drm/i915: Improved w/a for rps on Baytrail
authorChris Wilson <chris@chris-wilson.co.uk>
Wed, 18 Mar 2015 09:48:22 +0000 (09:48 +0000)
committerDaniel Vetter <daniel.vetter@ffwll.ch>
Fri, 20 Mar 2015 10:48:14 +0000 (11:48 +0100)
Rewrite commit 31685c258e0b0ad6aa486c5ec001382cf8a64212
Author: Deepak S <deepak.s@linux.intel.com>
Date:   Thu Jul 3 17:33:01 2014 -0400

    drm/i915/vlv: WA for Turbo and RC6 to work together.

Other than code clarity, the major improvement is to disable the extra
interrupts generated when idle.  However, the reclocking remains rather
slow under the new manual regime, in particular it fails to downclock as
quickly as desired. The second major improvement is that for certain
workloads, like games, we need to combine render+media activity counters
as the work of displaying the frame is split across the engines and both
need to be taken into account when deciding the global GPU frequency as
memory cycles are shared.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Deepak S <deepak.s@linux.intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Rodrigo Vivi <rodrigo.vivi@intel.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Deepak S<deepak.s@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
drivers/gpu/drm/i915/i915_irq.c
drivers/gpu/drm/i915/i915_reg.h
drivers/gpu/drm/i915/intel_display.c
drivers/gpu/drm/i915/intel_drv.h
drivers/gpu/drm/i915/intel_pm.c

index 49ad5fb82ace27909e3e5e49a48fa8c1d86b627d..8d8d33d068dd0c55d3442aeb517cd6a950670fa4 100644 (file)
@@ -997,129 +997,84 @@ static void notify_ring(struct drm_device *dev,
        wake_up_all(&ring->irq_queue);
 }
 
-static u32 vlv_c0_residency(struct drm_i915_private *dev_priv,
-                           struct intel_rps_ei *rps_ei)
+static void vlv_c0_read(struct drm_i915_private *dev_priv,
+                       struct intel_rps_ei *ei)
 {
-       u32 cz_ts, cz_freq_khz;
-       u32 render_count, media_count;
-       u32 elapsed_render, elapsed_media, elapsed_time;
-       u32 residency = 0;
-
-       cz_ts = vlv_punit_read(dev_priv, PUNIT_REG_CZ_TIMESTAMP);
-       cz_freq_khz = DIV_ROUND_CLOSEST(dev_priv->mem_freq * 1000, 4);
-
-       render_count = I915_READ(VLV_RENDER_C0_COUNT_REG);
-       media_count = I915_READ(VLV_MEDIA_C0_COUNT_REG);
-
-       if (rps_ei->cz_clock == 0) {
-               rps_ei->cz_clock = cz_ts;
-               rps_ei->render_c0 = render_count;
-               rps_ei->media_c0 = media_count;
-
-               return dev_priv->rps.cur_freq;
-       }
-
-       elapsed_time = cz_ts - rps_ei->cz_clock;
-       rps_ei->cz_clock = cz_ts;
+       ei->cz_clock = vlv_punit_read(dev_priv, PUNIT_REG_CZ_TIMESTAMP);
+       ei->render_c0 = I915_READ(VLV_RENDER_C0_COUNT);
+       ei->media_c0 = I915_READ(VLV_MEDIA_C0_COUNT);
+}
 
-       elapsed_render = render_count - rps_ei->render_c0;
-       rps_ei->render_c0 = render_count;
+static bool vlv_c0_above(struct drm_i915_private *dev_priv,
+                        const struct intel_rps_ei *old,
+                        const struct intel_rps_ei *now,
+                        int threshold)
+{
+       u64 time, c0;
 
-       elapsed_media = media_count - rps_ei->media_c0;
-       rps_ei->media_c0 = media_count;
+       if (old->cz_clock == 0)
+               return false;
 
-       /* Convert all the counters into common unit of milli sec */
-       elapsed_time /= VLV_CZ_CLOCK_TO_MILLI_SEC;
-       elapsed_render /=  cz_freq_khz;
-       elapsed_media /= cz_freq_khz;
+       time = now->cz_clock - old->cz_clock;
+       time *= threshold * dev_priv->mem_freq;
 
-       /*
-        * Calculate overall C0 residency percentage
-        * only if elapsed time is non zero
+       /* Workload can be split between render + media, e.g. SwapBuffers
+        * being blitted in X after being rendered in mesa. To account for
+        * this we need to combine both engines into our activity counter.
         */
-       if (elapsed_time) {
-               residency =
-                       ((max(elapsed_render, elapsed_media) * 100)
-                               / elapsed_time);
-       }
+       c0 = now->render_c0 - old->render_c0;
+       c0 += now->media_c0 - old->media_c0;
+       c0 *= 100 * VLV_CZ_CLOCK_TO_MILLI_SEC * 4 / 1000;
 
-       return residency;
+       return c0 >= time;
 }
 
-/**
- * vlv_calc_delay_from_C0_counters - Increase/Decrease freq based on GPU
- * busy-ness calculated from C0 counters of render & media power wells
- * @dev_priv: DRM device private
- *
- */
-static int vlv_calc_delay_from_C0_counters(struct drm_i915_private *dev_priv)
+void gen6_rps_reset_ei(struct drm_i915_private *dev_priv)
 {
-       u32 residency_C0_up = 0, residency_C0_down = 0;
-       int new_delay, adj;
-
-       dev_priv->rps.ei_interrupt_count++;
-
-       WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
+       vlv_c0_read(dev_priv, &dev_priv->rps.down_ei);
+       dev_priv->rps.up_ei = dev_priv->rps.down_ei;
+       dev_priv->rps.ei_interrupt_count = 0;
+}
 
+static u32 vlv_wa_c0_ei(struct drm_i915_private *dev_priv, u32 pm_iir)
+{
+       struct intel_rps_ei now;
+       u32 events = 0;
 
-       if (dev_priv->rps.up_ei.cz_clock == 0) {
-               vlv_c0_residency(dev_priv, &dev_priv->rps.up_ei);
-               vlv_c0_residency(dev_priv, &dev_priv->rps.down_ei);
-               return dev_priv->rps.cur_freq;
-       }
+       if ((pm_iir & GEN6_PM_RP_UP_EI_EXPIRED) == 0)
+               return 0;
 
+       vlv_c0_read(dev_priv, &now);
+       if (now.cz_clock == 0)
+               return 0;
 
        /*
         * To down throttle, C0 residency should be less than down threshold
         * for continous EI intervals. So calculate down EI counters
         * once in VLV_INT_COUNT_FOR_DOWN_EI
         */
-       if (dev_priv->rps.ei_interrupt_count == VLV_INT_COUNT_FOR_DOWN_EI) {
-
+       if (++dev_priv->rps.ei_interrupt_count >= VLV_INT_COUNT_FOR_DOWN_EI) {
+               pm_iir |= GEN6_PM_RP_DOWN_EI_EXPIRED;
                dev_priv->rps.ei_interrupt_count = 0;
-
-               residency_C0_down = vlv_c0_residency(dev_priv,
-                                                    &dev_priv->rps.down_ei);
-       } else {
-               residency_C0_up = vlv_c0_residency(dev_priv,
-                                                  &dev_priv->rps.up_ei);
        }
 
-       new_delay = dev_priv->rps.cur_freq;
-
-       adj = dev_priv->rps.last_adj;
-       /* C0 residency is greater than UP threshold. Increase Frequency */
-       if (residency_C0_up >= VLV_RP_UP_EI_THRESHOLD) {
-               if (adj > 0)
-                       adj *= 2;
-               else
-                       adj = 1;
-
-               if (dev_priv->rps.cur_freq < dev_priv->rps.max_freq_softlimit)
-                       new_delay = dev_priv->rps.cur_freq + adj;
-
-               /*
-                * For better performance, jump directly
-                * to RPe if we're below it.
-                */
-               if (new_delay < dev_priv->rps.efficient_freq)
-                       new_delay = dev_priv->rps.efficient_freq;
+       if (pm_iir & GEN6_PM_RP_DOWN_EI_EXPIRED) {
+               if (!vlv_c0_above(dev_priv,
+                                 &dev_priv->rps.down_ei, &now,
+                                 VLV_RP_DOWN_EI_THRESHOLD))
+                       events |= GEN6_PM_RP_DOWN_THRESHOLD;
+               dev_priv->rps.down_ei = now;
+       }
 
-       } else if (!dev_priv->rps.ei_interrupt_count &&
-                       (residency_C0_down < VLV_RP_DOWN_EI_THRESHOLD)) {
-               if (adj < 0)
-                       adj *= 2;
-               else
-                       adj = -1;
-               /*
-                * This means, C0 residency is less than down threshold over
-                * a period of VLV_INT_COUNT_FOR_DOWN_EI. So, reduce the freq
-                */
-               if (dev_priv->rps.cur_freq > dev_priv->rps.min_freq_softlimit)
-                       new_delay = dev_priv->rps.cur_freq + adj;
+       if (pm_iir & GEN6_PM_RP_UP_EI_EXPIRED) {
+               if (vlv_c0_above(dev_priv,
+                                &dev_priv->rps.up_ei, &now,
+                                VLV_RP_UP_EI_THRESHOLD))
+                       events |= GEN6_PM_RP_UP_THRESHOLD;
+               dev_priv->rps.up_ei = now;
        }
 
-       return new_delay;
+       return events;
 }
 
 static void gen6_pm_rps_work(struct work_struct *work)
@@ -1149,6 +1104,8 @@ static void gen6_pm_rps_work(struct work_struct *work)
 
        mutex_lock(&dev_priv->rps.hw_lock);
 
+       pm_iir |= vlv_wa_c0_ei(dev_priv, pm_iir);
+
        adj = dev_priv->rps.last_adj;
        if (pm_iir & GEN6_PM_RP_UP_THRESHOLD) {
                if (adj > 0)
@@ -1171,8 +1128,6 @@ static void gen6_pm_rps_work(struct work_struct *work)
                else
                        new_delay = dev_priv->rps.min_freq_softlimit;
                adj = 0;
-       } else if (pm_iir & GEN6_PM_RP_UP_EI_EXPIRED) {
-               new_delay = vlv_calc_delay_from_C0_counters(dev_priv);
        } else if (pm_iir & GEN6_PM_RP_DOWN_THRESHOLD) {
                if (adj < 0)
                        adj *= 2;
index cc8ebabc488dd23e3e69b9020ad6a121bd7fc94b..2d76c566d84348f71eecc81e56596a147db725bf 100644 (file)
@@ -6220,8 +6220,8 @@ enum skl_disp_power_wells {
 
 #define GEN6_GT_GFX_RC6p                       0x13810C
 #define GEN6_GT_GFX_RC6pp                      0x138110
-#define VLV_RENDER_C0_COUNT_REG                0x138118
-#define VLV_MEDIA_C0_COUNT_REG                 0x13811C
+#define VLV_RENDER_C0_COUNT                    0x138118
+#define VLV_MEDIA_C0_COUNT                     0x13811C
 
 #define GEN6_PCODE_MAILBOX                     0x138124
 #define   GEN6_PCODE_READY                     (1<<31)
index 90b460cf2b57d275ce1b9d49d58a0d28ecd19e89..f1c0295f69e5c4da7aed74222dd88a97ad3079f4 100644 (file)
@@ -9201,6 +9201,8 @@ void intel_mark_busy(struct drm_device *dev)
 
        intel_runtime_pm_get(dev_priv);
        i915_update_gfx_val(dev_priv);
+       if (INTEL_INFO(dev)->gen >= 6)
+               gen6_rps_busy(dev_priv);
        dev_priv->mm.busy = true;
 }
 
index a1baaa188b0add1d50ec8ba75a11e0e049b9bdf1..8bb18e507f5fc1332589f1637018df41f8dfae48 100644 (file)
@@ -1242,6 +1242,8 @@ void intel_disable_gt_powersave(struct drm_device *dev);
 void intel_suspend_gt_powersave(struct drm_device *dev);
 void intel_reset_gt_powersave(struct drm_device *dev);
 void gen6_update_ring_freq(struct drm_device *dev);
+void gen6_rps_busy(struct drm_i915_private *dev_priv);
+void gen6_rps_reset_ei(struct drm_i915_private *dev_priv);
 void gen6_rps_idle(struct drm_i915_private *dev_priv);
 void gen6_rps_boost(struct drm_i915_private *dev_priv);
 void ilk_wm_get_hw_state(struct drm_device *dev);
index beab305e320db35b42527330c7a8a4b02cc16532..68c9cc252d36b625e2222a9d5a5eac3b6a183839 100644 (file)
@@ -4041,6 +4041,18 @@ static void vlv_set_rps_idle(struct drm_i915_private *dev_priv)
        I915_WRITE(GEN6_PMINTRMSK, gen6_rps_pm_mask(dev_priv, val));
 }
 
+void gen6_rps_busy(struct drm_i915_private *dev_priv)
+{
+       mutex_lock(&dev_priv->rps.hw_lock);
+       if (dev_priv->rps.enabled) {
+               if (dev_priv->pm_rps_events & (GEN6_PM_RP_DOWN_EI_EXPIRED | GEN6_PM_RP_UP_EI_EXPIRED))
+                       gen6_rps_reset_ei(dev_priv);
+               I915_WRITE(GEN6_PMINTRMSK,
+                          gen6_rps_pm_mask(dev_priv, dev_priv->rps.cur_freq));
+       }
+       mutex_unlock(&dev_priv->rps.hw_lock);
+}
+
 void gen6_rps_idle(struct drm_i915_private *dev_priv)
 {
        struct drm_device *dev = dev_priv->dev;
@@ -4052,15 +4064,21 @@ void gen6_rps_idle(struct drm_i915_private *dev_priv)
                else
                        gen6_set_rps(dev_priv->dev, dev_priv->rps.idle_freq);
                dev_priv->rps.last_adj = 0;
+               I915_WRITE(GEN6_PMINTRMSK, 0xffffffff);
        }
        mutex_unlock(&dev_priv->rps.hw_lock);
 }
 
 void gen6_rps_boost(struct drm_i915_private *dev_priv)
 {
+       u32 val;
+
        mutex_lock(&dev_priv->rps.hw_lock);
-       if (dev_priv->rps.enabled) {
-               intel_set_rps(dev_priv->dev, dev_priv->rps.max_freq_softlimit);
+       val = dev_priv->rps.max_freq_softlimit;
+       if (dev_priv->rps.enabled &&
+           dev_priv->mm.busy &&
+           dev_priv->rps.cur_freq < val) {
+               intel_set_rps(dev_priv->dev, val);
                dev_priv->rps.last_adj = 0;
        }
        mutex_unlock(&dev_priv->rps.hw_lock);