Re: [PATCH net-next 2/3] net: qualcomm: rmnet: add tx packets aggregation

From: Jakub Kicinski <kuba@kernel.org>
To: "Subash Abhinov Kasiviswanathan (KS)" <quic_subashab@quicinc.com>
Cc: "Alexander Lobakin" <alexandr.lobakin@intel.com>,
	"Daniele Palmas" <dnlplm@gmail.com>,
	"David Miller" <davem@davemloft.net>,
	"Paolo Abeni" <pabeni@redhat.com>,
	"Eric Dumazet" <edumazet@google.com>,
	"Sean Tranchetti" <quic_stranche@quicinc.com>,
	"Jonathan Corbet" <corbet@lwn.net>, "Bjørn Mork" <bjorn@mork.no>,
	"Greg Kroah-Hartman" <gregkh@linuxfoundation.org>,
	netdev@vger.kernel.org
Subject: Re: [PATCH net-next 2/3] net: qualcomm: rmnet: add tx packets aggregation
Date: Fri, 11 Nov 2022 09:11:41 -0800	[thread overview]
Message-ID: <20221111091141.784c88f3@kernel.org> (raw)
In-Reply-To: <b84e45e0-55e0-a1f5-e1cc-980983946019@quicinc.com>

On Thu, 10 Nov 2022 18:17:09 -0700 Subash Abhinov Kasiviswanathan (KS)
wrote:
> The difference here is that hardware would use a single descriptor for 
> aggregation vs multiple descriptors for scatter gather.
> 
> I wonder if this issue is related to pacing though.
> Daniele, perhaps you can try this hack without enabling EGRESS 
> AGGREGATION and check if you are able to reach the same level of 
> performance for your scenario.
> 
> --- a/drivers/net/ethernet/qualcomm/rmnet/rmnet_handlers.c
> +++ b/drivers/net/ethernet/qualcomm/rmnet/rmnet_handlers.c
> @@ -236,7 +236,7 @@ void rmnet_egress_handler(struct sk_buff *skb)
>          struct rmnet_priv *priv;
>          u8 mux_id;
> 
> -       sk_pacing_shift_update(skb->sk, 8);
> +       skb_orphan(skb);
> 
>          orig_dev = skb->dev;
>          priv = netdev_priv(orig_dev);

The pacing shift update is much cleaner than orphaning packets, IMHO.
And the aggregation should hold onto the original skbs (a list?) and
free them once the aggregate was transmitted.