[v4,3/6] i2c: Introduce OF component probe function

Some devices are designed and manufactured with some components having
multiple drop-in replacement options. These components are often
connected to the mainboard via ribbon cables, having the same signals
and pin assignments across all options. These may include the display
panel and touchscreen on laptops and tablets, and the trackpad on
laptops. Sometimes which component option is used in a particular device
can be detected by some firmware provided identifier, other times that
information is not available, and the kernel has to try to probe each
device.

This change attempts to make the "probe each device" case cleaner. The
current approach is to have all options added and enabled in the device
tree. The kernel would then bind each device and run each driver's probe
function. This works, but has been broken before due to the introduction
of asynchronous probing, causing multiple instances requesting "shared"
resources, such as pinmuxes, GPIO pins, interrupt lines, at the same
time, with only one instance succeeding. Work arounds for these include
moving the pinmux to the parent I2C controller, using GPIO hogs or
pinmux settings to keep the GPIO pins in some fixed configuration, and
requesting the interrupt line very late. Such configurations can be seen
on the MT8183 Krane Chromebook tablets, and the Qualcomm sc8280xp-based
Lenovo Thinkpad 13S.

Instead of this delicate dance between drivers and device tree quirks,
this change introduces a simple I2C component probe. function For a
given class of devices on the same I2C bus, it will go through all of
them, doing a simple I2C read transfer and see which one of them responds.
It will then enable the device that responds.

This requires some minor modifications in the existing device tree. The
status for all the device nodes for the component options must be set
to "failed-needs-probe". This makes it clear that some mechanism is
needed to enable one of them, and also prevents the prober and device
drivers running at the same time.

Signed-off-by: Chen-Yu Tsai <wenst@chromium.org>
---
Changes since v3:
- Complete kernel-doc
- Return different error if I2C controller is disabled
- Expand comment to explain assumptions and constraints
- Split for-loop finding target node and operations on target node
- Add missing i2c_put_adapter()
- Move prober code to separate file

Rob also asked why there was a limitation of "exactly one touchscreen
will be enabled across the whole tree".

The use case this prober currently targets is a component on consumer
electronics (tablet or laptop) being swapped out due to cost or supply
reasons. Designs with multiple components of the same type are pretty
rare. The way the next patch is written also assumes this for efficiency
reasons.

Changes since v2:
- New patch split out from "of: Introduce hardware prober driver"
- Addressed Rob's comments
  - Move i2c prober to i2c subsystem
  - Use of_node_is_available() to check if node is enabled.
  - Use OF changeset API to update status property
- Addressed Andy's comments
  - Probe function now accepts "struct device *dev" instead to reduce
    line length and dereferences
  - Move "ret = 0" to just before for_each_child_of_node(i2c_node, node)
---
 drivers/i2c/Makefile             |   1 +
 drivers/i2c/i2c-core-of-prober.c | 155 +++++++++++++++++++++++++++++++
 include/linux/i2c.h              |   4 +
 3 files changed, 160 insertions(+)
 create mode 100644 drivers/i2c/i2c-core-of-prober.c

On Tue, Aug 13, 2024 at 7:26 PM Andy Shevchenko
<andriy.shevchenko@linux.intel.com> wrote:
>
> On Thu, Aug 08, 2024 at 05:59:26PM +0800, Chen-Yu Tsai wrote:
> > Some devices are designed and manufactured with some components having
> > multiple drop-in replacement options. These components are often
> > connected to the mainboard via ribbon cables, having the same signals
> > and pin assignments across all options. These may include the display
> > panel and touchscreen on laptops and tablets, and the trackpad on
> > laptops. Sometimes which component option is used in a particular device
> > can be detected by some firmware provided identifier, other times that
> > information is not available, and the kernel has to try to probe each
> > device.
> >
> > This change attempts to make the "probe each device" case cleaner. The
> > current approach is to have all options added and enabled in the device
> > tree. The kernel would then bind each device and run each driver's probe
> > function. This works, but has been broken before due to the introduction
> > of asynchronous probing, causing multiple instances requesting "shared"
> > resources, such as pinmuxes, GPIO pins, interrupt lines, at the same
> > time, with only one instance succeeding. Work arounds for these include
> > moving the pinmux to the parent I2C controller, using GPIO hogs or
> > pinmux settings to keep the GPIO pins in some fixed configuration, and
> > requesting the interrupt line very late. Such configurations can be seen
> > on the MT8183 Krane Chromebook tablets, and the Qualcomm sc8280xp-based
> > Lenovo Thinkpad 13S.
> >
> > Instead of this delicate dance between drivers and device tree quirks,
> > this change introduces a simple I2C component probe. function For a
> > given class of devices on the same I2C bus, it will go through all of
> > them, doing a simple I2C read transfer and see which one of them responds.
> > It will then enable the device that responds.
> >
> > This requires some minor modifications in the existing device tree. The
> > status for all the device nodes for the component options must be set
> > to "failed-needs-probe". This makes it clear that some mechanism is
> > needed to enable one of them, and also prevents the prober and device
> > drivers running at the same time.
>
> ...
>
> > +int i2c_of_probe_component(struct device *dev, const char *type)
>
> Use respective scoped variants and remove the related of_node_put() calls.

Ack. Will also try splitting and reworking the code so they have
tighter scopes.

> ...
>
> > +             ocs = kzalloc(sizeof(*ocs), GFP_KERNEL);
>
> Use __free()

This ends up only a bit better than with gotos once this small section
is split out into a helper function.

AFAIK ocs, or at least the underlying properties, have to be left around
once the changeset is applied as they are now part of the dynamic live
tree. And that's fine since once applied, nothing is going to un-apply
it. OTOH in the error path it needs extra cleanup if any actions were
added.

So I end up with the following to silence the "must check return value"
warning on success, and cleanup on error:

ret = of_changeset_apply(ocs);
if (!ret) {
        void *ptr __always_unused = no_free_ptr(ocs);
} else {
        of_changeset_destroy(ocs);
}

In my case it might actually be safe to do of_changeset_destroy(ocs) and
free it regardless, but I'm not really confident.


ChenYu



> > +             if (!ocs) {
> > +                     ret = -ENOMEM;
> > +                     goto err_put_node;
> > +             }
>
> > +err_free_ocs:
> > +     kfree(ocs);
>
> --
> With Best Regards,
> Andy Shevchenko
>
>