![Thomas Gleixner](/assets/img/avatar_default.png)
The hardlockup detector on x86 uses a performance counter based on unhalted
CPU cycles and a periodic hrtimer. The hrtimer period is about 2/5 of the
performance counter period, so the hrtimer should fire 2-3 times before the
performance counter NMI fires. The NMI code checks whether the hrtimer
fired since the last invocation. If not, it assumess a hard lockup.
The calculation of those periods is based on the nominal CPU
frequency. Turbo modes increase the CPU clock frequency and therefore
shorten the period of the perf/NMI watchdog. With extreme Turbo-modes (3x
nominal frequency) the perf/NMI period is shorter than the hrtimer period
which leads to false positives.
A simple fix would be to shorten the hrtimer period, but that comes with
the side effect of more frequent hrtimer and softlockup thread wakeups,
which is not desired.
Implement a low pass filter, which checks the perf/NMI period against
kernel time. If the perf/NMI fires before 4/5 of the watchdog period has
elapsed then the event is ignored and postponed to the next perf/NMI.
That solves the problem and avoids the overhead of shorter hrtimer periods
and more frequent softlockup thread wakeups.
Fixes: 58687acba5
("lockup_detector: Combine nmi_watchdog and softlockup detector")
Reported-and-tested-by: Kan Liang <Kan.liang@intel.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: dzickus@redhat.com
Cc: prarit@redhat.com
Cc: ak@linux.intel.com
Cc: babu.moger@oracle.com
Cc: peterz@infradead.org
Cc: eranian@google.com
Cc: acme@redhat.com
Cc: stable@vger.kernel.org
Cc: atomlin@redhat.com
Cc: akpm@linux-foundation.org
Cc: torvalds@linux-foundation.org
Link: http://lkml.kernel.org/r/alpine.DEB.2.20.1708150931310.1886@nanos
264 lines
7.3 KiB
C
264 lines
7.3 KiB
C
/*
|
|
* Detect hard lockups on a system
|
|
*
|
|
* started by Don Zickus, Copyright (C) 2010 Red Hat, Inc.
|
|
*
|
|
* Note: Most of this code is borrowed heavily from the original softlockup
|
|
* detector, so thanks to Ingo for the initial implementation.
|
|
* Some chunks also taken from the old x86-specific nmi watchdog code, thanks
|
|
* to those contributors as well.
|
|
*/
|
|
|
|
#define pr_fmt(fmt) "NMI watchdog: " fmt
|
|
|
|
#include <linux/nmi.h>
|
|
#include <linux/module.h>
|
|
#include <linux/sched/debug.h>
|
|
|
|
#include <asm/irq_regs.h>
|
|
#include <linux/perf_event.h>
|
|
|
|
static DEFINE_PER_CPU(bool, hard_watchdog_warn);
|
|
static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
|
|
static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
|
|
|
|
static unsigned long hardlockup_allcpu_dumped;
|
|
|
|
void arch_touch_nmi_watchdog(void)
|
|
{
|
|
/*
|
|
* Using __raw here because some code paths have
|
|
* preemption enabled. If preemption is enabled
|
|
* then interrupts should be enabled too, in which
|
|
* case we shouldn't have to worry about the watchdog
|
|
* going off.
|
|
*/
|
|
raw_cpu_write(watchdog_nmi_touch, true);
|
|
}
|
|
EXPORT_SYMBOL(arch_touch_nmi_watchdog);
|
|
|
|
#ifdef CONFIG_HARDLOCKUP_CHECK_TIMESTAMP
|
|
static DEFINE_PER_CPU(ktime_t, last_timestamp);
|
|
static DEFINE_PER_CPU(unsigned int, nmi_rearmed);
|
|
static ktime_t watchdog_hrtimer_sample_threshold __read_mostly;
|
|
|
|
void watchdog_update_hrtimer_threshold(u64 period)
|
|
{
|
|
/*
|
|
* The hrtimer runs with a period of (watchdog_threshold * 2) / 5
|
|
*
|
|
* So it runs effectively with 2.5 times the rate of the NMI
|
|
* watchdog. That means the hrtimer should fire 2-3 times before
|
|
* the NMI watchdog expires. The NMI watchdog on x86 is based on
|
|
* unhalted CPU cycles, so if Turbo-Mode is enabled the CPU cycles
|
|
* might run way faster than expected and the NMI fires in a
|
|
* smaller period than the one deduced from the nominal CPU
|
|
* frequency. Depending on the Turbo-Mode factor this might be fast
|
|
* enough to get the NMI period smaller than the hrtimer watchdog
|
|
* period and trigger false positives.
|
|
*
|
|
* The sample threshold is used to check in the NMI handler whether
|
|
* the minimum time between two NMI samples has elapsed. That
|
|
* prevents false positives.
|
|
*
|
|
* Set this to 4/5 of the actual watchdog threshold period so the
|
|
* hrtimer is guaranteed to fire at least once within the real
|
|
* watchdog threshold.
|
|
*/
|
|
watchdog_hrtimer_sample_threshold = period * 2;
|
|
}
|
|
|
|
static bool watchdog_check_timestamp(void)
|
|
{
|
|
ktime_t delta, now = ktime_get_mono_fast_ns();
|
|
|
|
delta = now - __this_cpu_read(last_timestamp);
|
|
if (delta < watchdog_hrtimer_sample_threshold) {
|
|
/*
|
|
* If ktime is jiffies based, a stalled timer would prevent
|
|
* jiffies from being incremented and the filter would look
|
|
* at a stale timestamp and never trigger.
|
|
*/
|
|
if (__this_cpu_inc_return(nmi_rearmed) < 10)
|
|
return false;
|
|
}
|
|
__this_cpu_write(nmi_rearmed, 0);
|
|
__this_cpu_write(last_timestamp, now);
|
|
return true;
|
|
}
|
|
#else
|
|
static inline bool watchdog_check_timestamp(void)
|
|
{
|
|
return true;
|
|
}
|
|
#endif
|
|
|
|
static struct perf_event_attr wd_hw_attr = {
|
|
.type = PERF_TYPE_HARDWARE,
|
|
.config = PERF_COUNT_HW_CPU_CYCLES,
|
|
.size = sizeof(struct perf_event_attr),
|
|
.pinned = 1,
|
|
.disabled = 1,
|
|
};
|
|
|
|
/* Callback function for perf event subsystem */
|
|
static void watchdog_overflow_callback(struct perf_event *event,
|
|
struct perf_sample_data *data,
|
|
struct pt_regs *regs)
|
|
{
|
|
/* Ensure the watchdog never gets throttled */
|
|
event->hw.interrupts = 0;
|
|
|
|
if (atomic_read(&watchdog_park_in_progress) != 0)
|
|
return;
|
|
|
|
if (__this_cpu_read(watchdog_nmi_touch) == true) {
|
|
__this_cpu_write(watchdog_nmi_touch, false);
|
|
return;
|
|
}
|
|
|
|
if (!watchdog_check_timestamp())
|
|
return;
|
|
|
|
/* check for a hardlockup
|
|
* This is done by making sure our timer interrupt
|
|
* is incrementing. The timer interrupt should have
|
|
* fired multiple times before we overflow'd. If it hasn't
|
|
* then this is a good indication the cpu is stuck
|
|
*/
|
|
if (is_hardlockup()) {
|
|
int this_cpu = smp_processor_id();
|
|
|
|
/* only print hardlockups once */
|
|
if (__this_cpu_read(hard_watchdog_warn) == true)
|
|
return;
|
|
|
|
pr_emerg("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
|
|
print_modules();
|
|
print_irqtrace_events(current);
|
|
if (regs)
|
|
show_regs(regs);
|
|
else
|
|
dump_stack();
|
|
|
|
/*
|
|
* Perform all-CPU dump only once to avoid multiple hardlockups
|
|
* generating interleaving traces
|
|
*/
|
|
if (sysctl_hardlockup_all_cpu_backtrace &&
|
|
!test_and_set_bit(0, &hardlockup_allcpu_dumped))
|
|
trigger_allbutself_cpu_backtrace();
|
|
|
|
if (hardlockup_panic)
|
|
nmi_panic(regs, "Hard LOCKUP");
|
|
|
|
__this_cpu_write(hard_watchdog_warn, true);
|
|
return;
|
|
}
|
|
|
|
__this_cpu_write(hard_watchdog_warn, false);
|
|
return;
|
|
}
|
|
|
|
/*
|
|
* People like the simple clean cpu node info on boot.
|
|
* Reduce the watchdog noise by only printing messages
|
|
* that are different from what cpu0 displayed.
|
|
*/
|
|
static unsigned long firstcpu_err;
|
|
static atomic_t watchdog_cpus;
|
|
|
|
int watchdog_nmi_enable(unsigned int cpu)
|
|
{
|
|
struct perf_event_attr *wd_attr;
|
|
struct perf_event *event = per_cpu(watchdog_ev, cpu);
|
|
int firstcpu = 0;
|
|
|
|
/* nothing to do if the hard lockup detector is disabled */
|
|
if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
|
|
goto out;
|
|
|
|
/* is it already setup and enabled? */
|
|
if (event && event->state > PERF_EVENT_STATE_OFF)
|
|
goto out;
|
|
|
|
/* it is setup but not enabled */
|
|
if (event != NULL)
|
|
goto out_enable;
|
|
|
|
if (atomic_inc_return(&watchdog_cpus) == 1)
|
|
firstcpu = 1;
|
|
|
|
wd_attr = &wd_hw_attr;
|
|
wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
|
|
|
|
/* Try to register using hardware perf events */
|
|
event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
|
|
|
|
/* save the first cpu's error for future comparision */
|
|
if (firstcpu && IS_ERR(event))
|
|
firstcpu_err = PTR_ERR(event);
|
|
|
|
if (!IS_ERR(event)) {
|
|
/* only print for the first cpu initialized */
|
|
if (firstcpu || firstcpu_err)
|
|
pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
|
|
goto out_save;
|
|
}
|
|
|
|
/*
|
|
* Disable the hard lockup detector if _any_ CPU fails to set up
|
|
* set up the hardware perf event. The watchdog() function checks
|
|
* the NMI_WATCHDOG_ENABLED bit periodically.
|
|
*
|
|
* The barriers are for syncing up watchdog_enabled across all the
|
|
* cpus, as clear_bit() does not use barriers.
|
|
*/
|
|
smp_mb__before_atomic();
|
|
clear_bit(NMI_WATCHDOG_ENABLED_BIT, &watchdog_enabled);
|
|
smp_mb__after_atomic();
|
|
|
|
/* skip displaying the same error again */
|
|
if (!firstcpu && (PTR_ERR(event) == firstcpu_err))
|
|
return PTR_ERR(event);
|
|
|
|
/* vary the KERN level based on the returned errno */
|
|
if (PTR_ERR(event) == -EOPNOTSUPP)
|
|
pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
|
|
else if (PTR_ERR(event) == -ENOENT)
|
|
pr_warn("disabled (cpu%i): hardware events not enabled\n",
|
|
cpu);
|
|
else
|
|
pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
|
|
cpu, PTR_ERR(event));
|
|
|
|
pr_info("Shutting down hard lockup detector on all cpus\n");
|
|
|
|
return PTR_ERR(event);
|
|
|
|
/* success path */
|
|
out_save:
|
|
per_cpu(watchdog_ev, cpu) = event;
|
|
out_enable:
|
|
perf_event_enable(per_cpu(watchdog_ev, cpu));
|
|
out:
|
|
return 0;
|
|
}
|
|
|
|
void watchdog_nmi_disable(unsigned int cpu)
|
|
{
|
|
struct perf_event *event = per_cpu(watchdog_ev, cpu);
|
|
|
|
if (event) {
|
|
perf_event_disable(event);
|
|
per_cpu(watchdog_ev, cpu) = NULL;
|
|
|
|
/* should be in cleanup, but blocks oprofile */
|
|
perf_event_release_kernel(event);
|
|
|
|
/* watchdog_nmi_enable() expects this to be zero initially. */
|
|
if (atomic_dec_and_test(&watchdog_cpus))
|
|
firstcpu_err = 0;
|
|
}
|
|
}
|