Re: [PATCH v3 3/6] mm: introduce pte_move_swp_offset() helper which can move offset bidirectionally

[Ryan Roberts]

On 06/05/2024 09:31, David Hildenbrand wrote:
> On 06.05.24 10:20, Barry Song wrote:
>> On Mon, May 6, 2024 at 8:06 PM David Hildenbrand <david@xxxxxxxxxx> wrote:
>>>
>>> On 04.05.24 01:40, Barry Song wrote:
>>>> On Fri, May 3, 2024 at 5:41 PM Ryan Roberts <ryan.roberts@xxxxxxx> wrote:
>>>>>
>>>>> On 03/05/2024 01:50, Barry Song wrote:
>>>>>> From: Barry Song <v-songbaohua@xxxxxxxx>
>>>>>>
>>>>>> There could arise a necessity to obtain the first pte_t from a swap
>>>>>> pte_t located in the middle. For instance, this may occur within the
>>>>>> context of do_swap_page(), where a page fault can potentially occur in
>>>>>> any PTE of a large folio. To address this, the following patch introduces
>>>>>> pte_move_swp_offset(), a function capable of bidirectional movement by
>>>>>> a specified delta argument. Consequently, pte_increment_swp_offset()
>>>>>
>>>>> You mean pte_next_swp_offset()?
>>>>
>>>> yes.
>>>>
>>>>>
>>>>>> will directly invoke it with delta = 1.
>>>>>>
>>>>>> Suggested-by: "Huang, Ying" <ying.huang@xxxxxxxxx>
>>>>>> Signed-off-by: Barry Song <v-songbaohua@xxxxxxxx>
>>>>>> ---
>>>>>>    mm/internal.h | 25 +++++++++++++++++++++----
>>>>>>    1 file changed, 21 insertions(+), 4 deletions(-)
>>>>>>
>>>>>> diff --git a/mm/internal.h b/mm/internal.h
>>>>>> index c5552d35d995..cfe4aed66a5c 100644
>>>>>> --- a/mm/internal.h
>>>>>> +++ b/mm/internal.h
>>>>>> @@ -211,18 +211,21 @@ static inline int folio_pte_batch(struct folio
>>>>>> *folio, unsigned long addr,
>>>>>>    }
>>>>>>
>>>>>>    /**
>>>>>> - * pte_next_swp_offset - Increment the swap entry offset field of a swap
>>>>>> pte.
>>>>>> + * pte_move_swp_offset - Move the swap entry offset field of a swap pte
>>>>>> + *    forward or backward by delta
>>>>>>     * @pte: The initial pte state; is_swap_pte(pte) must be true and
>>>>>>     *    non_swap_entry() must be false.
>>>>>> + * @delta: The direction and the offset we are moving; forward if delta
>>>>>> + *    is positive; backward if delta is negative
>>>>>>     *
>>>>>> - * Increments the swap offset, while maintaining all other fields, including
>>>>>> + * Moves the swap offset, while maintaining all other fields, including
>>>>>>     * swap type, and any swp pte bits. The resulting pte is returned.
>>>>>>     */
>>>>>> -static inline pte_t pte_next_swp_offset(pte_t pte)
>>>>>> +static inline pte_t pte_move_swp_offset(pte_t pte, long delta)
>>>>>
>>>>> We have equivalent functions for pfn:
>>>>>
>>>>>     pte_next_pfn()
>>>>>     pte_advance_pfn()
>>>>>
>>>>> Although the latter takes an unsigned long and only moves forward currently. I
>>>>> wonder if it makes sense to have their naming and semantics match? i.e. change
>>>>> pte_advance_pfn() to pte_move_pfn() and let it move backwards too.
>>>>>
>>>>> I guess we don't have a need for that and it adds more churn.
>>>>
>>>> we might have a need in the below case.
>>>> A forks B, then A and B share large folios. B unmap/exit, then large
>>>> folios of process
>>>> A become single-mapped.
>>>> Right now, while writing A's folios, we are CoWing A's large folios
>>>> into many small
>>>> folios. I believe we can reuse the entire large folios instead of doing
>>>> nr_pages
>>>> CoW and page faults.
>>>> In this case, we might want to get the first PTE from vmf->pte.
>>>
>>> Once we have COW reuse for large folios in place (I think you know that
>>> I am working on that), it might make sense to "COW-reuse around",
>>
>> TBH, I don't know if you are working on that. please Cc me next time :-)
> 
> I could have sworn I mentioned it to you already :)
> 
> See
> 
> https://lore.kernel.org/linux-mm/a9922f58-8129-4f15-b160-e0ace581bcbe@xxxxxxxxxx/T/
> 
> I'll follow-up on that soonish (now that batching is upstream and the large
> mapcount is on its way upstream).
> 
>>
>>> meaning we look if some neighboring PTEs map the same large folio and
>>> map them writable as well. But if it's really worth it, increasing page
>>> fault latency, is to be decided separately.
>>
>> On the other hand, we eliminate latency for the remaining nr_pages - 1 PTEs.
>> Perhaps we can discover a more cost-effective method to signify that a large
>> folio is probably singly mapped?
> 
> Yes, precisely what I am up to!
> 
>> and only call "multi-PTEs" reuse while that
>> condition is true in PF and avoid increasing latency always?
> 
> I'm thinking along those lines:
> 
> If we detect that it's exclusive, we can certainly mapped the current PTE
> writable. Then, we can decide how much (and if) we want to fault-around writable
> as an optimization.
> 
> For smallish large folios, it might make sense to try faulting around most of
> the folio.
> 
> For large large folios (e.g., PTE-mapped 2MiB THP and bigger), we might not want
> to fault around the whole thing -- especially if there is little benefit to be
> had from contig-pte bits.
> 
>>
>>>
>>>
>>>>
>>>> Another case, might be
>>>> A forks B, and we write either A or B, we might CoW an entire large
>>>> folios instead
>>>> CoWing nr_pages small folios.
>>>>
>>>> case 1 seems more useful, I might have a go after some days. then we might
>>>> see pte_move_pfn().
>>> pte_move_pfn() does sound odd to me. 

Yes, I agree the name is odd. pte_move_swp_offset() sounds similarly odd tbh.
Perhaps just pte_advance_swp_offset() with a negative value is clearer about
what its doing?

>>> It might not be required to
>>> implement the optimization described above. (it's easier to simply read
>>> another PTE, check if it maps the same large folio, and to batch from there)

Yes agreed.

>>>
>>
>> It appears that your proposal suggests potential reusability as follows: if we
>> have a large folio containing 16 PTEs, you might consider reusing only 4 by
>> examining PTEs "around" but not necessarily all 16 PTEs. please correct me
>> if my understanding is wrong.
>>
>> Initially, my idea was to obtain the first PTE using pte_move_pfn() and then
>> utilize folio_pte_batch() with the first PTE as arguments to ensure consistency
>> in nr_pages, thus enabling complete reuse of the whole folio.
> 
> Simply doing an vm_normal_folio(pte - X) == folio and then trying to batch from
> there might be easier and cleaner.
>