Re: [PATCH v4 01/12] x86/rtc: drop code related to strict mode

[Jan Beulich <jbeulich@suse.com>]

On 03.05.2021 11:28, Roger Pau Monné wrote:
> On Thu, Apr 29, 2021 at 04:53:07PM +0200, Jan Beulich wrote:
>> On 20.04.2021 16:07, Roger Pau Monne wrote:
>>> --- a/xen/arch/x86/hvm/rtc.c
>>> +++ b/xen/arch/x86/hvm/rtc.c
>>> @@ -46,15 +46,6 @@
>>>  #define epoch_year     1900
>>>  #define get_year(x)    (x + epoch_year)
>>>  
>>> -enum rtc_mode {
>>> -   rtc_mode_no_ack,
>>> -   rtc_mode_strict
>>> -};
>>> -
>>> -/* This must be in sync with how hvmloader sets the ACPI WAET flags. */
>>> -#define mode_is(d, m) ((void)(d), rtc_mode_##m == rtc_mode_no_ack)
>>> -#define rtc_mode_is(s, m) mode_is(vrtc_domain(s), m)
>>
>> Leaving aside my concerns about this removal, I think some form of
>> reference to hvmloader and its respective behavior should remain
>> here, presumably in form of a (replacement) comment.
> 
> What about adding a comment in rtc_pf_callback:
> 
> /*
>  * The current RTC implementation will inject an interrupt regardless
>  * of whether REG_C has been read since the last interrupt was
>  * injected. This is why the ACPI WAET 'RTC good' flag must be
>  * unconditionally set by hvmloader.
>  */

For one I'm unconvinced this is "must"; I think it is "may". We're
producing excess interrupts for an unaware guest, aiui. Presumably most
guests can tolerate this, but - second - it may be unnecessary overhead.
Which in turn may be why nobody has complained so far, as this sort of
overhead my be hard to notice. I also suspect the RTC may not be used
very often for generating a periodic interrupt. (I've also not seen the
flag named "RTC good" - the ACPI constant is ACPI_WAET_RTC_NO_ACK, for
example.)

>>> @@ -337,8 +336,7 @@ int pt_update_irq(struct vcpu *v)
>>>      {
>>>          if ( pt->pending_intr_nr )
>>>          {
>>> -            /* RTC code takes care of disabling the timer itself. */
>>> -            if ( (pt->irq != RTC_IRQ || !pt->priv) && pt_irq_masked(pt) &&
>>> +            if ( pt_irq_masked(pt) &&
>>>                   /* Level interrupts should be asserted even if masked. */
>>>                   !pt->level )
>>>              {
>>
>> I'm struggling to relate this to any other part of the patch. In
>> particular I can't find the case where a periodic timer would be
>> registered with RTC_IRQ and a NULL private pointer. The only use
>> I can find is with a non-NULL pointer, which would mean the "else"
>> path is always taken at present for the RTC case (which you now
>> change).
> 
> Right, the else case was always taken because as the comment noted RTC
> would take care of disabling itself (by calling destroy_periodic_time
> from the callback when using strict_mode). When no_ack mode was
> implemented this wasn't taken into account AFAICT, and thus the RTC
> was never removed from the list even when masked.
> 
> I think with no_ack mode the RTC shouldn't have this specific handling
> in pt_update_irq, as it should behave like any other virtual timer.
> I could try to split this as a separate bugfix, but then I would have
> to teach pt_update_irq to differentiate between strict_mode and no_ack
> mode.

A fair part of my confusion was about "&& !pt->priv". I've looked back
at 9607327abbd3 ("x86/HVM: properly handle RTC periodic timer even when
!RTC_PIE"), where this was added. It was, afaict, to cover for
hpet_set_timer() passing NULL with RTC_IRQ. Which makes me suspect that
be07023be115 ("x86/vhpet: add support for level triggered interrupts")
may have subtly broken things.

> Would you be fine if the following is added to the commit message
> instead:
> 
> "Note that the special handling of the RTC timer done in pt_update_irq
> is wrong for the no_ack mode, as the RTC timer callback won't disable
> the timer anymore when it detects the guest is not reading REG_C. As
> such remove the code as part of the removal of strict_mode, and don't
> special case the RTC timer anymore in pt_update_irq."

Not sure yet - as per above I'm still not convinced this part of the
change is correct.

Jan