Skip to contentSkip to conversation

drm/i915: Use natural width type for VMA pin count

RFC In review Tvrtko Ursulin 4 participants 10 messages Apr 21, 2016 – Apr 26, 2016

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

RFC 1/3 drm/i915: Use natural width type for VMA pin count

From: Tvrtko Ursulin <[email protected]>

Using four bits for the pin count in the middle of the data structure just makes the compiler generate verbose code.

---
 drivers/gpu/drm/i915/i915_gem_gtt.h | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
drivers/gpu/drm/i915/i915_gem_gtt.h +3 -3
index d7dd3d8a8758..b4766c4c4bac 100644
@@ -213,9 +213,9 @@ struct i915_vma {
 	 * and the framebuffer code. When switching/pageflipping, the
 	 * framebuffer code has at most two buffers pinned per crtc.
 	 *
-	 * In the worst case this is 1 + 1 + 1 + 2*2 = 7. That would fit into 3
-	 * bits with absolutely no headroom. So use 4 bits. */
-	unsigned int pin_count:4;
+	 * In the worst case this is 1 + 1 + 1 + 2*2 = 7.
+	 */
+	unsigned int pin_count;
 #define DRM_I915_GEM_OBJECT_MAX_PIN_COUNT 0xf
 };
 
Signature · 6 lines
1.9.1

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

RFC 2/3 drm/i915: Track aggregate per-object VMA pin count

From: Tvrtko Ursulin <[email protected]>

This allows trivial (non-iterating) i915_gem_obj_is_pinned implementation which in turns prevents i915_gem_madvise_ioctl showing up in profiles in benchmarks like SynMark2/OglDrvCtx.

---
 drivers/gpu/drm/i915/i915_drv.h            |  9 ++++++++-
 drivers/gpu/drm/i915/i915_gem.c            | 15 +++++----------
 drivers/gpu/drm/i915/i915_gem_execbuffer.c |  4 +++-
 3 files changed, 16 insertions(+), 12 deletions(-)
drivers/gpu/drm/i915/i915_drv.h +8 -1
index 6f1e0f127c0a..589256e2ebe9 100644
@@ -2104,6 +2104,9 @@ struct drm_i915_gem_object {
 	/** List of VMAs backed by this object */
 	struct list_head vma_list;
 
+	/** Aggregate pin count of all VMAs backed by this object. */
+	unsigned int vma_pin_count;
+
 	/** Stolen memory for this object, instead of being backed by shmem. */
 	struct drm_mm_node *stolen;
 	struct list_head global_list;
@@ -3236,7 +3239,11 @@ i915_gem_obj_to_ggtt(struct drm_i915_gem_object *obj)
 {
 	return i915_gem_obj_to_ggtt_view(obj, &i915_ggtt_view_normal);
 }
-bool i915_gem_obj_is_pinned(struct drm_i915_gem_object *obj);
+
+static inline bool i915_gem_obj_is_pinned(struct drm_i915_gem_object *obj)
+{
+	return obj->vma_pin_count > 0;
+}
 
 /* Some GGTT VM helpers */
 static inline struct i915_hw_ppgtt *
drivers/gpu/drm/i915/i915_gem.c +5 -10
index 261a3ef72828..0549dea683e1 100644
@@ -4311,6 +4311,8 @@ i915_gem_object_do_pin(struct drm_i915_gem_object *obj,
 	}
 
 	vma->pin_count++;
+	obj->vma_pin_count++;
+
 	return 0;
 }
 
@@ -4351,6 +4353,7 @@ i915_gem_object_ggtt_unpin_view(struct drm_i915_gem_object *obj,
 	WARN_ON(!i915_gem_obj_ggtt_bound_view(obj, view));
 
 	--vma->pin_count;
+	--obj->vma_pin_count;
 }
 
 int
@@ -4583,6 +4586,8 @@ void i915_gem_free_object(struct drm_gem_object *gem_obj)
 	list_for_each_entry_safe(vma, next, &obj->vma_list, obj_link) {
 		int ret;
 
+		GEM_BUG_ON(obj->vma_pin_count < vma->pin_count);
+		obj->vma_pin_count -= vma->pin_count;
 		vma->pin_count = 0;
 		ret = i915_vma_unbind(vma);
 		if (WARN_ON(ret == -ERESTARTSYS)) {
@@ -5326,16 +5331,6 @@ unsigned long i915_gem_obj_size(struct drm_i915_gem_object *o,
 	return 0;
 }
 
-bool i915_gem_obj_is_pinned(struct drm_i915_gem_object *obj)
-{
-	struct i915_vma *vma;
-	list_for_each_entry(vma, &obj->vma_list, obj_link)
-		if (vma->pin_count > 0)
-			return true;
-
-	return false;
-}
-
 /* Like i915_gem_object_get_page(), but mark the returned page dirty */
 struct page *
 i915_gem_object_get_dirty_page(struct drm_i915_gem_object *obj, int n)
drivers/gpu/drm/i915/i915_gem_execbuffer.c +3 -1
index 6f4f2a6cdf93..b4c787763955 100644
@@ -218,8 +218,10 @@ i915_gem_execbuffer_unreserve_vma(struct i915_vma *vma)
 	if (entry->flags & __EXEC_OBJECT_HAS_FENCE)
 		i915_gem_object_unpin_fence(obj);
 
-	if (entry->flags & __EXEC_OBJECT_HAS_PIN)
+	if (entry->flags & __EXEC_OBJECT_HAS_PIN) {
 		vma->pin_count--;
+		obj->vma_pin_count--;
+	}
 
 	entry->flags &= ~(__EXEC_OBJECT_HAS_FENCE | __EXEC_OBJECT_HAS_PIN);
 }
Signature · 6 lines
1.9.1

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

Re: [RFC 2/3] drm/i915: Track aggregate per-object VMA pin count

Quoted 5 lines · Tvrtko Ursulin

From: Tvrtko Ursulin <[email protected]>

This allows trivial (non-iterating) i915_gem_obj_is_pinned implementation which in turns prevents i915_gem_madvise_ioctl showing up in profiles in benchmarks like SynMark2/OglDrvCtx.

madvise doesn't need to check is-pinned at all.

And I prefer obj->bind_count.
-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: [RFC 2/3] drm/i915: Track aggregate per-object VMA pin count

Quoted 8 lines · Chris Wilson
Quoted 5 lines · Tvrtko Ursulin

From: Tvrtko Ursulin <[email protected]>

This allows trivial (non-iterating) i915_gem_obj_is_pinned implementation which in turns prevents i915_gem_madvise_ioctl showing up in profiles in benchmarks like SynMark2/OglDrvCtx.

madvise doesn't need to check is-pinned at all.

And if you are keen on solving the madvise contention, I've some patches to migrate it from struct_mutex to its own obj->mm.mutex. The patch is in very early (functional) state:
https://cgit.freedesktop.org/~ickle/linux-2.6/commit/?h=tasklet&id=239350fdd06f5aca4112c3d046979b281b7adf60
-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
#4 lore raw

RFC 3/3 drm/i915: Micro-optimize i915_gem_obj_to_vma

In reply to Tvrtko Ursulin

From: Tvrtko Ursulin <[email protected]>

i915_gem_obj_to_vma is one of the most expensive functions in our profiles. Could avoiding some branching by replacing it with arithmetic be beneficial? Some benchmarks suggest it slightly might.

---
 drivers/gpu/drm/i915/i915_gem.c | 14 ++++++++++++--
 1 file changed, 12 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/i915_gem.c +12 -2
index 0549dea683e1..243bfb922eb3 100644
@@ -4642,11 +4642,21 @@ struct i915_vma *i915_gem_obj_to_vma(struct drm_i915_gem_object *obj,
 				     struct i915_address_space *vm)
 {
 	struct i915_vma *vma;
+
+	BUILD_BUG_ON(I915_GGTT_VIEW_NORMAL != 0);
+
 	list_for_each_entry(vma, &obj->vma_list, obj_link) {
-		if (vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL &&
-		    vma->vm == vm)
+		/*
+		 * Below is just a branching avoiding way of saying:
+		 * vma_ggtt_view.type == I915_GGTT_VIEW_NORMAL && vma->vm == vm,
+		 * which relies on the fact I915_GGTT_VIEW_NORMAL has to be
+		 * zero.
+		 */
+		if (!((unsigned long)vma->ggtt_view.type |
+		    ((unsigned long)vma->vm ^ (unsigned long)vm)))
 			return vma;
 	}
+
 	return NULL;
 }
 
Signature · 6 lines
1.9.1

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

Re: [RFC 3/3] drm/i915: Micro-optimize i915_gem_obj_to_vma

Quoted 6 lines · Tvrtko Ursulin

From: Tvrtko Ursulin <[email protected]>

i915_gem_obj_to_vma is one of the most expensive functions in our profiles. Could avoiding some branching by replacing it with arithmetic be beneficial? Some benchmarks suggest it slightly might.

We can do much better by changing the algorithm here (here and higher up the call chain).
-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
#6 lore raw

Re: [RFC 3/3] drm/i915: Micro-optimize i915_gem_obj_to_vma

In reply to Tvrtko Ursulin
Quoted 29 lines · Tvrtko Ursulin

From: Tvrtko Ursulin <[email protected]>

i915_gem_obj_to_vma is one of the most expensive functions in our profiles. Could avoiding some branching by replacing it with arithmetic be beneficial? Some benchmarks suggest it slightly might.

---
  drivers/gpu/drm/i915/i915_gem.c | 14 ++++++++++++--
  1 file changed, 12 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/i915_gem.c +6 -2
index 0549dea683e1..243bfb922eb3 100644
@@ -4642,11 +4642,21 @@ struct i915_vma *i915_gem_obj_to_vma(struct drm_i915_gem_object *obj,
  				     struct i915_address_space *vm)
  {
  	struct i915_vma *vma;
+
+	BUILD_BUG_ON(I915_GGTT_VIEW_NORMAL != 0);
+
  	list_for_each_entry(vma, &obj->vma_list, obj_link) {
-		if (vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL &&
-		    vma->vm == vm)
+		/*
+		 * Below is just a branching avoiding way of saying:
+		 * vma_ggtt_view.type == I915_GGTT_VIEW_NORMAL && vma->vm == vm,

drivers/gpu/drm/i915/i915_gem.c struct i915_vma *i915_gem_obj_to_vma(struct drm_i915_gem_object *obj, · line 4661

Quoted 10 lines
drivers/gpu/drm/i915/i915_gem.c +6 -0
+		 * which relies on the fact I915_GGTT_VIEW_NORMAL has to be
+		 * zero.
+		 */
+		if (!((unsigned long)vma->ggtt_view.type |
+		    ((unsigned long)vma->vm ^ (unsigned long)vm)))
  			return vma;
  	}
+
  	return NULL;
  }

Other alternatives might include splitting the vma_list, so that we have one list for the most-frequently searched-for entries (GGTT view NORMAL) and for everything else, so the above would just need a single test for equality.

Or, slightly less effectively, add GGTT/NORMAL entries at the head of the list and others at the tail (and search backwards if you *don't* want a GGTT/NORMAL entry). That would still need the comparisons, but would likely hit an early match.

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

#7 lore raw

Re: [RFC 3/3] drm/i915: Micro-optimize i915_gem_obj_to_vma

Quoted 46 lines · Dave Gordon
Quoted 29 lines · Tvrtko Ursulin

From: Tvrtko Ursulin <[email protected]>

i915_gem_obj_to_vma is one of the most expensive functions in our profiles. Could avoiding some branching by replacing it with arithmetic be beneficial? Some benchmarks suggest it slightly might.

---
 drivers/gpu/drm/i915/i915_gem.c | 14 ++++++++++++--
 1 file changed, 12 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/i915_gem.c +6 -2
index 0549dea683e1..243bfb922eb3 100644
@@ -4642,11 +4642,21 @@ struct i915_vma *i915_gem_obj_to_vma(struct drm_i915_gem_object *obj,
 				     struct i915_address_space *vm)
 {
 	struct i915_vma *vma;
+
+	BUILD_BUG_ON(I915_GGTT_VIEW_NORMAL != 0);
+
 	list_for_each_entry(vma, &obj->vma_list, obj_link) {
-		if (vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL &&
-		    vma->vm == vm)
+		/*
+		 * Below is just a branching avoiding way of saying:
+		 * vma_ggtt_view.type == I915_GGTT_VIEW_NORMAL && vma->vm == vm,

drivers/gpu/drm/i915/i915_gem.c struct i915_vma *i915_gem_obj_to_vma(struct drm_i915_gem_object *obj, · line 4661

Quoted 10 lines
drivers/gpu/drm/i915/i915_gem.c +6 -0
+		 * which relies on the fact I915_GGTT_VIEW_NORMAL has to be
+		 * zero.
+		 */
+		if (!((unsigned long)vma->ggtt_view.type |
+		    ((unsigned long)vma->vm ^ (unsigned long)vm)))
 			return vma;
 	}
+
 	return NULL;
 }

Other alternatives might include splitting the vma_list, so that we have one list for the most-frequently searched-for entries (GGTT view NORMAL) and for everything else, so the above would just need a single test for equality.

Quoted 4 lines

Or, slightly less effectively, add GGTT/NORMAL entries at the head of the list and others at the tail (and search backwards if you *don't* want a GGTT/NORMAL entry). That would still need the comparisons, but would likely hit an early match.

We want one list for convenience elsewhere, but can keep a rht in parallel. This is not as effective/important as keeping a hashtable to translate from handle to vma, but is still useful for some stress cases. -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
#8 lore raw
In reply to Tvrtko Ursulin
Quoted 4 lines · Tvrtko Ursulin

From: Tvrtko Ursulin <[email protected]>

Using four bits for the pin count in the middle of the data structure just makes the compiler generate verbose code.

Or rather by opening coding this we can do much better and simultaneously checking pin and binding inside the hot i915_vma_pin function.
-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

✓ Fi.CI.BAT: success for series starting with [RFC,1/3] drm/i915: Use natural width type for VMA pin count

In reply to Tvrtko Ursulin
== Series Details ==

== Series Details ==

Series: series starting with [RFC,1/3] drm/i915: Use natural width type for VMA pin count
URL : https://patchwork.freedesktop.org/series/6061/
State : success

== Summary ==

Series 6061v1 Series without cover letter
http://patchwork.freedesktop.org/api/1.0/series/6061/revisions/1/mbox/

bdw-nuci7        total:193  pass:181  dwarn:0   dfail:0   fail:0   skip:12 
bdw-ultra        total:193  pass:170  dwarn:0   dfail:0   fail:0   skip:23 
bsw-nuc-2        total:192  pass:153  dwarn:0   dfail:0   fail:0   skip:39 
byt-nuc          total:192  pass:154  dwarn:0   dfail:0   fail:0   skip:38 
ilk-hp8440p      total:193  pass:136  dwarn:0   dfail:0   fail:0   skip:57 
ivb-t430s        total:193  pass:165  dwarn:0   dfail:0   fail:0   skip:28 
skl-i7k-2        total:193  pass:168  dwarn:0   dfail:0   fail:0   skip:25 
skl-nuci5        total:193  pass:182  dwarn:0   dfail:0   fail:0   skip:11 
snb-dellxps      total:193  pass:155  dwarn:0   dfail:0   fail:0   skip:38 

snb-x220t failed to connect after reboot

Results at /archive/results/CI_IGT_test/Patchwork_2022/

340c485ad98d0ec0369a3b18d4a09938f3f5537d drm-intel-nightly: 2016y-04m-22d-17h-32m-25s UTC integration manifest 88b2604 drm/i915: Micro-optimize i915_gem_obj_to_vma
1e27bbf drm/i915: Track aggregate per-object VMA pin count
8a4b065 drm/i915: Use natural width type for VMA pin count

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