Re: [PATCH 2/2] nvmem: expose NVMEM cells in sysfs

[Rafał Miłecki]

On 21.12.2021 15:27, Greg Kroah-Hartman wrote:
> On Tue, Dec 21, 2021 at 02:52:05PM +0100, Rafał Miłecki wrote:
> > > How are nvmem devices named?
> >
> > $ ls /sys/bus/nvmem/devices/
> > brcm-nvram0
> > mtd0
> > mtd1
> > u-boot-envvar0
>
> So no naming scheme at all.
>
> {sigh}
>
> > > > Example:
> > > > $ cat /sys/bus/nvmem/devices/foo/cells/bootcmd
> > > > tftp
> > > > $ cat /sys/bus/nvmem/devices/foo/cells/bootdelay
> > > > 5
> > > >
> > > > As you can see above NVMEM cells are not known at compilation time.
> > >
> > > Why do you want to expose these in a way that forces the kernel to parse
> > > these key/value pairs?  Why not just do it all in userspace like you can
> > > today?  What forces the kernel to do it and not a perl script?
> > >
> > > > So I believe the question is: how can I expose cells in sysfs?
> > >
> > > You can do this by dynamically creating the attributes on the fly, but
> > > your show function is going to be rough and it's not going to be simple
> > > to do so.  One example will be the code that creates the
> > > /sys/devices/system/machinecheck/machinecheckXX/bank* files.
> > >
> > > But I will push back again, why not just do it all in userspace?  What
> > > userspace tool is requiring the kernel to do this work for it?
> >
> > Environment data contains info that may be required by kernel.
> >
> > For example some home routers store two firmwares on flash. Kernel needs
> > to read index of currently booted firmware to make sure MTD subsystem
> > creates partitions correctly.
>
> You are talking about a kernel<->kernel api here, that's not what sysfs
> is for at all.
>
> > Another example: MAC address. Ethernet subsystem supports reading MAC
> > from NVMEM cell.
>
> Again, internal kernel api, nothing sysfs is ever involved in.
>
> > One could argue those tasks could be handled from userspace but that
> > would get tricky. Sure - we have API for setting MAC address. However
> > other cases (like setting active firmware partition and asking MTD to
> > parse it into subpartitions) would require new user <-> kernel
> > interfaces.
>
> Ok, but again, sysfs is for userspace to get access to these values.
> That's what I'm concerned about.  If you want to make an in-kernel api
> for other subsystems to get these key/value pairs, wonderful, that has
> nothing to do with sysfs.
>
> So I ask again, why do you want to expose these to userspace through
> sysfs in a new format from what you have today.  Who is going to use
> that information and what is it going to be used for.

Sorry, you asked why I need parsing in kernel and I focused on that part
only.

For user space there are also some relevant U-Boot environment entries
(NVMEM cells). Examples: serial number, country code (e.g. WiFi
purposes), default passwords (as printed on labels).

So both: kernel and user space need to access U-Boot environment
variables (NVMEM cells). Each for its own purposes.

Kernel goes first so it needs its own parser of NVMEM content (data).

User space can either: get NVMEM cells exposed by kernel OR parse NVMEM
content on its own. I thought it'd be nice to avoid parsing code
duplication in user space and let kernel expose NVMEM cells.