[OE-core] [PATCH] bzip2-native: fix problems when bzip2-native is installed in parallel

[Richard Purdie]

On Tue, 2012-07-24 at 13:39 -0500, Mark Hatle wrote:
> On 7/24/12 8:57 AM, Burton, Ross wrote:
> > On 24 July 2012 14:49, Yao Zhao <yao.zhao at windriver.com> wrote:
> >> when bzip2-native is installed in parallel to sysroot, it is possible that
> >> some packages are using bzip2 to unpack, there are chances that bzip2 is
> >> installed to sysroot but libbz2.so.0 not installed yet because parallel
> >> installation.
> >> link bzip2 and bzip2recover statically to avoid this problem and don't lose
> >> parallel installation. libbz2.so is still available.
> >
> > Is it me, or is this officially getting silly?  This probably happens
> > for *every* binary in the sysroot that links to a library, which is
> > probably a fair proportion of them.  Statically linking every single
> > one and then special-casing further problems where a static link isn't
> > sufficient (see pythonnative) just isn't going to scale.
> 
> The problem is that there are a handful of things that are needed, and for some 
> reason (valid or not) packages are not "requiring", either because they assume 
> the system is valid or they simply don't know they have the requirement.
> 
> Both python and perl may be used by a number of auto* utilities as well as by 
> packages themselves.  We could attempt to add DEPENDS for all of the cases, but 
> is it really worth it?  I suspect we'll end up specify the DEPENDS for 90% of 
> the things in the system, simply due to them incidentally running some system 
> level python or perl script.  If we make the python/perl into external items 
> that only get loaded into the environment when we're building python/perl 
> modules then things become increasingly easy.
> 
> As for the bzip2 issue.. This is another of those "we assume it's valid 
> binaries"..  And the problem appears to be when it gets build, it's not valid 
> for a small window of time.
> 
>  From what Yao has explained to me (offline) the issue is that the 
> assume_provided for bzip2-native works just fine, we use the system 
> version...but then python (or other) utilities come in and say they need 
> "bzip2-full-native".  Which triggers the build.. and opens a small window of 
> time where bzip2 is being installed, where it isn't functional -- something else 
> comes in and has a requirement of bzip2-native, which is satisfied by the 
> assume_provided and we get into the race where it executes a broken binary.
> 
> The underlying issue is simply that if we're installing tooling that is either 
> assume-provided (based on an alternative 'full' version) or incidental usage 
> like python or perl we need to take precautions to ensure that the build system 
> never sees the "broken" version during the short install window triggering the 
> race condition.
> 
> I'll let Yao comment further on this particular issue, but there is an overall 
> class of issues with -native that we need to track to get rid of these subtle races.

The set of problems are cases where we have things in ASSUME_PROVIDED.
It is near impossible to have an empty set there and sometimes we need
to build things ourselves which are already listed there. python-native
and perl-native are the pathological cases but as also have issues with
bzip2, tar and likely undiscovered yet, now git.

I'd been searching for bzip2-replacement-native rather than full which
is why I didn't see this earlier. I suspect in the python/bzip2 case,
not installing the bzip2 binary might well solve many of the problems as
python really wants a native libbz2 and the host bzip2 binary is ok.

The races themselves are over binaries in PATH. As far as I
know/understand things we're close to having them resolved. What I am
considering as an improvement is instead of the pythonnative/perlnative
bbclass files, having a variable like "NATIVEPATHADDITIONS" which would
take values like "python-native", "perl-native" and would simply add
${STAGING_BINDIR}/xxx to PATH for each item listed. The bbclass files
are useful for the dependency additions too though which sometimes are
not trivial. This would at least remove the case by case bbclass
addition issue we currently have for the simple cases though.

Cheers,

Richard