Re: [PATCH] mm: Introduce GFP_PGTABLE

[Anshuman Khandual]

On 01/12/2019 05:42 PM, Matthew Wilcox wrote:
> On Sat, Jan 12, 2019 at 03:56:38PM +0530, Anshuman Khandual wrote:
>> All architectures have been defining their own PGALLOC_GFP as (GFP_KERNEL |
>> __GFP_ZERO) and using it for allocating page table pages.
> 
> Except that's not true.
> 
>> +++ b/arch/x86/mm/pgtable.c
>> @@ -13,19 +13,17 @@ phys_addr_t physical_mask __ro_after_init = (1ULL << __PHYSICAL_MASK_SHIFT) - 1;
>>  EXPORT_SYMBOL(physical_mask);
>>  #endif
>>  
>> -#define PGALLOC_GFP (GFP_KERNEL_ACCOUNT | __GFP_ZERO)
>> -
>>  #ifdef CONFIG_HIGHPTE
> 
> ...
> 
>>  pte_t *pte_alloc_one_kernel(struct mm_struct *mm)
>>  {
>> -	return (pte_t *)__get_free_page(PGALLOC_GFP & ~__GFP_ACCOUNT);
>> +	return (pte_t *)__get_free_page(GFP_PGTABLE & ~__GFP_ACCOUNT);
>>  }
> 
> I think x86 was the only odd one out here, but you'll need to try again ...

IIUC the user page table pages need __GFP_ACCOUNT not the kernel ones. Hence
in the above function it clears out __GFP_ACCOUNT for kernel page table page
allocations but where as by default it has got __GFP_ACCOUNT which would be
used for user page tables. Instead we can make X86 user allocations add
__GFP_ACCOUNT (like other archs) to generic GFP_PGTABLE when ever required.