Use of contiguous physical memory in cxgbe driver

[Adrian Chadd]

On 13 February 2014 10:15, Garrett Wollman <wollman at bimajority.org> wrote:
> <<On Thu, 13 Feb 2014 10:07:36 -0800, John-Mark Gurney <jmg at funkthat.com> said:
>
>>> [I wrote:]
>>> It's very easy to get to that state if the driver insists on getting
>>> three physically contiguous pages (which is what it takes to allocate
>>> a single 9k cluster)
>
>> Well, if you're using a cheap NIC that can't do scatter/gather DMA,
>> then you get what you pay for...
>
> No, I'm using expensive 10G NICs, like the Chelsio ones mentioned in
> the subject header.  They have no trouble doing s/g (subject to some
> reasonable limits on chain length).  The problem is when the drivers
> insist on physical contiguity anyway.
>
> I finally gave up and hacked the cxgbe driver never to use anything
> longer than a physical page.

Have you tried it at 40G? ISTR np@ commenting on there being benefits
to larger (up to 64K) allocations for the 40G T5 NICs that do packet
-> buffer packing on receive side.


-a