Re: [PATCH v3 00/23] Add generic vdso_base tracking

[H. Peter Anvin]

On 6/11/21 11:02, Dmitry Safonov wrote:
> Patches set is based on linux-next (next-20201123) and it depends on
> changes in x86/cleanups (those reclaim TIF_IA32/TIF_X32) and also
> on my changes in akpm (fixing several mremap() issues).
>
> Logically, the patches set divides on:
> - patch       1: a cleanup for patches in x86/cleanups
> - patches  2-13: cleanups for arch_setup_additional_pages()
> - patches 13-14: x86 signal changes for unmapped vdso
> - patches 15-22: provide generic vdso_base in mm_struct
> - patch      23: selftest for unmapped vDSO & fast syscalls
>
> In the end, besides cleanups, it's now more predictable what happens for
> applications with unmapped vdso on architectures those support .mremap()
> for vdso/sigpage.
>
> I'm aware of only one user that unmaps vdso - Valgrind [2].
> (there possibly are more, but this one is "special", it unmaps vdso, but
>   not vvar, which confuses CRIU [Checkpoint Restore In Userspace], that's
>   why I'm aware of it)

There was another discussion that might be relevant: actually 
associating the vdso with an actual file, and allowing a program to map 
said file normally if it want access to one that it normally wouldn't 
have (say, /proc/vdso/x86_64.so versus /proc/vdso/i386.so on the same 
system.)

The "catch", of course, is that this file will need to be mapped as 
MAP_SHARED because of vdso data.

	-hpa