drm/i915/gt: Switch to manual evaluation of RPS

As with the realisation for soft-rc6, we respond to idling the engines
within microseconds, far faster than the response times for HW RC6 and
RPS. Furthermore, our fast parking upon idle, prevents HW RPS from
running for many desktop workloads, as the RPS evaluation intervals are
on the order of tens of milliseconds, but the typical workload is just a
couple of milliseconds, but yet we still need to determine the best
frequency for user latency versus power.

Recognising that the HW evaluation intervals are a poor fit, and that
they were deprecated [in bspec at least] from gen10, start to wean
ourselves off them and replace the EI with a timer and our accurate
busy-stats. The principle benefit of manually evaluating RPS intervals
is that we can be more responsive for better performance and powersaving
for both spiky workloads and steady-state.

Closes: https://gitlab.freedesktop.org/drm/intel/-/issues/1698
Fixes: 98479ada421a ("drm/i915/gt: Treat idling as a RPS downclock event")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Cc: Andi Shyti <andi.shyti@intel.com>
Reviewed-by: Andi Shyti <andi.shyti@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200429205446.3259-4-chris@chris-wilson.co.uk

1 files changed, 122 insertions, 16 deletions

diff --git a/drivers/gpu/drm/i915/gt/intel_rps.c b/drivers/gpu/drm/i915/gt/intel_rps.c
index 52151001d7ab..8b2991de1c97 100644
--- a/drivers/gpu/drm/i915/gt/intel_rps.c
+++ b/drivers/gpu/drm/i915/gt/intel_rps.c
@@ -15,6 +15,8 @@
 #include "intel_sideband.h"
 #include "../../../platform/x86/intel_ips.h"
 
+#define BUSY_MAX_EI	20u /* ms */
+
 /*
  * Lock protecting IPS related data structures
  */
@@ -45,6 +47,100 @@ static inline void set(struct intel_uncore *uncore, i915_reg_t reg, u32 val)
 	intel_uncore_write_fw(uncore, reg, val);
 }
 
+static void rps_timer(struct timer_list *t)
+{
+	struct intel_rps *rps = from_timer(rps, t, timer);
+	struct intel_engine_cs *engine;
+	enum intel_engine_id id;
+	s64 max_busy[3] = {};
+	ktime_t dt, last;
+
+	for_each_engine(engine, rps_to_gt(rps), id) {
+		s64 busy;
+		int i;
+
+		dt = intel_engine_get_busy_time(engine);
+		last = engine->stats.rps;
+		engine->stats.rps = dt;
+
+		busy = ktime_to_ns(ktime_sub(dt, last));
+		for (i = 0; i < ARRAY_SIZE(max_busy); i++) {
+			if (busy > max_busy[i])
+				swap(busy, max_busy[i]);
+		}
+	}
+
+	dt = ktime_get();
+	last = rps->pm_timestamp;
+	rps->pm_timestamp = dt;
+
+	if (intel_rps_is_active(rps)) {
+		s64 busy;
+		int i;
+
+		dt = ktime_sub(dt, last);
+
+		/*
+		 * Our goal is to evaluate each engine independently, so we run
+		 * at the lowest clocks required to sustain the heaviest
+		 * workload. However, a task may be split into sequential
+		 * dependent operations across a set of engines, such that
+		 * the independent contributions do not account for high load,
+		 * but overall the task is GPU bound. For example, consider
+		 * video decode on vcs followed by colour post-processing
+		 * on vecs, followed by general post-processing on rcs.
+		 * Since multi-engines being active does imply a single
+		 * continuous workload across all engines, we hedge our
+		 * bets by only contributing a factor of the distributed
+		 * load into our busyness calculation.
+		 */
+		busy = max_busy[0];
+		for (i = 1; i < ARRAY_SIZE(max_busy); i++) {
+			if (!max_busy[i])
+				break;
+
+			busy += div_u64(max_busy[i], 1 << i);
+		}
+		GT_TRACE(rps_to_gt(rps),
+			 "busy:%lld [%d%%], max:[%lld, %lld, %lld], interval:%d\n",
+			 busy, (int)div64_u64(100 * busy, dt),
+			 max_busy[0], max_busy[1], max_busy[2],
+			 rps->pm_interval);
+
+		if (100 * busy > rps->power.up_threshold * dt &&
+		    rps->cur_freq < rps->max_freq_softlimit) {
+			rps->pm_iir |= GEN6_PM_RP_UP_THRESHOLD;
+			rps->pm_interval = 1;
+			schedule_work(&rps->work);
+		} else if (100 * busy < rps->power.down_threshold * dt &&
+			   rps->cur_freq > rps->min_freq_softlimit) {
+			rps->pm_iir |= GEN6_PM_RP_DOWN_THRESHOLD;
+			rps->pm_interval = 1;
+			schedule_work(&rps->work);
+		} else {
+			rps->last_adj = 0;
+		}
+
+		mod_timer(&rps->timer,
+			  jiffies + msecs_to_jiffies(rps->pm_interval));
+		rps->pm_interval = min(rps->pm_interval * 2, BUSY_MAX_EI);
+	}
+}
+
+static void rps_start_timer(struct intel_rps *rps)
+{
+	rps->pm_timestamp = ktime_sub(ktime_get(), rps->pm_timestamp);
+	rps->pm_interval = 1;
+	mod_timer(&rps->timer, jiffies + 1);
+}
+
+static void rps_stop_timer(struct intel_rps *rps)
+{
+	del_timer_sync(&rps->timer);
+	rps->pm_timestamp = ktime_sub(ktime_get(), rps->pm_timestamp);
+	cancel_work_sync(&rps->work);
+}
+
 static u32 rps_pm_mask(struct intel_rps *rps, u8 val)
 {
 	u32 mask = 0;
@@ -535,36 +631,24 @@ static void rps_set_power(struct intel_rps *rps, int new_power)
 	if (new_power == rps->power.mode)
 		return;
 
+	threshold_up = 95;
+	threshold_down = 85;
+
 	/* Note the units here are not exactly 1us, but 1280ns. */
 	switch (new_power) {
 	case LOW_POWER:
-		/* Upclock if more than 95% busy over 16ms */
 		ei_up = 16000;
-		threshold_up = 95;
-
-		/* Downclock if less than 85% busy over 32ms */
 		ei_down = 32000;
-		threshold_down = 85;
 		break;
 
 	case BETWEEN:
-		/* Upclock if more than 90% busy over 13ms */
 		ei_up = 13000;
-		threshold_up = 90;
-
-		/* Downclock if less than 75% busy over 32ms */
 		ei_down = 32000;
-		threshold_down = 75;
 		break;
 
 	case HIGH_POWER:
-		/* Upclock if more than 85% busy over 10ms */
 		ei_up = 10000;
-		threshold_up = 85;
-
-		/* Downclock if less than 60% busy over 32ms */
 		ei_down = 32000;
-		threshold_down = 60;
 		break;
 	}
 
@@ -742,8 +826,11 @@ void intel_rps_unpark(struct intel_rps *rps)
 
 	mutex_unlock(&rps->lock);
 
+	rps->pm_iir = 0;
 	if (intel_rps_has_interrupts(rps))
 		rps_enable_interrupts(rps);
+	if (intel_rps_uses_timer(rps))
+		rps_start_timer(rps);
 
 	if (IS_GEN(rps_to_i915(rps), 5))
 		gen5_rps_update(rps);
@@ -754,6 +841,8 @@ void intel_rps_park(struct intel_rps *rps)
 	if (!intel_rps_clear_active(rps))
 		return;
 
+	if (intel_rps_uses_timer(rps))
+		rps_stop_timer(rps);
 	if (intel_rps_has_interrupts(rps))
 		rps_disable_interrupts(rps);
 
@@ -1211,6 +1300,19 @@ static unsigned long __ips_gfx_val(struct intel_ips *ips)
 	return ips->gfx_power + state2;
 }
 
+static bool has_busy_stats(struct intel_rps *rps)
+{
+	struct intel_engine_cs *engine;
+	enum intel_engine_id id;
+
+	for_each_engine(engine, rps_to_gt(rps), id) {
+		if (!intel_engine_supports_stats(engine))
+			return false;
+	}
+
+	return true;
+}
+
 void intel_rps_enable(struct intel_rps *rps)
 {
 	struct drm_i915_private *i915 = rps_to_i915(rps);
@@ -1255,7 +1357,9 @@ void intel_rps_enable(struct intel_rps *rps)
 	GEM_BUG_ON(rps->efficient_freq < rps->min_freq);
 	GEM_BUG_ON(rps->efficient_freq > rps->max_freq);
 
-	if (INTEL_GEN(i915) >= 6)
+	if (has_busy_stats(rps))
+		intel_rps_set_timer(rps);
+	else if (INTEL_GEN(i915) >= 6)
 		intel_rps_set_interrupts(rps);
 	else
 		/* Ironlake currently uses intel_ips.ko */ {}
@@ -1274,6 +1378,7 @@ void intel_rps_disable(struct intel_rps *rps)
 
 	intel_rps_clear_enabled(rps);
 	intel_rps_clear_interrupts(rps);
+	intel_rps_clear_timer(rps);
 
 	if (INTEL_GEN(i915) >= 6)
 		gen6_rps_disable(rps);
@@ -1689,6 +1794,7 @@ void intel_rps_init_early(struct intel_rps *rps)
 	mutex_init(&rps->power.mutex);
 
 	INIT_WORK(&rps->work, rps_work);
+	timer_setup(&rps->timer, rps_timer, 0);
 
 	atomic_set(&rps->num_waiters, 0);
 }