[pulseaudio-discuss] [PATCH v2 1/2] conf-parser: add support for .d directories

[Tanu Kaskinen]

On Sat, 2015-12-05 at 19:50 +0200, Tanu Kaskinen wrote:
> On Sat, 2015-12-05 at 12:24 -0300, Felipe Sateler wrote:
> > On 4 December 2015 at 09:46, Tanu Kaskinen <tanuk at iki.fi> wrote:
> > > On Fri, 2015-12-04 at 09:29 -0300, Felipe Sateler wrote:
> > > > On 4 December 2015 at 02:58, Tanu Kaskinen <tanuk at iki.fi> wrote:
> > > > > + * If use_dot_d is true, then before parsing the file named by the filename
> > > > > + * argument, the function will parse all files ending with ".conf" in
> > > > > + * alphabetical order from a directory whose name is filename + ".d", if such
> > > > > + * directory exists.
> > > > 
> > > > This is opposite to how other software (eg systemd) work: first the
> > > > main file is read, and then the .d/*.conf files are read to override
> > > > the configurations.
> > > > 
> > > > So if a distro default has an uncommented line, the .d should override
> > > > that line. This would allow (in the future) to extend the search path
> > > > to /usr/share/pulseaudio, and ship the defaults (uncommented) there.
> > > > Then local configuration can be done via dropin files.
> > > 
> > > All advice given to users so far has been to modify the main files,
> > > because nothing else has existed. If users put their changes to the
> > > main files and distros put their changes to the .d files, that will be
> > > compatible with existing advice on the internet.
> > 
> > The alternative does not preclude editing the main file either.
> > 
> > Maybe it would be best to step back a bit. Desirable properties for
> > the configuration loading are (IMO, of course):
> > 
> > 1. Any upstream default should be easily overridable by a downstream
> >    distributor.
> > 2. Any upstream/distribution default should be easily overridable by a
> >    local admin.
> > 3. Any upstream/distribution/admin default should be easily
> >    overridable by a local user.
> > 4. Admin overriding a setting should not require writing all the options again.
> > 5. Local user overriding a setting should not require writing all the
> >     options again.
> > 6. Third party packages should be able to provide configuration,
> >     without overriding other options.
> > 7. Local admin/user should be able to override such configuration.
> > 8. Local admin changes to configuration should be clearly visible.
> 
> I agree on all of these.
> 
> > Currently, pulseaudio only covers 1-3. Your proposed patch addresses
> > point 6, but only in the case the distro default is "everything
> > commented".
> 
> I guess you mean that my proposal breaks if the distro overrides
> upstream defaults in the main file, because packages can't override
> those in .d files.
> 
> My suggestion is to not patch the main files in distros. Instead, put
> distro defaults that diverge from the upstream defaults to .d files.
> Use 00- prefix to make these defaults have the lowest precedence, and
> use higher numbers in third-party packages.
> 
> > If the load order is inverted, then there is a path to achieving all 8
> > properties:
> > 
> > - If the dropins were read after the main configuration file, then 6
> >   would be fully addressed.
> > - If then the path loading is extended to load dropins in the entire
> >   search path, keeping only the highest precedence one, then point 4 and
> >   5 would be addressed.
> 
> I guess you mean merging the file list in all .d directories before
> starting processing. For example:
> 
> $datadir/client.conf.d/10-foo.conf
> $sysconfdir/client.conf.d/30-bar.conf
> $userconfdir/client.conf.d/20-blub.conf
> 
> The processing order would be 10-foo.conf -> 20-blub.conf -> 30-
> bar.conf. That ordering is needed when dealing with scripts (like udev
> configuration): 10-foo.conf may set some variable that the admin uses
> as a condition in 30-bar.conf. If the user is happy with the code in
> 30-bar.conf, but wants to change the variable set in 10-foo.conf before
> 30-bar.conf uses it, the user has to be able to insert code in between.
> 
> When dealing with plain key-value stuff without conditional constructs,
> however, this kind of "insert in between" support is not needed. In
> this situation I prefer to just load everything from $datadir first,
> then from $sysconfdir, and finally from $userconfdir. Simple to
> understand, simple to implement.
> 
> > - If then the loading stops looking at the search path at the first
> >   file with same name found, then point 7 would be addressed.
> > - And then point 8 can be implemented by shipping the full
> >   configuration file, uncommented, in $datadir. Thus only admin changes
> >   would live in /etc, and local user changes would live in ~/.config.
> >  
> > Not inverting the change, though, makes it cumbersome to provide a
> > distro default that can be overriden in a package: some arbitrary
> > prefix should be determined so that packages ship their defaults
> > ordered after that.
> 
> I'm not sure what you mean. Do you mean that if distro defaults are in
> client.conf.d/00-distrodefaults.conf, it's hard to make sure that
> packages ship files that are ordered after that?
> 
> > This will necessarily be distro-dependent. It also
> > makes it harder to have properties 4 and 5, since a setting written in
> > the main file necessarily requires modifying or copying it.
> 
> Why would overriding a setting in the main file require modifying or
> copying the file? Do you refer to the masking, i.e. if
> $userconfir/client.conf exists, $sysconfdir/client.conf won't be read?
> I want to get rid of the masking eventually. Both files should be read.
> 
> > This is the way many other packages work (eg, systemd and dbus), so
> > people already familiar with that configuration would be surprised by
> > pa's scheme.
> 
> You mean that $userconfdir/mainfile.conf masks
> $sysconfdir/mainfile.conf in systemd? If so, I certainly don't want to
> (continue to) copy systemd.
> 
> > There is a small wrinkle to observe with this load order: if a
> > package ships a setting in a dropin file, then the user needs to
> > override that dropin (or ship a new one with a higher ordering) for
> > his change to work, because the dropin is processed after the main
> > file. For example, a dropin in /etc that sets autospawn=no, will take
> > precedence over the main file in ~/.config/pulse that says
> > autospawn=yes.
> > 
> > This can be addressed by making the dropins from later in the
> > search path be loaded before the main file being read (eg, a dropin in
> > /etc would be loaded before reading ~/.config/pulse/client.conf). This
> > is not done AFAIK in other software; I'm not sure it is a good idea
> > to deviate.
> 
> I had some trouble following what scheme you were talking about in
> different parts of your mail, but I got the impression that you
> basically have two objections to my patch: the scheme is different from
> systemd, and it doesn't work when upstream ships its defaults in the
> main files. The solution to the latter is simple: ship the distro
> defaults in .d files.
> 
> Being different from systemd is a valid concern. However, if I
> understood corrently, the "systemd scheme" implies masking with the
> main files just like we currently do, and I think that sucks badly
> enough to justify not caring about the systemd scheme.

I'm starting to have doubts about the superiority of loading the main
file after the .d files... You have advocated putting the upstream
defaults to the main file under $datadir instead of having everything
commented out. I have dismissed that, because it doesn't have any
particular benefit when talking about client.conf and daemon.conf,
since they can be empty as well. But it occurred to me that the alsa
mixer files can't be empty by default. Overriding the alsa mixer
defaults from .d files wouldn't work with my original proposal.

Third-party packages aren't likely to extend the alsa mixer files
(these patches won't allow that anyway), but in principle I don't see
any reason why we couldn't some day have non-empty default
configuration files that should be overridable from .d files.

The problem with loading the main file before the .d files is that user
configuration should be writable to the main file without worrying that
the changes don't have any effect due to overrides in .d files that the
user isn't aware of. That was the main reason why I wrote the patch as
I did. But now I realize that in practice .d files don't get written to
$userconfdir behind the user's back, because the .d files get installed
in $sysconfdir. If $userconfdir has .d files, they are written by the
user, and if that happens, we can assume that the user knows the
semantics.

The same would apply to the admin-controlled files in $sysconfdir, if
we supported $datadir as a configuration source, but we don't. That
means that there may be confusion due to .d files being installed in
$sysconfdir, which override the admin config in the main file. That
makes me think that it would be best to first implement loading
configuration from $datadir, and also remove the masking, and only then
apply these patches (modified to reverse the file loading order). Is it
a good idea to wait for that to happen, though? Or should we apply
these patches first (again, modified to reverse order), and just live
with the possible confusion that .d files in $sysconfdir can cause for
admins?

-- 
Tanu