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

[shuah <shuah@kernel.org>]

On 11/27/19 10:57 AM, Tim.Bird@sony.com wrote:
> 
> 
>> -----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.
> 

Sorry Cristian! You reported this a couple of weeks ago and I missed
your email. I was away on family emergency and this email got buried
in my Inbox. No excuses though. My bad.

Cristian suggested this fix. There is another issue here.

INSTALL_PATH conflicts with main Makefile variable. The recent
kbuild changes impacted kselftest use-cases and I had to drop
support for INSTALL_PATH in kselftest_install.sh and move
to KSFT_INSTALL_PATH - sorry the document is outdated.

I should update it. I would like to see CIs migrate to using
KSFT_INSTALL_PATH

If we go with the below fix, does it fix for both O and KBUILD_OUTPUT
cases, does it cover all cases?

A possible fix would be (but duplicates in fact the main Makefile logic)


diff --git a/tools/testing/selftests/Makefile 
b/tools/testing/selftests/Makefile
index 319e094c3212..491d8b3ef1c7 100644
--- a/tools/testing/selftests/Makefile
+++ b/tools/testing/selftests/Makefile
@@ -88,6 +88,15 @@ ifdef building_out_of_srctree
  override LDFLAGS =
  endif

+ifeq ($(abs_objtree),)
+ifneq ($(KBUILD_OUTPUT),)
+abs_objtree := $(shell cd $(KBUILD_OUTPUT) && pwd)
+abs_objtree := $(realpath $(abs_objtree))
+else
+abs_objtree := $(shell pwd)
+endif
+endif #ifeq ($(abs_objtree),)
+
  ifneq ($(O),)
         BUILD := $(abs_objtree)
  else

thanks,
-- Shuah