8:57 a.m.
On 09/30/22 15:35, Daniel P. Berrangé wrote:
On Fri, Sep 30, 2022 at 03:23:31PM +0200, Laszlo Ersek wrote:
> On 09/30/22 14:11, Richard W.M. Jones wrote:
>> On Fri, Sep 30, 2022 at 12:56:40PM +0200, Laszlo Ersek wrote:
>
>>> (2d) I started icewm with "icewm --replace" (as recommended by the
error
>>> message from (2c)), and lo and behold, two changes had come into effect:
>>>
>>> - the spinner started working
>>
>> This is strange ..? Did librsvg get installed at some point?
>
> So, "ldd `which icewm`" reports that icewm depends on librsvg-2.so.2;
> therefore it's not surprising that RPM generated the proper Requires:
> directive for the binary icewm package. I guess with Metacity, this is
> not the case -- which should be fine, if Metacity itself does not use
> librsvg2.
>
> So it looks like a package dependency bug somewhere in or around gtk3; I
> guess it has a kind of plugin architecture, and the SPEC file does not
> properly capture the (dynamic) dependency.
>
> Interestingly, in the usptream gtk repository, on the gtk-3-24 branch,
> the file ".gitlab-ci/fedora-gtk3.Dockerfile" spells out
> "librsvg2-common" and "librsvg2", from commit 81c4fa5d505ec
(which is
> quite non-descript in this regard). So it is a *known* hidden dependency
> in a way. Sigh. Packaging bug then? :/
"its complicated" :-)
GTK doesn't directly know about any image formats. It is built on top
of GDK-Pixbuf which is a library that provides an API for loading images,
with plugins for various formats, most outsourced to other packages.
librsvg2 provides a plugin for GDK-Pixbuf for SVG files.
Meanwhile the usage of SVG files doesn't actually exist in GTK code either,
because GTK widget rendering is theme based, and both widget and icon themes
are separate. Even when the widget theme requests an icon, it doesn't specify
a format, because icon themes are pluggable too and it will look for icons
in whatever format happens to exist on disk.
The icon theme doesn't include a dep on librsvg2 either, because the icons'
package doesn't dictate what is used to load them.
IOW, it is a tragedy of a highly modular framework that nothing individually
needs/want to have a dep on librsvg2, but the combined work does need such
a dep.
In a typical Fedora/RHEL install, all the right bits will get installed
either because of deps elsewhere in the default "desktop" install profile
group, or because of comps groups listing librsvg. p2v though is picking
a minimal set of individual packages and so misses the dep.
This is horrible. The librsvg2 package provides this file:
/usr/lib64/gdk-pixbuf-2.0/2.10.0/loaders/libpixbufloader-svg.so
which, if present, enables the whole stack to work, but if absent,
breaks the stack without a peep.
This is the same terrible "push" (or "registration") model (rather
than
the "pull") model that plagues systemd: if a kernel module is missing
from the initrd that's needed for driving a device or a filesystem, the
boot gets stuck without any indication of what *should* happen (i.e.
what dependency is not being satisfied). There is no technical
dependency whatsoever, it's just that an event does not occur that
"might" otherwise "cause progress" from a global perspective.
The exact same silly pattern exists in UEFI, with PPI and protocol GUID
dependencies. Very powerful as long as you are developing a particular
module. Once you forget about it, *and* a particular protocol GUID is
provided outside of your current repository (so you can't simply grep
for it), heaven help you resolve the stuck boot (as, of course, it is
not easy to all to get the set of those protocol GUIDs that *might* get
the DXE dispatcher unstuck via their registration in the protocol database).
How is it better that the librsvg2 package, which is a *low level
package*, includes a *semantic dependency* (basically: *knowledge*) of
one of its high-profile comsumers (namely, gdk-pixbuf) -- a high level
package --, and not the other way around? Since when is it a good idea
to encode reverse dependencies in packages? The mind boggles.
All of this registration stuff is for the 5% more comfort of
programmers, at the 90% more discomfort of sysadmins / users.
Laszlo