cpuidle: pseries: Fixup exit latency for CEDE(0)

We are currently assuming that CEDE(0) has exit latency 10us, since
there is no way for us to query from the platform. However, if the
wakeup latency of an Extended CEDE state is smaller than 10us, then we
can be sure that the exit latency of CEDE(0) cannot be more than that.

In this patch, we fix the exit latency of CEDE(0) if we discover an
Extended CEDE state with wakeup latency smaller than 10us.

Benchmark results:

On POWER8, this patch does not have any impact since the advertized
latency of Extended CEDE (1) is 30us which is higher than the default
latency of CEDE (0) which is 10us.

On POWER9 we see improvement the single-threaded performance of
ebizzy, and no regression in the wakeup latency or the number of
context-switches.

ebizzy:
2 ebizzy threads bound to the same big-core. 25% improvement in the
avg records/s with patch.

  x without_patch
  * with_patch
      N           Min           Max        Median           Avg        Stddev
  x  10       2491089       5834307       5398375       4244335     1596244.9
  *  10       2893813       5834474       5832448     5327281.3     1055941.4

context_switch2:
There is no major regression observed with this patch as seen from the
context_switch2 benchmark.

context_switch2 across CPU0 CPU1 (Both belong to same big-core, but
different small cores). We observe a minor 0.14% regression in the
number of context-switches (higher is better).

  x without_patch
  * with_patch
      N           Min           Max        Median           Avg        Stddev
  x 500        348872        362236        354712     354745.69      2711.827
  * 500        349422        361452        353942      354215.4     2576.9258

  Difference at 99.0% confidence
    -530.288 +/- 430.963
    -0.149484% +/- 0.121485%
    (Student's t, pooled s = 2645.24)

context_switch2 across CPU0 CPU8 (Different big-cores). We observe a
0.37% improvement in the number of context-switches (higher is
better).

  x without_patch
  * with_patch
      N           Min           Max        Median           Avg        Stddev
  x 500        287956        294940        288896     288977.23     646.59295
  * 500        288300        294646        289582     290064.76     1161.9992

  Difference at 99.0% confidence
    1087.53 +/- 153.194
    0.376337% +/- 0.0530125%
    (Student's t, pooled s = 940.299)

schbench:
No major difference could be seen until the 99.9th percentile.

Without-patch:
  Latency percentiles (usec)
        50.0th: 29
        75.0th: 39
        90.0th: 49
        95.0th: 59
        *99.0th: 13104
        99.5th: 14672
        99.9th: 15824
        min=0, max=17993

With-patch:
  Latency percentiles (usec)
        50.0th: 29
        75.0th: 40
        90.0th: 50
        95.0th: 61
        *99.0th: 13648
        99.5th: 14768
        99.9th: 15664
        min=0, max=29812

Signed-off-by: Gautham R. Shenoy <ego@linux.vnet.ibm.com>
[mpe: Minor formatting]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/1596087177-30329-4-git-send-email-ego@linux.vnet.ibm.com

drivers/cpuidle/cpuidle-pseries.c

@@ -343,13 +343,52 @@ static int pseries_cpuidle_driver_init(void)
 	return 0;
 }
 
-static void __init parse_xcede_idle_states(void)
+static void __init fixup_cede0_latency(void)
 {
+	struct xcede_latency_payload *payload;
+	u64 min_latency_us;
+	int i;
+
+	min_latency_us = dedicated_states[1].exit_latency; // CEDE latency
+
 	if (parse_cede_parameters())
 		return;
 
-	pr_info("cpuidle : Skipping the %d Extended CEDE idle states\n",
+	pr_info("cpuidle: Skipping the %d Extended CEDE idle states\n",
 		nr_xcede_records);
+
+	payload = &xcede_latency_parameter.payload;
+	for (i = 0; i < nr_xcede_records; i++) {
+		struct xcede_latency_record *record = &payload->records[i];
+		u64 latency_tb = be64_to_cpu(record->latency_ticks);
+		u64 latency_us = tb_to_ns(latency_tb) / NSEC_PER_USEC;
+
+		if (latency_us < min_latency_us)
+			min_latency_us = latency_us;
+	}
+
+	/*
+	 * By default, we assume that CEDE(0) has exit latency 10us,
+	 * since there is no way for us to query from the platform.
+	 *
+	 * However, if the wakeup latency of an Extended CEDE state is
+	 * smaller than 10us, then we can be sure that CEDE(0)
+	 * requires no more than that.
+	 *
+	 * Perform the fix-up.
+	 */
+	if (min_latency_us < dedicated_states[1].exit_latency) {
+		u64 cede0_latency = min_latency_us - 1;
+
+		if (cede0_latency <= 0)
+			cede0_latency = min_latency_us;
+
+		dedicated_states[1].exit_latency = cede0_latency;
+		dedicated_states[1].target_residency = 10 * (cede0_latency);
+		pr_info("cpuidle: Fixed up CEDE exit latency to %llu us\n",
+			cede0_latency);
+	}
+
 }
 
 /*
@@ -373,7 +412,7 @@ static int pseries_idle_probe(void)
 			cpuidle_state_table = shared_states;
 			max_idle_state = ARRAY_SIZE(shared_states);
 		} else {
-			parse_xcede_idle_states();
+			fixup_cede0_latency();
 			cpuidle_state_table = dedicated_states;
 			max_idle_state = NR_DEDICATED_STATES;
 		}