Re: [PATCH 1/2] stmmac: introduce flag to dynamically disable TX offload for rockchip devices

[Robin Murphy]

On 2019-04-02 12:53 pm, Jose Abreu wrote:
> From: Robin Murphy <robin.murphy@xxxxxxx>
> Date: Tue, Apr 02, 2019 at 12:49:36
>
> > On 02/04/2019 08:59, Jose Abreu wrote:
> > > From: Philipp Tomsich <philipp.tomsich@xxxxxxxxxxxxxxxxxxxxx>
> > > Date: Mon, Apr 01, 2019 at 20:12:21
> > >
> > > > + Christoph.
> > > >
> > > > > On 01.04.2019, at 21:06, Heiko StÃbner <heiko@xxxxxxxxx> wrote:
> > > > >
> > > > > Am Montag, 1. April 2019, 20:54:45 CEST schrieb Robin Murphy:
> > > > > > On 01/04/2019 19:18, Leonidas P. Papadakos wrote:
> > > > > > > From: =?UTF-8?q?Kamil=20Trzci=C5=84ski?= <ayufan@xxxxxxxxx>
> > > > > > >
> > > > > > > Some rockchip boards exhibit an issue where tx checksumming does not
> > work
> > > > with
> > > > > > > packets larger than 1498.
> > > > > >
> > > > > > Is it really a board-level problem? I'm no networking expert, but the
> > > > > > nature of the workaround suggests this is more likely to be some
> > > > > > inherent limitation of the IP block in the SoC, rather than something to
> > > > > > do with how the external pins get wired up. Does anyone have an RK3328
> > > > > > or RK3399 board that provably *does* checksum large packets correctly?
> > > > >
> > > > > I don't have that many rk3399-boards with actual ethernet and even only
> > > > > the rock64 from rk3328-land, but at least my rk3399-firefly also seems
> > > > > affected by this.
> > > > >
> > > > > But so far the rk3399-puma board from Theobroma did not show that
> > ethernet
> > > > > issue for me, so I've added two Theobroma people who may or may not tell
> > > > > if they've also seen that issue.
> > > > >
> > > > > > > This is bad for network stability.
> > > > > > >
> > > > > > > The previous approach was using force_thresh_dma_mode in the board dts,
> > > > which
> > > > > > > does more than we need.
> > > > > >
> > > > > > If indeed it is a SoC-level thing (or at least we want to treat it as
> > > > > > such), then couldn't we just hang it off the existing SoC-specific
> > > > > > compatibles in dwmac-rk.c and avoid the need for a new DT property at
> > > > > > all? After all, that's precisely why SoC-specific compatibles are a
> > > > > > thing in the first place.
> > >
> > > This can happen when FIFO size + PBL settings are not big enough for COE.
> > >
> > > Can you please share the above settings ?
> >
> > Can the FIFO size be discovered by dumping registers, or does someone
> > from Rockchip need to look up the IP configuration details?
> >
> > FWIW, taking a look at the RK3399 TRM, this (p788) jumps out:
> >
> > "PBL
> > ...
> > For TxFIFO, valid PBL range in full duplex mode and duplex mode is
> > 128 or less.
> > For RxFIFO, valid PBL range in full duplex mode is all."
> >
> >
> > Does that suggest that it's worth fiddling with the "snps,txpbl" value
> > in DT?
>
> Yes, please try with PBL = 0x1 and no-pbl-x8. Performance will be lower
> but at least we will know if thatâs the cause.

OK, this looks promising - I've never encountered the problem 
'naturally' myself, but I managed to contrive a test setup with my 
RK3399 board wired directly to a laptop running an iperf3 server. As 
standard with an MTU of 1500, I get ~940Mbps as expected; with MTU 
bumped up to 1550 at both ends, the client grinds to a halt after ~100KB 
sent with a dozen or so retries, and the server reports 0 bytes 
received; with "snps,no-pbl-x8" present and the MTU still at 1550, it's 
back to ~940Mbps with 0 retries again.

I'll experiment further with txpbl, and my other boards, as and when I 
have time, but at this point I'm already pretty much convinced you're 
right about that Tx FIFO.

Cheers,
Robin.