scylladb/tools/toolchain

Official toolchain for ScyllaDB

While we aim to build out-of-the-box on recent distributions, this isn't always possible and not everyone runs a recent distribution. For this reason a version-controlled toolchain is provided as a docker image.

Quick start

If your workstation supports docker (without requiring sudo), you can build and run Scylla easily without setting up the build dependencies beforehand:

./tools/toolchain/dbuild ./configure.py
./tools/toolchain/dbuild ninja build/release/scylla
./tools/toolchain/dbuild ./build/release/scylla --developer-mode 1

The dbuild script

The script dbuild allows you to run any command in that toolchain with the working directory mounted:

./tools/toolchain/dbuild ./configure.py
./tools/toolchain/dbuild ninja

The script will bind-mount ~/.cache and ~/.config so sccache within the container will access a cache directory on the host, so the cache is persistent across runs.

You can adjust the docker run command by adding more flags before the command to be executed, separating the flags and the command with --. This can be useful to attach more volumes (say, for /var/lib/scylla) and to set environment variables. For example, to mount /var/lib/scylla:

./tools/toolchain/dbuild -e MYVAR=foo -v $HOME/data:/var/lib/scylla:z -- ninja

To pass the same options to every run of dbuild, put them in the file ~/.config/scylladb/dbuild, which should contain a bash array assignment:

SCYLLADB_DBUILD=(-e PATH=/usr/lib64/ccache:/usr/bin:/usr/local/bin -v $HOME/.ccache:$HOME/.ccache:z)

The script also works from other directories, so if you have scylla-ccm checked out alongside scylla, you can write

../scylla/tools/toolchain/dbuild ./ccm ...

You will have access to both scylla and scylla-ccm in the container.

Interactive mode is also supported: running dbuild with no arguments will drop you into a shell, with all of the toolchain accessible.

Obtaining the current toolchain

The toolchain is stored in a file called tools/toolchain/image. Normally, dbuild will fetch the toolchain automatically. If you want to access the toolchain explicitly, pull that image:

docker pull $(<tools/toolchain/image)

Building the toolchain

If you add dependencies (to install-dependencies.sh or seastar/install-dependencies.sh) you should update the toolchain.

Run the command

podman build --no-cache --pull -f tools/toolchain/Dockerfile .

and use the resulting image.

Publishing an image

If you're a maintainer, you can tag the image and push it using podman push. Tags follow the format scylladb/scylla-toolchain:fedora-29-[branch-3.0-]20181128.

For master toolchains, the branch designation is omitted. In a branch, if there is a need to update a toolchain, the branch designation is added to the tag to avoid ambiguity.

Publishing an image is complicated since multiple architectures are supported. There are two procedures, one using emulation (can run on any x86 machine) and another using native systems, which requires access to aarch64 and s390x machines.

The sources for the toolchain are Internet builds of open-source projects, based on the current Fedora release for most packages, with supplements from pip (Python), cargo (Rust) and other binary repositories for projects not packaged by Fedora. We also package a ScyllaDB build of clang that is optimized for faster compilation, see tools/toolchain/optimized_clang.sh.

Because the clang binary is self-packaged, there are different procedures depending on whether the clang version changed since the last toolchain generation or not.

To obtain the clang version, use ./tools/toolchain/dbuild clang --version for the current toolchain, and podman run --rm docker.io/fedora:<version> dnf info clang for the to-be-packaged version. If they are different, you must use the procedure that also regenerates clang.

Emulated publishing procedure (slow, no clang regeneration)

1. Pick a new name for the image (in tools/toolchain/image) and commit it. The commit updating install-dependencies.sh should include the toolchain change, for atomicity. Do not push the commit to next yet.
2. Run tools/toolchain/prepare --clang-build-mode INSTALL_FROM --clang-archive-x86_64 <filename to the archive> --clang-archive-aarch64 <filename to the archive> and wait. The clang archive needs to be downloaded prior to build. It requires buildah and qemu-user-static to be installed (and will complain if they are not). The clang archive is recorded in each commit that changes the toolchain (git log -1 tools/toolchain/image). The URLs point to an object storage bucket we maintain.
3. Publish the image using the instructions printed by the previous step.
4. Push the next branch that refers to the new toolchain.

Native publishing procedure (complicated, no clang regeneration)

1. Pick a new name for the image (in tools/toolchain/image) and commit it. The commit updating install-dependencies.sh should include the toolchain change, for atomicity. Do not push the commit to next yet.
2. Push the commit to a personal repository/branch.
3. Perform the following on an x86 and an ARM machine:
   1. check out the branch containing the new toolchain name
   2. Run git submodule update --init --recursive to make sure all the submodules are synchronized
   3. Run tools/toolchain/prepare --clang-build-mode INSTALL_FROM --clang-archive-x86_64 <filename to the archive> --clang-archive-aarch64 <filename to the archive> --disable-multiarch. This should complete relatively quickly. The clang archive is recorded in each commit that changes the toolchain (git log -1 tools/toolchain/image). The URLs point to an object storage bucket we maintain.
4. Now, create a multiarch image with the following:
   1. Push one of the images using the podman manifest push command suggested by tools/toolchain/prepare.
   2. For the other image, first merge the other image into it. This is done by using the command from step 1, but replacing push with add. For example, if in step 1 you pushed the x86_64 image, in step 2 you add the x86_64 image to the local aarch64 image. This creates a local image supporting the two architectures.
   3. Push the combined image using the podman manifest push command suggested by tools/toolchain/prepare. This replaces the single-architecture image with a two-architecture image.
5. Now push the commit that updated the toolchain with git push.

Native publishing procedure (complicated, with clang regeneration)

1. Pick a new name for the image (in tools/toolchain/image) and commit it. The commit updating install-dependencies.sh should include the toolchain change, for atomicity. Do not push the commit to next yet.
2. In tools/toolchain/optimized_clang.sh, adjust the variable LLVM_CLANG_TAG to point to the version of clang you want to build. It should match what is available in Fedora at this point in time. Amend the commit with this, but don't push it to next yet.
3. Push the commit to a personal repository/branch.
4. Perform the following on an x86 and an ARM machine:
   1. check out the branch containing the new toolchain name
   2. Run git submodule update --init --recursive to make sure all the submodules are synchronized
   3. Make sure the clang generation directories are removed: ./build_profile and ./clang_build
   4. Run tools/toolchain/prepare --clang-build-mode INSTALL --clang-archive-x86_64 <filename to the archive> --clang-archive-aarch64 <filename to the archive>. This will be quite slow as clang and scylla are built multiple times. Pick a new name for the clang archive based on previous names. The names are recorded in each commit that changes the toolchain (git log -1 tools/toolchain/image). The names include the Fedora version this is built on, the clang version, and the architecture. The new name must be unique. The URLs point to an object storage bucket we maintain.
   5. Upload the generated clang image to its URL. Use gsutil cp <filename> <GSURL> where GSURL is the same as URL except the protocol is gs instead of https.
5. Now, create a multiarch image with the following:
   1. Push one of the images using the podman manifest push command suggested by tools/toolchain/prepare.
   2. For the other image, first merge the other image into it. This is done by using the command from step 1, but replacing push with add. For example, if in step 1 you pushed the x86_64 image, in step 2 you add the x86_64 image to the local aarch64 image. This creates a local image supporting the two architectures.
   3. Push the combined image using the podman manifest push command suggested by tools/toolchain/prepare. This replaces the single-architecture image with a two-architecture image.
6. Now push the commit that updated the toolchain with git push. Remember to record the clang archive URLs for future reference.

Troubleshooting

When running sudo inside the container fails like this:

$ tools/toolchain/dbuild /bin/bash
bash-4.4$ sudo dnf install gdb
sudo: unknown uid 1000: who are you?

You can work it around by disabling SELinux on the host before running dbuild:

$ sudo setenforce 0

The future toolchain

To prevent surprises when new Fedora/libstdc++/clang are made available, a spec for a "future" toolchain is available in tools/toolchain/future.dockerfile. See that file for details.