Re: [PATCH v2 3/4] target/arm: Prepare generic timer for per-platform CNTFRQ

["Andrew Jeffery" <andrew@aj.id.au>]

On Tue, 3 Dec 2019, at 16:49, Philippe Mathieu-Daudé wrote:
> On 12/3/19 5:14 AM, Andrew Jeffery wrote:
> > The ASPEED AST2600 clocks the generic timer at the rate of HPLL. On
> > recent firmwares this is at 1125MHz, which is considerably quicker than
> > the assumed 62.5MHz of the current generic timer implementation. The
> > delta between the value as read from CNTFRQ and the true rate of the
> > underlying QEMUTimer leads to sticky behaviour in AST2600 guests.
> > 
> > Add a feature-gated property exposing CNTFRQ for ARM CPUs providing the
> > generic timer. This allows platforms to configure CNTFRQ (and the
> > associated QEMUTimer) to the appropriate frequency prior to starting the
> > guest.
> > 
> > As the platform can now determine the rate of CNTFRQ we're exposed to
> > limitations of QEMUTimer that didn't previously materialise: In the
> > course of emulation we need to arbitrarily and accurately convert
> > between guest ticks and time, but we're constrained by QEMUTimer's use
> > of an integer scaling factor. The effect is QEMUTimer cannot exactly
> > capture the period of frequencies that do not cleanly divide
> > NANOSECONDS_PER_SECOND for scaling ticks to time. As such, provide an
> > equally inaccurate scaling factor for scaling time to ticks so at least
> > a self-consistent inverse relationship holds.
> > 
> > Signed-off-by: Andrew Jeffery <andrew@aj.id.au>
> > Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
> > ---
> >   target/arm/cpu.c    | 43 +++++++++++++++++++++++++++++++++----------
> >   target/arm/cpu.h    | 18 ++++++++++++++++++
> >   target/arm/helper.c |  9 ++++++++-
> >   3 files changed, 59 insertions(+), 11 deletions(-)
> > 
> > diff --git a/target/arm/cpu.c b/target/arm/cpu.c
> > index 5698a74061bb..f186019a77fd 100644
> > --- a/target/arm/cpu.c
> > +++ b/target/arm/cpu.c
> > @@ -974,10 +974,12 @@ static void arm_cpu_initfn(Object *obj)
> >       if (tcg_enabled()) {
> >           cpu->psci_version = 2; /* TCG implements PSCI 0.2 */
> >       }
> > -
> > -    cpu->gt_cntfrq = NANOSECONDS_PER_SECOND / GTIMER_SCALE;
> >   }
> >   
> > +static Property arm_cpu_gt_cntfrq_property =
> > +            DEFINE_PROP_UINT64("cntfrq", ARMCPU, gt_cntfrq,
> > +                               NANOSECONDS_PER_SECOND / GTIMER_SCALE);
> > +
> >   static Property arm_cpu_reset_cbar_property =
> >               DEFINE_PROP_UINT64("reset-cbar", ARMCPU, reset_cbar, 0);
> >   
> > @@ -1174,6 +1176,11 @@ void arm_cpu_post_init(Object *obj)
> >   
> >       qdev_property_add_static(DEVICE(obj), &arm_cpu_cfgend_property,
> >                                &error_abort);
> > +
> > +    if (arm_feature(&cpu->env, ARM_FEATURE_GENERIC_TIMER)) {
> > +        qdev_property_add_static(DEVICE(cpu), &arm_cpu_gt_cntfrq_property,
> > +                                 &error_abort);
> > +    }
> >   }
> >   
> >   static void arm_cpu_finalizefn(Object *obj)
> > @@ -1253,14 +1260,30 @@ static void arm_cpu_realizefn(DeviceState *dev, Error **errp)
> >           }
> >       }
> >   
> > -    cpu->gt_timer[GTIMER_PHYS] = timer_new(QEMU_CLOCK_VIRTUAL, GTIMER_SCALE,
> > -                                           arm_gt_ptimer_cb, cpu);
> > -    cpu->gt_timer[GTIMER_VIRT] = timer_new(QEMU_CLOCK_VIRTUAL, GTIMER_SCALE,
> > -                                           arm_gt_vtimer_cb, cpu);
> > -    cpu->gt_timer[GTIMER_HYP] = timer_new(QEMU_CLOCK_VIRTUAL, GTIMER_SCALE,
> > -                                          arm_gt_htimer_cb, cpu);
> > -    cpu->gt_timer[GTIMER_SEC] = timer_new(QEMU_CLOCK_VIRTUAL, GTIMER_SCALE,
> > -                                          arm_gt_stimer_cb, cpu);
> > +
> > +    {
> > +        uint64_t scale;
> 
> Apparently you have to use this odd indent due to the '#ifndef 
> CONFIG_USER_ONLY'. Well, acceptable.

It's the indent associated with the block scope for the scale variable to limit its lifetime
to where I needed it.

> 
> > +
> > +        if (arm_feature(env, ARM_FEATURE_GENERIC_TIMER)) {
> > +            if (!cpu->gt_cntfrq) {
> > +                error_setg(errp, "Invalid CNTFRQ: %"PRId64"Hz",
> > +                           cpu->gt_cntfrq);
> > +                return;
> > +            }
> > +            scale = gt_cntfrq_period_ns(cpu);
> > +        } else {
> > +            scale = GTIMER_SCALE;
> > +        }
> > +
> > +        cpu->gt_timer[GTIMER_PHYS] = timer_new(QEMU_CLOCK_VIRTUAL, scale,
> > +                                               arm_gt_ptimer_cb, cpu);
> > +        cpu->gt_timer[GTIMER_VIRT] = timer_new(QEMU_CLOCK_VIRTUAL, scale,
> > +                                               arm_gt_vtimer_cb, cpu);
> > +        cpu->gt_timer[GTIMER_HYP] = timer_new(QEMU_CLOCK_VIRTUAL, scale,
> > +                                              arm_gt_htimer_cb, cpu);
> > +        cpu->gt_timer[GTIMER_SEC] = timer_new(QEMU_CLOCK_VIRTUAL, scale,
> > +                                              arm_gt_stimer_cb, cpu);
> > +    }
> >   #endif
> >   
> >       cpu_exec_realizefn(cs, &local_err);
> > diff --git a/target/arm/cpu.h b/target/arm/cpu.h
> > index 666c03871fdf..0bcd13dcac81 100644
> > --- a/target/arm/cpu.h
> > +++ b/target/arm/cpu.h
> > @@ -939,6 +939,24 @@ struct ARMCPU {
> >   
> >   static inline unsigned int gt_cntfrq_period_ns(ARMCPU *cpu)
> >   {
> > +    /*
> > +     * The exact approach to calculating guest ticks is:
> > +     *
> > +     *     muldiv64(qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL), cpu->gt_cntfrq,
> > +     *              NANOSECONDS_PER_SECOND);
> > +     *
> > +     * We don't do that. Rather we intentionally use integer division
> > +     * truncation below and in the caller for the conversion of host monotonic
> > +     * time to guest ticks to provide the exact inverse for the semantics of
> > +     * the QEMUTimer scale factor. QEMUTimer's scale facter is an integer, so
> > +     * it loses precision when representing frequencies where
> > +     * `(NANOSECONDS_PER_SECOND % cpu->gt_cntfrq) > 0` holds. Failing to
> > +     * provide an exact inverse leads to scheduling timers with negative
> > +     * periods, which in turn leads to sticky behaviour in the guest.
> > +     *
> > +     * Finally, CNTFRQ is effectively capped at 1GHz to ensure our scale factor
> > +     * cannot become zero.
> > +     */
> 
> This comment belong to the previous patch.

Sort of, but also sort of not. We don't expose the limitation until this patch as
NANOSECONDS_PER_SECOND is an integer multiple of GTIMER_SCALE, which
is what gt_cntfrq is set to until we add the property to configure it to arbitrary
values in this patch. So I added the comment in this patch rather than the
previous one which adds the code.

Andrew