WSL2-Linux-Kernel/arch/x86/kernel/tsc.c

1016 строки
26 KiB
C
Исходник Обычный вид История

#include <linux/kernel.h>
#include <linux/sched.h>
#include <linux/init.h>
#include <linux/module.h>
#include <linux/timer.h>
#include <linux/acpi_pmtmr.h>
#include <linux/cpufreq.h>
#include <linux/delay.h>
#include <linux/clocksource.h>
#include <linux/percpu.h>
#include <linux/timex.h>
#include <asm/hpet.h>
#include <asm/timer.h>
#include <asm/vgtod.h>
#include <asm/time.h>
#include <asm/delay.h>
x86: Hypervisor detection and get tsc_freq from hypervisor Impact: Changes timebase calibration on Vmware. v3->v2 : Abstract the hypervisor detection and feature (tsc_freq) request behind a hypervisor.c file v2->v1 : Add a x86_hyper_vendor field to the cpuinfo_x86 structure. This avoids multiple calls to the hypervisor detection function. This patch adds function to detect if we are running under VMware. The current way to check if we are on VMware is following, # check if "hypervisor present bit" is set, if so read the 0x40000000 cpuid leaf and check for "VMwareVMware" signature. # if the above fails, check the DMI vendors name for "VMware" string if we find one we query the VMware hypervisor port to check if we are under VMware. The DMI + "VMware hypervisor port check" is needed for older VMware products, which don't implement the hypervisor signature cpuid leaf. Also note that since we are checking for the DMI signature the hypervisor port should never be accessed on native hardware. This patch also adds a hypervisor_get_tsc_freq function, instead of calibrating the frequency which can be error prone in virtualized environment, we ask the hypervisor for it. We get the frequency from the hypervisor by accessing the hypervisor port if we are running on VMware. Other hypervisors too can add code to the generic routine to get frequency on their platform. Signed-off-by: Alok N Kataria <akataria@vmware.com> Signed-off-by: Dan Hecht <dhecht@vmware.com> Signed-off-by: H. Peter Anvin <hpa@zytor.com>
2008-10-27 20:41:46 +03:00
#include <asm/hypervisor.h>
#include <asm/nmi.h>
#include <asm/x86_init.h>
unsigned int __read_mostly cpu_khz; /* TSC clocks / usec, not used here */
EXPORT_SYMBOL(cpu_khz);
unsigned int __read_mostly tsc_khz;
EXPORT_SYMBOL(tsc_khz);
/*
* TSC can be unstable due to cpufreq or due to unsynced TSCs
*/
static int __read_mostly tsc_unstable;
/* native_sched_clock() is called before tsc_init(), so
we must start with the TSC soft disabled to prevent
erroneous rdtsc usage on !cpu_has_tsc processors */
static int __read_mostly tsc_disabled = -1;
int tsc_clocksource_reliable;
/*
* Scheduler clock - returns current time in nanosec units.
*/
u64 native_sched_clock(void)
{
u64 this_offset;
/*
* Fall back to jiffies if there's no TSC available:
* ( But note that we still use it if the TSC is marked
* unstable. We do this because unlike Time Of Day,
* the scheduler clock tolerates small errors and it's
* very important for it to be as fast as the platform
* can achieve it. )
*/
if (unlikely(tsc_disabled)) {
/* No locking but a rare wrong value is not a big deal: */
return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ);
}
/* read the Time Stamp Counter: */
rdtscll(this_offset);
/* return the value in ns */
return __cycles_2_ns(this_offset);
}
/* We need to define a real function for sched_clock, to override the
weak default version */
#ifdef CONFIG_PARAVIRT
unsigned long long sched_clock(void)
{
return paravirt_sched_clock();
}
#else
unsigned long long
sched_clock(void) __attribute__((alias("native_sched_clock")));
#endif
int check_tsc_unstable(void)
{
return tsc_unstable;
}
EXPORT_SYMBOL_GPL(check_tsc_unstable);
#ifdef CONFIG_X86_TSC
int __init notsc_setup(char *str)
{
printk(KERN_WARNING "notsc: Kernel compiled with CONFIG_X86_TSC, "
"cannot disable TSC completely.\n");
tsc_disabled = 1;
return 1;
}
#else
/*
* disable flag for tsc. Takes effect by clearing the TSC cpu flag
* in cpu/common.c
*/
int __init notsc_setup(char *str)
{
setup_clear_cpu_cap(X86_FEATURE_TSC);
return 1;
}
#endif
__setup("notsc", notsc_setup);
static int no_sched_irq_time;
x86: Skip verification by the watchdog for TSC clocksource. Impact: Changes timekeeping on Vmware (or with tsc=reliable). This is achieved by resetting the CLOCKSOURCE_MUST_VERIFY flag. We add a tsc=reliable commandline option to enable this. This enables legacy hardware without HPET, LAPIC, or ACPI timers to enter high-resolution timer mode. Along with that have extended this to be used in virtualization environement too. Now we also set this flag if the X86_FEATURE_TSC_RELIABLE bit is set. This is important since there is a wrap-around problem with the acpi_pm timer. The acpi_pm counter is just 24bits and this can overflow in ~4 seconds. With the NO_HZ kernels in virtualized environment, there can be situations when the guest is descheduled for longer duration, as a result we may miss the wrap of the acpi counter. When TSC is used as a clocksource and acpi_pm timer is being used as the watchdog clocksource this error in acpi_pm results in TSC being marked as unstable, and essentially results in time dropping in chunks of 4 seconds whenever this wrap is missed. Since the virtualized TSC is reliable on VMware, we should always use the TSCs clocksource on VMware, so we skip the verfication at runtime, by checking for the feature bit. Since we reset the flag for mgeode systems too, i have combined the mgeode case with the feature bit check. Signed-off-by: Jeff Hansen <jhansen@cardaccess-inc.com> Signed-off-by: Alok N Kataria <akataria@vmware.com> Signed-off-by: Dan Hecht <dhecht@vmware.com> Signed-off-by: H. Peter Anvin <hpa@zytor.com>
2008-10-25 04:22:01 +04:00
static int __init tsc_setup(char *str)
{
if (!strcmp(str, "reliable"))
tsc_clocksource_reliable = 1;
if (!strncmp(str, "noirqtime", 9))
no_sched_irq_time = 1;
x86: Skip verification by the watchdog for TSC clocksource. Impact: Changes timekeeping on Vmware (or with tsc=reliable). This is achieved by resetting the CLOCKSOURCE_MUST_VERIFY flag. We add a tsc=reliable commandline option to enable this. This enables legacy hardware without HPET, LAPIC, or ACPI timers to enter high-resolution timer mode. Along with that have extended this to be used in virtualization environement too. Now we also set this flag if the X86_FEATURE_TSC_RELIABLE bit is set. This is important since there is a wrap-around problem with the acpi_pm timer. The acpi_pm counter is just 24bits and this can overflow in ~4 seconds. With the NO_HZ kernels in virtualized environment, there can be situations when the guest is descheduled for longer duration, as a result we may miss the wrap of the acpi counter. When TSC is used as a clocksource and acpi_pm timer is being used as the watchdog clocksource this error in acpi_pm results in TSC being marked as unstable, and essentially results in time dropping in chunks of 4 seconds whenever this wrap is missed. Since the virtualized TSC is reliable on VMware, we should always use the TSCs clocksource on VMware, so we skip the verfication at runtime, by checking for the feature bit. Since we reset the flag for mgeode systems too, i have combined the mgeode case with the feature bit check. Signed-off-by: Jeff Hansen <jhansen@cardaccess-inc.com> Signed-off-by: Alok N Kataria <akataria@vmware.com> Signed-off-by: Dan Hecht <dhecht@vmware.com> Signed-off-by: H. Peter Anvin <hpa@zytor.com>
2008-10-25 04:22:01 +04:00
return 1;
}
__setup("tsc=", tsc_setup);
#define MAX_RETRIES 5
#define SMI_TRESHOLD 50000
/*
* Read TSC and the reference counters. Take care of SMI disturbance
*/
static u64 tsc_read_refs(u64 *p, int hpet)
{
u64 t1, t2;
int i;
for (i = 0; i < MAX_RETRIES; i++) {
t1 = get_cycles();
if (hpet)
*p = hpet_readl(HPET_COUNTER) & 0xFFFFFFFF;
else
*p = acpi_pm_read_early();
t2 = get_cycles();
if ((t2 - t1) < SMI_TRESHOLD)
return t2;
}
return ULLONG_MAX;
}
/*
* Calculate the TSC frequency from HPET reference
*/
static unsigned long calc_hpet_ref(u64 deltatsc, u64 hpet1, u64 hpet2)
{
u64 tmp;
if (hpet2 < hpet1)
hpet2 += 0x100000000ULL;
hpet2 -= hpet1;
tmp = ((u64)hpet2 * hpet_readl(HPET_PERIOD));
do_div(tmp, 1000000);
do_div(deltatsc, tmp);
return (unsigned long) deltatsc;
}
/*
* Calculate the TSC frequency from PMTimer reference
*/
static unsigned long calc_pmtimer_ref(u64 deltatsc, u64 pm1, u64 pm2)
{
u64 tmp;
if (!pm1 && !pm2)
return ULONG_MAX;
if (pm2 < pm1)
pm2 += (u64)ACPI_PM_OVRRUN;
pm2 -= pm1;
tmp = pm2 * 1000000000LL;
do_div(tmp, PMTMR_TICKS_PER_SEC);
do_div(deltatsc, tmp);
return (unsigned long) deltatsc;
}
#define CAL_MS 10
#define CAL_LATCH (PIT_TICK_RATE / (1000 / CAL_MS))
#define CAL_PIT_LOOPS 1000
#define CAL2_MS 50
#define CAL2_LATCH (PIT_TICK_RATE / (1000 / CAL2_MS))
#define CAL2_PIT_LOOPS 5000
/*
* Try to calibrate the TSC against the Programmable
* Interrupt Timer and return the frequency of the TSC
* in kHz.
*
* Return ULONG_MAX on failure to calibrate.
*/
static unsigned long pit_calibrate_tsc(u32 latch, unsigned long ms, int loopmin)
{
u64 tsc, t1, t2, delta;
unsigned long tscmin, tscmax;
int pitcnt;
/* Set the Gate high, disable speaker */
outb((inb(0x61) & ~0x02) | 0x01, 0x61);
/*
* Setup CTC channel 2* for mode 0, (interrupt on terminal
* count mode), binary count. Set the latch register to 50ms
* (LSB then MSB) to begin countdown.
*/
outb(0xb0, 0x43);
outb(latch & 0xff, 0x42);
outb(latch >> 8, 0x42);
tsc = t1 = t2 = get_cycles();
pitcnt = 0;
tscmax = 0;
tscmin = ULONG_MAX;
while ((inb(0x61) & 0x20) == 0) {
t2 = get_cycles();
delta = t2 - tsc;
tsc = t2;
if ((unsigned long) delta < tscmin)
tscmin = (unsigned int) delta;
if ((unsigned long) delta > tscmax)
tscmax = (unsigned int) delta;
pitcnt++;
}
/*
* Sanity checks:
*
* If we were not able to read the PIT more than loopmin
* times, then we have been hit by a massive SMI
*
* If the maximum is 10 times larger than the minimum,
* then we got hit by an SMI as well.
*/
if (pitcnt < loopmin || tscmax > 10 * tscmin)
return ULONG_MAX;
/* Calculate the PIT value */
delta = t2 - t1;
do_div(delta, ms);
return delta;
}
x86: quick TSC calibration Introduce a fast TSC-calibration method on sane hardware. It only uses 17920 PIT timer ticks to calibrate the TSC, plus 256 ticks on each side to make sure the TSC values were very close to the tick, so the whole calibration takes 15ms. Yet, despite only takign 15ms, we can actually give pretty stringent guarantees of accuracy: - the code requires that we hit each 256-counter block at least 50 times, so the TSC error is basically at *MOST* just a few PIT cycles off in any direction. In practice, it's going to be about one microseconds off (which is how long it takes to read the counter) - so over 17920 PIT cycles, we can pretty much guarantee that the calibration error is less than one half of a percent. My testing bears this out: on my machine, the quick-calibration reports 2934.085kHz, while the slow one reports 2933.415. Yes, the slower calibration is still more precise. For me, the slow calibration is stable to within about one hundreth of a percent, so it's (at a guess) roughly an order-and-a-half of magnitude more precise. The longer you wait, the more precise you can be. However, the nice thing about the fast TSC PIT synchronization is that it's pretty much _guaranteed_ to give that 0.5% precision, and fail gracefully (and very quickly) if it doesn't get it. And it really is fairly simple (even if there's a lot of _details_ there, and I didn't get all of those right ont he first try or even the second ;) The patch says "110 insertions", but 63 of those new lines are actually comments. Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Ingo Molnar <mingo@elte.hu> --- arch/x86/kernel/tsc.c | 111 ++++++++++++++++++++++++++++++++++++++++++++++++- 1 files changed, 110 insertions(+), 1 deletions(-)
2008-09-04 21:41:22 +04:00
/*
* This reads the current MSB of the PIT counter, and
* checks if we are running on sufficiently fast and
* non-virtualized hardware.
*
* Our expectations are:
*
* - the PIT is running at roughly 1.19MHz
*
* - each IO is going to take about 1us on real hardware,
* but we allow it to be much faster (by a factor of 10) or
* _slightly_ slower (ie we allow up to a 2us read+counter
* update - anything else implies a unacceptably slow CPU
* or PIT for the fast calibration to work.
*
* - with 256 PIT ticks to read the value, we have 214us to
* see the same MSB (and overhead like doing a single TSC
* read per MSB value etc).
*
* - We're doing 2 reads per loop (LSB, MSB), and we expect
* them each to take about a microsecond on real hardware.
* So we expect a count value of around 100. But we'll be
* generous, and accept anything over 50.
*
* - if the PIT is stuck, and we see *many* more reads, we
* return early (and the next caller of pit_expect_msb()
* then consider it a failure when they don't see the
* next expected value).
*
* These expectations mean that we know that we have seen the
* transition from one expected value to another with a fairly
* high accuracy, and we didn't miss any events. We can thus
* use the TSC value at the transitions to calculate a pretty
* good value for the TSC frequencty.
*/
2009-07-31 23:45:41 +04:00
static inline int pit_verify_msb(unsigned char val)
{
/* Ignore LSB */
inb(0x42);
return inb(0x42) == val;
}
Fast TSC calibration: calculate proper frequency error bounds In order for ntpd to correctly synchronize the clocks, the frequency of the system clock must not be off by more than 500 ppm (or, put another way, 1:2000), or ntpd will end up giving up on trying to synchronize properly, and ends up reseting the clock in jumps instead. The fast TSC PIT calibration sometimes failed this test - it was assuming that the PIT reads always took about one microsecond each (2us for the two reads to get a 16-bit timer), and that calibrating TSC to the PIT over 15ms should thus be sufficient to get much closer than 500ppm (max 2us error on both sides giving 4us over 15ms: a 270 ppm error value). However, that assumption does not always hold: apparently some hardware is either very much slower at reading the PIT registers, or there was other noise causing at least one machine to get 700+ ppm errors. So instead of using a fixed 15ms timing loop, this changes the fast PIT calibration to read the TSC delta over the individual PIT timer reads, and use the result to calculate the error bars on the PIT read timing properly. We then successfully calibrate the TSC only if the maximum error bars fall below 500ppm. In the process, we also relax the timing to allow up to 25ms for the calibration, although it can happen much faster depending on hardware. Reported-and-tested-by: Jesper Krogh <jesper@krogh.cc> Cc: john stultz <johnstul@us.ibm.com> Cc: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-03-17 18:13:17 +03:00
static inline int pit_expect_msb(unsigned char val, u64 *tscp, unsigned long *deltap)
x86: quick TSC calibration Introduce a fast TSC-calibration method on sane hardware. It only uses 17920 PIT timer ticks to calibrate the TSC, plus 256 ticks on each side to make sure the TSC values were very close to the tick, so the whole calibration takes 15ms. Yet, despite only takign 15ms, we can actually give pretty stringent guarantees of accuracy: - the code requires that we hit each 256-counter block at least 50 times, so the TSC error is basically at *MOST* just a few PIT cycles off in any direction. In practice, it's going to be about one microseconds off (which is how long it takes to read the counter) - so over 17920 PIT cycles, we can pretty much guarantee that the calibration error is less than one half of a percent. My testing bears this out: on my machine, the quick-calibration reports 2934.085kHz, while the slow one reports 2933.415. Yes, the slower calibration is still more precise. For me, the slow calibration is stable to within about one hundreth of a percent, so it's (at a guess) roughly an order-and-a-half of magnitude more precise. The longer you wait, the more precise you can be. However, the nice thing about the fast TSC PIT synchronization is that it's pretty much _guaranteed_ to give that 0.5% precision, and fail gracefully (and very quickly) if it doesn't get it. And it really is fairly simple (even if there's a lot of _details_ there, and I didn't get all of those right ont he first try or even the second ;) The patch says "110 insertions", but 63 of those new lines are actually comments. Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Ingo Molnar <mingo@elte.hu> --- arch/x86/kernel/tsc.c | 111 ++++++++++++++++++++++++++++++++++++++++++++++++- 1 files changed, 110 insertions(+), 1 deletions(-)
2008-09-04 21:41:22 +04:00
{
Fast TSC calibration: calculate proper frequency error bounds In order for ntpd to correctly synchronize the clocks, the frequency of the system clock must not be off by more than 500 ppm (or, put another way, 1:2000), or ntpd will end up giving up on trying to synchronize properly, and ends up reseting the clock in jumps instead. The fast TSC PIT calibration sometimes failed this test - it was assuming that the PIT reads always took about one microsecond each (2us for the two reads to get a 16-bit timer), and that calibrating TSC to the PIT over 15ms should thus be sufficient to get much closer than 500ppm (max 2us error on both sides giving 4us over 15ms: a 270 ppm error value). However, that assumption does not always hold: apparently some hardware is either very much slower at reading the PIT registers, or there was other noise causing at least one machine to get 700+ ppm errors. So instead of using a fixed 15ms timing loop, this changes the fast PIT calibration to read the TSC delta over the individual PIT timer reads, and use the result to calculate the error bars on the PIT read timing properly. We then successfully calibrate the TSC only if the maximum error bars fall below 500ppm. In the process, we also relax the timing to allow up to 25ms for the calibration, although it can happen much faster depending on hardware. Reported-and-tested-by: Jesper Krogh <jesper@krogh.cc> Cc: john stultz <johnstul@us.ibm.com> Cc: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-03-17 18:13:17 +03:00
int count;
x86, tsc: Fix SMI induced variation in quick_pit_calibrate() pit_expect_msb() returns success wrongly in the below SMI scenario: a. pit_verify_msb() has not yet seen the MSB transition. b. we are close to the MSB transition though and got a SMI immediately after returning from pit_verify_msb() which didn't see the MSB transition. PIT MSB transition has happened somewhere during SMI execution. c. returned from SMI and we noted down the 'tsc', saw the pit MSB change now and exited the loop to calculate 'deltatsc'. Instead of noting the TSC at the MSB transition, we are way off because of the SMI. And as the SMI happened between the pit_verify_msb() and before the 'tsc' is recorded in the for loop, 'delattsc' (d1/d2 in quick_pit_calibrate()) will be small and quick_pit_calibrate() will not notice this error. Depending on whether SMI disturbance happens while computing d1 or d2, we will see the TSC calibrated value smaller or bigger than the expected value. As a result, in a cluster we were seeing a variation of approximately +/- 20MHz in the calibrated values, resulting in NTP failures. [ As far as the SMI source is concerned, this is a periodic SMI that gets disabled after ACPI is enabled by the OS. But the TSC calibration happens before the ACPI is enabled. ] To address this, change pit_expect_msb() so that - the 'tsc' is the TSC in between the two reads that read the MSB change from the PIT (same as before) - the 'delta' is the difference in TSC from *before* the MSB changed to *after* the MSB changed. Now the delta is twice as big as before (it covers four PIT accesses, roughly 4us) and quick_pit_calibrate() will loop a bit longer to get the calibrated value with in the 500ppm precision. As the delta (d1/d2) covers four PIT accesses, actual calibrated result might be closer to 250ppm precision. As the loop now takes longer to stabilize, double MAX_QUICK_PIT_MS to 50. SMI disturbance will showup as much larger delta's and the loop will take longer than usual for the result to be with in the accepted precision. Or will fallback to slow PIT calibration if it takes more than 50msec. Also while we are at this, remove the calibration correction that aims to get the result to the middle of the error bars. We really don't know which direction to correct into, so remove it. Reported-and-tested-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Link: http://lkml.kernel.org/r/1326843337.5291.4.camel@sbsiddha-mobl2 Signed-off-by: H. Peter Anvin <hpa@zytor.com>
2012-01-18 03:35:37 +04:00
u64 tsc = 0, prev_tsc = 0;
x86: quick TSC calibration Introduce a fast TSC-calibration method on sane hardware. It only uses 17920 PIT timer ticks to calibrate the TSC, plus 256 ticks on each side to make sure the TSC values were very close to the tick, so the whole calibration takes 15ms. Yet, despite only takign 15ms, we can actually give pretty stringent guarantees of accuracy: - the code requires that we hit each 256-counter block at least 50 times, so the TSC error is basically at *MOST* just a few PIT cycles off in any direction. In practice, it's going to be about one microseconds off (which is how long it takes to read the counter) - so over 17920 PIT cycles, we can pretty much guarantee that the calibration error is less than one half of a percent. My testing bears this out: on my machine, the quick-calibration reports 2934.085kHz, while the slow one reports 2933.415. Yes, the slower calibration is still more precise. For me, the slow calibration is stable to within about one hundreth of a percent, so it's (at a guess) roughly an order-and-a-half of magnitude more precise. The longer you wait, the more precise you can be. However, the nice thing about the fast TSC PIT synchronization is that it's pretty much _guaranteed_ to give that 0.5% precision, and fail gracefully (and very quickly) if it doesn't get it. And it really is fairly simple (even if there's a lot of _details_ there, and I didn't get all of those right ont he first try or even the second ;) The patch says "110 insertions", but 63 of those new lines are actually comments. Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Ingo Molnar <mingo@elte.hu> --- arch/x86/kernel/tsc.c | 111 ++++++++++++++++++++++++++++++++++++++++++++++++- 1 files changed, 110 insertions(+), 1 deletions(-)
2008-09-04 21:41:22 +04:00
for (count = 0; count < 50000; count++) {
2009-07-31 23:45:41 +04:00
if (!pit_verify_msb(val))
x86: quick TSC calibration Introduce a fast TSC-calibration method on sane hardware. It only uses 17920 PIT timer ticks to calibrate the TSC, plus 256 ticks on each side to make sure the TSC values were very close to the tick, so the whole calibration takes 15ms. Yet, despite only takign 15ms, we can actually give pretty stringent guarantees of accuracy: - the code requires that we hit each 256-counter block at least 50 times, so the TSC error is basically at *MOST* just a few PIT cycles off in any direction. In practice, it's going to be about one microseconds off (which is how long it takes to read the counter) - so over 17920 PIT cycles, we can pretty much guarantee that the calibration error is less than one half of a percent. My testing bears this out: on my machine, the quick-calibration reports 2934.085kHz, while the slow one reports 2933.415. Yes, the slower calibration is still more precise. For me, the slow calibration is stable to within about one hundreth of a percent, so it's (at a guess) roughly an order-and-a-half of magnitude more precise. The longer you wait, the more precise you can be. However, the nice thing about the fast TSC PIT synchronization is that it's pretty much _guaranteed_ to give that 0.5% precision, and fail gracefully (and very quickly) if it doesn't get it. And it really is fairly simple (even if there's a lot of _details_ there, and I didn't get all of those right ont he first try or even the second ;) The patch says "110 insertions", but 63 of those new lines are actually comments. Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Ingo Molnar <mingo@elte.hu> --- arch/x86/kernel/tsc.c | 111 ++++++++++++++++++++++++++++++++++++++++++++++++- 1 files changed, 110 insertions(+), 1 deletions(-)
2008-09-04 21:41:22 +04:00
break;
x86, tsc: Fix SMI induced variation in quick_pit_calibrate() pit_expect_msb() returns success wrongly in the below SMI scenario: a. pit_verify_msb() has not yet seen the MSB transition. b. we are close to the MSB transition though and got a SMI immediately after returning from pit_verify_msb() which didn't see the MSB transition. PIT MSB transition has happened somewhere during SMI execution. c. returned from SMI and we noted down the 'tsc', saw the pit MSB change now and exited the loop to calculate 'deltatsc'. Instead of noting the TSC at the MSB transition, we are way off because of the SMI. And as the SMI happened between the pit_verify_msb() and before the 'tsc' is recorded in the for loop, 'delattsc' (d1/d2 in quick_pit_calibrate()) will be small and quick_pit_calibrate() will not notice this error. Depending on whether SMI disturbance happens while computing d1 or d2, we will see the TSC calibrated value smaller or bigger than the expected value. As a result, in a cluster we were seeing a variation of approximately +/- 20MHz in the calibrated values, resulting in NTP failures. [ As far as the SMI source is concerned, this is a periodic SMI that gets disabled after ACPI is enabled by the OS. But the TSC calibration happens before the ACPI is enabled. ] To address this, change pit_expect_msb() so that - the 'tsc' is the TSC in between the two reads that read the MSB change from the PIT (same as before) - the 'delta' is the difference in TSC from *before* the MSB changed to *after* the MSB changed. Now the delta is twice as big as before (it covers four PIT accesses, roughly 4us) and quick_pit_calibrate() will loop a bit longer to get the calibrated value with in the 500ppm precision. As the delta (d1/d2) covers four PIT accesses, actual calibrated result might be closer to 250ppm precision. As the loop now takes longer to stabilize, double MAX_QUICK_PIT_MS to 50. SMI disturbance will showup as much larger delta's and the loop will take longer than usual for the result to be with in the accepted precision. Or will fallback to slow PIT calibration if it takes more than 50msec. Also while we are at this, remove the calibration correction that aims to get the result to the middle of the error bars. We really don't know which direction to correct into, so remove it. Reported-and-tested-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Link: http://lkml.kernel.org/r/1326843337.5291.4.camel@sbsiddha-mobl2 Signed-off-by: H. Peter Anvin <hpa@zytor.com>
2012-01-18 03:35:37 +04:00
prev_tsc = tsc;
Fast TSC calibration: calculate proper frequency error bounds In order for ntpd to correctly synchronize the clocks, the frequency of the system clock must not be off by more than 500 ppm (or, put another way, 1:2000), or ntpd will end up giving up on trying to synchronize properly, and ends up reseting the clock in jumps instead. The fast TSC PIT calibration sometimes failed this test - it was assuming that the PIT reads always took about one microsecond each (2us for the two reads to get a 16-bit timer), and that calibrating TSC to the PIT over 15ms should thus be sufficient to get much closer than 500ppm (max 2us error on both sides giving 4us over 15ms: a 270 ppm error value). However, that assumption does not always hold: apparently some hardware is either very much slower at reading the PIT registers, or there was other noise causing at least one machine to get 700+ ppm errors. So instead of using a fixed 15ms timing loop, this changes the fast PIT calibration to read the TSC delta over the individual PIT timer reads, and use the result to calculate the error bars on the PIT read timing properly. We then successfully calibrate the TSC only if the maximum error bars fall below 500ppm. In the process, we also relax the timing to allow up to 25ms for the calibration, although it can happen much faster depending on hardware. Reported-and-tested-by: Jesper Krogh <jesper@krogh.cc> Cc: john stultz <johnstul@us.ibm.com> Cc: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-03-17 18:13:17 +03:00
tsc = get_cycles();
x86: quick TSC calibration Introduce a fast TSC-calibration method on sane hardware. It only uses 17920 PIT timer ticks to calibrate the TSC, plus 256 ticks on each side to make sure the TSC values were very close to the tick, so the whole calibration takes 15ms. Yet, despite only takign 15ms, we can actually give pretty stringent guarantees of accuracy: - the code requires that we hit each 256-counter block at least 50 times, so the TSC error is basically at *MOST* just a few PIT cycles off in any direction. In practice, it's going to be about one microseconds off (which is how long it takes to read the counter) - so over 17920 PIT cycles, we can pretty much guarantee that the calibration error is less than one half of a percent. My testing bears this out: on my machine, the quick-calibration reports 2934.085kHz, while the slow one reports 2933.415. Yes, the slower calibration is still more precise. For me, the slow calibration is stable to within about one hundreth of a percent, so it's (at a guess) roughly an order-and-a-half of magnitude more precise. The longer you wait, the more precise you can be. However, the nice thing about the fast TSC PIT synchronization is that it's pretty much _guaranteed_ to give that 0.5% precision, and fail gracefully (and very quickly) if it doesn't get it. And it really is fairly simple (even if there's a lot of _details_ there, and I didn't get all of those right ont he first try or even the second ;) The patch says "110 insertions", but 63 of those new lines are actually comments. Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Ingo Molnar <mingo@elte.hu> --- arch/x86/kernel/tsc.c | 111 ++++++++++++++++++++++++++++++++++++++++++++++++- 1 files changed, 110 insertions(+), 1 deletions(-)
2008-09-04 21:41:22 +04:00
}
x86, tsc: Fix SMI induced variation in quick_pit_calibrate() pit_expect_msb() returns success wrongly in the below SMI scenario: a. pit_verify_msb() has not yet seen the MSB transition. b. we are close to the MSB transition though and got a SMI immediately after returning from pit_verify_msb() which didn't see the MSB transition. PIT MSB transition has happened somewhere during SMI execution. c. returned from SMI and we noted down the 'tsc', saw the pit MSB change now and exited the loop to calculate 'deltatsc'. Instead of noting the TSC at the MSB transition, we are way off because of the SMI. And as the SMI happened between the pit_verify_msb() and before the 'tsc' is recorded in the for loop, 'delattsc' (d1/d2 in quick_pit_calibrate()) will be small and quick_pit_calibrate() will not notice this error. Depending on whether SMI disturbance happens while computing d1 or d2, we will see the TSC calibrated value smaller or bigger than the expected value. As a result, in a cluster we were seeing a variation of approximately +/- 20MHz in the calibrated values, resulting in NTP failures. [ As far as the SMI source is concerned, this is a periodic SMI that gets disabled after ACPI is enabled by the OS. But the TSC calibration happens before the ACPI is enabled. ] To address this, change pit_expect_msb() so that - the 'tsc' is the TSC in between the two reads that read the MSB change from the PIT (same as before) - the 'delta' is the difference in TSC from *before* the MSB changed to *after* the MSB changed. Now the delta is twice as big as before (it covers four PIT accesses, roughly 4us) and quick_pit_calibrate() will loop a bit longer to get the calibrated value with in the 500ppm precision. As the delta (d1/d2) covers four PIT accesses, actual calibrated result might be closer to 250ppm precision. As the loop now takes longer to stabilize, double MAX_QUICK_PIT_MS to 50. SMI disturbance will showup as much larger delta's and the loop will take longer than usual for the result to be with in the accepted precision. Or will fallback to slow PIT calibration if it takes more than 50msec. Also while we are at this, remove the calibration correction that aims to get the result to the middle of the error bars. We really don't know which direction to correct into, so remove it. Reported-and-tested-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Link: http://lkml.kernel.org/r/1326843337.5291.4.camel@sbsiddha-mobl2 Signed-off-by: H. Peter Anvin <hpa@zytor.com>
2012-01-18 03:35:37 +04:00
*deltap = get_cycles() - prev_tsc;
Fast TSC calibration: calculate proper frequency error bounds In order for ntpd to correctly synchronize the clocks, the frequency of the system clock must not be off by more than 500 ppm (or, put another way, 1:2000), or ntpd will end up giving up on trying to synchronize properly, and ends up reseting the clock in jumps instead. The fast TSC PIT calibration sometimes failed this test - it was assuming that the PIT reads always took about one microsecond each (2us for the two reads to get a 16-bit timer), and that calibrating TSC to the PIT over 15ms should thus be sufficient to get much closer than 500ppm (max 2us error on both sides giving 4us over 15ms: a 270 ppm error value). However, that assumption does not always hold: apparently some hardware is either very much slower at reading the PIT registers, or there was other noise causing at least one machine to get 700+ ppm errors. So instead of using a fixed 15ms timing loop, this changes the fast PIT calibration to read the TSC delta over the individual PIT timer reads, and use the result to calculate the error bars on the PIT read timing properly. We then successfully calibrate the TSC only if the maximum error bars fall below 500ppm. In the process, we also relax the timing to allow up to 25ms for the calibration, although it can happen much faster depending on hardware. Reported-and-tested-by: Jesper Krogh <jesper@krogh.cc> Cc: john stultz <johnstul@us.ibm.com> Cc: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-03-17 18:13:17 +03:00
*tscp = tsc;
/*
* We require _some_ success, but the quality control
* will be based on the error terms on the TSC values.
*/
return count > 5;
x86: quick TSC calibration Introduce a fast TSC-calibration method on sane hardware. It only uses 17920 PIT timer ticks to calibrate the TSC, plus 256 ticks on each side to make sure the TSC values were very close to the tick, so the whole calibration takes 15ms. Yet, despite only takign 15ms, we can actually give pretty stringent guarantees of accuracy: - the code requires that we hit each 256-counter block at least 50 times, so the TSC error is basically at *MOST* just a few PIT cycles off in any direction. In practice, it's going to be about one microseconds off (which is how long it takes to read the counter) - so over 17920 PIT cycles, we can pretty much guarantee that the calibration error is less than one half of a percent. My testing bears this out: on my machine, the quick-calibration reports 2934.085kHz, while the slow one reports 2933.415. Yes, the slower calibration is still more precise. For me, the slow calibration is stable to within about one hundreth of a percent, so it's (at a guess) roughly an order-and-a-half of magnitude more precise. The longer you wait, the more precise you can be. However, the nice thing about the fast TSC PIT synchronization is that it's pretty much _guaranteed_ to give that 0.5% precision, and fail gracefully (and very quickly) if it doesn't get it. And it really is fairly simple (even if there's a lot of _details_ there, and I didn't get all of those right ont he first try or even the second ;) The patch says "110 insertions", but 63 of those new lines are actually comments. Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Ingo Molnar <mingo@elte.hu> --- arch/x86/kernel/tsc.c | 111 ++++++++++++++++++++++++++++++++++++++++++++++++- 1 files changed, 110 insertions(+), 1 deletions(-)
2008-09-04 21:41:22 +04:00
}
/*
Fast TSC calibration: calculate proper frequency error bounds In order for ntpd to correctly synchronize the clocks, the frequency of the system clock must not be off by more than 500 ppm (or, put another way, 1:2000), or ntpd will end up giving up on trying to synchronize properly, and ends up reseting the clock in jumps instead. The fast TSC PIT calibration sometimes failed this test - it was assuming that the PIT reads always took about one microsecond each (2us for the two reads to get a 16-bit timer), and that calibrating TSC to the PIT over 15ms should thus be sufficient to get much closer than 500ppm (max 2us error on both sides giving 4us over 15ms: a 270 ppm error value). However, that assumption does not always hold: apparently some hardware is either very much slower at reading the PIT registers, or there was other noise causing at least one machine to get 700+ ppm errors. So instead of using a fixed 15ms timing loop, this changes the fast PIT calibration to read the TSC delta over the individual PIT timer reads, and use the result to calculate the error bars on the PIT read timing properly. We then successfully calibrate the TSC only if the maximum error bars fall below 500ppm. In the process, we also relax the timing to allow up to 25ms for the calibration, although it can happen much faster depending on hardware. Reported-and-tested-by: Jesper Krogh <jesper@krogh.cc> Cc: john stultz <johnstul@us.ibm.com> Cc: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-03-17 18:13:17 +03:00
* How many MSB values do we want to see? We aim for
* a maximum error rate of 500ppm (in practice the
* real error is much smaller), but refuse to spend
x86, tsc: Fix SMI induced variation in quick_pit_calibrate() pit_expect_msb() returns success wrongly in the below SMI scenario: a. pit_verify_msb() has not yet seen the MSB transition. b. we are close to the MSB transition though and got a SMI immediately after returning from pit_verify_msb() which didn't see the MSB transition. PIT MSB transition has happened somewhere during SMI execution. c. returned from SMI and we noted down the 'tsc', saw the pit MSB change now and exited the loop to calculate 'deltatsc'. Instead of noting the TSC at the MSB transition, we are way off because of the SMI. And as the SMI happened between the pit_verify_msb() and before the 'tsc' is recorded in the for loop, 'delattsc' (d1/d2 in quick_pit_calibrate()) will be small and quick_pit_calibrate() will not notice this error. Depending on whether SMI disturbance happens while computing d1 or d2, we will see the TSC calibrated value smaller or bigger than the expected value. As a result, in a cluster we were seeing a variation of approximately +/- 20MHz in the calibrated values, resulting in NTP failures. [ As far as the SMI source is concerned, this is a periodic SMI that gets disabled after ACPI is enabled by the OS. But the TSC calibration happens before the ACPI is enabled. ] To address this, change pit_expect_msb() so that - the 'tsc' is the TSC in between the two reads that read the MSB change from the PIT (same as before) - the 'delta' is the difference in TSC from *before* the MSB changed to *after* the MSB changed. Now the delta is twice as big as before (it covers four PIT accesses, roughly 4us) and quick_pit_calibrate() will loop a bit longer to get the calibrated value with in the 500ppm precision. As the delta (d1/d2) covers four PIT accesses, actual calibrated result might be closer to 250ppm precision. As the loop now takes longer to stabilize, double MAX_QUICK_PIT_MS to 50. SMI disturbance will showup as much larger delta's and the loop will take longer than usual for the result to be with in the accepted precision. Or will fallback to slow PIT calibration if it takes more than 50msec. Also while we are at this, remove the calibration correction that aims to get the result to the middle of the error bars. We really don't know which direction to correct into, so remove it. Reported-and-tested-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Link: http://lkml.kernel.org/r/1326843337.5291.4.camel@sbsiddha-mobl2 Signed-off-by: H. Peter Anvin <hpa@zytor.com>
2012-01-18 03:35:37 +04:00
* more than 50ms on it.
x86: quick TSC calibration Introduce a fast TSC-calibration method on sane hardware. It only uses 17920 PIT timer ticks to calibrate the TSC, plus 256 ticks on each side to make sure the TSC values were very close to the tick, so the whole calibration takes 15ms. Yet, despite only takign 15ms, we can actually give pretty stringent guarantees of accuracy: - the code requires that we hit each 256-counter block at least 50 times, so the TSC error is basically at *MOST* just a few PIT cycles off in any direction. In practice, it's going to be about one microseconds off (which is how long it takes to read the counter) - so over 17920 PIT cycles, we can pretty much guarantee that the calibration error is less than one half of a percent. My testing bears this out: on my machine, the quick-calibration reports 2934.085kHz, while the slow one reports 2933.415. Yes, the slower calibration is still more precise. For me, the slow calibration is stable to within about one hundreth of a percent, so it's (at a guess) roughly an order-and-a-half of magnitude more precise. The longer you wait, the more precise you can be. However, the nice thing about the fast TSC PIT synchronization is that it's pretty much _guaranteed_ to give that 0.5% precision, and fail gracefully (and very quickly) if it doesn't get it. And it really is fairly simple (even if there's a lot of _details_ there, and I didn't get all of those right ont he first try or even the second ;) The patch says "110 insertions", but 63 of those new lines are actually comments. Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Ingo Molnar <mingo@elte.hu> --- arch/x86/kernel/tsc.c | 111 ++++++++++++++++++++++++++++++++++++++++++++++++- 1 files changed, 110 insertions(+), 1 deletions(-)
2008-09-04 21:41:22 +04:00
*/
x86, tsc: Fix SMI induced variation in quick_pit_calibrate() pit_expect_msb() returns success wrongly in the below SMI scenario: a. pit_verify_msb() has not yet seen the MSB transition. b. we are close to the MSB transition though and got a SMI immediately after returning from pit_verify_msb() which didn't see the MSB transition. PIT MSB transition has happened somewhere during SMI execution. c. returned from SMI and we noted down the 'tsc', saw the pit MSB change now and exited the loop to calculate 'deltatsc'. Instead of noting the TSC at the MSB transition, we are way off because of the SMI. And as the SMI happened between the pit_verify_msb() and before the 'tsc' is recorded in the for loop, 'delattsc' (d1/d2 in quick_pit_calibrate()) will be small and quick_pit_calibrate() will not notice this error. Depending on whether SMI disturbance happens while computing d1 or d2, we will see the TSC calibrated value smaller or bigger than the expected value. As a result, in a cluster we were seeing a variation of approximately +/- 20MHz in the calibrated values, resulting in NTP failures. [ As far as the SMI source is concerned, this is a periodic SMI that gets disabled after ACPI is enabled by the OS. But the TSC calibration happens before the ACPI is enabled. ] To address this, change pit_expect_msb() so that - the 'tsc' is the TSC in between the two reads that read the MSB change from the PIT (same as before) - the 'delta' is the difference in TSC from *before* the MSB changed to *after* the MSB changed. Now the delta is twice as big as before (it covers four PIT accesses, roughly 4us) and quick_pit_calibrate() will loop a bit longer to get the calibrated value with in the 500ppm precision. As the delta (d1/d2) covers four PIT accesses, actual calibrated result might be closer to 250ppm precision. As the loop now takes longer to stabilize, double MAX_QUICK_PIT_MS to 50. SMI disturbance will showup as much larger delta's and the loop will take longer than usual for the result to be with in the accepted precision. Or will fallback to slow PIT calibration if it takes more than 50msec. Also while we are at this, remove the calibration correction that aims to get the result to the middle of the error bars. We really don't know which direction to correct into, so remove it. Reported-and-tested-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Link: http://lkml.kernel.org/r/1326843337.5291.4.camel@sbsiddha-mobl2 Signed-off-by: H. Peter Anvin <hpa@zytor.com>
2012-01-18 03:35:37 +04:00
#define MAX_QUICK_PIT_MS 50
Fast TSC calibration: calculate proper frequency error bounds In order for ntpd to correctly synchronize the clocks, the frequency of the system clock must not be off by more than 500 ppm (or, put another way, 1:2000), or ntpd will end up giving up on trying to synchronize properly, and ends up reseting the clock in jumps instead. The fast TSC PIT calibration sometimes failed this test - it was assuming that the PIT reads always took about one microsecond each (2us for the two reads to get a 16-bit timer), and that calibrating TSC to the PIT over 15ms should thus be sufficient to get much closer than 500ppm (max 2us error on both sides giving 4us over 15ms: a 270 ppm error value). However, that assumption does not always hold: apparently some hardware is either very much slower at reading the PIT registers, or there was other noise causing at least one machine to get 700+ ppm errors. So instead of using a fixed 15ms timing loop, this changes the fast PIT calibration to read the TSC delta over the individual PIT timer reads, and use the result to calculate the error bars on the PIT read timing properly. We then successfully calibrate the TSC only if the maximum error bars fall below 500ppm. In the process, we also relax the timing to allow up to 25ms for the calibration, although it can happen much faster depending on hardware. Reported-and-tested-by: Jesper Krogh <jesper@krogh.cc> Cc: john stultz <johnstul@us.ibm.com> Cc: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-03-17 18:13:17 +03:00
#define MAX_QUICK_PIT_ITERATIONS (MAX_QUICK_PIT_MS * PIT_TICK_RATE / 1000 / 256)
x86: quick TSC calibration Introduce a fast TSC-calibration method on sane hardware. It only uses 17920 PIT timer ticks to calibrate the TSC, plus 256 ticks on each side to make sure the TSC values were very close to the tick, so the whole calibration takes 15ms. Yet, despite only takign 15ms, we can actually give pretty stringent guarantees of accuracy: - the code requires that we hit each 256-counter block at least 50 times, so the TSC error is basically at *MOST* just a few PIT cycles off in any direction. In practice, it's going to be about one microseconds off (which is how long it takes to read the counter) - so over 17920 PIT cycles, we can pretty much guarantee that the calibration error is less than one half of a percent. My testing bears this out: on my machine, the quick-calibration reports 2934.085kHz, while the slow one reports 2933.415. Yes, the slower calibration is still more precise. For me, the slow calibration is stable to within about one hundreth of a percent, so it's (at a guess) roughly an order-and-a-half of magnitude more precise. The longer you wait, the more precise you can be. However, the nice thing about the fast TSC PIT synchronization is that it's pretty much _guaranteed_ to give that 0.5% precision, and fail gracefully (and very quickly) if it doesn't get it. And it really is fairly simple (even if there's a lot of _details_ there, and I didn't get all of those right ont he first try or even the second ;) The patch says "110 insertions", but 63 of those new lines are actually comments. Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Ingo Molnar <mingo@elte.hu> --- arch/x86/kernel/tsc.c | 111 ++++++++++++++++++++++++++++++++++++++++++++++++- 1 files changed, 110 insertions(+), 1 deletions(-)
2008-09-04 21:41:22 +04:00
static unsigned long quick_pit_calibrate(void)
{
Fast TSC calibration: calculate proper frequency error bounds In order for ntpd to correctly synchronize the clocks, the frequency of the system clock must not be off by more than 500 ppm (or, put another way, 1:2000), or ntpd will end up giving up on trying to synchronize properly, and ends up reseting the clock in jumps instead. The fast TSC PIT calibration sometimes failed this test - it was assuming that the PIT reads always took about one microsecond each (2us for the two reads to get a 16-bit timer), and that calibrating TSC to the PIT over 15ms should thus be sufficient to get much closer than 500ppm (max 2us error on both sides giving 4us over 15ms: a 270 ppm error value). However, that assumption does not always hold: apparently some hardware is either very much slower at reading the PIT registers, or there was other noise causing at least one machine to get 700+ ppm errors. So instead of using a fixed 15ms timing loop, this changes the fast PIT calibration to read the TSC delta over the individual PIT timer reads, and use the result to calculate the error bars on the PIT read timing properly. We then successfully calibrate the TSC only if the maximum error bars fall below 500ppm. In the process, we also relax the timing to allow up to 25ms for the calibration, although it can happen much faster depending on hardware. Reported-and-tested-by: Jesper Krogh <jesper@krogh.cc> Cc: john stultz <johnstul@us.ibm.com> Cc: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-03-17 18:13:17 +03:00
int i;
u64 tsc, delta;
unsigned long d1, d2;
x86: quick TSC calibration Introduce a fast TSC-calibration method on sane hardware. It only uses 17920 PIT timer ticks to calibrate the TSC, plus 256 ticks on each side to make sure the TSC values were very close to the tick, so the whole calibration takes 15ms. Yet, despite only takign 15ms, we can actually give pretty stringent guarantees of accuracy: - the code requires that we hit each 256-counter block at least 50 times, so the TSC error is basically at *MOST* just a few PIT cycles off in any direction. In practice, it's going to be about one microseconds off (which is how long it takes to read the counter) - so over 17920 PIT cycles, we can pretty much guarantee that the calibration error is less than one half of a percent. My testing bears this out: on my machine, the quick-calibration reports 2934.085kHz, while the slow one reports 2933.415. Yes, the slower calibration is still more precise. For me, the slow calibration is stable to within about one hundreth of a percent, so it's (at a guess) roughly an order-and-a-half of magnitude more precise. The longer you wait, the more precise you can be. However, the nice thing about the fast TSC PIT synchronization is that it's pretty much _guaranteed_ to give that 0.5% precision, and fail gracefully (and very quickly) if it doesn't get it. And it really is fairly simple (even if there's a lot of _details_ there, and I didn't get all of those right ont he first try or even the second ;) The patch says "110 insertions", but 63 of those new lines are actually comments. Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Ingo Molnar <mingo@elte.hu> --- arch/x86/kernel/tsc.c | 111 ++++++++++++++++++++++++++++++++++++++++++++++++- 1 files changed, 110 insertions(+), 1 deletions(-)
2008-09-04 21:41:22 +04:00
/* Set the Gate high, disable speaker */
outb((inb(0x61) & ~0x02) | 0x01, 0x61);
x86: quick TSC calibration Introduce a fast TSC-calibration method on sane hardware. It only uses 17920 PIT timer ticks to calibrate the TSC, plus 256 ticks on each side to make sure the TSC values were very close to the tick, so the whole calibration takes 15ms. Yet, despite only takign 15ms, we can actually give pretty stringent guarantees of accuracy: - the code requires that we hit each 256-counter block at least 50 times, so the TSC error is basically at *MOST* just a few PIT cycles off in any direction. In practice, it's going to be about one microseconds off (which is how long it takes to read the counter) - so over 17920 PIT cycles, we can pretty much guarantee that the calibration error is less than one half of a percent. My testing bears this out: on my machine, the quick-calibration reports 2934.085kHz, while the slow one reports 2933.415. Yes, the slower calibration is still more precise. For me, the slow calibration is stable to within about one hundreth of a percent, so it's (at a guess) roughly an order-and-a-half of magnitude more precise. The longer you wait, the more precise you can be. However, the nice thing about the fast TSC PIT synchronization is that it's pretty much _guaranteed_ to give that 0.5% precision, and fail gracefully (and very quickly) if it doesn't get it. And it really is fairly simple (even if there's a lot of _details_ there, and I didn't get all of those right ont he first try or even the second ;) The patch says "110 insertions", but 63 of those new lines are actually comments. Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Ingo Molnar <mingo@elte.hu> --- arch/x86/kernel/tsc.c | 111 ++++++++++++++++++++++++++++++++++++++++++++++++- 1 files changed, 110 insertions(+), 1 deletions(-)
2008-09-04 21:41:22 +04:00
/*
* Counter 2, mode 0 (one-shot), binary count
*
* NOTE! Mode 2 decrements by two (and then the
* output is flipped each time, giving the same
* final output frequency as a decrement-by-one),
* so mode 0 is much better when looking at the
* individual counts.
*/
outb(0xb0, 0x43);
x86: quick TSC calibration Introduce a fast TSC-calibration method on sane hardware. It only uses 17920 PIT timer ticks to calibrate the TSC, plus 256 ticks on each side to make sure the TSC values were very close to the tick, so the whole calibration takes 15ms. Yet, despite only takign 15ms, we can actually give pretty stringent guarantees of accuracy: - the code requires that we hit each 256-counter block at least 50 times, so the TSC error is basically at *MOST* just a few PIT cycles off in any direction. In practice, it's going to be about one microseconds off (which is how long it takes to read the counter) - so over 17920 PIT cycles, we can pretty much guarantee that the calibration error is less than one half of a percent. My testing bears this out: on my machine, the quick-calibration reports 2934.085kHz, while the slow one reports 2933.415. Yes, the slower calibration is still more precise. For me, the slow calibration is stable to within about one hundreth of a percent, so it's (at a guess) roughly an order-and-a-half of magnitude more precise. The longer you wait, the more precise you can be. However, the nice thing about the fast TSC PIT synchronization is that it's pretty much _guaranteed_ to give that 0.5% precision, and fail gracefully (and very quickly) if it doesn't get it. And it really is fairly simple (even if there's a lot of _details_ there, and I didn't get all of those right ont he first try or even the second ;) The patch says "110 insertions", but 63 of those new lines are actually comments. Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Ingo Molnar <mingo@elte.hu> --- arch/x86/kernel/tsc.c | 111 ++++++++++++++++++++++++++++++++++++++++++++++++- 1 files changed, 110 insertions(+), 1 deletions(-)
2008-09-04 21:41:22 +04:00
/* Start at 0xffff */
outb(0xff, 0x42);
outb(0xff, 0x42);
/*
* The PIT starts counting at the next edge, so we
* need to delay for a microsecond. The easiest way
* to do that is to just read back the 16-bit counter
* once from the PIT.
*/
2009-07-31 23:45:41 +04:00
pit_verify_msb(0);
Fast TSC calibration: calculate proper frequency error bounds In order for ntpd to correctly synchronize the clocks, the frequency of the system clock must not be off by more than 500 ppm (or, put another way, 1:2000), or ntpd will end up giving up on trying to synchronize properly, and ends up reseting the clock in jumps instead. The fast TSC PIT calibration sometimes failed this test - it was assuming that the PIT reads always took about one microsecond each (2us for the two reads to get a 16-bit timer), and that calibrating TSC to the PIT over 15ms should thus be sufficient to get much closer than 500ppm (max 2us error on both sides giving 4us over 15ms: a 270 ppm error value). However, that assumption does not always hold: apparently some hardware is either very much slower at reading the PIT registers, or there was other noise causing at least one machine to get 700+ ppm errors. So instead of using a fixed 15ms timing loop, this changes the fast PIT calibration to read the TSC delta over the individual PIT timer reads, and use the result to calculate the error bars on the PIT read timing properly. We then successfully calibrate the TSC only if the maximum error bars fall below 500ppm. In the process, we also relax the timing to allow up to 25ms for the calibration, although it can happen much faster depending on hardware. Reported-and-tested-by: Jesper Krogh <jesper@krogh.cc> Cc: john stultz <johnstul@us.ibm.com> Cc: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-03-17 18:13:17 +03:00
if (pit_expect_msb(0xff, &tsc, &d1)) {
for (i = 1; i <= MAX_QUICK_PIT_ITERATIONS; i++) {
if (!pit_expect_msb(0xff-i, &delta, &d2))
break;
/*
* Iterate until the error is less than 500 ppm
*/
delta -= tsc;
2009-07-31 23:45:41 +04:00
if (d1+d2 >= delta >> 11)
continue;
/*
* Check the PIT one more time to verify that
* all TSC reads were stable wrt the PIT.
*
* This also guarantees serialization of the
* last cycle read ('d2') in pit_expect_msb.
*/
if (!pit_verify_msb(0xfe - i))
break;
goto success;
x86: quick TSC calibration Introduce a fast TSC-calibration method on sane hardware. It only uses 17920 PIT timer ticks to calibrate the TSC, plus 256 ticks on each side to make sure the TSC values were very close to the tick, so the whole calibration takes 15ms. Yet, despite only takign 15ms, we can actually give pretty stringent guarantees of accuracy: - the code requires that we hit each 256-counter block at least 50 times, so the TSC error is basically at *MOST* just a few PIT cycles off in any direction. In practice, it's going to be about one microseconds off (which is how long it takes to read the counter) - so over 17920 PIT cycles, we can pretty much guarantee that the calibration error is less than one half of a percent. My testing bears this out: on my machine, the quick-calibration reports 2934.085kHz, while the slow one reports 2933.415. Yes, the slower calibration is still more precise. For me, the slow calibration is stable to within about one hundreth of a percent, so it's (at a guess) roughly an order-and-a-half of magnitude more precise. The longer you wait, the more precise you can be. However, the nice thing about the fast TSC PIT synchronization is that it's pretty much _guaranteed_ to give that 0.5% precision, and fail gracefully (and very quickly) if it doesn't get it. And it really is fairly simple (even if there's a lot of _details_ there, and I didn't get all of those right ont he first try or even the second ;) The patch says "110 insertions", but 63 of those new lines are actually comments. Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Ingo Molnar <mingo@elte.hu> --- arch/x86/kernel/tsc.c | 111 ++++++++++++++++++++++++++++++++++++++++++++++++- 1 files changed, 110 insertions(+), 1 deletions(-)
2008-09-04 21:41:22 +04:00
}
}
Fast TSC calibration: calculate proper frequency error bounds In order for ntpd to correctly synchronize the clocks, the frequency of the system clock must not be off by more than 500 ppm (or, put another way, 1:2000), or ntpd will end up giving up on trying to synchronize properly, and ends up reseting the clock in jumps instead. The fast TSC PIT calibration sometimes failed this test - it was assuming that the PIT reads always took about one microsecond each (2us for the two reads to get a 16-bit timer), and that calibrating TSC to the PIT over 15ms should thus be sufficient to get much closer than 500ppm (max 2us error on both sides giving 4us over 15ms: a 270 ppm error value). However, that assumption does not always hold: apparently some hardware is either very much slower at reading the PIT registers, or there was other noise causing at least one machine to get 700+ ppm errors. So instead of using a fixed 15ms timing loop, this changes the fast PIT calibration to read the TSC delta over the individual PIT timer reads, and use the result to calculate the error bars on the PIT read timing properly. We then successfully calibrate the TSC only if the maximum error bars fall below 500ppm. In the process, we also relax the timing to allow up to 25ms for the calibration, although it can happen much faster depending on hardware. Reported-and-tested-by: Jesper Krogh <jesper@krogh.cc> Cc: john stultz <johnstul@us.ibm.com> Cc: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-03-17 18:13:17 +03:00
printk("Fast TSC calibration failed\n");
x86: quick TSC calibration Introduce a fast TSC-calibration method on sane hardware. It only uses 17920 PIT timer ticks to calibrate the TSC, plus 256 ticks on each side to make sure the TSC values were very close to the tick, so the whole calibration takes 15ms. Yet, despite only takign 15ms, we can actually give pretty stringent guarantees of accuracy: - the code requires that we hit each 256-counter block at least 50 times, so the TSC error is basically at *MOST* just a few PIT cycles off in any direction. In practice, it's going to be about one microseconds off (which is how long it takes to read the counter) - so over 17920 PIT cycles, we can pretty much guarantee that the calibration error is less than one half of a percent. My testing bears this out: on my machine, the quick-calibration reports 2934.085kHz, while the slow one reports 2933.415. Yes, the slower calibration is still more precise. For me, the slow calibration is stable to within about one hundreth of a percent, so it's (at a guess) roughly an order-and-a-half of magnitude more precise. The longer you wait, the more precise you can be. However, the nice thing about the fast TSC PIT synchronization is that it's pretty much _guaranteed_ to give that 0.5% precision, and fail gracefully (and very quickly) if it doesn't get it. And it really is fairly simple (even if there's a lot of _details_ there, and I didn't get all of those right ont he first try or even the second ;) The patch says "110 insertions", but 63 of those new lines are actually comments. Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Ingo Molnar <mingo@elte.hu> --- arch/x86/kernel/tsc.c | 111 ++++++++++++++++++++++++++++++++++++++++++++++++- 1 files changed, 110 insertions(+), 1 deletions(-)
2008-09-04 21:41:22 +04:00
return 0;
Fast TSC calibration: calculate proper frequency error bounds In order for ntpd to correctly synchronize the clocks, the frequency of the system clock must not be off by more than 500 ppm (or, put another way, 1:2000), or ntpd will end up giving up on trying to synchronize properly, and ends up reseting the clock in jumps instead. The fast TSC PIT calibration sometimes failed this test - it was assuming that the PIT reads always took about one microsecond each (2us for the two reads to get a 16-bit timer), and that calibrating TSC to the PIT over 15ms should thus be sufficient to get much closer than 500ppm (max 2us error on both sides giving 4us over 15ms: a 270 ppm error value). However, that assumption does not always hold: apparently some hardware is either very much slower at reading the PIT registers, or there was other noise causing at least one machine to get 700+ ppm errors. So instead of using a fixed 15ms timing loop, this changes the fast PIT calibration to read the TSC delta over the individual PIT timer reads, and use the result to calculate the error bars on the PIT read timing properly. We then successfully calibrate the TSC only if the maximum error bars fall below 500ppm. In the process, we also relax the timing to allow up to 25ms for the calibration, although it can happen much faster depending on hardware. Reported-and-tested-by: Jesper Krogh <jesper@krogh.cc> Cc: john stultz <johnstul@us.ibm.com> Cc: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-03-17 18:13:17 +03:00
success:
/*
* Ok, if we get here, then we've seen the
* MSB of the PIT decrement 'i' times, and the
* error has shrunk to less than 500 ppm.
*
* As a result, we can depend on there not being
* any odd delays anywhere, and the TSC reads are
x86, tsc: Fix SMI induced variation in quick_pit_calibrate() pit_expect_msb() returns success wrongly in the below SMI scenario: a. pit_verify_msb() has not yet seen the MSB transition. b. we are close to the MSB transition though and got a SMI immediately after returning from pit_verify_msb() which didn't see the MSB transition. PIT MSB transition has happened somewhere during SMI execution. c. returned from SMI and we noted down the 'tsc', saw the pit MSB change now and exited the loop to calculate 'deltatsc'. Instead of noting the TSC at the MSB transition, we are way off because of the SMI. And as the SMI happened between the pit_verify_msb() and before the 'tsc' is recorded in the for loop, 'delattsc' (d1/d2 in quick_pit_calibrate()) will be small and quick_pit_calibrate() will not notice this error. Depending on whether SMI disturbance happens while computing d1 or d2, we will see the TSC calibrated value smaller or bigger than the expected value. As a result, in a cluster we were seeing a variation of approximately +/- 20MHz in the calibrated values, resulting in NTP failures. [ As far as the SMI source is concerned, this is a periodic SMI that gets disabled after ACPI is enabled by the OS. But the TSC calibration happens before the ACPI is enabled. ] To address this, change pit_expect_msb() so that - the 'tsc' is the TSC in between the two reads that read the MSB change from the PIT (same as before) - the 'delta' is the difference in TSC from *before* the MSB changed to *after* the MSB changed. Now the delta is twice as big as before (it covers four PIT accesses, roughly 4us) and quick_pit_calibrate() will loop a bit longer to get the calibrated value with in the 500ppm precision. As the delta (d1/d2) covers four PIT accesses, actual calibrated result might be closer to 250ppm precision. As the loop now takes longer to stabilize, double MAX_QUICK_PIT_MS to 50. SMI disturbance will showup as much larger delta's and the loop will take longer than usual for the result to be with in the accepted precision. Or will fallback to slow PIT calibration if it takes more than 50msec. Also while we are at this, remove the calibration correction that aims to get the result to the middle of the error bars. We really don't know which direction to correct into, so remove it. Reported-and-tested-by: Suresh Siddha <suresh.b.siddha@intel.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com> Link: http://lkml.kernel.org/r/1326843337.5291.4.camel@sbsiddha-mobl2 Signed-off-by: H. Peter Anvin <hpa@zytor.com>
2012-01-18 03:35:37 +04:00
* reliable (within the error).
Fast TSC calibration: calculate proper frequency error bounds In order for ntpd to correctly synchronize the clocks, the frequency of the system clock must not be off by more than 500 ppm (or, put another way, 1:2000), or ntpd will end up giving up on trying to synchronize properly, and ends up reseting the clock in jumps instead. The fast TSC PIT calibration sometimes failed this test - it was assuming that the PIT reads always took about one microsecond each (2us for the two reads to get a 16-bit timer), and that calibrating TSC to the PIT over 15ms should thus be sufficient to get much closer than 500ppm (max 2us error on both sides giving 4us over 15ms: a 270 ppm error value). However, that assumption does not always hold: apparently some hardware is either very much slower at reading the PIT registers, or there was other noise causing at least one machine to get 700+ ppm errors. So instead of using a fixed 15ms timing loop, this changes the fast PIT calibration to read the TSC delta over the individual PIT timer reads, and use the result to calculate the error bars on the PIT read timing properly. We then successfully calibrate the TSC only if the maximum error bars fall below 500ppm. In the process, we also relax the timing to allow up to 25ms for the calibration, although it can happen much faster depending on hardware. Reported-and-tested-by: Jesper Krogh <jesper@krogh.cc> Cc: john stultz <johnstul@us.ibm.com> Cc: Thomas Gleixner <tglx@linutronix.de> Acked-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-03-17 18:13:17 +03:00
*
* kHz = ticks / time-in-seconds / 1000;
* kHz = (t2 - t1) / (I * 256 / PIT_TICK_RATE) / 1000
* kHz = ((t2 - t1) * PIT_TICK_RATE) / (I * 256 * 1000)
*/
delta *= PIT_TICK_RATE;
do_div(delta, i*256*1000);
printk("Fast TSC calibration using PIT\n");
return delta;
x86: quick TSC calibration Introduce a fast TSC-calibration method on sane hardware. It only uses 17920 PIT timer ticks to calibrate the TSC, plus 256 ticks on each side to make sure the TSC values were very close to the tick, so the whole calibration takes 15ms. Yet, despite only takign 15ms, we can actually give pretty stringent guarantees of accuracy: - the code requires that we hit each 256-counter block at least 50 times, so the TSC error is basically at *MOST* just a few PIT cycles off in any direction. In practice, it's going to be about one microseconds off (which is how long it takes to read the counter) - so over 17920 PIT cycles, we can pretty much guarantee that the calibration error is less than one half of a percent. My testing bears this out: on my machine, the quick-calibration reports 2934.085kHz, while the slow one reports 2933.415. Yes, the slower calibration is still more precise. For me, the slow calibration is stable to within about one hundreth of a percent, so it's (at a guess) roughly an order-and-a-half of magnitude more precise. The longer you wait, the more precise you can be. However, the nice thing about the fast TSC PIT synchronization is that it's pretty much _guaranteed_ to give that 0.5% precision, and fail gracefully (and very quickly) if it doesn't get it. And it really is fairly simple (even if there's a lot of _details_ there, and I didn't get all of those right ont he first try or even the second ;) The patch says "110 insertions", but 63 of those new lines are actually comments. Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Ingo Molnar <mingo@elte.hu> --- arch/x86/kernel/tsc.c | 111 ++++++++++++++++++++++++++++++++++++++++++++++++- 1 files changed, 110 insertions(+), 1 deletions(-)
2008-09-04 21:41:22 +04:00
}
/**
* native_calibrate_tsc - calibrate the tsc on boot
*/
unsigned long native_calibrate_tsc(void)
{
u64 tsc1, tsc2, delta, ref1, ref2;
[x86] Fix TSC calibration issues Larry Finger reported at http://lkml.org/lkml/2008/9/1/90: An ancient laptop of mine started throwing errors from b43legacy when I started using 2.6.27 on it. This has been bisected to commit bfc0f59 "x86: merge tsc calibration". The unification of the TSC code adopted mostly the 64bit code, which prefers PMTIMER/HPET over the PIT calibration. Larrys system has an AMD K6 CPU. Such systems are known to have PMTIMER incarnations which run at double speed. This results in a miscalibration of the TSC by factor 0.5. So the resulting calibrated CPU/TSC speed is half of the real CPU speed, which means that the TSC based delay loop will run half the time it should run. That might explain why the b43legacy driver went berserk. On the other hand we know about systems, where the PIT based calibration results in random crap due to heavy SMI/SMM disturbance. On those systems the PMTIMER/HPET based calibration logic with SMI detection shows better results. According to Alok also virtualized systems suffer from the PIT calibration method. The solution is to use a more wreckage aware aproach than the current either/or decision. 1) reimplement the retry loop which was dropped from the 32bit code during the merge. It repeats the calibration and selects the lowest frequency value as this is probably the closest estimate to the real frequency 2) Monitor the delta of the TSC values in the delay loop which waits for the PIT counter to reach zero. If the maximum value is significantly different from the minimum, then we have a pretty safe indicator that the loop was disturbed by an SMI. 3) keep the pmtimer/hpet reference as a backup solution for systems where the SMI disturbance is a permanent point of failure for PIT based calibration 4) do the loop iteration for both methods, record the lowest value and decide after all iterations finished. 5) Set a clear preference to PIT based calibration when the result makes sense. The implementation does the reference calibration based on HPET/PMTIMER around the delay, which is necessary for the PIT anyway, but keeps separate TSC values to ensure the "independency" of the resulting calibration values. Tested on various 32bit/64bit machines including Geode 266Mhz, AMD K6 (affected machine with a double speed pmtimer which I grabbed out of the dump), Pentium class machines and AMD/Intel 64 bit boxen. Bisected-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-09-03 02:54:47 +04:00
unsigned long tsc_pit_min = ULONG_MAX, tsc_ref_min = ULONG_MAX;
unsigned long flags, latch, ms, fast_calibrate;
int hpet = is_hpet_enabled(), i, loopmin;
x86: quick TSC calibration Introduce a fast TSC-calibration method on sane hardware. It only uses 17920 PIT timer ticks to calibrate the TSC, plus 256 ticks on each side to make sure the TSC values were very close to the tick, so the whole calibration takes 15ms. Yet, despite only takign 15ms, we can actually give pretty stringent guarantees of accuracy: - the code requires that we hit each 256-counter block at least 50 times, so the TSC error is basically at *MOST* just a few PIT cycles off in any direction. In practice, it's going to be about one microseconds off (which is how long it takes to read the counter) - so over 17920 PIT cycles, we can pretty much guarantee that the calibration error is less than one half of a percent. My testing bears this out: on my machine, the quick-calibration reports 2934.085kHz, while the slow one reports 2933.415. Yes, the slower calibration is still more precise. For me, the slow calibration is stable to within about one hundreth of a percent, so it's (at a guess) roughly an order-and-a-half of magnitude more precise. The longer you wait, the more precise you can be. However, the nice thing about the fast TSC PIT synchronization is that it's pretty much _guaranteed_ to give that 0.5% precision, and fail gracefully (and very quickly) if it doesn't get it. And it really is fairly simple (even if there's a lot of _details_ there, and I didn't get all of those right ont he first try or even the second ;) The patch says "110 insertions", but 63 of those new lines are actually comments. Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Ingo Molnar <mingo@elte.hu> --- arch/x86/kernel/tsc.c | 111 ++++++++++++++++++++++++++++++++++++++++++++++++- 1 files changed, 110 insertions(+), 1 deletions(-)
2008-09-04 21:41:22 +04:00
local_irq_save(flags);
fast_calibrate = quick_pit_calibrate();
local_irq_restore(flags);
x86: quick TSC calibration Introduce a fast TSC-calibration method on sane hardware. It only uses 17920 PIT timer ticks to calibrate the TSC, plus 256 ticks on each side to make sure the TSC values were very close to the tick, so the whole calibration takes 15ms. Yet, despite only takign 15ms, we can actually give pretty stringent guarantees of accuracy: - the code requires that we hit each 256-counter block at least 50 times, so the TSC error is basically at *MOST* just a few PIT cycles off in any direction. In practice, it's going to be about one microseconds off (which is how long it takes to read the counter) - so over 17920 PIT cycles, we can pretty much guarantee that the calibration error is less than one half of a percent. My testing bears this out: on my machine, the quick-calibration reports 2934.085kHz, while the slow one reports 2933.415. Yes, the slower calibration is still more precise. For me, the slow calibration is stable to within about one hundreth of a percent, so it's (at a guess) roughly an order-and-a-half of magnitude more precise. The longer you wait, the more precise you can be. However, the nice thing about the fast TSC PIT synchronization is that it's pretty much _guaranteed_ to give that 0.5% precision, and fail gracefully (and very quickly) if it doesn't get it. And it really is fairly simple (even if there's a lot of _details_ there, and I didn't get all of those right ont he first try or even the second ;) The patch says "110 insertions", but 63 of those new lines are actually comments. Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Ingo Molnar <mingo@elte.hu> --- arch/x86/kernel/tsc.c | 111 ++++++++++++++++++++++++++++++++++++++++++++++++- 1 files changed, 110 insertions(+), 1 deletions(-)
2008-09-04 21:41:22 +04:00
if (fast_calibrate)
return fast_calibrate;
[x86] Fix TSC calibration issues Larry Finger reported at http://lkml.org/lkml/2008/9/1/90: An ancient laptop of mine started throwing errors from b43legacy when I started using 2.6.27 on it. This has been bisected to commit bfc0f59 "x86: merge tsc calibration". The unification of the TSC code adopted mostly the 64bit code, which prefers PMTIMER/HPET over the PIT calibration. Larrys system has an AMD K6 CPU. Such systems are known to have PMTIMER incarnations which run at double speed. This results in a miscalibration of the TSC by factor 0.5. So the resulting calibrated CPU/TSC speed is half of the real CPU speed, which means that the TSC based delay loop will run half the time it should run. That might explain why the b43legacy driver went berserk. On the other hand we know about systems, where the PIT based calibration results in random crap due to heavy SMI/SMM disturbance. On those systems the PMTIMER/HPET based calibration logic with SMI detection shows better results. According to Alok also virtualized systems suffer from the PIT calibration method. The solution is to use a more wreckage aware aproach than the current either/or decision. 1) reimplement the retry loop which was dropped from the 32bit code during the merge. It repeats the calibration and selects the lowest frequency value as this is probably the closest estimate to the real frequency 2) Monitor the delta of the TSC values in the delay loop which waits for the PIT counter to reach zero. If the maximum value is significantly different from the minimum, then we have a pretty safe indicator that the loop was disturbed by an SMI. 3) keep the pmtimer/hpet reference as a backup solution for systems where the SMI disturbance is a permanent point of failure for PIT based calibration 4) do the loop iteration for both methods, record the lowest value and decide after all iterations finished. 5) Set a clear preference to PIT based calibration when the result makes sense. The implementation does the reference calibration based on HPET/PMTIMER around the delay, which is necessary for the PIT anyway, but keeps separate TSC values to ensure the "independency" of the resulting calibration values. Tested on various 32bit/64bit machines including Geode 266Mhz, AMD K6 (affected machine with a double speed pmtimer which I grabbed out of the dump), Pentium class machines and AMD/Intel 64 bit boxen. Bisected-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-09-03 02:54:47 +04:00
/*
* Run 5 calibration loops to get the lowest frequency value
* (the best estimate). We use two different calibration modes
* here:
*
* 1) PIT loop. We set the PIT Channel 2 to oneshot mode and
* load a timeout of 50ms. We read the time right after we
* started the timer and wait until the PIT count down reaches
* 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
[x86] Fix TSC calibration issues Larry Finger reported at http://lkml.org/lkml/2008/9/1/90: An ancient laptop of mine started throwing errors from b43legacy when I started using 2.6.27 on it. This has been bisected to commit bfc0f59 "x86: merge tsc calibration". The unification of the TSC code adopted mostly the 64bit code, which prefers PMTIMER/HPET over the PIT calibration. Larrys system has an AMD K6 CPU. Such systems are known to have PMTIMER incarnations which run at double speed. This results in a miscalibration of the TSC by factor 0.5. So the resulting calibrated CPU/TSC speed is half of the real CPU speed, which means that the TSC based delay loop will run half the time it should run. That might explain why the b43legacy driver went berserk. On the other hand we know about systems, where the PIT based calibration results in random crap due to heavy SMI/SMM disturbance. On those systems the PMTIMER/HPET based calibration logic with SMI detection shows better results. According to Alok also virtualized systems suffer from the PIT calibration method. The solution is to use a more wreckage aware aproach than the current either/or decision. 1) reimplement the retry loop which was dropped from the 32bit code during the merge. It repeats the calibration and selects the lowest frequency value as this is probably the closest estimate to the real frequency 2) Monitor the delta of the TSC values in the delay loop which waits for the PIT counter to reach zero. If the maximum value is significantly different from the minimum, then we have a pretty safe indicator that the loop was disturbed by an SMI. 3) keep the pmtimer/hpet reference as a backup solution for systems where the SMI disturbance is a permanent point of failure for PIT based calibration 4) do the loop iteration for both methods, record the lowest value and decide after all iterations finished. 5) Set a clear preference to PIT based calibration when the result makes sense. The implementation does the reference calibration based on HPET/PMTIMER around the delay, which is necessary for the PIT anyway, but keeps separate TSC values to ensure the "independency" of the resulting calibration values. Tested on various 32bit/64bit machines including Geode 266Mhz, AMD K6 (affected machine with a double speed pmtimer which I grabbed out of the dump), Pentium class machines and AMD/Intel 64 bit boxen. Bisected-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-09-03 02:54:47 +04:00
* 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
* 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.
*/
/* Preset PIT loop values */
latch = CAL_LATCH;
ms = CAL_MS;
loopmin = CAL_PIT_LOOPS;
for (i = 0; i < 3; i++) {
unsigned long tsc_pit_khz;
[x86] Fix TSC calibration issues Larry Finger reported at http://lkml.org/lkml/2008/9/1/90: An ancient laptop of mine started throwing errors from b43legacy when I started using 2.6.27 on it. This has been bisected to commit bfc0f59 "x86: merge tsc calibration". The unification of the TSC code adopted mostly the 64bit code, which prefers PMTIMER/HPET over the PIT calibration. Larrys system has an AMD K6 CPU. Such systems are known to have PMTIMER incarnations which run at double speed. This results in a miscalibration of the TSC by factor 0.5. So the resulting calibrated CPU/TSC speed is half of the real CPU speed, which means that the TSC based delay loop will run half the time it should run. That might explain why the b43legacy driver went berserk. On the other hand we know about systems, where the PIT based calibration results in random crap due to heavy SMI/SMM disturbance. On those systems the PMTIMER/HPET based calibration logic with SMI detection shows better results. According to Alok also virtualized systems suffer from the PIT calibration method. The solution is to use a more wreckage aware aproach than the current either/or decision. 1) reimplement the retry loop which was dropped from the 32bit code during the merge. It repeats the calibration and selects the lowest frequency value as this is probably the closest estimate to the real frequency 2) Monitor the delta of the TSC values in the delay loop which waits for the PIT counter to reach zero. If the maximum value is significantly different from the minimum, then we have a pretty safe indicator that the loop was disturbed by an SMI. 3) keep the pmtimer/hpet reference as a backup solution for systems where the SMI disturbance is a permanent point of failure for PIT based calibration 4) do the loop iteration for both methods, record the lowest value and decide after all iterations finished. 5) Set a clear preference to PIT based calibration when the result makes sense. The implementation does the reference calibration based on HPET/PMTIMER around the delay, which is necessary for the PIT anyway, but keeps separate TSC values to ensure the "independency" of the resulting calibration values. Tested on various 32bit/64bit machines including Geode 266Mhz, AMD K6 (affected machine with a double speed pmtimer which I grabbed out of the dump), Pentium class machines and AMD/Intel 64 bit boxen. Bisected-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-09-03 02:54:47 +04:00
/*
* Read the start value and the reference count of
* hpet/pmtimer when available. Then do the PIT
* calibration, which will take at least 50ms, and
* read the end value.
[x86] Fix TSC calibration issues Larry Finger reported at http://lkml.org/lkml/2008/9/1/90: An ancient laptop of mine started throwing errors from b43legacy when I started using 2.6.27 on it. This has been bisected to commit bfc0f59 "x86: merge tsc calibration". The unification of the TSC code adopted mostly the 64bit code, which prefers PMTIMER/HPET over the PIT calibration. Larrys system has an AMD K6 CPU. Such systems are known to have PMTIMER incarnations which run at double speed. This results in a miscalibration of the TSC by factor 0.5. So the resulting calibrated CPU/TSC speed is half of the real CPU speed, which means that the TSC based delay loop will run half the time it should run. That might explain why the b43legacy driver went berserk. On the other hand we know about systems, where the PIT based calibration results in random crap due to heavy SMI/SMM disturbance. On those systems the PMTIMER/HPET based calibration logic with SMI detection shows better results. According to Alok also virtualized systems suffer from the PIT calibration method. The solution is to use a more wreckage aware aproach than the current either/or decision. 1) reimplement the retry loop which was dropped from the 32bit code during the merge. It repeats the calibration and selects the lowest frequency value as this is probably the closest estimate to the real frequency 2) Monitor the delta of the TSC values in the delay loop which waits for the PIT counter to reach zero. If the maximum value is significantly different from the minimum, then we have a pretty safe indicator that the loop was disturbed by an SMI. 3) keep the pmtimer/hpet reference as a backup solution for systems where the SMI disturbance is a permanent point of failure for PIT based calibration 4) do the loop iteration for both methods, record the lowest value and decide after all iterations finished. 5) Set a clear preference to PIT based calibration when the result makes sense. The implementation does the reference calibration based on HPET/PMTIMER around the delay, which is necessary for the PIT anyway, but keeps separate TSC values to ensure the "independency" of the resulting calibration values. Tested on various 32bit/64bit machines including Geode 266Mhz, AMD K6 (affected machine with a double speed pmtimer which I grabbed out of the dump), Pentium class machines and AMD/Intel 64 bit boxen. Bisected-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-09-03 02:54:47 +04:00
*/
local_irq_save(flags);
tsc1 = tsc_read_refs(&ref1, hpet);
tsc_pit_khz = pit_calibrate_tsc(latch, ms, loopmin);
tsc2 = tsc_read_refs(&ref2, hpet);
[x86] Fix TSC calibration issues Larry Finger reported at http://lkml.org/lkml/2008/9/1/90: An ancient laptop of mine started throwing errors from b43legacy when I started using 2.6.27 on it. This has been bisected to commit bfc0f59 "x86: merge tsc calibration". The unification of the TSC code adopted mostly the 64bit code, which prefers PMTIMER/HPET over the PIT calibration. Larrys system has an AMD K6 CPU. Such systems are known to have PMTIMER incarnations which run at double speed. This results in a miscalibration of the TSC by factor 0.5. So the resulting calibrated CPU/TSC speed is half of the real CPU speed, which means that the TSC based delay loop will run half the time it should run. That might explain why the b43legacy driver went berserk. On the other hand we know about systems, where the PIT based calibration results in random crap due to heavy SMI/SMM disturbance. On those systems the PMTIMER/HPET based calibration logic with SMI detection shows better results. According to Alok also virtualized systems suffer from the PIT calibration method. The solution is to use a more wreckage aware aproach than the current either/or decision. 1) reimplement the retry loop which was dropped from the 32bit code during the merge. It repeats the calibration and selects the lowest frequency value as this is probably the closest estimate to the real frequency 2) Monitor the delta of the TSC values in the delay loop which waits for the PIT counter to reach zero. If the maximum value is significantly different from the minimum, then we have a pretty safe indicator that the loop was disturbed by an SMI. 3) keep the pmtimer/hpet reference as a backup solution for systems where the SMI disturbance is a permanent point of failure for PIT based calibration 4) do the loop iteration for both methods, record the lowest value and decide after all iterations finished. 5) Set a clear preference to PIT based calibration when the result makes sense. The implementation does the reference calibration based on HPET/PMTIMER around the delay, which is necessary for the PIT anyway, but keeps separate TSC values to ensure the "independency" of the resulting calibration values. Tested on various 32bit/64bit machines including Geode 266Mhz, AMD K6 (affected machine with a double speed pmtimer which I grabbed out of the dump), Pentium class machines and AMD/Intel 64 bit boxen. Bisected-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-09-03 02:54:47 +04:00
local_irq_restore(flags);
/* Pick the lowest PIT TSC calibration so far */
tsc_pit_min = min(tsc_pit_min, tsc_pit_khz);
[x86] Fix TSC calibration issues Larry Finger reported at http://lkml.org/lkml/2008/9/1/90: An ancient laptop of mine started throwing errors from b43legacy when I started using 2.6.27 on it. This has been bisected to commit bfc0f59 "x86: merge tsc calibration". The unification of the TSC code adopted mostly the 64bit code, which prefers PMTIMER/HPET over the PIT calibration. Larrys system has an AMD K6 CPU. Such systems are known to have PMTIMER incarnations which run at double speed. This results in a miscalibration of the TSC by factor 0.5. So the resulting calibrated CPU/TSC speed is half of the real CPU speed, which means that the TSC based delay loop will run half the time it should run. That might explain why the b43legacy driver went berserk. On the other hand we know about systems, where the PIT based calibration results in random crap due to heavy SMI/SMM disturbance. On those systems the PMTIMER/HPET based calibration logic with SMI detection shows better results. According to Alok also virtualized systems suffer from the PIT calibration method. The solution is to use a more wreckage aware aproach than the current either/or decision. 1) reimplement the retry loop which was dropped from the 32bit code during the merge. It repeats the calibration and selects the lowest frequency value as this is probably the closest estimate to the real frequency 2) Monitor the delta of the TSC values in the delay loop which waits for the PIT counter to reach zero. If the maximum value is significantly different from the minimum, then we have a pretty safe indicator that the loop was disturbed by an SMI. 3) keep the pmtimer/hpet reference as a backup solution for systems where the SMI disturbance is a permanent point of failure for PIT based calibration 4) do the loop iteration for both methods, record the lowest value and decide after all iterations finished. 5) Set a clear preference to PIT based calibration when the result makes sense. The implementation does the reference calibration based on HPET/PMTIMER around the delay, which is necessary for the PIT anyway, but keeps separate TSC values to ensure the "independency" of the resulting calibration values. Tested on various 32bit/64bit machines including Geode 266Mhz, AMD K6 (affected machine with a double speed pmtimer which I grabbed out of the dump), Pentium class machines and AMD/Intel 64 bit boxen. Bisected-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-09-03 02:54:47 +04:00
/* hpet or pmtimer available ? */
if (ref1 == ref2)
[x86] Fix TSC calibration issues Larry Finger reported at http://lkml.org/lkml/2008/9/1/90: An ancient laptop of mine started throwing errors from b43legacy when I started using 2.6.27 on it. This has been bisected to commit bfc0f59 "x86: merge tsc calibration". The unification of the TSC code adopted mostly the 64bit code, which prefers PMTIMER/HPET over the PIT calibration. Larrys system has an AMD K6 CPU. Such systems are known to have PMTIMER incarnations which run at double speed. This results in a miscalibration of the TSC by factor 0.5. So the resulting calibrated CPU/TSC speed is half of the real CPU speed, which means that the TSC based delay loop will run half the time it should run. That might explain why the b43legacy driver went berserk. On the other hand we know about systems, where the PIT based calibration results in random crap due to heavy SMI/SMM disturbance. On those systems the PMTIMER/HPET based calibration logic with SMI detection shows better results. According to Alok also virtualized systems suffer from the PIT calibration method. The solution is to use a more wreckage aware aproach than the current either/or decision. 1) reimplement the retry loop which was dropped from the 32bit code during the merge. It repeats the calibration and selects the lowest frequency value as this is probably the closest estimate to the real frequency 2) Monitor the delta of the TSC values in the delay loop which waits for the PIT counter to reach zero. If the maximum value is significantly different from the minimum, then we have a pretty safe indicator that the loop was disturbed by an SMI. 3) keep the pmtimer/hpet reference as a backup solution for systems where the SMI disturbance is a permanent point of failure for PIT based calibration 4) do the loop iteration for both methods, record the lowest value and decide after all iterations finished. 5) Set a clear preference to PIT based calibration when the result makes sense. The implementation does the reference calibration based on HPET/PMTIMER around the delay, which is necessary for the PIT anyway, but keeps separate TSC values to ensure the "independency" of the resulting calibration values. Tested on various 32bit/64bit machines including Geode 266Mhz, AMD K6 (affected machine with a double speed pmtimer which I grabbed out of the dump), Pentium class machines and AMD/Intel 64 bit boxen. Bisected-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-09-03 02:54:47 +04:00
continue;
/* Check, whether the sampling was disturbed by an SMI */
if (tsc1 == ULLONG_MAX || tsc2 == ULLONG_MAX)
continue;
tsc2 = (tsc2 - tsc1) * 1000000LL;
if (hpet)
tsc2 = calc_hpet_ref(tsc2, ref1, ref2);
else
tsc2 = calc_pmtimer_ref(tsc2, ref1, ref2);
[x86] Fix TSC calibration issues Larry Finger reported at http://lkml.org/lkml/2008/9/1/90: An ancient laptop of mine started throwing errors from b43legacy when I started using 2.6.27 on it. This has been bisected to commit bfc0f59 "x86: merge tsc calibration". The unification of the TSC code adopted mostly the 64bit code, which prefers PMTIMER/HPET over the PIT calibration. Larrys system has an AMD K6 CPU. Such systems are known to have PMTIMER incarnations which run at double speed. This results in a miscalibration of the TSC by factor 0.5. So the resulting calibrated CPU/TSC speed is half of the real CPU speed, which means that the TSC based delay loop will run half the time it should run. That might explain why the b43legacy driver went berserk. On the other hand we know about systems, where the PIT based calibration results in random crap due to heavy SMI/SMM disturbance. On those systems the PMTIMER/HPET based calibration logic with SMI detection shows better results. According to Alok also virtualized systems suffer from the PIT calibration method. The solution is to use a more wreckage aware aproach than the current either/or decision. 1) reimplement the retry loop which was dropped from the 32bit code during the merge. It repeats the calibration and selects the lowest frequency value as this is probably the closest estimate to the real frequency 2) Monitor the delta of the TSC values in the delay loop which waits for the PIT counter to reach zero. If the maximum value is significantly different from the minimum, then we have a pretty safe indicator that the loop was disturbed by an SMI. 3) keep the pmtimer/hpet reference as a backup solution for systems where the SMI disturbance is a permanent point of failure for PIT based calibration 4) do the loop iteration for both methods, record the lowest value and decide after all iterations finished. 5) Set a clear preference to PIT based calibration when the result makes sense. The implementation does the reference calibration based on HPET/PMTIMER around the delay, which is necessary for the PIT anyway, but keeps separate TSC values to ensure the "independency" of the resulting calibration values. Tested on various 32bit/64bit machines including Geode 266Mhz, AMD K6 (affected machine with a double speed pmtimer which I grabbed out of the dump), Pentium class machines and AMD/Intel 64 bit boxen. Bisected-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-09-03 02:54:47 +04:00
tsc_ref_min = min(tsc_ref_min, (unsigned long) tsc2);
/* Check the reference deviation */
delta = ((u64) tsc_pit_min) * 100;
do_div(delta, tsc_ref_min);
/*
* If both calibration results are inside a 10% window
* then we can be sure, that the calibration
* succeeded. We break out of the loop right away. We
* use the reference value, as it is more precise.
*/
if (delta >= 90 && delta <= 110) {
printk(KERN_INFO
"TSC: PIT calibration matches %s. %d loops\n",
hpet ? "HPET" : "PMTIMER", i + 1);
return tsc_ref_min;
[x86] Fix TSC calibration issues Larry Finger reported at http://lkml.org/lkml/2008/9/1/90: An ancient laptop of mine started throwing errors from b43legacy when I started using 2.6.27 on it. This has been bisected to commit bfc0f59 "x86: merge tsc calibration". The unification of the TSC code adopted mostly the 64bit code, which prefers PMTIMER/HPET over the PIT calibration. Larrys system has an AMD K6 CPU. Such systems are known to have PMTIMER incarnations which run at double speed. This results in a miscalibration of the TSC by factor 0.5. So the resulting calibrated CPU/TSC speed is half of the real CPU speed, which means that the TSC based delay loop will run half the time it should run. That might explain why the b43legacy driver went berserk. On the other hand we know about systems, where the PIT based calibration results in random crap due to heavy SMI/SMM disturbance. On those systems the PMTIMER/HPET based calibration logic with SMI detection shows better results. According to Alok also virtualized systems suffer from the PIT calibration method. The solution is to use a more wreckage aware aproach than the current either/or decision. 1) reimplement the retry loop which was dropped from the 32bit code during the merge. It repeats the calibration and selects the lowest frequency value as this is probably the closest estimate to the real frequency 2) Monitor the delta of the TSC values in the delay loop which waits for the PIT counter to reach zero. If the maximum value is significantly different from the minimum, then we have a pretty safe indicator that the loop was disturbed by an SMI. 3) keep the pmtimer/hpet reference as a backup solution for systems where the SMI disturbance is a permanent point of failure for PIT based calibration 4) do the loop iteration for both methods, record the lowest value and decide after all iterations finished. 5) Set a clear preference to PIT based calibration when the result makes sense. The implementation does the reference calibration based on HPET/PMTIMER around the delay, which is necessary for the PIT anyway, but keeps separate TSC values to ensure the "independency" of the resulting calibration values. Tested on various 32bit/64bit machines including Geode 266Mhz, AMD K6 (affected machine with a double speed pmtimer which I grabbed out of the dump), Pentium class machines and AMD/Intel 64 bit boxen. Bisected-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-09-03 02:54:47 +04:00
}
/*
* Check whether PIT failed more than once. This
* happens in virtualized environments. We need to
* give the virtual PC a slightly longer timeframe for
* the HPET/PMTIMER to make the result precise.
*/
if (i == 1 && tsc_pit_min == ULONG_MAX) {
latch = CAL2_LATCH;
ms = CAL2_MS;
loopmin = CAL2_PIT_LOOPS;
}
[x86] Fix TSC calibration issues Larry Finger reported at http://lkml.org/lkml/2008/9/1/90: An ancient laptop of mine started throwing errors from b43legacy when I started using 2.6.27 on it. This has been bisected to commit bfc0f59 "x86: merge tsc calibration". The unification of the TSC code adopted mostly the 64bit code, which prefers PMTIMER/HPET over the PIT calibration. Larrys system has an AMD K6 CPU. Such systems are known to have PMTIMER incarnations which run at double speed. This results in a miscalibration of the TSC by factor 0.5. So the resulting calibrated CPU/TSC speed is half of the real CPU speed, which means that the TSC based delay loop will run half the time it should run. That might explain why the b43legacy driver went berserk. On the other hand we know about systems, where the PIT based calibration results in random crap due to heavy SMI/SMM disturbance. On those systems the PMTIMER/HPET based calibration logic with SMI detection shows better results. According to Alok also virtualized systems suffer from the PIT calibration method. The solution is to use a more wreckage aware aproach than the current either/or decision. 1) reimplement the retry loop which was dropped from the 32bit code during the merge. It repeats the calibration and selects the lowest frequency value as this is probably the closest estimate to the real frequency 2) Monitor the delta of the TSC values in the delay loop which waits for the PIT counter to reach zero. If the maximum value is significantly different from the minimum, then we have a pretty safe indicator that the loop was disturbed by an SMI. 3) keep the pmtimer/hpet reference as a backup solution for systems where the SMI disturbance is a permanent point of failure for PIT based calibration 4) do the loop iteration for both methods, record the lowest value and decide after all iterations finished. 5) Set a clear preference to PIT based calibration when the result makes sense. The implementation does the reference calibration based on HPET/PMTIMER around the delay, which is necessary for the PIT anyway, but keeps separate TSC values to ensure the "independency" of the resulting calibration values. Tested on various 32bit/64bit machines including Geode 266Mhz, AMD K6 (affected machine with a double speed pmtimer which I grabbed out of the dump), Pentium class machines and AMD/Intel 64 bit boxen. Bisected-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-09-03 02:54:47 +04:00
}
/*
[x86] Fix TSC calibration issues Larry Finger reported at http://lkml.org/lkml/2008/9/1/90: An ancient laptop of mine started throwing errors from b43legacy when I started using 2.6.27 on it. This has been bisected to commit bfc0f59 "x86: merge tsc calibration". The unification of the TSC code adopted mostly the 64bit code, which prefers PMTIMER/HPET over the PIT calibration. Larrys system has an AMD K6 CPU. Such systems are known to have PMTIMER incarnations which run at double speed. This results in a miscalibration of the TSC by factor 0.5. So the resulting calibrated CPU/TSC speed is half of the real CPU speed, which means that the TSC based delay loop will run half the time it should run. That might explain why the b43legacy driver went berserk. On the other hand we know about systems, where the PIT based calibration results in random crap due to heavy SMI/SMM disturbance. On those systems the PMTIMER/HPET based calibration logic with SMI detection shows better results. According to Alok also virtualized systems suffer from the PIT calibration method. The solution is to use a more wreckage aware aproach than the current either/or decision. 1) reimplement the retry loop which was dropped from the 32bit code during the merge. It repeats the calibration and selects the lowest frequency value as this is probably the closest estimate to the real frequency 2) Monitor the delta of the TSC values in the delay loop which waits for the PIT counter to reach zero. If the maximum value is significantly different from the minimum, then we have a pretty safe indicator that the loop was disturbed by an SMI. 3) keep the pmtimer/hpet reference as a backup solution for systems where the SMI disturbance is a permanent point of failure for PIT based calibration 4) do the loop iteration for both methods, record the lowest value and decide after all iterations finished. 5) Set a clear preference to PIT based calibration when the result makes sense. The implementation does the reference calibration based on HPET/PMTIMER around the delay, which is necessary for the PIT anyway, but keeps separate TSC values to ensure the "independency" of the resulting calibration values. Tested on various 32bit/64bit machines including Geode 266Mhz, AMD K6 (affected machine with a double speed pmtimer which I grabbed out of the dump), Pentium class machines and AMD/Intel 64 bit boxen. Bisected-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-09-03 02:54:47 +04:00
* Now check the results.
*/
[x86] Fix TSC calibration issues Larry Finger reported at http://lkml.org/lkml/2008/9/1/90: An ancient laptop of mine started throwing errors from b43legacy when I started using 2.6.27 on it. This has been bisected to commit bfc0f59 "x86: merge tsc calibration". The unification of the TSC code adopted mostly the 64bit code, which prefers PMTIMER/HPET over the PIT calibration. Larrys system has an AMD K6 CPU. Such systems are known to have PMTIMER incarnations which run at double speed. This results in a miscalibration of the TSC by factor 0.5. So the resulting calibrated CPU/TSC speed is half of the real CPU speed, which means that the TSC based delay loop will run half the time it should run. That might explain why the b43legacy driver went berserk. On the other hand we know about systems, where the PIT based calibration results in random crap due to heavy SMI/SMM disturbance. On those systems the PMTIMER/HPET based calibration logic with SMI detection shows better results. According to Alok also virtualized systems suffer from the PIT calibration method. The solution is to use a more wreckage aware aproach than the current either/or decision. 1) reimplement the retry loop which was dropped from the 32bit code during the merge. It repeats the calibration and selects the lowest frequency value as this is probably the closest estimate to the real frequency 2) Monitor the delta of the TSC values in the delay loop which waits for the PIT counter to reach zero. If the maximum value is significantly different from the minimum, then we have a pretty safe indicator that the loop was disturbed by an SMI. 3) keep the pmtimer/hpet reference as a backup solution for systems where the SMI disturbance is a permanent point of failure for PIT based calibration 4) do the loop iteration for both methods, record the lowest value and decide after all iterations finished. 5) Set a clear preference to PIT based calibration when the result makes sense. The implementation does the reference calibration based on HPET/PMTIMER around the delay, which is necessary for the PIT anyway, but keeps separate TSC values to ensure the "independency" of the resulting calibration values. Tested on various 32bit/64bit machines including Geode 266Mhz, AMD K6 (affected machine with a double speed pmtimer which I grabbed out of the dump), Pentium class machines and AMD/Intel 64 bit boxen. Bisected-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-09-03 02:54:47 +04:00
if (tsc_pit_min == ULONG_MAX) {
/* PIT gave no useful value */
printk(KERN_WARNING "TSC: Unable to calibrate against PIT\n");
[x86] Fix TSC calibration issues Larry Finger reported at http://lkml.org/lkml/2008/9/1/90: An ancient laptop of mine started throwing errors from b43legacy when I started using 2.6.27 on it. This has been bisected to commit bfc0f59 "x86: merge tsc calibration". The unification of the TSC code adopted mostly the 64bit code, which prefers PMTIMER/HPET over the PIT calibration. Larrys system has an AMD K6 CPU. Such systems are known to have PMTIMER incarnations which run at double speed. This results in a miscalibration of the TSC by factor 0.5. So the resulting calibrated CPU/TSC speed is half of the real CPU speed, which means that the TSC based delay loop will run half the time it should run. That might explain why the b43legacy driver went berserk. On the other hand we know about systems, where the PIT based calibration results in random crap due to heavy SMI/SMM disturbance. On those systems the PMTIMER/HPET based calibration logic with SMI detection shows better results. According to Alok also virtualized systems suffer from the PIT calibration method. The solution is to use a more wreckage aware aproach than the current either/or decision. 1) reimplement the retry loop which was dropped from the 32bit code during the merge. It repeats the calibration and selects the lowest frequency value as this is probably the closest estimate to the real frequency 2) Monitor the delta of the TSC values in the delay loop which waits for the PIT counter to reach zero. If the maximum value is significantly different from the minimum, then we have a pretty safe indicator that the loop was disturbed by an SMI. 3) keep the pmtimer/hpet reference as a backup solution for systems where the SMI disturbance is a permanent point of failure for PIT based calibration 4) do the loop iteration for both methods, record the lowest value and decide after all iterations finished. 5) Set a clear preference to PIT based calibration when the result makes sense. The implementation does the reference calibration based on HPET/PMTIMER around the delay, which is necessary for the PIT anyway, but keeps separate TSC values to ensure the "independency" of the resulting calibration values. Tested on various 32bit/64bit machines including Geode 266Mhz, AMD K6 (affected machine with a double speed pmtimer which I grabbed out of the dump), Pentium class machines and AMD/Intel 64 bit boxen. Bisected-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-09-03 02:54:47 +04:00
/* We don't have an alternative source, disable TSC */
if (!hpet && !ref1 && !ref2) {
[x86] Fix TSC calibration issues Larry Finger reported at http://lkml.org/lkml/2008/9/1/90: An ancient laptop of mine started throwing errors from b43legacy when I started using 2.6.27 on it. This has been bisected to commit bfc0f59 "x86: merge tsc calibration". The unification of the TSC code adopted mostly the 64bit code, which prefers PMTIMER/HPET over the PIT calibration. Larrys system has an AMD K6 CPU. Such systems are known to have PMTIMER incarnations which run at double speed. This results in a miscalibration of the TSC by factor 0.5. So the resulting calibrated CPU/TSC speed is half of the real CPU speed, which means that the TSC based delay loop will run half the time it should run. That might explain why the b43legacy driver went berserk. On the other hand we know about systems, where the PIT based calibration results in random crap due to heavy SMI/SMM disturbance. On those systems the PMTIMER/HPET based calibration logic with SMI detection shows better results. According to Alok also virtualized systems suffer from the PIT calibration method. The solution is to use a more wreckage aware aproach than the current either/or decision. 1) reimplement the retry loop which was dropped from the 32bit code during the merge. It repeats the calibration and selects the lowest frequency value as this is probably the closest estimate to the real frequency 2) Monitor the delta of the TSC values in the delay loop which waits for the PIT counter to reach zero. If the maximum value is significantly different from the minimum, then we have a pretty safe indicator that the loop was disturbed by an SMI. 3) keep the pmtimer/hpet reference as a backup solution for systems where the SMI disturbance is a permanent point of failure for PIT based calibration 4) do the loop iteration for both methods, record the lowest value and decide after all iterations finished. 5) Set a clear preference to PIT based calibration when the result makes sense. The implementation does the reference calibration based on HPET/PMTIMER around the delay, which is necessary for the PIT anyway, but keeps separate TSC values to ensure the "independency" of the resulting calibration values. Tested on various 32bit/64bit machines including Geode 266Mhz, AMD K6 (affected machine with a double speed pmtimer which I grabbed out of the dump), Pentium class machines and AMD/Intel 64 bit boxen. Bisected-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-09-03 02:54:47 +04:00
printk("TSC: No reference (HPET/PMTIMER) available\n");
return 0;
}
/* The alternative source failed as well, disable TSC */
if (tsc_ref_min == ULONG_MAX) {
printk(KERN_WARNING "TSC: HPET/PMTIMER calibration "
"failed.\n");
[x86] Fix TSC calibration issues Larry Finger reported at http://lkml.org/lkml/2008/9/1/90: An ancient laptop of mine started throwing errors from b43legacy when I started using 2.6.27 on it. This has been bisected to commit bfc0f59 "x86: merge tsc calibration". The unification of the TSC code adopted mostly the 64bit code, which prefers PMTIMER/HPET over the PIT calibration. Larrys system has an AMD K6 CPU. Such systems are known to have PMTIMER incarnations which run at double speed. This results in a miscalibration of the TSC by factor 0.5. So the resulting calibrated CPU/TSC speed is half of the real CPU speed, which means that the TSC based delay loop will run half the time it should run. That might explain why the b43legacy driver went berserk. On the other hand we know about systems, where the PIT based calibration results in random crap due to heavy SMI/SMM disturbance. On those systems the PMTIMER/HPET based calibration logic with SMI detection shows better results. According to Alok also virtualized systems suffer from the PIT calibration method. The solution is to use a more wreckage aware aproach than the current either/or decision. 1) reimplement the retry loop which was dropped from the 32bit code during the merge. It repeats the calibration and selects the lowest frequency value as this is probably the closest estimate to the real frequency 2) Monitor the delta of the TSC values in the delay loop which waits for the PIT counter to reach zero. If the maximum value is significantly different from the minimum, then we have a pretty safe indicator that the loop was disturbed by an SMI. 3) keep the pmtimer/hpet reference as a backup solution for systems where the SMI disturbance is a permanent point of failure for PIT based calibration 4) do the loop iteration for both methods, record the lowest value and decide after all iterations finished. 5) Set a clear preference to PIT based calibration when the result makes sense. The implementation does the reference calibration based on HPET/PMTIMER around the delay, which is necessary for the PIT anyway, but keeps separate TSC values to ensure the "independency" of the resulting calibration values. Tested on various 32bit/64bit machines including Geode 266Mhz, AMD K6 (affected machine with a double speed pmtimer which I grabbed out of the dump), Pentium class machines and AMD/Intel 64 bit boxen. Bisected-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-09-03 02:54:47 +04:00
return 0;
}
/* Use the alternative source */
printk(KERN_INFO "TSC: using %s reference calibration\n",
hpet ? "HPET" : "PMTIMER");
return tsc_ref_min;
}
[x86] Fix TSC calibration issues Larry Finger reported at http://lkml.org/lkml/2008/9/1/90: An ancient laptop of mine started throwing errors from b43legacy when I started using 2.6.27 on it. This has been bisected to commit bfc0f59 "x86: merge tsc calibration". The unification of the TSC code adopted mostly the 64bit code, which prefers PMTIMER/HPET over the PIT calibration. Larrys system has an AMD K6 CPU. Such systems are known to have PMTIMER incarnations which run at double speed. This results in a miscalibration of the TSC by factor 0.5. So the resulting calibrated CPU/TSC speed is half of the real CPU speed, which means that the TSC based delay loop will run half the time it should run. That might explain why the b43legacy driver went berserk. On the other hand we know about systems, where the PIT based calibration results in random crap due to heavy SMI/SMM disturbance. On those systems the PMTIMER/HPET based calibration logic with SMI detection shows better results. According to Alok also virtualized systems suffer from the PIT calibration method. The solution is to use a more wreckage aware aproach than the current either/or decision. 1) reimplement the retry loop which was dropped from the 32bit code during the merge. It repeats the calibration and selects the lowest frequency value as this is probably the closest estimate to the real frequency 2) Monitor the delta of the TSC values in the delay loop which waits for the PIT counter to reach zero. If the maximum value is significantly different from the minimum, then we have a pretty safe indicator that the loop was disturbed by an SMI. 3) keep the pmtimer/hpet reference as a backup solution for systems where the SMI disturbance is a permanent point of failure for PIT based calibration 4) do the loop iteration for both methods, record the lowest value and decide after all iterations finished. 5) Set a clear preference to PIT based calibration when the result makes sense. The implementation does the reference calibration based on HPET/PMTIMER around the delay, which is necessary for the PIT anyway, but keeps separate TSC values to ensure the "independency" of the resulting calibration values. Tested on various 32bit/64bit machines including Geode 266Mhz, AMD K6 (affected machine with a double speed pmtimer which I grabbed out of the dump), Pentium class machines and AMD/Intel 64 bit boxen. Bisected-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-09-03 02:54:47 +04:00
/* We don't have an alternative source, use the PIT calibration value */
if (!hpet && !ref1 && !ref2) {
[x86] Fix TSC calibration issues Larry Finger reported at http://lkml.org/lkml/2008/9/1/90: An ancient laptop of mine started throwing errors from b43legacy when I started using 2.6.27 on it. This has been bisected to commit bfc0f59 "x86: merge tsc calibration". The unification of the TSC code adopted mostly the 64bit code, which prefers PMTIMER/HPET over the PIT calibration. Larrys system has an AMD K6 CPU. Such systems are known to have PMTIMER incarnations which run at double speed. This results in a miscalibration of the TSC by factor 0.5. So the resulting calibrated CPU/TSC speed is half of the real CPU speed, which means that the TSC based delay loop will run half the time it should run. That might explain why the b43legacy driver went berserk. On the other hand we know about systems, where the PIT based calibration results in random crap due to heavy SMI/SMM disturbance. On those systems the PMTIMER/HPET based calibration logic with SMI detection shows better results. According to Alok also virtualized systems suffer from the PIT calibration method. The solution is to use a more wreckage aware aproach than the current either/or decision. 1) reimplement the retry loop which was dropped from the 32bit code during the merge. It repeats the calibration and selects the lowest frequency value as this is probably the closest estimate to the real frequency 2) Monitor the delta of the TSC values in the delay loop which waits for the PIT counter to reach zero. If the maximum value is significantly different from the minimum, then we have a pretty safe indicator that the loop was disturbed by an SMI. 3) keep the pmtimer/hpet reference as a backup solution for systems where the SMI disturbance is a permanent point of failure for PIT based calibration 4) do the loop iteration for both methods, record the lowest value and decide after all iterations finished. 5) Set a clear preference to PIT based calibration when the result makes sense. The implementation does the reference calibration based on HPET/PMTIMER around the delay, which is necessary for the PIT anyway, but keeps separate TSC values to ensure the "independency" of the resulting calibration values. Tested on various 32bit/64bit machines including Geode 266Mhz, AMD K6 (affected machine with a double speed pmtimer which I grabbed out of the dump), Pentium class machines and AMD/Intel 64 bit boxen. Bisected-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-09-03 02:54:47 +04:00
printk(KERN_INFO "TSC: Using PIT calibration value\n");
return tsc_pit_min;
}
[x86] Fix TSC calibration issues Larry Finger reported at http://lkml.org/lkml/2008/9/1/90: An ancient laptop of mine started throwing errors from b43legacy when I started using 2.6.27 on it. This has been bisected to commit bfc0f59 "x86: merge tsc calibration". The unification of the TSC code adopted mostly the 64bit code, which prefers PMTIMER/HPET over the PIT calibration. Larrys system has an AMD K6 CPU. Such systems are known to have PMTIMER incarnations which run at double speed. This results in a miscalibration of the TSC by factor 0.5. So the resulting calibrated CPU/TSC speed is half of the real CPU speed, which means that the TSC based delay loop will run half the time it should run. That might explain why the b43legacy driver went berserk. On the other hand we know about systems, where the PIT based calibration results in random crap due to heavy SMI/SMM disturbance. On those systems the PMTIMER/HPET based calibration logic with SMI detection shows better results. According to Alok also virtualized systems suffer from the PIT calibration method. The solution is to use a more wreckage aware aproach than the current either/or decision. 1) reimplement the retry loop which was dropped from the 32bit code during the merge. It repeats the calibration and selects the lowest frequency value as this is probably the closest estimate to the real frequency 2) Monitor the delta of the TSC values in the delay loop which waits for the PIT counter to reach zero. If the maximum value is significantly different from the minimum, then we have a pretty safe indicator that the loop was disturbed by an SMI. 3) keep the pmtimer/hpet reference as a backup solution for systems where the SMI disturbance is a permanent point of failure for PIT based calibration 4) do the loop iteration for both methods, record the lowest value and decide after all iterations finished. 5) Set a clear preference to PIT based calibration when the result makes sense. The implementation does the reference calibration based on HPET/PMTIMER around the delay, which is necessary for the PIT anyway, but keeps separate TSC values to ensure the "independency" of the resulting calibration values. Tested on various 32bit/64bit machines including Geode 266Mhz, AMD K6 (affected machine with a double speed pmtimer which I grabbed out of the dump), Pentium class machines and AMD/Intel 64 bit boxen. Bisected-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-09-03 02:54:47 +04:00
/* The alternative source failed, use the PIT calibration value */
if (tsc_ref_min == ULONG_MAX) {
printk(KERN_WARNING "TSC: HPET/PMTIMER calibration failed. "
"Using PIT calibration\n");
[x86] Fix TSC calibration issues Larry Finger reported at http://lkml.org/lkml/2008/9/1/90: An ancient laptop of mine started throwing errors from b43legacy when I started using 2.6.27 on it. This has been bisected to commit bfc0f59 "x86: merge tsc calibration". The unification of the TSC code adopted mostly the 64bit code, which prefers PMTIMER/HPET over the PIT calibration. Larrys system has an AMD K6 CPU. Such systems are known to have PMTIMER incarnations which run at double speed. This results in a miscalibration of the TSC by factor 0.5. So the resulting calibrated CPU/TSC speed is half of the real CPU speed, which means that the TSC based delay loop will run half the time it should run. That might explain why the b43legacy driver went berserk. On the other hand we know about systems, where the PIT based calibration results in random crap due to heavy SMI/SMM disturbance. On those systems the PMTIMER/HPET based calibration logic with SMI detection shows better results. According to Alok also virtualized systems suffer from the PIT calibration method. The solution is to use a more wreckage aware aproach than the current either/or decision. 1) reimplement the retry loop which was dropped from the 32bit code during the merge. It repeats the calibration and selects the lowest frequency value as this is probably the closest estimate to the real frequency 2) Monitor the delta of the TSC values in the delay loop which waits for the PIT counter to reach zero. If the maximum value is significantly different from the minimum, then we have a pretty safe indicator that the loop was disturbed by an SMI. 3) keep the pmtimer/hpet reference as a backup solution for systems where the SMI disturbance is a permanent point of failure for PIT based calibration 4) do the loop iteration for both methods, record the lowest value and decide after all iterations finished. 5) Set a clear preference to PIT based calibration when the result makes sense. The implementation does the reference calibration based on HPET/PMTIMER around the delay, which is necessary for the PIT anyway, but keeps separate TSC values to ensure the "independency" of the resulting calibration values. Tested on various 32bit/64bit machines including Geode 266Mhz, AMD K6 (affected machine with a double speed pmtimer which I grabbed out of the dump), Pentium class machines and AMD/Intel 64 bit boxen. Bisected-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-09-03 02:54:47 +04:00
return tsc_pit_min;
}
[x86] Fix TSC calibration issues Larry Finger reported at http://lkml.org/lkml/2008/9/1/90: An ancient laptop of mine started throwing errors from b43legacy when I started using 2.6.27 on it. This has been bisected to commit bfc0f59 "x86: merge tsc calibration". The unification of the TSC code adopted mostly the 64bit code, which prefers PMTIMER/HPET over the PIT calibration. Larrys system has an AMD K6 CPU. Such systems are known to have PMTIMER incarnations which run at double speed. This results in a miscalibration of the TSC by factor 0.5. So the resulting calibrated CPU/TSC speed is half of the real CPU speed, which means that the TSC based delay loop will run half the time it should run. That might explain why the b43legacy driver went berserk. On the other hand we know about systems, where the PIT based calibration results in random crap due to heavy SMI/SMM disturbance. On those systems the PMTIMER/HPET based calibration logic with SMI detection shows better results. According to Alok also virtualized systems suffer from the PIT calibration method. The solution is to use a more wreckage aware aproach than the current either/or decision. 1) reimplement the retry loop which was dropped from the 32bit code during the merge. It repeats the calibration and selects the lowest frequency value as this is probably the closest estimate to the real frequency 2) Monitor the delta of the TSC values in the delay loop which waits for the PIT counter to reach zero. If the maximum value is significantly different from the minimum, then we have a pretty safe indicator that the loop was disturbed by an SMI. 3) keep the pmtimer/hpet reference as a backup solution for systems where the SMI disturbance is a permanent point of failure for PIT based calibration 4) do the loop iteration for both methods, record the lowest value and decide after all iterations finished. 5) Set a clear preference to PIT based calibration when the result makes sense. The implementation does the reference calibration based on HPET/PMTIMER around the delay, which is necessary for the PIT anyway, but keeps separate TSC values to ensure the "independency" of the resulting calibration values. Tested on various 32bit/64bit machines including Geode 266Mhz, AMD K6 (affected machine with a double speed pmtimer which I grabbed out of the dump), Pentium class machines and AMD/Intel 64 bit boxen. Bisected-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-09-03 02:54:47 +04:00
/*
* The calibration values differ too much. In doubt, we use
* the PIT value as we know that there are PMTIMERs around
* running at double speed. At least we let the user know:
[x86] Fix TSC calibration issues Larry Finger reported at http://lkml.org/lkml/2008/9/1/90: An ancient laptop of mine started throwing errors from b43legacy when I started using 2.6.27 on it. This has been bisected to commit bfc0f59 "x86: merge tsc calibration". The unification of the TSC code adopted mostly the 64bit code, which prefers PMTIMER/HPET over the PIT calibration. Larrys system has an AMD K6 CPU. Such systems are known to have PMTIMER incarnations which run at double speed. This results in a miscalibration of the TSC by factor 0.5. So the resulting calibrated CPU/TSC speed is half of the real CPU speed, which means that the TSC based delay loop will run half the time it should run. That might explain why the b43legacy driver went berserk. On the other hand we know about systems, where the PIT based calibration results in random crap due to heavy SMI/SMM disturbance. On those systems the PMTIMER/HPET based calibration logic with SMI detection shows better results. According to Alok also virtualized systems suffer from the PIT calibration method. The solution is to use a more wreckage aware aproach than the current either/or decision. 1) reimplement the retry loop which was dropped from the 32bit code during the merge. It repeats the calibration and selects the lowest frequency value as this is probably the closest estimate to the real frequency 2) Monitor the delta of the TSC values in the delay loop which waits for the PIT counter to reach zero. If the maximum value is significantly different from the minimum, then we have a pretty safe indicator that the loop was disturbed by an SMI. 3) keep the pmtimer/hpet reference as a backup solution for systems where the SMI disturbance is a permanent point of failure for PIT based calibration 4) do the loop iteration for both methods, record the lowest value and decide after all iterations finished. 5) Set a clear preference to PIT based calibration when the result makes sense. The implementation does the reference calibration based on HPET/PMTIMER around the delay, which is necessary for the PIT anyway, but keeps separate TSC values to ensure the "independency" of the resulting calibration values. Tested on various 32bit/64bit machines including Geode 266Mhz, AMD K6 (affected machine with a double speed pmtimer which I grabbed out of the dump), Pentium class machines and AMD/Intel 64 bit boxen. Bisected-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-09-03 02:54:47 +04:00
*/
printk(KERN_WARNING "TSC: PIT calibration deviates from %s: %lu %lu.\n",
hpet ? "HPET" : "PMTIMER", tsc_pit_min, tsc_ref_min);
[x86] Fix TSC calibration issues Larry Finger reported at http://lkml.org/lkml/2008/9/1/90: An ancient laptop of mine started throwing errors from b43legacy when I started using 2.6.27 on it. This has been bisected to commit bfc0f59 "x86: merge tsc calibration". The unification of the TSC code adopted mostly the 64bit code, which prefers PMTIMER/HPET over the PIT calibration. Larrys system has an AMD K6 CPU. Such systems are known to have PMTIMER incarnations which run at double speed. This results in a miscalibration of the TSC by factor 0.5. So the resulting calibrated CPU/TSC speed is half of the real CPU speed, which means that the TSC based delay loop will run half the time it should run. That might explain why the b43legacy driver went berserk. On the other hand we know about systems, where the PIT based calibration results in random crap due to heavy SMI/SMM disturbance. On those systems the PMTIMER/HPET based calibration logic with SMI detection shows better results. According to Alok also virtualized systems suffer from the PIT calibration method. The solution is to use a more wreckage aware aproach than the current either/or decision. 1) reimplement the retry loop which was dropped from the 32bit code during the merge. It repeats the calibration and selects the lowest frequency value as this is probably the closest estimate to the real frequency 2) Monitor the delta of the TSC values in the delay loop which waits for the PIT counter to reach zero. If the maximum value is significantly different from the minimum, then we have a pretty safe indicator that the loop was disturbed by an SMI. 3) keep the pmtimer/hpet reference as a backup solution for systems where the SMI disturbance is a permanent point of failure for PIT based calibration 4) do the loop iteration for both methods, record the lowest value and decide after all iterations finished. 5) Set a clear preference to PIT based calibration when the result makes sense. The implementation does the reference calibration based on HPET/PMTIMER around the delay, which is necessary for the PIT anyway, but keeps separate TSC values to ensure the "independency" of the resulting calibration values. Tested on various 32bit/64bit machines including Geode 266Mhz, AMD K6 (affected machine with a double speed pmtimer which I grabbed out of the dump), Pentium class machines and AMD/Intel 64 bit boxen. Bisected-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Tested-by: Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-09-03 02:54:47 +04:00
printk(KERN_INFO "TSC: Using PIT calibration value\n");
return tsc_pit_min;
}
int recalibrate_cpu_khz(void)
{
#ifndef CONFIG_SMP
unsigned long cpu_khz_old = cpu_khz;
if (cpu_has_tsc) {
tsc_khz = x86_platform.calibrate_tsc();
cpu_khz = tsc_khz;
cpu_data(0).loops_per_jiffy =
cpufreq_scale(cpu_data(0).loops_per_jiffy,
cpu_khz_old, cpu_khz);
return 0;
} else
return -ENODEV;
#else
return -ENODEV;
#endif
}
EXPORT_SYMBOL(recalibrate_cpu_khz);
/* Accelerators for sched_clock()
* convert from cycles(64bits) => nanoseconds (64bits)
* basic equation:
* ns = cycles / (freq / ns_per_sec)
* ns = cycles * (ns_per_sec / freq)
* ns = cycles * (10^9 / (cpu_khz * 10^3))
* ns = cycles * (10^6 / cpu_khz)
*
* Then we use scaling math (suggested by george@mvista.com) to get:
* ns = cycles * (10^6 * SC / cpu_khz) / SC
* ns = cycles * cyc2ns_scale / SC
*
* And since SC is a constant power of two, we can convert the div
* into a shift.
*
* We can use khz divisor instead of mhz to keep a better precision, since
* cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits.
* (mathieu.desnoyers@polymtl.ca)
*
* -johnstul@us.ibm.com "math is hard, lets go shopping!"
*/
DEFINE_PER_CPU(unsigned long, cyc2ns);
DEFINE_PER_CPU(unsigned long long, cyc2ns_offset);
static void set_cyc2ns_scale(unsigned long cpu_khz, int cpu)
{
unsigned long long tsc_now, ns_now, *offset;
unsigned long flags, *scale;
local_irq_save(flags);
sched_clock_idle_sleep_event();
scale = &per_cpu(cyc2ns, cpu);
offset = &per_cpu(cyc2ns_offset, cpu);
rdtscll(tsc_now);
ns_now = __cycles_2_ns(tsc_now);
if (cpu_khz) {
*scale = (NSEC_PER_MSEC << CYC2NS_SCALE_FACTOR)/cpu_khz;
*offset = ns_now - (tsc_now * *scale >> CYC2NS_SCALE_FACTOR);
}
sched_clock_idle_wakeup_event(0);
local_irq_restore(flags);
}
static unsigned long long cyc2ns_suspend;
void save_sched_clock_state(void)
{
if (!sched_clock_stable)
return;
cyc2ns_suspend = sched_clock();
}
/*
* Even on processors with invariant TSC, TSC gets reset in some the
* ACPI system sleep states. And in some systems BIOS seem to reinit TSC to
* arbitrary value (still sync'd across cpu's) during resume from such sleep
* states. To cope up with this, recompute the cyc2ns_offset for each cpu so
* that sched_clock() continues from the point where it was left off during
* suspend.
*/
void restore_sched_clock_state(void)
{
unsigned long long offset;
unsigned long flags;
int cpu;
if (!sched_clock_stable)
return;
local_irq_save(flags);
__this_cpu_write(cyc2ns_offset, 0);
offset = cyc2ns_suspend - sched_clock();
for_each_possible_cpu(cpu)
per_cpu(cyc2ns_offset, cpu) = offset;
local_irq_restore(flags);
}
#ifdef CONFIG_CPU_FREQ
/* Frequency scaling support. Adjust the TSC based timer when the cpu frequency
* changes.
*
* RED-PEN: On SMP we assume all CPUs run with the same frequency. It's
* not that important because current Opteron setups do not support
* scaling on SMP anyroads.
*
* Should fix up last_tsc too. Currently gettimeofday in the
* first tick after the change will be slightly wrong.
*/
static unsigned int ref_freq;
static unsigned long loops_per_jiffy_ref;
static unsigned long tsc_khz_ref;
static int time_cpufreq_notifier(struct notifier_block *nb, unsigned long val,
void *data)
{
struct cpufreq_freqs *freq = data;
unsigned long *lpj;
if (cpu_has(&cpu_data(freq->cpu), X86_FEATURE_CONSTANT_TSC))
return 0;
lpj = &boot_cpu_data.loops_per_jiffy;
#ifdef CONFIG_SMP
if (!(freq->flags & CPUFREQ_CONST_LOOPS))
lpj = &cpu_data(freq->cpu).loops_per_jiffy;
#endif
if (!ref_freq) {
ref_freq = freq->old;
loops_per_jiffy_ref = *lpj;
tsc_khz_ref = tsc_khz;
}
if ((val == CPUFREQ_PRECHANGE && freq->old < freq->new) ||
(val == CPUFREQ_POSTCHANGE && freq->old > freq->new) ||
(val == CPUFREQ_RESUMECHANGE)) {
*lpj = cpufreq_scale(loops_per_jiffy_ref, ref_freq, freq->new);
tsc_khz = cpufreq_scale(tsc_khz_ref, ref_freq, freq->new);
if (!(freq->flags & CPUFREQ_CONST_LOOPS))
mark_tsc_unstable("cpufreq changes");
}
set_cyc2ns_scale(tsc_khz, freq->cpu);
return 0;
}
static struct notifier_block time_cpufreq_notifier_block = {
.notifier_call = time_cpufreq_notifier
};
static int __init cpufreq_tsc(void)
{
if (!cpu_has_tsc)
return 0;
if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
return 0;
cpufreq_register_notifier(&time_cpufreq_notifier_block,
CPUFREQ_TRANSITION_NOTIFIER);
return 0;
}
core_initcall(cpufreq_tsc);
#endif /* CONFIG_CPU_FREQ */
/* clocksource code */
static struct clocksource clocksource_tsc;
/*
* We compare the TSC to the cycle_last value in the clocksource
* structure to avoid a nasty time-warp. This can be observed in a
* very small window right after one CPU updated cycle_last under
* xtime/vsyscall_gtod lock and the other CPU reads a TSC value which
* is smaller than the cycle_last reference value due to a TSC which
* is slighty behind. This delta is nowhere else observable, but in
* that case it results in a forward time jump in the range of hours
* due to the unsigned delta calculation of the time keeping core
* code, which is necessary to support wrapping clocksources like pm
* timer.
*/
static cycle_t read_tsc(struct clocksource *cs)
{
cycle_t ret = (cycle_t)get_cycles();
return ret >= clocksource_tsc.cycle_last ?
ret : clocksource_tsc.cycle_last;
}
static void resume_tsc(struct clocksource *cs)
{
clocksource_tsc.cycle_last = 0;
}
static struct clocksource clocksource_tsc = {
.name = "tsc",
.rating = 300,
.read = read_tsc,
.resume = resume_tsc,
.mask = CLOCKSOURCE_MASK(64),
.flags = CLOCK_SOURCE_IS_CONTINUOUS |
CLOCK_SOURCE_MUST_VERIFY,
#ifdef CONFIG_X86_64
.archdata = { .vclock_mode = VCLOCK_TSC },
#endif
};
void mark_tsc_unstable(char *reason)
{
if (!tsc_unstable) {
tsc_unstable = 1;
sched_clock_stable = 0;
disable_sched_clock_irqtime();
printk(KERN_INFO "Marking TSC unstable due to %s\n", reason);
/* Change only the rating, when not registered */
if (clocksource_tsc.mult)
clocksource_mark_unstable(&clocksource_tsc);
else {
clocksource_tsc.flags |= CLOCK_SOURCE_UNSTABLE;
clocksource_tsc.rating = 0;
}
}
}
EXPORT_SYMBOL_GPL(mark_tsc_unstable);
x86: Skip verification by the watchdog for TSC clocksource. Impact: Changes timekeeping on Vmware (or with tsc=reliable). This is achieved by resetting the CLOCKSOURCE_MUST_VERIFY flag. We add a tsc=reliable commandline option to enable this. This enables legacy hardware without HPET, LAPIC, or ACPI timers to enter high-resolution timer mode. Along with that have extended this to be used in virtualization environement too. Now we also set this flag if the X86_FEATURE_TSC_RELIABLE bit is set. This is important since there is a wrap-around problem with the acpi_pm timer. The acpi_pm counter is just 24bits and this can overflow in ~4 seconds. With the NO_HZ kernels in virtualized environment, there can be situations when the guest is descheduled for longer duration, as a result we may miss the wrap of the acpi counter. When TSC is used as a clocksource and acpi_pm timer is being used as the watchdog clocksource this error in acpi_pm results in TSC being marked as unstable, and essentially results in time dropping in chunks of 4 seconds whenever this wrap is missed. Since the virtualized TSC is reliable on VMware, we should always use the TSCs clocksource on VMware, so we skip the verfication at runtime, by checking for the feature bit. Since we reset the flag for mgeode systems too, i have combined the mgeode case with the feature bit check. Signed-off-by: Jeff Hansen <jhansen@cardaccess-inc.com> Signed-off-by: Alok N Kataria <akataria@vmware.com> Signed-off-by: Dan Hecht <dhecht@vmware.com> Signed-off-by: H. Peter Anvin <hpa@zytor.com>
2008-10-25 04:22:01 +04:00
static void __init check_system_tsc_reliable(void)
{
#ifdef CONFIG_MGEODE_LX
x86: Skip verification by the watchdog for TSC clocksource. Impact: Changes timekeeping on Vmware (or with tsc=reliable). This is achieved by resetting the CLOCKSOURCE_MUST_VERIFY flag. We add a tsc=reliable commandline option to enable this. This enables legacy hardware without HPET, LAPIC, or ACPI timers to enter high-resolution timer mode. Along with that have extended this to be used in virtualization environement too. Now we also set this flag if the X86_FEATURE_TSC_RELIABLE bit is set. This is important since there is a wrap-around problem with the acpi_pm timer. The acpi_pm counter is just 24bits and this can overflow in ~4 seconds. With the NO_HZ kernels in virtualized environment, there can be situations when the guest is descheduled for longer duration, as a result we may miss the wrap of the acpi counter. When TSC is used as a clocksource and acpi_pm timer is being used as the watchdog clocksource this error in acpi_pm results in TSC being marked as unstable, and essentially results in time dropping in chunks of 4 seconds whenever this wrap is missed. Since the virtualized TSC is reliable on VMware, we should always use the TSCs clocksource on VMware, so we skip the verfication at runtime, by checking for the feature bit. Since we reset the flag for mgeode systems too, i have combined the mgeode case with the feature bit check. Signed-off-by: Jeff Hansen <jhansen@cardaccess-inc.com> Signed-off-by: Alok N Kataria <akataria@vmware.com> Signed-off-by: Dan Hecht <dhecht@vmware.com> Signed-off-by: H. Peter Anvin <hpa@zytor.com>
2008-10-25 04:22:01 +04:00
/* RTSC counts during suspend */
#define RTSC_SUSP 0x100
unsigned long res_low, res_high;
rdmsr_safe(MSR_GEODE_BUSCONT_CONF0, &res_low, &res_high);
/* Geode_LX - the OLPC CPU has a very reliable TSC */
if (res_low & RTSC_SUSP)
x86: Skip verification by the watchdog for TSC clocksource. Impact: Changes timekeeping on Vmware (or with tsc=reliable). This is achieved by resetting the CLOCKSOURCE_MUST_VERIFY flag. We add a tsc=reliable commandline option to enable this. This enables legacy hardware without HPET, LAPIC, or ACPI timers to enter high-resolution timer mode. Along with that have extended this to be used in virtualization environement too. Now we also set this flag if the X86_FEATURE_TSC_RELIABLE bit is set. This is important since there is a wrap-around problem with the acpi_pm timer. The acpi_pm counter is just 24bits and this can overflow in ~4 seconds. With the NO_HZ kernels in virtualized environment, there can be situations when the guest is descheduled for longer duration, as a result we may miss the wrap of the acpi counter. When TSC is used as a clocksource and acpi_pm timer is being used as the watchdog clocksource this error in acpi_pm results in TSC being marked as unstable, and essentially results in time dropping in chunks of 4 seconds whenever this wrap is missed. Since the virtualized TSC is reliable on VMware, we should always use the TSCs clocksource on VMware, so we skip the verfication at runtime, by checking for the feature bit. Since we reset the flag for mgeode systems too, i have combined the mgeode case with the feature bit check. Signed-off-by: Jeff Hansen <jhansen@cardaccess-inc.com> Signed-off-by: Alok N Kataria <akataria@vmware.com> Signed-off-by: Dan Hecht <dhecht@vmware.com> Signed-off-by: H. Peter Anvin <hpa@zytor.com>
2008-10-25 04:22:01 +04:00
tsc_clocksource_reliable = 1;
#endif
x86: Skip verification by the watchdog for TSC clocksource. Impact: Changes timekeeping on Vmware (or with tsc=reliable). This is achieved by resetting the CLOCKSOURCE_MUST_VERIFY flag. We add a tsc=reliable commandline option to enable this. This enables legacy hardware without HPET, LAPIC, or ACPI timers to enter high-resolution timer mode. Along with that have extended this to be used in virtualization environement too. Now we also set this flag if the X86_FEATURE_TSC_RELIABLE bit is set. This is important since there is a wrap-around problem with the acpi_pm timer. The acpi_pm counter is just 24bits and this can overflow in ~4 seconds. With the NO_HZ kernels in virtualized environment, there can be situations when the guest is descheduled for longer duration, as a result we may miss the wrap of the acpi counter. When TSC is used as a clocksource and acpi_pm timer is being used as the watchdog clocksource this error in acpi_pm results in TSC being marked as unstable, and essentially results in time dropping in chunks of 4 seconds whenever this wrap is missed. Since the virtualized TSC is reliable on VMware, we should always use the TSCs clocksource on VMware, so we skip the verfication at runtime, by checking for the feature bit. Since we reset the flag for mgeode systems too, i have combined the mgeode case with the feature bit check. Signed-off-by: Jeff Hansen <jhansen@cardaccess-inc.com> Signed-off-by: Alok N Kataria <akataria@vmware.com> Signed-off-by: Dan Hecht <dhecht@vmware.com> Signed-off-by: H. Peter Anvin <hpa@zytor.com>
2008-10-25 04:22:01 +04:00
if (boot_cpu_has(X86_FEATURE_TSC_RELIABLE))
tsc_clocksource_reliable = 1;
}
/*
* Make an educated guess if the TSC is trustworthy and synchronized
* over all CPUs.
*/
__cpuinit int unsynchronized_tsc(void)
{
if (!cpu_has_tsc || tsc_unstable)
return 1;
#ifdef CONFIG_SMP
if (apic_is_clustered_box())
return 1;
#endif
if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
return 0;
if (tsc_clocksource_reliable)
return 0;
/*
* Intel systems are normally all synchronized.
* Exceptions must mark TSC as unstable:
*/
if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL) {
/* assume multi socket systems are not synchronized: */
if (num_possible_cpus() > 1)
return 1;
}
return 0;
}
x86: Improve TSC calibration using a delayed workqueue Boot to boot the TSC calibration may vary by quite a large amount. While normal variance of 50-100ppm can easily be seen, the quick calibration code only requires 500ppm accuracy, which is the limit of what NTP can correct for. This can cause problems for systems being used as NTP servers, as every time they reboot it can take hours for them to calculate the new drift error caused by the calibration. The classic trade-off here is calibration accuracy vs slow boot times, as during the calibration nothing else can run. This patch uses a delayed workqueue to calibrate the TSC over the period of a second. This allows very accurate calibration (in my tests only varying by 1khz or 0.4ppm boot to boot). Additionally this refined calibration step does not block the boot process, and only delays the TSC clocksoure registration by a few seconds in early boot. If the refined calibration strays 1% from the early boot calibration value, the system will fall back to already calculated early boot calibration. Credit to Andi Kleen who suggested using a timer quite awhile back, but I dismissed it thinking the timer calibration would be done after the clocksource was registered (which would break things). Forgive me for my short-sightedness. This patch has worked very well in my testing, but TSC hardware is quite varied so it would probably be good to get some extended testing, possibly pushing inclusion out to 2.6.39. Signed-off-by: John Stultz <johnstul@us.ibm.com> LKML-Reference: <1289003985-29060-1-git-send-email-johnstul@us.ibm.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> CC: Thomas Gleixner <tglx@linutronix.de> CC: Ingo Molnar <mingo@elte.hu> CC: Martin Schwidefsky <schwidefsky@de.ibm.com> CC: Clark Williams <williams@redhat.com> CC: Andi Kleen <andi@firstfloor.org>
2010-07-28 04:00:00 +04:00
static void tsc_refine_calibration_work(struct work_struct *work);
static DECLARE_DELAYED_WORK(tsc_irqwork, tsc_refine_calibration_work);
/**
* tsc_refine_calibration_work - Further refine tsc freq calibration
* @work - ignored.
*
* This functions uses delayed work over a period of a
* second to further refine the TSC freq value. Since this is
* timer based, instead of loop based, we don't block the boot
* process while this longer calibration is done.
*
* If there are any calibration anomalies (too many SMIs, etc),
x86: Improve TSC calibration using a delayed workqueue Boot to boot the TSC calibration may vary by quite a large amount. While normal variance of 50-100ppm can easily be seen, the quick calibration code only requires 500ppm accuracy, which is the limit of what NTP can correct for. This can cause problems for systems being used as NTP servers, as every time they reboot it can take hours for them to calculate the new drift error caused by the calibration. The classic trade-off here is calibration accuracy vs slow boot times, as during the calibration nothing else can run. This patch uses a delayed workqueue to calibrate the TSC over the period of a second. This allows very accurate calibration (in my tests only varying by 1khz or 0.4ppm boot to boot). Additionally this refined calibration step does not block the boot process, and only delays the TSC clocksoure registration by a few seconds in early boot. If the refined calibration strays 1% from the early boot calibration value, the system will fall back to already calculated early boot calibration. Credit to Andi Kleen who suggested using a timer quite awhile back, but I dismissed it thinking the timer calibration would be done after the clocksource was registered (which would break things). Forgive me for my short-sightedness. This patch has worked very well in my testing, but TSC hardware is quite varied so it would probably be good to get some extended testing, possibly pushing inclusion out to 2.6.39. Signed-off-by: John Stultz <johnstul@us.ibm.com> LKML-Reference: <1289003985-29060-1-git-send-email-johnstul@us.ibm.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> CC: Thomas Gleixner <tglx@linutronix.de> CC: Ingo Molnar <mingo@elte.hu> CC: Martin Schwidefsky <schwidefsky@de.ibm.com> CC: Clark Williams <williams@redhat.com> CC: Andi Kleen <andi@firstfloor.org>
2010-07-28 04:00:00 +04:00
* or the refined calibration is off by 1% of the fast early
* calibration, we throw out the new calibration and use the
* early calibration.
*/
static void tsc_refine_calibration_work(struct work_struct *work)
{
static u64 tsc_start = -1, ref_start;
static int hpet;
u64 tsc_stop, ref_stop, delta;
unsigned long freq;
/* Don't bother refining TSC on unstable systems */
if (check_tsc_unstable())
goto out;
/*
* Since the work is started early in boot, we may be
* delayed the first time we expire. So set the workqueue
* again once we know timers are working.
*/
if (tsc_start == -1) {
/*
* 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);
return;
}
tsc_stop = tsc_read_refs(&ref_stop, hpet);
/* hpet or pmtimer available ? */
if (ref_start == ref_stop)
x86: Improve TSC calibration using a delayed workqueue Boot to boot the TSC calibration may vary by quite a large amount. While normal variance of 50-100ppm can easily be seen, the quick calibration code only requires 500ppm accuracy, which is the limit of what NTP can correct for. This can cause problems for systems being used as NTP servers, as every time they reboot it can take hours for them to calculate the new drift error caused by the calibration. The classic trade-off here is calibration accuracy vs slow boot times, as during the calibration nothing else can run. This patch uses a delayed workqueue to calibrate the TSC over the period of a second. This allows very accurate calibration (in my tests only varying by 1khz or 0.4ppm boot to boot). Additionally this refined calibration step does not block the boot process, and only delays the TSC clocksoure registration by a few seconds in early boot. If the refined calibration strays 1% from the early boot calibration value, the system will fall back to already calculated early boot calibration. Credit to Andi Kleen who suggested using a timer quite awhile back, but I dismissed it thinking the timer calibration would be done after the clocksource was registered (which would break things). Forgive me for my short-sightedness. This patch has worked very well in my testing, but TSC hardware is quite varied so it would probably be good to get some extended testing, possibly pushing inclusion out to 2.6.39. Signed-off-by: John Stultz <johnstul@us.ibm.com> LKML-Reference: <1289003985-29060-1-git-send-email-johnstul@us.ibm.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> CC: Thomas Gleixner <tglx@linutronix.de> CC: Ingo Molnar <mingo@elte.hu> CC: Martin Schwidefsky <schwidefsky@de.ibm.com> CC: Clark Williams <williams@redhat.com> CC: Andi Kleen <andi@firstfloor.org>
2010-07-28 04:00:00 +04:00
goto out;
/* Check, whether the sampling was disturbed by an SMI */
if (tsc_start == ULLONG_MAX || tsc_stop == ULLONG_MAX)
goto out;
delta = tsc_stop - tsc_start;
delta *= 1000000LL;
if (hpet)
freq = calc_hpet_ref(delta, ref_start, ref_stop);
else
freq = calc_pmtimer_ref(delta, ref_start, ref_stop);
/* Make sure we're within 1% */
if (abs(tsc_khz - freq) > tsc_khz/100)
goto out;
tsc_khz = freq;
printk(KERN_INFO "Refined TSC clocksource calibration: "
"%lu.%03lu MHz.\n", (unsigned long)tsc_khz / 1000,
(unsigned long)tsc_khz % 1000);
out:
clocksource_register_khz(&clocksource_tsc, tsc_khz);
}
static int __init init_tsc_clocksource(void)
{
if (!cpu_has_tsc || tsc_disabled > 0 || !tsc_khz)
return 0;
x86: Skip verification by the watchdog for TSC clocksource. Impact: Changes timekeeping on Vmware (or with tsc=reliable). This is achieved by resetting the CLOCKSOURCE_MUST_VERIFY flag. We add a tsc=reliable commandline option to enable this. This enables legacy hardware without HPET, LAPIC, or ACPI timers to enter high-resolution timer mode. Along with that have extended this to be used in virtualization environement too. Now we also set this flag if the X86_FEATURE_TSC_RELIABLE bit is set. This is important since there is a wrap-around problem with the acpi_pm timer. The acpi_pm counter is just 24bits and this can overflow in ~4 seconds. With the NO_HZ kernels in virtualized environment, there can be situations when the guest is descheduled for longer duration, as a result we may miss the wrap of the acpi counter. When TSC is used as a clocksource and acpi_pm timer is being used as the watchdog clocksource this error in acpi_pm results in TSC being marked as unstable, and essentially results in time dropping in chunks of 4 seconds whenever this wrap is missed. Since the virtualized TSC is reliable on VMware, we should always use the TSCs clocksource on VMware, so we skip the verfication at runtime, by checking for the feature bit. Since we reset the flag for mgeode systems too, i have combined the mgeode case with the feature bit check. Signed-off-by: Jeff Hansen <jhansen@cardaccess-inc.com> Signed-off-by: Alok N Kataria <akataria@vmware.com> Signed-off-by: Dan Hecht <dhecht@vmware.com> Signed-off-by: H. Peter Anvin <hpa@zytor.com>
2008-10-25 04:22:01 +04:00
if (tsc_clocksource_reliable)
clocksource_tsc.flags &= ~CLOCK_SOURCE_MUST_VERIFY;
/* lower the rating if we already know its unstable: */
if (check_tsc_unstable()) {
clocksource_tsc.rating = 0;
clocksource_tsc.flags &= ~CLOCK_SOURCE_IS_CONTINUOUS;
}
x86: Improve TSC calibration using a delayed workqueue Boot to boot the TSC calibration may vary by quite a large amount. While normal variance of 50-100ppm can easily be seen, the quick calibration code only requires 500ppm accuracy, which is the limit of what NTP can correct for. This can cause problems for systems being used as NTP servers, as every time they reboot it can take hours for them to calculate the new drift error caused by the calibration. The classic trade-off here is calibration accuracy vs slow boot times, as during the calibration nothing else can run. This patch uses a delayed workqueue to calibrate the TSC over the period of a second. This allows very accurate calibration (in my tests only varying by 1khz or 0.4ppm boot to boot). Additionally this refined calibration step does not block the boot process, and only delays the TSC clocksoure registration by a few seconds in early boot. If the refined calibration strays 1% from the early boot calibration value, the system will fall back to already calculated early boot calibration. Credit to Andi Kleen who suggested using a timer quite awhile back, but I dismissed it thinking the timer calibration would be done after the clocksource was registered (which would break things). Forgive me for my short-sightedness. This patch has worked very well in my testing, but TSC hardware is quite varied so it would probably be good to get some extended testing, possibly pushing inclusion out to 2.6.39. Signed-off-by: John Stultz <johnstul@us.ibm.com> LKML-Reference: <1289003985-29060-1-git-send-email-johnstul@us.ibm.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> CC: Thomas Gleixner <tglx@linutronix.de> CC: Ingo Molnar <mingo@elte.hu> CC: Martin Schwidefsky <schwidefsky@de.ibm.com> CC: Clark Williams <williams@redhat.com> CC: Andi Kleen <andi@firstfloor.org>
2010-07-28 04:00:00 +04:00
schedule_delayed_work(&tsc_irqwork, 0);
return 0;
}
x86: Improve TSC calibration using a delayed workqueue Boot to boot the TSC calibration may vary by quite a large amount. While normal variance of 50-100ppm can easily be seen, the quick calibration code only requires 500ppm accuracy, which is the limit of what NTP can correct for. This can cause problems for systems being used as NTP servers, as every time they reboot it can take hours for them to calculate the new drift error caused by the calibration. The classic trade-off here is calibration accuracy vs slow boot times, as during the calibration nothing else can run. This patch uses a delayed workqueue to calibrate the TSC over the period of a second. This allows very accurate calibration (in my tests only varying by 1khz or 0.4ppm boot to boot). Additionally this refined calibration step does not block the boot process, and only delays the TSC clocksoure registration by a few seconds in early boot. If the refined calibration strays 1% from the early boot calibration value, the system will fall back to already calculated early boot calibration. Credit to Andi Kleen who suggested using a timer quite awhile back, but I dismissed it thinking the timer calibration would be done after the clocksource was registered (which would break things). Forgive me for my short-sightedness. This patch has worked very well in my testing, but TSC hardware is quite varied so it would probably be good to get some extended testing, possibly pushing inclusion out to 2.6.39. Signed-off-by: John Stultz <johnstul@us.ibm.com> LKML-Reference: <1289003985-29060-1-git-send-email-johnstul@us.ibm.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> CC: Thomas Gleixner <tglx@linutronix.de> CC: Ingo Molnar <mingo@elte.hu> CC: Martin Schwidefsky <schwidefsky@de.ibm.com> CC: Clark Williams <williams@redhat.com> CC: Andi Kleen <andi@firstfloor.org>
2010-07-28 04:00:00 +04:00
/*
* We use device_initcall here, to ensure we run after the hpet
* is fully initialized, which may occur at fs_initcall time.
*/
device_initcall(init_tsc_clocksource);
void __init tsc_init(void)
{
u64 lpj;
int cpu;
x86_init.timers.tsc_pre_init();
if (!cpu_has_tsc)
return;
tsc_khz = x86_platform.calibrate_tsc();
cpu_khz = tsc_khz;
if (!tsc_khz) {
mark_tsc_unstable("could not calculate TSC khz");
return;
}
printk("Detected %lu.%03lu MHz processor.\n",
(unsigned long)cpu_khz / 1000,
(unsigned long)cpu_khz % 1000);
/*
* Secondary CPUs do not run through tsc_init(), so set up
* all the scale factors for all CPUs, assuming the same
* speed as the bootup CPU. (cpufreq notifiers will fix this
* up if their speed diverges)
*/
for_each_possible_cpu(cpu)
set_cyc2ns_scale(cpu_khz, cpu);
if (tsc_disabled > 0)
return;
/* now allow native_sched_clock() to use rdtsc */
tsc_disabled = 0;
if (!no_sched_irq_time)
enable_sched_clock_irqtime();
lpj = ((u64)tsc_khz * 1000);
do_div(lpj, HZ);
lpj_fine = lpj;
use_tsc_delay();
if (unsynchronized_tsc())
mark_tsc_unstable("TSCs unsynchronized");
x86: Skip verification by the watchdog for TSC clocksource. Impact: Changes timekeeping on Vmware (or with tsc=reliable). This is achieved by resetting the CLOCKSOURCE_MUST_VERIFY flag. We add a tsc=reliable commandline option to enable this. This enables legacy hardware without HPET, LAPIC, or ACPI timers to enter high-resolution timer mode. Along with that have extended this to be used in virtualization environement too. Now we also set this flag if the X86_FEATURE_TSC_RELIABLE bit is set. This is important since there is a wrap-around problem with the acpi_pm timer. The acpi_pm counter is just 24bits and this can overflow in ~4 seconds. With the NO_HZ kernels in virtualized environment, there can be situations when the guest is descheduled for longer duration, as a result we may miss the wrap of the acpi counter. When TSC is used as a clocksource and acpi_pm timer is being used as the watchdog clocksource this error in acpi_pm results in TSC being marked as unstable, and essentially results in time dropping in chunks of 4 seconds whenever this wrap is missed. Since the virtualized TSC is reliable on VMware, we should always use the TSCs clocksource on VMware, so we skip the verfication at runtime, by checking for the feature bit. Since we reset the flag for mgeode systems too, i have combined the mgeode case with the feature bit check. Signed-off-by: Jeff Hansen <jhansen@cardaccess-inc.com> Signed-off-by: Alok N Kataria <akataria@vmware.com> Signed-off-by: Dan Hecht <dhecht@vmware.com> Signed-off-by: H. Peter Anvin <hpa@zytor.com>
2008-10-25 04:22:01 +04:00
check_system_tsc_reliable();
}
#ifdef CONFIG_SMP
/*
* If we have a constant TSC and are using the TSC for the delay loop,
* we can skip clock calibration if another cpu in the same socket has already
* been calibrated. This assumes that CONSTANT_TSC applies to all
* cpus in the socket - this should be a safe assumption.
*/
unsigned long __cpuinit calibrate_delay_is_known(void)
{
int i, cpu = smp_processor_id();
if (!tsc_disabled && !cpu_has(&cpu_data(cpu), X86_FEATURE_CONSTANT_TSC))
return 0;
for_each_online_cpu(i)
if (cpu_data(i).phys_proc_id == cpu_data(cpu).phys_proc_id)
return cpu_data(i).loops_per_jiffy;
return 0;
}
#endif