[dpdk-dev] [PATCH v2 5/5] eal/timer: honor architecture specific rdtsc hz function

Thomas Monjalon thomas at monjalon.net
Wed Oct 11 19:36:11 CEST 2017

22/09/2017 10:25, Gowrishankar:
> From: Jerin Jacob <jerin.jacob at caviumnetworks.com>
> When calibrating the tsc frequency, first, probe the architecture specific
> rdtsc hz function. if not available, use the existing calibrate scheme
> to calibrate the tsc frequency.
> Signed-off-by: Jerin Jacob <jerin.jacob at caviumnetworks.com>

I agree on the idea.

The namespace of cycles related function in DPDK is a real mess.
I think we can choose better names in this series as a first step
to tidy this mess.
I will explain below.

At first, we should avoid TSC and RDTSC which are Intel-only wording.
The generic word could be "cycles" (the word used in arch headers),
or "ticks".
We should also name the timer sources or their function in a generic way.
Examples: CPU cycles? fast counter? precise counter?

Sometimes we use "hz", sometimes "freq".
It would better to keep one of them.

> --- a/lib/librte_eal/common/eal_common_timer.c
> +++ b/lib/librte_eal/common/eal_common_timer.c
> @@ -80,8 +80,11 @@
>  void
>  set_tsc_freq(void)
>  {
> -	uint64_t freq = get_tsc_freq();
> +	uint64_t freq;
> +	freq = rte_rdtsc_arch_hz();

This new function is arch-specific and exported as a new API.

> +	if (!freq)
> +		freq = get_tsc_freq();

The function get_tsc_freq is guessing the freq with OS-specific method.

>  	if (!freq)
>  		freq = estimate_tsc_freq();

The function estimate_tsc_freq is doing an estimation based on sleep().

At the end, the most accurate frequency is saved in eal_tsc_resolution_hz
and can be retrieved with rte_get_tsc_hz().
I don't understand why rte_rdtsc_arch_hz() is also exported to the apps.

TSC and HPET timer sources are wrapped in rte_get_timer_hz() in the
generic code despite HPET is Intel specific.

Similarly we can get the current timer with rte_get_timer_cycles().
In the case of TSC, it calls rte_get_tsc_cycles() which is an alias
of rte_rdtsc().
Some code is still using directly rte_rdtsc().
There is also rte_rdtsc_precise which adds a memory barrier.

The real question is what is the right abstraction for the application?
Do we want the fastest timer? the CPU timer? a precise timer?

I would like to see a real discussion on this topic, in order of building
a new timer API which would alias the old one for some time.

If you don't want to bother with all these questions, I suggest to not
export the new function rte_rdtsc_arch_hz() and rename it to tsc_arch_hz.

More information about the dev mailing list