Re: [RFC 00/10] hw/mos6522: VIA timer emulation fixes and improvements

[Finn Thain]

On Tue, 24 Aug 2021, Philippe Mathieu-Daudé wrote:

> On 8/24/21 12:09 PM, Finn Thain wrote:
> 
> > On a real Quadra, accesses to the SY6522 chips are slow because they are 
> > synchronous with the 783360 Hz "phase 2" clock. In QEMU, they are slow 
> > only because of the division operation in the timer count calculation.
> > 
> > This patch series improves the fidelity of the emulated chip, but the 
> > price is more division ops. I haven't tried to measure this yet.
> > 
> > The emulated 6522 still deviates from the behaviour of the real thing, 
> > however. For example, two consecutive accesses to a real 6522 timer 
> > counter can never yield the same value. This is not true of the 6522 in 
> > QEMU 6 wherein two consecutive accesses to a timer count register have 
> > been observed to yield the same value.
> > 
> > Linux is not particularly robust in the face of a 6522 that deviates 
> > from the usual behaviour. The problem presently affecting a Linux guest 
> > is that its 'via' clocksource is prone to monotonicity failure. That is, 
> > the clocksource counter can jump backwards. This can be observed by 
> > patching Linux like so:
> > 
> > diff --git a/arch/m68k/mac/via.c b/arch/m68k/mac/via.c
> > --- a/arch/m68k/mac/via.c
> > +++ b/arch/m68k/mac/via.c
> > @@ -606,6 +606,8 @@ void __init via_init_clock(void)
> >     clocksource_register_hz(&mac_clk, VIA_CLOCK_FREQ);
> >  }
> >  
> > +static u32 prev_ticks;
> > +
> >  static u64 mac_read_clk(struct clocksource *cs)
> >  {
> >     unsigned long flags;
> > @@ -631,6 +633,8 @@ static u64 mac_read_clk(struct clocksource *cs)
> >     count = count_high << 8;
> >     ticks = VIA_TIMER_CYCLES - count;
> >     ticks += clk_offset + clk_total;
> > +if (ticks < prev_ticks) pr_warn("%s: %u < %u\n", __func__, ticks, 
> > prev_ticks);
> > +prev_ticks = ticks;
> >     local_irq_restore(flags);
> >  
> >     return ticks;
> > 
> > This problem can be partly blamed on a 6522 design limitation, which is 
> > that the timer counter has no overflow register. Hence, if a timer 
> > counter wraps around and the kernel is late to handle the subsequent 
> > interrupt, the kernel can't account for any missed ticks.
> > 
> > On a real Quadra, the kernel mitigates this limitation by minimizing 
> > interrupt latency. But on QEMU, interrupt latency is unbounded. This 
> > can't be mitigated by the guest kernel at all and leads to clock drift. 
> > This can be observed by patching QEMU like so:
> > 
> > diff --git a/hw/misc/mos6522.c b/hw/misc/mos6522.c
> > --- a/hw/misc/mos6522.c
> > +++ b/hw/misc/mos6522.c
> > @@ -379,6 +379,12 @@ void mos6522_write(void *opaque, hwaddr addr, uint64_t 
> > val, unsigned size)
> >          s->pcr = val;
> >          break;
> >      case VIA_REG_IFR:
> > +        if (val & T1_INT) {
> > +            static int64_t last_t1_int_cleared;
> > +            int64_t now = qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL);
> > +            if (now - last_t1_int_cleared > 20000000) printf("\t%s: t1 int 
> > clear is late\n", __func__);
> > +            last_t1_int_cleared = now;
> > +        }
> >          /* reset bits */
> >          s->ifr &= ~val;
> >          mos6522_update_irq(s);
> > 
> > This logic asserts that, given that Linux/m68k sets CONFIG_HZ to 100, 
> > the emulator will theoretically see each timer 1 interrupt cleared 
> > within 20 ms of the previous one. But that deadline is often missed on 
> > my QEMU host [4].
> 
> I wonder if using QEMU ptimer wouldn't help here, instead of
> calling qemu_clock_get_ns() and doing the math by hand:
> 
> /* Starting to run with/setting counter to "0" won't trigger immediately,
>  * but after a one period for both oneshot and periodic modes.  */
> #define PTIMER_POLICY_NO_IMMEDIATE_TRIGGER  (1 << 2)
> 
> /* Starting to run with/setting counter to "0" won't re-load counter
>  * immediately, but after a one period.  */
> #define PTIMER_POLICY_NO_IMMEDIATE_RELOAD   (1 << 3)
> 
> /* Make counter value of the running timer represent the actual value and
>  * not the one less.  */
> #define PTIMER_POLICY_NO_COUNTER_ROUND_DOWN (1 << 4)
> 

It's often the case that a conversion to a new API is trivial for someone 
who understands that API. But I still haven't got my head around the 
implementation (hw/core/ptimer.c).

So I agree the ptimer API could simplify mos6522.c but adopting it is not 
trivial for me.

QEMU's 6522 device does not attempt to model the relationship between the 
"phase two" clock and timer counters and timer interrupts. I haven't 
attempted to fix these deviations at all, as that's not trivial either.

But using the ptimer API could potentially make it easier to address those 
fidelity issues.