[lttng-dev] LTTng - Xenomai : different results between timestamp-lttng and rt_time_read()

[Mathieu Desnoyers]

----- On May 20, 2021, at 9:56 AM, Mathieu Desnoyers mathieu.desnoyers at efficios.com wrote:

> ----- On May 20, 2021, at 9:54 AM, lttng-dev lttng-dev at lists.lttng.org wrote:
> 
>> ----- On May 20, 2021, at 5:11 AM, lttng-dev lttng-dev at lists.lttng.org wrote:
>> 
>>> Am Do., 20. Mai 2021 um 10:28 Uhr schrieb MONTET Julien
>>> <julien.montet at reseau.eseo.fr>:
>>>>
>>>> Hi Norbert,
>>>>
>>>> Thank you for your answer !
>>>>
>>>> Yes, I am using a Xenomai cobalt - xenomai is 3.1
>>>> cat /proc/xenomai/version => 3.1
>>>>
>>>> After the installation, I tested "test tools" in /proc/xenomai/ and it worked
>>>> nice.
>>> 
>>> Just asked to make sure, thought the scripts usual add some -xeno tag
>>> to the kernel version.
>>> 
>>>> What do you mean by "it might deadlock really good" ?
>>> 
>>> clock_gettime will either use a syscall (kills realtime always) or is
>>> optimized via VDSO (which very likely is your case).
>>> 
>>> What happens is that the kernel will take a spinlock, then write new
>>> values, then releases the spinlock.
>>> your program will aswell spin (but just to see if the spinlock is
>>> free), read the values and interpolates them.
>>> 
>>> But if your program interrupts the kernel while the kernel holds the
>>> lock (all on the same cpu core), then it will spin forever and the
>>> kernel will never execute.
>> 
>> Just one clarification: the specific locking strategy used by the
>> Linux kernel monotonic clock vDSO is a "seqlock", where the kernel
>> sets a bit which keeps concurrent readers looping until they observe
> 
> When I say "sets a bit", I actually mean "increment a sequence counter",
> and readers observe either odd or even state, thus knowing whether
> they need to retry, and whether the value read before/after reading
> the data structure changed.

Looking again at the Linux kernel's kernel/time/vsyscall.c implementation
of vdso_update_{begin,end}, I notice that interrupts are disabled across
the entire update. So I understand that the Interrupt pipeline (I-pipe)
interrupt gets delivered even when the kernel disables interrupts. Did
you consider modifying the I-pipe kernel patch to change the vdso update so
it updates the vdso from within an I-pipe virq handler ?

AFAIU this would allow Xenomai userspace to use the Linux kernel vDSO
clock sources.

Thanks,

Mathieu

> 
> Thanks,
> 
> Mathieu
> 
>> a consistent value. With Xenomai it indeed appears to be prone to
>> deadlock if a high priority Xenomai thread interrupts the kernel
>> while the write seqlock is held, and then proceeds to loop forever
>> on the read-side of the seqlock.
>> 
>> Note that for the in-kernel tracer clock read use-case, which
>> needs to be able to happen from NMI context, I've contributed a
>> modified version of the seqlock to the Linux kernel:
>> 
>> https://lwn.net/Articles/831540/ The seqcount latch lock type
>> 
>> It basically keeps two copies of the clock data structures, so the
>> read-side never has to loop waiting for the updater: it simply gets
>> redirected to the "stable" copy of the data.
>> 
>> The trade-off here is that with the latch lock used for clocks, a
>> reader may observe time going slightly backwards between two clock
>> reads when reading while specific clock rate adjustments are made
>> by an updater. The clock user needs to be aware of this.
>> 
>> Thanks,
>> 
>> Mathieu
>> 
>> --
>> Mathieu Desnoyers
>> EfficiOS Inc.
>> http://www.efficios.com
>> _______________________________________________
>> lttng-dev mailing list
>> lttng-dev at lists.lttng.org
>> https://lists.lttng.org/cgi-bin/mailman/listinfo/lttng-dev
> 
> --
> Mathieu Desnoyers
> EfficiOS Inc.
> http://www.efficios.com

-- 
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com