[PATCH V2] powerpc/mm: Initialize kernel pagetable memory for PTE fragments

[PATCH V2] powerpc/mm: Initialize kernel pagetable memory for PTE fragments

[Anshuman Khandual]

Kernel pagetable pages for PTE fragments never go through the standard init
sequence which can cause inaccuracies in utilization statistics reported at
places like /proc and /sysfs interfaces etc. Also the allocated page misses
out on pagetable lock and page flag init as well. Fix it by making sure all
pages allocated for either user process or kernel PTE fragments go through
same initialization.

Signed-off-by: Anshuman Khandual <khandual@linux.vnet.ibm.com>
---
Changes in V2:

- Call the destructor function during free for all cases

 arch/powerpc/mm/pgtable-book3s64.c | 27 ++++++++++++---------------
 1 file changed, 12 insertions(+), 15 deletions(-)

diff --git a/arch/powerpc/mm/pgtable-book3s64.c b/arch/powerpc/mm/pgtable-book3s64.c
index c1f4ca4..a820ee6 100644
--- a/arch/powerpc/mm/pgtable-book3s64.c
+++ b/arch/powerpc/mm/pgtable-book3s64.c
@@ -335,23 +335,21 @@ static pte_t *get_pte_from_cache(struct mm_struct *mm)
 
 static pte_t *__alloc_for_ptecache(struct mm_struct *mm, int kernel)
 {
+	gfp_t gfp_mask = PGALLOC_GFP;
 	void *ret = NULL;
 	struct page *page;
 
-	if (!kernel) {
-		page = alloc_page(PGALLOC_GFP | __GFP_ACCOUNT);
-		if (!page)
-			return NULL;
-		if (!pgtable_page_ctor(page)) {
-			__free_page(page);
-			return NULL;
-		}
-	} else {
-		page = alloc_page(PGALLOC_GFP);
-		if (!page)
-			return NULL;
-	}
+	if (!kernel)
+		gfp_mask |= __GFP_ACCOUNT;
 
+	page = alloc_page(gfp_mask);
+	if (!page)
+		return NULL;
+
+	if (!pgtable_page_ctor(page)) {
+		__free_page(page);
+		return NULL;
+	}
 
 	ret = page_address(page);
 	/*
@@ -391,8 +389,7 @@ void pte_fragment_free(unsigned long *table, int kernel)
 	struct page *page = virt_to_page(table);
 
 	if (put_page_testzero(page)) {
-		if (!kernel)
-			pgtable_page_dtor(page);
+		pgtable_page_dtor(page);
 		free_unref_page(page);
 	}
 }
-- 
1.8.3.1

Re: [PATCH V2] powerpc/mm: Initialize kernel pagetable memory for PTE fragments

[Aneesh Kumar K.V]

On 06/20/2018 02:52 PM, Anshuman Khandual wrote:
> Kernel pagetable pages for PTE fragments never go through the standard init
> sequence which can cause inaccuracies in utilization statistics reported at
> places like /proc and /sysfs interfaces etc. Also the allocated page misses
> out on pagetable lock and page flag init as well. Fix it by making sure all
> pages allocated for either user process or kernel PTE fragments go through
> same initialization.
> 
> Signed-off-by: Anshuman Khandual <khandual@linux.vnet.ibm.com>
> ---
> Changes in V2:
> 
> - Call the destructor function during free for all cases
> 
>   arch/powerpc/mm/pgtable-book3s64.c | 27 ++++++++++++---------------
>   1 file changed, 12 insertions(+), 15 deletions(-)
> 
> diff --git a/arch/powerpc/mm/pgtable-book3s64.c b/arch/powerpc/mm/pgtable-book3s64.c
> index c1f4ca4..a820ee6 100644
> --- a/arch/powerpc/mm/pgtable-book3s64.c
> +++ b/arch/powerpc/mm/pgtable-book3s64.c
> @@ -335,23 +335,21 @@ static pte_t *get_pte_from_cache(struct mm_struct *mm)
>   
>   static pte_t *__alloc_for_ptecache(struct mm_struct *mm, int kernel)
>   {
> +	gfp_t gfp_mask = PGALLOC_GFP;
>   	void *ret = NULL;
>   	struct page *page;
>   
> -	if (!kernel) {
> -		page = alloc_page(PGALLOC_GFP | __GFP_ACCOUNT);
> -		if (!page)
> -			return NULL;
> -		if (!pgtable_page_ctor(page)) {
> -			__free_page(page);
> -			return NULL;
> -		}
> -	} else {
> -		page = alloc_page(PGALLOC_GFP);
> -		if (!page)
> -			return NULL;
> -	}
> +	if (!kernel)
> +		gfp_mask |= __GFP_ACCOUNT;
>   
> +	page = alloc_page(gfp_mask);
> +	if (!page)
> +		return NULL;
> +
> +	if (!pgtable_page_ctor(page)) {
> +		__free_page(page);
> +		return NULL;
> +	}
>   
>   	ret = page_address(page);
>   	/*
> @@ -391,8 +389,7 @@ void pte_fragment_free(unsigned long *table, int kernel)
>   	struct page *page = virt_to_page(table);
>   
>   	if (put_page_testzero(page)) {
> -		if (!kernel)
> -			pgtable_page_dtor(page);
> +		pgtable_page_dtor(page);
>   		free_unref_page(page);
>   	}
>   }
> 

If we really are going to do this, may be we can  do it as 
alloc_for_pmdcache, kill the kernel arg and compare against init_mm?

-aneesh