Re: [PATCH v2 2/2] x86/delay: Introduce TPAUSE delay

[Andy Lutomirski]

On Thu, Mar 19, 2020 at 9:13 PM Kyung Min Park <kyung.min.park@xxxxxxxxx> wrote:
>
> TPAUSE instructs the processor to enter an implementation-dependent
> optimized state. The instruction execution wakes up when the time-stamp
> counter reaches or exceeds the implicit EDX:EAX 64-bit input value.
> The instruction execution also wakes up due to the expiration of
> the operating system time-limit or by an external interrupt
> or exceptions such as a debug exception or a machine check exception.
>
> TPAUSE offers a choice of two lower power states:
>  1. Light-weight power/performance optimized state C0.1
>  2. Improved power/performance optimized state C0.2
> This way, it can save power with low wake-up latency in comparison to
> spinloop based delay. The selection between the two is governed by the
> input register.
>
> TPAUSE is available on processors with X86_FEATURE_WAITPKG.
>
> Reviewed-by: Tony Luck <tony.luck@xxxxxxxxx>
> Co-developed-by: Fenghua Yu <fenghua.yu@xxxxxxxxx>
> Signed-off-by: Fenghua Yu <fenghua.yu@xxxxxxxxx>
> Signed-off-by: Kyung Min Park <kyung.min.park@xxxxxxxxx>
> ---
>  arch/x86/include/asm/mwait.h | 17 +++++++++++++++++
>  arch/x86/lib/delay.c         | 27 ++++++++++++++++++++++++++-
>  2 files changed, 43 insertions(+), 1 deletion(-)
>
> diff --git a/arch/x86/include/asm/mwait.h b/arch/x86/include/asm/mwait.h
> index aaf6643..fd59db0 100644
> --- a/arch/x86/include/asm/mwait.h
> +++ b/arch/x86/include/asm/mwait.h
> @@ -22,6 +22,8 @@
>  #define MWAITX_ECX_TIMER_ENABLE                BIT(1)
>  #define MWAITX_MAX_WAIT_CYCLES         UINT_MAX
>  #define MWAITX_DISABLE_CSTATES         0xf0
> +#define TPAUSE_C01_STATE               1
> +#define TPAUSE_C02_STATE               0
>
>  static inline void __monitor(const void *eax, unsigned long ecx,
>                              unsigned long edx)
> @@ -120,4 +122,19 @@ static inline void mwait_idle_with_hints(unsigned long eax, unsigned long ecx)
>         current_clr_polling();
>  }
>
> +/*
> + * Caller can specify whether to enter C0.1 (low latency, less
> + * power saving) or C0.2 state (saves more power, but longer wakeup
> + * latency). This may be overridden by the IA32_UMWAIT_CONTROL MSR
> + * which can force requests for C0.2 to be downgraded to C0.1.
> + */
> +static inline void __tpause(unsigned int ecx, unsigned int edx,
> +                           unsigned int eax)
> +{
> +       /* "tpause %ecx, %edx, %eax;" */
> +       asm volatile(".byte 0x66, 0x0f, 0xae, 0xf1\t\n"
> +                    :
> +                    : "c"(ecx), "d"(edx), "a"(eax));
> +}
> +
>  #endif /* _ASM_X86_MWAIT_H */
> diff --git a/arch/x86/lib/delay.c b/arch/x86/lib/delay.c
> index e6db855..5f11f0a 100644
> --- a/arch/x86/lib/delay.c
> +++ b/arch/x86/lib/delay.c
> @@ -97,6 +97,27 @@ static void delay_tsc(u64 cycles)
>  }
>
>  /*
> + * On Intel the TPAUSE instruction waits until any of:
> + * 1) the TSC counter exceeds the value provided in EAX:EDX
> + * 2) global timeout in IA32_UMWAIT_CONTROL is exceeded
> + * 3) an external interrupt occurs
> + */
> +static void delay_halt_tpause(u64 start, u64 cycles)
> +{
> +       u64 until = start + cycles;
> +       unsigned int eax, edx;
> +
> +       eax = (unsigned int)(until & 0xffffffff);
> +       edx = (unsigned int)(until >> 32);
> +
> +       /*
> +        * Hard code the deeper (C0.2) sleep state because exit latency is
> +        * small compared to the "microseconds" that usleep() will delay.
> +        */
> +       __tpause(TPAUSE_C02_STATE, edx, eax);
> +}
> +
> +/*
>   * On some AMD platforms, MWAITX has a configurable 32-bit timer, that
>   * counts with TSC frequency. The input value is the number of TSC cycles
>   * to wait. MWAITX will also exit when the timer expires.
> @@ -152,8 +173,12 @@ static void delay_halt(u64 __cycles)
>
>  void use_tsc_delay(void)
>  {
> -       if (delay_fn == delay_loop)
> +       if (static_cpu_has(X86_FEATURE_WAITPKG)) {
> +               delay_halt_fn = delay_halt_tpause;
> +               delay_fn = delay_halt;
> +       } else if (delay_fn == delay_loop) {
>                 delay_fn = delay_tsc;
> +       }
>  }

This is an odd way to dispatch: you're using static_cpu_has(), but
you're using it once to populate a function pointer.  Why not just put
the static_cpu_has() directly into delay_halt() and open-code the
three variants?  That will also make it a lot easier to understand the
oddity with start and cycles.

--Andy

> --
> 2.7.4
>