__alloc_pages() failures reported due to fragmentation

__alloc_pages() failures reported due to fragmentation

[Larry Woodman]

We have seen a couple of __alloc_pages() failures due to fragmentation, 
there is plenty
of free memory but no large order pages available.  I think the problem 
is in
sock_alloc_send_pskb(), the gfp_mask includes __GFP_REPEAT but its never
used/passed to the page allocator.  Shouldnt the gfp_mask be passed to 
alloc_skb() ?

-------------------------------------------------------------------
struct sk_buff *sock_alloc_send_pskb()
{
       unsigned int gfp_mask;
... 
                                                                          
       gfp_mask = sk->sk_allocation;
       if (gfp_mask & __GFP_WAIT)
               gfp_mask |= __GFP_REPEAT;
...
       skb = alloc_skb(header_len, sk->sk_allocation);
--------------------------------------------------------------------


--- linux-2.6.18.noarch/net/core/sock.c.orig
+++ linux-2.6.18.noarch/net/core/sock.c
@@ -1154,7 +1154,7 @@ static struct sk_buff *sock_alloc_send_p
 			goto failure;
 
 		if (atomic_read(&sk->sk_wmem_alloc) < sk->sk_sndbuf) {
-			skb = alloc_skb(header_len, sk->sk_allocation);
+			skb = alloc_skb(header_len, gfp_mask);
 			if (skb) {
 				int npages;
 				int i;

Re: __alloc_pages() failures reported due to fragmentation

[Arjan van de Ven]

> --- linux-2.6.18.noarch/net/core/sock.c.orig
> +++ linux-2.6.18.noarch/net/core/sock.c
> @@ -1154,7 +1154,7 @@ static struct sk_buff *sock_alloc_send_p
>  			goto failure;
>  
>  		if (atomic_read(&sk->sk_wmem_alloc) < sk->sk_sndbuf) {
> -			skb = alloc_skb(header_len, sk->sk_allocation);
> +			skb = alloc_skb(header_len, gfp_mask);
>  			if (skb) {
>  				int npages;
>  				int i;

Hi,

this is not actually right though... sk_allocation is very possible to
have a restricting mask compared to the one passed in (say "no highmem"
or even GFP_DMA) and you now discard this... probably better would be to
calculate a set of "transient" flags that you then or into the
sk_allocation mask at this time...

Greetings,
   Arjan van de Ven

-- 
if you want to mail me at work (you don't), use arjan (at) linux.intel.com
Test the interaction between Linux and your BIOS via http://www.linuxfirmwarekit.org

Re: __alloc_pages() failures reported due to fragmentation

[Larry Woodman]

Arjan van de Ven wrote:

>>--- linux-2.6.18.noarch/net/core/sock.c.orig
>>+++ linux-2.6.18.noarch/net/core/sock.c
>>@@ -1154,7 +1154,7 @@ static struct sk_buff *sock_alloc_send_p
>> 			goto failure;
>> 
>> 		if (atomic_read(&sk->sk_wmem_alloc) < sk->sk_sndbuf) {
>>-			skb = alloc_skb(header_len, sk->sk_allocation);
>>+			skb = alloc_skb(header_len, gfp_mask);
>> 			if (skb) {
>> 				int npages;
>> 				int i;
>>    
>>
>
>Hi,
>
>this is not actually right though... sk_allocation is very possible to
>have a restricting mask compared to the one passed in (say "no highmem"
>or even GFP_DMA) and you now discard this... probably better would be to
>calculate a set of "transient" flags that you then or into the
>sk_allocation mask at this time...
>
>Greetings,
>   Arjan van de Ven
>
>  
>
Hi Arjan.  Right but this just includes __GFP_REPEAT in the mask so we can
defrag in __alloc_pages and only if GFP_WAIT was passed in origionally.

Larry

Re: __alloc_pages() failures reported due to fragmentation

[David Miller]

From: Larry Woodman <lwoodman@redhat.com>
Date: Fri, 03 Nov 2006 07:39:44 -0500

> Hi Arjan.  Right but this just includes __GFP_REPEAT in the mask so we can
> defrag in __alloc_pages and only if GFP_WAIT was passed in origionally.

Indeed, quoting that small snippet of the patch was deceptive :-)

Arjan, gfp_mask is set to sk->sk_allocation, and we just add a bit
into it conditionally.