Hi Rafał,
rafal@xxxxxxxxxx wrote on Thu, 09 Mar 2023 09:39:54 +0100:
> On 2023-03-09 09:34, Miquel Raynal wrote:
> > Hi Rafał,
> >
> > zajec5@xxxxxxxxx wrote on Thu, 9 Mar 2023 07:56:05 +0100:
> >
> >> On 8.03.2023 19:31, Miquel Raynal wrote:
> >> > Hi Rafał,
> >> >
> >> > rafal@xxxxxxxxxx wrote on Wed, 08 Mar 2023 19:12:32 +0100:
> >> >
> >> >> On 2023-03-08 19:06, Miquel Raynal wrote:
> >> >>> Hi Rafał,
> >> >>>
> >> >>> rafal@xxxxxxxxxx wrote on Wed, 08 Mar 2023 17:55:46 +0100:
> >> >>> >>>> On 2023-03-08 17:34, Miquel Raynal wrote:
> >> >>>>> Hi Rafał,
> >> >>>>>
> >> >>>>> zajec5@xxxxxxxxx wrote on Fri, 24 Feb 2023 08:29:03 +0100:
> >> >>>>> >>>>>> From: Rafał Miłecki <rafal@xxxxxxxxxx>
> >> >>>>>>>> NVMEM subsystem looks for fixed NVMEM cells (specified in DT) by
> >> >>>>>> default. This behaviour made sense in early days before adding support
> >> >>>>>> for dynamic cells.
> >> >>>>>>>> With every new supported NVMEM device with dynamic cells current
> >> >>>>>> behaviour becomes non-optimal. It results in unneeded iterating over >> DT
> >> >>>>>> nodes and may result in false discovery of cells (depending on used DT
> >> >>>>>> properties).
> >> >>>>>>>> This behaviour has actually caused a problem already with the MTD
> >> >>>>>> subsystem. MTD subpartitions were incorrectly treated as NVMEM cells.
> >> >>>>>
> >> >>>>> That's true, but I expect this to be really MTD specific.
> >> >>>>>
> >> >>>>> A concrete proposal below.
> >> >>>>> >>>>>> Also with upcoming support for NVMEM layouts no new binding or driver
> >> >>>>>> should support fixed cells defined in device node.
> >> >>>>>
> >> >>>>> I'm not sure I agree with this statement. We are not preventing new
> >> >>>>> binding/driver to use fixed cells, or...? We offer a new way to expose
> >> >>>>> nvmem cells with another way than "fixed-offset" and "fixed-size" OF
> >> >>>>> nodes.
> >> >>>>>> From what I understood all new NVMEM bindings should have cells >> defined
> >> >>>> in the nvmem-layout { } node. That's what I mean by saying they should
> >> >>>> not be defined in device node (but its "nvmem-layout" instead).
> >> >>>
> >> >>> Layouts are just another possibility, either you user the nvmem-cells
> >> >>> compatible and produce nvmem cells with fixed OF nodes, or you use the
> >> >>> nvmem-layout container. I don't think all new bindings should have
> >> >>> cells in layouts. It depends if the content is static or not.
> >> >>> >>>>>> Solve this by modifying drivers for bindings that support specifying
> >> >>>>>> fixed NVMEM cells in DT. Make them explicitly tell NVMEM subsystem to
> >> >>>>>> read cells from DT.
> >> >>>>>>>> It wasn't clear (to me) if rtc and w1 code actually uses fixed cells. >> I
> >> >>>>>> enabled them to don't risk any breakage.
> >> >>>>>>>> Signed-off-by: Rafał Miłecki <rafal@xxxxxxxxxx>
> >> >>>>>> [for drivers/nvmem/meson-{efuse,mx-efuse}.c]
> >> >>>>>> Acked-by: Martin Blumenstingl <martin.blumenstingl@xxxxxxxxxxxxxx>
> >> >>>>>> ---
> >> >>>>>> V2: Fix stm32-romem.c typo breaking its compilation
> >> >>>>>> Pick Martin's Acked-by
> >> >>>>>> Add paragraph about layouts deprecating use_fixed_of_cells
> >> >>>>>> ---
> >> >>>>>> drivers/mtd/mtdcore.c | 2 ++
> >> >>>>>> drivers/nvmem/apple-efuses.c | 1 +
> >> >>>>>> drivers/nvmem/core.c | 8 +++++---
> >> >>>>>> drivers/nvmem/imx-ocotp-scu.c | 1 +
> >> >>>>>> drivers/nvmem/imx-ocotp.c | 1 +
> >> >>>>>> drivers/nvmem/meson-efuse.c | 1 +
> >> >>>>>> drivers/nvmem/meson-mx-efuse.c | 1 +
> >> >>>>>> drivers/nvmem/microchip-otpc.c | 1 +
> >> >>>>>> drivers/nvmem/mtk-efuse.c | 1 +
> >> >>>>>> drivers/nvmem/qcom-spmi-sdam.c | 1 +
> >> >>>>>> drivers/nvmem/qfprom.c | 1 +
> >> >>>>>> drivers/nvmem/rave-sp-eeprom.c | 1 +
> >> >>>>>> drivers/nvmem/rockchip-efuse.c | 1 +
> >> >>>>>> drivers/nvmem/sc27xx-efuse.c | 1 +
> >> >>>>>> drivers/nvmem/sprd-efuse.c | 1 +
> >> >>>>>> drivers/nvmem/stm32-romem.c | 1 +
> >> >>>>>> drivers/nvmem/sunplus-ocotp.c | 1 +
> >> >>>>>> drivers/nvmem/sunxi_sid.c | 1 +
> >> >>>>>> drivers/nvmem/uniphier-efuse.c | 1 +
> >> >>>>>> drivers/nvmem/zynqmp_nvmem.c | 1 +
> >> >>>>>> drivers/rtc/nvmem.c | 1 +
> >> >>>>>> drivers/w1/slaves/w1_ds250x.c | 1 +
> >> >>>>>> include/linux/nvmem-provider.h | 2 ++
> >> >>>>>> 23 files changed, 29 insertions(+), 3 deletions(-)
> >> >>>>>>>> diff --git a/drivers/mtd/mtdcore.c b/drivers/mtd/mtdcore.c
> >> >>>>>> index 0feacb9fbdac..1bb479c0f758 100644
> >> >>>>>> --- a/drivers/mtd/mtdcore.c
> >> >>>>>> +++ b/drivers/mtd/mtdcore.c
> >> >>>>>> @@ -523,6 +523,7 @@ static int mtd_nvmem_add(struct mtd_info *mtd)
> >> >>>>>> config.dev = &mtd->dev;
> >> >>>>>> config.name = dev_name(&mtd->dev);
> >> >>>>>> config.owner = THIS_MODULE;
> >> >>>>>> + config.use_fixed_of_cells = of_device_is_compatible(node, >> "nvmem-cells");
> >> >>>>>
> >> >>>>> I am wondering how mtd specific this is? For me all OF nodes containing
> >> >>>>> the nvmem-cells compatible should be treated as cells providers and
> >> >>>>> populate nvmem cells as for each children.
> >> >>>>>
> >> >>>>> Why don't we just check for this compatible to be present? in
> >> >>>>> nvmem_add_cells_from_of() ? And if not we just skip the operation.
> >> >>>>>
> >> >>>>> This way we still follow the bindings (even though using nvmem-cells in
> >> >>>>> the compatible property to require cells population was a mistake in
> >> >>>>> the first place, as discussed in the devlink thread recently) but there
> >> >>>>> is no need for a per-driver config option?
> >> >>>>>> This isn't mtd specific. Please check this patch for all occurrences >> of
> >> >>>> the:
> >> >>>> use_fixed_of_cells = true
> >> >>>>>> The very first one: drivers/nvmem/apple-efuses.c driver for the
> >> >>>> "apple,efuses" binding. That binding supports fixed OF cells, see:
> >> >>>> Documentation/devicetree/bindings/nvmem/apple,efuses.yaml
> >> >>>
> >> >>> I'm saying: based on what has been enforced so far, I would expect all
> >> >>> fixed cell providers to come with nvmem-cells as compatible, no?
> >> >>>
> >> >>> If that's the case we could use that as a common denominator?
> >> >>
> >> >> Sorry, I don't get it. Have you checked
> >> >> Documentation/devicetree/bindings/nvmem/apple,efuses.yaml
> >> >> ?
> >> >>
> >> >> It's a NVMEM provied binding with fixed cells that doesn't use
> >> >> nvmem-cells as compatible. There are many more.
> >> >
> >> > Oh yeah you're right, I'm mixing things. Well I guess you're right
> >> > then, it's such a mess, we have to tell the core the parsing method.
> >> >
> >> > So maybe another question: do we have other situations than mtd which
> >> > sometimes expect the nvmem core to parse the OF nodes to populate cells,
> >> > and sometimes not?
> >> >> I'm not aware of that. Please also check my patch. The only case I set
> >> "use_fixed_of_cells" conditionally is mtd code. In other cases it's
> >> hardcoded to "true".
> >> >> >> > Also, what about "of_children_are_cells" ? Because actually in most
> >> > cases it's a "fixed of cell", so I don't find the current naming
> >> > descriptive enough for something so touchy.
> >> >> That would be just incorrect because this new config property
> >> ("use_fixed_of_cells") is only about FIXED cells.
> >> >> There are cases of OF children being cells but NOT being fixed cells.
> >> They should NOT be parsed by the nvmem_add_cells_from_of().
> >> >> Example:
> >> a607a850ba1f ("dt-bindings: nvmem: u-boot,env: add basic NVMEM cells")
> >> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=a607a850ba1fa966cbb035544c1588e24a6307df
> >
> > This is backwards. That's why layouts have been proposed: having
> > a clear framework were the nvmem core should or should not play with
> > the OF children nodes. If each binding is different, you end up with
> > the mess we have today, where nobody knows how to properly
> > populate the cells.
> >
> > Anyway, it's not a big deal either, if the cells are empty we can
> > easily check that and have *yet another* specific case in the core.
> >
> >> So that would result in U-Boot env:
> >> 1. Having OF children nodes being cells
> >> 2. Setting "of_children_are_cells" to false (counter-intuitive) to >> avoid nvmem_add_cells_from_of()
> >
> > Agreed. So what about config.ignore_of_children?
> > - mtd sets this to !is_compatible(nvmem-cells)
> > - nobody else touches it (the core don't try to parse the cell if it's
> > empty so U-Boot env driver works)
>
> "ignore_of_children" would have opposite (reversed) meaning to the
> "use_fixed_of_cells":
> 1. By default it would be 0 / false
> 2. By default NVMEM code would NOT ignore OF children nodes
>
> That is what I actually *don't* want.
>
> Having NVMEM core look for fixed cells in device node is undesirable.
Is it?
I think this is the standard case. Most of the time fixed cells are
the simplest and most direct approach. I don't get why we would like
to prevent it at some point? Only the more advanced stuff that does not
fit the purpose of fixed OF cells should go through layouts.
> I want that feature to be off by default. I want devices to have to
> enable it explicitly only when it's needed.
Well, that's exactly the opposite of what nvmem does right now, that's
maybe why reviewers don't get the interest of the current
implementation: it has many impacts on users which should not be
impacted just because we messed with mtd.
Thanks,
Miquèl
