On Thu, Aug 01, 2019 at 11:40:33AM -0400, Aaron Conole wrote: > Bruce Richardson <bruce.richard...@intel.com> writes: > > > On Wed, Jul 31, 2019 at 12:10:55PM -0400, Aaron Conole wrote: > >> Aaron Conole <acon...@redhat.com> writes: > >> > >> > Bruce Richardson <bruce.richard...@intel.com> writes: > >> > > >> >> On Wed, Jul 31, 2019 at 10:50:29AM -0400, Aaron Conole wrote: > >> >>> From: Michael Santana <msant...@redhat.com> > >> >>> > >> >>> Currently many unit tests fail when running tests under shared builds. > >> >>> This happens because of missing driver dependencies. This is fixed by > >> >>> explicitly linking in missing drivers for the test application. > >> >>> > >> >>> before and after (clang): > >> >>> https://travis-ci.com/Maickii/dpdk-2/jobs/212329160#L623 > >> >>> https://travis-ci.com/Maickii/dpdk-2/jobs/212335912#L620 > >> >>> > >> >>> Suggested-by: Bruce Richardson <bruce.richard...@intel.com> > >> >>> Suggested-by: David Marchand <david.march...@redhat.com> > >> >>> Signed-off-by: Michael Santana <msant...@redhat.com> > >> >>> Signed-off-by: Aaron Conole <acon...@redhat.com> > >> >>> --- > >> >> Rather than linking in the libraries explicitly, can you have the build > >> >> do > >> >> a "ninja install" at the end to place the libraries and drivers in their > >> >> correct paths. That should mean that the test app (via eal) auto-loads > >> >> all > >> >> drivers from EAL_PMD_PATH (/usr/local/...). It would save having to make > >> >> further changes to this file to link in any additional drivers. > >> > >> Oops, forgot to include a link to a build where I did this. It's > >> failing (you can see the corresponding commit at > >> https://github.com/orgcandman/dpdk/commit/ccba975bdfe851b4c8ec3f208451bb105317b76d): > >> > >> https://travis-ci.org/orgcandman/dpdk/jobs/566044409 > >> > > I think the error may be due to having some drivers already linked in while > > trying to reload all drivers dynamically. The dynamic loading of drivers I > > actually think we could make more robust. For example: > > > > * only load .so files. If you pass in the drivers path from the build, it > > errors out trying to load .a files > > * when loading directories, maybe skip any files which don't have a "librte" > > prefix - anyone who has their own drivers they want loaded can add the > > prefix or load it explicitly via -d > > * alternatively, skip files which don't have a pmdinfo section in them > > * allow skipping of drivers which are already linked in > > * don't fail the whole process if one driver fails to load from a > > directory. > > What change do you suggest I make to the steps I outlined? I think it > seems to be more complex than a simple 'ninja install'. > > Maybe it makes sense to merge these patches now and fix the library > loading code as a separate action? Or is there a different procedure I > (and other developers) should follow? Does it really need a 'ninja > install' even for a drive-by developer to make a small patch and test > before submitting? > I think your patch is the best solution for 19.08. Fixing the library loading for developers is a larger job beyond that.
Another item for future consideration: since meson/ninja automatically sets the rpath for binaries in build directory - and then strips that out on install - it would be really nice if we could leverage that somehow for finding drivers. /Bruce
