tsc: make calibration refinement more robust
PATCH 1/1 tsc: make calibration refinement more robust
The threshold in tsc_read_refs() is constant which may favor slower CPUs but may not be optimal for simple reading of reference on faster ones. Hence make it proportional to tsc_khz to compensate for this. The threshold guards against any disturbance like IRQs, NMIs, SMIs or CPU stealing by host on guest systems so rename it accordingly and fix comments as well.
Also on some systems there is noticeable DMI bus contention at some point during boot keeping the readout failing (observed with about one in ~300 boots when testing). In that case retry also the second readout instead of simply bailing out unrefined. Usually the next second the readout returns fast just fine without any issues.
--- arch/x86/kernel/tsc.c | 21 +++++++++++---------- 1 file changed, 11 insertions(+), 10 deletions(-)
arch/x86/kernel/tsc.c +11 -10
@@ -297,11 +297,11 @@ static int __init tsc_setup(char *str)__setup ("tsc=" , tsc_setup); -#define MAX_RETRIES 5 -#define SMI_TRESHOLD 50000 +#define MAX_RETRIES 5 +#define TSC_THRESHOLD (tsc_khz >> 5)/* -* Read TSC and the reference counters. Take care of SMI disturbance +* Read TSC and the reference counters. Take care of any disturbances */ static u64 tsc_read_refs (u64 *p,int hpet) {
@@ -315,7 +315,7 @@ static u64 tsc_read_refs(u64 *p, int hpet)else *p =acpi_pm_read_early (); t2 =get_cycles (); -if ((t2 - t1) <SMI_TRESHOLD ) +if ((t2 - t1) <TSC_THRESHOLD )return t2; }return ULLONG_MAX ;
@@ -1268,7 +1268,7 @@ struct system_counterval_t convert_art_ns_to_tsc(u64 art_ns)*/ static void tsc_refine_calibration_work (struct work_struct *work) { -static u64 tsc_start =-1 , ref_start; +static u64 tsc_start =ULLONG_MAX , ref_start;static int hpet;u64 tsc_stop, ref_stop, delta;unsigned long freq;
@@ -1283,14 +1283,15 @@ static void tsc_refine_calibration_work(struct work_struct *work)* delayed the first time we expire. So set the workqueue * again once we know timers are working. */ -if (tsc_start ==-1 ) { +if (tsc_start ==ULLONG_MAX ) { +restart:/* * Only set hpet once, to avoid mixing hardware * if the hpet becomes enabled later. */ hpet =is_hpet_enabled (); -schedule_delayed_work (&tsc_irqwork,HZ ); tsc_start =tsc_read_refs (&ref_start, hpet); +schedule_delayed_work (&tsc_irqwork,HZ );return ; }
@@ -1300,9 +1301,9 @@ static void tsc_refine_calibration_work(struct work_struct *work)if (ref_start == ref_stop) goto out; -/* Check, whether the sampling was disturbed by an SMI */ -if (tsc_start ==ULLONG_MAX || tsc_stop ==ULLONG_MAX ) - goto out; +/* Check, whether the sampling was disturbed */ +if (tsc_stop ==ULLONG_MAX ) + goto restart; delta = tsc_stop - tsc_start; delta *=1000000LL ;
Signature · 1 line
2.19.1
Daniel,
On Thu, 1 Nov 2018, Daniel Vacek wrote:
Please use 'x86/tsc:' as prefix. git log path/to/file usually gives you a reasonable hint about prefixes.
arch/x86/kernel/tsc.c static int __init tsc_setup(char *str) · line 301
Quoted 4 lines
arch/x86/kernel/tsc.c +2
-2-#define MAX_RETRIES 5 -#define SMI_TRESHOLD 50000 +#define MAX_RETRIES 5 +#define TSC_THRESHOLD (tsc_khz >> 5)
This breaks pit_hpet_ptimer_calibrate_cpu() because at that point tsc_hkz is 0.
Thanks,
tglx
Hi Thomas,
thanks for checking.
Quoted 6 lines · Thomas Gleixner
Daniel,
On Thu, 1 Nov 2018, Daniel Vacek wrote:
Please use 'x86/tsc:' as prefix. git log path/to/file usually gives you a reasonable hint about prefixes.
Oh, sure thing. The dmesg always prints 'tsc:' - I somehow sticked to it...
Quoted 6 lines
arch/x86/kernel/tsc.c static int __init tsc_setup(char *str) · line 301
Quoted 4 lines
arch/x86/kernel/tsc.c +2
-2-#define MAX_RETRIES 5 -#define SMI_TRESHOLD 50000 +#define MAX_RETRIES 5 +#define TSC_THRESHOLD (tsc_khz >> 5)This breaks
pit_hpet_ptimer_calibrate_cpu()because at that pointtsc_hkzis 0.
That did not show up with my testing, sorry. I guess pit_calibrate_tsc() never failed for me. Hmm, actually it looks like quick_pit_calibrate() does the job for me so pit_hpet_ptimer_calibrate_cpu() is likely not even called. Would this:
#define TSC_THRESHOLD (tsc_khz? tsc_khz >> 5: 0x20000)
work for you instead? Or alternatively at some point when chasing this down I used:
#define TSC_THRESHOLD (0x10000 + (tsc_khz >> 6))
The first one seems better though. I can send v2 next week if you like it.
--nX
Quoted 3 lines
Thanks,
tglx
Daniel,
Quoted 8 lines · Daniel Vacek
Quoted 6 lines · Thomas Gleixner
arch/x86/kernel/tsc.c static int __init tsc_setup(char *str) · line 301
Quoted 4 lines
arch/x86/kernel/tsc.c +2
-2-#define MAX_RETRIES 5 -#define SMI_TRESHOLD 50000 +#define MAX_RETRIES 5 +#define TSC_THRESHOLD (tsc_khz >> 5)This breaks
pit_hpet_ptimer_calibrate_cpu()because at that pointtsc_hkzis 0.
Quoted 4 lines
That did not show up with my testing, sorry. I guess
pit_calibrate_tsc()never failed for me. Hmm, actually it looks likequick_pit_calibrate()does the job for me sopit_hpet_ptimer_calibrate_cpu()is likely not even called.
Right. It's only called when quick calibration fails. Testing does not replace code inspection :)
Quoted 10 lines
Would this:
#define
TSC_THRESHOLD(tsc_khz?tsc_khz>> 5:0x20000)work for you instead? Or alternatively at some point when chasing this down I used:
#define
TSC_THRESHOLD(0x10000+ (tsc_khz>> 6))The first one seems better though. I can send v2 next week if you like it.
Can you please avoid hiding the logic in a macro? Just use a local variable:
u64 thresh = tsc_khz ? tsc_khz >> 5 : TSC_DEFAULT_THRESHOLD;
and use that in the comparison.
Thanks,
tglx
Quoted 11 lines · Thomas Gleixner
Daniel,
Quoted 8 lines · Daniel Vacek
Quoted 6 lines · Thomas Gleixner
arch/x86/kernel/tsc.c static int __init tsc_setup(char *str) · line 301
Quoted 4 lines
arch/x86/kernel/tsc.c +2
-2-#define MAX_RETRIES 5 -#define SMI_TRESHOLD 50000 +#define MAX_RETRIES 5 +#define TSC_THRESHOLD (tsc_khz >> 5)This breaks
pit_hpet_ptimer_calibrate_cpu()because at that pointtsc_hkzis 0.
Quoted 7 lines
Quoted 4 lines
That did not show up with my testing, sorry. I guess
pit_calibrate_tsc()never failed for me. Hmm, actually it looks likequick_pit_calibrate()does the job for me sopit_hpet_ptimer_calibrate_cpu()is likely not even called.Right. It's only called when quick calibration fails. Testing does not replace code inspection :)
Agreed. I was not 100% sure about this early init and order of execution as it's dynamically changed with x86_platform.calibrate_cpu and x86_platform.calibrate_tsc. Thanks again for the review, Thomas.
Quoted 6 lines
Can you please avoid hiding the logic in a macro? Just use a local variable:
u64 thresh = tsc_khz ? tsc_khz >> 5 : TSC_DEFAULT_THRESHOLD;and use that in the comparison.
Sweet, I'll do that :)
--nX
Quoted 3 lines
Thanks,
tglx
PATCH v2 1/1 x86/tsc: make calibration refinement more robust
In reply to Daniel VacekThe threshold in tsc_read_refs() is constant which may favor slower CPUs but may not be optimal for simple reading of reference on faster ones. Hence make it proportional to tsc_khz when available to compensate for this. The threshold guards against any disturbance like IRQs, NMIs, SMIs or CPU stealing by host on guest systems so rename it accordingly and fix comments as well.
Also on some systems there is noticeable DMI bus contention at some point during boot keeping the readout failing (observed with about one in ~300 boots when testing). In that case retry also the second readout instead of simply bailing out unrefined. Usually the next second the readout returns fast just fine without any issues.
v2: keep using the constant early when the tsc_khz is not available yet
as suggested by tglx
--- arch/x86/kernel/tsc.c | 30 ++++++++++++++++-------------- 1 file changed, 16 insertions(+), 14 deletions(-)
arch/x86/kernel/tsc.c +16 -14
@@ -297,15 +297,16 @@ static int __init tsc_setup(char *str)__setup ("tsc=" , tsc_setup); -#define MAX_RETRIES 5 -#define SMI_TRESHOLD 50000 +#define MAX_RETRIES 5 +#define TSC_DEFAULT_THRESHOLD 0x20000/* -* Read TSC and the reference counters. Take care of SMI disturbance +* Read TSC and the reference counters. Take care of any disturbances */ static u64 tsc_read_refs (u64 *p,int hpet) {u64 t1, t2; +u64 thresh = tsc_khz ? tsc_khz >>5 :TSC_DEFAULT_THRESHOLD ;int i;for (i =0 ; i <MAX_RETRIES ; i++) {
@@ -315,7 +316,7 @@ static u64 tsc_read_refs(u64 *p, int hpet)else *p =acpi_pm_read_early (); t2 =get_cycles (); -if ((t2 - t1) <SMI_TRESHOLD ) +if ((t2 - t1) < thresh)return t2; }return ULLONG_MAX ;
@@ -703,15 +704,15 @@ static unsigned long pit_hpet_ptimer_calibrate_cpu(void)* zero. In each wait loop iteration we read the TSC and check * the delta to the previous read. We keep track of the min * and max values of that delta. The delta is mostly defined -* by the IO time of the PIT access, so we can detect when a -* SMI/SMM disturbance happened between the two reads. If the +* by the IO time of the PIT access, so we can detect when +* any disturbance happened between the two reads. If the * maximum time is significantly larger than the minimum time, * then we discard the result and have another try. * * 2) Reference counter. If available we use the HPET or the * PMTIMER as a reference to check the sanity of that value. * We use separate TSC readouts and check inside of the -* reference read for a SMI/SMM disturbance. We dicard +* reference read for any possible disturbance. We dicard * disturbed values here as well. We do that around the PIT * calibration delay loop as we have to wait for a certain * amount of time anyway.
@@ -744,7 +745,7 @@ static unsigned long pit_hpet_ptimer_calibrate_cpu(void)if (ref1 == ref2)continue ; -/* Check, whether the sampling was disturbed by an SMI */ +/* Check, whether the sampling was disturbed */ if (tsc1 ==ULLONG_MAX || tsc2 ==ULLONG_MAX )continue ;
@@ -1268,7 +1269,7 @@ struct system_counterval_t convert_art_ns_to_tsc(u64 art_ns)*/ static void tsc_refine_calibration_work (struct work_struct *work) { -static u64 tsc_start =-1 , ref_start; +static u64 tsc_start =ULLONG_MAX , ref_start;static int hpet;u64 tsc_stop, ref_stop, delta;unsigned long freq;
@@ -1283,14 +1284,15 @@ static void tsc_refine_calibration_work(struct work_struct *work)* delayed the first time we expire. So set the workqueue * again once we know timers are working. */ -if (tsc_start ==-1 ) { +if (tsc_start ==ULLONG_MAX ) { +restart:/* * Only set hpet once, to avoid mixing hardware * if the hpet becomes enabled later. */ hpet =is_hpet_enabled (); -schedule_delayed_work (&tsc_irqwork,HZ ); tsc_start =tsc_read_refs (&ref_start, hpet); +schedule_delayed_work (&tsc_irqwork,HZ );return ; }
@@ -1300,9 +1302,9 @@ static void tsc_refine_calibration_work(struct work_struct *work)if (ref_start == ref_stop) goto out; -/* Check, whether the sampling was disturbed by an SMI */ -if (tsc_start ==ULLONG_MAX || tsc_stop ==ULLONG_MAX ) - goto out; +/* Check, whether the sampling was disturbed */ +if (tsc_stop ==ULLONG_MAX ) + goto restart; delta = tsc_stop - tsc_start; delta *=1000000LL ;
Signature · 1 line
2.19.1
tip-bot for Daniel Vacek
merge bot[tip:x86/timers] x86/tsc: Make calibration refinement more robust
x86/tsc: Make calibration refinement more robust
Commit-ID: a786ef152cdcfebc923a67f63c7815806eefcf81 Gitweb: https://git.kernel.org/tip/a786ef152cdcfebc923a67f63c7815806eefcf81 Author: Daniel Vacek <[email protected]> AuthorDate: Mon, 5 Nov 2018 18:10:40 +0100 Committer: Thomas Gleixner <[email protected]> CommitDate: Tue, 6 Nov 2018 21:53:15 +0100
x86/tsc: Make calibration refinement more robust
The threshold in tsc_read_refs() is constant which may favor slower CPUs but may not be optimal for simple reading of reference on faster ones.
Hence make it proportional to tsc_khz when available to compensate for this. The threshold guards against any disturbance like IRQs, NMIs, SMIs or CPU stealing by host on guest systems so rename it accordingly and fix comments as well.
Also on some systems there is noticeable DMI bus contention at some point during boot keeping the readout failing (observed with about one in ~300 boots when testing). In that case retry also the second readout instead of simply bailing out unrefined. Usually the next second the readout returns fast just fine without any issues.
--- arch/x86/kernel/tsc.c | 30 ++++++++++++++++-------------- 1 file changed, 16 insertions(+), 14 deletions(-)
arch/x86/kernel/tsc.c +16 -14
@@ -297,15 +297,16 @@ static int __init tsc_setup(char *str)__setup ("tsc=" , tsc_setup); -#define MAX_RETRIES 5 -#define SMI_TRESHOLD 50000 +#define MAX_RETRIES 5 +#define TSC_DEFAULT_THRESHOLD 0x20000/* -* Read TSC and the reference counters. Take care of SMI disturbance +* Read TSC and the reference counters. Take care of any disturbances */ static u64 tsc_read_refs (u64 *p,int hpet) {u64 t1, t2; +u64 thresh = tsc_khz ? tsc_khz >>5 :TSC_DEFAULT_THRESHOLD ;int i;for (i =0 ; i <MAX_RETRIES ; i++) {
@@ -315,7 +316,7 @@ static u64 tsc_read_refs(u64 *p, int hpet)else *p =acpi_pm_read_early (); t2 =get_cycles (); -if ((t2 - t1) <SMI_TRESHOLD ) +if ((t2 - t1) < thresh)return t2; }return ULLONG_MAX ;
@@ -703,15 +704,15 @@ static unsigned long pit_hpet_ptimer_calibrate_cpu(void)* zero. In each wait loop iteration we read the TSC and check * the delta to the previous read. We keep track of the min * and max values of that delta. The delta is mostly defined -* by the IO time of the PIT access, so we can detect when a -* SMI/SMM disturbance happened between the two reads. If the +* by the IO time of the PIT access, so we can detect when +* any disturbance happened between the two reads. If the * maximum time is significantly larger than the minimum time, * then we discard the result and have another try. * * 2) Reference counter. If available we use the HPET or the * PMTIMER as a reference to check the sanity of that value. * We use separate TSC readouts and check inside of the -* reference read for a SMI/SMM disturbance. We dicard +* reference read for any possible disturbance. We dicard * disturbed values here as well. We do that around the PIT * calibration delay loop as we have to wait for a certain * amount of time anyway.
@@ -744,7 +745,7 @@ static unsigned long pit_hpet_ptimer_calibrate_cpu(void)if (ref1 == ref2)continue ; -/* Check, whether the sampling was disturbed by an SMI */ +/* Check, whether the sampling was disturbed */ if (tsc1 ==ULLONG_MAX || tsc2 ==ULLONG_MAX )continue ;
@@ -1268,7 +1269,7 @@ static DECLARE_DELAYED_WORK(tsc_irqwork, tsc_refine_calibration_work);*/ static void tsc_refine_calibration_work (struct work_struct *work) { -static u64 tsc_start =-1 , ref_start; +static u64 tsc_start =ULLONG_MAX , ref_start;static int hpet;u64 tsc_stop, ref_stop, delta;unsigned long freq;
@@ -1283,14 +1284,15 @@ static void tsc_refine_calibration_work(struct work_struct *work)* delayed the first time we expire. So set the workqueue * again once we know timers are working. */ -if (tsc_start ==-1 ) { +if (tsc_start ==ULLONG_MAX ) { +restart:/* * Only set hpet once, to avoid mixing hardware * if the hpet becomes enabled later. */ hpet =is_hpet_enabled (); -schedule_delayed_work (&tsc_irqwork,HZ ); tsc_start =tsc_read_refs (&ref_start, hpet); +schedule_delayed_work (&tsc_irqwork,HZ );return ; }
@@ -1300,9 +1302,9 @@ static void tsc_refine_calibration_work(struct work_struct *work)if (ref_start == ref_stop) goto out; -/* Check, whether the sampling was disturbed by an SMI */ -if (tsc_start ==ULLONG_MAX || tsc_stop ==ULLONG_MAX ) - goto out; +/* Check, whether the sampling was disturbed */ +if (tsc_stop ==ULLONG_MAX ) + goto restart; delta = tsc_stop - tsc_start; delta *=1000000LL ;
Re: [tip:x86/timers] x86/tsc: Make calibration refinement more robust
Hi Thomas, Ingo, Peter.
I'm wondering, was x86/timers branch of tip tree merged to linus' tree for v5.0-rc1? Somehow I do not see this patch make it through...
Am I doing something wrong?
--nX
Quoted 135 lines · tip-bot for Daniel Vacek
Commit-ID: a786ef152cdcfebc923a67f63c7815806eefcf81 Gitweb: https://git.kernel.org/tip/a786ef152cdcfebc923a67f63c7815806eefcf81 Author: Daniel Vacek <[email protected]> AuthorDate: Mon, 5 Nov 2018 18:10:40 +0100 Committer: Thomas Gleixner <[email protected]> CommitDate: Tue, 6 Nov 2018 21:53:15 +0100x86/tsc: Make calibration refinement more robust
The threshold in
tsc_read_refs()is constant which may favor slower CPUs but may not be optimal for simple reading of reference on faster ones.Hence make it proportional to
tsc_khzwhen available to compensate for this. The threshold guards against any disturbance like IRQs, NMIs, SMIs or CPU stealing by host on guest systems so rename it accordingly and fix comments as well.Also on some systems there is noticeable DMI bus contention at some point during boot keeping the readout failing (observed with about one in ~300 boots when testing). In that case retry also the second readout instead of simply bailing out unrefined. Usually the next second the readout returns fast just fine without any issues.
Signed-off-by Daniel Vacek <[email protected]>Signed-off-by Thomas Gleixner <[email protected]>Cc Borislav Petkov <[email protected]>Cc "H. Peter Anvin" <[email protected]>--- arch/x86/kernel/tsc.c | 30 ++++++++++++++++-------------- 1 file changed, 16 insertions(+), 14 deletions(-)arch/x86/kernel/tsc.c +16
-14@@ -297,15 +297,16 @@ static int __init tsc_setup(char *str)__setup ("tsc=" , tsc_setup); -#define MAX_RETRIES 5 -#define SMI_TRESHOLD 50000 +#define MAX_RETRIES 5 +#define TSC_DEFAULT_THRESHOLD 0x20000/* -* Read TSC and the reference counters. Take care of SMI disturbance +* Read TSC and the reference counters. Take care of any disturbances */ static u64 tsc_read_refs (u64 *p,int hpet) {u64 t1, t2; +u64 thresh = tsc_khz ? tsc_khz >>5 :TSC_DEFAULT_THRESHOLD ;int i;for (i =0 ; i <MAX_RETRIES ; i++) {@@ -315,7 +316,7 @@ static u64 tsc_read_refs(u64 *p, int hpet)else *p =acpi_pm_read_early (); t2 =get_cycles (); -if ((t2 - t1) <SMI_TRESHOLD ) +if ((t2 - t1) < thresh)return t2; }return ULLONG_MAX ;@@ -703,15 +704,15 @@ static unsigned long pit_hpet_ptimer_calibrate_cpu(void)* zero. In each wait loop iteration we read the TSC and check * the delta to the previous read. We keep track of the min * and max values of that delta. The delta is mostly defined -* by the IO time of the PIT access, so we can detect when a -* SMI/SMM disturbance happened between the two reads. If the +* by the IO time of the PIT access, so we can detect when +* any disturbance happened between the two reads. If the * maximum time is significantly larger than the minimum time, * then we discard the result and have another try. * * 2) Reference counter. If available we use the HPET or the * PMTIMER as a reference to check the sanity of that value. * We use separate TSC readouts and check inside of the -* reference read for a SMI/SMM disturbance. We dicard +* reference read for any possible disturbance. We dicard * disturbed values here as well. We do that around the PIT * calibration delay loop as we have to wait for a certain * amount of time anyway. @@ -744,7 +745,7 @@ static unsigned long pit_hpet_ptimer_calibrate_cpu(void)if (ref1 == ref2)continue ; -/* Check, whether the sampling was disturbed by an SMI */ +/* Check, whether the sampling was disturbed */ if (tsc1 ==ULLONG_MAX || tsc2 ==ULLONG_MAX )continue ;@@ -1268,7 +1269,7 @@ static DECLARE_DELAYED_WORK(tsc_irqwork, tsc_refine_calibration_work);*/ static void tsc_refine_calibration_work (struct work_struct *work) { -static u64 tsc_start =-1 , ref_start; +static u64 tsc_start =ULLONG_MAX , ref_start;static int hpet;u64 tsc_stop, ref_stop, delta;unsigned long freq;@@ -1283,14 +1284,15 @@ static void tsc_refine_calibration_work(struct work_struct *work)* delayed the first time we expire. So set the workqueue * again once we know timers are working. */ -if (tsc_start ==-1 ) { +if (tsc_start ==ULLONG_MAX ) { +restart:/* * Only set hpet once, to avoid mixing hardware * if the hpet becomes enabled later. */ hpet =is_hpet_enabled (); -schedule_delayed_work (&tsc_irqwork,HZ ); tsc_start =tsc_read_refs (&ref_start, hpet); +schedule_delayed_work (&tsc_irqwork,HZ );return ; }@@ -1300,9 +1302,9 @@ static void tsc_refine_calibration_work(struct work_struct *work)if (ref_start == ref_stop) goto out; -/* Check, whether the sampling was disturbed by an SMI */ -if (tsc_start ==ULLONG_MAX || tsc_stop ==ULLONG_MAX ) - goto out; +/* Check, whether the sampling was disturbed */ +if (tsc_stop ==ULLONG_MAX ) + goto restart; delta = tsc_stop - tsc_start; delta *=1000000LL ;
Re: [tip:x86/timers] x86/tsc: Make calibration refinement more robust
Daniel,
Quoted 4 lines · Daniel Vacek
I'm wondering, was x86/timers branch of tip tree merged to linus' tree for v5.0-rc1? Somehow I do not see this patch make it through...
Am I doing something wrong?
No. The branch was not sent to Linus during the merge window due to holiday season induced oversight.
Thanks,
tglx
Re: [tip:x86/timers] x86/tsc: Make calibration refinement more robust
Quoted 3 lines · Thomas Gleixner
No. The branch was not sent to Linus during the merge window due to holiday season induced oversight.
So does it wait for the next merge window now or is it still going to be sent? This is actually worth stable branches. I've confirmations from Hitachi and Nec that this really helps latest Skylake-X systems boot reliably.
--nX
Re: [tip:x86/timers] x86/tsc: Make calibration refinement more robust
Quoted 9 lines · Daniel Vacek
Quoted 3 lines · Thomas Gleixner
No. The branch was not sent to Linus during the merge window due to holiday season induced oversight.
So does it wait for the next merge window now or is it still going to be sent? This is actually worth stable branches. I've confirmations from Hitachi and Nec that this really helps latest Skylake-X systems boot reliably.
It's in 5.0-rc4 now.
Thanks,
tglx