Re: [PATCH v1] memblock: Initialized the memory of memblock.reserve to the MIGRATE_MOVABL

[Mike Rapoport <rppt@kernel.org>]

On Wed, Sep 25, 2024 at 07:02:35PM +0800, suhua wrote:
> After sparse_init function requests memory for struct page in memblock and
> adds it to memblock.reserved, this memory area is present in both
> memblock.memory and memblock.reserved.
> 
> When CONFIG_DEFERRED_STRUCT_PAGE_INIT is not set. The memmap_init function
> is called during the initialization of the free area of the zone, this
> function calls for_each_mem_pfn_range to initialize all memblock.memory,
> excluding memory that is also placed in memblock.reserved, such as the
> struct page metadata that describes the page, 1TB memory is about 16GB,
> and generally this part of reserved memory occupies more than 90% of the
> total reserved memory of the system. So all memory in memblock.memory is
> set to MIGRATE_MOVABLE according to the alignment of pageblock_nr_pages.
> For example, if hugetlb_optimize_vmemmap=1, huge pages are allocated, the
> freed pages are placed on buddy's MIGRATE_MOVABL list for use.

Please make sure you spell MIGRATE_MOVABLE and MIGRATE_UNMOVABLE correctly.
 
> When CONFIG_DEFERRED_STRUCT_PAGE_INIT=y, only the first_deferred_pfn range
> is initialized in memmap_init. The subsequent free_low_memory_core_early
> initializes all memblock.reserved memory but not MIGRATE_MOVABL. All
> memblock.memory is set to MIGRATE_MOVABL when it is placed in buddy via
> free_low_memory_core_early and deferred_init_memmap. As a result, when
> hugetlb_optimize_vmemmap=1 and huge pages are allocated, the freed pages
> will be placed on buddy's MIGRATE_UNMOVABL list (For example, on machines
> with 1TB of memory, alloc 2MB huge page size of 1000GB frees up about 15GB
> to MIGRATE_UNMOVABL). Since the huge page alloc requires a MIGRATE_MOVABL
> page, a fallback is performed to alloc memory from MIGRATE_UNMOVABL for
> MIGRATE_MOVABL.
> 
> Large amount of UNMOVABL memory is not conducive to defragmentation, so
> the reserved memory is also set to MIGRATE_MOVABLE in the
> free_low_memory_core_early phase following the alignment of
> pageblock_nr_pages.
> 
> Eg：
> echo 500000 > /proc/sys/vm/nr_hugepages
> cat /proc/pagetypeinfo
> 
> before：
> Free pages count per migrate type at order       0      1      2      3      4      5      6      7      8      9     10
> …
> Node    0, zone   Normal, type    Unmovable     51      2      1     28     53     35     35     43     40     69   3852
> Node    0, zone   Normal, type      Movable   6485   4610    666    202    200    185    208     87     54      2    240
> Node    0, zone   Normal, type  Reclaimable      2      2      1     23     13      1      2      1      0      1      0
> Node    0, zone   Normal, type   HighAtomic      0      0      0      0      0      0      0      0      0      0      0
> Node    0, zone   Normal, type      Isolate      0      0      0      0      0      0      0      0      0      0      0
> Unmovable ≈ 15GB
> 
> after：
> Free pages count per migrate type at order       0      1      2      3      4      5      6      7      8      9     10
> …
> Node    0, zone   Normal, type    Unmovable      0      1      1      0      0      0      0      1      1      1      0
> Node    0, zone   Normal, type      Movable   1563   4107   1119    189    256    368    286    132    109      4   3841
> Node    0, zone   Normal, type  Reclaimable      2      2      1     23     13      1      2      1      0      1      0
> Node    0, zone   Normal, type   HighAtomic      0      0      0      0      0      0      0      0      0      0      0
> Node    0, zone   Normal, type      Isolate      0      0      0      0      0      0      0      0      0      0      0
> 
> Signed-off-by: suhua <suhua1@kingsoft.com>
> ---
>  mm/mm_init.c | 6 ++++++
>  1 file changed, 6 insertions(+)
> 
> diff --git a/mm/mm_init.c b/mm/mm_init.c
> index 4ba5607aaf19..e0190e3f8f26 100644
> --- a/mm/mm_init.c
> +++ b/mm/mm_init.c
> @@ -722,6 +722,12 @@ static void __meminit init_reserved_page(unsigned long pfn, int nid)
>  		if (zone_spans_pfn(zone, pfn))
>  			break;
>  	}
> +
> +	if (pageblock_aligned(pfn)) {
> +		set_pageblock_migratetype(pfn_to_page(pfn), MIGRATE_MOVABLE);
> +		cond_resched();
> +	}

If you are trying to make initialization of pageblock migrate type
consistent with or without CONFIG_DEFERRED_STRUCT_PAGE_INIT, move setting
of migrate type from deferred_free_pages() to deferred_init_pages().

> +
>  	__init_single_page(pfn_to_page(pfn), pfn, zid, nid);
>  }
>  #else
> -- 
> 2.34.1
> 

-- 
Sincerely yours,
Mike.