Re: [PATCH RFC v2] mm: use per_vma lock for MADV_DONTNEED

[David Hildenbrand]

On 09.06.25 08:40, Qi Zheng wrote:
> Hi Lorenzo,
>
> On 6/6/25 6:44 PM, Lorenzo Stoakes wrote:
>
> [snip]
>
> > > > We could in theory always add another callback .pmd_entry_sleep or
> > > > something for this one case and document the requirement...
> > >
> > > Maybe, but the SRCU critical section cannot prevent the PTE page from
> > > being freed via RCU. :(
> >
> > Idea is we'd fall back to non-RCU in this case and take locks... but then
> > ugh we'd race everything RCU and no it's all or nothing isn't it?
>
> So maybe the RCU+refcount method is feasible. We can release the RCU
> lock after incrementing the reference count, which can ensure that the
> page table page is not freed.

I'll not that maybe after the memdesc rework, page tables will no longer 
have a refcount. Maybe.

I mean, it kind-of makes sense, because nobody should really be taking 
references on page table.

So finding something that doesn't depend on page-table refcounts might 
be better.

--
Cheers,

David / dhildenb