Re: [PATCH v1 08/19] xen/sysctl: introduce CONFIG_PM_STATISTIC

[Jan Beulich <jbeulich@xxxxxxxx>]

On 12.03.2025 05:06, Penny Zheng wrote:
> This commit introduces CONFIG_PM_STATISTIC for wrapping all operations
> regarding performance management statistic operations.

Please can descriptions not use "This patch ..." or "This commit ..." or
anyhting like these?

> --- a/xen/common/Kconfig
> +++ b/xen/common/Kconfig
> @@ -107,6 +107,11 @@ config NEEDS_LIBELF
>  config NUMA
>       bool
>  
> +config PM_STATISTIC

I think this wants to use plural, i.e. PM_STATISTICS or PM_STATS.

> +        bool "Enable Performance Management Statistic Operations"

"Enable Performance Management Statistics" ?

> +        depends on ACPI && HAS_CPUFREQ
> +        default y

Nit: Indentation.

> --- a/xen/common/sysctl.c
> +++ b/xen/common/sysctl.c
> @@ -170,7 +170,7 @@ long do_sysctl(XEN_GUEST_HANDLE_PARAM(xen_sysctl_t) 
> u_sysctl)
>          op->u.availheap.avail_bytes <<= PAGE_SHIFT;
>          break;
>  
> -#if defined (CONFIG_ACPI) && defined (CONFIG_HAS_CPUFREQ)
> +#ifdef CONFIG_PM_STATISTIC
>      case XEN_SYSCTL_get_pmstat:
>          ret = do_get_pm_info(&op->u.get_pmstat);
>          break;
> @@ -180,7 +180,7 @@ long do_sysctl(XEN_GUEST_HANDLE_PARAM(xen_sysctl_t) 
> u_sysctl)
>          if ( ret == -EAGAIN )
>              copyback = 1;
>          break;
> -#endif
> +#endif /* CONFIG_PM_STATISTIC */
>  
>      case XEN_SYSCTL_page_offline_op:
>      {
> diff --git a/xen/drivers/acpi/Makefile b/xen/drivers/acpi/Makefile
> index 2fc5230253..70156ee0a4 100644
> --- a/xen/drivers/acpi/Makefile
> +++ b/xen/drivers/acpi/Makefile
> @@ -5,7 +5,7 @@ obj-$(CONFIG_X86) += apei/
>  obj-bin-y += tables.init.o
>  obj-$(CONFIG_ACPI_NUMA) += numa.o
>  obj-y += osl.o
> -obj-$(CONFIG_HAS_CPUFREQ) += pmstat.o
> +obj-$(CONFIG_PM_STATISTIC) += pmstat.o
>  
>  obj-$(CONFIG_X86) += hwregs.o
>  obj-$(CONFIG_X86) += reboot.o
> diff --git a/xen/drivers/acpi/pmstat.c b/xen/drivers/acpi/pmstat.c
> index df309e27b4..58cccd589b 100644
> --- a/xen/drivers/acpi/pmstat.c
> +++ b/xen/drivers/acpi/pmstat.c
> @@ -43,6 +43,167 @@
>  
>  DEFINE_PER_CPU_READ_MOSTLY(struct pm_px *, cpufreq_statistic_data);
>  
> +DEFINE_PER_CPU(spinlock_t, cpufreq_statistic_lock);
> +
> +/*********************************************************************
> + *                    Px STATISTIC INFO                              *
> + *********************************************************************/
> +
> +static void cpufreq_residency_update(unsigned int cpu, uint8_t state)
> +{
> +    uint64_t now, total_idle_ns;
> +    int64_t delta;
> +    struct pm_px *pxpt = per_cpu(cpufreq_statistic_data, cpu);
> +
> +    total_idle_ns = get_cpu_idle_time(cpu);
> +    now = NOW();
> +
> +    delta = (now - pxpt->prev_state_wall) -
> +            (total_idle_ns - pxpt->prev_idle_wall);
> +
> +    if ( likely(delta >= 0) )
> +        pxpt->u.pt[state].residency += delta;
> +
> +    pxpt->prev_state_wall = now;
> +    pxpt->prev_idle_wall = total_idle_ns;
> +}
> +
> +void cpufreq_statistic_update(unsigned int cpu, uint8_t from, uint8_t to)
> +{
> +    struct pm_px *pxpt;
> +    struct processor_pminfo *pmpt = processor_pminfo[cpu];
> +    spinlock_t *cpufreq_statistic_lock =
> +               &per_cpu(cpufreq_statistic_lock, cpu);
> +
> +    spin_lock(cpufreq_statistic_lock);
> +
> +    pxpt = per_cpu(cpufreq_statistic_data, cpu);
> +    if ( !pxpt || !pmpt ) {
> +        spin_unlock(cpufreq_statistic_lock);
> +        return;
> +    }
> +
> +    pxpt->u.last = from;
> +    pxpt->u.cur = to;
> +    pxpt->u.pt[to].count++;
> +
> +    cpufreq_residency_update(cpu, from);
> +
> +    (*(pxpt->u.trans_pt + from * pmpt->perf.state_count + to))++;
> +
> +    spin_unlock(cpufreq_statistic_lock);
> +}
> +
> +int cpufreq_statistic_init(unsigned int cpu)
> +{
> +    uint32_t i, count;
> +    struct pm_px *pxpt;
> +    const struct processor_pminfo *pmpt = processor_pminfo[cpu];
> +    spinlock_t *cpufreq_statistic_lock = &per_cpu(cpufreq_statistic_lock, 
> cpu);
> +
> +    spin_lock_init(cpufreq_statistic_lock);
> +
> +    if ( !pmpt )
> +        return -EINVAL;
> +
> +    spin_lock(cpufreq_statistic_lock);
> +
> +    pxpt = per_cpu(cpufreq_statistic_data, cpu);
> +    if ( pxpt ) {
> +        spin_unlock(cpufreq_statistic_lock);
> +        return 0;
> +    }
> +
> +    count = pmpt->perf.state_count;
> +
> +    pxpt = xzalloc(struct pm_px);
> +    if ( !pxpt ) {
> +        spin_unlock(cpufreq_statistic_lock);
> +        return -ENOMEM;
> +    }
> +    per_cpu(cpufreq_statistic_data, cpu) = pxpt;
> +
> +    pxpt->u.trans_pt = xzalloc_array(uint64_t, count * count);
> +    if (!pxpt->u.trans_pt) {
> +        xfree(pxpt);
> +        spin_unlock(cpufreq_statistic_lock);
> +        return -ENOMEM;
> +    }
> +
> +    pxpt->u.pt = xzalloc_array(struct pm_px_val, count);
> +    if (!pxpt->u.pt) {
> +        xfree(pxpt->u.trans_pt);
> +        xfree(pxpt);
> +        spin_unlock(cpufreq_statistic_lock);
> +        return -ENOMEM;
> +    }
> +
> +    pxpt->u.total = pmpt->perf.state_count;
> +    pxpt->u.usable = pmpt->perf.state_count - pmpt->perf.platform_limit;
> +
> +    for (i=0; i < pmpt->perf.state_count; i++)
> +        pxpt->u.pt[i].freq = pmpt->perf.states[i].core_frequency;
> +
> +    pxpt->prev_state_wall = NOW();
> +    pxpt->prev_idle_wall = get_cpu_idle_time(cpu);
> +
> +    spin_unlock(cpufreq_statistic_lock);
> +
> +    return 0;
> +}
> +
> +void cpufreq_statistic_exit(unsigned int cpu)
> +{
> +    struct pm_px *pxpt;
> +    spinlock_t *cpufreq_statistic_lock = &per_cpu(cpufreq_statistic_lock, 
> cpu);
> +
> +    spin_lock(cpufreq_statistic_lock);
> +
> +    pxpt = per_cpu(cpufreq_statistic_data, cpu);
> +    if (!pxpt) {
> +        spin_unlock(cpufreq_statistic_lock);
> +        return;
> +    }
> +
> +    xfree(pxpt->u.trans_pt);
> +    xfree(pxpt->u.pt);
> +    xfree(pxpt);
> +    per_cpu(cpufreq_statistic_data, cpu) = NULL;
> +
> +    spin_unlock(cpufreq_statistic_lock);
> +}
> +
> +static void cpufreq_statistic_reset(unsigned int cpu)
> +{
> +    uint32_t i, j, count;
> +    struct pm_px *pxpt;
> +    const struct processor_pminfo *pmpt = processor_pminfo[cpu];
> +    spinlock_t *cpufreq_statistic_lock = &per_cpu(cpufreq_statistic_lock, 
> cpu);
> +
> +    spin_lock(cpufreq_statistic_lock);
> +
> +    pxpt = per_cpu(cpufreq_statistic_data, cpu);
> +    if ( !pmpt || !pxpt || !pxpt->u.pt || !pxpt->u.trans_pt ) {
> +        spin_unlock(cpufreq_statistic_lock);
> +        return;
> +    }
> +
> +    count = pmpt->perf.state_count;
> +
> +    for (i=0; i < count; i++) {
> +        pxpt->u.pt[i].residency = 0;
> +        pxpt->u.pt[i].count = 0;
> +
> +        for (j=0; j < count; j++)
> +            *(pxpt->u.trans_pt + i*count + j) = 0;
> +    }
> +
> +    pxpt->prev_state_wall = NOW();
> +    pxpt->prev_idle_wall = get_cpu_idle_time(cpu);
> +
> +    spin_unlock(cpufreq_statistic_lock);
> +}

The want/need for this code movement wants mentioning in the description imo.
It may even be desirable to split this out. The more that while you move it,
it would be quite nice if various style corrections could be applied.

> @@ -522,34 +683,3 @@ int do_pm_op(struct xen_sysctl_pm_op *op)
>  
>      return ret;
>  }
> -
> -int acpi_set_pdc_bits(uint32_t acpi_id, XEN_GUEST_HANDLE(uint32) pdc)
> -{
> -    u32 bits[3];
> -    int ret;
> -
> -    if ( copy_from_guest(bits, pdc, 2) )
> -        ret = -EFAULT;
> -    else if ( bits[0] != ACPI_PDC_REVISION_ID || !bits[1] )
> -        ret = -EINVAL;
> -    else if ( copy_from_guest_offset(bits + 2, pdc, 2, 1) )
> -        ret = -EFAULT;
> -    else
> -    {
> -        u32 mask = 0;
> -
> -        if ( xen_processor_pmbits & XEN_PROCESSOR_PM_CX )
> -            mask |= ACPI_PDC_C_MASK | ACPI_PDC_SMP_C1PT;
> -        if ( xen_processor_pmbits & XEN_PROCESSOR_PM_PX )
> -            mask |= ACPI_PDC_P_MASK | ACPI_PDC_SMP_C1PT;
> -        if ( xen_processor_pmbits & XEN_PROCESSOR_PM_TX )
> -            mask |= ACPI_PDC_T_MASK | ACPI_PDC_SMP_C1PT;
> -        bits[2] &= (ACPI_PDC_C_MASK | ACPI_PDC_P_MASK | ACPI_PDC_T_MASK |
> -                    ACPI_PDC_SMP_C1PT) & ~mask;
> -        ret = arch_acpi_set_pdc_bits(acpi_id, bits, mask);
> -    }
> -    if ( !ret && __copy_to_guest_offset(pdc, 2, bits + 2, 1) )
> -        ret = -EFAULT;
> -
> -    return ret;
> -}
> --- a/xen/drivers/cpufreq/cpufreq.c
> +++ b/xen/drivers/cpufreq/cpufreq.c
> @@ -582,6 +582,37 @@ out:
>      return ret;
>  }
>  
> +int acpi_set_pdc_bits(uint32_t acpi_id, XEN_GUEST_HANDLE(uint32) pdc)
> +{
> +    u32 bits[3];
> +    int ret;
> +
> +    if ( copy_from_guest(bits, pdc, 2) )
> +        ret = -EFAULT;
> +    else if ( bits[0] != ACPI_PDC_REVISION_ID || !bits[1] )
> +        ret = -EINVAL;
> +    else if ( copy_from_guest_offset(bits + 2, pdc, 2, 1) )
> +        ret = -EFAULT;
> +    else
> +    {
> +        u32 mask = 0;
> +
> +        if ( xen_processor_pmbits & XEN_PROCESSOR_PM_CX )
> +            mask |= ACPI_PDC_C_MASK | ACPI_PDC_SMP_C1PT;
> +        if ( xen_processor_pmbits & XEN_PROCESSOR_PM_PX )
> +            mask |= ACPI_PDC_P_MASK | ACPI_PDC_SMP_C1PT;
> +        if ( xen_processor_pmbits & XEN_PROCESSOR_PM_TX )
> +            mask |= ACPI_PDC_T_MASK | ACPI_PDC_SMP_C1PT;
> +        bits[2] &= (ACPI_PDC_C_MASK | ACPI_PDC_P_MASK | ACPI_PDC_T_MASK |
> +                    ACPI_PDC_SMP_C1PT) & ~mask;
> +        ret = arch_acpi_set_pdc_bits(acpi_id, bits, mask);
> +    }
> +    if ( !ret && __copy_to_guest_offset(pdc, 2, bits + 2, 1) )
> +        ret = -EFAULT;
> +
> +    return ret;
> +}

Same here - looks pretty independent.

> --- a/xen/include/acpi/cpufreq/processor_perf.h
> +++ b/xen/include/acpi/cpufreq/processor_perf.h
> @@ -9,11 +9,19 @@
>  
>  unsigned int powernow_register_driver(void);
>  unsigned int get_measured_perf(unsigned int cpu, unsigned int flag);
> -void cpufreq_residency_update(unsigned int cpu, uint8_t state);
> +#ifdef CONFIG_PM_STATISTIC
>  void cpufreq_statistic_update(unsigned int cpu, uint8_t from, uint8_t to);
>  int  cpufreq_statistic_init(unsigned int cpu);
>  void cpufreq_statistic_exit(unsigned int cpu);
> -void cpufreq_statistic_reset(unsigned int cpu);
> +#else
> +static inline  void cpufreq_statistic_update(unsigned int cpu, uint8_t from,
> +                                             uint8_t to) {};

Nit: Stray semicolon. I'm also uncertain whether we like this kind of
formatting (i.e. if already things don't fit on a single line, I'm unsure
whether we like then having the braces not on their own line).

> +static inline int cpufreq_statistic_init(unsigned int cpu)
> +{
> +    return 0;
> +}
> +static inline void cpufreq_statistic_exit(unsigned int cpu) {};

Stray semicolon again.

> --- a/xen/include/xen/acpi.h
> +++ b/xen/include/xen/acpi.h
> @@ -158,6 +158,7 @@ int acpi_gsi_to_irq (u32 gsi, unsigned int *irq);
>  extern unsigned int max_cstate;
>  extern unsigned int max_csubstate;
>  
> +#ifdef CONFIG_PM_STATISTIC
>  static inline unsigned int acpi_get_cstate_limit(void)
>  {
>       return max_cstate;
> @@ -177,6 +178,7 @@ static inline void acpi_set_csubstate_limit(unsigned int 
> new_limit)
>  {
>       max_csubstate = new_limit;
>  }
> +#endif /* CONFIG_PM_STATISTIC */

Is this really necessary? Afaict these inline functions would still
compile fine; they'd merely end up without any user for now. (Not sure
what Misra's take is on unused inline functions.)

> --- a/xen/include/xen/pmstat.h
> +++ b/xen/include/xen/pmstat.h
> @@ -15,11 +15,13 @@ struct compat_processor_power;
>  long compat_set_cx_pminfo(uint32_t acpi_id, struct compat_processor_power 
> *power);
>  #endif
>  
> +#ifdef CONFIG_PM_STATISTIC
>  uint32_t pmstat_get_cx_nr(unsigned int cpu);
>  int pmstat_get_cx_stat(unsigned int cpu, struct pm_cx_stat *stat);
>  int pmstat_reset_cx_stat(unsigned int cpu);
>  
>  int do_get_pm_info(struct xen_sysctl_get_pmstat *op);
>  int do_pm_op(struct xen_sysctl_pm_op *op);
> +#endif /* CONFIG_PM_STATISTIC */

Similarly leaving these declarations visible isn't going to be a problem.
We do so quite frequently elsewhere.

Jan