On Wed, Aug 4, 2021 at 1:50 AM Peter Rosin [off-list ref] wrote:

On 2021-08-02 23:51, Eddie James wrote:

quoted

On Mon, 2021-08-02 at 14:46 -0600, Rob Herring wrote:

quoted

On Tue, Jul 27, 2021 at 11:03:15AM -0500, Eddie James wrote:

quoted

Some systems connect several PCA954x devices to a single reset
GPIO. For
these devices to get out of reset and probe successfully, each
device must
defer the probe until the GPIO has been hogged. Accomplish this by
attempting to grab a new "reset-shared-hogged" devicetree property,
but
expect it to fail with EPROBE_DEFER or EBUSY.

Signed-off-by: Eddie James <eajames@linux.ibm.com>
---
 drivers/i2c/muxes/i2c-mux-pca954x.c | 46 +++++++++++++++++++++++
------
 1 file changed, 37 insertions(+), 9 deletions(-)

diff --git a/drivers/i2c/muxes/i2c-mux-pca954x.c

b/drivers/i2c/muxes/i2c-mux-pca954x.c
index 4ad665757dd8..376b54ffb590 100644

--- a/drivers/i2c/muxes/i2c-mux-pca954x.c
+++ b/drivers/i2c/muxes/i2c-mux-pca954x.c

@@ -434,15 +434,43 @@ static int pca954x_probe(struct i2c_client

*client,
    i2c_set_clientdata(client, muxc);
    data->client = client;

-   /* Reset the mux if a reset GPIO is specified. */
-   gpio = devm_gpiod_get_optional(dev, "reset", GPIOD_OUT_HIGH);
-   if (IS_ERR(gpio))
-           return PTR_ERR(gpio);
-   if (gpio) {
-           udelay(1);
-           gpiod_set_value_cansleep(gpio, 0);
-           /* Give the chip some time to recover. */
-           udelay(1);
+   /*
+    * Grab the shared, hogged gpio that controls the mux reset. We
expect
+    * this to fail with either EPROBE_DEFER or EBUSY. The only
purpose of
+    * trying to get it is to make sure the gpio controller has
probed up
+    * and hogged the line to take the mux out of reset, meaning
that the
+    * mux is ready to be probed up. Don't try and set the line any
way; in
+    * the event we actually successfully get the line (if it
wasn't
+    * hogged) then we immediately release it, since there is no
way to
+    * sync up the line between muxes.
+    */
+   gpio = gpiod_get_optional(dev, "reset-shared-hogged", 0);
+   if (IS_ERR(gpio)) {
+           ret = PTR_ERR(gpio);
+           if (ret != -EBUSY)
+                   return ret;

Why can't you just do this with the existing 'reset-gpios' property?
What's the usecase where you'd want to fail probe because EBUSY
other
than an error in your DT.

Hi, thanks for the reply.

Are you suggesting I use "reset-gpios" and change the driver to ignore
EBUSY? I don't know any other usecase, I just didn't think it would be
acceptable to ignore EBUSY on that, but perhaps it is a better
solution.

Hi!

From a device-tree point of view that might seem simple. But it becomes
a mess when several driver instances need to coordinate. If one instance
is grabbing the reset line but is then stalled while other instances
race ahead, they might be clobbered by a late reset from the stalled
first instance.

And while it might be possible to arrange the code such that those dragons
are dodged and that the reset is properly coordinated, what if the gpio is
supposed to be shared with some other totally unrelated driver? It might
seem to work when everything is normal, but as soon as anything out of the
ordinary happens, all bets are off. I expect subtle problems in the
furture.

All of this is true, but a different reset GPIO property name does
nothing to solve it.

I see no simple solution to this, and I also expect that if gpios need
to be shared, there will eventually need to be some kind of layer that
helps with coordination such that it becomes explicit rather than
implicit and fragile.

Yes, like making the reset subsystem handle 'reset-gpios' properties
as I suggested.

Rob