Re: [dpdk-dev] [dpdk-stable] [PATCH] app: fix plugin load on static builds

[Thomas Monjalon <thomas@monjalon.net>]

26/11/2020 17:37, Bruce Richardson:
> On Thu, Nov 26, 2020 at 05:32:26PM +0100, Olivier Matz wrote:
> > Hi Bruce,
> > 
> > Thanks for the feedback.
> > 
> > On Thu, Nov 26, 2020 at 03:28:46PM +0000, Bruce Richardson wrote:
> > > On Thu, Nov 26, 2020 at 03:20:42PM +0100, Olivier Matz wrote:
> > > > When dpdk is compiled as static libraries, it is not possible
> > > > to load a plugin from an application. We get the following error:
> > > > 
> > > >   EAL: librte_pmd_xxxx.so: undefined symbol: per_lcore__rte_errno
> > > > 
> > > > This happens because the dpdk symbols are not exported. Add them to the
> > > > dynamic symbol table by using '-Wl,--export-dynamic'. This option was
> > > > previously present when compiled with Makefiles, it was introduced in
> > > > commit f9a08f650211 ("eal: add support for shared object drivers")
> > > > 
> > > > Fixes: 16ade738fd0d ("app/testpmd: build with meson")
> > > > Fixes: 89f0711f9ddf ("examples: build some samples with meson")
> > > > Cc: stable@dpdk.org
> > > > 
> > > > Signed-off-by: Olivier Matz <olivier.matz@6wind.com>
> > > > ---
> > > > 
> > > > Hi,
> > > > 
> > > > Maybe the patch can be improved: I suppose that --export-dynamic
> > > > should only be added in case we are linking in static mode. Any
> > > > help about how to do that is welcome.
> > > 
> > > get_option('default_library') == 'static'
> > > 
> > > However, if the flag is harmless in shared linking mode, I don't think we
> > > need bother with this.
> > 
> > ok
> > 
> > > > 
> > > > There is probably a better place to define the default ldflags for
> > > > all applications (to factorize between app and example).
> > > > 
> > > > Also, should this flag be advertised in pkg-config?
> > > > 
> > > Perhaps. However, I'm not sure how common it would be for people to static
> > > link their own apps with DPDK and then want to plug in extra drivers after
> > > the fact? Are there any possible negative impacts to making that change?
> > 
> > I don't know if it is common, but this is something we do :)
> > 
> > > If we weren't right before the release deadline I'd definitely suggest
> > > adding it to the pkg-config files. At this late stage in release, I'm more
> > > cautious.
> > 
> > Yes, it is too late for 20.11. Maybe even for this patch without
> > updating pkg-config.
> > 
> 
> Well, since this is something clearly broken that is of use to you, I think
> we should strive to get some fix for it in. Based on the lateness of the
> hour, I think your patch is pretty close to the least-risky option we could
> take here. Therefore, despite my previous comment about not needing to
> limit the additional linking flag to static-builds only, I would suggest -
> out of an abundance of caution - that we do so, and keep the rest of your
> patch as-is for 20.11. Thereafter we can look at other possible approaches
> and simplifications in 21.02 and backport them.
> 
> So, for your patch with "if get_option('default_library') ..." checks added
> appropriately:
> 
> Acked-by: Bruce Richardson <bruce.richardson@intel.com>

The most critical issue is for externally built applications
statically linked using external plugins.
This patch fixes only internal apps built with meson.

I am for keeping the issue as-is for 20.11.0,
because it is really too late to change a compilation option.
We have already been bitten by compilers not supporting an option
or being buggy in specific contexts. I prefer not taking this risk
for a fix which helps nobody really.