Re: [PATCH v5] x86/tsc: Add option to force frequency recalibration with HW timer

[Feng Tang <feng.tang@intel.com>]

On Wed, Jan 04, 2023 at 09:46:34AM -0800, Paul E. McKenney wrote:
> On Wed, Jan 04, 2023 at 06:32:27AM -0800, Paul E. McKenney wrote:
> > On Wed, Jan 04, 2023 at 04:19:38PM +0800, Feng Tang wrote:
> > > The kernel assumes that the TSC frequency which is provided by the
> > > hardware / firmware via MSRs or CPUID(0x15) is correct after applying
> > > a few basic consistency checks. This disables the TSC recalibration
> > > against HPET or PM timer.
> > > 
> > > As a result there is no mechanism to validate that frequency in cases
> > > where a firmware or hardware defect is suspected. And there was case
> > > that some user used atomic clock to measure the TSC frequency and
> > > reported an inaccuracy issue, which was later fixed in firmware.
> > > 
> > > Add an option 'recalibrate' for 'tsc' kernel parameter to force the
> > > tsc freq recalibration with HPET or PM timer, and warn if the
> > > deviation from previous value is more than about 500 PPM, which
> > > provides a way to verify the data from hardware / firmware.
> > > 
> > > There is no functional change to existing work flow.
> > > 
> > > Recently there was a real-world case: "The 40ms/s divergence between
> > > TSC and HPET was observed on hardware that is quite recent" [1], on
> > > that platform the TSC frequence 1896 MHz was got from CPUID(0x15),
> > > and the force-reclibration with HPET/PMTIMER both calibrated out
> > > value of 1975 MHz, which also matched with check from software
> > > 'chronyd', indicating it's a problem of BIOS or firmware.
> > > 
> > > [Thanks tglx for helping improving the commit log]
> > > 
> > > [1]. https://lore.kernel.org/lkml/20221117230910.GI4001@paulmck-ThinkPad-P17-Gen-1/
> > > Signed-off-by: Feng Tang <feng.tang@intel.com>
> > 
> > Nice!!!
> > 
> > Tested-by: Paul E. McKenney <paulmck@kernel.org>
> 
> And I have queued this on -rcu for further review and testing, in
> particular, to get it into -next sooner rather than later.  Hope that
> is OK!
> 
> I was thinking that this recalibrate patch made mine unnecessary:
> 
> b32498162f5c ("clocksource: Verify HPET and PMTMR when TSC unverified")
> 
> But upon further thought, I remembered that what we here at Meta need is
> for TSC to remain in use on systems for which it is deemed trustworthy.
> The reason is that even a short switch to HPET can terminally annoy some
> of our systems.  So I must therefore keep b32498162f5c.

Yes, my patch only adds an optional kernel cmdline parameter, and it
can't use both HPET and PM_TIMER to do the calibration in one turn.
Also, the purpose of the 2 patches are completely different.

Thanks,
Feng

> 
> 							Thanx, Paul
> 
> > > ---
> > > Changelog:
> > > 
> > >   since v4:
> > >     * add the real world case, where the patch helped to root
> > >       caused a BIOS/FW problem of inaccurate CPUID-0x15 info
> > >     * rebase against v6.2-rc1
> > > 
> > >   since v3:
> > >     * add some real world case into commit log
> > >     * rebase against v6.0-rc1
> > > 
> > >   since v2:
> > >     * revise the option description in kernel-parameters.txt
> > >     * rebase against v5.19-rc2
> > > 
> > >   since v1:
> > >     * refine commit log to state clearly the problem and intention
> > >       of the patch by copying Thomas' words.
> > > 
> > > .../admin-guide/kernel-parameters.txt         |  4 +++
> > >  arch/x86/kernel/tsc.c                         | 34 ++++++++++++++++---
> > >  2 files changed, 34 insertions(+), 4 deletions(-)
> > > 
> > > diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt
> > > index 6cfa6e3996cf..d9eb98e748d5 100644
> > > --- a/Documentation/admin-guide/kernel-parameters.txt
> > > +++ b/Documentation/admin-guide/kernel-parameters.txt
> > > @@ -6369,6 +6369,10 @@
> > >  			in situations with strict latency requirements (where
> > >  			interruptions from clocksource watchdog are not
> > >  			acceptable).
> > > +			[x86] recalibrate: force to do frequency recalibration
> > > +			with a HW timer (HPET or PM timer) for systems whose
> > > +			TSC frequency comes from HW or FW through MSR or CPUID(0x15),
> > > +			and warn if the difference is more than 500 ppm.
> > >  
> > >  	tsc_early_khz=  [X86] Skip early TSC calibration and use the given
> > >  			value instead. Useful when the early TSC frequency discovery
> > > diff --git a/arch/x86/kernel/tsc.c b/arch/x86/kernel/tsc.c
> > > index a78e73da4a74..92bbc4a6b3fc 100644
> > > --- a/arch/x86/kernel/tsc.c
> > > +++ b/arch/x86/kernel/tsc.c
> > > @@ -48,6 +48,8 @@ static DEFINE_STATIC_KEY_FALSE(__use_tsc);
> > >  
> > >  int tsc_clocksource_reliable;
> > >  
> > > +static int __read_mostly tsc_force_recalibrate;
> > > +
> > >  static u32 art_to_tsc_numerator;
> > >  static u32 art_to_tsc_denominator;
> > >  static u64 art_to_tsc_offset;
> > > @@ -303,6 +305,8 @@ static int __init tsc_setup(char *str)
> > >  		mark_tsc_unstable("boot parameter");
> > >  	if (!strcmp(str, "nowatchdog"))
> > >  		no_tsc_watchdog = 1;
> > > +	if (!strcmp(str, "recalibrate"))
> > > +		tsc_force_recalibrate = 1;
> > >  	return 1;
> > >  }
> > >  
> > > @@ -1374,6 +1378,25 @@ static void tsc_refine_calibration_work(struct work_struct *work)
> > >  	else
> > >  		freq = calc_pmtimer_ref(delta, ref_start, ref_stop);
> > >  
> > > +	/* Will hit this only if tsc_force_recalibrate has been set */
> > > +	if (boot_cpu_has(X86_FEATURE_TSC_KNOWN_FREQ)) {
> > > +
> > > +		/* Warn if the deviation exceeds 500 ppm */
> > > +		if (abs(tsc_khz - freq) > (tsc_khz >> 11)) {
> > > +			pr_warn("Warning: TSC freq calibrated by CPUID/MSR differs from what is calibrated by HW timer, please check with vendor!!\n");
> > > +			pr_info("Previous calibrated TSC freq:\t %lu.%03lu MHz\n",
> > > +				(unsigned long)tsc_khz / 1000,
> > > +				(unsigned long)tsc_khz % 1000);
> > > +		}
> > > +
> > > +		pr_info("TSC freq recalibrated by [%s]:\t %lu.%03lu MHz\n",
> > > +			hpet ? "HPET" : "PM_TIMER",
> > > +			(unsigned long)freq / 1000,
> > > +			(unsigned long)freq % 1000);
> > > +
> > > +		return;
> > > +	}
> > > +
> > >  	/* Make sure we're within 1% */
> > >  	if (abs(tsc_khz - freq) > tsc_khz/100)
> > >  		goto out;
> > > @@ -1407,8 +1430,10 @@ static int __init init_tsc_clocksource(void)
> > >  	if (!boot_cpu_has(X86_FEATURE_TSC) || !tsc_khz)
> > >  		return 0;
> > >  
> > > -	if (tsc_unstable)
> > > -		goto unreg;
> > > +	if (tsc_unstable) {
> > > +		clocksource_unregister(&clocksource_tsc_early);
> > > +		return 0;
> > > +	}
> > >  
> > >  	if (boot_cpu_has(X86_FEATURE_NONSTOP_TSC_S3))
> > >  		clocksource_tsc.flags |= CLOCK_SOURCE_SUSPEND_NONSTOP;
> > > @@ -1421,9 +1446,10 @@ static int __init init_tsc_clocksource(void)
> > >  		if (boot_cpu_has(X86_FEATURE_ART))
> > >  			art_related_clocksource = &clocksource_tsc;
> > >  		clocksource_register_khz(&clocksource_tsc, tsc_khz);
> > > -unreg:
> > >  		clocksource_unregister(&clocksource_tsc_early);
> > > -		return 0;
> > > +
> > > +		if (!tsc_force_recalibrate)
> > > +			return 0;
> > >  	}
> > >  
> > >  	schedule_delayed_work(&tsc_irqwork, 0);
> > > -- 
> > > 2.34.1
> > >