[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <1415041448.31570.0.camel@edumazet-glaptop2.roam.corp.google.com>
Date: Mon, 03 Nov 2014 11:04:08 -0800
From: Eric Dumazet <eric.dumazet@...il.com>
To: David Laight <David.Laight@...LAB.COM>
Cc: Ezequiel Garcia <ezequiel.garcia@...e-electrons.com>,
"netdev@...r.kernel.org" <netdev@...r.kernel.org>,
David Miller <davem@...emloft.net>,
Thomas Petazzoni <thomas.petazzoni@...e-electrons.com>,
Gregory Clement <gregory.clement@...e-electrons.com>,
Tawfik Bayouk <tawfik@...vell.com>,
Lior Amsalem <alior@...vell.com>,
Nadav Haklai <nadavh@...vell.com>
Subject: Re: [PATCH 0/1] mv643xx_eth: Disable TSO by default
On Mon, 2014-11-03 at 14:51 +0000, David Laight wrote:
> From: Eric Dumazet
> > On Sat, 2014-11-01 at 12:30 -0300, Ezequiel Garcia wrote:
> > > Several users ([1], [2]) have been reporting data corruption with TSO on
> > > Kirkwood platforms (i.e. using the mv643xx_eth driver).
> > >
> > > Until we manage to find what's causing this, this simple patch will make
> > > the TSO path disabled by default. This patch should be queued for stable,
> > > fixing the TSO feature introduced in v3.16.
> > >
> > > The corruption itself is very easy to reproduce: checking md5sum on a mounted
> > > NFS directory gives a different result each time. Same tests using the mvneta
> > > driver (Armada 370/38x/XP SoC) pass with no issues.
> > >
> > > Frankly, I'm a bit puzzled about this, and so any ideas or debugging hints
> > > are well received.
> >
> > lack of barriers maybe ?
> >
> > It seems you might need to populate all TX descriptors but delay the
> > first, like doing the populate in descending order.
> >
> > If you take a look at txq_submit_skb(), you'll see the final
> > desc->cmd_sts = cmd_sts (line 959) is done _after_ frags were cooked by
> > txq_submit_frag_skb()
> >
> > You should kick the nick only when all TX descriptors are ready and
> > committed to memory.
>
> Don't forget that the nick might process the first descriptor without
> being given a 'kick' - it will read it when it finishes processing the
> previous frame.
> This also means that you have to be careful about the order of the writes
> to the first descriptor.
This is what I implied and implemented in the patch ;)
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Powered by blists - more mailing lists