[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20240906183844.2e8226f3@kernel.org>
Date: Fri, 6 Sep 2024 18:38:44 -0700
From: Jakub Kicinski <kuba@...nel.org>
To: Taehee Yoo <ap420073@...il.com>
Cc: davem@...emloft.net, pabeni@...hat.com, edumazet@...gle.com,
michael.chan@...adcom.com, netdev@...r.kernel.org
Subject: Re: [PATCH net-next 0/2] bnxt_en: implement tcp-data-split ethtool
command
On Fri, 6 Sep 2024 08:07:48 +0000 Taehee Yoo wrote:
> The approach of this patch is to support the bnxt_en driver setting up
> enable/disable HDS explicitly, not rely on LRO/GRO, JUMBO.
> In addition, hds_threshold no longer follows rx-copybreak.
> By this patch, hds_threshold always be 0.
That may make sense for zero-copy use cases, where you want to make
sure that all of the data lands in target page pool. But in general
using the data buffers may waste quite a bit of memory, and PCIe bus
bandwidth (two small transfers instead of one medium size).
I think we should add a user-controlled setting in ethtool -g for
hds-threshold.
Also please make sure you describe the level of testing you have done
in the commit message. I remember discussing this a few years back
and at that time HDS was tied to GRO for bnxt at the FW level.
A lot has changed since but please describe what you tested..
--
pw-bot: cr
Powered by blists - more mailing lists