1 files changed, 80 insertions, 2 deletions

diff --git a/arch/mips/kernel/cevt-r4k.c b/arch/mips/kernel/cevt-r4k.c
index 8dfe6a6e1480..e4c21bbf9422 100644
--- a/arch/mips/kernel/cevt-r4k.c
+++ b/arch/mips/kernel/cevt-r4k.c
@@ -28,6 +28,83 @@ static int mips_next_event(unsigned long delta,
 	return res;
 }
 
+/**
+ * calculate_min_delta() - Calculate a good minimum delta for mips_next_event().
+ *
+ * Running under virtualisation can introduce overhead into mips_next_event() in
+ * the form of hypervisor emulation of CP0_Count/CP0_Compare registers,
+ * potentially with an unnatural frequency, which makes a fixed min_delta_ns
+ * value inappropriate as it may be too small.
+ *
+ * It can also introduce occasional latency from the guest being descheduled.
+ *
+ * This function calculates a good minimum delta based roughly on the 75th
+ * percentile of the time taken to do the mips_next_event() sequence, in order
+ * to handle potentially higher overhead while also eliminating outliers due to
+ * unpredictable hypervisor latency (which can be handled by retries).
+ *
+ * Return:	An appropriate minimum delta for the clock event device.
+ */
+static unsigned int calculate_min_delta(void)
+{
+	unsigned int cnt, i, j, k, l;
+	unsigned int buf1[4], buf2[3];
+	unsigned int min_delta;
+
+	/*
+	 * Calculate the median of 5 75th percentiles of 5 samples of how long
+	 * it takes to set CP0_Compare = CP0_Count + delta.
+	 */
+	for (i = 0; i < 5; ++i) {
+		for (j = 0; j < 5; ++j) {
+			/*
+			 * This is like the code in mips_next_event(), and
+			 * directly measures the borderline "safe" delta.
+			 */
+			cnt = read_c0_count();
+			write_c0_compare(cnt);
+			cnt = read_c0_count() - cnt;
+
+			/* Sorted insert into buf1 */
+			for (k = 0; k < j; ++k) {
+				if (cnt < buf1[k]) {
+					l = min_t(unsigned int,
+						  j, ARRAY_SIZE(buf1) - 1);
+					for (; l > k; --l)
+						buf1[l] = buf1[l - 1];
+					break;
+				}
+			}
+			if (k < ARRAY_SIZE(buf1))
+				buf1[k] = cnt;
+		}
+
+		/* Sorted insert of 75th percentile into buf2 */
+		for (k = 0; k < i; ++k) {
+			if (buf1[ARRAY_SIZE(buf1) - 1] < buf2[k]) {
+				l = min_t(unsigned int,
+					  i, ARRAY_SIZE(buf2) - 1);
+				for (; l > k; --l)
+					buf2[l] = buf2[l - 1];
+				break;
+			}
+		}
+		if (k < ARRAY_SIZE(buf2))
+			buf2[k] = buf1[ARRAY_SIZE(buf1) - 1];
+	}
+
+	/* Use 2 * median of 75th percentiles */
+	min_delta = buf2[ARRAY_SIZE(buf2) - 1] * 2;
+
+	/* Don't go too low */
+	if (min_delta < 0x300)
+		min_delta = 0x300;
+
+	pr_debug("%s: median 75th percentile=%#x, min_delta=%#x\n",
+		 __func__, buf2[ARRAY_SIZE(buf2) - 1], min_delta);
+	return min_delta;
+}
+
 DEFINE_PER_CPU(struct clock_event_device, mips_clockevent_device);
 int cp0_timer_irq_installed;
 
@@ -177,7 +254,7 @@ int r4k_clockevent_init(void)
 {
 	unsigned int cpu = smp_processor_id();
 	struct clock_event_device *cd;
-	unsigned int irq;
+	unsigned int irq, min_delta;
 
 	if (!cpu_has_counter || !mips_hpt_frequency)
 		return -ENXIO;
@@ -203,7 +280,8 @@ int r4k_clockevent_init(void)
 
 	/* Calculate the min / max delta */
 	cd->max_delta_ns	= clockevent_delta2ns(0x7fffffff, cd);
-	cd->min_delta_ns	= clockevent_delta2ns(0x300, cd);
+	min_delta		= calculate_min_delta();
+	cd->min_delta_ns	= clockevent_delta2ns(min_delta, cd);
 
 	cd->rating		= 300;
 	cd->irq			= irq;