[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <3d2db7f2-9e3c-929c-de41-c3f0011839b8@gmail.com>
Date: Mon, 9 Jan 2023 13:46:13 +0200
From: Tariq Toukan <ttoukan.linux@...il.com>
To: Jesper Dangaard Brouer <jbrouer@...hat.com>,
"Matthew Wilcox (Oracle)" <willy@...radead.org>,
Jesper Dangaard Brouer <hawk@...nel.org>,
Ilias Apalodimas <ilias.apalodimas@...aro.org>, gal@...dia.com,
Dragos Tatulea <dtatulea@...dia.com>
Cc: brouer@...hat.com, netdev@...r.kernel.org, linux-mm@...ck.org,
Shakeel Butt <shakeelb@...gle.com>,
Saeed Mahameed <saeed@...nel.org>
Subject: Re: [PATCH v2 24/24] mlx5: Convert to netmem
On 06/01/2023 18:31, Jesper Dangaard Brouer wrote:
>
> To Saeed and Tariq, please review.
>
Adding Dragos, Gal.
Hi Jesper,
Thanks for the ping. I'm on it.
> This reminds me, that IMHO we/nvidia/mellanox should remove the local
> mlx5e_page_cache functionality, as SKBs can now recycle page_pool pages.
> This should simplify the driver and we get rid of the head-of-line
> blocking issue with the local page cache (refcnt elevation tricks).
Totally agree.
Dragos is currently working on this task. This should clean up
significant amount of code, and improve performance. We target this for
the next submission window, to kernel v6.4.
> It might look good in microbencmarks, but my experience from prod
> systems are that this local cache isn't utilized. And I believe we
> should be able to get good/similar microbenchmark with page_pool, which
> will continue to recycle and have no HoL issues for prod use-cases.
>
100%.
Thanks,
Tariq
Powered by blists - more mailing lists