Re: [PATCH] accel/tcg: Make round-robin kick period configurable

[Ethan Chen via <qemu-devel@nongnu.org>]

Hi Philippe,

Thank you for your response.

Since TCGState is currently defined only in tcg-all.c, we’re unable to access
rr_kick_delay_ns from other .c files through the struct.

One potential approach would be to introduce a tcg-all.h header and move the
TCGState definition there. This would allow us to access the struct in other
components by including the header and using TCG_STATE(current_accel()).

Would you be open to this direction, or do you have any alternative 
suggestions?

Thanks,
Ethan

On Thu, Jun 05, 2025 at 09:21:55AM +0200, Philippe Mathieu-Daudé wrote:
> Hi Ethan,
> 
> On 5/6/25 08:18, Ethan Chen via wrote:
> > This change introduces a configurable round-robin kick period, giving users the
> > flexibility to balance SMP simulation accuracy and performance according to
> > their specific needs.
> > 
> > The round-robin kick period is the time one vCPU can run before scheduler
> > switches to another vCPU when using a single thread TCG. The default value of
> > 0.1 seconds may allow one vCPU to run for too long before the scheduler
> > switches to another. This behavior may not be suitable for workloads with
> > strict timing requirements.
> > 
> > Reducing the period can improve the fidelity of SMP simulation by allowing
> > more frequent vCPU switching, though it may negatively impact overall
> > simulation performance.
> > 
> > Signed-off-by: Ethan Chen <ethan84@andestech.com>
> > ---
> >   accel/tcg/tcg-accel-ops-rr.c |  2 +-
> >   accel/tcg/tcg-accel-ops-rr.h |  2 +-
> >   accel/tcg/tcg-all.c          | 35 +++++++++++++++++++++++++++++++++++
> >   qemu-options.hx              |  9 ++++++++-
> >   4 files changed, 45 insertions(+), 3 deletions(-)
> 
> 
> > diff --git a/accel/tcg/tcg-accel-ops-rr.h b/accel/tcg/tcg-accel-ops-rr.h
> > index 2a76a29612..324bb772cb 100644
> > --- a/accel/tcg/tcg-accel-ops-rr.h
> > +++ b/accel/tcg/tcg-accel-ops-rr.h
> > @@ -10,7 +10,7 @@
> >   #ifndef TCG_ACCEL_OPS_RR_H
> >   #define TCG_ACCEL_OPS_RR_H
> > 
> > -#define TCG_KICK_PERIOD (NANOSECONDS_PER_SECOND / 10)
> > +extern uint64_t rr_kick_period;
> 
> No need for another extern, pass it as argument:
> 
> -- >8 --
> diff --git a/accel/tcg/tcg-accel-ops-rr.c b/accel/tcg/tcg-accel-ops-rr.c
> index f62cf24e1d4..551864b5509 100644
> --- a/accel/tcg/tcg-accel-ops-rr.c
> +++ b/accel/tcg/tcg-accel-ops-rr.c
> @@ -62,9 +62,9 @@ void rr_kick_vcpu_thread(CPUState *unused)
>  static QEMUTimer *rr_kick_vcpu_timer;
>  static CPUState *rr_current_cpu;
> 
> -static inline int64_t rr_next_kick_time(void)
> +static inline int64_t rr_next_kick_time(uint64_t kick_delay_ns)
>  {
> -    return qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + TCG_KICK_PERIOD;
> +    return qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + kick_delay_ns;
>  }
> 
>  /* Kick the currently round-robin scheduled vCPU to next */
> @@ -83,15 +83,20 @@ static void rr_kick_next_cpu(void)
> 
>  static void rr_kick_thread(void *opaque)
>  {
> -    timer_mod(rr_kick_vcpu_timer, rr_next_kick_time());
> +    CPUState *cpu = opaque;
> +
> +    // here use rr_next_kick_time(cpu->accel->rr_kick_delay_ns):
> +    timer_mod(rr_kick_vcpu_timer, rr_next_kick_time(TCG_KICK_PERIOD));
>      rr_kick_next_cpu();
>  }
> 
>  static void rr_start_kick_timer(void)
>  {
> -    if (!rr_kick_vcpu_timer && CPU_NEXT(first_cpu)) {
> +    CPUState *next_cpu = CPU_NEXT(first_cpu);
> +
> +    if (!rr_kick_vcpu_timer && next_cpu) {
>          rr_kick_vcpu_timer = timer_new_ns(QEMU_CLOCK_VIRTUAL,
> -                                           rr_kick_thread, NULL);
> +                                           rr_kick_thread, next_cpu);
>      }
>      if (rr_kick_vcpu_timer && !timer_pending(rr_kick_vcpu_timer)) {
>          timer_mod(rr_kick_vcpu_timer, rr_next_kick_time());
> ---
> 
> > diff --git a/accel/tcg/tcg-all.c b/accel/tcg/tcg-all.c
> > index 6e5dc333d5..69390020aa 100644
> > --- a/accel/tcg/tcg-all.c
> > +++ b/accel/tcg/tcg-all.c
> > @@ -36,6 +36,7 @@
> >   #include "qapi/qapi-builtin-visit.h"
> >   #include "qemu/units.h"
> >   #include "qemu/target-info.h"
> > +#include "qemu/timer.h"
> >   #ifndef CONFIG_USER_ONLY
> >   #include "hw/boards.h"
> >   #endif
> > @@ -50,6 +51,7 @@ struct TCGState {
> >       bool one_insn_per_tb;
> >       int splitwx_enabled;
> >       unsigned long tb_size;
> > +    uint64_t rr_kick_period;
> 
> 'rr_kick_delay_ns' seems more accurate.
> 
> >   };
> >   typedef struct TCGState TCGState;
> > 
> > @@ -76,9 +78,11 @@ static void tcg_accel_instance_init(Object *obj)
> >   #else
> >       s->splitwx_enabled = 0;
> >   #endif
> > +    s->rr_kick_period = NANOSECONDS_PER_SECOND / 10;
> >   }
> > 
> >   bool one_insn_per_tb;
> > +uint64_t rr_kick_period;
> 
> (Drop)
> 
> > 
> >   static int tcg_init_machine(MachineState *ms)
> >   {
> > @@ -125,6 +129,7 @@ static int tcg_init_machine(MachineState *ms)
> >   #endif
> > 
> >       tcg_allowed = true;
> > +    rr_kick_period = s->rr_kick_period;
> 
> (Drop)
> 
> > 
> >       page_init();
> >       tb_htable_init();
> > @@ -234,6 +239,30 @@ static int tcg_gdbstub_supported_sstep_flags(void)
> >       }
> >   }
> 
> Patch LGTM otherwise.
> 
> Regards,
> 
> Phil.