[PATCH] arm64: mm: decrease the section size to reduce the memory reserved for the page map

[Anshuman Khandual]

On 12/7/20 7:10 AM, Song Bao Hua (Barry Song) wrote:
> 
> 
>> -----Original Message-----
>> From: Mike Rapoport [mailto:rppt at linux.ibm.com]
>> Sent: Saturday, December 5, 2020 12:44 AM
>> To: Will Deacon <will at kernel.org>
>> Cc: liwei (CM) <liwei213 at huawei.com>; catalin.marinas at arm.com; fengbaopeng
>> <fengbaopeng2 at hisilicon.com>; nsaenzjulienne at suse.de; steve.capper at arm.com;
>> Song Bao Hua (Barry Song) <song.bao.hua at hisilicon.com>;
>> linux-arm-kernel at lists.infradead.org; linux-kernel at vger.kernel.org; butao
>> <butao at hisilicon.com>
>> Subject: Re: [PATCH] arm64: mm: decrease the section size to reduce the memory
>> reserved for the page map
>>
>> On Fri, Dec 04, 2020 at 11:13:47AM +0000, Will Deacon wrote:
>>> On Fri, Dec 04, 2020 at 09:44:43AM +0800, Wei Li wrote:
>>>> For the memory hole, sparse memory model that define SPARSEMEM_VMEMMAP
>>>> do not free the reserved memory for the page map, decrease the section
>>>> size can reduce the waste of reserved memory.
>>>>
>>>> Signed-off-by: Wei Li <liwei213 at huawei.com>
>>>> Signed-off-by: Baopeng Feng <fengbaopeng2 at hisilicon.com>
>>>> Signed-off-by: Xia Qing <saberlily.xia at hisilicon.com>
>>>> ---
>>>>  arch/arm64/include/asm/sparsemem.h | 2 +-
>>>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>>>
>>>> diff --git a/arch/arm64/include/asm/sparsemem.h
>> b/arch/arm64/include/asm/sparsemem.h
>>>> index 1f43fcc79738..8963bd3def28 100644
>>>> --- a/arch/arm64/include/asm/sparsemem.h
>>>> +++ b/arch/arm64/include/asm/sparsemem.h
>>>> @@ -7,7 +7,7 @@
>>>>
>>>>  #ifdef CONFIG_SPARSEMEM
>>>>  #define MAX_PHYSMEM_BITS	CONFIG_ARM64_PA_BITS
>>>> -#define SECTION_SIZE_BITS	30
>>>> +#define SECTION_SIZE_BITS	27
>>>
>>> We chose '30' to avoid running out of bits in the page flags. What changed?
>>
>> I think that for 64-bit there are still plenty of free bits. I didn't
>> check now, but when I played with SPARSEMEM on m68k there were 8 bits
>> for section out of 32.
>>
>>> With this patch, I can trigger:
>>>
>>> ./include/linux/mmzone.h:1170:2: error: Allocator MAX_ORDER exceeds
>> SECTION_SIZE
>>> #error Allocator MAX_ORDER exceeds SECTION_SIZE
>>>
>>> if I bump up NR_CPUS and NODES_SHIFT.
>>
>> I don't think it's related to NR_CPUS and NODES_SHIFT.
>> This seems rather 64K pages that cause this.
>>
>> Not that is shouldn't be addressed.
> 
> Right now, only 4K PAGES will define ARM64_SWAPPER_USES_SECTION_MAPS.
> Other cases will use vmemmap_populate_basepages().
> The original patch should be only addressing the issue in 4K pages:
> https://lore.kernel.org/lkml/20200812010655.96339-1-liwei213@huawei.com/
> 
> would we do something like the below?
> #ifdef CONFIG_ARM64_4K_PAGE
> #define SECTION_SIZE_BITS	27
> #else
> #define SECTION_SIZE_BITS	30
> #endif

This is bit arbitrary. Probably 27 can be further reduced for 4K page size.
Instead, we should make SECTION_SIZE_BITS explicitly depend upon MAX_ORDER.
IOW section size should be the same as the highest order page in the buddy.
CONFIG_FORCE_MAX_ZONEORDER is always defined on arm64. A quick test shows
SECTION_SIZE_BITS would be 22 on 4K pages and 29 for 64K pages. As a fall
back SECTION_SIZE_BITS can still be 30 in case CONFIG_FORCE_MAX_ZONEORDER
is not defined.

--- a/arch/arm64/include/asm/sparsemem.h
+++ b/arch/arm64/include/asm/sparsemem.h
@@ -7,7 +7,7 @@
 
 #ifdef CONFIG_SPARSEMEM
 #define MAX_PHYSMEM_BITS       CONFIG_ARM64_PA_BITS
-#define SECTION_SIZE_BITS      30
+#define SECTION_SIZE_BITS      (CONFIG_FORCE_MAX_ZONEORDER - 1 + PAGE_SHIFT)
 #endif
 
 #endif

A similar approach exists on ia64 platform as well.