]> git.openfabrics.org - ~shefty/rdma-dev.git/commitdiff
drm/i915: Always update RPS interrupts thresholds along with frequency
authorChris Wilson <chris@chris-wilson.co.uk>
Sat, 28 Apr 2012 07:56:39 +0000 (08:56 +0100)
committerDaniel Vetter <daniel.vetter@ffwll.ch>
Thu, 24 May 2012 15:54:00 +0000 (17:54 +0200)
In order to avoid missed down-interrupts when coming out of RC6, it is
advised that we always reset the down-threshold upon a PM event. This is
due to that the PM unit goes through a little dance when coming out of
RC6, it first brings the GPU up at the lowest frequency then a short
time later it restores the thresholds. During that interval, the
down-interval may expire and the interrupt be suppressed.

Now aware of the dance taking place within the GPU when coming out of
RC6, one wonders what other writes need to be queued in the fifo buffer
in order to be properly sequenced; setting the RP state appears to be
one.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=44006
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
drivers/gpu/drm/i915/i915_irq.c
drivers/gpu/drm/i915/intel_pm.c

index cc4a633076110bce0f844d1c24e8c739c9b6917c..1417660a93ec00a0a8a24cc797acc7b7754db063 100644 (file)
@@ -350,8 +350,8 @@ static void gen6_pm_rps_work(struct work_struct *work)
 {
        drm_i915_private_t *dev_priv = container_of(work, drm_i915_private_t,
                                                    rps_work);
-       u8 new_delay = dev_priv->cur_delay;
        u32 pm_iir, pm_imr;
+       u8 new_delay;
 
        spin_lock_irq(&dev_priv->rps_lock);
        pm_iir = dev_priv->pm_iir;
@@ -360,41 +360,18 @@ static void gen6_pm_rps_work(struct work_struct *work)
        I915_WRITE(GEN6_PMIMR, 0);
        spin_unlock_irq(&dev_priv->rps_lock);
 
-       if (!pm_iir)
+       if ((pm_iir & GEN6_PM_DEFERRED_EVENTS) == 0)
                return;
 
        mutex_lock(&dev_priv->dev->struct_mutex);
-       if (pm_iir & GEN6_PM_RP_UP_THRESHOLD) {
-               if (dev_priv->cur_delay != dev_priv->max_delay)
-                       new_delay = dev_priv->cur_delay + 1;
-               if (new_delay > dev_priv->max_delay)
-                       new_delay = dev_priv->max_delay;
-       } else if (pm_iir & (GEN6_PM_RP_DOWN_THRESHOLD | GEN6_PM_RP_DOWN_TIMEOUT)) {
-               gen6_gt_force_wake_get(dev_priv);
-               if (dev_priv->cur_delay != dev_priv->min_delay)
-                       new_delay = dev_priv->cur_delay - 1;
-               if (new_delay < dev_priv->min_delay) {
-                       new_delay = dev_priv->min_delay;
-                       I915_WRITE(GEN6_RP_INTERRUPT_LIMITS,
-                                  I915_READ(GEN6_RP_INTERRUPT_LIMITS) |
-                                  ((new_delay << 16) & 0x3f0000));
-               } else {
-                       /* Make sure we continue to get down interrupts
-                        * until we hit the minimum frequency */
-                       I915_WRITE(GEN6_RP_INTERRUPT_LIMITS,
-                                  I915_READ(GEN6_RP_INTERRUPT_LIMITS) & ~0x3f0000);
-               }
-               gen6_gt_force_wake_put(dev_priv);
-       }
+
+       if (pm_iir & GEN6_PM_RP_UP_THRESHOLD)
+               new_delay = dev_priv->cur_delay + 1;
+       else
+               new_delay = dev_priv->cur_delay - 1;
 
        gen6_set_rps(dev_priv->dev, new_delay);
-       dev_priv->cur_delay = new_delay;
 
-       /*
-        * rps_lock not held here because clearing is non-destructive. There is
-        * an *extremely* unlikely race with gen6_rps_enable() that is prevented
-        * by holding struct_mutex for the duration of the write.
-        */
        mutex_unlock(&dev_priv->dev->struct_mutex);
 }
 
index 8e79ff67ec98931e6a7680d8b9e28375d8db1978..bbca4e403b84bc3f67cb30bb26c72ad9e9672369 100644 (file)
@@ -2270,10 +2270,33 @@ void ironlake_disable_drps(struct drm_device *dev)
 void gen6_set_rps(struct drm_device *dev, u8 val)
 {
        struct drm_i915_private *dev_priv = dev->dev_private;
-       u32 swreq;
+       u32 limits;
 
-       swreq = (val & 0x3ff) << 25;
-       I915_WRITE(GEN6_RPNSWREQ, swreq);
+       limits = 0;
+       if (val >= dev_priv->max_delay)
+               val = dev_priv->max_delay;
+       else
+               limits |= dev_priv->max_delay << 24;
+
+       if (val <= dev_priv->min_delay)
+               val = dev_priv->min_delay;
+       else
+               limits |= dev_priv->min_delay << 16;
+
+       if (val == dev_priv->cur_delay)
+               return;
+
+       I915_WRITE(GEN6_RPNSWREQ,
+                  GEN6_FREQUENCY(val) |
+                  GEN6_OFFSET(0) |
+                  GEN6_AGGRESSIVE_TURBO);
+
+       /* Make sure we continue to get interrupts
+        * until we hit the minimum or maximum frequencies.
+        */
+       I915_WRITE(GEN6_RP_INTERRUPT_LIMITS, limits);
+
+       dev_priv->cur_delay = val;
 }
 
 void gen6_disable_rps(struct drm_device *dev)
@@ -2327,11 +2350,10 @@ int intel_enable_rc6(const struct drm_device *dev)
 void gen6_enable_rps(struct drm_i915_private *dev_priv)
 {
        struct intel_ring_buffer *ring;
-       u32 rp_state_cap = I915_READ(GEN6_RP_STATE_CAP);
-       u32 gt_perf_status = I915_READ(GEN6_GT_PERF_STATUS);
+       u32 rp_state_cap;
+       u32 gt_perf_status;
        u32 pcu_mbox, rc6_mask = 0;
        u32 gtfifodbg;
-       int cur_freq, min_freq, max_freq;
        int rc6_mode;
        int i;
 
@@ -2352,6 +2374,14 @@ void gen6_enable_rps(struct drm_i915_private *dev_priv)
 
        gen6_gt_force_wake_get(dev_priv);
 
+       rp_state_cap = I915_READ(GEN6_RP_STATE_CAP);
+       gt_perf_status = I915_READ(GEN6_GT_PERF_STATUS);
+
+       /* In units of 100MHz */
+       dev_priv->max_delay = rp_state_cap & 0xff;
+       dev_priv->min_delay = (rp_state_cap & 0xff0000) >> 16;
+       dev_priv->cur_delay = 0;
+
        /* disable the counters and set deterministic thresholds */
        I915_WRITE(GEN6_RC_CONTROL, 0);
 
@@ -2399,8 +2429,8 @@ void gen6_enable_rps(struct drm_i915_private *dev_priv)
 
        I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 1000000);
        I915_WRITE(GEN6_RP_INTERRUPT_LIMITS,
-                  18 << 24 |
-                  6 << 16);
+                  dev_priv->max_delay << 24 |
+                  dev_priv->min_delay << 16);
        I915_WRITE(GEN6_RP_UP_THRESHOLD, 10000);
        I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 1000000);
        I915_WRITE(GEN6_RP_UP_EI, 100000);
@@ -2426,10 +2456,6 @@ void gen6_enable_rps(struct drm_i915_private *dev_priv)
                     500))
                DRM_ERROR("timeout waiting for pcode mailbox to finish\n");
 
-       min_freq = (rp_state_cap & 0xff0000) >> 16;
-       max_freq = rp_state_cap & 0xff;
-       cur_freq = (gt_perf_status & 0xff00) >> 8;
-
        /* Check for overclock support */
        if (wait_for((I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) == 0,
                     500))
@@ -2440,14 +2466,11 @@ void gen6_enable_rps(struct drm_i915_private *dev_priv)
                     500))
                DRM_ERROR("timeout waiting for pcode mailbox to finish\n");
        if (pcu_mbox & (1<<31)) { /* OC supported */
-               max_freq = pcu_mbox & 0xff;
+               dev_priv->max_delay = pcu_mbox & 0xff;
                DRM_DEBUG_DRIVER("overclocking supported, adjusting frequency max to %dMHz\n", pcu_mbox * 50);
        }
 
-       /* In units of 100MHz */
-       dev_priv->max_delay = max_freq;
-       dev_priv->min_delay = min_freq;
-       dev_priv->cur_delay = cur_freq;
+       gen6_set_rps(dev_priv->dev, (gt_perf_status & 0xff00) >> 8);
 
        /* requires MSI enabled */
        I915_WRITE(GEN6_PMIER,