On Thu, Oct 16, 2025 at 01:57:41PM +0200, Jan Beulich wrote:
> On 16.10.2025 13:42, Roger Pau Monné wrote:
> > On Thu, Oct 16, 2025 at 09:31:42AM +0200, Jan Beulich wrote:
> >> Keeping channels enabled when they're unused is only causing problems:
> >> Extra interrupts harm performance, and extra nested interrupts could even
> >> have caused worse problems.
> >>
> >> Note that no explicit "enable" is necessary - that's implicitly done by
> >> set_channel_irq_affinity() once the channel goes into use again.
> >>
> >> Along with disabling the counter, also "clear" the channel's "next event",
> >> for it to be properly written by whatever the next user is going to want
> >> (possibly avoiding too early an IRQ).
> >>
> >> Further, along the same lines, don't enable channels early when starting
> >> up an IRQ. This similarly should happen no earlier than from
> >> set_channel_irq_affinity() (here: once a channel goes into use the very
> >> first time). This eliminates a single instance of
> >>
> >> (XEN) [VT-D]INTR-REMAP: Request device [0000:00:1f.0] fault index 0
> >> (XEN) [VT-D]INTR-REMAP: reason 25 - Blocked a compatibility format
> >> interrupt request
> >>
> >> during boot. (Why exactly there's only one instance, when we use multiple
> >> counters and hence multiple IRQs, I can't tell. My understanding would be
> >> that this was due to __hpet_setup_msi_irq() being called only after
> >> request_irq() [and hence the .startup handler], yet that should have
> >> affected all channels.)
> >>
> >> Fixes: 3ba523ff957c ("CPUIDLE: enable MSI capable HPET for timer
> >> broadcast")
> >> Signed-off-by: Jan Beulich <[email protected]>
> >> ---
> >> A window still remains for IRQs to be caused by stale comparator values:
> >> hpet_attach_channel() is called ahead of reprogram_hpet_evt_channel().
> >> Should we also write the comparator to "far into the future"?
> >
> > It might be helpful to reprogram the comparator as far ahead as
> > possible in hpet_attach_channel() ahead of enabling it, or
> > alternatively in hpet_detach_channel().
>
> The downside is yet another (slow) MMIO access. Hence why I didn't make
> such a change right away. Plus I wasn't quite sure about the locking there:
> Imo if we did so, it would be better if the lock wasn't dropped
> intermediately.
>
> >> @@ -542,6 +540,8 @@ static void hpet_detach_channel(unsigned
> >> spin_unlock_irq(&ch->lock);
> >> else if ( (next = cpumask_first(ch->cpumask)) >= nr_cpu_ids )
> >> {
> >> + hpet_disable_channel(ch);
> >> + ch->next_event = STIME_MAX;
> >> ch->cpu = -1;
> >> clear_bit(HPET_EVT_USED_BIT, &ch->flags);
> >> spin_unlock_irq(&ch->lock);
> >
> > I'm a bit confused with what the HPET code does here (don't know
> > enough about it, and there are no comments). Why is the timer rotated
> > to a CPU in ch->cpumask once disabled, instead of just being plain
> > disabled?
>
> Because it will still be needed by the other CPUs that the channel is
> shared with.
Yeah, missed that part, the channel is migrated to a different CPU. I
wonder however: since an active channel can be migrated around between
CPUs, isn't there a risk of the timer firing just in the middle of
migration (when interrupt generation is disabled), and hence Xen
possibly missing a deadline?
In hpet_broadcast_exit() we need to check whether the timer has
expired after the migration, and manually trigger a broadcast if
needed. This also risks doing to broadcasts also back-to-back, but
it's the only option I see to avoid missing a deadline.
Maybe there's something I'm missing, this is all fairly complex.
Thanks, Roger.
