* Re: Posix periodic time first tick
2020-08-10 7:52 ` Petr Červenka
@ 2020-08-10 8:54 ` Per Oberg
2020-08-10 9:49 ` Philippe Gerum
1 sibling, 0 replies; 6+ messages in thread
From: Per Oberg @ 2020-08-10 8:54 UTC (permalink / raw)
To: xenomai; +Cc: Petr Červenka
----- Den 10 aug 2020, på kl 9:52, Petr Červenka grugh@centrum.cz skrev:
> Hello.
> The problem is only when I use Xenomai linking flags and only with
> thestart time. Period is OK.
> That aprox. 16s time delay can be "compensated" by substracting of any
> number of seconds from the start time.
You could try using CLOCK_REALTIME or CLOCK_HOST_REALTIME instead of CLOCK_MONOTONIC. I use timerfd and similar heuristics without issues afaik.
>From the top of my head I don't recall exactly how Xenomai timers correspond to Linux timers but you may have an issue if you check your starting time a non RT thread and use that value in a RT thread, depending on your choice of timer.
Check out: https://gitlab.denx.de/Xenomai/xenomai/-/wikis/Porting_To_Xenomai_POSIX
I have had some issues with time-keeping of my machines and I had to do a bit of manual tuning for it to work properly (talking 0.x pct drifts but also a bit of a offset). For debugging I used the following:
* Hard coding freq using boot time kernel argument: "clocksource=tsc tsc=reliable xenomai.clockfreq=2807909955"
* Using (in my case ) "adjtimex -f 2080674" to adjust the time-keeping scaling
* "clocktest" from Xenomai test-suite.
Right now I have (for a setting tuned to some machine and then run on another)
CPU ToD offset [us] ToD drift [us/s] warps max delta [us]
--- -------------------- ---------------- ---------- --------------
0 110881217.1 316239.964 0 0.0
1 110881224.7 316239.943 0 0.0
2 110881216.8 316239.928 0 0.0
3 110881217.4 316239.921 0 0.0
Which is about 110s and drifting 0.3s per second.
Regards
Per Öberg
> // Create timer
> timer_fd = timerfd_create(CLOCK_MONOTONIC, 0);
> // Get current time
> struct timespec start;
> clock_gettime(CLOCK_MONOTONIC, &start);
> // Set periodic
> struct itimerspec timer_conf;
> memset(&timer_conf, 0, sizeof (timer_conf));
> timer_conf.it_value = timespec_add(start, -15 * NANOSEC_PER_SEC);
> // compensation of the delay
> ...
> timerfd_settime(timer_fd, TFD_TIMER_ABSTIME, &timer_conf, NULL);
> There is strange unexpected difference/time shift between product of
> clock_gettime and timerfd_settime. Only with Xenomai posix wrappers.
> Best regards
> Petr Cervenka
> Dne 10.08.2020 v 8:45 Per Oberg napsal(a):
> > ----- Den 7 aug 2020, på kl 17:12, xenomai xenomai@xenomai.org skrev:
> >> Hello.
> >> I have created small example of periodic task in posix.
> >> If I compile it without xenomai it works fine.
> >> When I add Xenomai linking flags, the first tick of periodic task take around
> >> 16s. After that it continues as it should.
> >> Used clock type for timer_fd is CLOCK_MONOTONIC. The example should in
> >> attachment.
> >> Do you have any suggestions?
>> Just had a extremely quick look at it. Are you sure you don't get any rounding
> > errors when doing your period arithmetics ?
> >> Best regards
> >> -------------- next part --------------
> >> A non-text attachment was scrubbed...
> >> Name: timer_test.cpp
> >> Type: text/x-c++src
> >> Size: 4351 bytes
> >> Desc: timer_test.cpp
> >> URL:
> >> <http://xenomai.org/pipermail/xenomai/attachments/20200807/d4203f59/attachment.cpp>
> > Per Öberg
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: Posix periodic time first tick
2020-08-10 7:52 ` Petr Červenka
2020-08-10 8:54 ` Per Oberg
@ 2020-08-10 9:49 ` Philippe Gerum
2020-08-18 15:21 ` Petr Červenka
1 sibling, 1 reply; 6+ messages in thread
From: Philippe Gerum @ 2020-08-10 9:49 UTC (permalink / raw)
To: Petr Červenka, Per Oberg, xenomai
On 8/10/20 9:52 AM, Petr Červenka via Xenomai wrote:
> Hello.
>
> The problem is only when I use Xenomai linking flags and only with thestart
> time. Period is OK.
>
> That aprox. 16s time delay can be "compensated" by substracting of any number
> of seconds from the start time.
>
> // Create timer
> timer_fd = timerfd_create(CLOCK_MONOTONIC, 0);
>
> // Get current time
> struct timespec start;
> clock_gettime(CLOCK_MONOTONIC, &start);
>
> // Set periodic
> struct itimerspec timer_conf;
> memset(&timer_conf, 0, sizeof (timer_conf));
> timer_conf.it_value = timespec_add(start, -15 * NANOSEC_PER_SEC); //
> compensation of the delay
> ...
> timerfd_settime(timer_fd, TFD_TIMER_ABSTIME, &timer_conf, NULL);
>
> There is strange unexpected difference/time shift between product of
> clock_gettime and timerfd_settime. Only with Xenomai posix wrappers.
>
To debug that kind of issues, you first need to check whether the symbol
wrapping is ok in the executable image. A common way to do this is to do a
visual inspection of the name list with nm, e.g. for the 'latency' program
which uses the timerfd services, one could do:
{rpm@cobalt} aarch64-linux-gnu-nm -o .../latency | grep -e __wrap
.../latency: U __wrap_clock_gettime
.../latency: U __wrap_close
.../latency: U __wrap_fclose
.../latency: U __wrap_fprintf
.../latency: U __wrap_fputc
.../latency: U __wrap_fputs
.../latency: U __wrap_free
.../latency: U __wrap_fwrite
.../latency: U __wrap_ioctl
.../latency: U __wrap_kill
.../latency:0000000000402728 W __wrap_main
.../latency: U __wrap_open
.../latency: U __wrap_printf
.../latency: U __wrap_pthread_attr_init
.../latency: U __wrap_pthread_create
.../latency: U __wrap_pthread_join
.../latency: U __wrap_pthread_setname_np
.../latency: U __wrap_puts
.../latency: U __wrap_read
.../latency: U __wrap_sem_close
.../latency: U __wrap_sem_open
.../latency: U __wrap_sem_post
.../latency: U __wrap_sem_unlink
.../latency: U __wrap_sem_wait
.../latency: U __wrap_time
.../latency: U __wrap_timerfd_create
.../latency: U __wrap_timerfd_settime
.../latency: U __wrap_write
We can see that all calls to the timerfd and clock_gettime services are
wrapped to libcobalt, which is right.
Weird delays may be caused by discrepancies in the call set, e.g. feeding
cobalt's timerfd services with the output of glibc's regular clock_gettime()
call. Since cobalt clocks defines epochs which differ from the common
kernel's, they won't mix nicely.
IOW, the issue may be in your build recipe, affecting such wrapping, not in
your code.
PS: you could also annotate the calls to clock_gettime() and friends in a way
which guarantees that the cobalt version is going to be used, e.g.
ret = __RT(clock_gettime(...));
...
ret = __RT(timerfd_ceate(...));
Granted, this looks ugly, but in some cases this is helpful, particularly in
order to figure out any issue with the wrapping.
--
Philippe.
^ permalink raw reply [flat|nested] 6+ messages in thread