Re: [RFC PATCH v2 2/7] x86/tsc: Add clocksource ID, set system_counterval_t.cs_id

[Simon Horman]

On Fri, Dec 15, 2023 at 11:06:07PM +0100, Peter Hilber wrote:
> Add a clocksource ID for TSC and a distinct one for the early TSC.
> 
> Use distinct IDs for TSC and early TSC, since those also have distinct
> clocksource structs. This should help to keep existing semantics when
> comparing clocksources.
> 
> Also, set the recently added struct system_counterval_t member cs_id to the
> TSC ID in the cases where the clocksource member is being set to the TSC
> clocksource. In the future, this will keep get_device_system_crosststamp()
> working, when it will compare the clocksource id in struct
> system_counterval_t, rather than the clocksource.
> 
> For the x86 ART related code, system_counterval_t.cs == NULL corresponds to
> system_counterval_t.cs_id == CSID_GENERIC (0).
> 
> Signed-off-by: Peter Hilber <peter.hilber@xxxxxxxxxxxxxxx>

Hi Peter,

some minor feedback from my side that you may consider for
a future revision.

> diff --git a/arch/x86/kernel/tsc.c b/arch/x86/kernel/tsc.c

...

> @@ -1327,12 +1334,15 @@ EXPORT_SYMBOL(convert_art_to_tsc);
>   * that this flag is set before conversion to TSC is attempted.
>   *
>   * Return:
> - * struct system_counterval_t - system counter value with the pointer to the
> + * struct system_counterval_t - system counter value with the ID of the
>   *	corresponding clocksource
>   *	@cycles:	System counter value
>   *	@cs:		Clocksource corresponding to system counter value. Used
>   *			by timekeeping code to verify comparability of two cycle
>   *			values.
> + *	@cs_id:		Clocksource ID corresponding to system counter value.
> + *			Used by timekeeping code to verify comparability of two
> + *			cycle values.

None of the documented parameters to convert_art_ns_to_tsc() above
correspond to the parameters of convert_art_ns_to_tsc() below.

I would suggest a separate patch to address this.
And dropping this hunk from this patch.

The same patch that corrects the kernel doc for convert_art_ns_to_tsc()
could also correct the kernel doc for tsc_refine_calibration_work()
by documenting it's work parameter.

>   */
>  
>  struct system_counterval_t convert_art_ns_to_tsc(u64 art_ns)
> @@ -1347,8 +1357,11 @@ struct system_counterval_t convert_art_ns_to_tsc(u64 art_ns)
>  	do_div(tmp, USEC_PER_SEC);
>  	res += tmp;
>  
> -	return (struct system_counterval_t) { .cs = art_related_clocksource,
> -					      .cycles = res};
> +	return (struct system_counterval_t) {
> +		.cs = art_related_clocksource,
> +		.cs_id = have_art ? CSID_X86_TSC : CSID_GENERIC,
> +		.cycles = res
> +	};
>  }
>  EXPORT_SYMBOL(convert_art_ns_to_tsc);
>  
> @@ -1454,8 +1467,10 @@ static void tsc_refine_calibration_work(struct work_struct *work)
>  	if (tsc_unstable)
>  		goto unreg;
>  
> -	if (boot_cpu_has(X86_FEATURE_ART))
> +	if (boot_cpu_has(X86_FEATURE_ART)) {
>  		art_related_clocksource = &clocksource_tsc;
> +		have_art = true;
> +	}
>  	clocksource_register_khz(&clocksource_tsc, tsc_khz);
>  unreg:
>  	clocksource_unregister(&clocksource_tsc_early);

...