[lm-sensors] RFC: Intel QST driver

[Simon J. Rowe]

I've written a driver for the Intel Quiet System Technology (QST)
function of the Management Engine Interface found on recent Intel
chipsets.

The git repo can be found here:

http://mose.dyndns.org/mei.git

A few questions / observations,

1) The code was developed and tested on 2.6.39. I would hope it
compiles and runs on 3.x but I haven't tried it.

2) The main hwmon code is in qst-hwmon.c. This implements QST v1, I
don't have access to hardware that supports v2. I would imagine that
implementing v2 would result in a new module (qst2-hwmon) which would
be almost identical.

3) I've gone a bit overboard with preprocessor macros but it does mean
that the amount of duplicated code is kept to a minimum.

3) I've not implemented any PWM methods yet.

4) I don't believe the MEI (HECI) implementation that Intel have
already submitted to the mainline kernel is usable by other kernel
modules. I have re-implemented it in a way that is accessible to
either the kernel or userspace.

5) I had to patch libsensors to work with a new bus type

diff -ur lm_sensors-3.3.1.org//lib/sysfs.c lm_sensors-3.3.1/lib/sysfs.c
--- lm_sensors-3.3.1.org//lib/sysfs.c 2011-03-04 20:37:43.000000000 +0000
+++ lm_sensors-3.3.1/lib/sysfs.c 2012-11-14 21:48:52.144860375 +0000
@@ -701,6 +701,12 @@
/* As of kernel 2.6.32, the hid device names don't look
good */
entry.chip.bus.nr = bus;
entry.chip.addr = id;
+ } else
+ if (subsys && !strcmp(subsys, "intel-mei") &&
+ sscanf(dev_name, "mei%d:%d", &bus, &fn) == 2) {
+ entry.chip.bus.type = SENSORS_BUS_TYPE_PCI;
+ entry.chip.bus.nr = bus;
+ entry.chip.addr = fn;
} else {
/* Ignore unknown device */
err = 0;

surely this sort of knowledge belongs in the driver not userspace?
Could drivers not provide another set of sysfs attributes which expose
bus type, number, addr etc?

Please let me know if there's any changes or improvements I can make
to it.

Simon

[Guenter Roeck]

Post by Simon J. Rowe
I've written a driver for the Intel Quiet System Technology (QST)
function of the Management Engine Interface found on recent Intel
chipsets.
http://mose.dyndns.org/mei.git
A few questions / observations,
1) The code was developed and tested on 2.6.39. I would hope it
compiles and runs on 3.x but I haven't tried it.

Hi Simon,

2.6.39 is pretty old. It doesn't include the new mei driver (which is now
in drivers/misc/mei).

Post by Simon J. Rowe
2) The main hwmon code is in qst-hwmon.c. This implements QST v1, I
don't have access to hardware that supports v2. I would imagine that
implementing v2 would result in a new module (qst2-hwmon) which would
be almost identical.

Is it that much different that it would require a separate driver, especially
if it is almost identical ?

Post by Simon J. Rowe
3) I've gone a bit overboard with preprocessor macros but it does mean
that the amount of duplicated code is kept to a minimum.

It also hides possible error cases, though, and makes review very complicated.
You would be better off using function parameters as much as possible. Plus,
after all, it is not correct that you avoid duplicate code. You only avoid
duplicate source code. Fore that, the additional cost (in terms of risk and
maintainability) is pretty high.

Post by Simon J. Rowe
3) I've not implemented any PWM methods yet.
4) I don't believe the MEI (HECI) implementation that Intel have
already submitted to the mainline kernel is usable by other kernel
modules. I have re-implemented it in a way that is accessible to
either the kernel or userspace.

2.6.39 does not include the new mei driver, so I don't think you can make
that claim.

I would suggest to have a look into the new driver and work with its
maintainer to make it better suitable for other drivers if needed.
Its main deficiency, as far as I could see when I looked into it, is that it
doesn't support out-of-directory drivers very well - which means the hwmon
driver might have to reside in drivers/misc/mei instead of drivers/hwmon.
Not an optimal solution, but better than nothing.

Post by Simon J. Rowe
5) I had to patch libsensors to work with a new bus type
diff -ur lm_sensors-3.3.1.org//lib/sysfs.c lm_sensors-3.3.1/lib/sysfs.c
--- lm_sensors-3.3.1.org//lib/sysfs.c 2011-03-04 20:37:43.000000000 +0000
+++ lm_sensors-3.3.1/lib/sysfs.c 2012-11-14 21:48:52.144860375 +0000
@@ -701,6 +701,12 @@
/* As of kernel 2.6.32, the hid device names don't
look good */
entry.chip.bus.nr = bus;
entry.chip.addr = id;
+ } else
+ if (subsys && !strcmp(subsys, "intel-mei") &&
+ sscanf(dev_name, "mei%d:%d", &bus, &fn) == 2) {
+ entry.chip.bus.type = SENSORS_BUS_TYPE_PCI;
+ entry.chip.bus.nr = bus;
+ entry.chip.addr = fn;
} else {
/* Ignore unknown device */
err = 0;

Is that really necessary ? Changing the ABI/API to userland is always
problematic.

Post by Simon J. Rowe
surely this sort of knowledge belongs in the driver not userspace?
Could drivers not provide another set of sysfs attributes which expose
bus type, number, addr etc?
Please let me know if there's any changes or improvements I can make
to it.

You might want to run your code through checkpatch and fix all errors and
warnings. The major problem, though, is that we can not have two instances
of mei core code in the kernel.

Copying the mei driver maintainer and the kernel mailing list for additional
input.

Thanks,
Guenter

[Tomas Winkler]

Post by Guenter Roeck

Post by Simon J. Rowe
4) I don't believe the MEI (HECI) implementation that Intel have
already submitted to the mainline kernel is usable by other kernel
modules. I have re-implemented it in a way that is accessible to
either the kernel or userspace.

2.6.39 does not include the new mei driver, so I don't think you can make
that claim.

We are working on adding possibility of stacking another drivers
above the MEI
in these days. I cannot commit to specific time line yet but it is on the way.


Thanks
Tomas