hello~
On 2025/1/23 6:21, Bjorn Helgaas wrote:
On Wed, Jan 15, 2025 at 03:06:37PM +0800, Chen Wang wrote:
From: Chen Wang <unicorn_wang@xxxxxxxxxxx>
Add binding for Sophgo SG2042 PCIe host controller.
+ sophgo,link-id:
+ $ref: /schemas/types.yaml#/definitions/uint32
+ description: |
+ SG2042 uses Cadence IP, every IP is composed of 2 cores (called link0
+ & link1 as Cadence's term). Each core corresponds to a host bridge,
+ and each host bridge has only one root port. Their configuration
+ registers are completely independent. SG2042 integrates two Cadence IPs,
+ so there can actually be up to four host bridges. "sophgo,link-id" is
+ used to identify which core/link the PCIe host bridge node corresponds to.
IIUC, the registers of Cadence IP 1 and IP 2 are completely
independent, and if you describe both of them, you would have separate
"pcie@62000000" stanzas with separate 'reg' and 'ranges' properties.
To be precise, for two cores of a cadence IP, each core has a separate
set of configuration registers, that is, the configuration of each core
is completely independent. This is also what I meant in the binding by
"Each core corresponds to a host bridge, and each host bridge has only
one root port. Their configuration registers are completely
independent.". Maybe the "Their" here is a bit unclear. My original
intention was to refer to the core. I can improve this description next
time.
From the driver, it does not look like the registers for Link0 and
Link1 are independent, since the driver claims the
"sophgo,sg2042-pcie-host", which includes two Cores, and it tests
pcie->link_id to select the correct register address and bit mask.
In the driver code, one "sophgo,sg2042-pcie-host" corresponds to one
core, not two. So, you can see in patch 4 of this pathset [1], 3 pcie
host-bridge nodes are defined, pcie_rc0 ~ pcie_rc2, each corresponding
to one core.
[1]:https://lore.kernel.org/linux-riscv/4a1f23e5426bfb56cad9c07f90d4efaad5eab976.1736923025.git.unicorn_wang@xxxxxxxxxxx/
I also need to explain that link0 and link1 are actually completely
independent in PCIE processing, but when sophgo implements the internal
msi controller for PCIE, its design is not good enough, and the
registers for processing msi are not made separately for link0 and
link1, but mixed together, which is what I said
cdns_pcie0_ctrl/cdns_pcie1_ctrl. In these two new register files added
by sophgo (only involving MSI processing), take the second cadence IP as
an example, some registers are used to control the msi controller of
pcie_rc1 (corresponding to link0), and some registers are used to
control the msi controller of pcie_rc2 (corresponding to link1). In a
more complicated situation, some bits in a register are used to control
pcie_rc1, and some bits are used to control pcie_rc2. This is why I have
to add the link_id attribute to know whether the current PCIe host
bridge corresponds to link0 or link1, so that when processing the msi
controller related to this pcie host bridge, we can find the
corresponding register or even the bit of a register in cdns_pcieX_ctrl.
"sophgo,link-id" corresponds to Cadence documentation, but I think it
is somewhat misleading in the binding because a PCIe "Link" refers to
the downstream side of a Root Port. If we use "link-id" to identify
either Core0 or Core1 of a Cadence IP, it sort of bakes in the
idea that there can never be more than one Root Port per Core.
The fact is that for the cadence IP used by sg2042, only one root port
is supported per core.
Since each Core is the root of a separate PCI hierarchy, it seems like
maybe there should be a stanza for the Core so there's a place where
per-hierarchy things like "linux,pci-domain" properties could go,
e.g.,
pcie@62000000 { // IP 1, single-link mode
compatible = "sophgo,sg2042-pcie-host";
reg = <...>;
ranges = <...>;
core0 {
sophgo,core-id = <0>;
linux,pci-domain = <0>;
port {
num-lanes = <4>; // all lanes
};
};
};
pcie@82000000 { // IP 2, dual-link mode
compatible = "sophgo,sg2042-pcie-host";
reg = <...>;
ranges = <...>;
core0 {
sophgo,core-id = <0>;
linux,pci-domain = <1>;
port {
num-lanes = <2>; // half of lanes
};
};
core1 {
sophgo,core-id = <1>;
linux,pci-domain = <2>;
port {
num-lanes = <2>; // half of lanes
};
};
};
Based on the above analysis, I think the introduction of a three-layer
structure (pcie-core-port) looks a bit too complicated for candence IP.
In fact, the source of the discussion at the beginning of this issue was
whether some attributes should be placed under the host bridge or the
root port. I suggest that adding the root port layer on the basis of the
existing patch may be enough. What do you think?
e.g.,
pcie_rc0: pcie@7060000000 {
compatible = "sophgo,sg2042-pcie-host";
...... // host bride level properties
sophgo,link-id = <0>;
port {
// port level properties
vendor-id = <0x1f1c>;
device-id = <0x2042>;
num-lanes = <4>;
}
};
pcie_rc1: pcie@7062000000 {
compatible = "sophgo,sg2042-pcie-host";
...... // host bride level properties
sophgo,link-id = <0>;
port {
// port level properties
vendor-id = <0x1f1c>;
device-id = <0x2042>;
num-lanes = <2>;
};
};
pcie_rc2: pcie@7062800000 {
compatible = "sophgo,sg2042-pcie-host";
...... // host bride level properties
sophgo,link-id = <0>;
port {
// port level properties
vendor-id = <0x1f1c>;
device-id = <0x2042>;
num-lanes = <2>;
}
};
[......]
+examples:
+ - |
+ #include <dt-bindings/interrupt-controller/irq.h>
+
+ pcie@62000000 {
+ compatible = "sophgo,sg2042-pcie-host";
+ device_type = "pci";
+ reg = <0x62000000 0x00800000>,
+ <0x48000000 0x00001000>;
+ reg-names = "reg", "cfg";
+ #address-cells = <3>;
+ #size-cells = <2>;
+ ranges = <0x81000000 0 0x00000000 0xde000000 0 0x00010000>,
+ <0x82000000 0 0xd0400000 0xd0400000 0 0x0d000000>;
+ bus-range = <0x00 0xff>;
+ vendor-id = <0x1f1c>;
+ device-id = <0x2042>;
+ cdns,no-bar-match-nbits = <48>;
+ sophgo,link-id = <0>;
+ sophgo,syscon-pcie-ctrl = <&cdns_pcie1_ctrl>;
+ msi-parent = <&msi_pcie>;
+ msi_pcie: msi {
+ compatible = "sophgo,sg2042-pcie-msi";
+ msi-controller;
+ interrupt-parent = <&intc>;
+ interrupts = <123 IRQ_TYPE_LEVEL_HIGH>;
+ interrupt-names = "msi";
+ };
+ };
It would be helpful for me if the example showed how both link-id 0
and link-id 1 would be used (or whatever they end up being named).
I assume both have to be somewhere in the same pcie@62000000 device to
make this work.
Bjorn
