Re: [PATCH v2 1/4] mm/page_isolation: make page isolation a standalone bit.

[Zi Yan <ziy@nvidia.com>]

On 14 Feb 2025, at 11:45, Johannes Weiner wrote:

> On Fri, Feb 14, 2025 at 10:42:12AM -0500, Zi Yan wrote:
>> During page isolation, the original migratetype is overwritten, since
>> MIGRATE_* are enums. Change MIGRATE_ISOLATE to be a standalone bit like
>> PB_migrate_skip. pageblock bits needs to be word aligned, so expand
>> the number of pageblock bits from 4 to 8 and make migrate isolate bit 7.
>>
>> Signed-off-by: Zi Yan <ziy@nvidia.com>
>> ---
>>  include/linux/mmzone.h          | 18 +++++++++++++-----
>>  include/linux/page-isolation.h  |  2 +-
>>  include/linux/pageblock-flags.h | 33 ++++++++++++++++++++++++++++++++-
>>  mm/page_alloc.c                 | 21 +++++++++++++++++++--
>>  4 files changed, 65 insertions(+), 9 deletions(-)
>>
>> diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
>> index 8aecbbb0b685..3c7d3f22ccb2 100644
>> --- a/include/linux/mmzone.h
>> +++ b/include/linux/mmzone.h
>> @@ -106,14 +106,22 @@ static inline bool migratetype_is_mergeable(int mt)
>>
>>  extern int page_group_by_mobility_disabled;
>>
>> -#define MIGRATETYPE_MASK ((1UL << PB_migratetype_bits) - 1)
>> +#ifdef CONFIG_MEMORY_ISOLATION
>> +#define MIGRATETYPE_NO_ISO_MASK (BIT(PB_migratetype_bits) - 1)
>> +#define MIGRATETYPE_MASK (MIGRATETYPE_NO_ISO_MASK | PB_migrate_isolate_bit)
>> +#else
>> +#define MIGRATETYPE_NO_ISO_MASK (BIT(PB_migratetype_bits) - 1)
>> +#define MIGRATETYPE_MASK (BIT(PB_migratetype_bits) - 1)
>> +#endif
>
> There is no user of the NO_ISO_MASK until the last patch. Can you
> please defer introduction until then?

Sure.

>
>> -#define get_pageblock_migratetype(page)					\
>> -	get_pfnblock_flags_mask(page, page_to_pfn(page), MIGRATETYPE_MASK)
>> +#define get_pageblock_migratetype(page) \
>> +		get_pfnblock_flags_mask(page, page_to_pfn(page), \
>> +			MIGRATETYPE_MASK)
>>
>> -#define folio_migratetype(folio)				\
>> -	get_pfnblock_flags_mask(&folio->page, folio_pfn(folio),		\
>> +#define folio_migratetype(folio) \
>> +		get_pfnblock_flags_mask(&folio->page, folio_pfn(folio), \
>>  			MIGRATETYPE_MASK)
>
> That's a spurious change currently. I assume you tweaked the
> MIGRATETYPE_MASK parameter during development, but I can't see a
> functional difference now.

Right, will remove it.
>
>> @@ -373,7 +374,13 @@ unsigned long get_pfnblock_flags_mask(const struct page *page,
>>  	 * racy, are not corrupted.
>>  	 */
>>  	word = READ_ONCE(bitmap[word_bitidx]);
>> -	return (word >> bitidx) & mask;
>> +	flags = (word >> bitidx) & mask;
>> +
>> +#ifdef CONFIG_MEMORY_ISOLATION
>> +	if (flags & PB_migrate_isolate_bit)
>> +		return MIGRATE_ISOLATE;
>> +#endif
>> +	return flags;
>>  }
>>
>>  static __always_inline int get_pfnblock_migratetype(const struct page *page,
>> @@ -397,8 +404,18 @@ void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
>>  	unsigned long bitidx, word_bitidx;
>>  	unsigned long word;
>>
>> +#ifdef CONFIG_MEMORY_ISOLATION
>> +	BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 8);
>> +	/* keep other migratetype bits if MIGRATE_ISOLATE is set */
>> +	if (flags == MIGRATE_ISOLATE) {
>> +		mask &= ~((1UL << PB_migratetype_bits) - 1);
>> +		flags = PB_migrate_isolate_bit;
>> +	}
>
> Please change the callers in both cases to pass the appropriate masks
> of interest instead.
>
> That's likely a bit of churn in the allocator code, but adding caller
> specifics to this function violates abstraction layering rules.

Basically, expose this to the caller. Namely, if it is setting
MIGRATE_ISOLATE, it should use the MIGRATETYPE_ISO_ONLY_MASK. Otherwise,
use MIGRATETYPE_MASK. Maybe adds two helper functions for
setting pageblock isolation and clearing pageblock isolation, like
{set,clear}_pageblock_skip()? I thought about it and did not do that
due to code churn, but now you ask for it explicitly, I am going to
do that.

Thank you for the review.

Best Regards,
Yan, Zi