Re: [PATCH v3 2/2] PCI: rockchip: add DesignWare based PCIe controller

[Robin Murphy <robin.murphy@arm.com>]

On 2021-01-25 09:01, Leon Romanovsky wrote:
> On Mon, Jan 25, 2021 at 02:40:10PM +0800, xxm wrote:
>> Hi Leon,
>>
>> Thanks for your reply.
>>
>> 在 2021/1/25 13:48, Leon Romanovsky 写道:
>>> On Mon, Jan 25, 2021 at 10:49:27AM +0800, Simon Xue wrote:
>>>> pcie-dw-rockchip is based on DWC IP. But pcie-rockchip-host
>>>> is Rockchip designed IP which is only used for RK3399. So all the following
>>>> non-RK3399 SoCs should use this driver.
>>>>
>>>> Signed-off-by: Simon Xue <xxm@rock-chips.com>
>>>> Signed-off-by: Shawn Lin <shawn.lin@rock-chips.com>
>>>> ---
>>>>    drivers/pci/controller/dwc/Kconfig            |   9 +
>>>>    drivers/pci/controller/dwc/Makefile           |   1 +
>>>>    drivers/pci/controller/dwc/pcie-dw-rockchip.c | 286 ++++++++++++++++++
>>>>    3 files changed, 296 insertions(+)
>>>>    create mode 100644 drivers/pci/controller/dwc/pcie-dw-rockchip.c
>>>>
>>>> diff --git a/drivers/pci/controller/dwc/Kconfig b/drivers/pci/controller/dwc/Kconfig
>>>> index 22c5529e9a65..aee408fe9283 100644
>>>> --- a/drivers/pci/controller/dwc/Kconfig
>>>> +++ b/drivers/pci/controller/dwc/Kconfig
>>>> @@ -214,6 +214,15 @@ config PCIE_ARTPEC6_EP
>>>>    	  Enables support for the PCIe controller in the ARTPEC-6 SoC to work in
>>>>    	  endpoint mode. This uses the DesignWare core.
>>>>
>>>> +config PCIE_ROCKCHIP_DW_HOST
>>>> +	bool "Rockchip DesignWare PCIe controller"
>>>> +	select PCIE_DW
>>>> +	select PCIE_DW_HOST
>>>> +	depends on ARCH_ROCKCHIP || COMPILE_TEST
>>>> +	depends on OF
>>>> +	help
>>>> +	  Enables support for the DW PCIe controller in the Rockchip SoC.
>>>> +
>>>>    config PCIE_INTEL_GW
>>>>    	bool "Intel Gateway PCIe host controller support"
>>>>    	depends on OF && (X86 || COMPILE_TEST)
>>>> diff --git a/drivers/pci/controller/dwc/Makefile b/drivers/pci/controller/dwc/Makefile
>>>> index a751553fa0db..30eef8e9ee8a 100644
>>>> --- a/drivers/pci/controller/dwc/Makefile
>>>> +++ b/drivers/pci/controller/dwc/Makefile
>>>> @@ -13,6 +13,7 @@ obj-$(CONFIG_PCI_LAYERSCAPE_EP) += pci-layerscape-ep.o
>>>>    obj-$(CONFIG_PCIE_QCOM) += pcie-qcom.o
>>>>    obj-$(CONFIG_PCIE_ARMADA_8K) += pcie-armada8k.o
>>>>    obj-$(CONFIG_PCIE_ARTPEC6) += pcie-artpec6.o
>>>> +obj-$(CONFIG_PCIE_ROCKCHIP_DW_HOST) += pcie-dw-rockchip.o
>>>>    obj-$(CONFIG_PCIE_INTEL_GW) += pcie-intel-gw.o
>>>>    obj-$(CONFIG_PCIE_KIRIN) += pcie-kirin.o
>>>>    obj-$(CONFIG_PCIE_HISI_STB) += pcie-histb.o
>>>> diff --git a/drivers/pci/controller/dwc/pcie-dw-rockchip.c b/drivers/pci/controller/dwc/pcie-dw-rockchip.c
>>>> new file mode 100644
>>>> index 000000000000..07f6d1cd5853
>>>> --- /dev/null
>>>> +++ b/drivers/pci/controller/dwc/pcie-dw-rockchip.c
>>>> @@ -0,0 +1,286 @@
>>>> +// SPDX-License-Identifier: GPL-2.0
>>>> +/*
>>>> + * PCIe host controller driver for Rockchip SoCs
>>>> + *
>>>> + * Copyright (C) 2021 Rockchip Electronics Co., Ltd.
>>>> + *		http://www.rock-chips.com
>>>> + *
>>>> + * Author: Simon Xue <xxm@rock-chips.com>
>>>> + */
>>>> +
>>>> +#include <linux/clk.h>
>>>> +#include <linux/gpio/consumer.h>
>>>> +#include <linux/mfd/syscon.h>
>>>> +#include <linux/module.h>
>>>> +#include <linux/of_device.h>
>>>> +#include <linux/phy/phy.h>
>>>> +#include <linux/platform_device.h>
>>>> +#include <linux/regmap.h>
>>>> +#include <linux/reset.h>
>>>> +
>>>> +#include "pcie-designware.h"
>>>> +
>>>> +/*
>>>> + * The upper 16 bits of PCIE_CLIENT_CONFIG are a write
>>>> + * mask for the lower 16 bits.  This allows atomic updates
>>>> + * of the register without locking.
>>>> + */
>>> This is correct only for the variables that naturally aligned, I imagine
>>> that this is the case here and in the Linux, but better do not write comments
>>> in the code that are not accurate.
>>
>> Ok, will remove.
>> I wonder what it would be when outside the Linux? Could you share some information?
> 
> The C standard says nothing about atomicity, integer assignment maybe atomic,
> maybe it isn’t. There is no guarantee, plain integer assignment in C is non-atomic
> by definition.
> 
> The atomicity of u32 is very dependent on hardware vendor, memory model and compiler,
> for example x86 and ARMs guarantee atomicity for u32. This is why I said that probably
> here (Linux) it is ok and you are not alone in expecting lockless write.

Huh? What do variables and the abstract machine of the C language 
environment have to do with the definition of *hardware MMIO registers*? 
We don't write to registers with plain integer assignment of u32, we use 
writel() (precisely in order to bypass that abstract C environment).

I appreciate that the comment is not universally true if taken 
completely out of context, but I that's true of pretty much all comments 
ever. If someone really were trying to learn basic programming 
principles from random comments in Linux drivers, then it's already a 
bit late for us to try and save them from themselves.

32-bit writes to these registers *will* be aligned, because the hardware 
decodes them at 32-bit-aligned addresses and there is nothing that can 
change that other than deliberately modifying the RTL in order to waste 
a large amount money fabbing a special broken version of the SoC. It can 
also be safely assumed that 32-bit writes to whichever part of the SoC 
memory map this device is placed *will* be issued atomically by the CPU 
and propagated atomically by the interconnect, because any SoCs 
integrating this device (or pretty much any modern peripheral IP) must 
be designed to meet those requirements for it to work correctly at all.

Robin.