RE: kselftest: failed to build with -C tool/testing/selftests when KBUILD_OUTPUT is set

[<Tim.Bird@sony.com>]

> -----Original Message-----
> From: linux-kselftest-owner@vger.kernel.org [mailto:linux-kselftest-
> owner@vger.kernel.org] On Behalf Of shuah
> Sent: Wednesday, November 27, 2019 5:33 PM
> To: Cristian Marussi <cristian.marussi@arm.com>; Michael Ellerman
> <mpe@ellerman.id.au>; Bird, Tim <Tim.Bird@sony.com>; linux-
> kselftest@vger.kernel.org; shuah <shuah@kernel.org>
> Subject: Re: kselftest: failed to build with -C tool/testing/selftests when
> KBUILD_OUTPUT is set
> 
> On 11/27/19 9:53 AM, Cristian Marussi wrote:
> > On 27/11/2019 16:42, shuah wrote:
> >> On 11/27/19 9:23 AM, Cristian Marussi wrote:
> >>> On 27/11/2019 16:04, shuah wrote:
> >>>> On 11/27/19 3:54 AM, Cristian Marussi wrote:
> >>>>> Hi
> >>>>>
> >>>>> On 27/11/2019 03:54, Michael Ellerman wrote:
> >>>>>> Cristian Marussi <cristian.marussi@arm.com> writes:
> >>>>>>> Hi
> >>>>>>>
> >>>>>>> while testing on linux-next
> >>>>>>>
> >>>>>>> I see that, when KBUILD_OUTPUT is set in the env, running
> something like (using TARGETS=exec as a random subsystem here...)
> >>>>>>>
> >>>>>>> $ make TARGETS=exec INSTALL_PATH=/nfs/LTP-official-debian-
> aarch64-rootfs/opt/KSFT_next kselftest-install
> >>>>>>>
> >>>>>>> works fine as usual, WHILE the alternative invocation (still
> documented in Documentation/dev-tools/kselftest.rst)
> >>>>>>>
> >>>>>>> make -C tools/testing/selftests/ TARGETS=exec
> INSTALL_PATH=/nfs/LTP-official-debian-aarch64-rootfs/opt/KSFT_next
> install
> >>>>>>>
> >>>>>>> fails miserably with:
> >>>>>>> ...
> >>>>>>> ...
> >>>>>>>     REMOVE  usr/include/rdma/cxgb3-abi.h usr/include/rdma/nes-
> abi.h
> >>>>>>>      HDRINST usr/include/asm/kvm.h
> >>>>>>>      INSTALL /kselftest/usr/include
> >>>>>>> mkdir: cannot create directory ‘/kselftest’: Permission denied
> >>>>>>> /home/crimar01/ARM/dev/src/pdsw/linux/Makefile:1187: recipe
> for target 'headers_install' failed
> >>>>>>> make[2]: *** [headers_install] Error 1
> >>>>>>>
> >>>>>>>
> >>>>>>> This is fixed by unsetting KBUILD_OUTPUT OR reverting:
> >>>>>>>
> >>>>>>> 303e6218ecec (ksft/fixes) selftests: Fix O= and KBUILD_OUTPUT
> handling for relative paths
> >>>>>>>
> >>>>>>> since bypassing top makefile with -C, the definition of abs-objtree
> used by the above patch
> >>>>>>> is no more available.
> >>>>>>>
> >>>>>>> As a side effect when KBUILD_OUTPUT is set, this breaks also the
> usage kselftest_install.sh.
> >>>>>>>
> >>>>>>>     $ ./kselftest_install.sh /home/crimar01/ARM/dev/nfs/LTP-
> official-debian-aarch64-rootfs/opt/KSFT_full_next
> >>>>>>> ./kselftest_install.sh: Installing in specified location -
> /home/crimar01/ARM/dev/nfs/LTP-official-debian-aarch64-
> rootfs/opt/KSFT_full_next ...
> >>>>>>> make --no-builtin-rules INSTALL_HDR_PATH=$BUILD/usr \
> >>>>>>> 	ARCH=arm64 -C ../../.. headers_install
> >>>>>>> make[1]: Entering directory
> '/home/crimar01/ARM/dev/src/pdsw/linux'
> >>>>>>> make[2]: Entering directory
> '/home/crimar01/ARM/dev/src/pdsw/out_linux'
> >>>>>>>      INSTALL /kselftest/usr/include
> >>>>>>> mkdir: cannot create directory ‘/kselftest’: Permission denied
> >>>>>>> /home/crimar01/ARM/dev/src/pdsw/linux/Makefile:1187: recipe
> for target 'headers_install' failed
> >>>>>>> make[2]: *** [headers_install] Error 1
> >>>>>>> make[2]: Leaving directory
> '/home/crimar01/ARM/dev/src/pdsw/out_linux'
> >>>>>>> Makefile:179: recipe for target 'sub-make' failed
> >>>>>>> make[1]: *** [sub-make] Error 2
> >>>>>>> make[1]: Leaving directory
> '/home/crimar01/ARM/dev/src/pdsw/linux'
> >>>>>>> Makefile:142: recipe for target 'khdr' failed
> >>>>>>> make: *** [khdr] Error 2
> >>>>>>>
> >>>>>>>
> >>>>>> ...
> >>>>>>> Any thoughts ? ... or am I missing something ?
> >>>>>>
> >>>>>> You're not missing anything, this is broken.
> >>>>> Thanks for the feedback !
> >>>>>
> >>>>> Cristian
> >>>>>
> >>>>>>
> >>>>>>> (I think I'm starting to see this in latest CI linaro kselftest while they
> cross-compile for arm64)
> >>>>>>
> >>>>>> It just hit my travis jobs when I merged up to master:
> >>>>>>
> >>>>>>      https://travis-ci.org/linuxppc/linux/jobs/617482001
> >>>>>>
> >>>>>> Shuah can we please get this reverted?
> >>>>>>
> >>>>
> >>>> I will take care of this. Supporting all these use-cases has been
> >>>> maintenance problem. Fixing one use-case breaks another. :(
> >>>>
> >>>> I will look into fixing this once for all.
> >>>
> >>> Thanks.
> >>>
> >>> In fact all of the above works fine with the new kselftest-install target
> you added recently
> >>> but I think a lot of CI is still using the old -C method.
> >>>
> >>> Regards
> >>
> >> If all of the above works with kselftest-install target, I would rather
> >> ask for CI's to change, so we can get rid of make -C usage for install.
> >>
> >> In which case, I would rather not revert this and request CIs to start
> >> using kselftest-install target.
> >>
> >>
> >> I would like to see CI's use kselftest-* targets from the main makefile
> >> for build/run/install.
> >>
> >> It has been very difficult to support "make -C" use-cases in general
> >> and especially for install and relocatable builds.
> >>
> >
> > Well I'm not sure if it covers all the build cases/scenario for CIs and I have
> > only verified that this specific failure seems to work using the toplevel
> Makefile
> > targets (since the offending commit uses abs-objtree).
> > Anyway $ ./kselftest_install.sh is broken too as of now.
> >
> 
> With the revert, it will go back to not having support for relative
> paths :(
> 
> I will look into reverting the offending commit or fixing it.

There may not be many people using relative paths for KBUILD_OUTPUT.
I was affected by it, and it's quite handy given some of the mechanisms in
my CI system, but it's not that hard to work around.  (That is, I can just
convert relative paths to absolute paths in a few key places in my
CI, before calling the kernel make.)

Given that I'm the only person (seemingly) who reported this, and I'm ok
working around it, I think a revert is OK for now. 

Thanks.
 -- Tim