Am 2020-06-08 12:02, schrieb Andy Shevchenko:
+Cc: some Intel people WRT our internal discussion about similar
problem and solutions.
On Mon, Jun 8, 2020 at 11:30 AM Lee Jones <lee.jones@xxxxxxxxxx> wrote:
On Sat, 06 Jun 2020, Michael Walle wrote:
> Am 2020-06-06 13:46, schrieb Mark Brown:
> > On Fri, Jun 05, 2020 at 10:07:36PM +0200, Michael Walle wrote:
> > > Am 2020-06-05 12:50, schrieb Mark Brown:
...
Right. I'm suggesting a means to extrapolate complex shared and
sometimes intertwined batches of register sets to be consumed by
multiple (sub-)devices spanning different subsystems.
Actually scrap that. The most common case I see is a single Regmap
covering all child-devices.
Yes, because often we need a synchronization across the entire address
space of the (parent) device in question.
It would be great if there was a way in
which we could make an assumption that the entire register address
space for a 'tagged' (MFD) device is to be shared (via Regmap) between
each of the devices described by its child-nodes. Probably by picking
up on the 'simple-mfd' compatible string in the first instance.
Rob, is the above something you would contemplate?
Michael, do your register addresses overlap i.e. are they intermingled
with one another? Do multiple child devices need access to the same
registers i.e. are they shared?
No they don't overlap, expect for maybe the version register, which is
just there once and not per function block.
> > > But, there is more in my driver:
> > > (1) there is a version check
If we can rid the Regmap dependency, then creating an entire driver to
conduct a version check is unjustifiable. This could become an inline
function which is called by each of the sub-devices instead, for
example.
sounds good to me. (although there would then be a probe fail per
sub-device
if the version is not supported)
> > > (2) there is another function for which there is no suitable linux
> > > subsystem I'm aware of and thus which I'd like to us sysfs
> > > attributes for: This controller supports 16 non-volatile
> > > configuration bits. (this is still TBD)
There is a place for everything in Linux.
What do these bits configure?
- hardware strappings which have to be there before the board powers up,
like clocking mode for different SerDes settings
- "keep-in-reset" bits for onboard peripherals if you want to save power
- disable watchdog bits (there is a watchdog which is active right from
the start and supervises the bootloader start and switches to failsafe
mode if it wasn't successfully started)
- special boot modes, like eMMC, etc.
Think of it as a 16bit configuration word.
> > TBH I'd also say that the enumeration of the subdevices for this
> > device should be in the device rather than the DT, they don't
> > seem to be things that exist outside of this one device.
>
> We're going circles here, formerly they were enumerated in the MFD.
> Yes, they are devices which aren't likely be used outside a
> "sl28cpld", but there might there might be other versions of the
> sl28cpld with other components on different base addresses. I
> don't care if they are enumerated in DT or MFD, actually, I'd
> prefer the latter. _But_ I would like to have the device tree
> properties for its subdevices, e.g. the ones for the watchdog or
> whatever components there might be in the future.
[...]
> MFD core can
> match a device tree node today; but only one per unique compatible
> string. So what should I use to differentiate the different
> subdevices?
Right. I have been aware of this issue. The only suitable solution
to this would be to match on 'reg'.
see below (1)
FYI: I plan to fix this.
If your register map needs to change, then I suggest that this is
either a new device or at least a different version of the device and
would also have to be represented as different (sub-)mfd_cell.
> Rob suggested the internal offset, which I did here.
FWIW, I don't like this idea. DTs should not have to be modified
(either in the first instance or subsequently) or specifically
designed to patch inadequacies in any given OS.
How does (1) play together with this? What do you propose the "reg"
property should contain?
> But then, there is less use in duplicating the offsets in the MFD
> just to have the MFD enumerate the subdevices and then match
> the device tree nodes against it. I can just use
> of_platform_populate() to enumerate the children and I won't
> have to duplicate the base addresses.
Which is fine. However this causes a different issue for you. By
stripping out the MFD code you render the MFD portion seemingly
superfluous. Another issue driver authors commonly contend with.
Yeah, I agree.
-michael
