On 09/01/2023 16.39, Jonathan Cameron wrote:
On Mon, 9 Jan 2023 08:54:05 +0100
Sean Nyekjaer <sean@xxxxxxxxxx> wrote:
Hi Jonathan and list,
We are getting some weird timestamps (with fxls8962af) in the iio
buffers when time is set back by NTP.
Here I have dumped old_timestamp, timestamp from the
fxls8962af_fifo_flush():
old_timestamp 1673008488434717043, timestamp 1673008489457042043
*old_timestamp 1673008489457042043, timestamp 1673008485198504719*
old_timestamp 1673008485198504719, timestamp 1673008486220571510
Time is set back in the middle data set, so the sample timestamp looks like:
tstamp0 3091988802491735451
tstamp1 4510969115526428858
tstamp2 5929949428561122265
tstamp3 7348929741595815672
tstamp4 8767910054630509079
tstamp5 -8259853706044349130
tstamp6 -6840873393009655723
tstamp7 -5421893079974962316
tstamp8 -4002912766940268909
tstamp9 -2583932453905575502
tstamp10 -1164952140870882095
tstamp11 254028172163811312
tstamp12 1673008485198504719
Possible solutions?
Could we store the sample time from the last collected data set?
And use that if old_timestamp > timestamp?
Ah. I guess this is the timestamp interpolation code going wrong.
It should be possible to make that code 'sort of cope' with negatives
though it would be a slightly nasty heuristic.
You would make it go backwards a little for each fifo entry..
Yeah agree, but then you could have 2 (or 3) samples with the same
timestamp :/
Or do you have any other idea's?
This mess is a side effect of me picking a silly default clock
a long time back (can't fix that as it's ABI...)
We could smooth it in the driver, but if you have
the option, I'd just change the default clock from userspace.
Take a look at current_timestamp_clock in sysfs.
You should be fine with "monotonic"
We will switch to the monotonic clock for now :)
Thanks
/Sean
