Re: [PATCH 0/4] mm: Avoid sharing high VMA flag bits

[Mark Brown <broonie@kernel.org>]

[-- Attachment #1: Type: text/plain, Size: 2214 bytes --]

On Wed, May 07, 2025 at 03:09:50PM +0200, Florent Revest wrote:

> I just want to make sure I fully understand the point you're making
> here, it seems that you are suggesting that 7677f7fd8be7
> ("userfaultfd: add minor fault registration mode") and ae80e1629aea
> ("mm: Define VM_SHADOW_STACK for arm64 when we support GCS") came in
> from two different trees and independently chose to use the same bit
> around the ~same time, is that right ? And that a conflict would have
> appeared when they'd eventually get merged into a shared tree ?

> I don't think that's what happened in this specific case though. As
> far as I can tell, the former was merged in 2021 and the latter was
> merged in late 2024. Also, since the hunks got added in very different

Yes, indeed - I wrote that initial reply before I saw the actual change.
I wasn't expecting a conflict but rather that what'd happened was that
the bit was allocated independently in two different trees and then due
to the way the allocations are scattered everywhere no conflict had
flagged the issue.

> parts of include/linux/mm.h, I don't think they caused a noticeable
> merge conflict. But I agree it would probably be preferable if these
> cases would cause some sort of noticeable merge conflict in the future

Putting all the flags somewhere in the vicinity of each other would make
the whole thing much more robust, or having some build assert for
uniqueness.

> I'll quickly respin this series to address my typos on patch 4 (sigh)
> and add your Reviewed-by tag but just to be clear, my refactorings in
> patches 2/3/4 currently focus on using VM_HIGH_ARCH macros
> consistently, to make it a bit more obvious to a reader if two
> features choose the same bit. But maybe what we would really need
> instead is a more obvious way for these bits to be mutually exclusive
> and to cause merge conflicts if they get added through independent
> trees ? For example, my colleague Brendan Jackman suggested using an
> enum for VMA flags bit offsets but I'm not sure what the sentiment is
> around that.

I think we need something here, although it wasn't what actually
happened here the potential for the scenario I mentioned above to occur
remains.

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 488 bytes --]