Time keeping Issues with the low-resolution TSC timecounter
Jung-uk Kim
jkim at FreeBSD.org
Tue Jun 21 20:53:51 UTC 2011
On Friday 17 June 2011 02:54 pm, Jung-uk Kim wrote:
> On Friday 17 June 2011 01:45 pm, Ian FREISLICH wrote:
> > Jung-uk Kim wrote:
> > > On Thursday 16 June 2011 03:10 am, Ian FREISLICH wrote:
> > > > Jung-uk Kim wrote:
> > > > > 1481522037 14459060 1.0098392393
> > > > > 1495969404 14447367 1.0090225853
> > > > >
> > > > > As you can see, HPET increases normally (within errors from
> > > > > sleep(3) accuracy, syscall overhead, etc.) but TSC-low is
> > > > > totally erratic (and too low). I don't know how this can
> > > > > happen, though.
> > > > >
> > > > > :-(
> > > > >
> > > > > I need some time to figure it out.
> > > >
> > > > Even though sleep states have been disabled in the past when
> > > > on AC power, they seem to have mysteriously been enabled.
> > > > Perhaps this accounts for the strangeness:
> > > >
> > > > /etc/rc.conf
> > > > performance_cx_lowest="HIGH"
> > > > performance_cpu_freq="HIGH"
> > > > economy_cx_lowest="LOW"
> > > > economy_cpu_freq="HIGH"
> > > >
> > > >
> > > > [mini] /usr/home/ianf $ sysctl dev.cpu
> > > > dev.cpu.0.%desc: ACPI CPU
> > > > dev.cpu.0.%driver: cpu
> > > > dev.cpu.0.%location: handle=\_PR_.CPU0
> > > > dev.cpu.0.%pnpinfo: _HID=none _UID=0
> > > > dev.cpu.0.%parent: acpi0
> > > > dev.cpu.0.freq: 1600
> > > > dev.cpu.0.freq_levels: 1600/2000 1400/1750 1333/1533
> > > > 1166/1341 1066/1066 932/932 800/600 700/525 600/450 500/375
> > > > 400/300 300/225 200/150 100/75 dev.cpu.0.cx_supported: C1/1
> > > > C2/1 C3/57 dev.cpu.0.cx_lowest: C3
> > > > dev.cpu.0.cx_usage: 0.00% 8.69% 91.30% last 693us
> > > > dev.cpu.1.%desc: ACPI CPU
> > > > dev.cpu.1.%driver: cpu
> > > > dev.cpu.1.%location: handle=\_PR_.CPU1
> > > > dev.cpu.1.%pnpinfo: _HID=none _UID=0
> > > > dev.cpu.1.%parent: acpi0
> > > > dev.cpu.1.cx_supported: C1/1 C2/1 C3/57
> > > > dev.cpu.1.cx_lowest: C3
> > > > dev.cpu.1.cx_usage: 0.00% 14.96% 85.03% last 2897us
> > > >
> > > > Pulling the power cord and re-inserting it has the cx_lowest
> > > > correctly trantsition to C1 and then TSC-low behaves properly
> > > > as the system timecounter. But, time will be wierd when on
> > > > battery.
> > > >
> > > > In light of this, I doubt the patch in your other email will
> > > > have any effect. Perhaps the thing to do is to have the
> > > > timecounter code aware of the lowest Cx sleep state and to
> > > > pick best time counter for that state and to re-evaluate the
> > > > choice on cx_lowest transitions.
> > > >
> > > > ie: TSC-low, HPET or ACPI-fast for C1 and HPET or ACPI-fast
> > > > for C2 and lower.
> > >
> > > Hmm... So, you are saying this CPU model is P-state invariant
> > > but not C-state invariant (i.e., it stops incrementing in C2
> > > state and above). If that's the case, it is really useless for
> > > timecounter. :-(
> > >
> > > What happens if you set it to C2, i.e.,
> > >
> > > economy_cx_lowest="C2"
> > >
> > > In other words, does it really stop in C2-state?
> >
> > The folowing is with timecounter=HPET, just to see what the
> > effect on TSC-low is. It looks like it does stop in C3.
> >
> > hw.acpi.cpu.cx_lowest=C3
> > [mini] /usr/home/ianf $ sh -c 'count=10; while [ $count -gt 0 ];
> > do count=$((count - 1)); sysctl
> > kern.timecounter.tc.TSC-low.counter; sleep 1; done'
> > kern.timecounter.tc.TSC-low.counter: 722687906
> > kern.timecounter.tc.TSC-low.counter: 724328394
> > kern.timecounter.tc.TSC-low.counter: 726038743
> > kern.timecounter.tc.TSC-low.counter: 727690855
> > kern.timecounter.tc.TSC-low.counter: 729245616
> > kern.timecounter.tc.TSC-low.counter: 730786569
> > kern.timecounter.tc.TSC-low.counter: 732398571
> > kern.timecounter.tc.TSC-low.counter: 733910987
> > kern.timecounter.tc.TSC-low.counter: 735711469
> > kern.timecounter.tc.TSC-low.counter: 737368279
> >
> > hw.acpi.cpu.cx_lowest=C2
> > kern.timecounter.tc.TSC-low.counter: 897318486
> > kern.timecounter.tc.TSC-low.counter: 909873821
> > kern.timecounter.tc.TSC-low.counter: 922416894
> > kern.timecounter.tc.TSC-low.counter: 934960462
> > kern.timecounter.tc.TSC-low.counter: 947504154
> > kern.timecounter.tc.TSC-low.counter: 960050573
> > kern.timecounter.tc.TSC-low.counter: 972590754
> > kern.timecounter.tc.TSC-low.counter: 985133990
> > kern.timecounter.tc.TSC-low.counter: 997677052
> > kern.timecounter.tc.TSC-low.counter: 1010220299
> >
> > CPU: Intel(R) Atom(TM) CPU N270 @ 1.60GHz (1596.04-MHz
> > 686-class CPU) Origin = "GenuineIntel" Id = 0x106c2 Family = 6
> > Model = 1c Stepping = 2
> > Features=0xbfe9fbff<FPU,VME,DE,PSE,TSC,MSR,PAE,MCE,CX8,APIC,SEP,M
> >TR
> > R,PGE,MCA,CMOV,PAT,CLFLUSH,DTS,ACPI,MMX,FXSR,SSE,SSE2,SS,HTT,TM,P
> >BE>
> > Features2=0x40c39d<SSE3,DTES64,MON,DS_CPL,EST,TM2,SSSE3,xTPR,PDCM
> >,M OVBE> AMD Features2=0x1<LAHF>
> > TSC: P-state invariant, performance statistics
>
> Thanks for the info, it confirmed my speculation.
>
> Somewhere from an Intel manual, I think I read TSC stops when
> DPSLP# pin is asserted for Core/Core2/Atom processors and I guess
> that means entering C3 stops TSC. :-(
Can you please try the attached patch? It should disable TSC/TSC-low
timecounter for your CPU models, I think.
Sorry for the delay,
Jung-uk Kim
-------------- next part --------------
Index: sys/dev/acpica/acpi_cpu.c
===================================================================
--- sys/dev/acpica/acpi_cpu.c (revision 223372)
+++ sys/dev/acpica/acpi_cpu.c (working copy)
@@ -856,6 +856,8 @@ acpi_cpu_cx_list(struct acpi_cpu_softc *sc)
sbuf_printf(&sb, "C%d/%d ", i + 1, sc->cpu_cx_states[i].trans_lat);
if (sc->cpu_cx_states[i].type < ACPI_STATE_C3)
sc->cpu_non_c3 = i;
+ else
+ cpu_can_deep_sleep = 1;
}
sbuf_trim(&sb);
sbuf_finish(&sb);
Index: sys/kern/kern_clocksource.c
===================================================================
--- sys/kern/kern_clocksource.c (revision 223372)
+++ sys/kern/kern_clocksource.c (working copy)
@@ -59,6 +59,7 @@ __FBSDID("$FreeBSD$");
cyclic_clock_func_t cyclic_clock_func = NULL;
#endif
+int cpu_can_deep_sleep = 0; /* C3 state is available. */
int cpu_disable_deep_sleep = 0; /* Timer dies in C3. */
static void setuptimer(void);
Index: sys/sys/systm.h
===================================================================
--- sys/sys/systm.h (revision 223372)
+++ sys/sys/systm.h (working copy)
@@ -253,6 +253,7 @@ void cpu_startprofclock(void);
void cpu_stopprofclock(void);
void cpu_idleclock(void);
void cpu_activeclock(void);
+extern int cpu_can_deep_sleep;
extern int cpu_disable_deep_sleep;
int cr_cansee(struct ucred *u1, struct ucred *u2);
Index: sys/x86/x86/tsc.c
===================================================================
--- sys/x86/x86/tsc.c (revision 223372)
+++ sys/x86/x86/tsc.c (working copy)
@@ -444,6 +444,19 @@ init_TSC_tc(void)
goto init;
}
+ /*
+ * We cannot use the TSC if it stops incrementing in deep sleep.
+ * Currently only Intel CPUs are known for this problem unless
+ * the invariant TSC bit is set.
+ */
+ if (cpu_vendor_id == CPU_VENDOR_INTEL && cpu_exthigh >= 0x80000007 &&
+ (amd_pminfo & AMDPM_TSC_INVARIANT) == 0 && cpu_can_deep_sleep) {
+ tsc_timecounter.tc_quality = -1000;
+ if (bootverbose)
+ printf("TSC timecounter disabled: C3 enabled.\n");
+ goto init;
+ }
+
#ifdef SMP
/*
* We can not use the TSC in SMP mode unless the TSCs on all CPUs are
More information about the freebsd-current
mailing list