RE: [PATCH v3 net-next 3/6] net: Add SW fallback infrastructure for offloaded sockets

From: Ilya Lesokhin <ilyal@mellanox.com>
To: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Cc: "netdev@vger.kernel.org" <netdev@vger.kernel.org>,
	"davem@davemloft.net" <davem@davemloft.net>,
	"davejwatson@fb.com" <davejwatson@fb.com>,
	"tom@herbertland.com" <tom@herbertland.com>,
	"hannes@stressinduktion.org" <hannes@stressinduktion.org>,
	Boris Pismenny <borisp@mellanox.com>,
	"Aviad Yehezkel" <aviadye@mellanox.com>,
	Liran Liss <liranl@mellanox.com>,
	"Steffen Klassert" <steffen.klassert@secunet.com>
Subject: RE: [PATCH v3 net-next 3/6] net: Add SW fallback infrastructure for offloaded sockets
Date: Tue, 19 Dec 2017 07:51:02 +0000	[thread overview]
Message-ID: <AM4PR0501MB2723B5452037B69253A0459BD40F0@AM4PR0501MB2723.eurprd05.prod.outlook.com> (raw)
In-Reply-To: <20171218191826.GA6122@localhost.localdomain>

On Monday, December 18, 2017 9:18 PM, Marcelo Ricardo Leitner wrote:

> > +
> > +	if (sk && sk_fullsock(sk) && sk->sk_offload_check)
> 
> Isn't this going to hurt the fast path, checking for sk fields here?
> 

We do add code to the fast path but it seems unavoidable if you want to have SW fallback.
The XFRM device offload also does that
http://elixir.free-electrons.com/linux/v4.14.7/source/net/core/dev.c#L3058

The check can be optimized but I didn't want to do that before I saw that it's an issue.
I'm also not sure what the correct solution is.
I don't like that fact that each "stateful protocol" we offload requires its own check. 
We need to think if we can find a generic way of doing it.

Perhaps we can hold the expected netdev somewhere in the SKB and only if we don't
Go out of the expected netdev go to a slow path that does a check for each protocol.

Thanks,
Ilya