On Mon, 30 Sep 2024 13:57:01 +0000
"Arnd Bergmann" [off-list ref] wrote:

On Mon, Sep 30, 2024, at 12:15, Herve Codina wrote:

quoted

In the LAN966x PCI device use case, syscon cannot be used as syscon
devices do not support removal [1]. A syscon device is a core "system"
device and not a device available in some addon boards and so, it is not
supposed to be removed.

In order to remove the syscon usage, use a local mapping of a reg
address range when cpu-syscon is not present.

Link: https://lore.kernel.org/all/20240923100741.11277439@bootlin.com/ (local) [1]
Signed-off-by: Herve Codina <herve.codina@bootlin.com>
---  

quoted

quoted

 	err = mchp_sparx5_map_syscon(pdev, "cpu-syscon", &ctx->cpu_ctrl);  

-	if (err)
+	switch (err) {
+	case 0:
+		break;
+	case -ENODEV:  

I was expecting a patch that would read the phandle and map the
syscon node to keep the behavior unchanged, but I guess this one
works as well.

The downside of your approach is that it requires an different
DT binding, which only works as long as there are no other
users of the syscon registers.

Yes, I knwow but keeping the binding with the syscon device (i.e. compatible
= "...", "syscon";) leads to confusion.
Indeed, the syscon API cannot be used because using this API leads issues
when the syscon device is removed.
That means the you have a "syscon" node (compatible = "syscon") but we cannot
use the syscon API (include/linux/mfd/syscon.h) with this node.

Also, in order to share resources between several consumers of the "syscon"
registers, we need exactly what is done in syscon. I mean we need to map
resources only once, provide this resource throught a regmap an share this
regmap between the consumers. Indeed a lock needs to be shared in order to
protect against registers RMW accesses done by several consumers.
In other word, we need to copy/paste syscon code with support for removal
implemented (feature needed in the LAN966x PCI device use case).

So, I found really simpler and less confusing to fully discard the syscon node
and handle registers directly in the only one consumer.

With all of these, do you thing my approach can be acceptable ?

Best regards,
Hervé