diff mbox series

[XEN] x86/hpet: Disable legacy replacement mode after IRQ test if not needed

Message ID 20230718122603.2002-1-simon@invisiblethingslab.com (mailing list archive)
State New, archived
Headers show
Series [XEN] x86/hpet: Disable legacy replacement mode after IRQ test if not needed | expand

Commit Message

Simon Gaiser July 18, 2023, 12:26 p.m. UTC
As far as I understand the HPET legacy mode is not required on systems
with ARAT after the timer IRQ test. For previous discussion see [1].
Keeping it enabled prevents reaching S0ix residency.

Link: https://lore.kernel.org/xen-devel/cb408368-077d-edb5-b4ad-f80086db48c1@invisiblethingslab.com/ # [1]
Signed-off-by: Simon Gaiser <simon@invisiblethingslab.com>
---
 xen/arch/x86/io_apic.c | 4 ++++
 1 file changed, 4 insertions(+)

Comments

Roger Pau Monne July 18, 2023, 12:53 p.m. UTC | #1
On Tue, Jul 18, 2023 at 02:26:03PM +0200, Simon Gaiser wrote:
> As far as I understand the HPET legacy mode is not required on systems
> with ARAT after the timer IRQ test.

What's the relation with ARAT here?

It would seem to me that keeping legacy replacement enabled should
only be done when opt_hpet_legacy_replacement > 0, and the currently
modified block is already in a opt_hpet_legacy_replacement < 0 gated
chunk.

Thanks, Roger.
Simon Gaiser July 18, 2023, 9:51 p.m. UTC | #2
Roger Pau Monné:
> On Tue, Jul 18, 2023 at 02:26:03PM +0200, Simon Gaiser wrote:
>> As far as I understand the HPET legacy mode is not required on systems
>> with ARAT after the timer IRQ test.
> 
> What's the relation with ARAT here?
> 
> It would seem to me that keeping legacy replacement enabled should
> only be done when opt_hpet_legacy_replacement > 0, and the currently
> modified block is already in a opt_hpet_legacy_replacement < 0 gated
> chunk.

I was concerned that on systems without ARAT cpuidle might rely on HPET
legacy mode being available. See _disable_pit_irq and lapic_timer_init.
But now that I stared at this again, I think that condition isn't
actually needed. If we reach that code we know that we have no working
PIT, but HPET is working. So _disable_pit_irq which is run after
check_timer (__start_xen first calls check_timer via smp_prepare_cpus
and only later disable_pit_irq via do_initcalls) will setup HPET
broadcast, which should succeed since HPET worked previously.

So I guess we can just drop the condition (please double check, that
code is quite tangled and I'm not familiar with it).

Simon
Jan Beulich July 19, 2023, 1:32 p.m. UTC | #3
On 18.07.2023 23:51, Simon Gaiser wrote:
> Roger Pau Monné:
>> On Tue, Jul 18, 2023 at 02:26:03PM +0200, Simon Gaiser wrote:
>>> As far as I understand the HPET legacy mode is not required on systems
>>> with ARAT after the timer IRQ test.
>>
>> What's the relation with ARAT here?
>>
>> It would seem to me that keeping legacy replacement enabled should
>> only be done when opt_hpet_legacy_replacement > 0, and the currently
>> modified block is already in a opt_hpet_legacy_replacement < 0 gated
>> chunk.
> 
> I was concerned that on systems without ARAT cpuidle might rely on HPET
> legacy mode being available. See _disable_pit_irq and lapic_timer_init.
> But now that I stared at this again, I think that condition isn't
> actually needed. If we reach that code we know that we have no working
> PIT, but HPET is working. So _disable_pit_irq which is run after
> check_timer (__start_xen first calls check_timer via smp_prepare_cpus
> and only later disable_pit_irq via do_initcalls) will setup HPET
> broadcast, which should succeed since HPET worked previously.
> 
> So I guess we can just drop the condition (please double check, that
> code is quite tangled and I'm not familiar with it).

What you want to respect instead though is opt_hpet_legacy_replacement.

Jan
Simon Gaiser July 24, 2023, 9:48 a.m. UTC | #4
Jan Beulich:
> On 18.07.2023 23:51, Simon Gaiser wrote:
>> Roger Pau Monné:
>>> On Tue, Jul 18, 2023 at 02:26:03PM +0200, Simon Gaiser wrote:
>>>> As far as I understand the HPET legacy mode is not required on systems
>>>> with ARAT after the timer IRQ test.
>>>
>>> What's the relation with ARAT here?
>>>
>>> It would seem to me that keeping legacy replacement enabled should
>>> only be done when opt_hpet_legacy_replacement > 0, and the currently
>>> modified block is already in a opt_hpet_legacy_replacement < 0 gated
>>> chunk.
>>
>> I was concerned that on systems without ARAT cpuidle might rely on HPET
>> legacy mode being available. See _disable_pit_irq and lapic_timer_init.
>> But now that I stared at this again, I think that condition isn't
>> actually needed. If we reach that code we know that we have no working
>> PIT, but HPET is working. So _disable_pit_irq which is run after
>> check_timer (__start_xen first calls check_timer via smp_prepare_cpus
>> and only later disable_pit_irq via do_initcalls) will setup HPET
>> broadcast, which should succeed since HPET worked previously.
>>
>> So I guess we can just drop the condition (please double check, that
>> code is quite tangled and I'm not familiar with it).
> 
> What you want to respect instead though is opt_hpet_legacy_replacement.

Can you please explain what behavior you expect? As Roger pointed out
this code only runs with opt_hpet_legacy_replacement < 0 so the user
didn't make an explicit choice. In that case enabling the legacy mode
for the timer IRQ test and then disabling it to allow lower power modes
seems reasonable to me.

Simon
Jan Beulich July 24, 2023, 10:02 a.m. UTC | #5
On 24.07.2023 11:48, Simon Gaiser wrote:
> Jan Beulich:
>> On 18.07.2023 23:51, Simon Gaiser wrote:
>>> Roger Pau Monné:
>>>> On Tue, Jul 18, 2023 at 02:26:03PM +0200, Simon Gaiser wrote:
>>>>> As far as I understand the HPET legacy mode is not required on systems
>>>>> with ARAT after the timer IRQ test.
>>>>
>>>> What's the relation with ARAT here?
>>>>
>>>> It would seem to me that keeping legacy replacement enabled should
>>>> only be done when opt_hpet_legacy_replacement > 0, and the currently
>>>> modified block is already in a opt_hpet_legacy_replacement < 0 gated
>>>> chunk.
>>>
>>> I was concerned that on systems without ARAT cpuidle might rely on HPET
>>> legacy mode being available. See _disable_pit_irq and lapic_timer_init.
>>> But now that I stared at this again, I think that condition isn't
>>> actually needed. If we reach that code we know that we have no working
>>> PIT, but HPET is working. So _disable_pit_irq which is run after
>>> check_timer (__start_xen first calls check_timer via smp_prepare_cpus
>>> and only later disable_pit_irq via do_initcalls) will setup HPET
>>> broadcast, which should succeed since HPET worked previously.
>>>
>>> So I guess we can just drop the condition (please double check, that
>>> code is quite tangled and I'm not familiar with it).
>>
>> What you want to respect instead though is opt_hpet_legacy_replacement.
> 
> Can you please explain what behavior you expect? As Roger pointed out
> this code only runs with opt_hpet_legacy_replacement < 0 so the user
> didn't make an explicit choice.

Oh, I'm sorry. Please disregard my comment then.

Jan
diff mbox series

Patch

diff --git a/xen/arch/x86/io_apic.c b/xen/arch/x86/io_apic.c
index 9b8a972cf5..ea98d717d0 100644
--- a/xen/arch/x86/io_apic.c
+++ b/xen/arch/x86/io_apic.c
@@ -1966,6 +1966,10 @@  static void __init check_timer(void)
 
             if ( timer_irq_works() )
             {
+                if ( boot_cpu_has(X86_FEATURE_ARAT) ) {
+                    printk(XENLOG_INFO "IRQ test with HPET Legacy Replacement Mode worked. Disabling it again.\n");
+                    hpet_disable_legacy_replacement_mode();
+                }
                 local_irq_restore(flags);
                 return;
             }