[V8,3/3] irq: Compute the periodic interval for interrupts

diff --git a/include/linux/interrupt.h b/include/linux/interrupt.h
index c050333..62bb0c9 100644
--- a/include/linux/interrupt.h
+++ b/include/linux/interrupt.h
@@ -707,6 +707,7 @@  static inline void init_irq_proc(void)
 #ifdef CONFIG_IRQ_TIMINGS
 void irq_timings_enable(void);
 void irq_timings_disable(void);
+u64 irq_timings_next_event(u64 now);
 #endif
 
 struct seq_file;
diff --git a/kernel/irq/internals.h b/kernel/irq/internals.h
index abaadaa..677c1a2 100644
--- a/kernel/irq/internals.h
+++ b/kernel/irq/internals.h
@@ -247,13 +247,21 @@  struct irq_timings {
 
 DECLARE_PER_CPU(struct irq_timings, irq_timings);
 
+extern void irq_timings_free(int irq);
+extern int irq_timings_alloc(int irq);
+
 static inline void remove_timings(struct irq_desc *desc)
 {
 	desc->istate &= ~IRQS_TIMINGS;
+
+	irq_timings_free(irq_desc_get_irq(desc));
 }
 
 static inline void setup_timings(struct irq_desc *desc, struct irqaction *act)
 {
+	int irq = irq_desc_get_irq(desc);
+	int ret;
+
 	/*
 	 * We don't need the measurement because the idle code already
 	 * knows the next expiry event.
@@ -261,6 +269,17 @@  static inline void setup_timings(struct irq_desc *desc, struct irqaction *act)
 	if (act->flags & __IRQF_TIMER)
 		return;
 
+	/*
+	 * In case the timing allocation fails, we just want to warn,
+	 * not fail, so letting the system boot anyway.
+	 */
+	ret = irq_timings_alloc(irq);
+	if (ret) {
+		pr_warn("Failed to allocate irq timing stats for irq%d (%d)",
+			irq, ret);
+		return;
+	}
+
 	desc->istate |= IRQS_TIMINGS;
 }
 
diff --git a/kernel/irq/timings.c b/kernel/irq/timings.c
index 21ceca8..a3879cb 100644
--- a/kernel/irq/timings.c
+++ b/kernel/irq/timings.c
@@ -9,9 +9,14 @@ 
  *
  */
 #include <linux/percpu.h>
+#include <linux/slab.h>
 #include <linux/static_key.h>
 #include <linux/interrupt.h>
+#include <linux/idr.h>
 #include <linux/irq.h>
+#include <linux/math64.h>
+
+#include <trace/events/irq.h>
 
 #include "internals.h"
 
@@ -19,6 +24,18 @@ 
 
 DEFINE_PER_CPU(struct irq_timings, irq_timings);
 
+struct irq_stat {
+	u64 ne;         /* next event                               */
+	u64 lts;	/* last timestamp                           */
+	u64 variance;	/* variance                                 */
+	u32 avg;	/* mean value                               */
+	u32 count;      /* number of samples                        */
+	int anomalies;  /* number of consecutives anomalies         */
+	int valid;      /* behaviour of the interrupt               */
+};
+
+static DEFINE_IDR(irq_stats);
+
 void irq_timings_enable(void)
 {
 	static_branch_inc(&irq_timing_enabled);
@@ -28,3 +45,330 @@  void irq_timings_disable(void)
 {
 	static_branch_dec(&irq_timing_enabled);
 }
+
+/**
+ * irqs_update - update the irq timing statistics with a new timestamp
+ *
+ * @irqs: an irq_stat struct pointer
+ * @ts: the new timestamp
+ *
+ * ** This function must be called with the local irq disabled **
+ *
+ * The statistics are computed online, in other words, the code is
+ * designed to compute the statistics on a stream of values rather
+ * than doing multiple passes on the values to compute the average,
+ * then the variance. The integer division introduces a loss of
+ * precision but with an acceptable error margin regarding the results
+ * we would have with the double floating precision: we are dealing
+ * with nanosec, so big numbers, consequently the mantisse is
+ * negligeable, especially when converting the time in usec
+ * afterwards.
+ *
+ * The computation happens at idle time. When the CPU is not idle, the
+ * interrupts' timestamps are stored in the circular buffer, when the
+ * CPU goes idle and this routine is called, all the buffer's values
+ * are injected in the statistical model continuying to extend the
+ * statistics from the previous busy-idle cycle.
+ *
+ * The observations showed a device will trigger a burst of periodic
+ * interrupts followed by one or two peaks of longer time, for
+ * instance when a SD card device flushes its cache, then the periodic
+ * intervals occur again. A one second inactivity period resets the
+ * stats, that gives us the certitude the statistical values won't
+ * exceed 1x10^9, thus the computation won't overflow.
+ *
+ * Basically, the purpose of the algorithm is to watch the periodic
+ * interrupts and eliminate the peaks.
+ *
+ * An interrupt is considered periodically stable if the interval of
+ * its occurences follow the normal distribution, thus the values
+ * comply with:
+ *
+ *      avg - 3 x stddev < value < avg + 3 x stddev
+ *
+ * Which can be simplified to:
+ *
+ *      -3 x stddev < value - avg < 3 x stddev
+ *
+ *      abs(value - avg) < 3 x stddev
+ *
+ * In order to save a costly square root computation, we use the
+ * variance. For the record, stddev = sqrt(variance). The equation
+ * above becomes:
+ *
+ *      abs(value - avg) < 3 x sqrt(variance)
+ *
+ * And finally we square it:
+ *
+ *      (value - avg) ^ 2 < (3 x sqrt(variance)) ^ 2
+ *
+ *      (value - avg) x (value - avg) < 9 x variance
+ *
+ * Statistically speaking, any values out of this interval is
+ * considered as an anomaly and is discarded. However, a normal
+ * distribution appears when the number of samples is 30 (it is the
+ * rule of thumb in statistics, cf. "30 samples" on Internet). When
+ * there are three consecutive anomalies, the statistics are resetted.
+ *
+ */
+static void irqs_update(struct irq_stat *irqs, u64 ts)
+{
+	u64 old_ts = irqs->lts;
+	u64 variance = 0;
+	u64 interval;
+	s64 diff;
+
+	/*
+	 * The timestamps are absolute time values, we need to compute
+	 * the timing interval between two interrupts.
+	 */
+	irqs->lts = ts;
+
+	/*
+	 * The interval type is u64 in order to deal with the same
+	 * type in our computation, that prevent mindfuck issues with
+	 * overflow, sign and do_div.
+	 */
+	interval = ts - old_ts;
+
+	/*
+	 * The interrupt triggered more than one second apart, that
+	 * ends the sequence as predictible for our purpose. In this
+	 * case, assume we have the beginning of a sequence and the
+	 * timestamp is the first value. As it is impossible to
+	 * predict anything at this point, return.
+	 *
+	 * Note the first timestamp of the sequence will always fall
+	 * in this test because the old_ts is zero. That is what we
+	 * want as we need another timestamp to compute an interval.
+	 */
+	if (interval >= NSEC_PER_SEC) {
+		memset(irqs, 0, sizeof(*irqs));
+		irqs->lts = ts;
+		return;
+	}
+
+	/*
+	 * Pre-compute the delta with the average as the result is
+	 * used several times in this function.
+	 */
+	diff = interval - irqs->avg;
+
+	/*
+	 * Increment the number of samples.
+	 */
+	irqs->count++;
+
+	/*
+	 * Online variance divided by the number of elements if there
+	 * is more than one sample.
+	 */
+	if (likely(irqs->count > 1))
+		variance = div_u64(irqs->variance, irqs->count - 1);
+
+	/*
+	 * The rule of thumb in statistics for the normal distribution
+	 * is having at least 30 samples in order to have the model to
+	 * apply. Values outside the interval are considered as an
+	 * anomaly.
+	 */
+	if ((irqs->count >= 30) && ((diff * diff) > (9 * variance))) {
+		/*
+		 * After three consecutive anomalies, we reset the
+		 * stats as it is no longer stable enough.
+		 */
+		if (irqs->anomalies++ >= 3) {
+			memset(irqs, 0, sizeof(*irqs));
+			irqs->lts = ts;
+			return;
+		}
+	} else {
+		/*
+		 * The anomalies must be consecutives, so at this
+		 * point, we reset the anomalies counter.
+		 */
+		irqs->anomalies = 0;
+	}
+
+	/*
+	 * The interrupt is considered stable enough to try to predict
+	 * the next event on it.
+	 */
+	irqs->valid = 1;
+
+	/*
+	 * Online average algorithm:
+	 *
+	 *  new_average = average + ((value - average) / count)
+	 *
+	 * The variance computation depends on the new average
+	 * to be computed here first.
+	 *
+	 */
+	irqs->avg = irqs->avg + div_s64(diff, irqs->count);
+
+	/*
+	 * Online variance algorithm:
+	 *
+	 *  new_variance = variance + (value - average) x (value - new_average)
+	 *
+	 * Warning: irqs->avg is updated with the line above, hence
+	 * 'interval - irqs->avg' is no longer equal to 'diff'
+	 */
+	irqs->variance = irqs->variance + (diff * (interval - irqs->avg));
+
+	/*
+	 * Update the next event
+	 */
+	irqs->ne = ts + irqs->avg;
+}
+
+/**
+ * irq_timings_next_event - Return when the next event is supposed to arrive
+ *
+ * *** This function must be called with the local irq disabled ***
+ *
+ * During the last busy cycle, the number of interrupts is incremented
+ * and stored in the irq_timings structure. This information is
+ * necessary to:
+ *
+ * - know if the index in the table wrapped up:
+ *
+ *      If more than the array size interrupts happened during the
+ *      last busy/idle cycle, the index wrapped up and we have to
+ *      begin with the next element in the array which is the last one
+ *      in the sequence, otherwise it is a the index 0.
+ *
+ * - have an indication of the interrupts activity on this CPU
+ *   (eg. irq/sec)
+ *
+ * The values are 'consumed' after inserting in the statistical model,
+ * thus the count is reinitialized.
+ *
+ * The array of values **must** be browsed in the time direction, the
+ * timestamp must increase between an element and the next one.
+ *
+ * Returns a nanosec time based estimation of the earliest interrupt,
+ * U64_MAX otherwise.
+ */
+u64 irq_timings_next_event(u64 now)
+{
+	struct irq_timings *irqts = this_cpu_ptr(&irq_timings);
+	struct irq_timing *irqt;
+	struct irq_stat *irqs;
+	struct irq_stat __percpu *s;
+	u64 ne = U64_MAX;
+	int index, count, i, irq = 0;
+
+	/*
+	 * Number of elements in the circular buffer. If it happens it
+	 * was flushed before, then the number of elements could be
+	 * smaller than IRQ_TIMINGS_SIZE, so the count is used,
+	 * otherwise the array size is used as we wrapped. The index
+	 * begins from zero when we did not wrap. That could be done
+	 * in a nicer way with the proper circular array structure
+	 * type but with the cost of extra computation in the
+	 * interrupt handler hot path. We choose efficiency.
+	 */
+	if (irqts->count >= IRQ_TIMINGS_SIZE) {
+		count = IRQ_TIMINGS_SIZE;
+		index = irqts->count & IRQ_TIMINGS_MASK;
+	} else {
+		count = irqts->count;
+		index = 0;
+	}
+
+	/*
+	 * Inject measured irq/timestamp to the statistical model.
+	 */
+	for (i = 0; i < count; i++) {
+
+		irqt = &irqts->values[(index + i) & IRQ_TIMINGS_MASK];
+
+		s = idr_find(&irq_stats, irqt->irq);
+
+		irqs = this_cpu_ptr(s);
+
+		irqs_update(irqs, irqt->ts);
+	}
+
+	/*
+	 * Reset the counter, we consumed all the data from our
+	 * circular buffer.
+	 */
+	irqts->count = 0;
+
+	/*
+	 * Look in the list of interrupts' statistics, the earliest
+	 * next event.
+	 */
+	idr_for_each_entry(&irq_stats, s, i) {
+
+		irqs = this_cpu_ptr(s);
+
+		if (!irqs->valid)
+			continue;
+
+		if (irqs->ne <= now) {
+			irq = i;
+			ne = now;
+
+			/*
+			 * This interrupt mustn't use in the future
+			 * until new events occur and update the
+			 * statistics.
+			 */
+			irqs->valid = 0;
+			break;
+		}
+
+		if (irqs->ne < ne) {
+			irq = i;
+			ne = irqs->ne;
+		}
+	}
+
+	return ne;
+}
+
+void irq_timings_free(int irq)
+{
+	struct irq_stat __percpu *s;
+
+	s = idr_find(&irq_stats, irq);
+	if (s) {
+		free_percpu(s);
+		idr_remove(&irq_stats, irq);
+	}
+}
+
+int irq_timings_alloc(int irq)
+{
+	int id;
+	struct irq_stat __percpu *s;
+
+	/*
+	 * Some platforms can have the same private interrupt per cpu,
+	 * so this function may be be called several times with the
+	 * same interrupt number. Just bail out in case the per cpu
+	 * stat structure is already allocated.
+	 */
+	s = idr_find(&irq_stats, irq);
+	if (s)
+		return 0;
+
+	s = alloc_percpu(*s);
+	if (!s)
+		return -ENOMEM;
+
+	idr_preload(GFP_KERNEL);
+	id = idr_alloc(&irq_stats, s, irq, irq + 1, GFP_NOWAIT);
+	idr_preload_end();
+
+	if (id < 0) {
+		free_percpu(s);
+		return id;
+	}
+
+	return 0;
+}