Re: [EXT] Re: [PATCH v1] net: mvpp2: divide fifo for dts-active ports only - Russell King

From: Russell King - ARM Linux admin <linux@armlinux.org.uk>
To: Stefan Chulski <stefanc@marvell.com>
Cc: "netdev@vger.kernel.org" <netdev@vger.kernel.org>,
	"thomas.petazzoni@bootlin.com" <thomas.petazzoni@bootlin.com>,
	"davem@davemloft.net" <davem@davemloft.net>,
	Nadav Haklai <nadavh@marvell.com>,
	Yan Markman <ymarkman@marvell.com>,
	"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
	"kuba@kernel.org" <kuba@kernel.org>,
	"mw@semihalf.com" <mw@semihalf.com>,
	"andrew@lunn.ch" <andrew@lunn.ch>
Subject: Re: [EXT] Re: [PATCH v1] net: mvpp2: divide fifo for dts-active ports only
Date: Mon, 23 Nov 2020 15:51:48 +0000	[thread overview]
Message-ID: <20201123155148.GX1551@shell.armlinux.org.uk> (raw)
In-Reply-To: <CO6PR18MB3873B4205ECAF2383F9539CCB0FC0@CO6PR18MB3873.namprd18.prod.outlook.com>

On Mon, Nov 23, 2020 at 03:44:05PM +0000, Stefan Chulski wrote:
> > > > On Mon, Nov 23, 2020 at 04:52:40PM +0200, stefanc@marvell.com wrote:
> > > > > From: Stefan Chulski <stefanc@marvell.com>
> > > > >
> > > > > Tx/Rx FIFO is a HW resource limited by total size, but shared by
> > > > > all ports of same CP110 and impacting port-performance.
> > > > > Do not divide the FIFO for ports which are not enabled in DTS, so
> > > > > active ports could have more FIFO.
> > > > >
> > > > > The active port mapping should be done in probe before FIFO-init.
> > > >
> > > > It would be nice to know what the effect is from this - is it a
> > > > small or large boost in performance?
> > >
> > > I didn't saw any significant improvement with LINUX bridge or
> > > forwarding, but this reduced PPv2 overruns drops, reduced CRC sent errors
> > with DPDK user space application.
> > > So this improved zero loss throughput. Probably with XDP we would see a
> > similar effect.
> > >
> > > > What is the effect when the ports on a CP110 are configured for 10G,
> > > > 1G, and 2.5G in that order, as is the case on the Macchiatobin board?
> > >
> > > Macchiatobin has two CP's.  CP1 has 3 ports, so the distribution of FIFO would
> > be the same as before.
> > > On CP0 which has a single port, all FIFO would be allocated for 10G port.
> > 
> > Your code allocates for CP1:
> > 
> > 32K to port 0 (the 10G port on Macchiatobin) 8K to port 1 (the 1G dedicated
> > ethernet port on Macchiatobin) 4K to port 2 (the 1G/2.5G SFP port on
> > Macchiatobin)
> > 
> > I'm questioning that allocation for port 1 and 2.
> 
> Yes, but this allocation exists also in current code.
> From HW point of view(MAC and PPv2) maximum supported speed
> in CP110: port 0 - 10G, port 1 - 2.5G, port 2 - 2.5G.
> in CP115: port 0 - 10G, port 1 - 5G, port 2 - 2.5G.
> 
> So this allocation looks correct at least for CP115.
> Problem that we cannot reallocate FIFO during runtime, after specific speed negotiation.

We could do much better. DT has a "max-speed" property for ethernet
controllers. If we have that property, then I think we should use
that to determine the initialisation time FIFO allocation.

As I say, on Macchiatobin, the allocations we end up with are just
crazy when you consider the port speeds that the hardware supports.
Maybe that should be done as a follow-on patch - but I think it
needs to be done.

-- 
RMK's Patch system: https://www.armlinux.org.uk/developer/patches/
FTTP is here! 40Mbps down 10Mbps up. Decent connectivity at last!