Re: Question regarding runtime pinctrl (2nd try)

["Niedermayr, BENEDIKT" <benedikt.niedermayr@xxxxxxxxxxx>]

On Wed, 2022-12-07 at 22:38 +0200, andriy.shevchenko@xxxxxxxxx wrote:
> On Wed, Dec 07, 2022 at 12:02:08PM +0000, Niedermayr, BENEDIKT wrote:
> > On Mon, 2022-12-05 at 14:58 +0200, andriy.shevchenko@xxxxxxxxx wrote:
> > > On Mon, Dec 05, 2022 at 10:47:27AM +0000, Niedermayr, BENEDIKT wrote:
> > > > On Wed, 2022-11-30 at 17:43 +0200, Andy Shevchenko wrote:
> > > > > On Wed, Nov 30, 2022 at 03:09:50PM +0000, Niedermayr, BENEDIKT
> > > > > wrote:
> 
> Something is wrong with your mail user agent or editor. The lines are
> quite
> long. Please, make sure you wrap them somewhere at ~76 characters per
> line.
> 
Thank for that. I fixed it now.

> > > > > > Hello,
> > > > > > 
> > > > > > I got no response since last time so I try it again, but with a
> > > > > > bit more
> > > > > > knowledge this time.
> > > > > > 
> > > > > > After carefully reading the pinctrl documentation
> > > > > > (driver-api/pin-control.rst) it was very clear for me that such
> > > > > > an interface
> > > > > > already exists and is accessable via debugfs. The documentation
> > > > > > is very clear
> > > > > > and self-explanatory. Thanks for that!
> > > > > > At the time of writing my last email [1] I took a look into an
> > > > > > older BSP
> > > > > > kernel where this feature has not been implemented, yet. I must
> > > > > > apologize for
> > > > > > that...
> > > > > > 
> > > > > > Now my last concern is using debugfs on a productive system.
> > > > > > IMHO debugfs is
> > > > > > not the right interface to interact on a productive system.
> > > > > 
> > > > > And this is a point. No-one should try it on the production
> > > > > systems.
> > > > > 
> > > > > > Especially when
> > > > > > when a unprivileged process wants to interact with an interface
> > > > > > offered by
> > > > > > debugfs. It's possible to change
> > > > > > permissions on files and folders there but nevertheless I think
> > > > > > that this is
> > > > > > not the way to go, since debugfs was designed to offer
> > > > > > interfaces to
> > > > > > privileged processes only.
> > > > > 
> > > > > Correct.
> > > > > 
> > > > > > My proposal would be to implement an chardev interface for that
> > > > > > and using
> > > > > > udev rules to assign correct permissions to that. With this
> > > > > > interface I can
> > > > > > then select the active pinctrl-groups which have been defined in
> > > > > > the device
> > > > > > tree before.
> > > > > > I could also imagine to put the interface into the sysfs (that
> > > > > > would be very
> > > > > > close to the debugfs implementation I think).
> > > > > > 
> > > > > > What do you think about it? Am I still missing something?
> > > > > 
> > > > > In my opinion -- no go.
> > > > > 
> > > > > The platform description (ACPI, DT, or board files) should know
> > > > > what they are
> > > > > doing. If something missing to achieve what you need via existing
> > > > > interfaces
> > > > > we rather think about that, but no, the debugfs stays and only for
> > > > > the purposes
> > > > > of development on the "I know what I'm doing" basis.
> > > > > 
> > > > Ok. If I got you right, you meant that there is no way to replace
> > > > the
> > > > debugfs interface?
> > > > 
> > > > So instead replacing the debugfs interface I would rather add a
> > > > second
> > > > interface that coexists with debugfs.
> > > 
> > > I meant that this feature quite likely will stay in the debugfs realm.
> > > No new
> > > interface is needed for sure.
> > > 
> > > > Unfortunatelly there is no interface available for runtime
> > > > configuration, yet.
> > > 
> > > There is no explanation why you need that.
> > > This is the main point of this discussion, right?
> > > 
> > > > The only alternative 
> > > > is to access "/dev/mem", but this is the most questionable solution
> > > > from
> > > > a security perspective.
> > > 
> > > It's not an alternative at all, it's simple no go variant.
> > > 
> > > > There should be a way to avoid unsecure "/dev/mem" implementations
> > > > but
> > > > currently this is the only way to achieve runtime configuration with
> > > > reasonable effort.  IMHO the current architecture leads to lot of
> > > > unsecure implementations out there.
> > > > 
> > > > For example the raspberrypi kernel tries to workaround this issue by
> > > > providing a "/dev/gpiomem"
> > > 
> > > This is even worse than more standard /dev/mem interface.
> > > 
> > > > interface that only provides mappings to the gpio register
> > > > set(drivers/char/broadcom/bcm2835-gpiomem.c). 
> > > > This reduces possible vulnerabilities but they still persist since:
> > > > 
> > > > - mmap() cannot map memory less than PAGE_SIZE, which means that
> > > > memory
> > > > outside of the GPIO registers is accessable. 
> > > > - it's possible to select untested  pin configurations which may not
> > > > be
> > > > electrical fine.
> > > > 
> > > > I like the current architecture since I define pingroups in the
> > > > platform
> > > > description which have been tested and then select one of them
> > > > during
> > > > runtime.  It's just the interface itself which is not sufficient
> > > > enough
> > > > when it comes to security.  
> > > 
> > > Still no clue, what you are trying to achieve and why. Use case,
> > > please?
> > It already mentioned the use case here [1].
> 
> (Too many [1]:s)
> 
> > But let me explain it again and
> > in slightly other words.
> 
> Thank you, it helps. See my comments below.
> 
> > We are currently working on platforms that can be extended with
> > different
> > types of IO-Shields. The pinmux configuration is currently done by a
> > userspace application that offers a ncurses-like GUI interface. There
> > you can
> > select the pinmux configuration for each pin regarding on the IO-Shield
> > you
> > are using.
> 
> This is very dangerous feature. While it might work in your case it may
> damage
> the users' hardware if they don't know what they are doing.
> 
> What prevents you to create a DT / ACPI overlay and load it?
> Okay, seems the answer is below...
> 
> > I already mentioned how pinmuxing works with this framework and we have
> > the
> > same opinion about using /dev/mem like you.  Well, now we're are looking
> > for
> > other solutions.
> > 
> > For example using device tree overlays (written statically or generated)
> > for
> > each different kind of shield would be technically fine, but doesn't
> > scale
> > very well with increasing number of shields and host platforms.
> 
> I agree with this.
> 
> > And if the host platform is non ARM based this approach may not work.
> 
> Why?
> 
> > Furthermore we need to deal also with ACPI stuff on x86.
> 
> And?..
> ACPI supports overlays.

Thanks! Important information. I wasn't aware of that.

> 
> > Recompiling the device-tree/kernel for each shield and host platform is
> > possible, but from a userspace developer point of view this means efford
> > and
> > may require more knowledge about the hardware (or you need to request
> > for
> > features from your BSP provider).
> 
> Yes, and it's a good thing, right?
Maybe that's a question we should ask them :-|

> 
> > I also think about the question, why are there frameworks out there that
> > try
> > to solve that part of the problem (wiring-pi, eclipse-mraa, etc.).  I
> > think
> > these frameworks try to address, among other things, those issues.  So
> > it's
> > not only about our special use case. IMHO there are many use cases where
> > this
> > would make sense, otherwise these frameworks would have never come into
> > being.
> > 
> > I hope this clarifies my point of view a bit better.
> 
> Definitely.
> 
> But again, what you are proposing is not for production, but for
> prototyping.
> That's why it's under debugfs. Moreover, it's dangerous from electrical
> point
> of view reconfigure pins at run-time. This feature very well may damage
> the
> hardware or even kill somebody (if you think of a heavy robots).
First of all, thank you very much for sharing your valuable opinion! 

There is one more question I have now.
So pin-configuration at runtime is a dangerous thing and there are
situations where I fully agree with you.

But what about the GPIO pinmuxing. I dived into the pinctrl/gpio subsystems
and it seems that a gpio pin is automatically muxed in when requesting it
(gpio_request_enable()). This is also potentially dangerous runtime
reconfiguration, right? But that landed into a stable production interface.

Know I ask myself if that could also be applied to i2c, spi, uart interfaces
as well?
For example: if one requests the the i2c interface by calling open() (or an
ioctl) to /dev/i2c-X. Would it then be possible to mux those pins in. 
It would be almost analog to the current gpio implementation.
It's just an idea and I don't know wether this is technically possible
or how much effort this means.

I'm currently just looking for different kinds of implementation
possibilities. And if there is one solution that may have the chance to get
upstream, I would rather stick to that one.

> 
> The solutions for the developers who KNOW what they ARE DOING are:
> 1) overlays;
> 2) debugfs;
> 3) reboot.
> 
> Choose one, suitable for you and go with it.
> 
> > [1] https://marc.info/?l=linux-gpio&m=166850640920120
> > 
> > > > > > [1] https://marc.info/?l=linux-gpio&m=166850640920120

cheers,
Benedikt