On Fri, Oct 13, 2017 at 04:26:57PM +0200, Peter Rosin wrote:

On 2017-10-13 15:50, Guenter Roeck wrote:

quoted

On 10/13/2017 02:27 AM, Peter Rosin wrote:

quoted

With a nxp,se97 chip on an atmel sama5d31 board, the I2C adapter driver
is not always capable of avoiding the 25-35 ms timeout as specified by
the SMBUS protocol. This may cause silent corruption of the last bit of
any transfer, e.g. a one is read instead of a zero if the sensor chip
times out. This also affects the eeprom half of the nxp-se97 chip, where
this silent corruption was originally noticed. Other I2C adapters probably
suffer similar issues, e.g. bit-banging comes to mind as risky...

The SMBUS register in the nxp chip is not a standard Jedec register, but
it is not special to the nxp chips either, at least the atmel chips
have the same mechanism. Therefore, do not special case this on the
manufacturer, it is opt-in via the device property anyway.

Signed-off-by: Peter Rosin <redacted>
---
  Documentation/devicetree/bindings/hwmon/jc42.txt |  4 ++++
  drivers/hwmon/jc42.c                             | 20 ++++++++++++++++++++
  2 files changed, 24 insertions(+)

diff --git a/Documentation/devicetree/bindings/hwmon/jc42.txt b/Documentation/devicetree/bindings/hwmon/jc42.txt
index 07a250498fbb..f569db58f64a 100644
--- a/Documentation/devicetree/bindings/hwmon/jc42.txt
+++ b/Documentation/devicetree/bindings/hwmon/jc42.txt

@@ -34,6 +34,10 @@ Required properties:
  
  - reg: I2C address
  
+Optional properties:
+- smbus-timeout-disable: When set, the smbus timeout function will be disabled.
+			 This is not supported on all chips.
+
  Example:
  
  temp-sensor@1a {

diff --git a/drivers/hwmon/jc42.c b/drivers/hwmon/jc42.c
index 1bf22eff0b08..fd816902fa30 100644
--- a/drivers/hwmon/jc42.c
+++ b/drivers/hwmon/jc42.c

@@ -45,6 +45,7 @@ static const unsigned short normal_i2c[] = {
  #define JC42_REG_TEMP		0x05
  #define JC42_REG_MANID		0x06
  #define JC42_REG_DEVICEID	0x07
+#define JC42_REG_SMBUS		0x22 /* NXP and Atmel, possibly others? */
  
  /* Status bits in temperature register */
  #define JC42_ALARM_CRIT_BIT	15

@@ -73,6 +74,9 @@ static const unsigned short normal_i2c[] = {
  #define ONS_MANID		0x1b09  /* ON Semiconductor */
  #define STM_MANID		0x104a  /* ST Microelectronics */
  
+/* SMBUS register */
+#define SMBUS_STMOUT		BIT(7)  /* SMBus time-out, active low */
+
  /* Supported chips */
  
  /* Analog Devices */

@@ -476,6 +480,22 @@ static int jc42_probe(struct i2c_client *client, const struct i2c_device_id *id)
  
  	data->extended = !!(cap & JC42_CAP_RANGE);
  
+	if (device_property_read_bool(dev, "smbus-timeout-disable")) {
+		int smbus;
+
+		/*
+		 * Not all chips support this register, but from a
+		 * quick read of various datasheets no chip appears
+		 * incompatible with the below attempt to disable
+		 * the timeout. And the whole thing is opt-in...
+		 */
+		smbus = i2c_smbus_read_word_swapped(client, JC42_REG_SMBUS);
+		if (smbus < 0)
+			return smbus;
+		i2c_smbus_write_word_swapped(client, JC42_REG_SMBUS,
+					     smbus | SMBUS_STMOUT);

Looking into the SE97 datasheet, the bit is only writable if the alarm bits
are not locked. Should we take this into account and unlock the alarm bits
if necessary ?

Right. And I thought about the case when the timeout was disabled before
probing but with the property not present (perhaps by someone trying things
out, like I have). Should the timeout be re-enabled in that case?

No, because the property only states that the timeout should be disabled.
It does not say that it should be _enabled_ if the property is not there.
That would require a different property. A -> B does not imply B -> A.

But, someone might have disabled the timeout by some previous arrangement
(e.g. in a boot-loader) but without having this newfangled property in the
device tree. Re-enabling the timeout in that case would break things. Slim
chance for that to be an issue, but perhaps not?

Unlocking the alarm bits is somewhat similar, since it should only be an
issue for warm starts. But the risk of breakage is perhaps not there at
all?

We would have to lock the alarm bits again, leaving them in a consistent
state.

Your call, I can fix thing however you like...

Let's just leave it as-is. If we encounter a problem later we can always
add code to unlock/lock the alarm bits.

Guenter
--
To unsubscribe from this list: send the line "unsubscribe devicetree" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html