Re: selftests/vm madv_populate.c test

From: David Hildenbrand
Date: Fri Oct 15 2021 - 12:28:49 EST


On 15.10.21 18:25, Shuah Khan wrote:
> On 10/15/21 10:19 AM, David Hildenbrand wrote:
>> On 15.10.21 18:15, David Hildenbrand wrote:
>>> On 15.10.21 18:06, David Hildenbrand wrote:
>>>> On 15.10.21 17:47, David Hildenbrand wrote:
>>>>> On 15.10.21 17:45, Shuah Khan wrote:
>>>>>> On 9/18/21 1:41 AM, David Hildenbrand wrote:
>>>>>>> On 18.09.21 00:45, Shuah Khan wrote:
>>>>>>>> Hi David,
>>>>>>>>
>>>>>>>> I am running into the following warning when try to build this test:
>>>>>>>>
>>>>>>>> madv_populate.c:334:2: warning: #warning "missing MADV_POPULATE_READ or MADV_POPULATE_WRITE definition" [-Wcpp]
>>>>>>>>     334 | #warning "missing MADV_POPULATE_READ or MADV_POPULATE_WRITE definition"
>>>>>>>>         |  ^~~~~~~
>>>>>>>>
>>>>>>>>
>>>>>>>> I see that the following handling is in place. However there is no
>>>>>>>> other information to explain why the check is necessary.
>>>>>>>>
>>>>>>>> #if defined(MADV_POPULATE_READ) && defined(MADV_POPULATE_WRITE)
>>>>>>>>
>>>>>>>> #else /* defined(MADV_POPULATE_READ) && defined(MADV_POPULATE_WRITE) */
>>>>>>>>
>>>>>>>> #warning "missing MADV_POPULATE_READ or MADV_POPULATE_WRITE definition"
>>>>>>>>
>>>>>>>> I do see these defined in:
>>>>>>>>
>>>>>>>> include/uapi/asm-generic/mman-common.h:#define MADV_POPULATE_READ       22
>>>>>>>> include/uapi/asm-generic/mman-common.h:#define MADV_POPULATE_WRITE      23
>>>>>>>>
>>>>>>>> Is this the case of missing include from madv_populate.c?
>>>>>>>
>>>>>>> Hi Shuan,
>>>>>>>
>>>>>>> note that we're including "#include <sys/mman.h>", which in my
>>>>>>> understanding maps to the version installed on your system instead
>>>>>>> of the one in our build environment.ing.
>>>>>>>
>>>>>>> So as soon as you have a proper kernel + the proper headers installed
>>>>>>> and try to build, it would pick up MADV_POPULATE_READ and
>>>>>>> MADV_POPULATE_WRITE from the updated headers. That makes sense: you
>>>>>>> annot run any MADV_POPULATE_READ/MADV_POPULATE_WRITE tests on a kernel
>>>>>>> that doesn't support it.
>>>>>>>
>>>>>>> See vm/userfaultfd.c where we do something similar.
>>>>>>>
>>>>>>
>>>>>> Kselftest is for testing the kernel with kernel headers. That is the
>>>>>> reason why there is the dependency on header install.
>>>>>>
>>>>>>>
>>>>>>> As soon as we have a proper environment, it seems to work just fine:
>>>>>>>
>>>>>>> Linux vm-0 5.15.0-0.rc1.20210915git3ca706c189db.13.fc36.x86_64 #1 SMP Thu Sep 16 11:32:54 UTC 2021 x86_64 x86_64 x86_64 GNU/Linux
>>>>>>> [root@vm-0 linux]# cat /etc/redhat-release
>>>>>>> Fedora release 36 (Rawhide)
>>>>>>
>>>>>> This is a distro release. We don't want to have dependency on headers
>>>>>> from the distro to run selftests. Hope this makes sense.
>>>>>>
>>>>>> I still see this on my test system running Linux 5.15-rc5.
>>>>>
>>>>> Did you also install Linux headers? I assume no, correct?
>>>>>
>>>>
>>>> What happens in your environment when compiling and running the
>>>> memfd_secret test?
>>>>
>>>> If assume you'll see a "skip" when executing, because it might also
>>>> refer to the local version of linux headers and although it builds, it
>>>> really cannot build something "functional". It just doesn't add a
>>>> "#warning" to make that obvious.
>>>>
>>>
>>> The following works but looks extremely hackish.
>>>
>>> diff --git a/tools/testing/selftests/vm/madv_populate.c
>>> b/tools/testing/selftests/vm/madv_populate.c
>>> index b959e4ebdad4..ab26163db540 100644
>>> --- a/tools/testing/selftests/vm/madv_populate.c
>>> +++ b/tools/testing/selftests/vm/madv_populate.c
>>> @@ -14,12 +14,11 @@
>>> #include <unistd.h>
>>> #include <errno.h>
>>> #include <fcntl.h>
>>> +#include "../../../../usr/include/linux/mman.h"
>>> #include <sys/mman.h>
>>>
>>> #include "../kselftest.h"
>>>
>>> -#if defined(MADV_POPULATE_READ) && defined(MADV_POPULATE_WRITE)
>>> -
>>> /*
>>> * For now, we're using 2 MiB of private anonymous memory for all tests.
>>> */
>>> @@ -328,15 +327,3 @@ int main(int argc, char **argv)
>>> err, ksft_test_num());
>>> return ksft_exit_pass();
>>> }
>>> -
>>> -#else /* defined(MADV_POPULATE_READ) && defined(MADV_POPULATE_WRITE) */
>>> -
>>> -#warning "missing MADV_POPULATE_READ or MADV_POPULATE_WRITE definition"
>>> -
>>> -int main(int argc, char **argv)
>>> -{
>>> - ksft_print_header();
>>> - ksft_exit_skip("MADV_POPULATE_READ or MADV_POPULATE_WRITE not
>>> defined\n");
>>> -}
>>> -
>>> -#endif /* defined(MADV_POPULATE_READ) && defined(MADV_POPULATE_WRITE) */
>>>
>>>
>>> There has to be some clean way to achieve the same.
>>>
>>
>> Sorry for the spam,
>>
>> diff --git a/tools/testing/selftests/vm/Makefile
>> b/tools/testing/selftests/vm/Makefile
>> index d9605bd10f2d..ce198b329ff5 100644
>> --- a/tools/testing/selftests/vm/Makefile
>> +++ b/tools/testing/selftests/vm/Makefile
>> @@ -23,7 +23,7 @@ MACHINE ?= $(shell echo $(uname_M) | sed -e
>> 's/aarch64.*/arm64/' -e 's/ppc64.*/p
>> # LDLIBS.
>> MAKEFLAGS += --no-builtin-rules
>>
>> -CFLAGS = -Wall -I ../../../../usr/include $(EXTRA_CFLAGS)
>> +CFLAGS = -Wall -idirafter ../../../../usr/include $(EXTRA_CFLAGS)
>> LDLIBS = -lrt -lpthread
>> TEST_GEN_FILES = compaction_test
>> TEST_GEN_FILES += gup_test
>>
>>
>> Seems to set the right include path priority.
>>
>>
>
> Yes. It works on linux-next-20211012
>
> Do you mind sending a me patch for this?

I just double-checked (after make clean) and there is still something
wrong :( the only think that seems to work is the

+#include "../../../../usr/include/linux/mman.h"
#include <sys/mman.h>

hack.

Using "-nostdinc" won't work because we need other headers :(

--
Thanks,

David / dhildenb