Skip to contentSkip to conversation

Resume time optimisation

PATCH In review David Weinehall 5 participants 12 messages Nov 18, 2016 – Nov 28, 2016

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

v2 cover Resume time optimisation

This patch series aims to reduce the time we spend in i915_gem_restore_gtt_mappings() on resume.

Now that we're getting the resume times down a bit this function started to stand out.

David Weinehall (3):
drm/i915: Cleanup i915_gem_restore_gtt_mappings() drm/i915: Take runtime pm in i915_gem_resume() drm/i915: optimise intel_runtime_pm_{get,put}

 drivers/gpu/drm/i915/i915_gem.c         |  2 ++
 drivers/gpu/drm/i915/i915_gem_gtt.c     | 17 +++++++++++------
 drivers/gpu/drm/i915/intel_runtime_pm.c | 13 ++++++++++++-
 3 files changed, 25 insertions(+), 7 deletions(-)
Signature · 6 lines
2.10.2

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#1 lore raw

PATCH v2 1/3 drm/i915: Cleanup i915_gem_restore_gtt_mappings()

On resume we unbind+bind our VMA-mappings. This patch simplifies this a bit by introducing a restore_vma() helper. As a nice side-effect this also makes the resume callgraph self-documenting.

v2: move the helper to i915_gem_gtt.c since it's only used by
i915_gem_restore_gtt_mappings(), rename to restore_vma(), and make static (Chris)

---
 drivers/gpu/drm/i915/i915_gem_gtt.c | 17 +++++++++++------
 1 file changed, 11 insertions(+), 6 deletions(-)
drivers/gpu/drm/i915/i915_gem_gtt.c +11 -6
index b4bde1452f2a..2711044e3bf2 100644
@@ -3277,6 +3277,16 @@ int i915_ggtt_enable_hw(struct drm_i915_private *dev_priv)
 	return 0;
 }
 
+static bool restore_vma(struct i915_vma *vma)
+{
+	if (i915_vma_is_pinned(vma)) {
+		WARN_ON(i915_vma_bind(vma, vma->obj->cache_level, PIN_UPDATE));
+		return true;
+	} else {
+		WARN_ON(i915_vma_unbind(vma));
+		return false;
+	}
+}
 void i915_gem_restore_gtt_mappings(struct drm_i915_private *dev_priv)
 {
 	struct i915_ggtt *ggtt = &dev_priv->ggtt;
@@ -3299,12 +3309,7 @@ void i915_gem_restore_gtt_mappings(struct drm_i915_private *dev_priv)
 			if (vma->vm != &ggtt->base)
 				continue;
 
-			if (!i915_vma_unbind(vma))
-				continue;
-
-			WARN_ON(i915_vma_bind(vma, obj->cache_level,
-					      PIN_UPDATE));
-			ggtt_bound = true;
+			ggtt_bound |= restore_vma(vma);
 		}
 
 		if (ggtt_bound)
Signature · 6 lines
2.10.2

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Reviewed-by on v2 1/3
#2 lore raw

Re: [PATCH v2 1/3] drm/i915: Cleanup i915_gem_restore_gtt_mappings()

Quoted 22 lines · David Weinehall

On resume we unbind+bind our VMA-mappings. This patch simplifies this a bit by introducing a restore_vma() helper. As a nice side-effect this also makes the resume callgraph self-documenting.

v2: move the helper to i915_gem_gtt.c since it's only used by
i915_gem_restore_gtt_mappings(), rename to restore_vma(), and make static (Chris)

---
 drivers/gpu/drm/i915/i915_gem_gtt.c | 17 +++++++++++------
 1 file changed, 11 insertions(+), 6 deletions(-)
drivers/gpu/drm/i915/i915_gem_gtt.c +0 -0
index b4bde1452f2a..2711044e3bf2 100644
@@ -3277,6 +3277,16 @@ int i915_ggtt_enable_hw(struct drm_i915_private *dev_priv)
 	return 0;
 }
 

drivers/gpu/drm/i915/i915_gem_gtt.c int i915_ggtt_enable_hw(struct drm_i915_private *dev_priv) · line 3289

Quoted 10 lines
drivers/gpu/drm/i915/i915_gem_gtt.c +10 -0
+static bool restore_vma(struct i915_vma *vma)
+{
+	if (i915_vma_is_pinned(vma)) {
+		WARN_ON(i915_vma_bind(vma, vma->obj->cache_level, PIN_UPDATE));
+		return true;
+	} else {
+		WARN_ON(i915_vma_unbind(vma));
+		return false;
+	}
+}

Missed a blank line [checkpatch]

-Chris

Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#3 lore raw

Re: [PATCH v2 1/3] drm/i915: Cleanup i915_gem_restore_gtt_mappings()

Quoted 23 lines · Chris Wilson
Quoted 22 lines · David Weinehall

On resume we unbind+bind our VMA-mappings. This patch simplifies this a bit by introducing a restore_vma() helper. As a nice side-effect this also makes the resume callgraph self-documenting.

v2: move the helper to i915_gem_gtt.c since it's only used by
i915_gem_restore_gtt_mappings(), rename to restore_vma(), and make static (Chris)

---
 drivers/gpu/drm/i915/i915_gem_gtt.c | 17 +++++++++++------
 1 file changed, 11 insertions(+), 6 deletions(-)
drivers/gpu/drm/i915/i915_gem_gtt.c +0 -0
index b4bde1452f2a..2711044e3bf2 100644
@@ -3277,6 +3277,16 @@ int i915_ggtt_enable_hw(struct drm_i915_private *dev_priv)
 	return 0;
 }
 
Quoted 12 lines

drivers/gpu/drm/i915/i915_gem_gtt.c int i915_ggtt_enable_hw(struct drm_i915_private *dev_priv) · line 3289

Quoted 10 lines
drivers/gpu/drm/i915/i915_gem_gtt.c +10 -0
+static bool restore_vma(struct i915_vma *vma)
+{
+	if (i915_vma_is_pinned(vma)) {
+		WARN_ON(i915_vma_bind(vma, vma->obj->cache_level, PIN_UPDATE));
+		return true;
+	} else {
+		WARN_ON(i915_vma_unbind(vma));
+		return false;
+	}
+}

Missed a blank line [checkpatch]

Dang.

Will you fix this manually for the merge, or do you want a fixed patch?

Kind regards, David
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#4 lore raw

Re: [PATCH v2 1/3] drm/i915: Cleanup i915_gem_restore_gtt_mappings()

Quoted 37 lines · David Weinehall
Quoted 23 lines · Chris Wilson
Quoted 22 lines · David Weinehall

On resume we unbind+bind our VMA-mappings. This patch simplifies this a bit by introducing a restore_vma() helper. As a nice side-effect this also makes the resume callgraph self-documenting.

v2: move the helper to i915_gem_gtt.c since it's only used by
i915_gem_restore_gtt_mappings(), rename to restore_vma(), and make static (Chris)

---
 drivers/gpu/drm/i915/i915_gem_gtt.c | 17 +++++++++++------
 1 file changed, 11 insertions(+), 6 deletions(-)
drivers/gpu/drm/i915/i915_gem_gtt.c +0 -0
index b4bde1452f2a..2711044e3bf2 100644
@@ -3277,6 +3277,16 @@ int i915_ggtt_enable_hw(struct drm_i915_private *dev_priv)
 	return 0;
 }
 
Quoted 12 lines

drivers/gpu/drm/i915/i915_gem_gtt.c int i915_ggtt_enable_hw(struct drm_i915_private *dev_priv) · line 3289

Quoted 10 lines
drivers/gpu/drm/i915/i915_gem_gtt.c +10 -0
+static bool restore_vma(struct i915_vma *vma)
+{
+	if (i915_vma_is_pinned(vma)) {
+		WARN_ON(i915_vma_bind(vma, vma->obj->cache_level, PIN_UPDATE));
+		return true;
+	} else {
+		WARN_ON(i915_vma_unbind(vma));
+		return false;
+	}
+}

Missed a blank line [checkpatch]

Quoted 3 lines

Dang.

Will you fix this manually for the merge, or do you want a fixed patch?

I was looking at how we get out of the pm_runtime mess. In part, we hurt ourselves because we are using wakeref_count to disable asserts, but it also seems that pushing the optimisation to pm_runtime is the right thing to do.
-Chris

Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#5 lore raw

Re: [PATCH v2 1/3] drm/i915: Cleanup i915_gem_restore_gtt_mappings()

Quoted 5 lines · Chris Wilson

I was looking at how we get out of the pm_runtime mess. In part, we hurt ourselves because we are using wakeref_count to disable asserts, but it also seems that pushing the optimisation to pm_runtime is the right thing to do.

Any more thoughts on this?

Kind regards, David Weinehall
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#6 lore raw

PATCH v2 2/3 drm/i915: Take runtime pm in i915_gem_resume()

In reply to David Weinehall

In i915_gem_resume(), before calling i915_gem_restore_gtt_mappings(), we want to take the runtime PM wakelocks for the entire duration instead of many short lived wakelocks around each access to the GGTT. We will take it in i915_gem_resume() as any other GEM operation is likely to require the gT powerwell as well.

v2: Better explanation of the patch (Chris)

---
 drivers/gpu/drm/i915/i915_gem.c | 2 ++
 1 file changed, 2 insertions(+)
drivers/gpu/drm/i915/i915_gem.c +2 -0
index d5b7723bb028..da6e949906ca 100644
@@ -4240,6 +4240,7 @@ void i915_gem_resume(struct drm_device *dev)
 
 	WARN_ON(dev_priv->gt.awake);
 
+	intel_runtime_pm_get(dev_priv);
 	mutex_lock(&dev->struct_mutex);
 	i915_gem_restore_gtt_mappings(dev_priv);
 
@@ -4250,6 +4251,7 @@ void i915_gem_resume(struct drm_device *dev)
 	dev_priv->gt.resume(dev_priv);
 
 	mutex_unlock(&dev->struct_mutex);
+	intel_runtime_pm_put(dev_priv);
 }
 
 void i915_gem_init_swizzling(struct drm_i915_private *dev_priv)
Signature · 6 lines
2.10.2

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#7 lore raw

PATCH v2 3/3 drm/i915: optimise intel_runtime_pm_{get, put}

In reply to David Weinehall

Benchmarking shows that on resume we spend quite a bit of time just taking and dropping these references, leaving us two options; either rewriting the code not to take these references more than once, which would be a rather invasive change since the involved functions are used from other places, or to optimise intel_runtime_pm_{get,put}(). This patch does the latter. Initial benchmarking indicate improvements of a couple of milliseconds on resume.

Original patch by Chris, with slight fixes by me.

v2: Fix missing return value (Patchwork)
Remove extra atomic_dec() (Chris)

---
 drivers/gpu/drm/i915/intel_runtime_pm.c | 13 ++++++++++++-
 1 file changed, 12 insertions(+), 1 deletion(-)
drivers/gpu/drm/i915/intel_runtime_pm.c +12 -1
index 356c662ad453..831fde343d10 100644
@@ -2632,6 +2632,9 @@ void intel_runtime_pm_get(struct drm_i915_private *dev_priv)
 	struct pci_dev *pdev = dev_priv->drm.pdev;
 	struct device *kdev = &pdev->dev;
 
+	if (atomic_inc_not_zero(&dev_priv->pm.wakeref_count))
+		return;
+
 	pm_runtime_get_sync(kdev);
 
 	atomic_inc(&dev_priv->pm.wakeref_count);
@@ -2653,6 +2656,9 @@ bool intel_runtime_pm_get_if_in_use(struct drm_i915_private *dev_priv)
 	struct pci_dev *pdev = dev_priv->drm.pdev;
 	struct device *kdev = &pdev->dev;
 
+	if (atomic_inc_not_zero(&dev_priv->pm.wakeref_count))
+		return true;
+
 	if (IS_ENABLED(CONFIG_PM)) {
 		int ret = pm_runtime_get_if_in_use(kdev);
 
@@ -2695,6 +2701,9 @@ void intel_runtime_pm_get_noresume(struct drm_i915_private *dev_priv)
 	struct pci_dev *pdev = dev_priv->drm.pdev;
 	struct device *kdev = &pdev->dev;
 
+	if (atomic_inc_not_zero(&dev_priv->pm.wakeref_count))
+		return;
+
 	assert_rpm_wakelock_held(dev_priv);
 	pm_runtime_get_noresume(kdev);
 
@@ -2714,8 +2723,10 @@ void intel_runtime_pm_put(struct drm_i915_private *dev_priv)
 	struct pci_dev *pdev = dev_priv->drm.pdev;
 	struct device *kdev = &pdev->dev;
 
+	if (!atomic_dec_and_test(&dev_priv->pm.wakeref_count))
+		return;
+
 	assert_rpm_wakelock_held(dev_priv);
-	atomic_dec(&dev_priv->pm.wakeref_count);
 
 	pm_runtime_mark_last_busy(kdev);
 	pm_runtime_put_autosuspend(kdev);
Signature · 6 lines
2.10.2

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#8 lore raw

Re: [PATCH v2 3/3] drm/i915: optimise intel_runtime_pm_{get, put}

Quoted 9 lines · David Weinehall

Benchmarking shows that on resume we spend quite a bit of time just taking and dropping these references, leaving us two options; either rewriting the code not to take these references more than once, which would be a rather invasive change since the involved functions are used from other places, or to optimise intel_runtime_pm_{get,put}(). This patch does the latter. Initial benchmarking indicate improvements of a couple of milliseconds on resume.

Quoted 7 lines

Original patch by Chris, with slight fixes by me.

v2: Fix missing return value (Patchwork)
Remove extra atomic_dec() (Chris)

Cc Imre Deak <[email protected]>

I'm happy with this. Not amused that it apparently saves quite a bit of overhead with frequent pm_runtime calls.

Imre?
-Chris

Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#9 lore raw

Re: [PATCH v2 3/3] drm/i915: optimise intel_runtime_pm_{get, put}

Quoted 15 lines · Chris Wilson
Quoted 14 lines · David Weinehall

Benchmarking shows that on resume we spend quite a bit of time just taking and dropping these references, leaving us two options; either rewriting the code not to take these references more than once, which would be a rather invasive change since the involved functions are used from other places, or to optimise intel_runtime_pm_{get,put}(). This patch does the latter. Initial benchmarking indicate improvements of a couple of milliseconds on resume.

Original patch by Chris, with slight fixes by me.

v2: Fix missing return value (Patchwork)
Remove extra atomic_dec() (Chris)

Quoted 6 lines
Quoted 2 lines
Cc Imre Deak <[email protected]>

I'm happy with this. Not amused that it apparently saves quite a bit of overhead with frequent pm_runtime calls.

We could eliminate some of those calls entirely by moving them from intel_display_power_{get,put}() into the always on well enable/disable hooks. But I'm not sure how much this overhead originates from the power well code as opposed to some gem/etc. stuff.

Quoted 10 lines

Imre?
-Chris

Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Signature · 6 lines
Ville Syrjälä
Intel OTC
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#10 lore raw

Re: [PATCH v2 3/3] drm/i915: optimise intel_runtime_pm_{get, put}

In reply to Chris Wilson
Quoted 17 lines · Chris Wilson
Quoted 9 lines · David Weinehall

Benchmarking shows that on resume we spend quite a bit of time just taking and dropping these references, leaving us two options; either rewriting the code not to take these references more than once, which would be a rather invasive change since the involved functions are used from other places, or to optimise intel_runtime_pm_{get,put}().  This patch does the latter. Initial benchmarking indicate improvements of a couple of milliseconds on resume.

Quoted 7 lines

Original patch by Chris, with slight fixes by me.

v2: Fix missing return value (Patchwork)
    Remove extra atomic_dec() (Chris)

Quoted 7 lines
Cc Imre Deak <[email protected]>

I'm happy with this. Not amused that it apparently saves quite a bit of
overhead with frequent pm_runtime calls.

Imre?

I think the overhead is because the RPM core takes a lock and checks if the device needs to be woken up even if the runtime_usage count is 0. But for us the device is awake whenever wakeref_count > 0, so yes we can optimize things.

Quoted 2 lines

-Chris

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

#11 lore raw

✗ Fi.CI.BAT: failure for Resume time optimisation (rev2)

In reply to David Weinehall
== Series Details ==

== Series Details ==

Series: Resume time optimisation (rev2)
URL : https://patchwork.freedesktop.org/series/15545/
State : failure

== Summary ==

Series 15545v2 Resume time optimisation
https://patchwork.freedesktop.org/api/1.0/series/15545/revisions/2/mbox/

Test pm_rpm:

        Subgroup basic-pci-d3-state:
                pass       -> INCOMPLETE (fi-bxt-t5700)
                pass       -> INCOMPLETE (fi-byt-j1900)
                pass       -> INCOMPLETE (fi-byt-n2820)
                pass       -> DMESG-WARN (fi-hsw-4770)
        Subgroup basic-rte:
                pass       -> INCOMPLETE (fi-bsw-n3050)
                pass       -> INCOMPLETE (fi-hsw-4770)
fi-bdw-5557u     total:244  pass:229  dwarn:0   dfail:0   fail:0   skip:15 
fi-bsw-n3050     total:213  pass:173  dwarn:0   dfail:0   fail:0   skip:39 
fi-bxt-t5700     total:212  pass:183  dwarn:0   dfail:0   fail:0   skip:28 
fi-byt-j1900     total:212  pass:183  dwarn:0   dfail:0   fail:0   skip:28 
fi-byt-n2820     total:212  pass:179  dwarn:0   dfail:0   fail:0   skip:32 
fi-hsw-4770      total:213  pass:191  dwarn:1   dfail:0   fail:0   skip:20 
fi-hsw-4770r     total:244  pass:224  dwarn:0   dfail:0   fail:0   skip:20 
fi-ilk-650       total:244  pass:191  dwarn:0   dfail:0   fail:0   skip:53 
fi-ivb-3520m     total:244  pass:222  dwarn:0   dfail:0   fail:0   skip:22 
fi-ivb-3770      total:244  pass:222  dwarn:0   dfail:0   fail:0   skip:22 
fi-kbl-7200u     total:244  pass:222  dwarn:0   dfail:0   fail:0   skip:22 
fi-skl-6260u     total:244  pass:230  dwarn:0   dfail:0   fail:0   skip:14 
fi-skl-6700hq    total:244  pass:223  dwarn:0   dfail:0   fail:0   skip:21 
fi-skl-6700k     total:244  pass:222  dwarn:1   dfail:0   fail:0   skip:21 
fi-skl-6770hq    total:244  pass:230  dwarn:0   dfail:0   fail:0   skip:14 
fi-snb-2520m     total:244  pass:212  dwarn:0   dfail:0   fail:0   skip:32 
fi-snb-2600      total:244  pass:211  dwarn:0   dfail:0   fail:0   skip:33 

05104adb7446f310fad4e36f22d4c84ffafa31fc drm-intel-nightly: 2016y-11m-18d-13h-10m-16s UTC integration manifest db3405c drm/i915: optimise intel_runtime_pm_{get, put}
1526407 drm/i915: Take runtime pm in i915_gem_resume()
c3ff9f5 drm/i915: Cleanup i915_gem_restore_gtt_mappings()

== Logs ==

For more details see: https://intel-gfx-ci.01.org/CI/Patchwork_3055/
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx