[dpdk-dev] [PATCH v2 0/3] Allow overriding of build-time constants

[Thomas Monjalon]

20/10/2020 10:34, Bruce Richardson:
> On Mon, Oct 19, 2020 at 11:04:54PM +0200, Thomas Monjalon wrote:
> > 19/10/2020 12:21, Bruce Richardson:
> > > On Fri, Oct 16, 2020 at 06:46:12PM +0200, David Marchand wrote:
> > > > On Fri, Oct 16, 2020 at 5:56 PM Bruce Richardson
> > > > <bruce.richardson at intel.com> wrote:
> > > > > > librte_eal.so is indeed built with the 64 value:
> > > > > > $ pahole -C rte_mem_config build/install/lib64/librte_eal.so |grep memsegs
> > > > > > die__process_function: tag not supported (INVALID)!
> > > > > >     struct rte_memseg_list     memsegs[64];          /*   136  8704 */
> > > > > >
> > > > > >
> > > > > > But no trace of the custom value for external applications:
> > > > > > $ grep -r RTE_MAX_MEMSEG_LISTS build/install
> > > > > > build/install/include/rte_config.h:#ifndef RTE_MAX_MEMSEG_LISTS
> > > > > > build/install/include/rte_config.h:#define RTE_MAX_MEMSEG_LISTS 128
> > > > > > Binary file build/install/lib64/librte_eal.a matches
> > > > > > Binary file build/install/lib64/librte_eal.so.21.0 matches
> > > > > >
> > > > > > I can see the same using the meson option -Dc_args.
> > > > >
> > > > > Good point, I had not thought of external apps using these values.
> > > > >
> > > > > They are mostly for internal use, so maybe its worthwhile looking to not
> > > > > have them in a public header file. What do you think? Is it likely that
> > > > > apps would be using some of these values, or needs to know the specifics?
> > > > 
> > > > Some are publicly exposed, like RTE_MEMPOOL_CACHE_MAX_SIZE,
> > > > RTE_PKTMBUF_HEADROOM, RTE_ETHDEV_RXTX_CALLBACKS,
> > > > For those, either we propagate the overriden value to the installed
> > > > rte_config.h or we refuse customisation.
> > > > 
> > > I'd suggest the first 2 of those should possibly be global meson options.
> > 
> > How the application is reading the meson options?
> > 
> The meson options are reflected in the rte_build_config.h file. It's not
> automatic, but they should be set there by the config/meson.build file. If
> some are missed, they can be added, but I disagree that all meson options
> should always be passed through to apps. It makes them part of the API,
> perhaps unnecessarily, and therefore harder to change or adjust.
> Furthermore why should an app ever need to care if a DPDK build included
> the docs, or the kmods?
> 
> > > Third should probably not be exposed at all.
> > 
> > I think everything should be exposed.
> > The application may need to know whether a feature is enabled or not,
> > and what is a specific tuning value.
> > 
> > I suspect it is the last blocker for meson adoption.  Now that we removed
> > the makefiles, we need to fill this gap urgently.
> > 
> I really don't see this as a gap. With "make" we struggled with massive
> amounts of build config, and we tried to remove as much as we can. While
> this is reporting what's there rather than tweaking it, surely many of
> these settings are just better as #defines in the individual header files -
> if they need to be exposed at all.

I agree with the goal of moving these #defines internally.
I just feel having wrong values in rte_config.h looks to be a bug.