Re: Bug#1040062: dpkg-dev: Please drop pie-{compile,link}.spec
On Tue, Jul 04, 2023 at 09:23:43AM +0200, Guillem Jover wrote:
> Hi!
Hi Guillem!
> On Sun, 2023-07-02 at 00:02:46 +0300, Adrian Bunk wrote:
>...
> > There are some problems with this:
> >
> > 1. PIE should either be default or not be used
> >
> > I suspect x32 might be able to default to PIE without problems
> > (there might just not be enough interest left to change the default).
> >
> > On alpha the toolchain has already become quite brittle
> > with frequent issues like (reproducible) linker segfaults,
> > any variations that affects the toolchain are bad.
> >
> > It is for the port maintainers to decide whether or not PIE
> > is considered stable on a port, and accordingly either make
> > it default (which also avoids the other issues below) or not.
> >
> > It is clear that a non-PIE architecture would no longer be
> > considered suitable as release architecture.
>
> In general the way this is handled in dpkg, is that if the flags
> supposedly work on that arch they are allowed, but if they are not
> supported or are broken then they are masked.
PIE is unusual in being enabled by default on all release architectures,
but not being enabled by default on some non-release architectures.
With some teams like Debian Med putting hardening=+all into every source
package, 20% of the source packages in the archive have it already set.
This makes both sides of the PIE/non-PIE archive split on the affected
semi-PIE architectures huge.
> > 2. It causes weird issues on undersupported architectures
> >
> > gluegen2 passes LDFLAGS to ld instead of gcc.
> >
> > Several packages have relocation errors only on affected
> > architectures.
> >
> > ...
> >
> > Such issues could be debugged and fixed, but in practice
> > trying to handle such issues that happen only with
> > pie-{compile,link}.spec creates additional work that frustrates
> > the few people keeping these non-release architectures alive.
>
> Regardless of this report, I think this would still be worthwhile,
> as otherwise you cannot for example disable them globally on arches
> where it is a builtin in the compiler (as those will also need the
> spec files.
Packages like python3.*-nopie could also set -fno-pie/-no-pie in
CFLAGS/LDFLAGS manually instead of using hardening=-pie, but different
from hardening=+all disabling PIE is not something that has unexpected
effects only on some ports architectures.
>...
> > 3. It breaks some cases of static linking
> >
> > Linking a package with hardening=+all against a static library
> > from a package not using hardening=+all cannot work on the
> > affected architectures.
> >
> > Static linking is relatively rare, but I remember requesting binNMUs
> > for static linking cases to fix FTBFS on release architectures when
> > the default changed before stretch.
>
> Hmm, ah or even packages not respecting DEB_BUILD_OPTIONS, right.
Bugs would fall under point 2.
Point 3 is about cases that cannot work despite no package doing
anything wrong.
Many packages needed fixing (or still have to be fixed) to enable
PIE because they had set hardening=+all,-pie before the default
changed in stretch. When all architectures defaulted to non-PIE,
hardening=+all did also cause PIE-related FTBFS on release architectures.
Semi-PIE architectures still have such problems between packages on
different sides of the PIE/non-PIE archive split.
> > Please drop pie-{compile,link}.spec, on the architectures
> > where it has any effect it is doing more harm than good.
>
> For example hppa has pie masked for build flags. If the porters for
> alpha and/or ia64 consider that they should also get pie masked for
> those arches, I'm fine doing the changes. Although that means on those
> ports it will not be possible to enable pie at all, even if asked for
> explicitly, as in «hardening=+pie».
Semi-PIE non-release architectures shouldn't exist, PIE on an
architecture should be a binary option decided by the porters
for this architecture.
If there are any porters left on an architecture and they consider PIE
stable, they can always ask the gcc maintainer to change the default.[1]
>...
> > The lowest effort fix would be to patch debian/rules of affected
> > packages to disable hardening=+pie on affected architectures,
> > but that would still be spending time on working around a problem
> > that shouldn't exist.
>
> Yeah, that does not look like the right thing to be spending time on.
>...
As long as we have at least one semi-PIE architecture, this is the only
realistic option for porters. If the porters for alpha and/or ia64 want
to have pie masked, such changes will still be required for x32.
> Thanks,
> Guillem
Thanks
Adrian
[1] no transition with rebuilds required, in the rare cases of FTBFS
when linking with a non-PIE static library this static library
package could be binNMU'ed when the problem appears
Reply to: