Re: [PATCH v4 17/21] mm/mmap: Drop arch_unmap() call from all archs

From: Liam R. Howlett
Date: Thu Jul 11 2024 - 12:00:17 EST


* LEROY Christophe <christophe.leroy2@xxxxxxxxxxxxxxxxxx> [240711 04:28]:
>
>
> Le 11/07/2024 à 01:26, Liam R. Howlett a écrit :
> > * LEROY Christophe <christophe.leroy2@xxxxxxxxxxxxxxxxxx> [240710 17:02]:
> >>
> >>
> >> Le 10/07/2024 à 21:22, Liam R. Howlett a écrit :
> >>> From: "Liam R. Howlett" <Liam.Howlett@xxxxxxxxxx>
> >>>
> >>> The arch_unmap call was previously moved above the rbtree modifications
> >>> in commit 5a28fc94c914 ("x86/mpx, mm/core: Fix recursive munmap()
> >>> corruption"). The move was motivated by an issue with calling
> >>> arch_unmap() after the rbtree was modified.
> >>>
> >>> Since the above commit, mpx was dropped from the kernel in 45fc24e89b7c
> >>> ("x86/mpx: remove MPX from arch/x86"), so the motivation for calling
> >>> arch_unmap() prior to modifying the vma tree no longer exists
> >>> (regardless of rbtree or maple tree implementations).
> >>>
> >>> Furthermore, the powerpc implementation is also no longer needed as per
> >>> [1] and [2]. So the arch_unmap() function can be completely removed.
> >>
> >> I'm not sure to understand. Is it replaced by something else ?
> >> We wanted to get rid of arch_unmap() but it was supposed to be replaced
> >> by some core function because the functionnality itself is still
> >> required and indeed all the discussion around [2] demonstrated that not
> >> only powerpc but at least arm and probably others needed to properly
> >> clean-up reference to VDSO mappings on unmapping.
> >>
> >> So as mentioned by Michael you can't just drop that without replacing it
> >> by something else. We need the VDSO signal handling to properly fallback
> >> on stack-based trampoline when the VDSO trampoline gets mapped out.
> >
> > I'll address this after the part I missed..
>
> After ? What do you mean ? It needs to be addressed _before_ removing
> arch_unmap()

After the later comments in this email, sorry that wasn't clear.

>
> >
> >>
> >> Or did I miss something ?
> >>
> >
> > I think I missed something in regards to what you need in ppc.
>
> It is not only powerpc. Powerpc is the only one doing it at the moment
> but investigation has demonstrated that other architectures are affected.
>
> >
> > From what I understand, other platforms still map and use the vdso
> > (context.vdso is set), but unmap_arch() does nothing. It is only the
> > powerpc version that clears the vdso pointer if it is unmapped.
>
> Yes on powerpc it works. On other platforms like arm it segfaults so it
> should be fixed
> (https://lore.kernel.org/lkml/87imd5h5kb.fsf@xxxxxxxxxxxxxxxxxx/)
>
> Could be fixed by properly implementing arch_unmap() on every arch, or
> carry-on with Dmitry's series.

Okay, I understand what you are saying now. I'm not going to tackle
that change within this series, so I'll just relocate the arch_munmap()
back to where it was, after the removal of the vmas in v5.

> I think you fully understand that it doesn't work as it is except on
> powerpc. Again the goal should be to make it work on all architectures.

Got it, thanks for clarifying.

Regards,
Liam