[OE-core] Autoconf problems in OE build of LTP realtime tests

[Peter A. Bigot]

On 09/05/2014 03:36 PM, Gary Robertson wrote:
> Dear OE developers,
>
> Linaro Networking Group is building root filesystem images containing  
> the LTP test suite, using the current recipe at 
> openembedded-core/meta/recipes-extended/ltp/ltp_20140422.bb 
> <http://ltp_20140422.bb>. This build uses autoconf to probe mutex 
> feature support for protocol specification, priority inheritance, and 
> robust mutexes.  The macros used for this can be found in 
> ltp/testcases/realtime/m4/check.m4 from the LTP git repository.
>
> We are cross-building for the armv7a architecture using the machine 
> definition found in 
> meta-linaro/meta-linaro/conf/machine/genericarmv7a.conf. The 
> development host platform is an Ubuntu 12.04 x86_64 platform with all 
> software kept regularly updated from the Ubuntu repositories.
>
> The macros in check.m4 use the AC_TRY_COMPILE function macro to 
> attempt compilation of POSIX function calls utilizing the features 
> being probed - such that if the compilation fails the features are 
> deemed to be unsupported... and if compilation succeeds the features 
> are deemed supported.
>
> Native compilation of the same code used in the probes, and execution 
> of the resulting program - shows support for all the probed mutex 
> features on both the development host as well as the armv7a target 
> platform... that is when this probe code is built on either the host 
> or target machine and then executed in that environment... so support 
> is there in both environments.
>
> Despite the presence of this support the AC_TRY_COMPILE macro is 
> indicating a lack of support for the probed features during the 
> cross-build process.
>
> I am not personally familiar with the detailed function of the 
> autoconf macros, and I can find no documentation about how OE 
> interacts with those macros in order to coerce them into configuring 
> for the target machine in a cross-build environment... so I am hoping 
> the autoconf experts at OE can shed some light on this and at least 
> provide some clues as to where to look for the problem and possible 
> resolution.  Any insight you can provide in this matter would be 
> appreciated.

Any mention of "armv7a" and "mutex" points me to 
http://cgit.openembedded.org/openembedded-core/commit/?id=0ba6ab39f187ecd4261f08e768f365f461384a3a 
which fixed a problem where the compiler and runtime disagreed about the 
capabilities of the target processor unless special flags were passed to 
the compiler.  Presence of armv7a enables mutex-related instructions 
which affect what support is provided in the C runtime system; absence 
of armv7 at compile-time will result in attempts to locate alternative 
implementations that are not present.

It's possible that a similar issue is hitting you, depending on what 
options are present in the test compile and whether the libraries linked 
to were built with compatible options.

I tried setting up a meta-linaro + meta-linaro-toolchain build to see if 
I could reproduce it, but meta-linaro appears to not be up to date with 
the removal of eglibc from oe-core.

The autoconf test should have left information in the config.log file 
showing how it tried to build the application.  Having that information 
would help with diagnosis.

Hope this is some help.

Peter
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openembedded.org/pipermail/openembedded-core/attachments/20140905/3c452d08/attachment-0002.html>