[RFC] Detect system XKB installation paths

[Dan Nicholson]

On Wed, Mar 24, 2010 at 11:42 AM, Yaakov (Cygwin/X)
<yselkowitz at users.sourceforge.net> wrote:
> On 2010-03-23 08:29, Dan Nicholson wrote:
>>
>> I've wanted to do this for a while, but there are a couple issues.
>
> FWIW, I'm trying to keep in mind several different scenarios:
>
> 1) where xserver is being built to update an existing version with the same
> prefix, e.g. by distributors.  This patch shouldn't change anything in that
> case.

Agreed.

> 2) where only xserver is built against system-installed dependencies but
> with a separate prefix, e.g. from git where no arguments are passed to
> autogen.sh (which then defaults to /usr/local vs. the system /usr). Without
> something like this, the server builds but fails to run because it expects
> these packages in its prefix.

Yeah, this is the one I've always wanted to fix. This is a spot where
xkb is a little special, though. You _can_ get the xserver to use your
system xkb files, but trying to use the system xkbcomp with local xkb
files can be problematic due to xkbcomp's handling of paths. This is
why "symlink to your host's xkbcomp in $xprefix/bin" is not actually a
useful suggestion. But that's a little orthogonal to this patch.

> 3) where a complete X.Org system is built in a separate prefix from the
> system-installed version, such as jhbuild.  Within jhbuild we could
> certainly add xkeyboard-config and xkbcomp as dependencies of xserver (if
> they are not already) without creating a hard dep for other scenarios.

If this patch lands, we would most certainly want to make them deps of
the xserver in jhbuild and build.sh.

> 4) where X.Org is being cross-compiled, we need to be sure not to pick up
> the build system's installation.
>
>> Why exclude cross compiling? Using PKG_CHECK_EXISTS or AC_PATH_PROG
>> have no problems in those situations.
>
> My concern was (4) above.  I'm not that familiar with cross-compiling; it
> makes sense that the pkg-config call should work, but wouldn't AC_PATH_PROG
> be prone to pick up the build system's xkbconf?

No more than situations 1-3 above, but I suppose that's the reason
AC_CHECK_FILE bails on cross compiling. Since we're actually encoding
runtime dependencies, I think we'd want to pick up the tools in the
cross prefix. Maybe it's safer to default to paths under $prefix in
that case. Let's leave it your way until someone that this actually
affects complains.

>> Check if DEFAULT_XKB_PATH is empty and set it to ${datadir}/X11/xkb if
>> so. Then we can fallback gracefully on older xkeyboard-config
>> installations.
>
> In what case would DEFAULT_XKB_PATH be empty?  The PKG_CHECK_EXISTS will do
> nothing if the .pc file is not present (meaning that either xkeyboard-config
> is not installed or it is an "older" version w/o that file).

Ah, right. There isn't an existing version of xkeyboard-config.pc that
doesn't have xkb_base variable. Probably fine.

>> The only drawback is that there's never been a hard requirement on
>> having xkeyboard-config installed before xserver, and we risk picking
>> up the host's installation instead of the one the user expects. Still
>> the CHECKING/RESULT is nice and informs people.
>
> This does not add a *hard* dep on xkeyboard-config at configure time; if
> xkeyboard-config.pc is not present, you end up with $datadir/X11/xkb just as
> before.  Adding a dep within the scope of jhbuild would fix that case but
> should not be necessary for other scenarios.

Sooner or later, though, we'll have system copies of
xkeyboard-config.pc, which means you'll get /usr/share/X11/xkb when
you might not have expected it. I think that's OK, but we need to fix
the build tools and make sure people know about it.

>> Same argument as above where we're likely to pick up the host's
>> xkbcomp since there was no hard requirement before. Hopefully they'd
>> see the result in the output.
>
> Which is why I didn't want to do this if cross-compiling.  Besides that and
> jhbuild (easily fixable), how else might this break things?

Cases 2 and 3 above. Things can get broken trying to mix your system
xkbcomp (with builtin path /usr/share/X11/xkb) and local xkb files. I
think we just have to make sure people understand this change. I just
wanted to note the drawback of it.

>> We should remove this stupid macro and just #define the path to
>> xkbcomp until some glorious future where it doesn't need to be forked
>> from the server. That's a separate patch, but --with-xkbcomp would be
>> better.
>
> That would require a more extensive patch affecting at least three .c files
> in xkb/.

Yep, some other day unless Daniel lands xkb2 and nukes xkbcomp from orbit.

--
Dan