Re: [PATCH v2 04/13] arm64/arch_timer: Provide noinstr sched_clock_read() functions

[Valentin Schneider <vschneid@redhat.com>]

On 19/05/23 12:21, Peter Zijlstra wrote:
> With the intent to provide local_clock_noinstr(), a variant of
> local_clock() that's safe to be called from noinstr code (with the
> assumption that any such code will already be non-preemptible),
> prepare for things by providing a noinstr sched_clock_read() function.
>
> Specifically, preempt_enable_*() calls out to schedule(), which upsets
> noinstr validation efforts.
>
> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
> ---
>  arch/arm64/include/asm/arch_timer.h  |    8 ----
>  drivers/clocksource/arm_arch_timer.c |   60 ++++++++++++++++++++++++++---------
>  2 files changed, 47 insertions(+), 21 deletions(-)
>
> --- a/arch/arm64/include/asm/arch_timer.h
> +++ b/arch/arm64/include/asm/arch_timer.h
> @@ -88,13 +88,7 @@ static inline notrace u64 arch_timer_rea
>
>  #define arch_timer_reg_read_stable(reg)					\
>       ({								\
> -		u64 _val;						\
> -									\
> -		preempt_disable_notrace();				\
> -		_val = erratum_handler(read_ ## reg)();			\
> -		preempt_enable_notrace();				\
> -									\
> -		_val;							\
> +		erratum_handler(read_ ## reg)();			\
>       })
>
>  /*
> --- a/drivers/clocksource/arm_arch_timer.c
> +++ b/drivers/clocksource/arm_arch_timer.c
> @@ -191,22 +191,40 @@ u32 arch_timer_reg_read(int access, enum
>       return val;
>  }
>
> -static notrace u64 arch_counter_get_cntpct_stable(void)
> +static noinstr u64 _arch_counter_get_cntpct_stable(void)
>  {
>       return __arch_counter_get_cntpct_stable();
>  }
>

I tripped over the underscores when reviewing this :( This must be
getting close to the symbol length limit, but could we give this a helpful
prefix or suffix? raw_* or *_noinstr?

> @@ -1074,6 +1092,13 @@ struct arch_timer_kvm_info *arch_timer_g
>
>  static void __init arch_counter_register(unsigned type)
>  {
> +	/*
> +	 * Default to cp15 based access because arm64 uses this function for
> +	 * sched_clock() before DT is probed and the cp15 method is guaranteed
> +	 * to exist on arm64. arm doesn't use this before DT is probed so even
> +	 * if we don't have the cp15 accessors we won't have a problem.
> +	 */
> +	u64 (*scr)(void) = arch_counter_get_cntvct;

So this bit sent me on a little spelunking session :-)

From a control flow perspective the initialization isn't required, but then
I looked into the comment and found it comes from the
arch_timer_read_counter() definition... Which itself doesn't get used by
sched_clock() until the sched_clock_register() below!

So AFAICT that comment was true as of

  220069945b29 ("clocksource: arch_timer: Add support for memory mapped timers")

but not after a commit that came 2 months later:

  65cd4f6c99c1 ("arch_timer: Move to generic sched_clock framework")

which IIUC made arm/arm64 follow the default approach of using the
jiffy-based sched_clock() before probing DT/ACPI and registering a "proper"
sched_clock.

All of that to say: the comment about arch_timer_read_counter() vs early
sched_clock() doesn't apply anymore, but I think we need to keep its
initalization around for stuff like get_cycles(). This initialization here
should be OK to put to the bin, though.