Re: [PATCH V1 net-next 00/12] Improving performance and reducing latencies, by using latest capabilities exposed in ENA device

[Jesper Dangaard Brouer <brouer@redhat.com>]

On Tue, 9 Oct 2018 21:18:30 +0000 "Bshara, Saeed" <saeedb@amazon.com> wrote:

> Currently the driver allocate page per rx buffer, but we are
> considering to support mode where page split to 2 buffers in order to
> overcome memory fragmentation issue on low memory systems. but, this
> won't work with XDP, right?
> what's your advice?

XDP is easiest to implement with 1-page per rx buffer, but given ixgbe
and i40e violated this, it is possible to use a page-split approach
with 2 frames per page, like they do. Do notice that you have to
deviate from the standard 256 bytes headroom for that to fit (which
basically killed my idea of placing the SKB in this headroom).

-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer

 
 
> From: Bshara, Nafea
> Sent: Tuesday, October 9, 2018 10:33 PM
> To: Jesper Dangaard Brouer; Kiyanovski, Arthur
> Cc: davem@davemloft.net; netdev@vger.kernel.org; Woodhouse, David; Machulsky, Zorik; Matushevsky, Alexander; Bshara, Saeed; Wilson, Matt; Liguori, Anthony; Tzalik, Guy; Belgazal, Netanel; Saidi, Ali; Björn Töpel
> Subject: Re: [PATCH V1 net-next 00/12] Improving performance and reducing latencies, by using latest capabilities exposed in ENA device
>     
> It is high priority for us right after this major release get merged.
> 
> On 10/9/18, 12:31 PM, "Jesper Dangaard Brouer" <brouer@redhat.com> wrote:
> 
>     
>     On Tue, 9 Oct 2018 21:44:57 +0300 <akiyano@amazon.com> wrote:
>     
>     > From: Arthur Kiyanovski <akiyano@amazon.com>
>     > 
>     > This patchset introduces the following:
>     > 1. A new placement policy of Tx headers and descriptors, which takes
>     > advantage of an option to place headers + descriptors in device memory
>     > space. This is sometimes referred to as LLQ - low latency queue.
>     > The patch set defines the admin capability, maps the device memory as
>     > write-combined, and adds a mode in transmit datapath to do header +
>     > descriptor placement on the device.
>     > 2. Support for RX checksum offloading
>     > 3. Miscelaneous small improvements and code cleanups
>     
>     What are your plans for XDP?
>     
>     You are unsure ask your-colleague David Woodhouse, who I've discussed
>     this with when he attended my talk at Kernel-Recipes[1], slide[2].
>     
>     [1]  https://kernel-recipes.org/en/2018/talks/xdp-a-new-programmable-network-layer/
>     [2]  http://people.netfilter.org/hawk/presentations/KernelRecipes2018/XDP_Kernel_Recipes_2018.pdf
>     -- 
>     Best regards,
>       Jesper Dangaard Brouer
>       MSc.CS, Principal Kernel Engineer at Red Hat
>       LinkedIn: http://www.linkedin.com/in/brouer
>     
> 
>