Re: [PATCH v5 00/21] nvmem: core: introduce NVMEM layouts

[Michael Walle]

Hi,

Am 2023-01-05 13:21, schrieb Alexander Stein:
> Am Donnerstag, 5. Januar 2023, 13:11:37 CET schrieb Michael Walle:
> > thanks for debugging. I'm not yet sure what is going wrong, so
> > I have some more questions below.
> >
> > >> This causes the following errors on existing boards (imx8mq-tqma8mq-
> > >> mba8mx.dtb):
> > >> root@tqma8-common:~# uname -r
> > >> 6.2.0-rc2-next-20230105
> > >>
> > >> > OF: /soc@0: could not get #nvmem-cell-cells for /soc@0/bus@30000000/
> > >>
> > >> efuse@30350000/soc-uid@4
> > >>
> > >> > OF: /soc@0/bus@30800000/ethernet@30be0000: could not get
> > >> > #nvmem-cell-cells
> > >>
> > >> for /soc@0/bus@30000000/efuse@30350000/mac-address@90
> > >>
> > >> These are caused because '#nvmem-cell-cells = <0>;' is not explicitly
> > >> set in
> > >> DT.
> > >>
> > >> > TI DP83867 30be0000.ethernet-1:0e: error -EINVAL: failed to get nvmem
> > >> > cell
> > >>
> > >> io_impedance_ctrl
> > >>
> > >> > TI DP83867: probe of 30be0000.ethernet-1:0e failed with error -22
> > >>
> > >> These are caused because of_nvmem_cell_get() now returns -EINVAL
> > >> instead of -
> > >> ENODEV if the requested nvmem cell is not available.
> >
> > What do you mean with not available? Not yet available because of 
> > probe
> > order?
>
> Ah, I was talking about there is no nvmem cell being used in my PHY 
> node, e.g.
> no 'nvmem-cells' nor 'nvmem-cell-names' (set to 'io_impedance_ctrl'). 
> That's
> why of_property_match_string returns -EINVAL.

Ahh I see. You mean ENOENT instead of ENODEV, right?

> > > Should we just assume #nvmem-cell-cells = <0> by default? I guess it's
> > > a safe assumption.
> >
> > Actually, that's what patch 2/21 is for.
> >
> > Alexander, did you verify that the EINVAL is returned by
> > of_parse_phandle_with_optional_args()?
>
> Yep.
>
> --8<--
> diff --git a/drivers/nvmem/core.c b/drivers/nvmem/core.c
> index 1b61c8bf0de4..f2a85a31d039 100644
> --- a/drivers/nvmem/core.c
> +++ b/drivers/nvmem/core.c
> @@ -1339,9 +1339,11 @@ struct nvmem_cell *of_nvmem_cell_get(struct 
> device_node
> *np, const char *id)
>         if (id)
>                 index = of_property_match_string(np, 
> "nvmem-cell-names", id);
>
> +       pr_info("%s: index: %d\n", __func__, index);
>         ret = of_parse_phandle_with_optional_args(np, "nvmem-cells",
>                                                   "#nvmem-cell-cells",
>                                                   index, &cell_spec);
> +       pr_info("%s: of_parse_phandle_with_optional_args: %d\n", 
> __func__,
> ret);
>         if (ret)
>                 return ERR_PTR(ret);
> --8<--
>
> Results in:
> > [    1.861896] of_nvmem_cell_get: index: -22
> > [    1.865934] of_nvmem_cell_get: of_parse_phandle_with_optional_args: 
> > -22
> > [    1.872595] TI DP83867 30be0000.ethernet-1:0e: error -EINVAL: 
> > failed to
> get nvmem cell io_impedance_ctrl
> > [    2.402575] TI DP83867: probe of 30be0000.ethernet-1:0e failed with 
> > error
> -22
>
> So, the index is wrong in the first place, but this was no problem 
> until now.

Thanks, could you try the following patch:

diff --git a/drivers/nvmem/core.c b/drivers/nvmem/core.c
index 1b61c8bf0de4..1085abfcd9b1 100644
--- a/drivers/nvmem/core.c
+++ b/drivers/nvmem/core.c
@@ -1336,8 +1336,11 @@ struct nvmem_cell *of_nvmem_cell_get(struct 
device_node *np, const char *id)
        int ret;

        /* if cell name exists, find index to the name */
-       if (id)
+       if (id) {
                index = of_property_match_string(np, "nvmem-cell-names", 
id);
+               if (index < 0)
+                       return ERR_PTR(-ENOENT);
+       }

        ret = of_parse_phandle_with_optional_args(np, "nvmem-cells",
                                                  "#nvmem-cell-cells",

Before patch 6/21, the -EINVAL was passed as index to of_parse_phandle()
which then returned NULL, which caused the nvmem core to return ENOENT.
I have a vague memory, that I made sure, that
of_parse_phandle_with_optional_args() will also propagate the
wrong index to its return code. But now, it won't be converted
to ENOENT.

-michael