Re: [PATCH 0/2] don't use mapcount() to check large folio sharing

[David Hildenbrand <david@redhat.com>]

On 02.08.23 13:51, Ryan Roberts wrote:
> On 02/08/2023 12:36, David Hildenbrand wrote:
>> On 02.08.23 13:20, Ryan Roberts wrote:
>>> On 02/08/2023 11:48, David Hildenbrand wrote:
>>>> On 02.08.23 12:27, Ryan Roberts wrote:
>>>>> On 28/07/2023 17:13, Yin Fengwei wrote:
>>>>>> In madvise_cold_or_pageout_pte_range() and madvise_free_pte_range(),
>>>>>> folio_mapcount() is used to check whether the folio is shared. But it's
>>>>>> not correct as folio_mapcount() returns total mapcount of large folio.
>>>>>>
>>>>>> Use folio_estimated_sharers() here as the estimated number is enough.
>>>>>>
>>>>>> Yin Fengwei (2):
>>>>>>      madvise: don't use mapcount() against large folio for sharing check
>>>>>>      madvise: don't use mapcount() against large folio for sharing check
>>>>>>
>>>>>>     mm/huge_memory.c | 2 +-
>>>>>>     mm/madvise.c     | 6 +++---
>>>>>>     2 files changed, 4 insertions(+), 4 deletions(-)
>>>>>>
>>>>>
>>>>> As a set of fixes, I agree this is definitely an improvement, so:
>>>>>
>>>>> Reviewed-By: Ryan Roberts
>>>>>
>>>>>
>>>>> But I have a couple of comments around further improvements;
>>>>>
>>>>> Once we have the scheme that David is working on to be able to provide precise
>>>>> exclusive vs shared info, we will probably want to move to that. Although that
>>>>> scheme will need access to the mm_struct of a process known to be mapping the
>>>>
>>>> There are probably ways to work around lack of mm_struct, but it would not be
>>>> completely for free. But passing the mm_struct should probably be an easy
>>>> refactoring.
>>>>
>>>>> folio. We have that info, but its not passed to folio_estimated_sharers() so we
>>>>> can't just reimplement folio_estimated_sharers() - we will need to rework these
>>>>> call sites again.
>>>>
>>>> We should probably just have a
>>>>
>>>> folio_maybe_mapped_shared()
>>>>
>>>> with proper documentation. Nobody should care about the exact number.
>>>>
>>>>
>>>> If my scheme for anon pages makes it in, that would be precise for anon pages
>>>> and we could document that. Once we can handle pagecache pages as well to get a
>>>> precise answer, we could change to folio_mapped_shared() and adjust the
>>>> documentation.
>>>
>>> Makes sense to me. I'm assuming your change would allow us to get rid of
>>> PG_anon_exclusive too? In which case we would also want a precise API
>>> specifically for anon folios for the CoW case, without waiting for pagecache
>>> page support.
>>
>> Not necessarily and I'm currently not planning that
>>
>> On the COW path, I'm planning on using it only when PG_anon_exclusive is clear
>> for a compound page, combined with a check that there are no other page
>> references besides from mappings: all mappings from me and #refs == #mappings ->
>> reuse (set PG_anon_exclusive). That keeps the default (no fork) as fast as
>> possible and simple.
>>
>>>>
>>>> I just saw
>>>>
>>>> https://lkml.kernel.org/r/20230802095346.87449-1-wangkefeng.wang@huawei.com
>>>>
>>>> that converts a lot of code to folio_estimated_sharers().
>>>>
>>>>
>>>> That patchset, for example, also does
>>>>
>>>> total_mapcount(page) > 1 -> folio_estimated_sharers(folio) > 1
>>>>
>>>> I'm not 100% sure what to think about that at this point. We eventually add
>>>> false negatives (actually shared but we fail to detect it) all over the place,
>>>> instead of having false positives (actually exclusive, but we fail to detect
>>>> it).
>>>>
>>>> And that patch set doesn't even spell that out.
>>>>
>>>>
>>>> Maybe it's as good as we will get, especially if my scheme doesn't make it in.
>>>
>>> I've been working on the assumption that your scheme is plan A, and I'm waiting
>>> for it to unblock forward progress on large anon folios. Is this the right
>>> approach, or do you think your scheme is sufficiently riskly and/or far out that
>>> I should aim not to depend on it?
>>
>> It is plan A. IMHO, it does not feel too risky and/or far out at this point --
>> and the implementation should not end up too complicated. But as always, I
>> cannot promise anything before it's been implemented and discussed upstream.
> 
> OK, good we are on the same folio... (stolen from Hugh; if a joke is worth
> telling once, its worth telling 1000 times ;-)

Heard it first the time :))

-- 
Cheers,

David / dhildenb