Re: Dovetail 7.1-rc6 updates pushed

On Mon, 2026-06-01 at 20:28 +0200, Philippe Gerum wrote:
> Florian Bezdeka <[email protected]> writes:
> 
> > On Mon, 2026-06-01 at 15:43 +0200, Philippe Gerum wrote:
> > > Florian Bezdeka <[email protected]> writes:
> > > 
> > > > Hi all,
> > > > 
> > > > I just pushed the following branch updates for Dovetail based on 7.1-
> > > > rc6:
> > > > 
> > > > Dovetail:
> > > > wip/flo/v7.1-dovetail-rebase
> > > > wip/flo/v7.1-dovetail-rebase+riscv
> > > > 
> > > > Evl:
> > > > wip/flo/v7.1-evl-rebase
> > > > wip/flo/v7.1-evl-rebase+riscv
> > > > 
> > > > The +riscv branches are containing v8 of Tobias' riscv enabling series,
> > > > maybe we need v9, let's see.
> > > > 
> > > > Any additional testing on those branches is highly appreciated.
> > > > 
> > > > @Philippe:
> > > > I'm still waiting for some feedback for [1]. That seems now to be a
> > > > invalid combination of kconfigs, which is dovetail specific. Upstream
> > > > thread at [2].
> > > > 
> > > > Can you remember why KVM_INTEL is not selecting X86_FRED in case
> > > > IRQ_PIPELINE is enabled? I haven't checked the details yet, nor tried to
> > > > remove the dovetail specific adjustment.
> > > > 
> > > 
> > > Maybe because the pipeline never supported X86_FRED, doing so would
> > > required to rework the irq-entry trampoline macros, which is deemed
> > > painful work.
> > 
> > I sent out a RFC patch for that, input welcome.
> > 
> > Meanwhile I can provide some updates on the testing part. The following
> > problems seem to be reproducible on 7.1-rc5 and 7.1-rc6 - while 7.1-rc4
> > was fine (at least once):
> > 
> > x86   x3: Test timeout. Something seems slower than expected.
> > arm64 x4: Same problem
> > 
> > To rule out problems in our CI setup I spawned some 7.0 tests, which
> > were fine, at least twice in a row. Seems someone has to look into that.
> > As I'm running out of time: Any resources available?
> > 
> 
> Not here, at least not yet.

I scheduled a couple of overnight test runs on -rc4. Those test are
looking OK. Something in -rc5 seems to trigger that. The problem seems
qemu/kvm specific, test runs on real hardware were OK.

I found a couple of recent discussions around VMX/SVM regressions, some
of them related to the delayed hrtimer arming. As we disable the delayed
arming completely, that should not be our problem. Maybe we should just
wait for -rc7?