On Tue, Aug 18, 2026 at 02:36:00PM +0000, Lorenzo Stoakes (ARM) wrote:
> But it might be worth checking in this case to see whether the problem here
> is in fact that we need PKEY_UNRESTRICTED defined over there?
>
> Hemanth - could you check please?
Checked. It is already defined over there:
tools/include/uapi/asm-generic/mman-common.h:88:#define PKEY_UNRESTRICTED
0x0
and it does reach the mm selftests without "make headers", through
TOOLS_INCLUDES. The path is not obvious, so for the record:
pkey-helpers.h
-> <linux/mman.h> tools/include/uapi/linux/mman.h
-> <asm/mman.h> /usr/include/asm/mman.h
(tools/include/uapi/asm/mman.h does not exist)
-> <asm-generic/mman.h> tools/include/uapi/asm-generic/mman.h
(-isystem puts it ahead of the system one)
-> <asm-generic/mman-common-tools.h>
-> <asm-generic/mman-common.h> PKEY_UNRESTRICTED
I built the pkey tests with KHDR_INCLUDES pointing at an empty directory,
that is, as if "make headers" had never run. They build, and cpp -dD
says:
defined in: tools/include/uapi/asm-generic/mman-common.h
Pointing TOOLS_INCLUDES at an empty directory as well is what breaks it:
pkey-helpers.h:188:30: error: 'PKEY_UNRESTRICTED' undeclared
So that is where the definition comes from, and dropping the local
#ifndef does not depend on "make headers" being run.
The one thing worth being aware of is that the redirection into
tools/include/uapi only happens because the system <asm/mman.h> includes
<asm-generic/mman.h>.
I had a look at the history you pointed at, thanks, that was the part I
had wrong. If I read it right this is exactly the mechanism from commit
e076eaca5906 ("selftests: break the dependency upon local header
files"), which added TOOLS_INCLUDES to the mm selftests so they build
without "make headers". So for this particular macro the value comes
from the in-tree snapshot rather than from whatever the build host has
installed, which is why it does not vary per system here.
That said, if you would still rather keep the #ifndef as a matter of
course, I have no objection to dropping the patch, the guard costs
nothing. I mainly wanted to answer the question with something
measured rather than assumed.
> This is incorrect. We explicitly do not have this requirement in mm selftests,
> and they're often built without having to do this step and tooling has been
> provided _explicitly_ to allow for that, which is why this kind of thing
> exists
> right now.
You are right, the commit message was wrong on that point. v2 drops
that paragraph and describes the above instead.
On the powerpc header David mentioned, I would rather leave
tools/testing/selftests/powerpc/include/pkeys.h alone. It is not the
same construct: it is an unconditional
#undef PKEY_UNRESTRICTED
#define PKEY_UNRESTRICTED 0x0
next to other powerpc specific overrides such as PKEY_DISABLE_EXECUTE
0x4, and that header only includes <sys/mman.h>, never <linux/mman.h>.
A probe including just <sys/mman.h> does not find PKEY_UNRESTRICTED,
with or without TOOLS_INCLUDES, so the definition there is load bearing
and removing it would break the build.
Thanks,
Hemanth
