Re: [PATCH] khugepaged: Reduce race probability between migration and khugepaged

[Dev Jain <dev.jain@arm.com>]

On 30/06/25 1:49 pm, David Hildenbrand wrote:
> On 30.06.25 10:12, Dev Jain wrote:
>>
>> On 30/06/25 1:25 pm, Anshuman Khandual wrote:
>>> On 30/06/25 10:18 AM, Dev Jain wrote:
>>>> Suppose a folio is under migration, and khugepaged is also trying to
>>>> collapse it. collapse_pte_mapped_thp() will retrieve the folio from 
>>>> the
>>>> page cache via filemap_lock_folio(), thus taking a reference on the 
>>>> folio
>>>> and sleeping on the folio lock, since the lock is held by the 
>>>> migration
>>>> path. Migration will then fail in
>>>> __folio_migrate_mapping -> folio_ref_freeze. Reduce the probability of
>>>> such a race happening (leading to migration failure) by bailing out
>>>> if we detect a PMD is marked with a migration entry.
>>> Could the migration be re-attempted after such failure ? Seems like
>>> the migration failure here is traded for a scan failure instead.
>>
>> We already re-attempt migration. See NR_MAX_MIGRATE_PAGES_RETRY and
>> NR_MAX_MIGRATE_ASYNC_RETRY. Also just before freezing the refcount,
>> we do a suitable refcount check in folio_migrate_mapping(). So the
>> race happens after this and folio_ref_freeze() in 
>> __folio_migrate_mapping(),
>> therefore the window for the race is already very small in the migration
>> path, but large in the khugepaged path.
>>
>>>
>>>> This fixes the migration-shared-anon-thp testcase failure on Apple M3.
>>> Could you please provide some more context why this test case was
>>> failing earlier and how does this change here fixes the problem ?
>>
>> IMHO the explanation I have given in the patch description is clear
>> and succinct: the testcase is failing due to the race. This patch
>> shortens the race window, and the test on this particular hardware
>> does not hit the race window again.
>>
>>>
>>>> Note that, this is not a "fix" since it only reduces the chance of
>>>> interference of khugepaged with migration, wherein both the kernel
>>>> functionalities are deemed "best-effort".
>>>>> Signed-off-by: Dev Jain <dev.jain@arm.com>
>>>> ---
>>>>
>>>> This patch was part of
>>>> https://lore.kernel.org/all/20250625055806.82645-1-dev.jain@arm.com/
>>>> but I have sent it separately on suggestion of Lorenzo, and also 
>>>> because
>>>> I plan to send the first two patches after David Hildenbrand's
>>>> folio_pte_batch series gets merged.
>>>>
>>>>    mm/khugepaged.c | 12 ++++++++++--
>>>>    1 file changed, 10 insertions(+), 2 deletions(-)
>>>>
>>>> diff --git a/mm/khugepaged.c b/mm/khugepaged.c
>>>> index 1aa7ca67c756..99977bb9bf6a 100644
>>>> --- a/mm/khugepaged.c
>>>> +++ b/mm/khugepaged.c
>>>> @@ -31,6 +31,7 @@ enum scan_result {
>>>>        SCAN_FAIL,
>>>>        SCAN_SUCCEED,
>>>>        SCAN_PMD_NULL,
>>>> +    SCAN_PMD_MIGRATION,
>>>>        SCAN_PMD_NONE,
>>>>        SCAN_PMD_MAPPED,
>>>>        SCAN_EXCEED_NONE_PTE,
>>>> @@ -941,6 +942,8 @@ static inline int check_pmd_state(pmd_t *pmd)
>>>>           if (pmd_none(pmde))
>>>>            return SCAN_PMD_NONE;
>>>> +    if (is_pmd_migration_entry(pmde))
>>>> +        return SCAN_PMD_MIGRATION;
>>>>        if (!pmd_present(pmde))
>>>>            return SCAN_PMD_NULL;
>>>>        if (pmd_trans_huge(pmde))
>>>> @@ -1502,9 +1505,12 @@ int collapse_pte_mapped_thp(struct mm_struct 
>>>> *mm, unsigned long addr,
>>>>            !range_in_vma(vma, haddr, haddr + HPAGE_PMD_SIZE))
>>>>            return SCAN_VMA_CHECK;
>>>>    -    /* Fast check before locking page if already PMD-mapped */
>>>> +    /*
>>>> +     * Fast check before locking folio if already PMD-mapped, or 
>>>> if the
>>>> +     * folio is under migration
>>>> +     */
>>>>        result = find_pmd_or_thp_or_none(mm, haddr, &pmd);
>>>> -    if (result == SCAN_PMD_MAPPED)
>>>> +    if (result == SCAN_PMD_MAPPED || result == SCAN_PMD_MIGRATION)
>>> Should mapped PMD and migrating PMD be treated equally while scanning ?
>>
>> SCAN_PMD_MAPPED is used as an indicator to change result to SCAN_SUCCEED
>> in khugepaged_scan_mm_slot: after the call to 
>> collapse_pte_mapped_thp. And,
>> it is also used in madvise_collapse() to do ++thps which is used to 
>> set the
>> return value of madvise_collapse. So I think this approach will be 
>> wrong.
>
> But if it already is PMD mapped (just temporarily through a migration 
> entry), isn't this exactly what we want?

Good point. I was about to say that what about PMD-folio splitting 
during migration, but then

during unmapping the folios, if we cannot migrate the PMD-folios, they 
will be splitted via

unmap_folio() along with the PMD, therefore we can be sure that if we 
encounter a PMD

migration entry, then eventually it will be converted to a PMD leaf 
entry on migration success

or failure.


I'll merge this into SCAN_PMD_MAPPED, thanks.