Re: Sporadic build failure of next

[Baurzhan Ismagulov <ibr@radix50.net>]

On Tue, Aug 20, 2019 at 12:40:16PM +0200, Jan Kiszka wrote:
> > Apparently, we instruct multiconfig to run the same task twice for the same
> > DISTRO and DISTRO_ARCH in the same directory, and they happen to run in
> > parallel. If this is the case, we should ideally build once; not sure whether
> > this is desirable for all packages, though (e.g., two boards building the
> > kernel from the same source and revision, but with different configs, etc.).
> 
> That usually happens when we feed in a variable into the build that makes
> the package different - without adjusting its name.

Ok, so this use case is solved. Although Debian does change the binary package
name, which I find better, since those are different binary packages.


> > Could we force building of the packages at the same time in CI? Ideas welcome.
> > In this way, we could perform regression testing for concurrency.
> 
> I suspect we can when we first build multiple multiconf targets up to the
> level needed to build that package and then request to build it for all
> those multiconf targets.

That sounds promising. Maybe it could require manual tailoring if there were
multiple candidates for building after the first stage, but at least we could
test it.


> > IIRC, the kernel or PREEMPT_RT has a static checker that can detect potential
> > deadlocks. That would be a nice addition to bitbake (although I doubt it is
> > implementable today, since it would require recipe introspection w.r.t. e.g.
> > build directory, etc.).
> 
> Maybe evaluating the stamps after a build helps. There should be different
> pattern of a recipe was built multiple times for the same distro-arch.

Ok, that might catch it at least when the critical section is violated
(depending on the point in time where the stamps are created).


> BTW, I'm about to adopt the OE layout for the stamps folder. Let's see what
> that brings (beside visual alignment).

Brings and hopefully not takes :) . We need that anyway to prevent *-* from
matching pkg-1.0 and pkg-abc-1.0 discussed recently. FWIW, Debian solved that
by using pkg_1.0 and pkg-abc_1.0.


With kind regards,
Baurzhan.