diff mbox series

[net-next,v9,07/13] mm: page_frag: some minor refactoring before adding new API

Message ID 20240625135216.47007-8-linyunsheng@huawei.com (mailing list archive)
State New
Headers show
Series [net-next,v9,01/13] mm: page_frag: add a test module for page_frag | expand

Commit Message

Yunsheng Lin June 25, 2024, 1:52 p.m. UTC 
Refactor common codes from __page_frag_alloc_va_align()
to __page_frag_cache_refill(), so that the new API can
make use of them.

CC: Alexander Duyck <alexander.duyck@gmail.com>
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
---
 mm/page_frag_cache.c | 61 ++++++++++++++++++++++----------------------
 1 file changed, 31 insertions(+), 30 deletions(-)

Comments

Alexander Duyck July 2, 2024, 3:30 p.m. UTC | #1 
On Tue, 2024-06-25 at 21:52 +0800, Yunsheng Lin wrote:
> Refactor common codes from __page_frag_alloc_va_align()
> to __page_frag_cache_refill(), so that the new API can
> make use of them.
> 
> CC: Alexander Duyck <alexander.duyck@gmail.com>
> Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>

I am generally not a fan of the concept behind this patch. I really
think we should keep the page_frag_cache_refill function to just
allocating the page, or in this case the encoded_va and populating only
that portion of the struct.

> ---
>  mm/page_frag_cache.c | 61 ++++++++++++++++++++++----------------------
>  1 file changed, 31 insertions(+), 30 deletions(-)
> 
> diff --git a/mm/page_frag_cache.c b/mm/page_frag_cache.c
> index a3316dd50eff..4fd421d4f22c 100644
> --- a/mm/page_frag_cache.c
> +++ b/mm/page_frag_cache.c
> @@ -29,10 +29,36 @@ static void *page_frag_cache_current_va(struct page_frag_cache *nc)
>  static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
>  					     gfp_t gfp_mask)
>  {
> -	struct page *page = NULL;
> +	struct encoded_va *encoded_va = nc->encoded_va;
>  	gfp_t gfp = gfp_mask;
>  	unsigned int order;
> +	struct page *page;
> +
> +	if (unlikely(!encoded_va))
> +		goto alloc;
> +
> +	page = virt_to_page(encoded_va);
> +	if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
> +		goto alloc;
> +
> +	if (unlikely(encoded_page_pfmemalloc(encoded_va))) {
> +		VM_BUG_ON(compound_order(page) !=
> +			  encoded_page_order(encoded_va));
> +		free_unref_page(page, encoded_page_order(encoded_va));
> +		goto alloc;
> +	}
> +
> +	/* OK, page count is 0, we can safely set it */
> +	set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);

Why not just make this block of code a function onto itself? You put an
if statement at the top that essentially is just merging two functions
into one. Perhaps this logic could be __page_frag_cache_recharge which
would return an error if the page is busy or the wrong type. Then
acting on that you could switch to the refill attempt.

Also thinking about it more the set_page_count in this function and
page_ref_add in the other can probably be merged into the recharge and
refill functions since they are acting directly on the encoded page and
not interacting with the other parts of the page_frag_cache.

> +
> +	/* reset page count bias and remaining of new frag */
> +	nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
> +	nc->remaining = page_frag_cache_page_size(encoded_va);

These two parts are more or less agnostic to the setup and could be
applied to refill or recharge. Also one thought occurs to me. You were
encoding "order" into the encoded VA. Why use that when your choices
are either PAGE_FRAG_CACHE_MAX_SIZE or PAGE_SIZE. It should be a single
bit and doesn't need to be a fully byte to store that. That would allow
you to reduce this down to just 2 bits, one for pfmemalloc and one for
max order vs order 0.

> +
> +	return page;
>  
> +alloc:
> +	page = NULL;
>  #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
>  	gfp_mask = (gfp_mask & ~__GFP_DIRECT_RECLAIM) |  __GFP_COMP |
>  		   __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC;
> @@ -89,40 +115,15 @@ void *__page_frag_alloc_va_align(struct page_frag_cache *nc,
>  				 unsigned int fragsz, gfp_t gfp_mask,
>  				 unsigned int align_mask)
>  {
> -	struct encoded_va *encoded_va = nc->encoded_va;
> -	struct page *page;
> -	int remaining;
> +	int remaining = nc->remaining & align_mask;
>  	void *va;
>  
> -	if (unlikely(!encoded_va)) {
> -refill:
> -		if (unlikely(!__page_frag_cache_refill(nc, gfp_mask)))
> -			return NULL;
> -
> -		encoded_va = nc->encoded_va;
> -	}
> -
> -	remaining = nc->remaining & align_mask;
>  	remaining -= fragsz;
>  	if (unlikely(remaining < 0)) {

I see, so this is why you were using the memset calls everywhere.

> -		page = virt_to_page(encoded_va);
> -		if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
> -			goto refill;
> -
> -		if (unlikely(encoded_page_pfmemalloc(encoded_va))) {
> -			VM_BUG_ON(compound_order(page) !=
> -				  encoded_page_order(encoded_va));
> -			free_unref_page(page, encoded_page_order(encoded_va));
> -			goto refill;
> -		}
> -
> -		/* OK, page count is 0, we can safely set it */
> -		set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
> +		if (unlikely(!__page_frag_cache_refill(nc, gfp_mask)))
> +			return NULL;
>  
> -		/* reset page count bias and remaining of new frag */
> -		nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
> -		nc->remaining = remaining = page_frag_cache_page_size(encoded_va);
> -		remaining -= fragsz;
> +		remaining = nc->remaining - fragsz;
>  		if (unlikely(remaining < 0)) {
>  			/*
>  			 * The caller is trying to allocate a fragment
Yunsheng Lin July 3, 2024, 12:36 p.m. UTC | #2 
On 2024/7/2 23:30, Alexander H Duyck wrote:
> On Tue, 2024-06-25 at 21:52 +0800, Yunsheng Lin wrote:
>> Refactor common codes from __page_frag_alloc_va_align()
>> to __page_frag_cache_refill(), so that the new API can
>> make use of them.
>>
>> CC: Alexander Duyck <alexander.duyck@gmail.com>
>> Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
> 
> I am generally not a fan of the concept behind this patch. I really
> think we should keep the page_frag_cache_refill function to just
> allocating the page, or in this case the encoded_va and populating only
> that portion of the struct.

As my understanding, the above mainly depends on how you look at it,
at least it seems odd to me that part of a struct is populated in
one function and other part of the same struct is populated in other
function.

> 
>> ---
>>  mm/page_frag_cache.c | 61 ++++++++++++++++++++++----------------------
>>  1 file changed, 31 insertions(+), 30 deletions(-)
>>
>> diff --git a/mm/page_frag_cache.c b/mm/page_frag_cache.c
>> index a3316dd50eff..4fd421d4f22c 100644
>> --- a/mm/page_frag_cache.c
>> +++ b/mm/page_frag_cache.c
>> @@ -29,10 +29,36 @@ static void *page_frag_cache_current_va(struct page_frag_cache *nc)
>>  static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
>>  					     gfp_t gfp_mask)
>>  {
>> -	struct page *page = NULL;
>> +	struct encoded_va *encoded_va = nc->encoded_va;
>>  	gfp_t gfp = gfp_mask;
>>  	unsigned int order;
>> +	struct page *page;
>> +
>> +	if (unlikely(!encoded_va))
>> +		goto alloc;
>> +
>> +	page = virt_to_page(encoded_va);
>> +	if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
>> +		goto alloc;
>> +
>> +	if (unlikely(encoded_page_pfmemalloc(encoded_va))) {
>> +		VM_BUG_ON(compound_order(page) !=
>> +			  encoded_page_order(encoded_va));
>> +		free_unref_page(page, encoded_page_order(encoded_va));
>> +		goto alloc;
>> +	}
>> +
>> +	/* OK, page count is 0, we can safely set it */
>> +	set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
> 
> Why not just make this block of code a function onto itself? You put an
> if statement at the top that essentially is just merging two functions
> into one. Perhaps this logic could be __page_frag_cache_recharge which
> would return an error if the page is busy or the wrong type. Then
> acting on that you could switch to the refill attempt.
> 
> Also thinking about it more the set_page_count in this function and
> page_ref_add in the other can probably be merged into the recharge and
> refill functions since they are acting directly on the encoded page and
> not interacting with the other parts of the page_frag_cache.

So we are agreed that the below is merged into __page_frag_cache_recharge()?
set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;

The below is merged into __page_frag_cache_refill()?
page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;

> 
>> +
>> +	/* reset page count bias and remaining of new frag */
>> +	nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
>> +	nc->remaining = page_frag_cache_page_size(encoded_va);
> 
> These two parts are more or less agnostic to the setup and could be
> applied to refill or recharge. Also one thought occurs to me. You were
> encoding "order" into the encoded VA. Why use that when your choices
> are either PAGE_FRAG_CACHE_MAX_SIZE or PAGE_SIZE. It should be a single
> bit and doesn't need to be a fully byte to store that. That would allow
> you to reduce this down to just 2 bits, one for pfmemalloc and one for
> max order vs order 0.

I thought about the above and implemented it actually, but it turned out
that it was not as good as encoding "order" into the encoded VA, at least
for the generated asm code size, it didn't seem better. Using one bit,
we need a checking to decide it is PAGE_FRAG_CACHE_MAX_SIZE or PAGE_SIZE.
And we can use that to replace compound_order(page) when calling
free_unref_page() too.

> 
>> +
>> +	return page;
diff mbox series

Patch

diff --git a/mm/page_frag_cache.c b/mm/page_frag_cache.c
index a3316dd50eff..4fd421d4f22c 100644
--- a/mm/page_frag_cache.c
+++ b/mm/page_frag_cache.c
@@ -29,10 +29,36 @@  static void *page_frag_cache_current_va(struct page_frag_cache *nc)
 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
 					     gfp_t gfp_mask)
 {
-	struct page *page = NULL;
+	struct encoded_va *encoded_va = nc->encoded_va;
 	gfp_t gfp = gfp_mask;
 	unsigned int order;
+	struct page *page;
+
+	if (unlikely(!encoded_va))
+		goto alloc;
+
+	page = virt_to_page(encoded_va);
+	if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
+		goto alloc;
+
+	if (unlikely(encoded_page_pfmemalloc(encoded_va))) {
+		VM_BUG_ON(compound_order(page) !=
+			  encoded_page_order(encoded_va));
+		free_unref_page(page, encoded_page_order(encoded_va));
+		goto alloc;
+	}
+
+	/* OK, page count is 0, we can safely set it */
+	set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
+
+	/* reset page count bias and remaining of new frag */
+	nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
+	nc->remaining = page_frag_cache_page_size(encoded_va);
+
+	return page;
 
+alloc:
+	page = NULL;
 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
 	gfp_mask = (gfp_mask & ~__GFP_DIRECT_RECLAIM) |  __GFP_COMP |
 		   __GFP_NOWARN | __GFP_NORETRY | __GFP_NOMEMALLOC;
@@ -89,40 +115,15 @@  void *__page_frag_alloc_va_align(struct page_frag_cache *nc,
 				 unsigned int fragsz, gfp_t gfp_mask,
 				 unsigned int align_mask)
 {
-	struct encoded_va *encoded_va = nc->encoded_va;
-	struct page *page;
-	int remaining;
+	int remaining = nc->remaining & align_mask;
 	void *va;
 
-	if (unlikely(!encoded_va)) {
-refill:
-		if (unlikely(!__page_frag_cache_refill(nc, gfp_mask)))
-			return NULL;
-
-		encoded_va = nc->encoded_va;
-	}
-
-	remaining = nc->remaining & align_mask;
 	remaining -= fragsz;
 	if (unlikely(remaining < 0)) {
-		page = virt_to_page(encoded_va);
-		if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
-			goto refill;
-
-		if (unlikely(encoded_page_pfmemalloc(encoded_va))) {
-			VM_BUG_ON(compound_order(page) !=
-				  encoded_page_order(encoded_va));
-			free_unref_page(page, encoded_page_order(encoded_va));
-			goto refill;
-		}
-
-		/* OK, page count is 0, we can safely set it */
-		set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
+		if (unlikely(!__page_frag_cache_refill(nc, gfp_mask)))
+			return NULL;
 
-		/* reset page count bias and remaining of new frag */
-		nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
-		nc->remaining = remaining = page_frag_cache_page_size(encoded_va);
-		remaining -= fragsz;
+		remaining = nc->remaining - fragsz;
 		if (unlikely(remaining < 0)) {
 			/*
 			 * The caller is trying to allocate a fragment