Re: [PATCH] libtool-native_2.4.2.bb: Always use /bin/sed for SED

[Jason Wessel <jason.wessel@windriver.com>]

On 02/12/2013 03:39 PM, Richard Purdie wrote:
> On Tue, 2013-02-12 at 13:36 -0600, Jason Wessel wrote:
>> If you never use sstate and always build everything from scratch you
>> will never see this problem.  However, if you use sstate and build
>> directories that last a long time eventually you can end up with the
>> scenario where libtool gets a hard coded path in it for sed, and sed
>> may not exist.  The reason you don't see this problem to often if you
>> generally build from scratch is that libtool builds before sed and
>> will pickup the host's /bin/sed.
>>
>> The way to reproduce the issue is:
>>
>> bitbake some_image
>> bitbake -c cleansstate libtool-native
>> bitbake sed-native
>> bitbake libtool-native
>> bitbake -c clean sed-native
>> bitbake ANY_PACKAGE_THAT_USES_LIBTOOL_NATIVE
>>
>> In my case I used modphp, which doesn't exist in the oe-core. You will
>> end up with a strange looking error like:
>>
>> | make[1]: *** [buckets/apr_buckets_alloc.lo] Error 1
>> | /opt/build/bitbake_build/tmp/sysroots/x86_64-linux/usr/bin/x86_64-linux-libtool: line 981: /opt/build/bitbake_build/tmp/sysroots/x86_64-linux//bin/sed: No such file or directory
>>
>> The solution is to always use /bin/sed for libtool-native.
>>
>> Signed-off-by: Jason Wessel <jason.wessel@windriver.com>
>> ---
>>  .../libtool/libtool-native_2.4.2.bb                |    3 ++-
>>  1 files changed, 2 insertions(+), 1 deletions(-)
>>
>> diff --git a/meta/recipes-devtools/libtool/libtool-native_2.4.2.bb b/meta/recipes-devtools/libtool/libtool-native_2.4.2.bb
>> index f12e6a1..18188ef 100644
>> --- a/meta/recipes-devtools/libtool/libtool-native_2.4.2.bb
>> +++ b/meta/recipes-devtools/libtool/libtool-native_2.4.2.bb
>> @@ -2,12 +2,13 @@ require libtool-${PV}.inc
>>  
>>  DEPENDS = ""
>>  
>> -PR = "${INC_PR}.0"
>> +PR = "${INC_PR}.1"
>>  SRC_URI += "file://prefix.patch"
>>  
>>  inherit native
>>  
>>  EXTRA_OECONF = " --with-libtool-sysroot=${STAGING_DIR_NATIVE}"
>> +CACHED_CONFIGUREVARS += "ac_cv_path_SED=/bin/sed"
> Do we have to set a path for it? Can't we rely on PATH being sane? I'm
> wondering if we should actually set this in the core site files.
> Hardcoding a path to utilities never usually ends well and this is just
> the tip of an iceberg.
>
> If we have to use a path, "${bindir}/env sed"?
>

The libtool seems to be in a class of its own with respect to internally hard coding things, so I am inclined to say this is a one off because A) it is libtool and B) it is part of the boot strap. For any other packages I have not observed any kind of problem with the sysroot sed vs host provided sed.

Unfortunately doing ${bindir}/env sed can lead to a fairly rare race where sed can be there an get removed later because it will prefer the sed in the sysroot area because it is in the path first.  I never hit any of these problems until recently while continuing to just randomly build things with the usual stream of updates from the git repository.

If we want libtool-native to use something other than /bin/sed on the host, the bootstrap needs some kind of overhaul to make libtool depend correctly on sed.  Currently we end up with a circular dependency if you try to make libtool-native depend on sed-native.

Jason.