Toolchain, Build and Packaging of Base

BSDCam session, 1pm - 2:45pm, August 26.

Topics for Discussion

LLVM Instrumentation

Traditional build turns a .c file into a .o file and then link the .o files. In the LLVM IR instrumentation world, you turn .c files into .ll/bc files, modify them, link them (then run link-time optimisation), then generate .o files, then link. In SOAAP, the instrumentation happens after the IR-link phase, for TESLA we run static analysis of the .c files, then instrument individual files.

There's no good way in make to express different routes from .c to .o. We want to be able to compose many of these things.

Things like TESLA and SOAAP require annotations, which are macros and can be compiled out. Things like the clang static analyser or KINT just check C-language semantics.

Some things in ports are also interesting to check, so it would be nice if we had a ${CC} and ${LD} that invoked some of these as part of a normal build. Some check API usage, including Apple kernel API checks, which could have some things.

How much annotation are we willing to accept in the kernel for documenting invariants and so on? Warren reminds us that having specification close to the source code is very useful. Robert proposes that we're willing to accept any changes that provide specs into the code, as long as we can run the checks that verify them.

Build Configuration

Configuration should be done by setting variables, not by executable make rules.

Kernel builds are currently very special - a generated rule for every .c file - but the kernel is just an ELF image.

Tinderboxes

Can we do the build-as-non-root and then boot in QEMU and so on to run regression tests? Should the tinderboxes run the static analyser?

DES is going to add scheduling to tinderboxes. For example, when secteam wants to test a security bug fix they will get to interrupt normal builds and build things.

Tinderbox should become a generic infrastructure for doing builds, testing, and static analysis. DES volunteers to own this gradual progression, but requires hosting for some machines.

Incremental Build

Touching certain headers forces a full rebuild. Our dependency tracking is hard because some headers are in src/include.

Probably a better discussion for Malta?

External Toolchains

Current version allows ${XCC}, ${XLD} and so on to replace compiler, binutils, and so on. This is important because new clang will need building with clang soon (clang 3.4 uses C++11 so won't build with gcc shipped in 9.x).

Currently no support for using X* to build the toolchain. This is useful for bootstrapping and so on.

How useful is this for building FreeBSD on Linux? You'll need to build a new LLVM (unless you have one in packages) and a binutils that targets FreeBSD. Requires bmake and so on as well - someone must enumerate everything that we need for the build.

GCC Removal

Our GCC and libstdc++ are from 2007. They predate the current C/C++ standards and are a pain to maintain. Big C++ ports like LibreOffice don't work with our libstdc++. We really, really don't want to be supporting a 2007 release in 2017 (by which time C++14 will be out) when 10.x is EOL'd.

Proposed plan:

No objections (except lots of people on mailing lists)...

Building old FreeBSD on Head

Brooks says: Yes. Umm. Probably work if you use fmake from ports... Bapt says it works. Or, at least, did. Documentation is probably required - the switch to bmake broke it (a yacc update broke it previously). Is this something we want to be checking automatically? We don't have automatic testing and we don't promise that it will work, but it's quite nice when it works.

Bootstrap infrastructure is not very principled. NetBSD's build-on-Linux infrastructure did this, but it's a lot of work.

Packaging Base

Bapt already builds the FreeBSD base system as pkgng packages. The package list is generated from the mtree that Brooks' work made it easier to generate. Configuration file upgrades are difficult - currently pkgng only allows blowing away the old ones or keeping them, no merging. Finer granularity distribution sets would be nice too, for example libraries, allowing easier testing. Origins are paths, would be nice for have FreeBSD/ports/whatever, FreeBSD/base/foo. Building infrastructure should ensure that we have reproduceable builds. Delta packages are also nice for not overloading the CDN, but they need a lot of thinking...

Adding users, for example, should be done by pkgng not by modifying files in /etc. Also allows renumbering of UIDs. Bapt plans on adding Fossil's merge code to pkgng for merging config files. Testing upgrades from latest-release to -head would be very good.

Adrian says /etc should be read-only and user-modifiable config files should be elsewhere, although making sure that /etc/defaults the read-only thing might be a good idea... Being better at .d directories for things is a good idea, things like pam.d exists but isn't defined for this sort of thing. How do we handle cases where formats (e.g. pf.conf) change significantly across versions? Removing non-config things (e.g. rc.d) from /etc would make sense.

Replacing freebsd-update with pkgng requires some key management fun and delta packages. Teaching freebsd-update to update pkgng-based installs is a middle ground.

Bumping a shared library version is hard because removing shared libraries doesn't happen automatically. Can we use PROVIDES so that ports libraries can either depend on the base system version or a ports version. pkgng can detect libraries that are required and find the provider packages for libraries that are not installed.

ABI compatibility decisions are hard. The pkgng ABI compatibility strings are not very flexible.

Adding /usr/lib/private that is not in LD_LIBRARY_PATH (explicit -rpath for things in the base system that use them) would be useful for cleaning up the base system.

Malta TODOs

201308DevSummit/ToolchainAndBuild (last edited 2013-09-25 15:57:03 by BrooksDavis)