Re: [RFC Patch v0 1/3] i2c-smbus: Add poll interface for smbus alert

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 27/03/14 21:50, Srinivas Pandruvada wrote:
On 03/27/2014 10:34 AM, Jonathan Cameron wrote:

On March 27, 2014 7:44:56 AM GMT+00:00, Jean Delvare<jdelvare@xxxxxxx>  wrote:
On Wed, 26 Mar 2014 17:42:10 -0700, Srinivas Pandruvada wrote:
The current i2c smbus alert module depends on smbus alert mechanism
supported by underlying bus drivers. By specifications, these alerts
can be polled if there is no hardware support.
Currently multiple drivers who needs smbus alerts are creating
a new i2c dummy device with address 0x0C (ARA register), by luck
they don't co-exist. Otherwise i2c device creation will fail.
Added a poll interface, so that all these driver can call a common
interface to poll. Even if they polli, all drivers bound to an
adapater will be notified by their alert callback if ARA register
read is successful.

Signed-off-by: Srinivas Pandruvada
<srinivas.pandruvada@xxxxxxxxxxxxxxx>
---
  drivers/i2c/i2c-smbus.c   | 23 +++++++++++++++++++++++
  include/linux/i2c-smbus.h |  3 +++
  2 files changed, 26 insertions(+)

diff --git a/drivers/i2c/i2c-smbus.c b/drivers/i2c/i2c-smbus.c
index fc99f0d..e274f20 100644
--- a/drivers/i2c/i2c-smbus.c
+++ b/drivers/i2c/i2c-smbus.c
@@ -72,6 +72,29 @@ static int smbus_do_alert(struct device *dev, void
*addrp)
  	return -EBUSY;
  }

+int i2c_smbus_ara_poll(const struct i2c_client *client)
+{
+	union i2c_smbus_data data;
+	int status;
+	struct alert_data alert_data;
+
+	status = i2c_smbus_xfer(client->adapter, 0x0C, 0,
+				I2C_SMBUS_READ, 0,
+				I2C_SMBUS_BYTE, &data);
+	if (status < 0)
+		return status;
+
+	alert_data.flag = data.byte & 1;
+	alert_data.addr = data.byte >> 1;
+
+	/* Notify driver for the device which issued the alert */
+	device_for_each_child(&client->adapter->dev, &alert_data,
+				smbus_do_alert);
+
+	return data.byte;
+}
This is essentially duplicating code from smbus_alert(), but in a
hackish way, as the ARA is never properly reserved. Your bus driver
should really register the ARA with i2c_setup_smbus_alert().

I see that the code may not properly deal with the polled case
everywhere but it should be pretty trivial to deal with. For example,
check for alert->irq > 0 before re-enabling the irq in smbus_alert(). I
don't immediately see any other change needed.

If SMBus alert polling is done from the i2c device driver, we'll have
to find a standard way for i2c device drivers to retrieve the ara
client associated with an i2c_adapter. However I still need to be
convinced that this makes any sense at all. Ultimately the alert will
call the i2c device drivers's alert() callback. If the i2c device
driver needs to do that, there's no need to go through ARA, it might as
well just call the callback by itself.

So can you please explain what problem exactly you are trying to solve?
As I understand it the issue is that some parts will not clear
their internal interrupt status unless an at a read occurs and
presumably the read has to get the correct address?

To my mind we should have polling inside the core and it should
ensure all devices that want to have replied. Have I miss
understood how Ara is supposed to work but should we not know which
device to call the callback on?
Ah, I was barking up the wrong tree with this one and had missed the
check that was conducted to ensure that only the right device is
notified.  Sorry about that.

This is summary: May be Jean can suggest some better solution: - We
have some sensor devices on i2C bus, which need to read ARA register
to ack smbus alert signal. Without acking this, they will not allow
further reads. They currently are calling i2c_new_dummy() for
reserving ARA (0x0C) register. But the problem is if there are two
devices on this bus both have to read ARA, they can't call
i2c_new_dummy() as one of them will fail (__i2c_check_addr_busy()
will return -EBUSY). - In above case, If they use hacked way to use
i2_transfer function, but it will not address the real issue. Because
one driver is reading ARA basically can acknowledge ALERT from wrong
device.

So their needs to be a core level poll so that devices can be
notified, if there is an alert for their device only.
Agreed.  Right now at a quick glance, all devices are notified which
seems odd...

From physical interface level many i2controllers (including x86s)
don't have separate signal for SMBALERT#. So somewhere it needs
polling. For some devices SMBALERT is wired through a GPIO for GPIO
interrupt. But they still need to read ARA to ACK.
If it is wired through a gpio interrupt (and that interrupt is shared
by various smbalert equipped devices - or there is only one present
on the i2c bus) then it should be handled using the existing
infrastructure. (perhaps modified slightly)

Also we can't blindly reserve 0x0c for every i2c-bus as 0x0C is a
valid i2C address. There are already occupied (e.g. AK8963 slave
address is 0x0C on the same platforms). So can't call
i2c_setup_smbus_alert from the bus driver for all i2c buses.
If platforms are wiring up both devices using this i2c address and
smbus alert devices then we are in an irritating corner case.
This is effectively the same as having two different smbus devices
on the same address and relying on the fact they don't have
overlapping address spaces.  Nuts but I guess we want to find some
way of coping with it?

Perhaps we need  away of allow multiple i2c devices to share a bus
in this case?  Is this something that could be done via a nefarious
hack of using a mux to pretend that they are separately wired
(when they aren't).

For polling, a bus driver can't decide periodic polling as this is
power/perf drain.
Agreed, so hints from the drivers to the core would be needed.
So my idea was to somehow kick a read on ARA on a
bus from interested clients may be when the receive GPIO interrupt.
Then you can end up with considerable waste if you have lots of devices
using this functionality on the same bus.  Each driver polls at say
50Hz, because it wants 50Hz, but you have 20 devices hence you get
uneven sampling at say 1kHz and might even block the i2c bus up entirely.
Hence the polling needs to be handled by the core.  I'd suggest
a call to request enabling of polling and a frequency.  Then the
core just uses the highest frequency currently requested.

It's a bit more complex than the nice sensible interrupt driven
version, but if board designers are going to do silly things
I guess we have to figure out how to make them work.

But the results are notified to interested drivers who needs smbus
alert processing.
Thanks,
Srinivas














+EXPORT_SYMBOL_GPL(i2c_smbus_ara_poll);
+
  /*
   * The alert IRQ handler needs to hand work off to a task which can
issue
   * SMBus calls, because those sleeping calls can't be made in IRQ
context.
diff --git a/include/linux/i2c-smbus.h b/include/linux/i2c-smbus.h
index 8f1b086..f70755d 100644
--- a/include/linux/i2c-smbus.h
+++ b/include/linux/i2c-smbus.h
@@ -48,4 +48,7 @@ struct i2c_client *i2c_setup_smbus_alert(struct
i2c_adapter *adapter,
  					 struct i2c_smbus_alert_setup *setup);
  int i2c_handle_smbus_alert(struct i2c_client *ara);

+/* Interface to poll smbus alert */
+int i2c_smbus_ara_poll(const struct i2c_client *client);
+
  #endif /* _LINUX_I2C_SMBUS_H */


--
To unsubscribe from this list: send the line "unsubscribe linux-i2c" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html




[Index of Archives]     [Linux GPIO]     [Linux SPI]     [Linux Hardward Monitoring]     [LM Sensors]     [Linux USB Devel]     [Linux Media]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]

  Powered by Linux