You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@httpd.apache.org by Rainer Jung <ra...@kippdata.de> on 2011/07/04 21:07:49 UTC

RUNPATH for module dependencies on Unix/Linux

Hi everyone,

a couple of modules have additional external dependencies:

mod_deflate: zlib
mod_lua: lua
mod_serf: serf
mod_socache_dc: distcache

At the moment, the compiled modules do not contain any RPATH/RUNPATH
info except when the libraries themselves are installed as libtool
libraries, i.e. have accompanying *.la files.

The original distcache, zlib and lua distributions by default do not
install .la libtool files. It is of course possible, that distros
provide such files.

I'd like to get the RPATH/RUNPATH included in the modules so that one
doesn't have to juggle with LD_LIBRARY_PATH or LoadFile when loading
those modules.

If the libtool .la files are installed for the dependency libs, the
RPATH/RUNPATH automatically gets added by libtool during module linking.
If it is not, we could add "-R ..." when calling libtool.

On Solaris we already add the "-R ..." unconditional via the
"ap_platform_runtime_link_flag" variable. Thsi goes back to at least
2001 and hasn't been used on any other platform since then. On other
Unix/Linux platforms we seem to assume that .la files are present or
users add LD_LIBRARY_PATH or LoadFile.

Since I don't wont to get into the business of writing even longer
dependency checking macros to find out, whether a .la file for a lib
exists or not, I suggest to always add the -R ... flags to the module
LDADD for the modules mentioned above. As a consequence we would also
remove the then no longer needed "ap_platform_runtime_link_flag".

Last remark: of course the "-R ..." will only be added if the libs are
installed in non-default locations, i.e. if --with-LIB=DIR is given.

Any objections?

Regards,

Rainer



Re: RUNPATH for module dependencies on Unix/Linux

Posted by Daniel Ruggeri <DR...@primary.net>.
On 7/5/2011 2:46 AM, Joe Orton wrote:
> IMO it is much better to leave this stuff in the control of people who 
> build the software, who can already set env vars or LDFLAGS as they 
> require.  (It would also be much better if everybody used libtool, since 
> as you say, the .la files make this problem goes away :)
Yup - also consider that a lot of folks build the software on different
machines (and potentially different environments/layouts) than the ones
the software runs on.

-- 
--
Daniel Ruggeri

Re: RUNPATH for module dependencies on Unix/Linux

Posted by Joe Orton <jo...@redhat.com>.
On Tue, Jul 05, 2011 at 11:25:53AM +0200, Rainer Jung wrote:
> > I will think about a good way, how users can pass additional LDADD
> > flags. Putting the "-R ..." into the LDFLAGS seems to be too heavy,
> > because then the RPATH of every module etc. will contain the given
> > directory. The real problematic cases are only mod_lua and in rare
> > circumstances (you want to link against a non-standard zlib)
> > mod_deflate. It might be enough to use APR_ADDTO() instead of direct
> > assignment or APR_SETVAR() when setting the LDADD variables. I'll
> > experiment.
> 
> Using APR_ADDTO() for the MOD_XXX_LDADD worked nicely, so I used that
> solution to allow users to easily prefix something to our LDADD detection.

Good idea, that looks useful.

> > So would you also remove the current "ap_platform_runtime_link_flag" 
> > hack, which we only use for Solaris and which adds "-R ..." 
> > unconditional? Or should we stick, because although it is not right 
> > it is what did work in the past (for Solaris)?
 
I'm neutral on this.  It is odd to do it specifically for Solaris, but 
maybe Solaris is "different" somehow.

Regards, Joe

Re: RUNPATH for module dependencies on Unix/Linux

Posted by Rainer Jung <ra...@kippdata.de>.
> I will think about a good way, how users can pass additional LDADD
> flags. Putting the "-R ..." into the LDFLAGS seems to be too heavy,
> because then the RPATH of every module etc. will contain the given
> directory. The real problematic cases are only mod_lua and in rare
> circumstances (you want to link against a non-standard zlib)
> mod_deflate. It might be enough to use APR_ADDTO() instead of direct
> assignment or APR_SETVAR() when setting the LDADD variables. I'll
> experiment.

Using APR_ADDTO() for the MOD_XXX_LDADD worked nicely, so I used that
solution to allow users to easily prefix something to our LDADD detection.

> So would you also remove the current "ap_platform_runtime_link_flag"
> hack, which we only use for Solaris and which adds "-R ..."
> unconditional? Or should we stick, because although it is not right it
> is what did work in the past (for Solaris)?

Thsi question still remains open for me.

Regards,

Rainer



Re: RUNPATH for module dependencies on Unix/Linux

Posted by Rainer Jung <ra...@kippdata.de>.
Hi Joe,

On 05.07.2011 09:46, Joe Orton wrote:
> On Mon, Jul 04, 2011 at 09:07:49PM +0200, Rainer Jung wrote:
>> Hi everyone,
>>
>> a couple of modules have additional external dependencies:
>>
>> mod_deflate: zlib
>> mod_lua: lua
>> mod_serf: serf
>> mod_socache_dc: distcache
>>
>> At the moment, the compiled modules do not contain any RPATH/RUNPATH
>> info except when the libraries themselves are installed as libtool
>> libraries, i.e. have accompanying *.la files.
>>
>> The original distcache, zlib and lua distributions by default do not
>> install .la libtool files. It is of course possible, that distros
>> provide such files.
> 
> It looks like distcache (1.4.5 at least) does install the .la files?

Yes, sorry, distcache and serf do, zlib and lua don't.

> ...
>> Last remark: of course the "-R ..." will only be added if the libs are 
>> installed in non-default locations, i.e. if --with-LIB=DIR is given.
> 
> That sounds simple but it rarely is, in my experience.  You need a bunch 
> of heuristics and as often as not, these will do the wrong thing in some 
> cases as well as the right thing in others; so the net gain is hard to 
> judge.  Notably, adding RPATHs by default can produce binaries which are 
> not relocatable within the filesystem, which can be surprising 
> behaviour.

OK

> 1) How can you tell whether you are linking against a static library or 
> a shared library?  RPATHs are only needed if using a shared library

Haven't thought about that case.

> 2) How can you tell whether a given path is a "non-default location"?  
> To be strictly correct, you need to add an RPATH iff it is not already 
> configured in the dynamic linker search path - which you'd need to parse 
> system config files to determine.  This gets messy very quickly.

I would have redefined "non-default location" by "explicit
--with-LIB=DIR provided for configure". That is not "strictly correct"
(far from it).

> IMO it is much better to leave this stuff in the control of people who 
> build the software, who can already set env vars or LDFLAGS as they 
> require.  (It would also be much better if everybody used libtool, since 
> as you say, the .la files make this problem goes away :)

So would you also remove the current "ap_platform_runtime_link_flag"
hack, which we only use for Solaris and which adds "-R ..."
unconditional? Or should we stick, because although it is not right it
is what did work in the past (for Solaris)?

I will think about a good way, how users can pass additional LDADD
flags. Putting the "-R ..." into the LDFLAGS seems to be too heavy,
because then the RPATH of every module etc. will contain the given
directory. The real problematic cases are only mod_lua and in rare
circumstances (you want to link against a non-standard zlib)
mod_deflate. It might be enough to use APR_ADDTO() instead of direct
assignment or APR_SETVAR() when setting the LDADD variables. I'll
experiment.

Regards,

Rainer

Re: RUNPATH for module dependencies on Unix/Linux

Posted by Joe Orton <jo...@redhat.com>.
On Mon, Jul 04, 2011 at 09:07:49PM +0200, Rainer Jung wrote:
> Hi everyone,
> 
> a couple of modules have additional external dependencies:
> 
> mod_deflate: zlib
> mod_lua: lua
> mod_serf: serf
> mod_socache_dc: distcache
> 
> At the moment, the compiled modules do not contain any RPATH/RUNPATH
> info except when the libraries themselves are installed as libtool
> libraries, i.e. have accompanying *.la files.
> 
> The original distcache, zlib and lua distributions by default do not
> install .la libtool files. It is of course possible, that distros
> provide such files.

It looks like distcache (1.4.5 at least) does install the .la files?

...
> Last remark: of course the "-R ..." will only be added if the libs are 
> installed in non-default locations, i.e. if --with-LIB=DIR is given.

That sounds simple but it rarely is, in my experience.  You need a bunch 
of heuristics and as often as not, these will do the wrong thing in some 
cases as well as the right thing in others; so the net gain is hard to 
judge.  Notably, adding RPATHs by default can produce binaries which are 
not relocatable within the filesystem, which can be surprising 
behaviour.

1) How can you tell whether you are linking against a static library or 
a shared library?  RPATHs are only needed if using a shared library

2) How can you tell whether a given path is a "non-default location"?  
To be strictly correct, you need to add an RPATH iff it is not already 
configured in the dynamic linker search path - which you'd need to parse 
system config files to determine.  This gets messy very quickly.

IMO it is much better to leave this stuff in the control of people who 
build the software, who can already set env vars or LDFLAGS as they 
require.  (It would also be much better if everybody used libtool, since 
as you say, the .la files make this problem goes away :)

Regards, Joe