Re: [PATCH v3 0/4] Yocto Gitlab CI

[Stefano Stabellini <sstabellini@xxxxxxxxxx>]

On Thu, 10 Nov 2022, Michal Orzel wrote:
> Hi Stefano,
> 
> On 10/11/2022 01:18, Stefano Stabellini wrote:
> > 
> > 
> > On Mon, 7 Nov 2022, Michal Orzel wrote:
> >> Hi Bertrand and Stefano,
> >>
> >> On 31/10/2022 16:00, Bertrand Marquis wrote:
> >>>
> >>>
> >>> Hi Michal,
> >>>
> >>>> On 31 Oct 2022, at 14:39, Michal Orzel <michal.orzel@xxxxxxx> wrote:
> >>>>
> >>>> Hi Bertrand,
> >>>>
> >>>> On 31/10/2022 15:00, Bertrand Marquis wrote:
> >>>>>
> >>>>>
> >>>>> This patch series is a first attempt to check if we could use Yocto in
> >>>>> gitlab ci to build and run xen on qemu for arm, arm64 and x86.
> >>>>>
> >>>>> The first patch is creating a container with all elements required to
> >>>>> build Yocto, a checkout of the yocto layers required and an helper
> >>>>> script to build and run xen on qemu with yocto.
> >>>>>
> >>>>> The second patch is creating containers with a first build of yocto done
> >>>>> so that susbsequent build with those containers would only rebuild what
> >>>>> was changed and take the rest from the cache.
> >>>>>
> >>>>> The third patch is adding a way to easily clean locally created
> >>>>> containers.
> >>>>>
> >>>>> This is is mainly for discussion and sharing as there are still some
> >>>>> issues/problem to solve:
> >>>>> - building the qemu* containers can take several hours depending on the
> >>>>>  network bandwith and computing power of the machine where those are
> >>>>>  created
> >>>> This is not really an issue as the build of the containers occurs on the 
> >>>> local
> >>>> machines before pushing them to registry. Also, building the containers
> >>>> will only be required for new Yocto releases.
> >>>>
> >>>>> - produced containers containing the cache have a size between 8 and
> >>>>>  12GB depending on the architecture. We might need to store the build
> >>>>>  cache somewhere else to reduce the size. If we choose to have one
> >>>>>  single image, the needed size is around 20GB and we need up to 40GB
> >>>>>  during the build, which is why I splitted them.
> >>>>> - during the build and run, we use a bit more then 20GB of disk which is
> >>>>>  over the allowed size in gitlab
> >>>> As we could see during v2 testing, we do not have any space restrictions
> >>>> on the Xen GitLab and I think we already decided to have the Yocto
> >>>> integrated into our CI.
> >>>
> >>> Right, I should have modified this chapter to be coherent with your 
> >>> latest tests.
> >>> Sorry for that.
> >>>
> >>>>
> >>>> I will do some testing and get back to you with results + review.
> >> I did some testing and here are the results:
> >>
> >> In the current form this series will fail when running CI because the 
> >> Yocto containers
> >> are based on "From ubuntu:22.04" (there is no platform prefix), which 
> >> means that the containers
> >> are built for the host architecture (in my case and in 99% of the cases of 
> >> the local build it will
> >> be x86). In Gitlab we have 2 runners (arm64 and x86_64). This means that 
> >> all the test jobs would need
> >> to specify x86_64 as a tag when keeping the current behavior.
> >> After I built all the containers on my x86 machine, I pushed them to 
> >> registry and the pipeline was successful:
> >> https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgitlab.com%2Fxen-project%2Fpeople%2Fmorzel%2Fxen-orzelmichal%2F-%2Fpipelines%2F686853939&amp;data=05%7C01%7Cmichal.orzel%40amd.com%7C2449f063e67341c3b95a08dac2b112a5%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C638036363027707274%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&amp;sdata=EwTJrW2vuwQIugKc7mnzG9NNbsYLP6tw5UODzBMmPEE%3D&amp;reserved=0
> > 
> > When I tested the previous version of this series I built the
> > containers natively on ARM64, so that is also an option.
> > 
> > 
> >> Here is the diff on patch no. 3 to make the series work (using x86 tag and 
> >> small improvement to include needs: []):
> >> ```
> >> diff --git a/automation/gitlab-ci/test.yaml 
> >> b/automation/gitlab-ci/test.yaml
> >> index 5c620fefce59..52cccec6f904 100644
> >> --- a/automation/gitlab-ci/test.yaml
> >> +++ b/automation/gitlab-ci/test.yaml
> >> @@ -65,6 +65,9 @@
> >>      paths:
> >>        - 'logs/*'
> >>      when: always
> >> +  needs: []
> >> +  tags:
> >> +    - x86_64
> >>
> >>  # Test jobs
> >>  build-each-commit-gcc:
> >> @@ -206,19 +209,13 @@ yocto-qemuarm64:
> >>    extends: .yocto-test
> >>    variables:
> >>      YOCTO_BOARD: qemuarm64
> >> -  tags:
> >> -    - arm64
> >>
> >>  yocto-qemuarm:
> >>    extends: .yocto-test
> >>    variables:
> >>      YOCTO_BOARD: qemuarm
> >> -  tags:
> >> -    - arm32
> >>
> >>  yocto-qemux86-64:
> >>    extends: .yocto-test
> >>    variables:
> >>      YOCTO_BOARD: qemux86-64
> >> -  tags:
> >> -    - x86_64
> >> ```
> >>
> >> Now, the logical way would be to build x86 yocto container for x86, arm64 
> >> for arm64 and arm32 on arm64 or x86.
> >> I tried building the container qemuarm64 specifying target arm64 on x86. 
> >> After 15h, only 70% of the Yocto build
> >> was completed and there was an error with glibc (the local build of the 
> >> container for the host arch takes on my machine max 2h).
> >> This enormous amount of time is due to the qemu docker emulation that 
> >> happens behind the scenes (I checked on 2 different machines).
> >>
> >> So we have 3 solutions:
> >> 1) Build and run these containers for/on x86_64:
> >>  - local users can build the containers on local machines that are almost 
> >> always x86 based, in short period of time,
> >>  - "everyone" can build/push the containers once there is a new Yocto 
> >> release
> >>  - slightly slower CI build time
> >> 2) Build and run these containers for specific architectures:
> >>  - almost no go for local users using x86 machine (unless using more than 
> >> 16 threads (which I used) and willing to wait 2 days for the build)
> >>  - faster CI build time (arm64 runner is faster than x86 one)
> >>  - someone with arm64 based machine (not that common) would have to build 
> >> and push the containers
> >> 3) Try to use CI to build and push the containers to registry
> >>  - it could be possible but what about local users
> > 
> > From a gitlab-ci perspective, given the runners we currently have, we
> > have to go with option 2). We don't have enough resources available on
> > the x86 runner to run the Yocto jobs on x86.
> > 
> That is what I reckon too. Running the Yocto build/test on CI using x86 
> runner will always be slower.
> So, if we go with this solution, then the following is needed:
> 1. Modify test jobs so that yocto-qemu{arm64/arm} uses arm64 tag to be taken 
> by arm64 runner and use tag x86_64 for yocto-qemux86-64.
> 2. Come up with a solution to build the yocto containers automatically for 
> the above platforms + possibility to specify the platform for local users.
>    Right now, these containers are being always build for the host machine 
> platform, so without doing tricks like adding --platform or prefix to image 
> name,
>    one cannot build the Yocto containers that would be ready to be pushed to 
> registry. We need to have a clean solution without requiring user to do 
> tricks.
> 
> The only drawback of this solution is that the person building the 
> yocto-qemu{arm64/arm} container and willing to push it to registry,
> needs to have access to arm64 machine.

I am fine with this drawback for now.

Due to resource constraints, we might want to avoid adding
yocto-qemux86-64 (if yocto-qemux86-64 has to run on x86) for now, I
worry it might choke the x86 ci-loop. Or we could add it but keep it
disabled. We'll enable it when we get better x86 runners.