Re: [PATCH bpf-next] selftests/bpf: fix uprobe_multi compilation error

[Alan Maguire <alan.maguire@oracle.com>]

On 27/09/2024 00:44, Eduard Zingerman wrote:
> On Thu, 2024-09-26 at 15:49 +0100, Alan Maguire wrote:
>> When building selftests, the following was seen:
>>
>> uprobe_multi.c: In function ‘trigger_uprobe’:
>> uprobe_multi.c:108:40: error: ‘MADV_PAGEOUT’ undeclared (first use in this function)
>>   108 |                 madvise(addr, page_sz, MADV_PAGEOUT);
>>       |                                        ^~~~~~~~~~~~
>> uprobe_multi.c:108:40: note: each undeclared identifier is reported only once for each function it appears in
>> make: *** [Makefile:850: bpf-next/tools/testing/selftests/bpf/uprobe_multi] Error 1
>>
>> ...even with updated UAPI headers. It seems the above value is
>> defined in UAPI <linux/mman.h> but including that file triggers
>> other redefinition errors.  Simplest solution is to add a
>> guarded definition, as was done for MADV_POPULATE_READ.
>>
>> Fixes: 3c217a182018 ("selftests/bpf: add build ID tests")
>> Signed-off-by: Alan Maguire <alan.maguire@oracle.com>
>> ---
> 
> I was curious why this error is not triggered on my local machine or CI.
> MADV_PAGEOUT is indeed defined in UAPI.
> Selftests build picks it from host system header, which is
> /usr/include/bits/mman-linux.h for my Fedora 40 setup.
> The MADV_PAGEOUT was added by commit [1] back in 2019
> (and should be available from Linux 5.4, I guess Alan uses a very old kernel).
> 
> I think that at some point in time we should adjust selftests to use
> UAPI headers that come from the kernel being tested, not from the host.
> Until that happens, I think this fix is fine.
> 

From what I can see at my end, /usr/include/bits/mman.h is provided by
glibc-headers. On distros the glibc-headers can get quite out of date. I
do a "make headers_install" prior to running bpf selftests to get latest
UAPI headers from the kernel in /usr/include/linux, but that wasn't
enough in this case since /usr/include/sys/mman.h and
/usr/include/bits/mman.h were provided via glibc-headers.

When I tried using <linux/mman.h> in the program, I ended up getting
complaints about madvise() being undefined, and when I used both
<sys/mman.h> and <linux/mman.h> there were complaints about mutiple
definitions so the approach taken in the patch seemed like the most
straightforward.

Thanks for taking a look!

Alan

> [1] 1a4e58cce84e ("mm: introduce MADV_PAGEOUT")
> 
> Acked-by: Eduard Zingerman <eddyz87@gmail.com>
> 
> (I want back to 2019...)
> 
> [...]
>