perfcounters: throttle on too high IRQ rates

Starting kerneltop with only -c 100 seems to be a bad idea, it can
easily lock the system due to perfcounter IRQ overload.

So add throttling: if a new IRQ arrives in a shorter than
PERFMON_MIN_PERIOD_NS time, turn off perfcounters and untrottle them
from the next timer tick.

Signed-off-by: Mike Galbraith <efault@gmx.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

3 files changed, 38 insertions, 6 deletions

diff --git a/arch/x86/kernel/apic.c b/arch/x86/kernel/apic.c
index 7b434e5b14c9..849c23009bf5 100644
--- a/arch/x86/kernel/apic.c
+++ b/arch/x86/kernel/apic.c
@@ -781,6 +781,8 @@ static void local_apic_timer_interrupt(void)
         inc_irq_stat(apic_timer_irqs);
 
         evt->event_handler(evt);
+
+        perf_counter_unthrottle();
 }
 
 /*
diff --git a/arch/x86/kernel/cpu/perf_counter.c b/arch/x86/kernel/cpu/perf_counter.c
index 9376771f757b..1a040b179b53 100644
--- a/arch/x86/kernel/cpu/perf_counter.c
+++ b/arch/x86/kernel/cpu/perf_counter.c
@@ -33,6 +33,9 @@ static int nr_counters_fixed __read_mostly;
 struct cpu_hw_counters {
         struct perf_counter     *counters[X86_PMC_IDX_MAX];
         unsigned long           used[BITS_TO_LONGS(X86_PMC_IDX_MAX)];
+        u64                     last_interrupt;
+        u64                     global_enable;
+        int                     throttled;
 };
 
 /*
@@ -474,16 +477,19 @@ perf_handle_group(struct perf_counter *sibling, u64 *status, u64 *overflown)
 static void __smp_perf_counter_interrupt(struct pt_regs *regs, int nmi)
 {
         int bit, cpu = smp_processor_id();
-        u64 ack, status, saved_global;
-        struct cpu_hw_counters *cpuc;
+        u64 ack, status, now;
+        struct cpu_hw_counters *cpuc = &per_cpu(cpu_hw_counters, cpu);
 
-        rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, saved_global);
+        rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, cpuc->global_enable);
 
         /* Disable counters globally */
         wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, 0);
         ack_APIC_irq();
 
-        cpuc = &per_cpu(cpu_hw_counters, cpu);
+        now = sched_clock();
+        if (now - cpuc->last_interrupt < PERFMON_MIN_PERIOD_NS)
+                cpuc->throttled = 1;
+        cpuc->last_interrupt = now;
 
         rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
         if (!status)
@@ -533,9 +539,29 @@ again:
                 goto again;
 out:
         /*
-         * Restore - do not reenable when global enable is off:
+         * Restore - do not reenable when global enable is off or throttled:
          */
-        wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, saved_global);
+        if (!cpuc->throttled)
+                wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, cpuc->global_enable);
+}
+
+void perf_counter_unthrottle(void)
+{
+        struct cpu_hw_counters *cpuc;
+
+        if (!cpu_has(&boot_cpu_data, X86_FEATURE_ARCH_PERFMON))
+                return;
+
+        if (unlikely(!perf_counters_initialized))
+                return;
+
+        cpuc = &per_cpu(cpu_hw_counters, smp_processor_id());
+        if (cpuc->throttled) {
+                if (printk_ratelimit())
+                        printk(KERN_WARNING "PERFMON: max event frequency exceeded!\n");
+                wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, cpuc->global_enable);
+                cpuc->throttled = 0;
+        }
 }
 
 void smp_perf_counter_interrupt(struct pt_regs *regs)
diff --git a/include/linux/perf_counter.h b/include/linux/perf_counter.h
index 33ba9fe0a781..91f1ca4c01c0 100644
--- a/include/linux/perf_counter.h
+++ b/include/linux/perf_counter.h
@@ -254,6 +254,7 @@ extern void perf_counter_init_task(struct task_struct *child);
 extern void perf_counter_exit_task(struct task_struct *child);
 extern void perf_counter_notify(struct pt_regs *regs);
 extern void perf_counter_print_debug(void);
+extern void perf_counter_unthrottle(void);
 extern u64 hw_perf_save_disable(void);
 extern void hw_perf_restore(u64 ctrl);
 extern int perf_counter_task_disable(void);
@@ -270,6 +271,8 @@ static inline int is_software_counter(struct perf_counter *counter)
         return !counter->hw_event.raw && counter->hw_event.type < 0;
 }
 
+#define PERFMON_MIN_PERIOD_NS 10000
+
 #else
 static inline void
 perf_counter_task_sched_in(struct task_struct *task, int cpu)           { }
@@ -281,6 +284,7 @@ static inline void perf_counter_init_task(struct task_struct *child)	{ }
 static inline void perf_counter_exit_task(struct task_struct *child)    { }
 static inline void perf_counter_notify(struct pt_regs *regs)            { }
 static inline void perf_counter_print_debug(void)                       { }
+static inline void perf_counter_unthrottle(void)                        { }
 static inline void hw_perf_restore(u64 ctrl)                    { }
 static inline u64 hw_perf_save_disable(void)                  { return 0; }
 static inline int perf_counter_task_disable(void)       { return -EINVAL; }