Re: [PATCH v2 06/29] mtd: Add support for reading MTD devices via the nvmem API

From: Alban
Date: Tue Aug 21 2018 - 09:58:36 EST


On Tue, 21 Aug 2018 14:57:25 +0200
Boris Brezillon <boris.brezillon@xxxxxxxxxxx> wrote:

> On Tue, 21 Aug 2018 14:27:16 +0200
> Alban <albeu@xxxxxxx> wrote:
>
> > On Tue, 21 Aug 2018 07:44:04 +0200
> > Boris Brezillon <boris.brezillon@xxxxxxxxxxx> wrote:
> >
> > > On Tue, 21 Aug 2018 00:53:27 +0200
> > > Alban <albeu@xxxxxxx> wrote:
> > >
> > > > On Sun, 19 Aug 2018 18:46:09 +0200
> > > > Boris Brezillon <boris.brezillon@xxxxxxxxxxx> wrote:
> > > >
> > > > > On Sun, 19 Aug 2018 13:31:06 +0200
> > > > > Alban <albeu@xxxxxxx> wrote:
> > > > >
> > > > > > On Fri, 17 Aug 2018 18:27:20 +0200
> > > > > > Boris Brezillon <boris.brezillon@xxxxxxxxxxx> wrote:
> > > > > >
> > > > > > > Hi Bartosz,
> > > > > > >
> > > > > > > On Fri, 10 Aug 2018 10:05:03 +0200
> > > > > > > Bartosz Golaszewski <brgl@xxxxxxxx> wrote:
> > > > > > >
> > > > > > > > From: Alban Bedel <albeu@xxxxxxx>
> > > > > > > >
> > > > > > > > Allow drivers that use the nvmem API to read data
> > > > > > > > stored on MTD devices. For this the mtd devices are
> > > > > > > > registered as read-only NVMEM providers.
> > > > > > > >
> > > > > > > > Signed-off-by: Alban Bedel <albeu@xxxxxxx>
> > > > > > > > [Bartosz:
> > > > > > > > - use the managed variant of nvmem_register(),
> > > > > > > > - set the nvmem name]
> > > > > > > > Signed-off-by: Bartosz Golaszewski
> > > > > > > > <bgolaszewski@xxxxxxxxxxxx>
> > > > > > >
> > > > > > > What happened to the 2 other patches of Alban's series?
> > > > > > > I'd really like the DT case to be handled/agreed on in
> > > > > > > the same patchset, but IIRC, Alban and Srinivas disagreed
> > > > > > > on how this should be represented. I hope this time we'll
> > > > > > > come to an agreement, because the MTD <-> NVMEM glue has
> > > > > > > been floating around for quite some time...
> > > > > >
> > > > > > These other patches were to fix what I consider a
> > > > > > fundamental flaw in the generic NVMEM bindings, however we
> > > > > > couldn't agree on this point. Bartosz later contacted me to
> > > > > > take over this series and I suggested to just change the
> > > > > > MTD NVMEM binding to use a compatible string on the NVMEM
> > > > > > cells as an alternative solution to fix the clash with the
> > > > > > old style MTD partition.
> > > > > >
> > > > > > However all this has no impact on the code needed to add
> > > > > > NVMEM support to MTD, so the above patch didn't change at
> > > > > > all.
> > > > >
> > > > > It does have an impact on the supported binding though.
> > > > > nvmem->dev.of_node is automatically assigned to
> > > > > mtd->dev.of_node, which means people will be able to define
> > > > > their NVMEM cells directly under the MTD device and reference
> > > > > them from other nodes (even if it's not documented), and as
> > > > > you said, it conflict with the old MTD partition bindings. So
> > > > > we'd better agree on this binding before merging this
> > > > > patch.
> > > >
> > > > Unless the nvmem cell node has a compatible string, then it
> > > > won't be considered as a partition by the MTD code. That is
> > > > were the clash is, both bindings allow free named child nodes
> > > > without a compatible string.
> > >
> > > Except the current nvmem cells parsing code does not enforce
> > > that, and existing DTs rely on this behavior, so we're screwed.
> > > Or are you suggesting to add a new "bool check_cells_compat;"
> > > field to nvmem_config?
> >
> > There is no nvmem cell parsing at the moment. The DT lookup just
> > resolve the phandle to the cell node, take the parent node and
> > search for the nvmem provider that has this OF node. So extending
> > it in case the node has a *new* compatible string would not break
> > users of the old binding, none of them has a compatible string.
>
> But we want to enforce the compat check on MTD devices, otherwise old
> MTD partitions (those defined under the MTD node) will be considered
> as NVMEM cells by the NVMEM framework. Hence the bool
> check_cells_compat field.

That would only be needed if the NVMEM framework would do "forward"
parsing, creating data structure for each NVMEM cell found under an
NVMEM provider. However currently it doesn't do that and only goes
"backward", starting by resolving a phandle pointing to a cell, then
finding the provider that the cell belongs to.

This also has the side effect that nvmem cells defined in DT don't
appear in sysfs, unlike those defined from board code.

> >
> > > >
> > > > > I see several options:
> > > > >
> > > > > 1/ provide a way to tell the NVMEM framework not to use
> > > > > parent->of_node even if it's != NULL. This way we really don't
> > > > > support defining NVMEM cells in the DT, and also don't support
> > > > > referencing the nvmem device using a phandle.
> > > >
> > > > I really don't get what the point of this would be. Make the
> > > > whole API useless?
> > >
> > > No, just allow Bartosz to get his changes merged without waiting
> > > for you and Srinivas to agree on how to handle the new binding.
> > > As I said earlier, this mtd <-> nvmem stuff has been around for
> > > quite some time, and instead of trying to find an approach that
> > > makes everyone happy, you decided to let the patchset die.
> >
> > As long as that wouldn't prevent using DT in the future I'm fine
> > with it.
> >
> > > >
> > > > > 2/ define a new binding where all nvmem-cells are placed in an
> > > > > "nvmem" subnode (just like we have this "partitions"
> > > > > subnode for partitions), and then add a config->of_node field
> > > > > so that the nvmem provider can explicitly specify the DT node
> > > > > representing the nvmem device. We'll also need to set this
> > > > > field to ERR_PTR(-ENOENT) in case this node does not exist so
> > > > > that the nvmem framework knows that it should not assign
> > > > > nvmem->dev.of_node to parent->of_node
> > > >
> > > > This is not good. First the NVMEM device is only a virtual
> > > > concept of the Linux kernel, it has no place in the DT.
> > >
> > > nvmem-cells is a virtual concept too, still, you define them in
> > > the DT.
> >
> > To be honest I also think that naming this concept "nvmem" in the
> > DT was a bad idea. Perhaps something like "driver-data" or
> > "data-cell" would have been better as that would make it clear what
> > this is about, nvmem is just the Linux implementation of this
> > concept.
>
> I'm fine using a different name.
>
> >
> > > > Secondly the NVMEM
> > > > provider (here the MTD device) then has to manually parse its DT
> > > > node to find this subnode, pass it to the NVMEM framework to
> > > > later again resolve it back to the MTD device.
> > >
> > > We don't resolve it back to the MTD device, because the MTD
> > > device is just the parent of the nvmem device.
> > >
> > > > Not very complex but still a lot of
> > > > useless code, just registering the MTD device is a lot simpler
> > > > and much more inline with most other kernel API that register a
> > > > "service" available from a device.
> > >
> > > I'm not a big fan of this option either, but I thought I had to
> > > propose it.
> > >
> > > >
> > > > > 3/ only declare partitions as nvmem providers. This would
> > > > > solve the problem we have with partitions defined in the DT
> > > > > since defining sub-partitions in the DT is not (yet?)
> > > > > supported and partition nodes are supposed to be leaf nodes.
> > > > > Still, I'm not a big fan of this solution because it will
> > > > > prevent us from supporting sub-partitions if we ever
> > > > > want/need to.
> > > >
> > > > That sound like a poor workaround.
> > >
> > > Yes, that's a workaround. And the reason I propose it, is, again,
> > > because I don't want to block Bartosz.
> > >
> > > > Remember that this problem could
> > > > appear with any device that has a binding that use child
> > > > nodes.
> > >
> > > I'm talking about partitions, and you're talking about mtd
> > > devices. Right now partitions don't have subnodes, and if we
> > > define that partition subnodes should describe nvmem-cells, then
> > > it becomes part of the official binding. So, no, the problem you
> > > mention does not (yet) exist.
> >
> > That would add another binding that allow free named child nodes
> > without compatible string although experience has repeatedly shown
> > that this was a bad idea.
>
> Yes, I agree. Just thought it was important to have this solution in
> the list, even if it's just to reject it.
>
> >
> > > >
> > > > > 4/ Add a ->of_xlate() hook that would be called if present by
> > > > > the framework instead of using the default parsing we have
> > > > > right now.
> > > >
> > > > That is a bit cleaner, but I don't think it would be worse the
> > > > complexity.
> > >
> > > But it's way more flexible than putting everything in the nvmem
> > > framework. BTW, did you notice that nvmem-cells parsing does not
> > > work with flashes bigger than 4GB, because the framework assumes
> > > #address-cells and #size-cells are always 1. That's probably
> > > something we'll have to fix for the MTD case.
> >
> > Yes, however that's just an implementation limitation which is
> > trivial to solve.
>
> Agree. I was just pointing it in case you hadn't noticed.
>
> >
> > > > Furthermore xlate functions are more about converting
> > > > from hardware parameters to internal kernel representation than
> > > > to hide extra DT parsing.
> > >
> > > Hm, how is that different? ->of_xlate() is just a way for drivers
> > > to have their own DT representation, which is exactly what we
> > > want here.
> >
> > There is a big difference. DT represent the hardware and the
> > relationship between the devices in an OS independent format. We
> > don't add extra stuff in there just to map back internal Linux API
> > details.
>
> And I'm not talking about adding SW information in the DT, I'm talking
> about HW specific description. We have the same solution for pinctrl
> configs (it's HW/driver specific).

For pinctrl I do understand, these beast can be very different from SoC
to SoC, having a single biding for all doesn't make much sense.

However here we are talking about a simple linear storage, nothing
special at all. I could see the need for an xlate to for example
support a device with several partitions, but not to just allow each
driver to have slightly incompatible bindings.

> >
> > > >
> > > > > 5/ Tell the nvmem framework the name of the subnode containing
> > > > > nvmem cell definitions (if NULL that means cells are directly
> > > > > defined under the nvmem provider node). We would set it to
> > > > > "nvmem-cells" (or whatever you like) for the MTD case.
> > > >
> > > > If so please match on compatible and not on the node name.
> > >
> > > If you like.
> > >
> > > >
> > > > 6/ Extend the current NVMEM cell lookup to check if the parent
> > > > node of the cell has a compatible string set to "nvmem-cells".
> > > > If it doesn't it mean we have the current binding and this node
> > > > is the NVMEM device. If it does the device node is just the
> > > > next parent. This is trivial to implement (literally 2 lines of
> > > > code) and cover all the cases currently known.
> > >
> > > Except Srinivas was not happy with this solution, and this
> > > stalled the discussion. I'm trying to find other options and you
> > > keep rejecting all of them to come back to this one.
> >
> > Well, I think this is the best solution :/
> >
> > > >
> > > > 7/ Just add a compatible string to the nvmem cell. No code
> > > > change is needed,
> > >
> > > That's not true!!!
> >
> > What is not true in this statement? The current nvmem lookup don't
> > care about compatible strings, so the cell lookup would just works
> > fine. The MTD partition parser won't consider them as a partition
> > because of the compatible string. Problem solved!
>
> No because partitions defined the old way (as direct subnodes of the
> MTD node) will be considered as NVMEM cells by the NVMEM framework,
> and I don't want that.

As I explained above that is not currently the case. If the NVMEM,
framework is ever changed to explicitly parse NVMEM cells in advance
we can first update the few existing users to add the compatible string.

> Plus, I don't want people to start defining their NVMEM cells and
> forget the compat string (which would work just fine because the
> NVMEM framework doesn't care).

A review of a new DTS should check that it use each binding correctly,
AFAIK the DT people do that. We could also add a warning when there is
no compatible string, that would also help pushing people to update
their DTS.

> >
> > > What forces people to add this compatible in their
> > > DT? Nothing. I'll tell you what will happen: people will start
> > > defining their nvmem cells directly under the MTD node because
> > > that *works*, and even if the binding is not documented and we
> > > consider it invalid, we'll be stuck supporting it forever.
> >
> > Do note that undocumented bindings are not allowed. DTS that use
> > undocumented bindings (normally) just get rejected.
>
> Except that's just in theory. In practice, if people can do something
> wrong, they'll complain if you later fix the bug and break their
> setup. So no, if we go for the "nvmem cells have an 'nvmem-cell'
> compat", then I'd like the NVMEM framework to enforce that somehow.

That should be trivial to implement.

Alban

Attachment: pgpnupoO4EKx6.pgp
Description: OpenPGP digital signature