[PATCH v2 1/2] mm: Allow lockless kernel pagetable walking

[David Hildenbrand]

On 10.06.25 15:27, Lorenzo Stoakes wrote:
> On Tue, Jun 10, 2025 at 03:24:16PM +0200, David Hildenbrand wrote:
>> On 10.06.25 14:07, Lorenzo Stoakes wrote:
>>> OK so I think the best solution here is to just update check_ops_valid(), which
>>> was kind of sucky anyway (we check everywhere but walk_page_range_mm() to
>>> enforce the install pte thing).
>>>
>>> Let's do something like:
>>>
>>> #define OPS_MAY_INSTALL_PTE	(1<<0)
>>> #define OPS_MAY_AVOID_LOCK	(1<<1)
>>>
>>> and update check_ops_valid() to take a flags or maybe 'capabilities' field.
>>>
>>> Then check based on this e.g.:
>>>
>>> if (ops->install_pte && !(capabilities & OPS_MAY_INSTALL_PTE))
>>> 	return false;
>>>
>>> if (ops->walk_lock == PGWALK_NOLOCK && !(capabilities & OPS_MAY_AVOID_LOCK))
>>> 	return false;
>>>
>>
>> Hm. I mean, we really only want to allow this lockless check for
>> walk_kernel_page_table_range(), right?
>>
>> Having a walk_kernel_page_table_range_lockeless() might (or might not) be
>> better, to really only special-case this specific path.
> 
> Agree completely, Dev - let's definitely do this.
> 
>>
>> So, I am wondering if we should further start splitting the
>> kernel-page-table walker up from the mm walker, at least on the "entry"
>> function for now.
> 
> How do you mean?

In particular, "struct mm_walk_ops"

does not quite make sense when not actually walking a "real" mm .

So maybe we should start having a separate structure where *vma, 
install_pte, walk_lock, hugetlb* does not even exist.

It might be a bit of churn, though ... not sure if there could be an 
easy translation layer for now.

-- 
Cheers,

David / dhildenb