[PATCH] PCI: j721e: Extend j721e_pcie_ctrl_init() for non syscon nodes

[PATCH] PCI: j721e: Extend j721e_pcie_ctrl_init() for non syscon nodes

[Siddharth Vadapalli]

The "ti,syscon-pcie-ctrl" device-tree property is used to obtain
reference to the "pcie_ctrl" register within the System Controller Module
in order to configure the link speed, number of lanes and the mode of
operation of the PCIe controller. The existing implementation of the
"j721e_pcie_ctrl_init()" function handles the case where the compatible for
the System Controller Module node specified using the "ti,syscon-pcie-ctrl"
property is "syscon". Since the System Controller Module can be modelled as
a "simple-bus" as well, extend the implementation of the
"j721e_pcie_ctrl_init()" function to handle the "simple-bus" case.

Signed-off-by: Siddharth Vadapalli <s-vadapalli@ti.com>
---

Hello,

This patch is based on linux-next tagged next-20240129.

The System Controller Module is modelled as a "simple-bus" in J784S4 SoC at
https://github.com/torvalds/linux/blob/master/arch/arm64/boot/dts/ti/k3-j784s4-main.dtsi#L45
The existing SoCs such as J721E and J7200 which currently model the node as
a Syscon node will also be updated to model it as a "simple-bus".
Therefore this patch aims to update the driver in order to handle the
migration of the System Controller Module to the "simple-bus" compatible
without breaking PCIe functionality on existing TI SoCs which make use
of the pci-j721e.c driver.

Regards,
Siddharth.

 drivers/pci/controller/cadence/pci-j721e.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

diff --git a/drivers/pci/controller/cadence/pci-j721e.c b/drivers/pci/controller/cadence/pci-j721e.c
index 85718246016b..2ace7e78a880 100644
--- a/drivers/pci/controller/cadence/pci-j721e.c
+++ b/drivers/pci/controller/cadence/pci-j721e.c
@@ -224,12 +224,20 @@ static int j721e_pcie_ctrl_init(struct j721e_pcie *pcie)
 {
 	struct device *dev = pcie->cdns_pcie->dev;
 	struct device_node *node = dev->of_node;
+	struct device_node *scm_conf;
 	struct of_phandle_args args;
 	unsigned int offset = 0;
 	struct regmap *syscon;
 	int ret;
 
-	syscon = syscon_regmap_lookup_by_phandle(node, "ti,syscon-pcie-ctrl");
+	scm_conf = of_parse_phandle(node, "ti,syscon-pcie-ctrl", 0);
+	if (!scm_conf) {
+		dev_err(dev, "unable to get System Controller node\n");
+		return -ENODEV;
+	}
+
+	syscon = device_node_to_regmap(scm_conf);
+	of_node_put(scm_conf);
 	if (IS_ERR(syscon)) {
 		dev_err(dev, "Unable to get ti,syscon-pcie-ctrl regmap\n");
 		return PTR_ERR(syscon);
-- 
2.34.1

Re: [PATCH] PCI: j721e: Extend j721e_pcie_ctrl_init() for non syscon nodes

[Andrew Davis]

On 1/29/24 4:49 AM, Siddharth Vadapalli wrote:
> The "ti,syscon-pcie-ctrl" device-tree property is used to obtain
> reference to the "pcie_ctrl" register within the System Controller Module
> in order to configure the link speed, number of lanes and the mode of
> operation of the PCIe controller. The existing implementation of the
> "j721e_pcie_ctrl_init()" function handles the case where the compatible for
> the System Controller Module node specified using the "ti,syscon-pcie-ctrl"
> property is "syscon". Since the System Controller Module can be modelled as
> a "simple-bus" as well, extend the implementation of the
> "j721e_pcie_ctrl_init()" function to handle the "simple-bus" case.
> 
> Signed-off-by: Siddharth Vadapalli <s-vadapalli@ti.com>
> ---
> 
> Hello,
> 
> This patch is based on linux-next tagged next-20240129.
> 
> The System Controller Module is modelled as a "simple-bus" in J784S4 SoC at
> https://github.com/torvalds/linux/blob/master/arch/arm64/boot/dts/ti/k3-j784s4-main.dtsi#L45
> The existing SoCs such as J721E and J7200 which currently model the node as
> a Syscon node will also be updated to model it as a "simple-bus".
> Therefore this patch aims to update the driver in order to handle the
> migration of the System Controller Module to the "simple-bus" compatible
> without breaking PCIe functionality on existing TI SoCs which make use
> of the pci-j721e.c driver.
> 
> Regards,
> Siddharth.
> 
>   drivers/pci/controller/cadence/pci-j721e.c | 10 +++++++++-
>   1 file changed, 9 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/pci/controller/cadence/pci-j721e.c b/drivers/pci/controller/cadence/pci-j721e.c
> index 85718246016b..2ace7e78a880 100644
> --- a/drivers/pci/controller/cadence/pci-j721e.c
> +++ b/drivers/pci/controller/cadence/pci-j721e.c
> @@ -224,12 +224,20 @@ static int j721e_pcie_ctrl_init(struct j721e_pcie *pcie)
>   {
>   	struct device *dev = pcie->cdns_pcie->dev;
>   	struct device_node *node = dev->of_node;
> +	struct device_node *scm_conf;
>   	struct of_phandle_args args;
>   	unsigned int offset = 0;
>   	struct regmap *syscon;
>   	int ret;
>   
> -	syscon = syscon_regmap_lookup_by_phandle(node, "ti,syscon-pcie-ctrl");
> +	scm_conf = of_parse_phandle(node, "ti,syscon-pcie-ctrl", 0);
> +	if (!scm_conf) {
> +		dev_err(dev, "unable to get System Controller node\n");
> +		return -ENODEV;
> +	}
> +
> +	syscon = device_node_to_regmap(scm_conf);

Turning the entire "simple-bus" region into a regmap using this
function is just as broken as having it as a "syscon". The core
problem we are solving by getting rid of the blanket syscon nodes
is that it causes multiple mappings of the same register. This
can cause issues with regmap caching, read–modify–write, etc..

What you want to do is add a subnode to the simple-bus, have that
encapsulate just the registers used for PCIe, and have the PCIe
node point to that. Then this patch isn't needed.

For an example, see how it's done for DSS[0].

Andrew

[0] https://github.com/torvalds/linux/blob/41bccc98fb7931d63d03f326a746ac4d429c1dd3/arch/arm64/boot/dts/ti/k3-am65-main.dtsi#L502

> +	of_node_put(scm_conf);
>   	if (IS_ERR(syscon)) {
>   		dev_err(dev, "Unable to get ti,syscon-pcie-ctrl regmap\n");
>   		return PTR_ERR(syscon);

Re: [PATCH] PCI: j721e: Extend j721e_pcie_ctrl_init() for non syscon nodes

[Siddharth Vadapalli]

Hello Andrew,

On 29/01/24 20:49, Andrew Davis wrote:
> On 1/29/24 4:49 AM, Siddharth Vadapalli wrote:

...

>>       int ret;
>>   -    syscon = syscon_regmap_lookup_by_phandle(node, "ti,syscon-pcie-ctrl");
>> +    scm_conf = of_parse_phandle(node, "ti,syscon-pcie-ctrl", 0);
>> +    if (!scm_conf) {
>> +        dev_err(dev, "unable to get System Controller node\n");
>> +        return -ENODEV;
>> +    }
>> +
>> +    syscon = device_node_to_regmap(scm_conf);
> 
> Turning the entire "simple-bus" region into a regmap using this
> function is just as broken as having it as a "syscon". The core
> problem we are solving by getting rid of the blanket syscon nodes
> is that it causes multiple mappings of the same register. This
> can cause issues with regmap caching, read–modify–write, etc..
> 
> What you want to do is add a subnode to the simple-bus, have that
> encapsulate just the registers used for PCIe, and have the PCIe
> node point to that. Then this patch isn't needed.
> 
> For an example, see how it's done for DSS[0].

Thank you for reviewing the patch. I will implement it similar to what's done
for DSS as you pointed out. However, what about the existing SoCs which make use
of the "ti,syscon-pcie-ctrl" property? Do you suggest that I add another
device-tree property for pointing to the PCIE_CTRL register within the CTRL_MMR
region, or do you suggest that I reuse the existing "ti,syscon-pcie-ctrl"
property differently in the SoCs like J784S4 where the scm_conf node is a
"simple-bus"?

The "ti,syscon-pcie-ctrl" property as defined in the device-tree bindings has
two elements with the first being the phandle to the scm_conf node and the
second being the offset of the PCIE_CTRL register. The newer implementation you
are suggesting will either require a new property which accepts only one element
namely the phandle to the node within scm_conf corresponding to the PCIE_CTRL
register. Will adding a new property be acceptable?

...

-- 
Regards,
Siddharth.

Re: [PATCH] PCI: j721e: Extend j721e_pcie_ctrl_init() for non syscon nodes

[Andrew Davis]

On 1/29/24 10:50 PM, Siddharth Vadapalli wrote:
> Hello Andrew,
> 
> On 29/01/24 20:49, Andrew Davis wrote:
>> On 1/29/24 4:49 AM, Siddharth Vadapalli wrote:
> 
> ...
> 
>>>        int ret;
>>>    -    syscon = syscon_regmap_lookup_by_phandle(node, "ti,syscon-pcie-ctrl");
>>> +    scm_conf = of_parse_phandle(node, "ti,syscon-pcie-ctrl", 0);
>>> +    if (!scm_conf) {
>>> +        dev_err(dev, "unable to get System Controller node\n");
>>> +        return -ENODEV;
>>> +    }
>>> +
>>> +    syscon = device_node_to_regmap(scm_conf);
>>
>> Turning the entire "simple-bus" region into a regmap using this
>> function is just as broken as having it as a "syscon". The core
>> problem we are solving by getting rid of the blanket syscon nodes
>> is that it causes multiple mappings of the same register. This
>> can cause issues with regmap caching, read–modify–write, etc..
>>
>> What you want to do is add a subnode to the simple-bus, have that
>> encapsulate just the registers used for PCIe, and have the PCIe
>> node point to that. Then this patch isn't needed.
>>
>> For an example, see how it's done for DSS[0].
> 
> Thank you for reviewing the patch. I will implement it similar to what's done
> for DSS as you pointed out. However, what about the existing SoCs which make use
> of the "ti,syscon-pcie-ctrl" property? Do you suggest that I add another
> device-tree property for pointing to the PCIE_CTRL register within the CTRL_MMR
> region, or do you suggest that I reuse the existing "ti,syscon-pcie-ctrl"
> property differently in the SoCs like J784S4 where the scm_conf node is a
> "simple-bus"?
> 
> The "ti,syscon-pcie-ctrl" property as defined in the device-tree bindings has
> two elements with the first being the phandle to the scm_conf node and the
> second being the offset of the PCIE_CTRL register. The newer implementation you
> are suggesting will either require a new property which accepts only one element
> namely the phandle to the node within scm_conf corresponding to the PCIE_CTRL
> register. Will adding a new property be acceptable?
> 

Why would you need a new property? If there is no offset to the PCIE_CTRL register
in the new syscon space then just set the offset = 0x0, easy.

Andrew

Re: [PATCH] PCI: j721e: Extend j721e_pcie_ctrl_init() for non syscon nodes

[Siddharth Vadapalli]

On 30/01/24 20:30, Andrew Davis wrote:
> On 1/29/24 10:50 PM, Siddharth Vadapalli wrote:
>> Hello Andrew,
>>
>> On 29/01/24 20:49, Andrew Davis wrote:
>>> On 1/29/24 4:49 AM, Siddharth Vadapalli wrote:
>>
>> ...
>>
>>>>        int ret;
>>>>    -    syscon = syscon_regmap_lookup_by_phandle(node, "ti,syscon-pcie-ctrl");
>>>> +    scm_conf = of_parse_phandle(node, "ti,syscon-pcie-ctrl", 0);
>>>> +    if (!scm_conf) {
>>>> +        dev_err(dev, "unable to get System Controller node\n");
>>>> +        return -ENODEV;
>>>> +    }
>>>> +
>>>> +    syscon = device_node_to_regmap(scm_conf);
>>>
>>> Turning the entire "simple-bus" region into a regmap using this
>>> function is just as broken as having it as a "syscon". The core
>>> problem we are solving by getting rid of the blanket syscon nodes
>>> is that it causes multiple mappings of the same register. This
>>> can cause issues with regmap caching, read–modify–write, etc..
>>>
>>> What you want to do is add a subnode to the simple-bus, have that
>>> encapsulate just the registers used for PCIe, and have the PCIe
>>> node point to that. Then this patch isn't needed.
>>>
>>> For an example, see how it's done for DSS[0].
>>
>> Thank you for reviewing the patch. I will implement it similar to what's done
>> for DSS as you pointed out. However, what about the existing SoCs which make use
>> of the "ti,syscon-pcie-ctrl" property? Do you suggest that I add another
>> device-tree property for pointing to the PCIE_CTRL register within the CTRL_MMR
>> region, or do you suggest that I reuse the existing "ti,syscon-pcie-ctrl"
>> property differently in the SoCs like J784S4 where the scm_conf node is a
>> "simple-bus"?
>>
>> The "ti,syscon-pcie-ctrl" property as defined in the device-tree bindings has
>> two elements with the first being the phandle to the scm_conf node and the
>> second being the offset of the PCIE_CTRL register. The newer implementation you
>> are suggesting will either require a new property which accepts only one element
>> namely the phandle to the node within scm_conf corresponding to the PCIE_CTRL
>> register. Will adding a new property be acceptable?
>>
> 
> Why would you need a new property? If there is no offset to the PCIE_CTRL register
> in the new syscon space then just set the offset = 0x0, easy.

Seems like a Hack to me considering that the offset will always be zero for
non-syscon parent nodes (which will be the convention going forward), implying
that the offset might as well be dropped and just the phandle is sufficient. For
now I shall implement it as you suggested. Maybe once the syscon nodes in
existing SoCs are also converted to simple-bus, then the property can be
redefined to just be the phandle to "pcie_ctrl" register.

> 
> Andrew

-- 
Regards,
Siddharth.

Re: [PATCH] PCI: j721e: Extend j721e_pcie_ctrl_init() for non syscon nodes

[Andrew Davis]

On 1/30/24 10:53 PM, Siddharth Vadapalli wrote:
> 
> 
> On 30/01/24 20:30, Andrew Davis wrote:
>> On 1/29/24 10:50 PM, Siddharth Vadapalli wrote:
>>> Hello Andrew,
>>>
>>> On 29/01/24 20:49, Andrew Davis wrote:
>>>> On 1/29/24 4:49 AM, Siddharth Vadapalli wrote:
>>>
>>> ...
>>>
>>>>>         int ret;
>>>>>     -    syscon = syscon_regmap_lookup_by_phandle(node, "ti,syscon-pcie-ctrl");
>>>>> +    scm_conf = of_parse_phandle(node, "ti,syscon-pcie-ctrl", 0);
>>>>> +    if (!scm_conf) {
>>>>> +        dev_err(dev, "unable to get System Controller node\n");
>>>>> +        return -ENODEV;
>>>>> +    }
>>>>> +
>>>>> +    syscon = device_node_to_regmap(scm_conf);
>>>>
>>>> Turning the entire "simple-bus" region into a regmap using this
>>>> function is just as broken as having it as a "syscon". The core
>>>> problem we are solving by getting rid of the blanket syscon nodes
>>>> is that it causes multiple mappings of the same register. This
>>>> can cause issues with regmap caching, read–modify–write, etc..
>>>>
>>>> What you want to do is add a subnode to the simple-bus, have that
>>>> encapsulate just the registers used for PCIe, and have the PCIe
>>>> node point to that. Then this patch isn't needed.
>>>>
>>>> For an example, see how it's done for DSS[0].
>>>
>>> Thank you for reviewing the patch. I will implement it similar to what's done
>>> for DSS as you pointed out. However, what about the existing SoCs which make use
>>> of the "ti,syscon-pcie-ctrl" property? Do you suggest that I add another
>>> device-tree property for pointing to the PCIE_CTRL register within the CTRL_MMR
>>> region, or do you suggest that I reuse the existing "ti,syscon-pcie-ctrl"
>>> property differently in the SoCs like J784S4 where the scm_conf node is a
>>> "simple-bus"?
>>>
>>> The "ti,syscon-pcie-ctrl" property as defined in the device-tree bindings has
>>> two elements with the first being the phandle to the scm_conf node and the
>>> second being the offset of the PCIE_CTRL register. The newer implementation you
>>> are suggesting will either require a new property which accepts only one element
>>> namely the phandle to the node within scm_conf corresponding to the PCIE_CTRL
>>> register. Will adding a new property be acceptable?
>>>
>>
>> Why would you need a new property? If there is no offset to the PCIE_CTRL register
>> in the new syscon space then just set the offset = 0x0, easy.
> 
> Seems like a Hack to me considering that the offset will always be zero for
> non-syscon parent nodes (which will be the convention going forward), implying
> that the offset might as well be dropped and just the phandle is sufficient. For

If we check the git history, this is actually how it used to be. The offset stuff
was added later[0]. Looks like for backwards compat it still works to not provide
an offset.

Andrew

[0] 7aa256234c4c ("PCI: j721e: Get offset within "syscon" from "ti,syscon-pcie-ctrl" phandle arg")

> now I shall implement it as you suggested. Maybe once the syscon nodes in
> existing SoCs are also converted to simple-bus, then the property can be
> redefined to just be the phandle to "pcie_ctrl" register.
> 
>>
>> Andrew
>