround time that we can get our MRs through marge-bot without the
pipeline backing up. As a result, we require that the test farm be
able to handle a whole pipeline's worth of jobs in less than 5 minutes
(to compare, the build stage is about 10 minutes, if you could get all
your jobs scheduled on the shared runners in time.).
If a test farm is short the HW to provide these guarantees, consider
dropping tests to reduce runtime.
``VK-GL-CTS/scripts/log/bottleneck_report.py`` can help you find what
tests were slow in a ``results.qpa`` file. Or, you can have a job with
no ``parallel`` field set and:
.. code-block:: yaml
variables:
CI_NODE_INDEX: 1
CI_NODE_TOTAL: 10
to just run 1/10th of the test list.
If a HW CI farm goes offline (network dies and all CI pipelines end up
stalled) or its runners are consistenly spuriously failing (disk
full?), and the maintainer is not immediately available to fix the
issue, please push through an MR disabling that farm's jobs by adding
'.' to the front of the jobs names until the maintainer can bring
things back up. If this happens, the farm maintainer should provide a
report to mesa-dev@lists.freedesktop.org after the fact explaining
what happened and what the mitigation plan is for that failure next
time.
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment