4.12-stable review patch. If anyone has any objections, please let me know.
------------------
From: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
commit 7edaeb6841dfb27e362288ab8466ebdc4972e867 upstream.
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: 58687acba592 ("lockup_detector: Combine nmi_watchdog and softlockup detector")
Reported-and-tested-by: Kan Liang <Kan.liang@xxxxxxxxx>
Signed-off-by: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
Cc: dzickus@xxxxxxxxxx
Cc: prarit@xxxxxxxxxx
Cc: ak@xxxxxxxxxxxxxxx
Cc: babu.moger@xxxxxxxxxx
Cc: peterz@xxxxxxxxxxxxx
Cc: eranian@xxxxxxxxxx
Cc: acme@xxxxxxxxxx
Cc: atomlin@xxxxxxxxxx
Cc: akpm@xxxxxxxxxxxxxxxxxxxx
Cc: torvalds@xxxxxxxxxxxxxxxxxxxx
Link: http://lkml.kernel.org/r/alpine.DEB.2.20.1708150931310.1886@nanos
Signed-off-by: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx>
---
arch/x86/Kconfig | 1
include/linux/nmi.h | 8 ++++++
kernel/watchdog.c | 1
kernel/watchdog_hld.c | 59 ++++++++++++++++++++++++++++++++++++++++++++++++++
lib/Kconfig.debug | 7 +++++
5 files changed, 76 insertions(+)
--- a/arch/x86/Kconfig
+++ b/arch/x86/Kconfig
@@ -94,6 +94,7 @@ config X86
select GENERIC_STRNCPY_FROM_USER
select GENERIC_STRNLEN_USER
select GENERIC_TIME_VSYSCALL
+ select HARDLOCKUP_CHECK_TIMESTAMP if X86_64
select HAVE_ACPI_APEI if ACPI
select HAVE_ACPI_APEI_NMI if ACPI
select HAVE_ALIGNED_STRUCT_PAGE if SLUB
--- a/include/linux/nmi.h
+++ b/include/linux/nmi.h
@@ -155,6 +155,14 @@ extern int sysctl_hardlockup_all_cpu_bac
#define sysctl_softlockup_all_cpu_backtrace 0
#define sysctl_hardlockup_all_cpu_backtrace 0
#endif
+
+#if defined(CONFIG_HARDLOCKUP_CHECK_TIMESTAMP) && \
+ defined(CONFIG_HARDLOCKUP_DETECTOR)
+void watchdog_update_hrtimer_threshold(u64 period);
+#else
+static inline void watchdog_update_hrtimer_threshold(u64 period) { }
+#endif
+
extern bool is_hardlockup(void);
struct ctl_table;
extern int proc_watchdog(struct ctl_table *, int ,
--- a/kernel/watchdog.c
+++ b/kernel/watchdog.c
@@ -161,6 +161,7 @@ static void set_sample_period(void)
* hardlockup detector generates a warning
*/
sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
+ watchdog_update_hrtimer_threshold(sample_period);
}
/* Commands for resetting the watchdog */
--- a/kernel/watchdog_hld.c
+++ b/kernel/watchdog_hld.c
@@ -70,6 +70,62 @@ void touch_nmi_watchdog(void)
}
EXPORT_SYMBOL(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,
@@ -94,6 +150,9 @@ static void watchdog_overflow_callback(s
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
--- a/lib/Kconfig.debug
+++ b/lib/Kconfig.debug
@@ -345,6 +345,13 @@ config SECTION_MISMATCH_WARN_ONLY
If unsure, say Y.
#
+# Enables a timestamp based low pass filter to compensate for perf based
+# hard lockup detection which runs too fast due to turbo modes.
+#
+config HARDLOCKUP_CHECK_TIMESTAMP
+ bool
+
+#
# Select this config option from the architecture Kconfig, if it
# is preferred to always offer frame pointers as a config
# option on the architecture (regardless of KERNEL_DEBUG):