Re: [PATCH] ARM64: Implement arch_report_meminfo()

["Christoph Lameter (Ampere)" <cl@linux.com>]

On Fri, 15 Dec 2023, Robin Murphy wrote:

> On 14/12/2023 9:35 pm, Christoph Lameter (Ampere) wrote:
>> On Thu, 14 Dec 2023, Robin Murphy wrote:
>> 
>>> It seems somewhat suspect that these counts only ever increase. It's not 
>>> often that we change or remove parts of the linear map, but it certainly 
>>> can happen.
>> 
>> Well yes in the case of hotplug I guess ... Ok here is V2
>
> There are also paths where we remove and reinstate parts of the linear map 
> via set_memory_valid(), set_direct_map_*(), and possibly others. If we're 
> exposing a user ABI that claims to be accounting kernel VA mappings, then I 
> think users are within their rights to expect it to actually account kernel 
> VA mappings, not just expose numbers whose only guaranteed significance is 
> whether they are zero or nonzero.

set_memory_valid() changes mappings via __change_memory_common()
and apply_to_page_range(). It seems that apply_to_page range() creates
PTEs as needed etc.

However, I do not see any accounting for direct map modification 
accounting on x86 either. Since this was satifactory for x86 I dont 
believe that more is needed. Introducing atomics in potentially 
performance sensitive functions run when the kernel is up is not that 
advisable and doing so would require core kernel changes going beyond the 
enablement of arch_report_meminfo() on ARM64.

> Looking again, am I also right in thinking that what I assumed were the 
> non-contiguous counts here are actually total counts of *either* type of 
> mapping at that level, and inclusive of the contiguous counts? If so, that 
> seems a bit non-obvious - my intuitive expectation would be for the sum of 
> all these numbers to represent the total amount of direct-mapped RAM, where 
> either we're intersted in each distinct type of mapping and accounting them 
> all separately, or we're simply interested in the general shape of the 
> pagetables, and thus would account per-level and ignore the contiguous bit 
> since we don't know whether it's actually doing anything useful anyweay

Yes, the CONT PTEs are a subset of the other counts since they are only a 
special case of the PTE type. They are important for performance on ARM64 
in particular with the anticipated use of them for the various sizes of 
pages supported in the kernel with the introduction of folios for the page 
cache.

The problem with CONT_PTE is that it is not clear whether the architecture 
supports it or now. The amount of CONT_PTE can influence the TLB coverage 
possible in kernel space.

We are generally interested in the shape of the page tables. If the user 
later uses processes that require a degradation through smaller mappings 
then that is load dependent.