[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20231211112332.2abc94ae@kernel.org>
Date: Mon, 11 Dec 2023 11:23:32 -0800
From: Jakub Kicinski <kuba@...nel.org>
To: Alexander Lobakin <aleksander.lobakin@...el.com>
Cc: Yunsheng Lin <linyunsheng@...wei.com>, "David S. Miller"
<davem@...emloft.net>, Eric Dumazet <edumazet@...gle.com>, Paolo Abeni
<pabeni@...hat.com>, Maciej Fijalkowski <maciej.fijalkowski@...el.com>,
Michal Kubiak <michal.kubiak@...el.com>, Larysa Zaremba
<larysa.zaremba@...el.com>, Alexander Duyck <alexanderduyck@...com>, "David
Christensen" <drc@...ux.vnet.ibm.com>, Jesper Dangaard Brouer
<hawk@...nel.org>, Ilias Apalodimas <ilias.apalodimas@...aro.org>, "Paul
Menzel" <pmenzel@...gen.mpg.de>, <netdev@...r.kernel.org>,
<intel-wired-lan@...ts.osuosl.org>, <linux-kernel@...r.kernel.org>, Amritha
Nambiar <amritha.nambiar@...el.com>
Subject: Re: [PATCH net-next v6 08/12] libie: add Rx buffer management (via
Page Pool)
On Mon, 11 Dec 2023 11:16:20 +0100 Alexander Lobakin wrote:
> Ideally, I'd like to pass a CPU ID this queue will be run on and use
> cpu_to_node(), but currently there's no NUMA-aware allocations in the
> Intel drivers and Rx queues don't get the corresponding CPU ID when
> configuring. I may revisit this later, but for now NUMA_NO_NODE is the
> most optimal solution here.
Hm, I've been wondering about persistent IRQ mappings. Drivers
resetting IRQ mapping on reconfiguration is a major PITA in production
clusters. You change the RSS hash and some NICs suddenly forget
affinitization 🤯️
The connection with memory allocations changes the math on that a bit.
The question is really whether we add CPU <> NAPI config as a netdev
Netlink API or build around the generic IRQ affinity API. The latter
is definitely better from "don't duplicate uAPI" perspective.
But we need to reset the queues and reallocate their state when
the mapping is changed. And shutting down queues on
echo $cpu > /../smp_affinity_list
seems moderately insane. Perhaps some middle-ground exists.
Anyway, if you do find cycles to tackle this - pls try to do it
generically not just for Intel? :)
Powered by blists - more mailing lists