Hi Marco,

m.felsch@pengutronix.de wrote on Wed, 22 Nov 2023 23:45:53 +0100:

Hi Miquel,

sorry for answering to my own mail, I forgot something I noticed later.

No problem :)

On 23-11-22, Marco Felsch wrote:

quoted

Hi Miquel,

thanks a lot for your effort on this. Please see my comments inline.

On 23-10-11, Miquel Raynal wrote:  

quoted

Current layout support was initially written without modules support in
mind. When the requirement for module support rose, the existing base
was improved to adopt modularization support, but kind of a design flaw
was introduced. With the existing implementation, when a storage device
registers into NVMEM, the core tries to hook a layout (if any) and
populates its cells immediately. This means, if the hardware description
expects a layout to be hooked up, but no driver was provided for that,
the storage medium will fail to probe and try later from
scratch. Even if we consider that the hardware description shall be
correct, we could still probe the storage device (especially if it
contains the rootfs).

One way to overcome this situation is to consider the layouts as
devices, and leverage the existing notifier mechanism. When a new NVMEM
device is registered, we can:
- populate its nvmem-layout child, if any
- try to modprobe the relevant driver, if relevant  

I'm not sure why we call of_request_module() the driver framework should
handle that right?

Actually that's right, it is no longer needed, we would expect udev to
do that now. Thanks for the pointer.

quoted

quoted

- try to hook the NVMEM device with a layout in the notifier  

The last part is no longer true since you don't use the notifier
anymore.

True, I've re-written this part.

quoted

quoted

And when a new layout is registered:
- try to hook all the existing NVMEM devices which are not yet hooked to
  a layout with the new layout
This way, there is no strong order to enforce, any NVMEM device creation
or NVMEM layout driver insertion will be observed as a new event which
may lead to the creation of additional cells, without disturbing the
probes with costly (and sometimes endless) deferrals.

In order to achieve that goal we need:
* To keep track of all nvmem devices
* To create a new bus for the nvmem-layouts with minimal logic to match
  nvmem-layout devices with nvmem-layout drivers.
All this infrastructure code is created in the layouts.c file.

Signed-off-by: Miquel Raynal <miquel.raynal@bootlin.com>
Tested-by: Rafał Miłecki <rafal@milecki.pl>  

...

quoted

quoted

@@ -944,19 +872,6 @@ struct nvmem_device *nvmem_register(const struct nvmem_config *config)
 			goto err_put_device;
 	}
 
-	/*
-	 * If the driver supplied a layout by config->layout, the module
-	 * pointer will be NULL and nvmem_layout_put() will be a noop.
-	 */
-	nvmem->layout = config->layout ?: nvmem_layout_get(nvmem);
-	if (IS_ERR(nvmem->layout)) {
-		rval = PTR_ERR(nvmem->layout);
-		nvmem->layout = NULL;
-
-		if (rval == -EPROBE_DEFER)
-			goto err_teardown_compat;
-	}  

Since this logic will be gone and the layout became a device the fixup
hook for the layout is more than confusing. E.g. the imx-ocotp driver
uses the layout to register a fixup for a cell which is fine but the
hook should be moved from the layout[-dev] to the config. Please see
below.

That is true.

quoted

quoted

-
 	if (config->cells) {
 		rval = nvmem_add_cells(nvmem, config->cells, config->ncells);
 		if (rval)

@@ -975,7 +890,7 @@ struct nvmem_device *nvmem_register(const struct nvmem_config *config)
 	if (rval)
 		goto err_remove_cells;
 
-	rval = nvmem_add_cells_from_layout(nvmem);
+	rval = nvmem_populate_layout(nvmem);
 	if (rval)
 		goto err_remove_cells;  

Also why do we populate the nvmem-layout device infront of the nvmem
device?

I'm not sure I get the question, there is nothing abnormal that stands
out to my eyes.

...

quoted

quoted

-	const char *name;
-	const struct of_device_id *of_match_table;
+	struct device dev;
+	struct nvmem_device *nvmem;
 	int (*add_cells)(struct device *dev, struct nvmem_device *nvmem,
 			 struct nvmem_layout *layout);
 	void (*fixup_cell_info)(struct nvmem_device *nvmem,
 				struct nvmem_layout *layout,
 				struct nvmem_cell_info *cell);  

I speak about this hook. This should be moved into the config, maybe
also renamed to fixup_dt_cell_info() or so to not confuse the users. If
we move that hook and remove the add_cells hook there are only two
members left for the cross-link.

It's not a bad idea, I've included this change in my series (for v14,
sic). I like your rename as well. Thanks for the hint.

Thanks,
Miquèl