Message-ID: <41dd848c-c27b-4373-9e89-3fda9e302cfb@linux.dev>
Date: Mon, 6 Oct 2025 23:02:42 +0800
From: Lance Yang <lance.yang@...ux.dev>
To: David Hildenbrand <david@...hat.com>, Wei Yang <richard.weiyang@...il.com>
Cc: lorenzo.stoakes@...cle.com, Liam.Howlett@...cle.com, baohua@...nel.org,
 baolin.wang@...ux.alibaba.com, dev.jain@....com, hughd@...gle.com,
 ioworker0@...il.com, kirill@...temov.name, linux-kernel@...r.kernel.org,
 linux-mm@...ck.org, mpenttil@...hat.com, npache@...hat.com,
 ryan.roberts@....com, ziy@...dia.com, akpm@...ux-foundation.org
Subject: Re: [PATCH mm-new v2 1/1] mm/khugepaged: abort collapse scan on
 non-swap entries



On 2025/10/6 22:18, David Hildenbrand wrote:
> On 05.10.25 04:12, Lance Yang wrote:
>>
>>
>> On 2025/10/5 09:05, Wei Yang wrote:
>>> On Wed, Oct 01, 2025 at 06:05:57PM +0800, Lance Yang wrote:
>>>>
>>>>
>>>> On 2025/10/1 16:54, Wei Yang wrote:
>>>>> On Wed, Oct 01, 2025 at 11:22:51AM +0800, Lance Yang wrote:
>>>>>> From: Lance Yang <lance.yang@...ux.dev>
>>>>>>
>>>>>> Currently, special non-swap entries (like migration, hwpoison, or PTE
>>>>>> markers) are not caught early in hpage_collapse_scan_pmd(), 
>>>>>> leading to
>>>>>> failures deep in the swap-in logic.
>>>>>>
>>>>>> hpage_collapse_scan_pmd()
>>>>>> `- collapse_huge_page()
>>>>>>        `- __collapse_huge_page_swapin() -> fails!
>>>>>>
>>>>>> As David suggested[1], this patch skips any such non-swap entries
>>>>>> early. If any one is found, the scan is aborted immediately with the
>>>>>> SCAN_PTE_NON_PRESENT result, as Lorenzo suggested[2], avoiding wasted
>>>>>> work.
>>>>>>
>>>>>> [1] https://lore.kernel.org/linux-mm/7840f68e-7580-42cb- 
>>>>>> a7c8-1ba64fd6df69@...hat.com
>>>>>> [2] https://lore.kernel.org/linux-mm/7df49fe7-c6b7-426a-8680- 
>>>>>> dcd55219c8bd@...ifer.local
>>>>>>
>>>>>> Suggested-by: David Hildenbrand <david@...hat.com>
>>>>>> Suggested-by: Lorenzo Stoakes <lorenzo.stoakes@...cle.com>
>>>>>> Signed-off-by: Lance Yang <lance.yang@...ux.dev>
>>>>>> ---
>>>>>> v1 -> v2:
>>>>>> - Skip all non-present entries except swap entries (per David) 
>>>>>> thanks!
>>>>>> - https://lore.kernel.org/linux-mm/20250924100207.28332-1- 
>>>>>> lance.yang@...ux.dev/
>>>>>>
>>>>>> mm/khugepaged.c | 32 ++++++++++++++++++--------------
>>>>>> 1 file changed, 18 insertions(+), 14 deletions(-)
>>>>>>
>>>>>> diff --git a/mm/khugepaged.c b/mm/khugepaged.c
>>>>>> index 7ab2d1a42df3..d0957648db19 100644
>>>>>> --- a/mm/khugepaged.c
>>>>>> +++ b/mm/khugepaged.c
>>>>>> @@ -1284,7 +1284,23 @@ static int hpage_collapse_scan_pmd(struct 
>>>>>> mm_struct *mm,
>>>>>>     for (addr = start_addr, _pte = pte; _pte < pte + HPAGE_PMD_NR;
>>>>>>          _pte++, addr += PAGE_SIZE) {
>>>>>>         pte_t pteval = ptep_get(_pte);
>>>>>> -        if (is_swap_pte(pteval)) {
>>>>>
>>>>> It looks is_swap_pte() is mis-leading?
>>>>
>>>> Hmm.. not to me, IMO. is_swap_pte() just means:
>>>>
>>>> !pte_none(pte) && !pte_present(pte)
>>>>
>>>
>>> Maybe it has some reason.
>>>
>>> I took another look into __collapse_huge_page_swapin(), which just check
>>> is_swap_pte() before do_swap_page().
> 
> Thanks for pointing that out.
> 
> A function that is called __collapse_huge_page_swapin() and documented 
> to "Bring missing pages in from swap" will handle other types as well.
> 
> Unbelievable horrible.
> 
> So let's think this through so we can document it in the changelog 
> properly.
> 
> We could have currently ended up in do_swap_page() with
> 
> (1) Migration entries. We would have waited.
> 
> -> Maybe worth it to wait, maybe not. I suspect we don't stumble into
>     that frequently such that we don't care. We could always unlock this
>     separately later.
> 
> 
> (2) Device-exclusive entries. We would have converted to non-exclusive.
> 
> -> See make_device_exclusive(), we cannot tolerate PMD entries and have
>     to split them through FOLL_SPLIT_PMD. As popped up during a recent
>     discussion, collapsing here is actually counter-productive, because
>     the next conversion will PTE-map it again. (until recently, it would
>     not have worked with large folios at all IIRC).
> 
> -> Ok to not collapse.
> 
> (3) Device-private entries. We would have migrated to RAM.
> 
> -> Device-private still does not support THPs, so collapsing right now 
> just means that the next device access would split the folio again.
> 
> -> Ok to not collapse.
> 
> (4) HWPoison entries
> 
> -> Cannot collapse
> 
> (5) Markers
> 
> -> Cannot collapse
> 
> 
> I suggest we add that in some form to the patch description, stating 
> that we can unlock later what we really need, and not account it towards 
> max_swap_ptes.

Cool!

I'll take a closer look and adjust the patch description accordingly ;)

Thanks a lot for the lesson!

> 
>>>
>>> We have filtered non-swap entries in hpage_collapse_scan_pmd(), but 
>>> we drop
>>> mmap lock before isolation. This looks we may have a chance to get 
>>> non-swap
>>> entry.
>>
>> Thanks for pointing that out!
>>
>> Yep, there is a theoretical window between dropping the mmap lock
>> after the initial scan and re-acquiring it for isolation.
>>
>>>
>>> Do you think it is reasonable to add a non_swap_entry() check before
>>> do_swap_page()?
>>
>> However, that seems unlikely in practice. IMHO, the early check in
>> hpage_collapse_scan_pmd() is sufficient for now, so I'd prefer to
>> keep it as-is :)
> 
> I think we really should add that check, as per reasoning above.
> 
> I was looking into some possible races with uffd-wp being set before we 
> enter do_swap_page(), but I think it might be okay (although very 
> confusing).

Ah, I see ;p

@Wei could you send a patch to add the non_swap_entry() check there?

Powered by blists - more mailing lists