Re: [dpdk-dev] [PATCH 1/1] build: allow disabling libs

[Bruce Richardson <bruce.richardson@intel.com>]

On Fri, Sep 18, 2020 at 10:49:23AM +0200, Mohammed Hawari wrote:
> Similarly to the disable_drivers option, the disable_libs option is
> introduced. This allows to selectively disable the build of elements
> in libs to speed-up the build process.
> 
> Signed-off-by: Mohammed Hawari <mohammed@hawari.fr>
> ---

While I don't particularly like allowing libs to be enabled and disabled
since it complicates the build, I can see why it's necessary. This is an
area that does need some discussion, as I believe others have some opinions
in this area too.

However, for now, some additional thoughts, both on this patch and in
general:

1. I see you included disabling apps if their required libs are not
   available. What about the drivers though?
2. A bigger issue is whether this is really what we want to do, guarantee a
   passing build even if vast chunks of DPDK are actually enabled? I'd tend
   towards "no" in this case, and I'd rather see disabling of libs more
   constrained.
3. To this end, I think I'd rather see us maintain a set of libs which are
   allowed to be disabled, and prevent the rest from being so. For example,
   it makes no sense in DPDK to disable the EAL or mempool libs, since nothing
   will build, while the bitrate_stats or latency_stats libs could likely
   be disabled with little or no impact.

Therefore, I think a better implementation is to start as in this patch
with a new config parameter to disable libs, but as part of that patch add
in an internal list of the libs which are allowed to be disabled (initially
empty). Telling the build system to disable a lib not on that list should
raise a configuration time error. As for how a lib gets on the list - that
should be done once the build has been tested with that lib disabled, i.e.
once testpmd and other apps have got #defines in the code for stripping out
the disabled blocks, and any drivers which depend on the lib have proper
checks and warnings in place about it being disabled (or also #defines in
the code if that can be done).

The other advantage of maintaining such a list is that it then becomes
somewhat feasible to test these build settings, in that (maybe once per
release), iterate through the list of disable-able libs and test that the
build passed with each one disabled individually. [I think for this purpose
we can ignore interactions of having two disabled simultaneously, in order
to have something testable]

What do others in the community think?

Regards,
/Bruce