śr., 27 lip 2022 o 23:15 Andy Shevchenko [off-list ref] napisał(a):

On Wednesday, July 27, 2022, Marcin Wojtas [off-list ref] wrote:

quoted

śr., 27 lip 2022 o 18:38 Vladimir Oltean [off-list ref] napisał(a):

quoted

On Wed, Jul 27, 2022 at 05:18:16PM +0200, Marcin Wojtas wrote:

quoted

Do you mean a situation analogous to what I addressed in:
[net-next: PATCH v3 4/8] net: mvpp2: initialize port fwnode pointer
?

Not sure if "analogous" is the right word. My estimation is that the
overwhelmingly vast majority of DSA masters can be found by DSA simply
due to the SET_NETDEV_DEV() call that the Ethernet drivers need to make
anyway.  I see that mvpp2 also needed commit c4053ef32208 ("net: mvpp2:
initialize port of_node pointer"), but that isn't needed in general, and
I can't tell you exactly why it is needed there, I don't know enough
about the mvpp2 driver.

SET_NETDEV_DEV() fills net_device->dev.parent with &pdev->dev
and in most cases it is sufficient apparently it is sufficient for
fwnode_find_parent_dev_match (at least tests with mvneta case proves
it's fine).

We have some corner cases though:
* mvpp2 -> single controller can handle up to 3 net_devices and
therefore we need device_set_node() to make this work. I think dpaa2
is a similar case
* PCIE drivers with extra DT description (I think that's the case of enetc).

quoted

quoted

I found indeed a couple of drivers that may require a similar change
(e.g. dpaa2).

There I can tell you why the dpaa2-mac code mangles with net_dev->dev.of_node,
but I'd rather not go into an explanation that essentially doesn't matter.
The point is that you'd be mistaken to think that only the drivers which
touch the net device's ->dev->of_node are the ones that need updating
for your series to not cause regressions.

As above - SET_NETDEV_DEV() should be fine in most cases, but we can
never be 100% sure untils it's verified.

quoted

quoted

IMO we have 2 options:
- update these drivers
- add some kind of fallback? If yes, I am wondering about an elegant
solution - maybe add an extra check inside
fwnode_find_parent_dev_match?

What would you suggest?

Fixing fwnode_find_parent_dev_match(), of course.

Fixing how?!

quoted

This change broke DSA

quoted

on my LS1028A system (master in drivers/net/ethernet/freescale/enetc/)
and LS1021A (master in drivers/net/ethernet/freescale/gianfar.c).

Can you please check applying following diff:

--- a/drivers/base/property.c
+++ b/drivers/base/property.c

@@ -695,20 +695,22 @@ EXPORT_SYMBOL_GPL(fwnode_get_nth_parent);
  * The routine can be used e.g. as a callback for class_find_device().
  *
  * Returns: %1 - match is found
  *          %0 - match not found
  */
 int fwnode_find_parent_dev_match(struct device *dev, const void *data)
 {
        for (; dev; dev = dev->parent) {
                if (device_match_fwnode(dev, data))
                        return 1;
+               else if (device_match_of_node(dev, to_of_node(data))
+                       return 1;
        }

This adds a piece of dead code. device_match_fwnode() covers this already.

Yes, indeed. After recent update, I think we can assume the current
implementation of fwnode_find_parent_dev_match should work fine with
all existing cases.

Thank you for all remarks and comments, I'll address them in v4 later today.

Best regards,
Marcin