Re: [PATCH 1/6] cpu: add t-head's c9xx

[Nils Le Roux <gilbsgilbert@gmail.com>]

Hi!

I tested this and it works fine. I also happen to have the glue layer ported
from the Linux somewhere on my machine. If you're interested, please 
reach out
to me and I'll arrange a patch. I have a few remarks on this patch:

On 3/27/24 9:07 AM, wefu@redhat.com wrote:
> From: Wei Fu <wefu@redhat.com>
>
> Signed-off-by: Wei Fu <wefu@redhat.com>
> Co-authored-by: Yixun Lan <dlan@gentoo.org>
> ---
>   arch/riscv/Kconfig               |  1 +
>   arch/riscv/cpu/c9xx/Kconfig      | 12 ++++++++
>   arch/riscv/cpu/c9xx/Makefile     |  5 ++++
>   arch/riscv/cpu/c9xx/cpu.c        | 51 ++++++++++++++++++++++++++++++++
>   arch/riscv/cpu/c9xx/dram.c       | 36 ++++++++++++++++++++++
>   arch/riscv/include/asm/csr.h     |  2 ++
>   board/thead/th1520_lpi4a/Kconfig |  3 +-
>   7 files changed, 109 insertions(+), 1 deletion(-)
>
> diff --git a/arch/riscv/Kconfig b/arch/riscv/Kconfig
> index ac52c5e6da..ac3b802abe 100644
> --- a/arch/riscv/Kconfig
> +++ b/arch/riscv/Kconfig
> @@ -97,6 +97,7 @@ source "arch/riscv/cpu/fu540/Kconfig"
>   source "arch/riscv/cpu/fu740/Kconfig"
>   source "arch/riscv/cpu/generic/Kconfig"
>   source "arch/riscv/cpu/jh7110/Kconfig"
> +source "arch/riscv/cpu/c9xx/Kconfig"
>   
>   # architecture-specific options below
>   
> diff --git a/arch/riscv/cpu/c9xx/Kconfig b/arch/riscv/cpu/c9xx/Kconfig
> new file mode 100644
> index 0000000000..5a84bcacd6
> --- /dev/null
> +++ b/arch/riscv/cpu/c9xx/Kconfig
> @@ -0,0 +1,12 @@
> +# SPDX-License-Identifier: GPL-2.0+
> +#
> +# Copyright (C) 2017-2020 Alibaba Group Holding Limited
> +
> +config RISCV_THEAD
> +	bool
> +	select ARCH_EARLY_INIT_R
> +	imply CPU
> +	imply CPU_RISCV
> +	imply RISCV_TIMER
> +	imply RISCV_RDTIME
> +	imply CMD_CPU
> diff --git a/arch/riscv/cpu/c9xx/Makefile b/arch/riscv/cpu/c9xx/Makefile
> new file mode 100644
> index 0000000000..e3f776cb41
> --- /dev/null
> +++ b/arch/riscv/cpu/c9xx/Makefile
> @@ -0,0 +1,5 @@
> +# SPDX-License-Identifier: GPL-2.0+
> +#
> +# Copyright (C) 2017-2020 Alibaba Group Holding Limited
> +
> +obj-y += cpu.o dram.o
> diff --git a/arch/riscv/cpu/c9xx/cpu.c b/arch/riscv/cpu/c9xx/cpu.c
> new file mode 100644
> index 0000000000..4b21edd62b
> --- /dev/null
> +++ b/arch/riscv/cpu/c9xx/cpu.c
> @@ -0,0 +1,51 @@
> +// SPDX-License-Identifier: GPL-2.0+
> +/*
> + * Copyright (C) 2017-2020 Alibaba Group Holding Limited
> + */
> +
> +#include <asm/cache.h>
> +#include <asm/csr.h>
> +#include <cpu_func.h>
> +
> +/*
> + * cleanup_before_linux() is called just before we call linux
> + * it prepares the processor for linux
> + *
> + * we disable interrupt and caches.
> + */
> +int cleanup_before_linux(void)
> +{
> +	cache_flush();
> +
> +	return 0;
> +}
> +
> +void flush_dcache_range(unsigned long start, unsigned long end)
> +{
> +	register unsigned long i asm("a0") = start & ~(CONFIG_SYS_CACHELINE_SIZE - 1);
Isn't this just round_down/ALIGN_DOWN macro?
> +
> +	for (; i < end; i += CONFIG_SYS_CACHELINE_SIZE)
> +		asm volatile(".long 0x0295000b");  /* dcache.cpa a0 */
> +
> +	sync_is();
> +}
> +
> +void invalidate_dcache_range(unsigned long start, unsigned long end)
> +{
> +	register unsigned long i asm("a0") = start & ~(CONFIG_SYS_CACHELINE_SIZE - 1);
> +
> +	for (; i < end; i += CONFIG_SYS_CACHELINE_SIZE)
> +		asm volatile(".long 0x02b5000b");  /* dcache.cipa a0 */

`th.dcache.cipa` not only invalidates the dcache, but also flushes it 
which is
not expected here. In practice, this leads to a bug with the dwmac 
driver: when
U-Boot handles an ARP request, it overwrites the ARP request with the ARP
response in-place, directly in the ring buffer[^1]. When re-using the same
buffer later on, the driver will flush this data instead of invalidating
it[^2], and you'll read the ARP response you just sent instead of the packet
just written by the chip.

This can easily be reproduced by downloading a big file using TFTP. As 
soon as
you receive an ARP request, U-Boot hangs, timeouts and requests the packet
again. It also makes TCP completely unusable.

You can use `th.dcache.ipa` instead, as implemented in 
`invalid_dcache_range`.

[^1] 
https://github.com/u-boot/u-boot/blob/a5ec56aea1a56737a4e124d058a6920d16f5e686/net/arp.c#L154-L160
[^2] 
https://github.com/u-boot/u-boot/blob/a5ec56aea1a56737a4e124d058a6920d16f5e686/drivers/net/designware.c#L542

> +
> +	sync_is();
> +}
> +
> +void invalid_dcache_range(unsigned long start, unsigned long end)

I think this isn't used anywhere.

> +{
> +	register unsigned long i asm("a0") = start & ~(CONFIG_SYS_CACHELINE_SIZE - 1);
> +
> +	for (; i < end; i += CONFIG_SYS_CACHELINE_SIZE)
> +		asm volatile(".long 0x02a5000b");  /* dcache.ipa a0 */
> +
> +	sync_is();
> +}
> diff --git a/arch/riscv/cpu/c9xx/dram.c b/arch/riscv/cpu/c9xx/dram.c
> new file mode 100644
> index 0000000000..614d7bf1cc
> --- /dev/null
> +++ b/arch/riscv/cpu/c9xx/dram.c
> @@ -0,0 +1,36 @@
> +// SPDX-License-Identifier: GPL-2.0+
> +/*
> + * Copyright (C) 2018, Bin Meng <bmeng.cn@gmail.com>
> + */
> +
> +#include <fdtdec.h>
> +#include <init.h>
> +#include <linux/sizes.h>
> +
> +DECLARE_GLOBAL_DATA_PTR;
> +
> +int dram_init(void)
> +{
> +	return fdtdec_setup_mem_size_base();
> +}
> +
> +int dram_init_banksize(void)
> +{
> +	return fdtdec_setup_memory_banksize();
> +}
> +
> +phys_size_t board_get_usable_ram_top(phys_size_t total_size)
> +{
> +	/*
> +	 * Ensure that we run from first 4GB so that all
> +	 * addresses used by U-Boot are 32bit addresses.
> +	 *
> +	 * This in-turn ensures that 32bit DMA capable
> +	 * devices work fine because DMA mapping APIs will
> +	 * provide 32bit DMA addresses only.
> +	 */
> +	if (gd->ram_top >= SZ_4G)
> +		return SZ_4G - 1;
> +
> +	return gd->ram_top;
> +}
> diff --git a/arch/riscv/include/asm/csr.h b/arch/riscv/include/asm/csr.h
> index 986f951c31..3102de6cbb 100644
> --- a/arch/riscv/include/asm/csr.h
> +++ b/arch/riscv/include/asm/csr.h
> @@ -145,6 +145,8 @@
>   #define CSR_MARCHID		0xf12
>   #define CSR_MHARTID		0xf14
>   
> +#define sync_is()   asm volatile (".long 0x01b0000b")
> +

I believe this instruction is also T-Head specific (at least my disassembler
doesn't know it). I doubt it belongs here.

Best regards,
Nils

>   #ifndef __ASSEMBLY__
>   
>   #define csr_swap(csr, val)					\
> diff --git a/board/thead/th1520_lpi4a/Kconfig b/board/thead/th1520_lpi4a/Kconfig
> index 622246127c..bef0638e80 100644
> --- a/board/thead/th1520_lpi4a/Kconfig
> +++ b/board/thead/th1520_lpi4a/Kconfig
> @@ -11,7 +11,7 @@ config SYS_VENDOR
>   	default "thead"
>   
>   config SYS_CPU
> -	default "generic"
> +	default "c9xx"
>   
>   config SYS_CONFIG_NAME
>   	default "th1520_lpi4a"
> @@ -30,6 +30,7 @@ config SPL_OPENSBI_LOAD_ADDR
>   config BOARD_SPECIFIC_OPTIONS
>   	def_bool y
>   	select ARCH_EARLY_INIT_R
> +	select RISCV_THEAD
>   	imply CPU
>   	imply CPU_RISCV
>   	imply RISCV_TIMER if RISCV_SMODE