Re: [RFC] allow skb->head to point/alias to first skb frag

["Maciej Żenczykowski" <maze@google.com>]

it would be very useful if there was an api for freeing skb->head memory...

I've fooled around with a single memory buffer for both the head and
the data and seen huge performance wins under heavy packet load (half
the amount of malloc/free), but could never quite get it to be 100%
stable.

On Thu, Apr 26, 2012 at 2:22 AM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
> On Thu, 2012-04-26 at 05:18 -0400, David Miller wrote:
>> From: Eric Dumazet <eric.dumazet@gmail.com>
>> Date: Thu, 26 Apr 2012 11:10:02 +0200
>>
>> > The 'frag' would have a known size : 2048 bytes
>> >
>> > But the end of it would be used by struct skb_shared_info
>> >
>> > so data_len would be 0 in fact.
>> >
>> > This would look like a regular linear skb.
>> >
>> > Just a bit set in skb to say : Warning, skb->head was not kmalloced :
>> > replace kfree(head) by put_page(...)
>> >
>> > And this bit would be tested in GRO or tcp merge to 'upgrade' this
>> > skb->head to proper page/frag
>>
>> And what happens if this ends up in a piece of code which wants to
>> append a page frag?
>>
>> Or a piece of code which copies an SKB, including page frag parts?
>>
>> All I'm saying is that the number of tests necessary to make this work
>> properly might become prohibitive.
>
> I'll cook a patch, I believe not so many parts assume skb->head can be
> kmalloc()/kfree(). This should be contained in net/core/skbuff.c only.
>
> Elsewhere we use skb->head as a pointer to memory. It will stay the
> same.
>
>
>



-- 
Maciej A. Żenczykowski
Kernel Networking Developer @ Google
1600 Amphitheatre Parkway, Mountain View, CA 94043
tel: +1 (650) 253-0062