Re: [PATCH net v4 3/3] net: stmmac: Specify hardware capability value when FIFO size isn't specified

[Andrew Lunn]

On Fri, Jan 31, 2025 at 03:54:16PM +0000, Steven Price wrote:
> On 31/01/2025 15:29, Andrew Lunn wrote:
> > On Fri, Jan 31, 2025 at 03:03:11PM +0000, Steven Price wrote:
> >> On 31/01/2025 14:47, Andrew Lunn wrote:
> >>>>> I'm guessing, but in your setup, i assume the value is never written
> >>>>> to a register, hence 0 is O.K. e.g. dwmac1000_dma_operation_mode_rx(),
> >>>>> the fifosz value is used to determine if flow control can be used, but
> >>>>> is otherwise ignored.
> >>>>
> >>>> I haven't traced the code, but that fits my assumptions too.
> >>>
> >>> I could probably figure it out using code review, but do you know
> >>> which set of DMA operations your hardware uses? A quick look at
> >>> dwmac-rk.c i see:
> >>>
> >>>         /* If the stmmac is not already selected as gmac4,
> >>>          * then make sure we fallback to gmac.
> >>>          */
> >>>         if (!plat_dat->has_gmac4)
> >>>                 plat_dat->has_gmac = true;
> >>
> >> has_gmac4 is false on this board, so has_gmac will be set to true here.
> > 
> > Thanks. Looking in hwif.c, that means dwmac1000_dma_ops are used.
> > 
> > dwmac1000_dma_operation_mode_rx() just does a check:
> > 
> > 	if (rxfifosz < 4096) {
> > 		csr6 &= ~DMA_CONTROL_EFC;
> > 
> > but otherwise does not use the value.
> > 
> > dwmac1000_dma_operation_mode_tx() appears to completely ignore fifosz.
> > 
> > So i would say all zero is valid for has_gmac == true, but you might
> > gain flow control if a value was passed.
> > 
> > A quick look at dwmac100_dma_operation_mode_tx() suggests fifosz is
> > also ignored, and dwmac100_dma_operation_mode_rx() does not exist. So
> > all 0 is also valid for gmac == false, gmac4 == false, and xgmac ==
> > false.
> > 
> > dwmac4_dma_rx_chan_op_mode() does use the value to determine mtl_rx_op
> > which gets written to a register. So gmac4 == true looks to need
> > values. dwxgmac2_dma_rx_mode() is the same, so xgmac = true also need
> > valid values.
> > 
> > Could you cook up a fix based on this?
> 
> The below works for me, I haven't got the hardware to actually test the 
> gmac4/xgmac paths:

This looks sensible. Hopefully Kunihiko can test on more platforms.

	Andrew