ixgbe truesize underestimation fix missing?

ixgbe truesize underestimation fix missing?

[Christian Brunner]

Back in Oktober, there was a patch by Eric Dumazet for the ixgbe
driver to fix the skb truesize underestimation (commit
98130646770db42cd14c44ba0d7f2d0eb8078820). It changed only one line in
ixgbe_main.c:

--- a/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
+++ b/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
@@ -1326,7 +1326,7 @@ static bool ixgbe_clean_rx_irq(struct
ixgbe_q_vector *q_vector,
                        skb->len += upper_len;
                        skb->data_len += upper_len;
-                       skb->truesize += upper_len;
+                       skb->truesize += PAGE_SIZE / 2;
                }                 i++;


But when I look at the current code (3.2.5 or 3.3rc), skb->truesize is
still increased by the same value as skb->data_len.

1228         skb->data_len += frag_list_size;
1229         skb->truesize += frag_list_size;

Is it posible that there was a merge conflict, or am I missing something here?

Thanks,
Christian

Re: ixgbe truesize underestimation fix missing?

[Eric Dumazet]

Le lundi 13 février 2012 à 09:45 +0100, Christian Brunner a écrit :
> Back in Oktober, there was a patch by Eric Dumazet for the ixgbe
> driver to fix the skb truesize underestimation (commit
> 98130646770db42cd14c44ba0d7f2d0eb8078820). It changed only one line in
> ixgbe_main.c:
> 
> --- a/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
> +++ b/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
> @@ -1326,7 +1326,7 @@ static bool ixgbe_clean_rx_irq(struct
> ixgbe_q_vector *q_vector,
>                         skb->len += upper_len;
>                         skb->data_len += upper_len;
> -                       skb->truesize += upper_len;
> +                       skb->truesize += PAGE_SIZE / 2;
>                 }                 i++;
> 
> 
> But when I look at the current code (3.2.5 or 3.3rc), skb->truesize is
> still increased by the same value as skb->data_len.
> 
> 1228         skb->data_len += frag_list_size;
> 1229         skb->truesize += frag_list_size;
> 
> Is it posible that there was a merge conflict, or am I missing something here?

Maybe a problem in your local tree ?

I can see proper fix in current tree.

Re: ixgbe truesize underestimation fix missing?

[Christian Brunner]

2012/2/13 Eric Dumazet <eric.dumazet@gmail.com>:
> Le lundi 13 février 2012 à 09:45 +0100, Christian Brunner a écrit :
>> Back in Oktober, there was a patch by Eric Dumazet for the ixgbe
>> driver to fix the skb truesize underestimation (commit
>> 98130646770db42cd14c44ba0d7f2d0eb8078820). It changed only one line in
>> ixgbe_main.c:
>>
>> --- a/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
>> +++ b/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
>> @@ -1326,7 +1326,7 @@ static bool ixgbe_clean_rx_irq(struct
>> ixgbe_q_vector *q_vector,
>>                         skb->len += upper_len;
>>                         skb->data_len += upper_len;
>> -                       skb->truesize += upper_len;
>> +                       skb->truesize += PAGE_SIZE / 2;
>>                 }                 i++;
>>
>>
>> But when I look at the current code (3.2.5 or 3.3rc), skb->truesize is
>> still increased by the same value as skb->data_len.
>>
>> 1228         skb->data_len += frag_list_size;
>> 1229         skb->truesize += frag_list_size;
>>
>> Is it posible that there was a merge conflict, or am I missing something here?
>
> Maybe a problem in your local tree ?
>
> I can see proper fix in current tree.

Ahh... Sorry. I Was looking at the wrong place -
ixgbe_transform_rsc_queue() has code that is quite similar.

This however makes me wonder, why we are using "skb->len" for
increasing the frag_list_size, as skb->len might be smaller than
skb->truesize.

Shouldn't we add a frag_list_truesize counter in ixgbe_transform_rsc_queue() ?

Thanks,
Christian

Re: ixgbe truesize underestimation fix missing?

[Eric Dumazet]

Le lundi 13 février 2012 à 10:57 +0100, Christian Brunner a écrit :
> 2012/2/13 Eric Dumazet <eric.dumazet@gmail.com>:
> > Le lundi 13 février 2012 à 09:45 +0100, Christian Brunner a écrit :
> >> Back in Oktober, there was a patch by Eric Dumazet for the ixgbe
> >> driver to fix the skb truesize underestimation (commit
> >> 98130646770db42cd14c44ba0d7f2d0eb8078820). It changed only one line in
> >> ixgbe_main.c:
> >>
> >> --- a/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
> >> +++ b/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
> >> @@ -1326,7 +1326,7 @@ static bool ixgbe_clean_rx_irq(struct
> >> ixgbe_q_vector *q_vector,
> >>                         skb->len += upper_len;
> >>                         skb->data_len += upper_len;
> >> -                       skb->truesize += upper_len;
> >> +                       skb->truesize += PAGE_SIZE / 2;
> >>                 }                 i++;
> >>
> >>
> >> But when I look at the current code (3.2.5 or 3.3rc), skb->truesize is
> >> still increased by the same value as skb->data_len.
> >>
> >> 1228         skb->data_len += frag_list_size;
> >> 1229         skb->truesize += frag_list_size;
> >>
> >> Is it posible that there was a merge conflict, or am I missing something here?
> >
> > Maybe a problem in your local tree ?
> >
> > I can see proper fix in current tree.
> 
> Ahh... Sorry. I Was looking at the wrong place -
> ixgbe_transform_rsc_queue() has code that is quite similar.
> 
> This however makes me wonder, why we are using "skb->len" for
> increasing the frag_list_size, as skb->len might be smaller than
> skb->truesize.
> 
> Shouldn't we add a frag_list_truesize counter in ixgbe_transform_rsc_queue() ?

Probably, but you should raise this discussion on netdev, not lkml, and
based on net-next tree, as ixgbe_transform_rsc_queue() doesnt exist
anymore ;)

Note that in this case (RSC) truesize underestimation is small since
typical segment size is very close to PAGE_SIZE/2