Re: [PATCH v9 10/21] mm: Add generic p?d_leaf() macros

[Anshuman Khandual]

On 07/29/2019 05:08 PM, Steven Price wrote:
> On 28/07/2019 12:44, Anshuman Khandual wrote:
>>
>>
>> On 07/23/2019 03:11 PM, Mark Rutland wrote:
>>> On Mon, Jul 22, 2019 at 04:41:59PM +0100, Steven Price wrote:
>>>> Exposing the pud/pgd levels of the page tables to walk_page_range() means
>>>> we may come across the exotic large mappings that come with large areas
>>>> of contiguous memory (such as the kernel's linear map).
>>>>
>>>> For architectures that don't provide all p?d_leaf() macros, provide
>>>> generic do nothing default that are suitable where there cannot be leaf
>>>> pages that that level.
>>>>
>>>> Signed-off-by: Steven Price <steven.price@xxxxxxx>
>>>
>>> Not a big deal, but it would probably make sense for this to be patch 1
>>> in the series, given it defines the semantic of p?d_leaf(), and they're
>>> not used until we provide all the architectural implemetnations anyway.
>>
>> Agreed.
>>
>>>
>>> It might also be worth pointing out the reasons for this naming, e.g.
>>> p?d_large() aren't currently generic, and this name minimizes potential
>>> confusion between p?d_{large,huge}().
>>
>> Agreed. But these fallback also need to first check non-availability of large
>> pages. I am not sure whether CONFIG_HUGETLB_PAGE config being clear indicates
>> that conclusively or not. Being a page table leaf entry has a broader meaning
>> than a large page but that is really not the case today. All leaf entries here
>> are large page entries from MMU perspective. This dependency can definitely be
>> removed when there are other types of leaf entries but for now IMHO it feels
>> bit problematic not to directly associate leaf entries with large pages in
>> config restriction while doing exactly the same.
> 
> The intention here is that the page walkers are able to walk any type of
> page table entry which the kernel may use. CONFIG_HUGETLB_PAGE only
> controls whether "huge TLB pages" are used by user space processes. It's
> quite possible that option to not be selected but the linear mapping to
> have been mapped using "large pages" (i.e. leaf entries further up the
> tree than normal).

I understand that kernel page table might use large pages where as user space
never enabled HugeTLB. The point to make here was CONFIG_HUGETLB approximately
indicates the presence of large pages though the absence of same does not
conclusively indicate that large pages are really absent on the MMU. Perhaps it
will requires something new like MMU_[LARGE|HUGE]_PAGES.

> 
> One of the goals was to avoid tying the new functions to a configuration
> option but instead match the hardware architecture. Of course this isn't
> possible in the most general case (e.g. an architecture may have
> multiple hardware page table formats). But to the extent that other
> functions like p?d_none() work the desire is that p?d_leaf() should also
> work.

It is fair enough to assume that a platform can decide wisely and provide
accurate definition for p?d_leaf() functions. Anyways its okay not to make
this more complex by tying with a new config option which does not exist.