Christoph Hellwig wrote: > On Sat, May 22, 2004 at 10:00:25PM +0200, Thiemo Seufer wrote: > > The real problem isn't the unified source tree, but how to create > > kernels from it. A single debian source package for all architectures > > means there is no way to keep an older version for one of them. So we > > lose the flexibility to skip known bad versions for one architecture. > > Umm, I don't thing known bad version for architecture makes any sense. > There's known bad changes and we can back changes out if nessecary.
An Example: Starting with 2.4.20, the mips cache code underwent major changes. This broke the r4k-kn04 subarchitecture, and some of the r4k-ip22 machines. Most of the latter were fixed relatively quickly, with improved performance, but some remained flaky, and the former remained broken until 2.4.25, when the mips kernels were synced up again. A single kernel package would elevate this from the architecture level to a problem for whole debian. This was for a supposedly stable kernel. For architectures catching up late I expect similiar things to happen again, at a time mainline development already moved to the next version, so it makes little sense to invest time for a fully portable backport (This will hopefully mostly hit some drivers). > > We also lose the flexibility to make an arch-specific update without > > forcing everyone to upgrade their kernels for nothing. > > Handle them the same as everywhere? Urgend per-arch changes become > NMU, if anything interesting for the others happened just make it a > normal release. You can't do a source NMU for a single arch. That is, unless you are proposing to do arch-specific source packages, which would waste about 500 MB mirror space and traffic per version. Thiemo
