Re: [PATCH 1/1] x86/vmemmap: Use direct-mapped VA instead of vmemmap-based VA

[Gwan-gyeong Mun]

Sorry for the late reply.

On 3/6/25 6:09 PM, Dave Hansen wrote:
> On 3/5/25 19:46, Andrew Morton wrote:
> > Can we please have review from x86 maintainers?
>
> I didn't respond here because I thought this was the same problem from
> the same contributor that we addressed in another thread:
>
> > https://lore.kernel.org/all/d1da214c-53d3-45ac-a8b6-51821c5416e4@xxxxxxxxx/
>
> I think this approach is a hack. It basically requires that every bit of
> code that _might_ update (and then use) a PGD in the init_mm know how to
> find the direct map alias and then use that instead. This would further
> specialize the x86 code. I have no reason to believe that this is truly
> an x86-specific problem. Are we really the only arch that has a
> per-process PGD that maps the shared kernel page tables?
currently the scenario that causes this issue only happens on x86, 
therfore I did a hotfix to make sure it would not affect other archtecture.

And I agree, I don't think this is an x86 only issue, and if there is a 
scenario that can cause the situation I reported on other platforms, I 
think it is possible.
> The right solution (like I mentioned in the other thread) is to sync the
> PGDs more aggressively. Syncing them is expensive, of course, but it's a
> pretty darn rare operation.
I'll send a new patch that introdues helper functions suggested by Dave 
above.

Thank you for not forgetting this issue and bringing it back to the surface.

Thanks,

G.G.