Several sandbox failures happen during configuration stage when cuda USE flag is enabled. -- Found LIBOMPTARGET_DEP_LIBFFI: /usr/lib64/libffi.so * ACCESS DENIED: open_wr: /dev/nvidiactl * ACCESS DENIED: open_wr: /dev/nvidiactl * ACCESS DENIED: open_wr: /dev/nvidiactl * ACCESS DENIED: open_wr: /dev/nvidiactl * ACCESS DENIED: open_wr: /dev/nvidiactl * ACCESS DENIED: open_wr: /dev/nvidiactl * ACCESS DENIED: open_wr: /dev/nvidiactl * ACCESS DENIED: open_wr: /dev/nvidiactl * ACCESS DENIED: open_wr: /dev/nvidiactl * ACCESS DENIED: open_wr: /dev/nvidiactl * ACCESS DENIED: open_wr: /dev/nvidiactl -- Automatic GPU detection failed. Building for common architectures. Reproducible: Always emerge --info attached due to comment size limit.
Created attachment 701319 [details] emerge-info-libomp-12.0.0.txt output of emerge --info libomp
Created attachment 701322 [details] libomp-12.0.0:20210420-123321.log build log
Created attachment 701343 [details] non-failing build.log i cannot trigger this issue. in my case i get "-- check-openmp does nothing." and the compilation goes on.
(In reply to Miroslav Šulc from comment #3) > i cannot trigger this issue. in my case i get "-- check-openmp does > nothing." and the compilation goes on. Need to test your configuration... What happens if you enable USE=offload as well? I have this enabled by default and the metadata.xml says for it's cuda entry, cuda needs offload to be enabled. Haven't verified this, but if it's true, there should probably a REQUIRED_USE dependency added for it as well.
Created attachment 701346 [details, diff] ebuild patch Could you try applying this change to the ebuild?
Comment on attachment 701346 [details, diff] ebuild patch addpretend /dev/nvidiactl Typo, should be "addpredict". I don't currently have cuda stuff setup to test myself though.
(In reply to Ionen Wolkens from comment #6) > I don't currently have cuda stuff setup to test myself though. Scratch that, I do now and could reproduce with same USE. "addpredict /dev/nvidiactl" works as expected for me.
(In reply to Michał Górny from comment #5) > Created attachment 701346 [details, diff] [details, diff] > ebuild patch > > Could you try applying this change to the ebuild? Adding the addpredict statement works for me too. (In reply to Bernd from comment #4) > default and the metadata.xml says for it's cuda entry, cuda needs offload to > be enabled. Haven't verified this, but if it's true, there should probably a > REQUIRED_USE dependency added for it as well. What about this mention in metadata.xml? Should we add a REQUIRED_USE=cuda? ( offload ) as well?
(In reply to Miroslav Šulc from comment #3) > Created attachment 701343 [details] > non-failing build.log > > i cannot trigger this issue. in my case i get "-- check-openmp does > nothing." and the compilation goes on. I can confirm, your configuration of USE="-offload cuda" does not trigger this. Probably due to USE=cuda does nothing without USE=offload, as noted in metadata.xml?
Created attachment 701361 [details, diff] libomp-12.0.0.patch I tried this patch, and it worked as expected on my machine.
The bug has been closed via the following commit(s): https://gitweb.gentoo.org/repo/gentoo.git/commit/?id=f2dd91d7e6fae86d831e9231ddda11d03a366a02 commit f2dd91d7e6fae86d831e9231ddda11d03a366a02 Author: Ionen Wolkens <sudinave@gmail.com> AuthorDate: 2021-04-21 06:47:35 +0000 Commit: Michał Górny <mgorny@gentoo.org> CommitDate: 2021-04-21 07:00:29 +0000 sys-libs/libomp: fix /dev/nvidiactl sandbox issues with cuda+offload Closes: https://bugs.gentoo.org/784539 Signed-off-by: Ionen Wolkens <sudinave@gmail.com> Closes: https://github.com/gentoo/gentoo/pull/20483 Signed-off-by: Michał Górny <mgorny@gentoo.org> sys-libs/libomp/libomp-12.0.0.9999.ebuild | 3 ++- sys-libs/libomp/libomp-12.0.0.ebuild | 1 + sys-libs/libomp/libomp-13.0.0.9999.ebuild | 1 + 3 files changed, 4 insertions(+), 1 deletion(-)
