Re: Suggestion - Configurable Source Clock Type for Line Event Timestamping

[Linus Walleij <linus.walleij@linaro.org>]

On Sun, Oct 11, 2020 at 5:11 PM Jack Winch <sunt.un.morcov@gmail.com> wrote:

> I recently noticed that in Linux 5.7, gpiolib was changed such that
> line events are now timestamped using the system 'monotonic' clock
> rather than the system realtime clock.  The rationale for this change
> appears to be due to the major use-case of the line event timestamp
> data in relation to the nature of the system realtime clock (which can
> jump backwards or forwards in time due to adjustments by third-parties
> - e.g., NTP or PTP clients, etc).

The background is in the commit:
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/drivers/gpio/gpiolib.c?id=f8850206e160bfe35de9ca2e726ab6d6b8cb77dd

Also study the solution in IIO that started the discussion about
all this:
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit?id=bc2b7dab629a51e8beb5fda4222c62a23b729f26

> I know there are users of the line event timestamp who actively
> rely on that value being obtained from the system realtime clock.

(then in follow-up mail)
> For context, these wall clock time sensitive users are running on
> systems which are PTPv2 clients, with their system realtime clock
> synchronised to that of a local PTP Grand Master clock.  In the past,
> I have used the TTL Pulse Per Second (PPS) output of the Grand Master
> to evaluate methods of timestamping line events with wall clock time
> and it was the kernel timestamping which was most suitable for our
> application.

As Arnd stated in the thread from 2018:
"most of these clocks make no sense at all for a random user
space interface, mainly because I wouldn't trust user space
programmers to make an informed decision which of those
seven to use."

So I suspect you actually managed to make a good argument
for using the realtime clock, we didn't see that coming. :)

But why can't userspace (whether your application or libgpiod)
just call the system clock_gettime() function at the arrival of a
lineevent in that case, and this will be quick due to using vDSO
on most arch:es so no user-to-kernelspace switch will be
required?

clock_gettime() can get you any of the things IIO can get
you.

If you really need the timestamp to be as close as possible
to the actual event then I see why you want the kernel to
make the timestamp in the hard IRQ handler already, but
I just want to confirm that you really really need this.

> My suggestion (which I would be happy to implement myself) is to allow
> users to select the clock to be used for line event timestamping on a
> per line handle basis.

I suppose that makes more sense than the "global switch" in sysfs
for the whole device that IIO uses. At least it is a clear sign that the
user wants this specific type of timestamp.

> 1. Increase in processing overhead and latency of timestamp
> acquisition on line event interrupts.  Implementing the proposed
> change requires a function call to be made to the appropriate ktime
> accessor function, based on what the user has configured as the
> timestamp clock source.  In kernel versions from 5.7 to current, a
> call is made to the ktime_get_ns() function which is most likely
> inlined by the compiler.  This change will result in an actual jump
> having to be made, which will have processor and memory access
> overhead (potential I$ and D$ misses).  Then there is of course the
> overhead of resolving which function to call - either a switch
> statement or call by function pointer (probably the latter option).

Given the overall latency of the kernelspace to userspace
switch and the whole kernel executing around us this is not of
any big concern I'd say, though I will stand corrected in the
face of real-world usecase.

Yours,
Linus Walleij