Re: [Qemu-arm] [PATCH v3 12/22] target/arm: Filter cycle counter based on PMCCFILTR_EL0

[Peter Maydell]

On 16 March 2018 at 20:31, Aaron Lindsay <address@hidden> wrote:
> The pmu_counter_filtered and pmu_op_start/finish functions are generic
> (as opposed to PMCCNTR-specific) to allow for the implementation of
> other events.
>
> Signed-off-by: Aaron Lindsay <address@hidden>
> ---
>  target/arm/cpu.c    |  3 ++
>  target/arm/cpu.h    | 37 +++++++++++++++++++----
>  target/arm/helper.c | 87 
> ++++++++++++++++++++++++++++++++++++++++++++++++++---
>  3 files changed, 116 insertions(+), 11 deletions(-)
>
> diff --git a/target/arm/cpu.c b/target/arm/cpu.c
> index a2cb21e..b0d032c 100644
> --- a/target/arm/cpu.c
> +++ b/target/arm/cpu.c
> @@ -887,6 +887,9 @@ static void arm_cpu_realizefn(DeviceState *dev, Error 
> **errp)
>      if (!cpu->has_pmu) {
>          unset_feature(env, ARM_FEATURE_PMU);
>          cpu->id_aa64dfr0 &= ~0xf00;
> +    } else {
> +        arm_register_pre_el_change_hook(cpu, &pmu_pre_el_change, 0);
> +        arm_register_el_change_hook(cpu, &pmu_post_el_change, 0);

You probably don't want to do this if we're using KVM, as there
are some code paths where we call do_interrupt() when using KVM
which will trigger the hooks and likely do unexpected things.
(For instance kvm_arm_handle_debug() does this to set the guest
up to take debug exceptions.)

>      }
>
>      if (!arm_feature(env, ARM_FEATURE_EL2)) {
> diff --git a/target/arm/cpu.h b/target/arm/cpu.h
> index b0ef727..9c3b5ef 100644
> --- a/target/arm/cpu.h
> +++ b/target/arm/cpu.h
> @@ -458,6 +458,11 @@ typedef struct CPUARMState {
>           * was reset. Otherwise it stores the counter value
>           */
>          uint64_t c15_ccnt;
> +        /* ccnt_cached_cycles is used to hold the last cycle count when
> +         * c15_ccnt holds the guest-visible count instead of the delta during
> +         * PMU operations which require this.
> +         */
> +        uint64_t ccnt_cached_cycles;

Can this ever hold valid state at a point when we need to do VM
migration, or is it purely temporary ?

>          uint64_t pmccfiltr_el0; /* Performance Monitor Filter Register */
>          uint64_t vpidr_el2; /* Virtualization Processor ID Register */
>          uint64_t vmpidr_el2; /* Virtualization Multiprocessor ID Register */
> @@ -896,15 +901,35 @@ int cpu_arm_signal_handler(int host_signum, void *pinfo,
>                             void *puc);
>
>  /**
> - * pmccntr_sync
> + * pmccntr_op_start/finish

Shouldn't this doc change and prototype change have gone with the earlier
patch that changed the implementation?

>   * @env: CPUARMState
>   *
> - * Synchronises the counter in the PMCCNTR. This must always be called twice,
> - * once before any action that might affect the timer and again afterwards.
> - * The function is used to swap the state of the register if required.
> - * This only happens when not in user mode (!CONFIG_USER_ONLY)
> + * Convert the counter in the PMCCNTR between its delta form (the typical 
> mode
> + * when it's enabled) and the guest-visible value. These two calls must 
> always
> + * surround any action which might affect the counter, and the return value
> + * from pmccntr_op_start must be supplied as the second argument to
> + * pmccntr_op_finish.
> + */
> +uint64_t pmccntr_op_start(CPUARMState *env);
> +void pmccntr_op_finish(CPUARMState *env, uint64_t prev_cycles);
> +
> +/**
> + * pmu_op_start/finish
> + * @env: CPUARMState
> + *
> + * Convert all PMU counters between their delta form (the typical mode when
> + * they are enabled) and the guest-visible values. These two calls must
> + * surround any action which might affect the counters, and the return value
> + * from pmu_op_start must be supplied as the second argument to 
> pmu_op_finish.
> + */
> +uint64_t pmu_op_start(CPUARMState *env);
> +void pmu_op_finish(CPUARMState *env, uint64_t prev_cycles);
> +
> +/**
> + * Functions to register as EL change hooks for PMU mode filtering
>   */
> -void pmccntr_sync(CPUARMState *env);
> +void pmu_pre_el_change(ARMCPU *cpu, void *ignored);
> +void pmu_post_el_change(ARMCPU *cpu, void *ignored);
>
>  /* SCTLR bit meanings. Several bits have been reused in newer
>   * versions of the architecture; in that case we define constants
> diff --git a/target/arm/helper.c b/target/arm/helper.c
> index 0102357..95b09d6 100644
> --- a/target/arm/helper.c
> +++ b/target/arm/helper.c
> @@ -908,6 +908,15 @@ static const ARMCPRegInfo v6_cp_reginfo[] = {
>  #define PMCRC   0x4
>  #define PMCRE   0x1
>
> +#define PMXEVTYPER_P          0x80000000
> +#define PMXEVTYPER_U          0x40000000
> +#define PMXEVTYPER_NSK        0x20000000
> +#define PMXEVTYPER_NSU        0x10000000
> +#define PMXEVTYPER_NSH        0x08000000
> +#define PMXEVTYPER_M          0x04000000
> +#define PMXEVTYPER_MT         0x02000000
> +#define PMXEVTYPER_EVTCOUNT   0x000003ff
> +
>  #define PMU_NUM_COUNTERS(env) ((env->cp15.c9_pmcr & PMCRN_MASK) >> 
> PMCRN_SHIFT)
>  /* Bits allowed to be set/cleared for PMCNTEN* and PMINTEN* */
>  #define PMU_COUNTER_MASK(env) ((1 << 31) | ((1 << PMU_NUM_COUNTERS(env)) - 
> 1))
> @@ -998,7 +1007,7 @@ static CPAccessResult pmreg_access_ccntr(CPUARMState 
> *env,
>
>  static inline bool arm_ccnt_enabled(CPUARMState *env)
>  {
> -    /* This does not support checking PMCCFILTR_EL0 register */
> +    /* Does not check PMCCFILTR_EL0, which is handled by 
> pmu_counter_filtered */
>
>      if (!(env->cp15.c9_pmcr & PMCRE) || !(env->cp15.c9_pmcnten & (1 << 31))) 
> {
>          return false;
> @@ -1006,6 +1015,44 @@ static inline bool arm_ccnt_enabled(CPUARMState *env)
>
>      return true;
>  }
> +
> +/* Returns true if the counter corresponding to the passed-in pmevtyper or
> + * pmccfiltr value is filtered using the current state */

"is filtered (ie does not count events)" would make it clearer.

Nit: */ should be on a line of its own.

> +static inline bool pmu_counter_filtered(CPUARMState *env, uint64_t 
> pmxevtyper)
> +{
> +    bool secure = arm_is_secure(env);
> +    int el = arm_current_el(env);
> +
> +    bool P   = pmxevtyper & PMXEVTYPER_P;
> +    bool U   = pmxevtyper & PMXEVTYPER_U;
> +    bool NSK = pmxevtyper & PMXEVTYPER_NSK;
> +    bool NSU = pmxevtyper & PMXEVTYPER_NSU;
> +    bool NSH = pmxevtyper & PMXEVTYPER_NSH;
> +    bool M   = pmxevtyper & PMXEVTYPER_M;

Lowercase for variable names, please.

> +
> +    if (el == 1 && P) {
> +        return true;
> +    } else if (el == 0 && U) {
> +        return true;
> +    }
> +
> +    if (arm_feature(env, ARM_FEATURE_EL3)) {
> +        if (el == 1 && !secure && NSK != P) {
> +            return true;
> +        } else if (el == 0 && !secure && NSU != U) {
> +            return true;
> +        } else if (el == 3 && secure && M != P) {
> +            return true;
> +        }
> +    }
> +
> +    if (arm_feature(env, ARM_FEATURE_EL2) && el == 2 && !secure && !NSH) {
> +        return true;
> +    }

This doesn't follow the structure that the Arm ARM pseudocode
uses, which makes it a bit tricky to review. (see AArch64.CountEvents()).
It also doesn't implement the same logic -- for instance this
code will return true if we're in NS EL1 and the U bit is set,
whereas the pseudocode returns true for NS EL1 only if U != NSU.

> +    return false;
> +}
> +
>  /*
>   * Ensure c15_ccnt is the guest-visible count so that operations such as
>   * enabling/disabling the counter or filtering, modifying the count itself,
> @@ -1023,7 +1070,8 @@ uint64_t pmccntr_op_start(CPUARMState *env)
>                            ARM_CPU_FREQ, NANOSECONDS_PER_SECOND);
>  #endif
>
> -    if (arm_ccnt_enabled(env)) {
> +    if (arm_ccnt_enabled(env) &&
> +          !pmu_counter_filtered(env, env->cp15.pmccfiltr_el0)) {
>
>          uint64_t eff_cycles = cycles;
>          if (env->cp15.c9_pmcr & PMCRD) {
> @@ -1043,7 +1091,8 @@ uint64_t pmccntr_op_start(CPUARMState *env)
>   */
>  void pmccntr_op_finish(CPUARMState *env, uint64_t prev_cycles)
>  {
> -    if (arm_ccnt_enabled(env)) {
> +    if (arm_ccnt_enabled(env) &&
> +          !pmu_counter_filtered(env, env->cp15.pmccfiltr_el0)) {
>          if (env->cp15.c9_pmcr & PMCRD) {
>              /* Increment once every 64 processor clock cycles */
> @@ -1054,10 +1103,30 @@ void pmccntr_op_finish(CPUARMState *env, uint64_t 
> prev_cycles)
>      }
>  }
>
> +uint64_t pmu_op_start(CPUARMState *env)
> +{
> +    return pmccntr_op_start(env);
> +}

Do these two functions end up being different at the end
of the patchset?

> +
> +void pmu_op_finish(CPUARMState *env, uint64_t prev_cycles)
> +{
> +    pmccntr_op_finish(env, prev_cycles);
> +}
> +
> +void pmu_pre_el_change(ARMCPU *cpu, void *ignored)
> +{
> +    cpu->env.cp15.ccnt_cached_cycles = pmu_op_start(&cpu->env);
> +}
> +
> +void pmu_post_el_change(ARMCPU *cpu, void *ignored)
> +{
> +    pmu_op_finish(&cpu->env, cpu->env.cp15.ccnt_cached_cycles);
> +}
> +
>  static void pmcr_write(CPUARMState *env, const ARMCPRegInfo *ri,
>                         uint64_t value)
>  {
> -    uint64_t saved_cycles = pmccntr_op_start(env);
> +    uint64_t saved_cycles = pmu_op_start(env);
>
>      if (value & PMCRC) {
>          /* The counter has been reset */
> @@ -1068,7 +1137,7 @@ static void pmcr_write(CPUARMState *env, const 
> ARMCPRegInfo *ri,
>      env->cp15.c9_pmcr &= ~0x39;
>      env->cp15.c9_pmcr |= (value & 0x39);
>
> -    pmccntr_op_finish(env, saved_cycles);
> +    pmu_op_finish(env, saved_cycles);
>  }
>
>  static uint64_t pmccntr_read(CPUARMState *env, const ARMCPRegInfo *ri)
> @@ -1117,6 +1186,14 @@ void pmccntr_op_finish(CPUARMState *env, uint64_t 
> prev_cycles)
>  {
>  }
>
> +uint64_t pmu_op_start(CPUARMState *env)
> +{
> +}
> +
> +void pmu_op_finish(CPUARMState *env, uint64_t prev_cycles)
> +{
> +}
> +
>  #endif
>
>  static void pmccfiltr_write(CPUARMState *env, const ARMCPRegInfo *ri,
> --

thanks
-- PMM