Re: [PATCH v1 2/2] mm: mTHP stats for pagecache folio allocations

[David Hildenbrand <david@redhat.com>]

On 17.07.24 12:18, Ryan Roberts wrote:
> On 17/07/2024 11:03, David Hildenbrand wrote:
>>>>>>>
>>>>>>> But today, controls and stats are exposed for:
>>>>>>>
>>>>>>>       anon:
>>>>>>>         min order: 2
>>>>>>>         max order: PMD_ORDER
>>>>>>>       anon-shmem:
>>>>>>>         min order: 2
>>>>>>>         max order: PMD_ORDER
>>>>>>>       tmpfs-shmem:
>>>>>>>         min order: PMD_ORDER
>>>>>>>         max order: PMD_ORDER
>>>>>>>       file:
>>>>>>>         min order: Nothing yet (this patch proposes 1)
>>>>>>>         max order: Nothing yet (this patch proposes MAX_PAGECACHE_ORDER)
>>>>>>>
>>>>>>> So I think there is definitely a bug for shmem where the minimum order
>>>>>>> control
>>>>>>> should be order-1 but its currently order-2.
>>>>>>
>>>>>> Maybe, did not play with that yet. Likely order-1 will work. (although
>>>>>> probably
>>>>>> of questionable use :) )
>>>>>
>>>>> You might have to expand on why its of "questionable use". I'd assume it has
>>>>> the
>>>>> same amount of value as using order-1 for regular page cache pages? i.e. half
>>>>> the number of objects to manage for the same amount of memory.
>>>>
>>>> order-1 was recently added for the pagecache to get some device setups running
>>>> (IIRC, where we cannot use order-0, because device blocksize > PAGE_SIZE).
>>>>
>>>> You might be right about "half the number of objects", but likely just going for
>>>> order-2, order-3, order-4 ... for shmem might be even better. And simply falling
>>>> back to order-0 when you cannot get the larger orders.
>>>
>>> Sure, but then you're into the territory of baking in policy. Remember that
>>> originally I was only interested in 64K but the concensus was to expose all the
>>> sizes. Same argument applies to 8K; expose it and let others decide policy.
>>
>> I don't disagree. The point I'm trying to make is that there was so far there
>> was no strong evidence that it is really required. Support for the pagecache had
>> a different motivation for these special devices.
> 
> Sure, but there was no clear need for anon mTHP orders other than order-2 and
> order-4 (for arm64's HPA and contpte, respectively), but we still chose to
> expose all the others.

order-2 and order-3 are valuable for AMD EPYC (depending on the 
generation 16 vs. 32 KiB coalescing).

But in general, at least for me, it's easier to argue why larger orders 
make more sense than very tiny ones.

For example, order-5 can be mapped using cont-pte as well and you get 
roughly half the memory allocation+page fault overhead compared to order-4.

order-1 ? No TLB optimization at least on any current HW I know.

But I believe we're in violent agreement here :)

-- 
Cheers,

David / dhildenb