Re: [PATCH v3 08/11] riscv: dts: add initial SpacemiT K1 SoC device tree

From: Jisheng Zhang
Date: Mon Jul 08 2024 - 08:46:11 EST

On Sat, Jul 06, 2024 at 02:24:03PM +0000, Yixun Lan wrote:
> On 18:40 Sat 06 Jul , Jisheng Zhang wrote:
> > On Sat, Jul 06, 2024 at 05:05:56AM +0000, Yixun Lan wrote:
> > >
> > > On 12:12 Sat 06 Jul , Jisheng Zhang wrote:
> > > > On Fri, Jul 05, 2024 at 06:38:39AM +0000, Yixun Lan wrote:
> > > > >
> > > > > On 21:46 Thu 04 Jul , Jisheng Zhang wrote:
> > > > > > On Wed, Jul 03, 2024 at 02:55:11PM +0000, Yixun Lan wrote:
> > > > > > > From: Yangyu Chen <cyy@xxxxxxxxxxxx>
> > > > > > >
> > > > > > > Banana Pi BPI-F3 motherboard is powered by SpacemiT K1[1].
> > > > > > >
> > > > > > > Key features:
> > > > > > > - 4 cores per cluster, 2 clusters on chip
> > > > > > > - UART IP is Intel XScale UART
> > > > > > >
> > > > > > > Some key considerations:
> > > > > > > - ISA string is inferred from vendor documentation[2]
> > > > > > > - Cluster topology is inferred from datasheet[1] and L2 in vendor dts[3]
> > > > > > > - No coherent DMA on this board
> > > > > > > Inferred by taking vendor ethernet and MMC drivers to the mainline
> > > > > > > kernel. Without dma-noncoherent in soc node, the driver fails.
> > > > > > > - No cache nodes now
> > > > > > > The parameters from vendor dts are likely to be wrong. It has 512
> > > > > > > sets for a 32KiB L1 Cache. In this case, each set is 64B in size.
> > > > > > > When the size of the cache line is 64B, it is a directly mapped
> > > > > > > cache rather than a set-associative cache, the latter is commonly
> > > > > > > used. Thus, I didn't use the parameters from vendor dts.
> > > > > > >
> > > > > > > Currently only support booting into console with only uart, other
> > > > > > > features will be added soon later.
> > > > > > >
> > > > > > > Link: https://docs.banana-pi.org/en/BPI-F3/SpacemiT_K1_datasheet [1]
> > > > > > > Link: https://developer.spacemit.com/#/documentation?token=BWbGwbx7liGW21kq9lucSA6Vnpb [2]
> > > > > > > Link: https://gitee.com/bianbu-linux/linux-6.1/blob/bl-v1.0.y/arch/riscv/boot/dts/spacemit/k1-x.dtsi [3]
> > > > > > > Signed-off-by: Yangyu Chen <cyy@xxxxxxxxxxxx>
> > > > > > > Signed-off-by: Yixun Lan <dlan@xxxxxxxxxx>
> > > > > > > ---
> > > > > > > arch/riscv/boot/dts/spacemit/k1.dtsi | 376 +++++++++++++++++++++++++++++++++++
> > > > > > > 1 file changed, 376 insertions(+)
> > > > > > >
> > > > > > > diff --git a/arch/riscv/boot/dts/spacemit/k1.dtsi b/arch/riscv/boot/dts/spacemit/k1.dtsi
> > > > > > > new file mode 100644
> > > > > > > index 0000000000000..a076e35855a2e
> > > > > > > --- /dev/null
> > > > > > > +++ b/arch/riscv/boot/dts/spacemit/k1.dtsi
> > > > > > > @@ -0,0 +1,376 @@
> > > > > > > +// SPDX-License-Identifier: GPL-2.0 OR MIT
> > > > > > > +/*
> > > > > > > + * Copyright (C) 2024 Yangyu Chen <cyy@xxxxxxxxxxxx>
> > > > > > > + */
> > > > > > > +
> > > > > > > +/dts-v1/;
> > > > > > > +/ {
> > > > > > > + #address-cells = <2>;
> > > > > > > + #size-cells = <2>;
> > > > > > > + model = "SpacemiT K1";
> > > > > > > + compatible = "spacemit,k1";
> > > > > > > +
> > > > > > > +
> > ...
> > > > > > > + soc {
> > > > > > > + compatible = "simple-bus";
> > > > > > > + interrupt-parent = <&plic>;
> > > > > > > + #address-cells = <2>;
> > > > > > > + #size-cells = <2>;
> > > > > > > + dma-noncoherent;
> > > > > > > + ranges;
> > > > > > > +
> > > > > > > + uart0: serial@d4017000 {
> > > > > > > + compatible = "spacemit,k1-uart", "intel,xscale-uart";
> > > > > >
> > > > > > no, this is not a correct hw modeling. The doc on spacemit says
> > > > > > all the uart support 64 bytes FIFO, declaring xscale only makes
> > > > > > use of 32 bytes FIFO.
> > > > > yes, I also noticed it's 64 bytes FIFO
> > > > >
> > > > > >
> > > > > > IIRC, 8250_pxa is a xscale uart with 64 bytes FIFO, so this should be
> > > > > > "mrvl,pxa-uart" or "mrvl,mmp-uart"
> > > > >
> > > > >
> > > > > for mrvl,pxa-uart, I think you imply to use drivers/tty/serial/8250/8250_pxa.c,
> > > > > which turn out doesn't work on k1 SoC, for the record, we need to adjust
> > > >
> > > > Really? I just tried "mrvl,pxa-uart" with rc6, it works perfectly, and the FIFO
> > > > in the driver logic is 64bytes now. Am I misssing something or you never tried it?
> > > >
> > > Ok, I realised it's the clock issue
> > >
> > > still, I'm not fully convinced about using "mrvl,pxa-uart",
> > > e.g this driver hardcoded tz_loadsz to 32, not sure if K1 suffer same problem
> > > 5208e7ced520 ("serial: 8250_pxa: Configure tx_loadsz to match FIFO IRQ level")
> >
> > I believe the problem commit 5208e7ced520 tries to solve is: the
> > mmp|pxa-uart only support threshold up to 32Bytes, tz_loadsz will be
> > fifo size by default, this will cause probleme with 64Bytes FIFO.
> >
> yes, exactly
>
> > >
> > > also, what's the preference when choosing driver between 8250_pxa.c vs 8250_of.c?
> >
> > Good question. I have no preference. But there are two problems with
> > 8250_of, I have sent out patches[1][2] to address them.
> >
> > After these two patches, both the earlycon and uart FIFO logic work too
> > with below dts properties:
> > uart0: serial@d4017000 {
> > compatible = "mrvl,mmp-uart";
> to be precise, I think here should be compatible = "mrvl,mmp-uart", "intel,xscale-uart"

see below.

>
> but can you check this patch below? it should be ok with your two proposed patches applied
> https://lore.kernel.org/all/20240703-k1-01-basic-dt-v3-6-12f73b47461e@xxxxxxxxxx/
>
> > ...
> > reg-shift = <2>;
> > reg-io-width = <4>;
> > tx-threshold = <32>;
> > fifo-size = <64>;

I just tried, the previous "spacemit,k1-uart", "intel,xscale-uart"; with above properties
work too, and the fifo size in driver logic seems correct as well.

> ..
> > no-loopback-test;
> need to check, from vendor docs, there is a loopback mode

oh, this property is from the UPF_SKIP_TEST of 8250_pxa.c. This
property can be removed.

> see 16.2.4.1 SSCR register description, bit12
>
> https://developer.spacemit.com/#/documentation?token=Rn9Kw3iFHirAMgkIpTAcV2Arnkf
> > ...
> > }
> >
> > Link: https://lore.kernel.org/linux-riscv/20240706082928.2238-1-jszhang@xxxxxxxxxx/ [1]
> I have some comments for this patch, and I believe it's a valid fix,
> without this patch, K1 will also have problem duo to "UART_CAP_UUE | UART_CAP_RTOIE" lost
>
> > Link: https://lore.kernel.org/linux-riscv/20240706101856.3077-1-jszhang@xxxxxxxxxx/ [2]
> >
> > > it occur to me that 8250_pxa.c is more specially tailored for pxa hardware, while
> > > 8250_of.c is more generic.. besides, should we consider one more step if we want to
> >
> > there's a work around for Erratum #74 in 8250_pxa, while I believe the
> do you have any link for this Erratum? let's double check it..

I can't find any link now :(

>
> > Errata doesn't exisit in K1, so from this PoV it seems 8250_of is
> > better, no?
> >
> > > support DMA mode in the future (vendor uart driver has DMA support)?
> >
> > Adding dma engine support to 8250_of is doable.
> Ok, sounds good to me
> >
> > >
> > >
> > > > > drivers/tty/serial/8250/Kconfig to enable the driver for ARCH_SPACEMIT,
> > > > > and change uart compatible to "spacemit,k1-uart", "mrvl,pxa-uart"
> > > > >
> > > > > for mrvl,mmp-uart, I see two choices, one using 8250_pxa.c which has same result
> > > > > as mrvl,pxa-uart, another choice would using the driver of 8250_of.c
> > > > > and it work as same as "intel,xscale-uart", I don't see any difference..
> > > > >
> > > > > P.S: there is possibly a side problem that "mrvl,mmp-uart" from 8250_of.c doesn't
> > > > > really compatile with "mrvl,mmp-uart" from 8250_pxa.c, but I think it's another story
> > >
> > > --
> > > Yixun Lan (dlan)
> > > Gentoo Linux Developer
> > > GPG Key ID AABEFD55
>
> --
> Yixun Lan (dlan)
> Gentoo Linux Developer
> GPG Key ID AABEFD55