[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20250307085946.244ddeb6@kernel.org>
Date: Fri, 7 Mar 2025 08:59:46 -0800
From: Jakub Kicinski <kuba@...nel.org>
To: Sebastian Andrzej Siewior <bigeasy@...utronix.de>
Cc: linux-rdma@...r.kernel.org, netdev@...r.kernel.org, "David S. Miller"
<davem@...emloft.net>, Andrew Lunn <andrew+netdev@...n.ch>, Eric Dumazet
<edumazet@...gle.com>, Ilias Apalodimas <ilias.apalodimas@...aro.org>,
Jesper Dangaard Brouer <hawk@...nel.org>, Joe Damato <jdamato@...tly.com>,
Leon Romanovsky <leon@...nel.org>, Paolo Abeni <pabeni@...hat.com>, Saeed
Mahameed <saeedm@...dia.com>, Simon Horman <horms@...nel.org>, Tariq Toukan
<tariqt@...dia.com>, Thomas Gleixner <tglx@...utronix.de>, Yunsheng Lin
<linyunsheng@...wei.com>
Subject: Re: [PATCH net-next v2 2/5] page_pool: Add per-queue statistics.
On Fri, 7 Mar 2025 17:50:46 +0100 Sebastian Andrzej Siewior wrote:
> On 2025-03-07 08:11:35 [-0800], Jakub Kicinski wrote:
> > On Fri, 7 Mar 2025 12:57:19 +0100 Sebastian Andrzej Siewior wrote:
> > > The mlx5 driver supports per-channel statistics. To make support generic
> > > it is required to have a template to fill the individual channel/ queue.
> > >
> > > Provide page_pool_ethtool_stats_get_strings_mq() to fill the strings for
> > > multiple queue.
> >
> > Sorry to say this is useless as a common helper, you should move it
> > to mlx5.
> >
> > The page pool stats have a standard interface, they are exposed over
> > netlink. If my grep-foo isn't failing me no driver uses the exact
> > strings mlx5 uses. "New drivers" are not supposed to add these stats
> > to ethtool -S, and should just steer users towards the netlink stats.
> >
> > IOW mlx5 is and will remain the only user of this helper forever.
>
> Okay, so per-runqueue stats is not something other/ new drivers are
> interested in?
> The strings are the same, except for the rx%d_ prefix, but yes this
> makes it unique.
> The mlx5 folks seem to be the only one interested in this. The veth
> driver for instance iterates over real_num_rx_queues and adds all
> per-queue stats into one counter. It could also expose this per-runqueue
> as it does with xdp_packets for instance. But then it uses the
> rx_queue_%d prefix…
What I'm saying is they are already available per queue, via netlink,
with no driver work necessary. See tools/net/ynl/samples/page-pool.c
The mlx5 stats predate the standard interface.
> I don't care, I just intended to provide some generic facility so we
> don't have every driver rolling its own thing. I have no problem to move
> this to the mlx5 driver.
Thanks, and sorry for not catching the conversation earlier.
Powered by blists - more mailing lists