Re: tip.today - scheduler bam boom crash (cpu hotplug)
From: Wanpeng Li
Date: Mon Feb 27 2017 - 20:52:03 EST
2017-02-28 0:11 GMT+08:00 Paolo Bonzini <pbonzini@xxxxxxxxxx>:
>
>
> On 27/02/2017 16:59, Peter Zijlstra wrote:
>> OK, so if !KVM_FEATURE_CLOCKSOURCE_STABLE_BIT nothing is stable, but if
>> it is set, TSC might still not be stable, but kvm_clock_read() is.
>>
>>> However, if kvmclock is stable, we know that the sched clock is stable.
>> Right, so the problem is that we only ever want to allow marking
>> unstable -- once its found unstable, for whatever reason, we should
>> never allow going stable. The corollary of this proposition is that we
>> must start out assuming it will become stable. And to avoid actually
>> using unstable TSC we do a 3 state bringup:
>>
>> 1) sched_clock_running = 0, __stable_early = 1, __stable = 0
>> 2) sched_clock_running = 1 (__stable is effective, iow, we run unstable)
>> 3) sched_clock_running = 2 (__stable <- __stable_early)
>>
>> 2) happens 'early' but is 'safe'.
>> 3) happens 'late', after we've brought up SMP and probed TSC
>>
>> Between there, we should have detected the most common TSC wreckage and
>> made sure to not then switch to 'stable' at 3.
>>
>> Now the problem appears to be that we assume sched_clock will use RDTSC
>> (native_sched_clock) while sched_clock is a paravirt op.
>>
>> Now, I've not yet figured out the ordering between when we set
>> pv_time_ops.sched_clock and when we do the 'normal' TSC init stuff.
>
> I think the ordering is fine:
>
> - pv_time_ops.sched_clock is set here:
>
> start_kernel (init/main.c line 509)
> setup_arch
> kvmclock_init
> kvm_sched_clock_init
>
> - TSC can be declared unstable only after this:
>
> start_kernel (init/main.c line 628)
> late_time_init
> tsc_init
>
> So by the time the tsc_cs_mark_unstable or mark_tsc_unstable can call
> clear_sched_clock_stable, pv_time_ops.sched_clock has been set.
>
>> But it appears to me, we should not be calling
>> clear_sched_clock_stable() on TSC bits when we don't end up using
>> native_sched_clock().
>
> Yes, this makes sense.
How about something like below, we delay sched clock stable check to
tsc_init() if we run in VM, it calls clear_sched_clock_stable() if
there is no invariant tsc bit when we end up using
native_sched_clock(), otherwise, kvmclock will determine if it is
stable depends on KVM_FEATURE_CLOCKSOURCE_STABLE_BIT /
PVCLOCK_TSC_STABLE_BIT :
diff --git a/arch/x86/kernel/cpu/amd.c b/arch/x86/kernel/cpu/amd.c
index 4e95b2e..ed8eda4 100644
--- a/arch/x86/kernel/cpu/amd.c
+++ b/arch/x86/kernel/cpu/amd.c
@@ -557,7 +557,7 @@ static void early_init_amd(struct cpuinfo_x86 *c)
set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
if (check_tsc_unstable())
clear_sched_clock_stable();
- } else {
+ } else if (!boot_cpu_has(X86_FEATURE_HYPERVISOR)) {
clear_sched_clock_stable();
}
diff --git a/arch/x86/kernel/cpu/intel.c b/arch/x86/kernel/cpu/intel.c
index 017ecd3..1927f5f 100644
--- a/arch/x86/kernel/cpu/intel.c
+++ b/arch/x86/kernel/cpu/intel.c
@@ -163,7 +163,7 @@ static void early_init_intel(struct cpuinfo_x86 *c)
set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
if (check_tsc_unstable())
clear_sched_clock_stable();
- } else {
+ } else if (!boot_cpu_has(X86_FEATURE_HYPERVISOR)) {
clear_sched_clock_stable();
}
diff --git a/arch/x86/kernel/tsc.c b/arch/x86/kernel/tsc.c
index 2724dc8..68149f5 100644
--- a/arch/x86/kernel/tsc.c
+++ b/arch/x86/kernel/tsc.c
@@ -326,9 +326,16 @@ unsigned long long sched_clock(void)
{
return paravirt_sched_clock();
}
+
+static inline bool using_native_sched_clock(void)
+{
+ return pv_time_ops.sched_clock == native_sched_clock;
+}
#else
unsigned long long
sched_clock(void) __attribute__((alias("native_sched_clock")));
+
+static inline bool using_native_sched_clock(void) { return true; }
#endif
int check_tsc_unstable(void)
@@ -1398,6 +1405,11 @@ void __init tsc_init(void)
use_tsc_delay();
+ if (using_native_sched_clock() &&
+ boot_cpu_has(X86_FEATURE_HYPERVISOR) &&
+ !boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
+ clear_sched_clock_stable();
+
if (unsynchronized_tsc())
mark_tsc_unstable("TSCs unsynchronized");
Regards,
Wanpeng Li