To build a package with
pmbootstrap, you will only need the command
pmbootstrap build hello-world. Optionally with
--force to build even if the package was already built, and with
--arch=armhf to build for another architecture, replace
armhf accordingly. This wiki page lists some more information about what is going on in the background.
The initramfs shows the boot splash images, and allows the root partition to be unlocked (currently via telnet). You can add a hook to inspect the initramfs running on the device, as described in Inspecting the initramfs.
To rebuild the initramfs, run
mkinitfs inside the chroot with the right parameters (or - much easier -
pmbootstrap initfs build). This gets done automatically, whenever a new kernel gets installed, or when the
postmarketos-mkinitfs package gets installed the first time.
pmbootstrap vs. abuild
pmbootstrapcan cross-compile out of the box, utilizing different chroots as needed (see below for details)
pmbootstrapdoes dependency parsing on its own (so it works across the
aportsfolder and the binary repository and can detect across chroots when a package is outdated and explicitly install them).
pmbootstrapdoes not honor operators in dependencies, such as:
!. These simply get ignored (
!packages don't count as dependencies). This may lead to errors, if it does please report them. However, since we're calling
apkto install the packages, it does the real dependency checking and so far it's working well enough.
APKBUILDfiles on its own.
APKINDEXparsing is considered to be pretty good (because the format is dead simple!)
APKBUILDparsing would require a shell to be done perfectly (which would in turn kill performance). The way it is implemented right now, is that the variables we care about are hardcoded inside the
pbm/config/__config__.pyor if not possible otherwise directly in
pmb/parse/apkbuild.py. That is really fast and works for all packages we care about. If it breaks somewhere, it should be easy to patch.
pmbootstrapdoes not remove build dependencies after a build is done (except when using
--strict). This is for performance reasons - if you want a clean start, run
pmbootstraphas a hack right now, that
gzipalways uses weak compression (also for speed)
These are the cross-compile types supported.
pmb.build.autodetect.crosscompile() figures out, which one is the right one for each build.
None: The target architecture is the same as the "native" architecture (e.g. compiling
"native": the build system of the package understands cross-compiling, like all kernel packages. We can use the native chroot with the cross-compiler inside that chroot for maximum speed.
"distcc": the build system of the package does not understand cross-compiling, so we run the whole compiling process inside a chroot with the target architecture (with qemu user mode emulation, this is slow). We avoid using qemu emulation for the compiler though, because we use the cross-compiler inside the native chroot through distcc (which was originally meant to share compiling efforts across the network, but it works for this case. Arch Linux ARM also does this and has a detailed description of how to set this up manually.) This method became a lot slower, as we had to run distcc behind a locally running SSH server, in order to avoid a security issue (!1649). This method is legacy, replaced by "crossdirect" below.
"crossdirect": similar to distcc, but the native chroot gets mounted in the foreign arch chroot, and with some magic we can redirect everything to the cross compilers without having distcc running (#1731). Therefore we side-step the security issue and the performance problems. As of writing, this method is brand new and only available in pmbootstrap master. If any problems arise, you can switch to "distcc-sshd" again by running
pmbootstrap --no-crossdirect ....
- Use "native" for (almost) all packages?
- Distcc: use pump mode for possible performance improvement
- Distcc: has integrated timeout, slowing down our builds ._.
pmbootstrap uses various caches. They can all be found inside the
work folder, and start with
cache_. All cache folders get mounted to the appropriate chroots, depending on $ARCH. They are shared among the chroots, when it makes sense (e.g.
cache_apk_$ARCH: APK files from binary repositories (see also: Local APK cache)
cache_ccache_$ARCH: ccache: Whenever you compile something with
pmbootstrap, the output gets cached in this folder (depending on the architecture). When you compile the same code for the second time, the cached output gets used, thus saving you a lot of time (think of re-compiling kernels, because you want to test another kernel config option etc.)
cache_distfiles: Whenever you build a package,
abuild(which pmbootstrap wraps) will download the source files to the
distfilescache (and skip these downloads, when they already exist). The exact file name can be controlled inside the APKBUILD (more info).
pmbootstrapcan download git repositories. This gets used in
pmbootstrap aportgen, which copies a package (aka. aport) from Alpine Linux and customizes it (for example:
gcc-armhf), so we inherit all patches and changes automatically, without much maintenance work. The git repos get stored inside this folder.
cache_http: This stores files, that get downloaded with
pmb.helpers.http.download(), so they don't need to be downloaded again every time. Currently, this gets used for the initial download of Alpine Linux' main repositories
APKINDEX.tar.gzand apk-tools-static (which is a static build of the
apkpackage manager, used to set up the chroot).
Why noarch packages are built for each architecture
noarch package, we could in theory compile it once and then create symlinks to all arches pointing to the package where we have created it. And previous versions of pmbootstrap used to do that. But that makes dependency parsing hard.
For example: device packages are always
noarch, so in theory you could compile the device package for your
armhf device in
x86_64, and then symlink it to the
armhf package folder. But it would still depend on the kernel, which is
armhf only for that device (assuming it is not mainlined, like most devices are right now). So it would not make sense to install it in
x86_64 anyway, and if you would want to build it, you could only really do it in
armhf anyway, because that's where the kernel is available.
So to simplify it, it is handled like in Alpine now; noarch packages just get built for each arch independently.