Re: [PATCH v13 4/6] nvmem: core: Rework layouts to become regular devices

From: Miquel Raynal
Date: Fri Nov 24 2023 - 14:25:18 EST


Hi Marco,

m.felsch@xxxxxxxxxxxxxx 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:
> > Hi Miquel,
> >
> > thanks a lot for your effort on this. Please see my comments inline.
> >
> > On 23-10-11, Miquel Raynal wrote:
> > > 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.

> > > - 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.

> > > 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@xxxxxxxxxxx>
> > > Tested-by: Rafał Miłecki <rafal@xxxxxxxxxx>
>
> ...
>
> > > @@ -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.

>
> > > -
> > > 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.

...

> >
> > > - 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