Re: [PATCH v2 1/8] x86: vdso: Introduce asm/vdso/mman.h

["Arnd Bergmann" <arnd@arndb.de>]

On Wed, Sep 25, 2024, at 06:51, Christophe Leroy wrote:
> Le 23/09/2024 à 16:19, Vincenzo Frascino a écrit :
>> @@ -0,0 +1,15 @@
>> +
>> +/* SPDX-License-Identifier: GPL-2.0 */
>> +#ifndef __ASM_VDSO_MMAN_H
>> +#define __ASM_VDSO_MMAN_H
>> +
>> +#ifndef __ASSEMBLY__
>> +
>> +#include <uapi/linux/mman.h>
>> +
>> +#define VDSO_MMAP_PROT	PROT_READ | PROT_WRITE
>> +#define VDSO_MMAP_FLAGS	MAP_DROPPABLE | MAP_ANONYMOUS
>
> I still can't see the point with that change.
>
> Today 4 architectures implement getrandom and none of them require that 
> indirection. Please leave prot and flags as they are in the code.
>
> Then this file is totally pointless, VDSO code can include 
> uapi/linux/mman.h directly.
>
> VDSO is userland code, it should be safe to include any UAPI file there.

I think we are hitting an unfortunate corner case in the build
system here, based on the way we handle the uapi/ file namespace
in the kernel:

include/uapi/linux/mman.h includes three headers: asm/mman.h,
asm-generic/hugetlb_encode.h and linux/types.h. Two of these
exist in both include/uapi/ and include/, so while building
kernel code we end up picking up the non-uapi version which
on some architectures includes many other headers.

I agree that moving the contents out of uapi/ into vdso/ namespace
is not a solution here because that removes the contents from
the installed user headers, but we still need to do something
to solve the issue.

The easiest workaround I see for this particular file is to
move the contents of arch/{arm,arm64,parisc,powerpc,sparc,x86}/\
include/asm/mman.h into a different file to ensure that the
only existing file is the uapi/ one. Unfortunately this does
not help to avoid it regressing again in the future.

To go a little step further I would also move
uapi/asm-generic/hugetlb_encode.h to uapi/linux/hugetlb_encode.h
or merge it into uapi/linux/mman.h. This file has no business
in asm-generic/* since there is only one copy.

After looking at this file for way too long, I somehow
ended up with a (completely unrelated) cleanup series that
I now posted at
https://lore.kernel.org/lkml/20240925210615.2572360-1-arnd@kernel.org/T/#t

     Arnd