Re: [PATCH v6 0/5] Introduce the Counter character device interface

[William Breathitt Gray]

On Sun, Dec 13, 2020 at 05:15:14PM -0600, David Lechner wrote:
> On 11/22/20 2:29 PM, William Breathitt Gray wrote:
> > 
> > 1. Should standard Counter component data types be defined as u8 or u32?
> > 
> >     Many standard Counter component types such COUNTER_COMP_SIGNAL_LEVEL
> >     have standard values defined (e.g. COUNTER_SIGNAL_LEVEL_LOW and
> >     COUNTER_SIGNAL_LEVEL_HIGH). These values are currently handled by the
> >     Counter subsystem code as u8 data types.
> > 
> >     If u32 is used for these values instead, C enum structures could be
> >     used by driver authors to implicitly cast these values via the driver
> >     callback parameters.
> > 
> >     This question is primarily addressed to David Lechner. I'm somewhat
> >     confused about how this setup would look in device drivers. I've gone
> >     ahead and refactored the code to support u32 enums, and pushed it to
> >     a separate branch on my repository called counter_chrdev_v6_u32_enum:
> >     https://gitlab.com/vilhelmgray/iio/-/tree/counter_chrdev_v6_u32_enum
> > 
> >     Please check it out and let me know what you think. Is this the
> >     support you had in mind? I'm curious to see an example of how would
> >     your driver callback functions would look in this case. If everything
> >     works out fine, then I'll submit this branch as v7 of this patchset.
> 
> I haven't had time to look at this in depth, but just superficially looking
> at it, it is mostly there. The driver callback would just use the enum type
> in place of u32. For example:
> 
> static int ti_eqep_function_write(struct counter_device *counter,
> 				  struct counter_count *count,
> 				  enum counter_function function)
> 
> and the COUNTER_FUNCTION_* constants would be defined as:
> 
> enum counter_function {
> 	COUNTER_FUNCTION_INCREASE,
> 	...
> };
> 
> instead of using #define macros.
> 
> One advantage I see to using u8, at least in the user API data structures,
> is that it increases the number of events that fit in the kfifo buffer by
> a significant factor.
> 
> And that is not to say that we couldn't do both: have the user API structs
> use u8 for enum values and still use u32/strong enum types internally in
> the callback functions.

I'm including David Laight because he initially opposed enums in favor
of fixed size types when we discussed this in an earlier revision:
https://lkml.org/lkml/2020/5/3/159

However, there have been significant changes to this patchset so the
context now is different than those earlier discussions (i.e. we're no
longer discussing ioctl calls).

I think reimplementing these constants as enums as described could work.
If we do so, should the enum constants be given specific values? For
example:

enum counter_function {
	COUNTER_FUNCTION_INCREASE = 0,
	COUNTER_FUNCTION_DECREASE = 1,
	...
};

> 
> > 
> > 2. How should we handle "raw" timestamps?
> > 
> >     Ahmad Fatoum brought up the possibility of returning "raw" timestamps
> >     similar to what the network stack offers (see the network stack
> >     SOF_TIMESTAMPING_{RAW,SYS}_HARDWARE support).
> > 
> >     I'm not very familiar with the networking stack code, but if I
> >     understand correctly the SOF_TIMESTAMPING_RAW_HARDWARE timestamps are
> >     values returned from the device. If so, I suspect we would be able to
> >     support these "raw" timestamps by defining them as Counter Extensions
> >     and returning them in struct counter_event elements similar to the
> >     other Extension values.
> 
> Is nanosecond resolution good enough? In the TI eQEP driver I considered
> returning the raw timer value, but quickly realized that it would not be
> very nice to expect the user code to know the clock rate of the timer. It
> was very easy to get the clock rate in the kernel and just convert the
> timer value to nanoseconds before returning it to userspace.
> 
> So if there is some specialized case where it can be solved no other way
> besides using raw timestamps, then sure, include it. Otherwise I think we
> should stick with nanoseconds for time values when possible.

Given that the struct counter_event 'timestamp' member serves as the
identification vessel for correlating component values to a single event
(i.e. component values of a given event will share the same unique
timestamp), I believe it's prudent to standardize this timestamp format
on the kernel monotonic time as we have currently done so via our
ktime_get_ns() call.

There are cases where it is understandably better to use a timestamp
provided directly by the hardware (e.g. keeping timestamping close to
data collection). For these cases, we can retrieve these "raw"
timestamps via a Counter Extension: users would get their "raw"
timestamp via the struct counter_event 'value' member, and just treat
the 'timestamp' member as a unique event identification number.

William Breathitt Gray

Attachment: signature.asc
Description: PGP signature