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

[Thomas Monjalon <thomas@monjalon.net>]

12/10/2017 10:48, Bruce Richardson:
> On Wed, Oct 11, 2017 at 09:25:58PM +0200, Thomas Monjalon wrote:
> > 11/10/2017 20:57, Jerin Jacob:
> > > From: Thomas Monjalon <thomas@monjalon.net>
> > > > 22/09/2017 10:25, Gowrishankar:
> > > > > From: Jerin Jacob <jerin.jacob@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@caviumnetworks.com>
> > > > 
> > > > I agree on the idea.
> > > 
> > > OK
> > > 
> > > > The namespace of cycles related function in DPDK is a real mess.
> > > 
> > > Absolutely!!
> > > 
> > > > 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.
> > > 
> > > I thought of avoid exporting it. But then if the function is in
> > > lib/librte_eal/common/include/arch/../rte_cycles.h it is anyway exposed to
> > > application. i.e whatever files in lib/librte_eal/common/include/arch/../
> > > anyway exposed to application.
> > 
> > Ah yes, you are right!
> > 
> > > See last comment.
> > > 
> > > > 
> > > > > +	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
> > > > 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.
> > > 
> > > I guess, we may need to see to how abstract vmware TSC support also in
> > > proper way
> > 
> > Yes
> > 
> > > > 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.
> > > 
> > > If I understand it correctly, You would like to create a header file 
> > > in lib/librte_eal/common/include/arch/../ which should not be exported and change
> > > the name to tsc_arch_hz.
> > 
> > I had not think about the way to do this.
> > What about having internal headers in lib/librte_eal/common/arch/ ?
> > 
> 
> Yes, this area needs cleanup, but I also think that this patchset (and
> follow-on patch for x86-specific implementation) does not make things
> significantly worse than they are now, while also giving significant
> benefits for users both in terms of improved clock accuracy and reduced
> startup time (due to not needing sleep). Therefore I'd like to see this
> merged into 17.11 as a definite improvement, even if it does not "fix" the
> bigger issues of poor naming etc. I think this is important enough an
> improvement that we need to see it in the LTS release.

I agree Bruce.
My initial point was to avoid exporting a new function with a wrong name.
It would be better to find a way to avoid this export.
If it cannot be done in 17.11 timeframe, we can at least add the
experimental tag.