On 5/14/25 11:03, Lei Wei wrote:> On 5/13/2025 6:56 AM, 
mr.nuke.me@gmail.com wrote:

quoted

On 2/19/25 4:46 AM, Lei Wei wrote:

I tried this PCS driver, and I am seeing a circular dependency in the 
clock init. If the clock tree is:
     GCC -> NSSCC -> PCS(uniphy) -> NSSCC -> PCS(mii)

The way I understand it, the UNIPHY probe depends on the MII probe. If 
MII .probe() returns -EPROBE_DEFER, then so will the UNIPHY .probe(). 
But the MII cannot probe until the UNIPHY is done, due to the clock 
dependency. How is it supposed to work?

The way I found to resolve this is to move the probing of the MII 
clocks to ipq_pcs_get().

This is the kernel log that I see:

[   12.008754] platform 39b00000.clock-controller: deferred probe 
pending: platform: supplier 7a00000.ethernet-pcs not ready
[   12.008788] mdio_bus 90000.mdio-1:18: deferred probe pending: 
mdio_bus: supplier 7a20000.ethernet-pcs not ready
[   12.018704] mdio_bus 90000.mdio-1:00: deferred probe pending: 
mdio_bus: supplier 90000.mdio-1:18 not ready
[   12.028588] mdio_bus 90000.mdio-1:01: deferred probe pending: 
mdio_bus: supplier 90000.mdio-1:18 not ready
[   12.038310] mdio_bus 90000.mdio-1:02: deferred probe pending: 
mdio_bus: supplier 90000.mdio-1:18 not ready
[   12.047943] mdio_bus 90000.mdio-1:03: deferred probe pending: 
mdio_bus: supplier 90000.mdio-1:18 not ready
[   12.057579] platform 7a00000.ethernet-pcs: deferred probe pending: 
ipq9574_pcs: Failed to get MII 0 RX clock
[   12.067209] platform 7a20000.ethernet-pcs: deferred probe pending: 
ipq9574_pcs: Failed to get MII 0 RX clock
[   12.077200] platform 3a000000.qcom-ppe: deferred probe pending: 
platform: supplier 39b00000.clock-controller not ready

Hello, thanks for bringing this to our notice. Let me try to understand 
the reason for the probe failure:

The merged NSSCC DTS does not reference the PCS node directly in the 
"clocks" property. It uses a placeholder phandle '<0>' for the 
reference. Please see below patch which is merged.
https://lore.kernel.org/all/20250313110359.242491-6-quic_mmanikan@quicinc.com/ (local)

Ideally there should be no direct dependency from NSSCC to PCS driver if
we use this version of the NSSCC DTS.

Hence it seems that you may have a modified patch here, and DTS changes 
have been applied to enable all the Ethernet components including PCS 
and NSSCC, and NSSCC modified to have a direct reference to PCS? However 
even in this case, I think the driver probe should work if the drivers 
are built as modules. Can you please confirm if the NSSCC and PCS 
drivers are built-in to the kernel and not built as modules

The NSSCC and PCS built-in. I also added the uniphy PCS clocks to the 
NSSCC in order to expose the issue.

I have a heavily patched tree with PPE driver and EDMA support. That's 
the final use case in order to support ethernet, right?

For the case where the drivers are built-in to kernel, and the NSSCC DTS
node has a direct reference to PCS node, we can use the below solution:
[Note that the 'UNIPHY' PCS clocks are not needed for NSSCC clocks
initialization/registration.]

     Enable 'post-init-providers' property in the NSSCC DTS node to mark
    'UNIPHY' PCS as post-initialization providers to NSSCC. This will
     ensure following probe order by the kernel:

     1.) NSSCC driver
     2.) PCS driver.

Please let me know if the above suggestion can help.

I see. Adding the 'post-init-providers' property does fix the circular 
dependency. Thank you!

I have another question. Do you have a public repository with the 
unmerged IPQ9574 patches, including, PCS, PPE, EDMA, QCA8084 ?

Later once the IPQ PCS driver is merged, we are planning to push the PCS 
DTS changes, along with an update of the NSSCC DTS to point to the PCS 
node and mark the "post-init-providers" property. This should work for 
all cases.

Also, in my view, it is not suitable to move PCS MII clocks get to
"ipq_pcs_get()" because the natural loading order for the drivers
is as below:

1) NSSCC driver
2) PCS driver
3) Ethernet driver.

Additionally, the community is currently working on an infrastructure to
provide a common pcs get method. (Christian and Sean Anderson has been 
working on this). Therefore, I expect "ipq_pcs_get" to be dropped in the 
future and replaced with the common pcs get method once this common 
infra is merged.

That makes sense. Thank you for clarifying.