
Building in a tmpfs speeds the build process and greatly reduces the disk io required during a build, leaving the host far more responsive to other tasks. Added option --tmpfs_percentage=N to the build-pkgs command to control how much memory can be used for tmpfs build environments. The valid range is 0-50%. The default is 0% resulting in a traditional disk only build. The available memory on the host is multiplied by the percentage to determine how much is availabe for tmpfs in total. This total is then assigned to up to N-1 of the N parallel build environments, with the minimum tmpfs environment size being 10GB. One environment is reserved for disk based builds. First time package builds are assigned to tmpfs based build environments in preference over the disk base environments. However, if a package has failed a prior build attempt, subsequents attemps will only occure on a disk based build environmnet. This may help if the build failed due to a too small tmpfs build environment. It also may leave the environment intact for mor indepth debugging. Testing has revealed that 20gb is required to build the largest packages (linux, ceph, kubernetes). In order to avoid costly rebuilds of these large packages, the choice of tmpfs percentage and number of parallel builds is important. e.g. on a single user host with 128GB of memory, an appropriate choice might be... --parallel=4 --tmpfs_percentage=50 ...yielding 3 tmpfs build environments of aprox 20GB each. A higher parallelism, or a lower tempfs percentage will result in build environments that drop below 20GB, and you might start seeing rebuilds of large packages. Further development is suggested. If we can add an advisory to the dsc metadata suggesting a minimum space requireemnt for the build of a package, we can proactively assign large package to a build environment that is large enough to support it, avoiding rebuilds. Testing: - build-pkgs without --tmpfs_percentage Only disk base build envoronments are used. - build-pkgs --parallel=4 --tmpfs_percentage=20 On 128GB machine, Only 2 10-12 GB tmpfs are used, the other two remanin disk based. Large packages fail in 10GB tmpfs, and a pass when retried on disk. - build-pkgs --parallel=4 --tmpfs_percentage=50 On 128GB machine, Creates 3 18-20 GB tmpfs. Large packages build in tmpfs without need for rebuild. partial-bug: 2081843 Change-Id: I09dd2f60afc3e866ec8f86b6898d41f19a419d87 Signed-off-by: Scott Little <scott.little@windriver.com>
tools
StarlingX Build Tools
The StarlingX build process is tightly tied to CentOS in a number of ways, doing the build inside a Docker container makes this much easier on other flavors of Linux. Basically, the StarlingX ISO image creation flow involves the following general steps.
- Build the StarlingX docker image.
- Package mirror creation.
- Build packages/ISO creation.
Build the Starlingx docker image
StarlingX docker image handles all steps related to StarlingX ISO creation. This section describes how to customize the docker image building process.
Container build image customization
You can start by customizing values for the StarlingX docker image build process. There are a pair of useful files that help to do this.
buildrc
localrc
The buildrc
file is a shell script that is used to set
the default configuration values. It is contained in the tbuilder repo
and should not need to be modified by users as it reads a
localrc
file that will not be overwritten by tbuilder
updates. This is where users should alter the default settings. This is
a sample of a localrc
file:
# tbuilder localrc
MYUNAME=<your user name>
PROJECT=starlingx
HOST_PREFIX=$HOME/starlingx/workspace
HOST_MIRROR_DIR=$HOME/starlingx/mirror
This project contains a Makefile that can be used to automate the
build lifecycle of a container. The Makefile will read the contents of
the buildrc
file.
StarlingX Builder container image are tied to your UID so image names should include your username.
Build image
Once the configuration files have been customized, it is possible to
build the docker image. This process is automated by the
tb.sh
script.
NOTE:
- Do NOT change the UID to be different from the one you have on your
host or things will go poorly. i.e. do not change
--build-arg MYUID=$(id -u)
- The Dockerfile needs MYUID and MYUNAME defined, the rest of the configuration is copied in via buildrc/localrc.
Package mirror creation
Once the StarlingX docker image has been built, you must create a mirror before creating the ISO image. Basically, a mirror is a directory that contains a series of packages. The packages are organized to be consumed by the ISO creation scripts.
The HOST_MIRROR_DIR
variable provides the path to the
mirror. The buildrc
file sets the value of this variable
unless the localrc
file has modified it.
The mirror creation involves a set of scripts and configuration files
required to download a group of RPMs, SRPMs, source code packages and so
forth. These tools live inside centos-mirror-tools
directory.
$ cd centos-mirror-tools
All items included in this directory must be visble inside the container environment. Then the container shall be run from the same directory where these tools are stored. Basically, we run a container with the previously created StarlingX docker image, using the following configuration:
$ docker run -it -v $(pwd):/localdisk <your_docker_image_name>:<your_image_version> bash
As /localdisk
is defined as the workdir of the
container, the same folder name should be used to define the volume. The
container will start to run and populate logs
and
output
folders in this directory.
Download packages
Inside the Docker container, enter the following commands to download the required packages to populate the CentOS mirror repository:
$ cd localdisk && bash download_mirror.sh
Monitor the download of packages until it is complete. When the download is complete, the following message appears:
totally 17 files are downloaded! step #3: done successfully IMPORTANT: The following 3 files are just bootstrap versions. Based on them, the workable images for StarlingX could be generated by running "update-pxe-network-installer" command after "build-iso" - out/stx/CentOS/Binary/LiveOS/squashfs.img - out/stx/CentOS/Binary/images/pxeboot/initrd.img - out/stx/CentOS/Binary/images/pxeboot/vmlinuz
Verify packages
Verify no missing or failed packages exist:
$ cat logs/_missing_.log $ cat logs/_failmove_.log
In case missing or failed packages do exist, which is usually caused by network instability (or timeout), you need to download the packages manually. Doing so assures you get all RPMs listed in rpms_3rdparties.lst/rpms_centos.lst/rpms_centos3rdparties.lst.
Copy the files to the mirror
After all downloads are complete, copy the downloaded files to mirror.
$ find ./output -name "*.i686.rpm" | xargs rm -f $ chown 751:751 -R ./output $ cp -rf output/stx/ <your_mirror_folder>/
In this case, <your_mirror_folder>
can be whatever
folder you want to use as mirror.
Tweaks in the StarlingX build system.
NOTE: You do not need to do the following step if you've synced the latest codebase.
Go into the StarlingX build system (i.e. another container that hosts the build system) and perform the following steps:
Build packages/ISO creation
StarlingX ISO image creation required some customized packages. In
this step, a set of patches and customizations are applied to the source
code to create the RPM packages. We have an script called
tb.sh
that helps with the process.
The tb.sh
script is used to manage the run/stop
lifecycle of working containers. Copy it to somewhere on your
PATH
, say $HOME/bin
if you have one, or maybe
/usr/local/bin
.
The basic workflow is to create a working directory for a particular
build, say a specific branch or whatever. Copy the buildrc
file from the tbuilder repo to your work directory and create a
localrc
if you need one. The current working directory is
assumed to be this work directory for all tb.sh
commands.
You switch projects by switching directories.
By default LOCALDISK
will be placed under the directory
pointed to by HOST_PREFIX
, which defaults to
$HOME/starlingx
.
The tb.sh
script uses sub-commands to select the
operation: * run
- Runs the container in a shell. It will
also create LOCALDISK
if it does not exist. *
stop
- Kills the running shell. * exec
-
Starts a shell inside the container.
You should name your running container with your username. tbuilder
does this automatically using the USER
environment
variable.
tb.sh run
will create LOCALDISK
if it does
not already exist before starting the container.
Set the mirror directory to the shared mirror pointed to by
HOST_MIRROR_DIR
. The mirror is LARGE, if you are on a
shared machine use the shared mirror. For example you could set the
default value for HOST_MIRROR_DIR
to
$HOME/starlingx/mirror
and share it.
Running the Container
Start the builder container:
or by hand:
docker run -it --rm \
--name ${TC_CONTAINER_NAME} \
--detach \
-v ${LOCALDISK}:${GUEST_LOCALDISK} \
-v ${HOST_MIRROR_DIR}:/import/mirrors:ro \
-v /sys/fs/cgroup:/sys/fs/cgroup:ro \
-v ~/.ssh:/mySSH:ro \
-e "container=docker" \
--security-opt seccomp=unconfined \
${TC_CONTAINER_TAG}
Running a Shell Inside the Container
Since running the container does not return to a shell prompt the exec into the container must be done from a different shell:
or by hand:
Notes:
- The above will reusult in a running container in systemd mode. It will have NO login.
- I tend to use tmux to keep a group of shells related to the build container
--user=${USER}
is the default username, setMYUNAME
inbuildrc
to change it.
Stop the Container
or by hand:
What to do to build from WITHIN the container
To make git cloning less painful
To start a fresh source tree
Instructions
Initialize the source tree.
cd $MY_REPO_ROOT_DIR
repo init -u https://opendev.org/starlingx/manifest.git -m default.xml
repo sync
To generate centos-repo
The centos-repo is a set of symbolic links to the packages in the mirror and the mock configuration file. It is needed to create these links if this is the first build or the mirror has been updated.
Where the argument to the script is the path of the mirror.
To build all packages:
To generate local-repo:
The local-repo has the dependency information that sequences the build order; To generate or update the information the following command needs to be executed after building modified or new packages.
To make an iso:
First time build
The entire project builds as a bootable image which means that the resulting ISO needs the boot files (initrd, vmlinuz, etc) that are also built by this build system. The symptom of this issue is that even if the build is successful, the ISO will be unable to boot.
For more specific instructions on how to solve this issue, please the
README on installer
folder in metal
repository.
WARNING HACK WARNING
Due to a lack of full udev support in the current build container, you need to do the following:
if you see complaints about udisksctl not being able to setup the loop device or not being able to mount it, you need to make sure the build-tools/update-efiboot-image is linked to the one in /usr/local/bin
Troubleshooting
if you see:
it's a docker bug. just kill the container and restart the it using a different name.
- I usually switch between -centos-builder and -centos-builder2. It's some kind of timeout (bind?) issue.