Re: tip.today - scheduler bam boom crash (cpu hotplug)

From: Wanpeng Li
Date: Mon Feb 27 2017 - 07:30:58 EST


Cc Paolo,
2017-01-19 21:36 GMT+08:00 Peter Zijlstra <peterz@xxxxxxxxxxxxx>:
> On Thu, Jan 19, 2017 at 11:19:24AM +0100, Peter Zijlstra wrote:
>
>> I'll go have a prod. Thanks!
>
> This seems to cure it for me.
>
> ---
> Subject: sched/clock: Fix hotplug issue
>
> Mike reported that he could trigger the WARN_ON_ONCE() in
> set_sched_clock_stable() using hotplug.
>
> This exposed a fundamental problem with the interface, we should never
> mark the TSC stable if we ever find it to be unstable. Therefore
> set_sched_clock_stable() is a broken interface.
>
> The reason it existed is that not having it is a pain, it means all
> relevant architecture code needs to call clear_sched_clock_stable()
> where appropriate.
>
> Of the three architectures that select HAVE_UNSTABLE_SCHED_CLOCK ia64
> and parisc are trivial in that they never called
> set_sched_clock_stable(), so add an unconditional call to
> clear_sched_clock_stable() to them.
>
> For x86 the story is a lot more involved, and what this patch tries to
> do is ensure we preserve the status quo. So even is Cyrix or Transmeta
> have usable TSC they never called set_sched_clock_stable() so they now
> get an explicit mark unstable.
>
> XXX: what about Xen ?
>
> Fixes: 9881b024b7d7 ("sched/clock: Delay switching sched_clock to stable")
> Reported-by: Mike Galbraith <efault@xxxxxx>
> Signed-off-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
> ---
> arch/ia64/kernel/setup.c | 2 ++
> arch/parisc/kernel/setup.c | 2 ++
> arch/x86/kernel/cpu/amd.c | 6 ++++--
> arch/x86/kernel/cpu/centaur.c | 2 ++
> arch/x86/kernel/cpu/common.c | 3 +++
> arch/x86/kernel/cpu/cyrix.c | 2 ++
> arch/x86/kernel/cpu/intel.c | 6 ++++--
> arch/x86/kernel/cpu/transmeta.c | 3 +++
> arch/x86/kernel/kvmclock.c | 2 +-
> include/linux/sched.h | 1 -
> kernel/sched/clock.c | 29 ++++++++---------------------
> 11 files changed, 31 insertions(+), 27 deletions(-)
>
> diff --git a/arch/ia64/kernel/setup.c b/arch/ia64/kernel/setup.c
> index 7ec7acc..c483ece 100644
> --- a/arch/ia64/kernel/setup.c
> +++ b/arch/ia64/kernel/setup.c
> @@ -619,6 +619,8 @@ setup_arch (char **cmdline_p)
> check_sal_cache_flush();
> #endif
> paging_init();
> +
> + clear_sched_clock_stable();
> }
>
> /*
> diff --git a/arch/parisc/kernel/setup.c b/arch/parisc/kernel/setup.c
> index 2e66a88..068ed36 100644
> --- a/arch/parisc/kernel/setup.c
> +++ b/arch/parisc/kernel/setup.c
> @@ -36,6 +36,7 @@
> #undef PCI_DEBUG
> #include <linux/proc_fs.h>
> #include <linux/export.h>
> +#include <linux/sched.h>
>
> #include <asm/processor.h>
> #include <asm/sections.h>
> @@ -176,6 +177,7 @@ void __init setup_arch(char **cmdline_p)
> conswitchp = &dummy_con; /* we use do_take_over_console() later ! */
> #endif
>
> + clear_sched_clock_stable();
> }
>
> /*
> diff --git a/arch/x86/kernel/cpu/amd.c b/arch/x86/kernel/cpu/amd.c
> index 1d31672..80e657e 100644
> --- a/arch/x86/kernel/cpu/amd.c
> +++ b/arch/x86/kernel/cpu/amd.c
> @@ -541,8 +541,10 @@ static void early_init_amd(struct cpuinfo_x86 *c)
> if (c->x86_power & (1 << 8)) {
> set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
> set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
> - if (!check_tsc_unstable())
> - set_sched_clock_stable();
> + if (check_tsc_unstable())
> + clear_sched_clock_stable();
> + } else {
> + clear_sched_clock_stable();
> }
>
> /* Bit 12 of 8000_0007 edx is accumulated power mechanism. */
> diff --git a/arch/x86/kernel/cpu/centaur.c b/arch/x86/kernel/cpu/centaur.c
> index 1661d8e..c4bab8c 100644
> --- a/arch/x86/kernel/cpu/centaur.c
> +++ b/arch/x86/kernel/cpu/centaur.c
> @@ -104,6 +104,8 @@ static void early_init_centaur(struct cpuinfo_x86 *c)
> #ifdef CONFIG_X86_64
> set_cpu_cap(c, X86_FEATURE_SYSENTER32);
> #endif
> +
> + clear_sched_clock_stable();
> }
>
> static void init_centaur(struct cpuinfo_x86 *c)
> diff --git a/arch/x86/kernel/cpu/common.c b/arch/x86/kernel/cpu/common.c
> index 9bab7a8..0bdb1ab 100644
> --- a/arch/x86/kernel/cpu/common.c
> +++ b/arch/x86/kernel/cpu/common.c
> @@ -83,6 +83,7 @@ static void default_init(struct cpuinfo_x86 *c)
> strcpy(c->x86_model_id, "386");
> }
> #endif
> + clear_sched_clock_stable();
> }
>
> static const struct cpu_dev default_cpu = {
> @@ -1055,6 +1056,8 @@ static void identify_cpu(struct cpuinfo_x86 *c)
> */
> if (this_cpu->c_init)
> this_cpu->c_init(c);
> + else
> + clear_sched_clock_stable();
>
> /* Disable the PN if appropriate */
> squash_the_stupid_serial_number(c);
> diff --git a/arch/x86/kernel/cpu/cyrix.c b/arch/x86/kernel/cpu/cyrix.c
> index bd9dcd6..47416f9 100644
> --- a/arch/x86/kernel/cpu/cyrix.c
> +++ b/arch/x86/kernel/cpu/cyrix.c
> @@ -9,6 +9,7 @@
> #include <asm/pci-direct.h>
> #include <asm/tsc.h>
> #include <asm/cpufeature.h>
> +#include <linux/sched.h>
>
> #include "cpu.h"
>
> @@ -183,6 +184,7 @@ static void early_init_cyrix(struct cpuinfo_x86 *c)
> set_cpu_cap(c, X86_FEATURE_CYRIX_ARR);
> break;
> }
> + clear_sched_clock_stable();
> }
>
> static void init_cyrix(struct cpuinfo_x86 *c)
> diff --git a/arch/x86/kernel/cpu/intel.c b/arch/x86/kernel/cpu/intel.c
> index 203f860..026c728 100644
> --- a/arch/x86/kernel/cpu/intel.c
> +++ b/arch/x86/kernel/cpu/intel.c
> @@ -119,8 +119,10 @@ static void early_init_intel(struct cpuinfo_x86 *c)
> if (c->x86_power & (1 << 8)) {
> set_cpu_cap(c, X86_FEATURE_CONSTANT_TSC);
> set_cpu_cap(c, X86_FEATURE_NONSTOP_TSC);
> - if (!check_tsc_unstable())
> - set_sched_clock_stable();
> + if (check_tsc_unstable())
> + clear_sched_clock_stable();
> + } else {
> + clear_sched_clock_stable();
> }
>
> /* Penwell and Cloverview have the TSC which doesn't sleep on S3 */
> diff --git a/arch/x86/kernel/cpu/transmeta.c b/arch/x86/kernel/cpu/transmeta.c
> index 3417856..c1ea5b9 100644
> --- a/arch/x86/kernel/cpu/transmeta.c
> +++ b/arch/x86/kernel/cpu/transmeta.c
> @@ -1,4 +1,5 @@
> #include <linux/kernel.h>
> +#include <linux/sched.h>
> #include <linux/mm.h>
> #include <asm/cpufeature.h>
> #include <asm/msr.h>
> @@ -14,6 +15,8 @@ static void early_init_transmeta(struct cpuinfo_x86 *c)
> if (xlvl >= 0x80860001)
> c->x86_capability[CPUID_8086_0001_EDX] = cpuid_edx(0x80860001);
> }
> +
> + clear_sched_clock_stable();
> }
>
> static void init_transmeta(struct cpuinfo_x86 *c)
> diff --git a/arch/x86/kernel/kvmclock.c b/arch/x86/kernel/kvmclock.c
> index 2a5cafd..542710b 100644
> --- a/arch/x86/kernel/kvmclock.c
> +++ b/arch/x86/kernel/kvmclock.c
> @@ -107,12 +107,12 @@ static inline void kvm_sched_clock_init(bool stable)
> {
> if (!stable) {
> pv_time_ops.sched_clock = kvm_clock_read;
> + clear_sched_clock_stable();
> return;
> }
>
> kvm_sched_clock_offset = kvm_clock_read();
> pv_time_ops.sched_clock = kvm_sched_clock_read;
> - set_sched_clock_stable();

This results in sched clock always unstable for kvm guest since there
is no invariant tsc cpuid bit exposed for kvm guest currently. The
blockage happened for several reasons:

1) Migration: to host with different TSC frequency.
2) Savevm: It is not safe to use the TSC for wall clock timer services.

How about something like below(untested):

diff --git a/arch/x86/kernel/kvmclock.c b/arch/x86/kernel/kvmclock.c
index bae6ea6..a61c477 100644
--- a/arch/x86/kernel/kvmclock.c
+++ b/arch/x86/kernel/kvmclock.c
@@ -115,6 +115,7 @@ static inline void kvm_sched_clock_init(bool stable)

kvm_sched_clock_offset = kvm_clock_read();
pv_time_ops.sched_clock = kvm_sched_clock_read;
+ hypervisor_sched_clock_stable();

printk(KERN_INFO "kvm-clock: using sched offset of %llu cycles\n",
kvm_sched_clock_offset);
diff --git a/include/linux/sched.h b/include/linux/sched.h
index 451e241..38c6edb 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -2499,6 +2499,10 @@ static inline void clear_sched_clock_stable(void)
{
}

+static inline void hypervisor_sched_clock_stable(void)
+{
+}
+
static inline void sched_clock_idle_sleep_event(void)
{
}
@@ -2526,6 +2530,7 @@ extern void sched_clock_init_late(void);
*/
extern int sched_clock_stable(void);
extern void clear_sched_clock_stable(void);
+extern void hypervisor_sched_clock_stable(void);

extern void sched_clock_tick(void);
extern void sched_clock_idle_sleep_event(void);
diff --git a/kernel/sched/clock.c b/kernel/sched/clock.c
index ad64efe..a46639e 100644
--- a/kernel/sched/clock.c
+++ b/kernel/sched/clock.c
@@ -77,9 +77,15 @@ EXPORT_SYMBOL_GPL(sched_clock);

__read_mostly int sched_clock_running;

+enum {
+ SCHED_CLOCK_INIT = 0,
+ HYPERVISOR_SCHED_CLOCK_STABLE,
+ SCHED_CLOCK_INIT_LATE
+};
+
void sched_clock_init(void)
{
- sched_clock_running = 1;
+ sched_clock_running = SCHED_CLOCK_INIT;
}

#ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
@@ -170,13 +176,14 @@ void clear_sched_clock_stable(void)

smp_mb(); /* matches sched_clock_init_late() */

- if (sched_clock_running == 2)
+ if (sched_clock_running == SCHED_CLOCK_INIT_LATE)
schedule_work(&sched_clock_work);
}

void sched_clock_init_late(void)
{
- sched_clock_running = 2;
+ if (sched_clock_running == SCHED_CLOCK_INIT)
+ sched_clock_running = SCHED_CLOCK_INIT_LATE;
/*
* Ensure that it is impossible to not do a static_key update.
*
@@ -186,8 +193,15 @@ void sched_clock_init_late(void)
*/
smp_mb(); /* matches {set,clear}_sched_clock_stable() */

- if (__sched_clock_stable_early)
+ if (__sched_clock_stable_early ||
+ sched_clock_running == HYPERVISOR_SCHED_CLOCK_STABLE) {
__set_sched_clock_stable();
+ }
+}
+
+void hypervisor_sched_clock_stable()
+{
+ sched_clock_running = HYPERVISOR_SCHED_CLOCK_STABLE;
}

/*

Regards,
Wanpeng Li


>
> printk(KERN_INFO "kvm-clock: using sched offset of %llu cycles\n",
> kvm_sched_clock_offset);
> diff --git a/include/linux/sched.h b/include/linux/sched.h
> index 68b35dd..8a9bbb6 100644
> --- a/include/linux/sched.h
> +++ b/include/linux/sched.h
> @@ -2522,7 +2522,6 @@ extern void sched_clock_init_late(void);
> * is reliable after all:
> */
> extern int sched_clock_stable(void);
> -extern void set_sched_clock_stable(void);
> extern void clear_sched_clock_stable(void);
>
> extern void sched_clock_tick(void);
> diff --git a/kernel/sched/clock.c b/kernel/sched/clock.c
> index 7713b2b..ad64efe 100644
> --- a/kernel/sched/clock.c
> +++ b/kernel/sched/clock.c
> @@ -83,8 +83,15 @@ void sched_clock_init(void)
> }
>
> #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
> +/*
> + * We must start with !__sched_clock_stable because the unstable -> stable
> + * transition is accurate, while the stable -> unstable transition is not.
> + *
> + * Similarly we start with __sched_clock_stable_early, thereby assuming we
> + * will become stable, such that there's only a single 1 -> 0 transition.
> + */
> static DEFINE_STATIC_KEY_FALSE(__sched_clock_stable);
> -static int __sched_clock_stable_early;
> +static int __sched_clock_stable_early = 1;
>
> /*
> * We want: ktime_get_ns() + gtod_offset == sched_clock() + raw_offset
> @@ -132,24 +139,6 @@ static void __set_sched_clock_stable(void)
> tick_dep_clear(TICK_DEP_BIT_CLOCK_UNSTABLE);
> }
>
> -void set_sched_clock_stable(void)
> -{
> - __sched_clock_stable_early = 1;
> -
> - smp_mb(); /* matches sched_clock_init_late() */
> -
> - /*
> - * This really should only be called early (before
> - * sched_clock_init_late()) when guestimating our sched_clock() is
> - * solid.
> - *
> - * After that we test stability and we can negate our guess using
> - * clear_sched_clock_stable, possibly from a watchdog.
> - */
> - if (WARN_ON_ONCE(sched_clock_running == 2))
> - __set_sched_clock_stable();
> -}
> -
> static void __clear_sched_clock_stable(struct work_struct *work)
> {
> struct sched_clock_data *scd = this_scd();
> @@ -199,8 +188,6 @@ void sched_clock_init_late(void)
>
> if (__sched_clock_stable_early)
> __set_sched_clock_stable();
> - else
> - __clear_sched_clock_stable(NULL);
> }
>
> /*