Re: Possible Xenomai 3 latency regression on arm64

Philippe Gerum <[email protected]> Mon, 22 Jun 2026 10:53:59 +0200
Newsgroups dev.linux.lists.xenomai
Message-ID <[email protected]>
Florian Bezdeka <[email protected]> writes:

> On Mon, 2026-06-22 at 09:13 +0200, Philippe Gerum wrote:
>> Philippe Gerum <[email protected]> writes:
>> 
>> > Jan Kiszka <[email protected]> writes:
>> > 
>> > > On 15.06.26 09:35, Philippe Gerum wrote:
>> > > > Jan Kiszka <[email protected]> writes:
>> > > > 
>> > > > > Hi all,
>> > > > > 
>> > > > > it seems we got some >10% worst-case latency increase over the last
>> > > > > month with Xenomai head and 6.12 - that is at least the setup we use in
>> > > > > our long-running weekly test.
>> > > > > 
>> > > > > Tonight, the RPi4 test triggered by exceeding the configured limit of
>> > > > > 150 µs [1]. About half a year ago, we were in 130 µs range with it [2].
>> > > > > x86 and beaglebone are apparently not seeing this change.
>> > > > > 
>> > > > 
>> > > > My rpi4 fried. Is the issue visible with x4 as well?
>> > > > 
>> > > 
>> > > We do not have long-running latency tests for x4 yet, do we, Tobias?
>> > > 
>> > 
>> > Ok, I can try this on an odroid c4. Did anyone succeeded in booting a
>> > mainline kernel on pi5? Last time I checked, root-over-nfs was not
>> > possible yet.
>> 
>> A follow up on the latency issue. I did notice a regression in the
>> latency figures with v4 too, specifically on arm64. The issue was
>> introduced in v6.12.85-cip22-dovetail3-rebase, where vDSO support for
>> oob callers was unexpectedly disabled due to applying 682c4aa14fe6. That
>> change is correct starting from v6.18.y, but should not have been
>> applied to v6.12.y. The bug caused a ~12% increase in the worst-case
>> latency figures on odroid c4.
>> 
>> With that change reverted, I'm back to normal on c4 with
>> v6.12.90-cip24-dovetail2-rebase. A couple of additional fixes to the
>> generic lib/vdso support went in as well, but those have no effect on
>> latency.
>> 
>> All dovetail and v4/evl branches have been updated accordingly.
>
> All those changes are not yet part of the 7.1 branches (which have been
> recently rebased to the first 7.1 release). Will you (Philippe) take
> care of integrating that or should I pick it?

If you already merged the latest changes from 7.0 into 7.1 on the fly,
please pick them, otherwise I'll do it.

-- 
Philippe.