Control: tags -1 + confirmed

On Fri, 2021-09-17 at 00:16 +0100, Colin Watson wrote:
> https://bugs.debian.org/984533 and its clone
> https://bugs.debian.org/985572 showed a buster-to-bullseye upgrade
> bug in which debconf was unable to execute whiptail between unpacking
> the new libslang2 and unpacking the new libc6.  Part of the fix for
> that involved adjusting debconf in bullseye to detect this situation
> and gracefully fall back to text mode.  The other part of the fix was
> to adjust libc6.preinst to do something similar, in case debconf has
> not yet been upgraded or the running debconf frontend is still from
> an old version.
> 
> Unfortunately, the code in libc6.preinst was somewhat broken,
> resulting in buster-to-bullseye upgrades that hang in some
> situations.  We only noticed this after bullseye released because the
> breakage is only apparent with certain package sets that provoke apt
> into choosing particular upgrade orderings; even with this, I only
> know of it happening for people who run "apt upgrade" as a separate
> step before "apt full-upgrade" (IMO unnecessarily, but it seems to be
> some people's habit).  https://bugs.debian.org/994042 has an analysis
> of the situation situation and a reproduction recipe.
> 
> While fixing this particular upgrade bug requires fixing
> libc6.preinst (because its broken logic happens before debconf has an
> opportunity to decide what to do), it's possible for apt to attempt
> to unpack some *other* package between unpacking the new libslang2
> and the new libc6 which also tries to use debconf in its preinst, and
> that would run into a similar bug.  (I admit to not having a concrete
> example of such an upgrade ordering.)  The only way to fix that
> situation is to cherry- pick the fix for #985572 into debconf in
> buster.  As Aurelien points out in
> #994042, we can't rely on people having applied all buster updates
> before starting the upgrade to bullseye.  Nevertheless, I think this
> change would make upgrades more robust, since debconf must take great
> care not to crash like this.
> 

Please go ahead; thanks.

Regards,

Adam

Reply via email to