On Mon, 2012-01-16 at 00:07 +0100, Vladimir 'φ-coder/phcoder' Serbinenko wrote: > On 14.01.2012 04:53, Ben Hutchings wrote: > > On Sat, 2012-01-14 at 02:06 +0100, Stéphane Aulery wrote: > >> Package: linux-2.6 > >> Version: 3.1.8-2 > >> Severity: normal > >> > >> Dear Maintainer, > >> > >> When I have updated package linux-image-3.1.0-1-686-pae 3.1.6-1 to > >> 3.1.8-2, aptitude sent some error messages like this : > >> > >> dm_mod: Unknown symbol scsi_verify_blk_ioctl (err 0) > > [...] > > > > It looks like this is due to GRUB running os-prober. > > > > We do not bump the kernel ABI when adding symbols, because that is > > backward-compatible with modules built for the earlier version. But > > loading new modules is liable to fail after an upgrade until the new > > kernel is running (i.e. until the next reboot.) So long as os-prober > > relies on loading a bunch of modules, GRUB *must not* run os-prober in > > its kernel postinst hook. > That doesn't sound as a reasonable requirement. As I see this, such an > install, overwriting older modules botches currently running kernel > which can lead to other unexpected failures as well.
It doesn't; the kernel runs just fine. This is just like upgrading libc6, when you might have to restart daemons that use NSS. > I think it would be > reasonable to have such updates versioned and have 3.1.X or 3.1.0-X as > Linux version rather than 3.1.0 No, because: 1. /boot would fill up really quickly (don't tell me about apt-get autoremove; it specifically doesn't work for kernel images). 2. This would break out-of-tree modules every time. Both of this problems are solveable, but it's not something we can just change unilaterally. Ben. -- Ben Hutchings When in doubt, use brute force. - Ken Thompson
signature.asc
Description: This is a digitally signed message part
