Re: [PATCH v8 04/11] net/mlx4_en: add support for fast rx drop bpf program

[Brenden Blanco <bblanco@plumgrid.com>]

On Wed, Jul 13, 2016 at 11:27:23AM +0000, David Laight wrote:
> From: Brenden Blanco
> > Sent: 12 July 2016 08:51
> > Add support for the BPF_PROG_TYPE_XDP hook in mlx4 driver.
> > 
> > In tc/socket bpf programs, helpers linearize skb fragments as needed
> > when the program touches the packet data. However, in the pursuit of
> > speed, XDP programs will not be allowed to use these slower functions,
> > especially if it involves allocating an skb.
> > 
> > Therefore, disallow MTU settings that would produce a multi-fragment
> > packet that XDP programs would fail to access. Future enhancements could
> > be done to increase the allowable MTU.
> 
> Maybe I'm misunderstanding what is going on here...
> But what has the MTU to do with how skb are fragmented?
This is mlx4 specific...depending on the MTU the driver will write data
into 1536, 1536+4096, 1536+4096+4096, etc. fragments.
> 
> If the skb come from a reasonably written USB ethernet interface they could
> easily have arbitrary fragment boundaries (the frames get packed into USB
> buffers).
The XDP program is operating directly on the packet memory, before any
skb has been allocated. The program also expects a continguous memory
region to inspect...it's too expensive to linearize the data like we do
in the tc hook case, that's a feature that costs too much for this type
of low level feature. Therefore, XDP can only be turned on in
combination with a cooperative driver, that's the performance tradeoff
we're imposing here.
> 
> Outbound skb can also have fragments depending on how they are generated.
Sure, but XDP won't run on those. This is an rx-only feature.
> 
> 	David