RE: [PATCH net-next v5] cadence: Add LSO support.

From: Rafal Ozieblo <rafalo@cadence.com>
To: Eric Dumazet <eric.dumazet@gmail.com>
Cc: David Miller <davem@davemloft.net>,
	"nicolas.ferre@atmel.com" <nicolas.ferre@atmel.com>,
	"netdev@vger.kernel.org" <netdev@vger.kernel.org>,
	"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>
Subject: RE: [PATCH net-next v5] cadence: Add LSO support.
Date: Tue, 15 Nov 2016 14:20:31 +0000	[thread overview]
Message-ID: <BN3PR07MB2516126E781EB2301E4E9325C9BF0@BN3PR07MB2516.namprd07.prod.outlook.com> (raw)
In-Reply-To: <1479215490.8455.122.camel@edumazet-glaptop3.roam.corp.google.com>

> From: Eric Dumazet [mailto:eric.dumazet@gmail.com] 
> Sent: 15 listopada 2016 14:12
> To: Rafal Ozieblo
> Cc: David Miller; nicolas.ferre@atmel.com; netdev@vger.kernel.org; linux-kernel@vger.kernel.org
> Subject: Re: [PATCH net-next v5] cadence: Add LSO support.
>
> On Tue, 2016-11-15 at 07:07 +0000, Rafal Ozieblo wrote:
> > > > > If UFO is in use it should not silently disable UDP checksums.
> > > > > 
> > > > > If you cannot support UFO with proper checksumming, then you cannot enable support for that feature.
> > > > 
> > > > According Cadence Gigabit Ethernet MAC documentation:
> > > > 
> > > > "Hardware will not calculate the UDP checksum or modify the UDP 
> > > > checksum field. Therefore software must set a value of zero in the 
> > > > checksum field in the UDP header (in the first payload buffer) to indicate to the receiver that the UDP datagram does not include a checksum."
> > > > 
> > > > It is hardware requirement.
> > >
> > > I do not doubt that it is a hardware restriction.
> > >
> > > But I am saying that you cannot enable this feature under Linux if this is how it operates on your hardware.
> > 
> > Would it be good to enable UFO conditionally with some internal define? Ex.:
> > 
> > +#ifdef MACB_ENABLE_UFO
> > +#define MACB_NETIF_LSO         (NETIF_F_TSO | NETIF_F_UFO)
> > +#else
> > +#define MACB_NETIF_LSO         (NETIF_F_TSO)
> > +#endif
> > 
> > I could add precise comment here that ufo is possible only without checksum.
> > 
> > Or maybe I could enable it from module_params or device-tree (like: drivers/net/ethernet/neterion/s2io.c).
>
> No you can not do that.
>
> 1) That would violate UDP specs.
> 2) Module params are no longer accepted.
> 3) Comments in a driver source code would only help the driver maintainer, not users to make their mind.
>
> Only way would be to propagate the intent of the sender.
>
> Only the sender application can decide to generate UDP checksums or not.
>
> Your driver ndo_features_check() could then force software segmentation fallback if the user did not asked to disable UDP checksums, and packet is UFO.
>
> (look for UDP_NO_CHECK6_TX, and SO_NO_CHECK )
>
> Problem is complex, because the skb has no marker, only the socket has.
>
> And socket state could change between packets, and packets can stay in an intermediate qdisc before hitting device driver. So looking at
> skb->sk from your ndo_features_check() would be racy.
>
> What use case would you have precisely ?

I have talked with hardware team who designed and created gem IP.  The conclusion is that there is no need to zeroed checksum for UFO. 
I have tested UFO without zeroing UDP checksum on cadence gem. It works fine.
I'll deeply investigate and test UFO again. I'll send version 6 of LSO PATCH either with UFO without zeroing or without UFO at all.