[v2,clocksource,6/7] clocksource: Verify HPET and PMTMR when TSC unverified

  On systems with two or fewer sockets, when the boot CPU has CONSTANT_TSC,
NONSTOP_TSC, and TSC_ADJUST, clocksource watchdog verification of the
TSC is disabled.  This works well much of the time, but there is the
occasional production-level system that meets all of these criteria, but
which still has a TSC that skews significantly from atomic-clock time.
This is usually attributed to a firmware or hardware fault.  Yes, the
various NTP daemons do express their opinions of userspace-to-atomic-clock
time skew, but they put them in various places, depending on the daemon
and distro in question.  It would therefore be good for the kernel to
have some clue that there is a problem.

The old behavior of marking the TSC unstable is a non-starter because a
great many workloads simply cannot tolerate the overheads and latencies
of the various non-TSC clocksources.  In addition, NTP-corrected systems
sometimes can tolerate significant kernel-space time skew as long as
the userspace time sources are within epsilon of atomic-clock time.

Therefore, when watchdog verification of TSC is disabled, enable it for
HPET and PMTMR (AKA ACPI PM timer).  This provides the needed in-kernel
time-skew diagnostic without degrading the system's performance.

Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Daniel Lezcano <daniel.lezcano@linaro.org>
Cc: Waiman Long <longman@redhat.com>
Cc: <x86@kernel.org>
Tested-by: Feng Tang <feng.tang@intel.com>
---
 arch/x86/include/asm/time.h   | 1 +
 arch/x86/kernel/hpet.c        | 2 ++
 arch/x86/kernel/tsc.c         | 5 +++++
 drivers/clocksource/acpi_pm.c | 6 ++++--
 4 files changed, 12 insertions(+), 2 deletions(-)
  

Hi Thomas,

are you ok with this patch ? Shall I pick it ?

Thanks

   -- Daniel


On 25/01/2023 01:27, Paul E. McKenney wrote:
> On systems with two or fewer sockets, when the boot CPU has CONSTANT_TSC,
> NONSTOP_TSC, and TSC_ADJUST, clocksource watchdog verification of the
> TSC is disabled.  This works well much of the time, but there is the
> occasional production-level system that meets all of these criteria, but
> which still has a TSC that skews significantly from atomic-clock time.
> This is usually attributed to a firmware or hardware fault.  Yes, the
> various NTP daemons do express their opinions of userspace-to-atomic-clock
> time skew, but they put them in various places, depending on the daemon
> and distro in question.  It would therefore be good for the kernel to
> have some clue that there is a problem.
> 
> The old behavior of marking the TSC unstable is a non-starter because a
> great many workloads simply cannot tolerate the overheads and latencies
> of the various non-TSC clocksources.  In addition, NTP-corrected systems
> sometimes can tolerate significant kernel-space time skew as long as
> the userspace time sources are within epsilon of atomic-clock time.
> 
> Therefore, when watchdog verification of TSC is disabled, enable it for
> HPET and PMTMR (AKA ACPI PM timer).  This provides the needed in-kernel
> time-skew diagnostic without degrading the system's performance.
> 
> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
> Cc: Thomas Gleixner <tglx@linutronix.de>
> Cc: Ingo Molnar <mingo@redhat.com>
> Cc: Borislav Petkov <bp@alien8.de>
> Cc: Dave Hansen <dave.hansen@linux.intel.com>
> Cc: "H. Peter Anvin" <hpa@zytor.com>
> Cc: Daniel Lezcano <daniel.lezcano@linaro.org>
> Cc: Waiman Long <longman@redhat.com>
> Cc: <x86@kernel.org>
> Tested-by: Feng Tang <feng.tang@intel.com>
> ---
>   arch/x86/include/asm/time.h   | 1 +
>   arch/x86/kernel/hpet.c        | 2 ++
>   arch/x86/kernel/tsc.c         | 5 +++++
>   drivers/clocksource/acpi_pm.c | 6 ++++--
>   4 files changed, 12 insertions(+), 2 deletions(-)
> 
> diff --git a/arch/x86/include/asm/time.h b/arch/x86/include/asm/time.h
> index 8ac563abb567b..a53961c64a567 100644
> --- a/arch/x86/include/asm/time.h
> +++ b/arch/x86/include/asm/time.h
> @@ -8,6 +8,7 @@
>   extern void hpet_time_init(void);
>   extern void time_init(void);
>   extern bool pit_timer_init(void);
> +extern bool tsc_clocksource_watchdog_disabled(void);
>   
>   extern struct clock_event_device *global_clock_event;
>   
> diff --git a/arch/x86/kernel/hpet.c b/arch/x86/kernel/hpet.c
> index 71f336425e58a..c8eb1ac5125ab 100644
> --- a/arch/x86/kernel/hpet.c
> +++ b/arch/x86/kernel/hpet.c
> @@ -1091,6 +1091,8 @@ int __init hpet_enable(void)
>   	if (!hpet_counting())
>   		goto out_nohpet;
>   
> +	if (tsc_clocksource_watchdog_disabled())
> +		clocksource_hpet.flags |= CLOCK_SOURCE_MUST_VERIFY;
>   	clocksource_register_hz(&clocksource_hpet, (u32)hpet_freq);
>   
>   	if (id & HPET_ID_LEGSUP) {
> diff --git a/arch/x86/kernel/tsc.c b/arch/x86/kernel/tsc.c
> index a78e73da4a74b..af3782fb6200c 100644
> --- a/arch/x86/kernel/tsc.c
> +++ b/arch/x86/kernel/tsc.c
> @@ -1186,6 +1186,11 @@ static void __init tsc_disable_clocksource_watchdog(void)
>   	clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
>   }
>   
> +bool tsc_clocksource_watchdog_disabled(void)
> +{
> +	return !(clocksource_tsc.flags & CLOCK_SOURCE_MUST_VERIFY);
> +}
> +
>   static void __init check_system_tsc_reliable(void)
>   {
>   #if defined(CONFIG_MGEODEGX1) || defined(CONFIG_MGEODE_LX) || defined(CONFIG_X86_GENERIC)
> diff --git a/drivers/clocksource/acpi_pm.c b/drivers/clocksource/acpi_pm.c
> index 279ddff81ab49..82338773602ca 100644
> --- a/drivers/clocksource/acpi_pm.c
> +++ b/drivers/clocksource/acpi_pm.c
> @@ -23,6 +23,7 @@
>   #include <linux/pci.h>
>   #include <linux/delay.h>
>   #include <asm/io.h>
> +#include <asm/time.h>
>   
>   /*
>    * The I/O port the PMTMR resides at.
> @@ -210,8 +211,9 @@ static int __init init_acpi_pm_clocksource(void)
>   		return -ENODEV;
>   	}
>   
> -	return clocksource_register_hz(&clocksource_acpi_pm,
> -						PMTMR_TICKS_PER_SEC);
> +	if (tsc_clocksource_watchdog_disabled())
> +		clocksource_acpi_pm.flags |= CLOCK_SOURCE_MUST_VERIFY;
> +	return clocksource_register_hz(&clocksource_acpi_pm, PMTMR_TICKS_PER_SEC);
>   }
>   
>   /* We use fs_initcall because we want the PCI fixups to have run
  

On Thu, Jan 26, 2023 at 11:57:41AM +0100, Daniel Lezcano wrote:
> 
> Hi Thomas,
> 
> are you ok with this patch ? Shall I pick it ?

Seeing no response, I sent a pull request.

If it would be helpful for me to make the pilgrimmage to (say) Intel
Jones Farm, I can easily make that trip.

							Thanx, Paul

> Thanks
> 
>   -- Daniel
> 
> 
> On 25/01/2023 01:27, Paul E. McKenney wrote:
> > On systems with two or fewer sockets, when the boot CPU has CONSTANT_TSC,
> > NONSTOP_TSC, and TSC_ADJUST, clocksource watchdog verification of the
> > TSC is disabled.  This works well much of the time, but there is the
> > occasional production-level system that meets all of these criteria, but
> > which still has a TSC that skews significantly from atomic-clock time.
> > This is usually attributed to a firmware or hardware fault.  Yes, the
> > various NTP daemons do express their opinions of userspace-to-atomic-clock
> > time skew, but they put them in various places, depending on the daemon
> > and distro in question.  It would therefore be good for the kernel to
> > have some clue that there is a problem.
> > 
> > The old behavior of marking the TSC unstable is a non-starter because a
> > great many workloads simply cannot tolerate the overheads and latencies
> > of the various non-TSC clocksources.  In addition, NTP-corrected systems
> > sometimes can tolerate significant kernel-space time skew as long as
> > the userspace time sources are within epsilon of atomic-clock time.
> > 
> > Therefore, when watchdog verification of TSC is disabled, enable it for
> > HPET and PMTMR (AKA ACPI PM timer).  This provides the needed in-kernel
> > time-skew diagnostic without degrading the system's performance.
> > 
> > Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
> > Cc: Thomas Gleixner <tglx@linutronix.de>
> > Cc: Ingo Molnar <mingo@redhat.com>
> > Cc: Borislav Petkov <bp@alien8.de>
> > Cc: Dave Hansen <dave.hansen@linux.intel.com>
> > Cc: "H. Peter Anvin" <hpa@zytor.com>
> > Cc: Daniel Lezcano <daniel.lezcano@linaro.org>
> > Cc: Waiman Long <longman@redhat.com>
> > Cc: <x86@kernel.org>
> > Tested-by: Feng Tang <feng.tang@intel.com>
> > ---
> >   arch/x86/include/asm/time.h   | 1 +
> >   arch/x86/kernel/hpet.c        | 2 ++
> >   arch/x86/kernel/tsc.c         | 5 +++++
> >   drivers/clocksource/acpi_pm.c | 6 ++++--
> >   4 files changed, 12 insertions(+), 2 deletions(-)
> > 
> > diff --git a/arch/x86/include/asm/time.h b/arch/x86/include/asm/time.h
> > index 8ac563abb567b..a53961c64a567 100644
> > --- a/arch/x86/include/asm/time.h
> > +++ b/arch/x86/include/asm/time.h
> > @@ -8,6 +8,7 @@
> >   extern void hpet_time_init(void);
> >   extern void time_init(void);
> >   extern bool pit_timer_init(void);
> > +extern bool tsc_clocksource_watchdog_disabled(void);
> >   extern struct clock_event_device *global_clock_event;
> > diff --git a/arch/x86/kernel/hpet.c b/arch/x86/kernel/hpet.c
> > index 71f336425e58a..c8eb1ac5125ab 100644
> > --- a/arch/x86/kernel/hpet.c
> > +++ b/arch/x86/kernel/hpet.c
> > @@ -1091,6 +1091,8 @@ int __init hpet_enable(void)
> >   	if (!hpet_counting())
> >   		goto out_nohpet;
> > +	if (tsc_clocksource_watchdog_disabled())
> > +		clocksource_hpet.flags |= CLOCK_SOURCE_MUST_VERIFY;
> >   	clocksource_register_hz(&clocksource_hpet, (u32)hpet_freq);
> >   	if (id & HPET_ID_LEGSUP) {
> > diff --git a/arch/x86/kernel/tsc.c b/arch/x86/kernel/tsc.c
> > index a78e73da4a74b..af3782fb6200c 100644
> > --- a/arch/x86/kernel/tsc.c
> > +++ b/arch/x86/kernel/tsc.c
> > @@ -1186,6 +1186,11 @@ static void __init tsc_disable_clocksource_watchdog(void)
> >   	clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
> >   }
> > +bool tsc_clocksource_watchdog_disabled(void)
> > +{
> > +	return !(clocksource_tsc.flags & CLOCK_SOURCE_MUST_VERIFY);
> > +}
> > +
> >   static void __init check_system_tsc_reliable(void)
> >   {
> >   #if defined(CONFIG_MGEODEGX1) || defined(CONFIG_MGEODE_LX) || defined(CONFIG_X86_GENERIC)
> > diff --git a/drivers/clocksource/acpi_pm.c b/drivers/clocksource/acpi_pm.c
> > index 279ddff81ab49..82338773602ca 100644
> > --- a/drivers/clocksource/acpi_pm.c
> > +++ b/drivers/clocksource/acpi_pm.c
> > @@ -23,6 +23,7 @@
> >   #include <linux/pci.h>
> >   #include <linux/delay.h>
> >   #include <asm/io.h>
> > +#include <asm/time.h>
> >   /*
> >    * The I/O port the PMTMR resides at.
> > @@ -210,8 +211,9 @@ static int __init init_acpi_pm_clocksource(void)
> >   		return -ENODEV;
> >   	}
> > -	return clocksource_register_hz(&clocksource_acpi_pm,
> > -						PMTMR_TICKS_PER_SEC);
> > +	if (tsc_clocksource_watchdog_disabled())
> > +		clocksource_acpi_pm.flags |= CLOCK_SOURCE_MUST_VERIFY;
> > +	return clocksource_register_hz(&clocksource_acpi_pm, PMTMR_TICKS_PER_SEC);
> >   }
> >   /* We use fs_initcall because we want the PCI fixups to have run
> 
> -- 
> <http://www.linaro.org/> Linaro.org │ Open source software for ARM SoCs
> 
> Follow Linaro:  <http://www.facebook.com/pages/Linaro> Facebook |
> <http://twitter.com/#!/linaroorg> Twitter |
> <http://www.linaro.org/linaro-blog/> Blog
>
  

Paul!

On Tue, Jan 24 2023 at 16:27, Paul E. McKenney wrote:
> On systems with two or fewer sockets, when the boot CPU has CONSTANT_TSC,
> NONSTOP_TSC, and TSC_ADJUST, clocksource watchdog verification of the
> TSC is disabled.  This works well much of the time, but there is the
> occasional production-level system that meets all of these criteria, but
> which still has a TSC that skews significantly from atomic-clock time.
> This is usually attributed to a firmware or hardware fault.  Yes, the
> various NTP daemons do express their opinions of userspace-to-atomic-clock
> time skew, but they put them in various places, depending on the daemon
> and distro in question.  It would therefore be good for the kernel to
> have some clue that there is a problem.
>
> The old behavior of marking the TSC unstable is a non-starter because a
> great many workloads simply cannot tolerate the overheads and latencies
> of the various non-TSC clocksources.  In addition, NTP-corrected systems
> sometimes can tolerate significant kernel-space time skew as long as
> the userspace time sources are within epsilon of atomic-clock time.
>
> Therefore, when watchdog verification of TSC is disabled, enable it for
> HPET and PMTMR (AKA ACPI PM timer).  This provides the needed in-kernel
> time-skew diagnostic without degrading the system's performance.

I'm more than unhappy about this. We finally have a point where the TSC
watchdog overhead can go away without adding TSC=reliable to the kernel
commandline.

Now you add an unconditionally enforce the watchdog again in a way which
even cannot be disabled on the kernel command line.

Patently bad idea, no cookies for you!

Thanks,

        tglx
  

Hi Thomas,

On Wed, Feb 01, 2023 at 11:24:14AM +0100, Thomas Gleixner wrote:
> Paul!
> 
> On Tue, Jan 24 2023 at 16:27, Paul E. McKenney wrote:
> > On systems with two or fewer sockets, when the boot CPU has CONSTANT_TSC,
> > NONSTOP_TSC, and TSC_ADJUST, clocksource watchdog verification of the
> > TSC is disabled.  This works well much of the time, but there is the
> > occasional production-level system that meets all of these criteria, but
> > which still has a TSC that skews significantly from atomic-clock time.
> > This is usually attributed to a firmware or hardware fault.  Yes, the
> > various NTP daemons do express their opinions of userspace-to-atomic-clock
> > time skew, but they put them in various places, depending on the daemon
> > and distro in question.  It would therefore be good for the kernel to
> > have some clue that there is a problem.
> >
> > The old behavior of marking the TSC unstable is a non-starter because a
> > great many workloads simply cannot tolerate the overheads and latencies
> > of the various non-TSC clocksources.  In addition, NTP-corrected systems
> > sometimes can tolerate significant kernel-space time skew as long as
> > the userspace time sources are within epsilon of atomic-clock time.
> >
> > Therefore, when watchdog verification of TSC is disabled, enable it for
> > HPET and PMTMR (AKA ACPI PM timer).  This provides the needed in-kernel
> > time-skew diagnostic without degrading the system's performance.
> 
> I'm more than unhappy about this. We finally have a point where the TSC
> watchdog overhead can go away without adding TSC=reliable to the kernel
> commandline.
> 
> Now you add an unconditionally enforce the watchdog again in a way which
> even cannot be disabled on the kernel command line.

Yes, this is a valid concern. Waiman, Paul and I discussed this and
had some proposal to handle this side effect, like only watchdoging 
HPET/ACPI-PM timer for a short period of time in this case.
https://lore.kernel.org/lkml/20221227183819.GI4001@paulmck-ThinkPad-P17-Gen-1/
My bad that I didn't follow up as my proposed code looked ugly as
bringing more complexsities. Does the idea of setting a watchdog
time limit sound fine to you?  

Thanks,
Feng

> Patently bad idea, no cookies for you!
> 
> Thanks,
> 
>         tglx
  

On 2/1/23 05:24, Thomas Gleixner wrote:
> Paul!
>
> On Tue, Jan 24 2023 at 16:27, Paul E. McKenney wrote:
>> On systems with two or fewer sockets, when the boot CPU has CONSTANT_TSC,
>> NONSTOP_TSC, and TSC_ADJUST, clocksource watchdog verification of the
>> TSC is disabled.  This works well much of the time, but there is the
>> occasional production-level system that meets all of these criteria, but
>> which still has a TSC that skews significantly from atomic-clock time.
>> This is usually attributed to a firmware or hardware fault.  Yes, the
>> various NTP daemons do express their opinions of userspace-to-atomic-clock
>> time skew, but they put them in various places, depending on the daemon
>> and distro in question.  It would therefore be good for the kernel to
>> have some clue that there is a problem.
>>
>> The old behavior of marking the TSC unstable is a non-starter because a
>> great many workloads simply cannot tolerate the overheads and latencies
>> of the various non-TSC clocksources.  In addition, NTP-corrected systems
>> sometimes can tolerate significant kernel-space time skew as long as
>> the userspace time sources are within epsilon of atomic-clock time.
>>
>> Therefore, when watchdog verification of TSC is disabled, enable it for
>> HPET and PMTMR (AKA ACPI PM timer).  This provides the needed in-kernel
>> time-skew diagnostic without degrading the system's performance.
> I'm more than unhappy about this. We finally have a point where the TSC
> watchdog overhead can go away without adding TSC=reliable to the kernel
> commandline.
>
> Now you add an unconditionally enforce the watchdog again in a way which
> even cannot be disabled on the kernel command line.
>
> Patently bad idea, no cookies for you!

I have a similar concern about this patch as well. That is why I was 
suggesting to have this enabled for a limited time after boot for sanity 
checking purpose only.

The previous "[PATCH clocksource 5/6] clocksource: Suspend the watchdog 
temporarily when high read latency detected" patch, however, should be 
fine. Right?

Cheers,
Longman
  

On Wed, Feb 01, 2023 at 11:24:14AM +0100, Thomas Gleixner wrote:
> Paul!
> 
> On Tue, Jan 24 2023 at 16:27, Paul E. McKenney wrote:
> > On systems with two or fewer sockets, when the boot CPU has CONSTANT_TSC,
> > NONSTOP_TSC, and TSC_ADJUST, clocksource watchdog verification of the
> > TSC is disabled.  This works well much of the time, but there is the
> > occasional production-level system that meets all of these criteria, but
> > which still has a TSC that skews significantly from atomic-clock time.
> > This is usually attributed to a firmware or hardware fault.  Yes, the
> > various NTP daemons do express their opinions of userspace-to-atomic-clock
> > time skew, but they put them in various places, depending on the daemon
> > and distro in question.  It would therefore be good for the kernel to
> > have some clue that there is a problem.
> >
> > The old behavior of marking the TSC unstable is a non-starter because a
> > great many workloads simply cannot tolerate the overheads and latencies
> > of the various non-TSC clocksources.  In addition, NTP-corrected systems
> > sometimes can tolerate significant kernel-space time skew as long as
> > the userspace time sources are within epsilon of atomic-clock time.
> >
> > Therefore, when watchdog verification of TSC is disabled, enable it for
> > HPET and PMTMR (AKA ACPI PM timer).  This provides the needed in-kernel
> > time-skew diagnostic without degrading the system's performance.
> 
> I'm more than unhappy about this. We finally have a point where the TSC
> watchdog overhead can go away without adding TSC=reliable to the kernel
> commandline.
> 
> Now you add an unconditionally enforce the watchdog again in a way which
> even cannot be disabled on the kernel command line.
> 
> Patently bad idea, no cookies for you!

What can I say?  40,000 parts per million TSC clock skew did raise some
eyebrows, and therefore the complete suppressing of that diagnostic
completely was not at all a welcome development.

So how about the (untested) patch below, either on top of the existing
series or folded into e57818b20b0b ("clocksource: Verify HPET and PMTMR
when TSC unverified")?

The idea is to provide TSC checking of HPET and PMTMR only when the
TSC is deemed reliable and the new tsc=watchdog kernel boot parameter
is provided.  If both tsc=watchdog and tsc=nowatchdog are provided,
tsc=watchdog wins and a console message is emitted (no splat).

To restate, with this patch, unless the sysadmin asks for it, there will
be no clocksource watchdog unless there also would have been one without
this patch.

Thoughts?

							Thanx, Paul

------------------------------------------------------------------------

diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt
index 115829d71d0ca..d681f9252aaa7 100644
--- a/Documentation/admin-guide/kernel-parameters.txt
+++ b/Documentation/admin-guide/kernel-parameters.txt
@@ -6382,6 +6382,12 @@
 			(HPET or PM timer) on systems whose TSC frequency was
 			obtained from HW or FW using either an MSR or CPUID(0x15).
 			Warn if the difference is more than 500 ppm.
+			[x86] watchdog: Use TSC as the watchdog clocksource with
+			which to check other HW timers (HPET or PM timer), but
+			only on systems where TSC has been deemed trustworthy.
+			This will be suppressed by an earlier tsc=nowatchdog and
+			can be overridden by a later tsc=nowatchdog.  A console
+			message will flag any such suppression or overriding.
 
 	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 a5371c6d4b64b..306c233c98d84 100644
--- a/arch/x86/kernel/tsc.c
+++ b/arch/x86/kernel/tsc.c
@@ -294,6 +294,7 @@ __setup("notsc", notsc_setup);
 
 static int no_sched_irq_time;
 static int no_tsc_watchdog;
+static int tsc_as_watchdog;
 
 static int __init tsc_setup(char *str)
 {
@@ -303,10 +304,22 @@ static int __init tsc_setup(char *str)
 		no_sched_irq_time = 1;
 	if (!strcmp(str, "unstable"))
 		mark_tsc_unstable("boot parameter");
-	if (!strcmp(str, "nowatchdog"))
+	if (!strcmp(str, "nowatchdog")) {
 		no_tsc_watchdog = 1;
+		if (tsc_as_watchdog)
+			pr_alert("%s: Overriding earlier tsc=watchdog with tsc=nowatchdog\n",
+				 __func__);
+		tsc_as_watchdog = 0;
+	}
 	if (!strcmp(str, "recalibrate"))
 		tsc_force_recalibrate = 1;
+	if (!strcmp(str, "watchdog")) {
+		if (no_tsc_watchdog)
+			pr_alert("%s: tsc=watchdog overridden by earlier tsc=nowatchdog\n",
+				 __func__);
+		else
+			tsc_as_watchdog = 1;
+	}
 	return 1;
 }
 
@@ -1192,7 +1205,8 @@ static void __init tsc_disable_clocksource_watchdog(void)
 
 bool tsc_clocksource_watchdog_disabled(void)
 {
-	return !(clocksource_tsc.flags & CLOCK_SOURCE_MUST_VERIFY);
+	return !(clocksource_tsc.flags & CLOCK_SOURCE_MUST_VERIFY) &&
+	       tsc_as_watchdog && !no_tsc_watchdog;
 }
 
 static void __init check_system_tsc_reliable(void)
  

On Wed, Feb 01, 2023 at 02:26:29PM -0500, Waiman Long wrote:
> On 2/1/23 05:24, Thomas Gleixner wrote:
> > Paul!
> > 
> > On Tue, Jan 24 2023 at 16:27, Paul E. McKenney wrote:
> > > On systems with two or fewer sockets, when the boot CPU has CONSTANT_TSC,
> > > NONSTOP_TSC, and TSC_ADJUST, clocksource watchdog verification of the
> > > TSC is disabled.  This works well much of the time, but there is the
> > > occasional production-level system that meets all of these criteria, but
> > > which still has a TSC that skews significantly from atomic-clock time.
> > > This is usually attributed to a firmware or hardware fault.  Yes, the
> > > various NTP daemons do express their opinions of userspace-to-atomic-clock
> > > time skew, but they put them in various places, depending on the daemon
> > > and distro in question.  It would therefore be good for the kernel to
> > > have some clue that there is a problem.
> > > 
> > > The old behavior of marking the TSC unstable is a non-starter because a
> > > great many workloads simply cannot tolerate the overheads and latencies
> > > of the various non-TSC clocksources.  In addition, NTP-corrected systems
> > > sometimes can tolerate significant kernel-space time skew as long as
> > > the userspace time sources are within epsilon of atomic-clock time.
> > > 
> > > Therefore, when watchdog verification of TSC is disabled, enable it for
> > > HPET and PMTMR (AKA ACPI PM timer).  This provides the needed in-kernel
> > > time-skew diagnostic without degrading the system's performance.
> > I'm more than unhappy about this. We finally have a point where the TSC
> > watchdog overhead can go away without adding TSC=reliable to the kernel
> > commandline.
> > 
> > Now you add an unconditionally enforce the watchdog again in a way which
> > even cannot be disabled on the kernel command line.
> > 
> > Patently bad idea, no cookies for you!
> 
> I have a similar concern about this patch as well. That is why I was
> suggesting to have this enabled for a limited time after boot for sanity
> checking purpose only.

Fair enough!

If the watchdog checking of HPET and/or PMTMR against TSC only happens
only when the sysadm asks for it, would you still want to have the ability
to enable such watchdog checking at boot time, and then to disable it
once the system had been running for some limited time?

							Thanx, Paul
  

On 2/1/23 14:55, Paul E. McKenney wrote:
> On Wed, Feb 01, 2023 at 02:26:29PM -0500, Waiman Long wrote:
>> On 2/1/23 05:24, Thomas Gleixner wrote:
>>> Paul!
>>>
>>> On Tue, Jan 24 2023 at 16:27, Paul E. McKenney wrote:
>>>> On systems with two or fewer sockets, when the boot CPU has CONSTANT_TSC,
>>>> NONSTOP_TSC, and TSC_ADJUST, clocksource watchdog verification of the
>>>> TSC is disabled.  This works well much of the time, but there is the
>>>> occasional production-level system that meets all of these criteria, but
>>>> which still has a TSC that skews significantly from atomic-clock time.
>>>> This is usually attributed to a firmware or hardware fault.  Yes, the
>>>> various NTP daemons do express their opinions of userspace-to-atomic-clock
>>>> time skew, but they put them in various places, depending on the daemon
>>>> and distro in question.  It would therefore be good for the kernel to
>>>> have some clue that there is a problem.
>>>>
>>>> The old behavior of marking the TSC unstable is a non-starter because a
>>>> great many workloads simply cannot tolerate the overheads and latencies
>>>> of the various non-TSC clocksources.  In addition, NTP-corrected systems
>>>> sometimes can tolerate significant kernel-space time skew as long as
>>>> the userspace time sources are within epsilon of atomic-clock time.
>>>>
>>>> Therefore, when watchdog verification of TSC is disabled, enable it for
>>>> HPET and PMTMR (AKA ACPI PM timer).  This provides the needed in-kernel
>>>> time-skew diagnostic without degrading the system's performance.
>>> I'm more than unhappy about this. We finally have a point where the TSC
>>> watchdog overhead can go away without adding TSC=reliable to the kernel
>>> commandline.
>>>
>>> Now you add an unconditionally enforce the watchdog again in a way which
>>> even cannot be disabled on the kernel command line.
>>>
>>> Patently bad idea, no cookies for you!
>> I have a similar concern about this patch as well. That is why I was
>> suggesting to have this enabled for a limited time after boot for sanity
>> checking purpose only.
> Fair enough!
>
> If the watchdog checking of HPET and/or PMTMR against TSC only happens
> only when the sysadm asks for it, would you still want to have the ability
> to enable such watchdog checking at boot time, and then to disable it
> once the system had been running for some limited time?

Yes, being optional is another way to avoid the overhead for the 
majority of users. The paranoids can turn it on if they want to.

Cheers,
Longman
  

On Wed, Feb 01, 2023 at 10:40:56PM -0500, Waiman Long wrote:
> On 2/1/23 14:55, Paul E. McKenney wrote:
> > On Wed, Feb 01, 2023 at 02:26:29PM -0500, Waiman Long wrote:
> > > On 2/1/23 05:24, Thomas Gleixner wrote:
> > > > Paul!
> > > > 
> > > > On Tue, Jan 24 2023 at 16:27, Paul E. McKenney wrote:
> > > > > On systems with two or fewer sockets, when the boot CPU has CONSTANT_TSC,
> > > > > NONSTOP_TSC, and TSC_ADJUST, clocksource watchdog verification of the
> > > > > TSC is disabled.  This works well much of the time, but there is the
> > > > > occasional production-level system that meets all of these criteria, but
> > > > > which still has a TSC that skews significantly from atomic-clock time.
> > > > > This is usually attributed to a firmware or hardware fault.  Yes, the
> > > > > various NTP daemons do express their opinions of userspace-to-atomic-clock
> > > > > time skew, but they put them in various places, depending on the daemon
> > > > > and distro in question.  It would therefore be good for the kernel to
> > > > > have some clue that there is a problem.
> > > > > 
> > > > > The old behavior of marking the TSC unstable is a non-starter because a
> > > > > great many workloads simply cannot tolerate the overheads and latencies
> > > > > of the various non-TSC clocksources.  In addition, NTP-corrected systems
> > > > > sometimes can tolerate significant kernel-space time skew as long as
> > > > > the userspace time sources are within epsilon of atomic-clock time.
> > > > > 
> > > > > Therefore, when watchdog verification of TSC is disabled, enable it for
> > > > > HPET and PMTMR (AKA ACPI PM timer).  This provides the needed in-kernel
> > > > > time-skew diagnostic without degrading the system's performance.
> > > > I'm more than unhappy about this. We finally have a point where the TSC
> > > > watchdog overhead can go away without adding TSC=reliable to the kernel
> > > > commandline.
> > > > 
> > > > Now you add an unconditionally enforce the watchdog again in a way which
> > > > even cannot be disabled on the kernel command line.
> > > > 
> > > > Patently bad idea, no cookies for you!
> > > I have a similar concern about this patch as well. That is why I was
> > > suggesting to have this enabled for a limited time after boot for sanity
> > > checking purpose only.
> > Fair enough!
> > 
> > If the watchdog checking of HPET and/or PMTMR against TSC only happens
> > only when the sysadm asks for it, would you still want to have the ability
> > to enable such watchdog checking at boot time, and then to disable it
> > once the system had been running for some limited time?
> 
> Yes, being optional is another way to avoid the overhead for the majority of
> users. The paranoids can turn it on if they want to.

Very good, thank you!

							Thanx, Paul
  

On Wed, Feb 01 2023 at 22:40, Waiman Long wrote:
> On 2/1/23 14:55, Paul E. McKenney wrote:
>>>>> Therefore, when watchdog verification of TSC is disabled, enable it for
>>>>> HPET and PMTMR (AKA ACPI PM timer).  This provides the needed in-kernel
>>>>> time-skew diagnostic without degrading the system's performance.
>>>> I'm more than unhappy about this. We finally have a point where the TSC
>>>> watchdog overhead can go away without adding TSC=reliable to the kernel
>>>> commandline.
>>>>
>>>> Now you add an unconditionally enforce the watchdog again in a way which
>>>> even cannot be disabled on the kernel command line.
>>>>
>>>> Patently bad idea, no cookies for you!
>>> I have a similar concern about this patch as well. That is why I was
>>> suggesting to have this enabled for a limited time after boot for sanity
>>> checking purpose only.
>> Fair enough!
>>
>> If the watchdog checking of HPET and/or PMTMR against TSC only happens
>> only when the sysadm asks for it, would you still want to have the ability
>> to enable such watchdog checking at boot time, and then to disable it
>> once the system had been running for some limited time?
>
> Yes, being optional is another way to avoid the overhead for the 
> majority of users. The paranoids can turn it on if they want to.

Yes, opt-in is good enough.

Thanks,

        tglx
  

On Thu, Feb 02, 2023 at 08:57:39AM +0100, Thomas Gleixner wrote:
> On Wed, Feb 01 2023 at 22:40, Waiman Long wrote:
> > On 2/1/23 14:55, Paul E. McKenney wrote:
> >>>>> Therefore, when watchdog verification of TSC is disabled, enable it for
> >>>>> HPET and PMTMR (AKA ACPI PM timer).  This provides the needed in-kernel
> >>>>> time-skew diagnostic without degrading the system's performance.
> >>>> I'm more than unhappy about this. We finally have a point where the TSC
> >>>> watchdog overhead can go away without adding TSC=reliable to the kernel
> >>>> commandline.
> >>>>
> >>>> Now you add an unconditionally enforce the watchdog again in a way which
> >>>> even cannot be disabled on the kernel command line.
> >>>>
> >>>> Patently bad idea, no cookies for you!
> >>> I have a similar concern about this patch as well. That is why I was
> >>> suggesting to have this enabled for a limited time after boot for sanity
> >>> checking purpose only.
> >> Fair enough!
> >>
> >> If the watchdog checking of HPET and/or PMTMR against TSC only happens
> >> only when the sysadm asks for it, would you still want to have the ability
> >> to enable such watchdog checking at boot time, and then to disable it
> >> once the system had been running for some limited time?
> >
> > Yes, being optional is another way to avoid the overhead for the 
> > majority of users. The paranoids can turn it on if they want to.
> 
> Yes, opt-in is good enough.

I have added this commit to my clocksource branch:

2ff7dacc88b0 clocksource: Enable TSC watchdog checking of HPET and PMTMR only when requested

It is passing my internal tests, and if it does fine for a couple of
days in -next, I will send an updated pull request.

							Thanx, Paul

------------------------------------------------------------------------

The full clocksource branch:

beaa1ffe551c3 clocksource: Print clocksource name when clocksource is tested unstable
c37e85c135cea clocksource: Loosen clocksource watchdog constraints
f092eb34b3304 clocksource: Improve read-back-delay message
dd029269947a3 clocksource: Improve "skew is too large" messages
b7082cdfc464b clocksource: Suspend the watchdog temporarily when high read latency detected
a7ec817d55421 x86/tsc: Add option to force frequency recalibration with HW timer
efc8b329c7fdc clocksource: Verify HPET and PMTMR when TSC unverified
2ff7dacc88b0c clocksource: Enable TSC watchdog checking of HPET and PMTMR only when requested