[dpdk-dev] [EXT] Re: [PATCH v2 2/2] eal: roundup tsc frequency when estimating it

[Pavan Nikhilesh Bhagavatula]

On Sat, 2019-03-16 at 17:18 +0000, Wiles, Keith wrote:
> > On Mar 16, 2019, at 10:06 AM, Pavan Nikhilesh Bhagavatula <
> > pbhagavatula at marvell.com> wrote:
> > 
> > On Sat, 2019-03-16 at 14:42 +0000, Wiles, Keith wrote:
> > > > On Mar 16, 2019, at 2:03 AM, Pavan Nikhilesh Bhagavatula <
> > > > pbhagavatula at marvell.com> wrote:
> > > > 
> > > > From: Pavan Nikhilesh <pbhagavatula at marvell.com>
> > > > 
> > > > When estimating tsc frequency using sleep/gettime round it up
> > > > to
> > > > the
> > > > nearest multiple of 10Mhz for more accuracy.
> 
> How does rounding up the TSC value become more accurate, If the value
> is 1 cycles more then it should be then your macro would round down
> and if it is 1 cycle greater than 1E7 it would round up.

Example in case of RTE_ARM_EAL_RDTSC_USE_PMU enabled 

Before roundup : 1400000979
After roundup : 1400000000
EAL: TSC frequency is ~1400000000 Hz


Before roundup : 1399999060
After roundup : 1400000000
EAL: TSC frequency is ~1400000000 Hz

> > > > Signed-off-by: Pavan Nikhilesh <pbhagavatula at marvell.com>
> > > > ---
> > > > Useful in case of ARM64 if we enable RTE_ARM_EAL_RDTSC_USE_PMU,
> > > > get_tsc_freq_arch() will return 0 as there is no instruction to
> > > > determine
> > > > the clk of PMU and eal falls back to sleep(1).
> > > > 
> > > > lib/librte_eal/common/eal_common_timer.c | 4 ++--
> > > > lib/librte_eal/linuxapp/eal/eal_timer.c  | 2 +-
> > > > 2 files changed, 3 insertions(+), 3 deletions(-)
> 
> It appears you did not use the head of the master as linuxapp is now
> just linux and freebsdapp is freebsd. You need to rebase to the head
> of master and send a new version.
> > > > diff --git a/lib/librte_eal/common/eal_common_timer.c
> > > > b/lib/librte_eal/common/eal_common_timer.c
> > > > index dcf26bfea..1358bbed0 100644
> > > > --- a/lib/librte_eal/common/eal_common_timer.c
> > > > +++ b/lib/librte_eal/common/eal_common_timer.c
> > > > @@ -69,7 +69,7 @@ estimate_tsc_freq(void)
> > > > 	/* assume that the sleep(1) will sleep for 1 second */
> > > > 	uint64_t start = rte_rdtsc();
> > > > 	sleep(1);
> > > > -	return rte_rdtsc() - start;
> > > > +	return RTE_ALIGN_MUL_NEAR(rte_rdtsc() - start, 1E7);
> 
> The 1E7 is a magic number convert this to a meaningful define.

1E7 ~ 10Mhz will convert to a macro.

> > > > }
> > > > 
> > > > void
> > > > @@ -83,7 +83,7 @@ set_tsc_freq(void)
> > > > 	if (!freq)
> > > > 		freq = estimate_tsc_freq();
> > > > 
> > > > -	RTE_LOG(DEBUG, EAL, "TSC frequency is ~%" PRIu64 "
> > > > KHz\n", freq
> > > > / 1000);
> > > > +	RTE_LOG(INFO, EAL, "TSC frequency is ~%" PRIu64 "
> > > > Hz\n", freq);
> > > > 	eal_tsc_resolution_hz = freq;
> > 
> > I missed this log will remove it in the next version.
> > 
> > > > }
> > > > 
> > > > diff --git a/lib/librte_eal/linuxapp/eal/eal_timer.c
> > > > b/lib/librte_eal/linuxapp/eal/eal_timer.c
> > > > index bc8f05199..864d6ef29 100644
> > > > --- a/lib/librte_eal/linuxapp/eal/eal_timer.c
> > > > +++ b/lib/librte_eal/linuxapp/eal/eal_timer.c
> > > > @@ -248,7 +248,7 @@ get_tsc_freq(void)
> > > > 
> > > > 		double secs = (double)ns/NS_PER_SEC;
> > > > 		tsc_hz = (uint64_t)((end - start)/secs);
> > > > -		return tsc_hz;
> > > > +		return RTE_ALIGN_MUL_NEAR(tsc_hz, 1E7);
> > > 
> > > Maybe I missed an email about this, but why would I want the TSC
> > > hz
> > > rounded here? I do not mind the macro just the fact that we are
> > > changing TSC hz value. If the TSC value is wrong then we need to
> > > fix
> > > the value, but I do not see it being wrong here.
> > 
> > Since in this function nanosleep might not be cycle accurate we
> > need to
> > round it up.
> > 
> > Please note that estimation only applies when  get_tsc_freq_arch()
> > fails. i.e there is no CPU instruction that specifies the cyc/sec.
> > 
> > As I mentioned in the patch notes
> > "Useful in case of ARM64 if we enable RTE_ARM_EAL_RDTSC_USE_PMU,
> > get_tsc_freq_arch() will return 0 as there is no instruction to
> > determine the clock of PMU and eal falls back to
> > sleep(1)/nanosleep.” 
> 
> OK, I looked at the changes and the code for setting the TSC again. I
> would have not handled the setting of TSC in the way it was done, but
> that is not your problem. I agree the changes do look ok, the only
> problem I have is the new macro will roundup or down depending on the
> value. In your statement you are wanting to roundup the values.
> 
> If the sleep/nanosleep is less than a second for some reason, then
> your macro would round it down is that what we wanted? I guess my
> point is you are assuming the TSC calculation will always be less
> than a second (with sleep) and the macro would round up depending on
> the value calculated using the sleep/nanosleep.
> 
> I was playing with these MUL macros and I am not sure they do what we
> expect in the case of the multiple value is much closer to the value
> passed.
> 
> If we have a v = 10001 and multiple to 1000 we have the following:
> 
> RTE_ALIGN_MUL_CEIL(10001, 1000)
> 	(10001 + (1000 - 1)) / (1000 * 1000)
((10001 + (1000 - 1)) / 1000) * 1000
> 	(10001 + 999)        / 1000000
> 	20000                / 1000000
> Result: 0

((10001 + (1000 - 1) / 1000) * 1000
((10001 + 999) / 1000) * 1000
(11000/1000) * 1000
11 * 1000 

Result : 11000

> 
> RTE_ALIGN_MUL_FLOOR(10001, 1000)
> 	(10001 / (1000 * 1000))
(10001 / 1000) * 1000
> 	(10001 / 1000000)
> Result: 0
10.001 * 1000

Result : 1000

> 
> Unless I am wrong here the value v must be over a 1,000,000 to make
> these macros work or the value v to be greater than (mul * mul) in
> all cases or zero is the result. It may work for the TSC values as we
> are using a small mul value compared to the TSC value. If DPDK was
> ported to a slower machine it could be also zero.

Unless we have machines that run at freq < 10Mhz this scheme will
always work.
If we have such machines lets hope that they have a CPU instruction
that tells us the cyc/sec.

> 
> I think we need to fix the macros and rethink how TSC is set here.
> 
> > > > 	}
> > > > #endif
> > > > 	return 0;
> > > > --
> > > > 2.21.0
> > > > 
> > > 
> > > Regards,
> > > Keith
> 
> Regards,
> Keith
> 

Regards,
Pavan.