Bugzilla Bug 221931

Passing --enable-runtime-module-search to configure will cause a segfault in
the Perl bindings on shutdown of i.e. git-svn.

See http://subversion.tigris.org/servlets/ReadMsg?list=dev&msgNo=138509 for
progress reports.

Reproducible: Always

Steps to Reproduce:

Please write a patch or wait until somebody else writes it.

--enable-runtime-module-search has many significant advantages and not using it
is unacceptable.

Created an attachment (id=153071) [edit]
Ebuild patch to disable runtime-module-search and die on configure errors 

I cannot write a patch, since even the exact cause is not yet know. ;)
Currently subversion devs suspect a generic showstopping bug in pool
destruction.
Since that functionality was already present in 1.4 and was not used by Gentoo,
I doubt it is now so extremely necessary to have it in a ~arch masked release
candidate. ;)

(In reply to comment #2)
> Created an attachment (id=153071) [edit]
> Ebuild patch to disable runtime-module-search

-1

> and die on configure errors

econf always dies on configure errors. This part is totally useless.

> Currently subversion devs suspect a generic showstopping bug in pool
> destruction.

I heard about it. I'm one of them:
https://svn.collab.net/viewvc/*checkout*/svn/trunk/COMMITTERS

> Since that functionality was already present in 1.4 and was not used by
> Gentoo, I doubt it is now so extremely necessary to have it

It avoids some problems which happen during incompatible upgrades of some
(sometimes indirect) dependencies of Subversion.

How about package.mask-ing the 1.5.0_rc* ebuilds? Why are they in ~arch anyway?

(In reply to comment #4)
> How about package.mask-ing the 1.5.0_rc* ebuilds?

Wrong idea.

> Why are they in ~arch anyway?

Because they are needed and useful here. Older versions don't work with
libtool-2.*.

The issue is *still* present in rc9.
Can you please either fix the segfaults immediately, or disable runtime module
search?
This is getting annoying...

Created an attachment (id=156927) [edit]
Workaround

(In reply to comment #7)

dev-util/git[subversion] should depend on dev-util/subversion[-dso].

For that matter it should be "anything-that-uses-subversion-bindings" should
depend on "subversion -dso"...
But the general idea seems like an acceptable compromise.

fixed in 1.5.0_rc9-r1

(In reply to comment #9)
> For that matter it should be "anything-that-uses-subversion-bindings" should
> depend on "subversion -dso"...

Maybe only Subversion Perl bindings. We didn't hear about problems with Python
or Ruby bindings.

I still see this issue in 1.5.1. (With USE=dso.) Is there anything known about
a real fix (besides disabling runtime-module-search)?

(In reply to comment #10)
> fixed in 1.5.0_rc9-r1
> 

Where is the fix? If you meant the introduction of the dso useflag, that seems
like a step in the right direction, but could you at least echo a big warning
if this is enabled? Git svn failed on me and I had now idea where this came
from...

Duplicated in bug #223747 ?

According to http://subversion.tigris.org/issues/show_bug.cgi?id=3289 the
configure flag in question is "unsupported, experimental and known to break
things"; this seems like a very bad idea to have as a default...