Re: systemd, ELF binaries and runtime dependency tracking


Khem Raj
 

On 6/1/21 9:13 AM, Luca Boccassi wrote:
On Tue, 2021-06-01 at 07:58 -0700, Khem Raj wrote:

On 5/31/21 3:40 PM, Paul Eggleton wrote:
Hi folks

Upstream in the systemd project, a proposal has been made to add a special
section to output ELF binaries to record soft runtime dependencies, so that
they could be read and utilised by distribution build systems such as ours
(they would be translated into RRECOMMENDS in our case). At the moment that
doesn't seem to have generated a huge amount of interest in the traditional
distro space, but would it be interesting for us?

https://github.com/systemd/systemd/pull/17416

For clarity, we (Microsoft) will volunteer to do the integration assuming the
above pull request gets reopened and merged, which is more likely if we
express our interest.
Finding dlopen dependencies is a neat idea, but it has to be accepted
cross distro, and also applications have to manually declare it in code
if I understand systemd's patch correctly. This will be hard to
accomplish as you can see changes are spread across apps from distro
point of view. Perhaps there is a smarter way of detecting adding them
in ELF spec itself and then have tools like linker help implement this
and also possibly collect the information or guide the users to achieve
this would be helpful.
Yes ideally ELF shared objects/the linker/the loader would support weak
symbols (like dylib on OSX). Unfortunately they do not, and it seems
there's no interest to add it becasue there's no concrete use case that
shows it's useful. But that cannot happen until there's some support
for it. Chicken and egg...
right and thats why I will be reluctant to go too far at distro level unless there is general interest in wider communities as it can make us an island.

There have been lots of theoretical discussions about pros and cons,
and my hope was that if at least one distro could find it useful, and
could show that it is in practice useful and the theoretical issues are
not that problematic and could be solved, others would follow suit.
you could tool it as a packageconfig for systemd alone and run with it and see how it pans out.

So leaving aside other distros, is this something that would concretely
benefit the Yocto project for handling the systemd recipe? There are
currently 12 dlopen()-based optional dependencies in systemd, and the
number grows with each release.
we could certainly try that, provided systemd upstream is supportive of it.

Join yocto@lists.yoctoproject.org to automatically receive all group messages.