Skip to contentSkip to conversation

GTT cleanups, rebased

PATCH Applied Ben Widawsky 3 participants 20 messages Jan 26, 2013 – Jan 29, 2013

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

cover GTT cleanups, rebased

This work sort of led me on the gtt vtable distraction which ended in Daniel doing his own version. I posted these originally about a month ago. I've rebased them on dinq. I'd like to see them get merged since they still help with some other work I'm doing on PPGTT.

Some of this work does conflict a bit with the gtt vtable stuff.

I did a decent amount of testing on the original post, less so on this rebased version.

Ben Widawsky (5):
drm/i915: trivial: kill-agp collateral cleanups
drm/i915: Reclaim GTT space for failed PPGTT
drm/i915: Extract gen6 aliasing ppgtt code
drm/i915: Aliased PPGTT size abstraction
drm/i915: Dynamically calculate dclv

 drivers/gpu/drm/i915/i915_debugfs.c |   2 +
 drivers/gpu/drm/i915/i915_drv.h     |   6 +--
 drivers/gpu/drm/i915/i915_gem_gtt.c | 101 +++++++++++++++++++++++-------------
 drivers/gpu/drm/i915/i915_reg.h     |   1 -
 4 files changed, 71 insertions(+), 39 deletions(-)
Signature · 1 line
1.8.1.1
#1 lore raw

PATCH 1/5 drm/i915: trivial: kill-agp collateral cleanups

- i915_gem_init_aliasing_ppgtt should now be static
- move i915_gem_init_ppgtt declaration to the right place

---
 drivers/gpu/drm/i915/i915_drv.h     | 3 +--
 drivers/gpu/drm/i915/i915_gem_gtt.c | 2 +-
 2 files changed, 2 insertions(+), 3 deletions(-)
drivers/gpu/drm/i915/i915_drv.h +1 -2
index 09753e5..1af8e73 100644
@@ -1598,7 +1598,6 @@ int __must_check i915_gem_init(struct drm_device *dev);
 int __must_check i915_gem_init_hw(struct drm_device *dev);
 void i915_gem_l3_remap(struct drm_device *dev);
 void i915_gem_init_swizzling(struct drm_device *dev);
-void i915_gem_init_ppgtt(struct drm_device *dev);
 void i915_gem_cleanup_ringbuffer(struct drm_device *dev);
 int __must_check i915_gpu_idle(struct drm_device *dev);
 int __must_check i915_gem_idle(struct drm_device *dev);
@@ -1653,7 +1652,7 @@ int i915_gem_context_destroy_ioctl(struct drm_device *dev, void *data,
 				   struct drm_file *file);
 
 /* i915_gem_gtt.c */
-int __must_check i915_gem_init_aliasing_ppgtt(struct drm_device *dev);
+void i915_gem_init_ppgtt(struct drm_device *dev);
 void i915_gem_cleanup_aliasing_ppgtt(struct drm_device *dev);
 void i915_ppgtt_bind_object(struct i915_hw_ppgtt *ppgtt,
 			    struct drm_i915_gem_object *obj,
drivers/gpu/drm/i915/i915_gem_gtt.c +1 -1
index a0ba4a9..462dec9 100644
@@ -108,7 +108,7 @@ static void i915_ppgtt_clear_range(struct i915_hw_ppgtt *ppgtt,
 	}
 }
 
-int i915_gem_init_aliasing_ppgtt(struct drm_device *dev)
+static int i915_gem_init_aliasing_ppgtt(struct drm_device *dev)
 {
 	struct drm_i915_private *dev_priv = dev->dev_private;
 	struct i915_hw_ppgtt *ppgtt;
Signature · 1 line
1.8.1.1
#2 lore raw

Re: [PATCH 1/5] drm/i915: trivial: kill-agp collateral cleanups

Quoted 5 lines · Ben Widawsky

- i915_gem_init_aliasing_ppgtt should now be static
- move i915_gem_init_ppgtt declaration to the right place

I've tried to apply this patch on top of latest dinq, and there's nothing left after resolving conflicts. Have I screwed something up? -Daniel

Quoted 46 lines
---
 drivers/gpu/drm/i915/i915_drv.h     | 3 +--
 drivers/gpu/drm/i915/i915_gem_gtt.c | 2 +-
 2 files changed, 2 insertions(+), 3 deletions(-)
drivers/gpu/drm/i915/i915_drv.h +1 -2
index 09753e5..1af8e73 100644
@@ -1598,7 +1598,6 @@ int __must_check i915_gem_init(struct drm_device *dev);
 int __must_check i915_gem_init_hw(struct drm_device *dev);
 void i915_gem_l3_remap(struct drm_device *dev);
 void i915_gem_init_swizzling(struct drm_device *dev);
-void i915_gem_init_ppgtt(struct drm_device *dev);
 void i915_gem_cleanup_ringbuffer(struct drm_device *dev);
 int __must_check i915_gpu_idle(struct drm_device *dev);
 int __must_check i915_gem_idle(struct drm_device *dev);
@@ -1653,7 +1652,7 @@ int i915_gem_context_destroy_ioctl(struct drm_device *dev, void *data,
 				   struct drm_file *file);
 
 /* i915_gem_gtt.c */
-int __must_check i915_gem_init_aliasing_ppgtt(struct drm_device *dev);
+void i915_gem_init_ppgtt(struct drm_device *dev);
 void i915_gem_cleanup_aliasing_ppgtt(struct drm_device *dev);
 void i915_ppgtt_bind_object(struct i915_hw_ppgtt *ppgtt,
 			    struct drm_i915_gem_object *obj,
drivers/gpu/drm/i915/i915_gem_gtt.c +1 -1
index a0ba4a9..462dec9 100644
@@ -108,7 +108,7 @@ static void i915_ppgtt_clear_range(struct i915_hw_ppgtt *ppgtt,
 	}
 }
 
-int i915_gem_init_aliasing_ppgtt(struct drm_device *dev)
+static int i915_gem_init_aliasing_ppgtt(struct drm_device *dev)
 {
 	struct drm_i915_private *dev_priv = dev->dev_private;
 	struct i915_hw_ppgtt *ppgtt;
Signature · 6 lines
1.8.1.1

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
#3 lore raw

Re: [PATCH 1/5] drm/i915: trivial: kill-agp collateral cleanups

Quoted 10 lines · Daniel Vetter
Quoted 5 lines · Ben Widawsky

- i915_gem_init_aliasing_ppgtt should now be static
- move i915_gem_init_ppgtt declaration to the right place

I've tried to apply this patch on top of latest dinq, and there's nothing left after resolving conflicts. Have I screwed something up? -Daniel

No, I think you're right. I think it no longer applies.

Signature · 1 line
Ben Widawsky, Intel Open Source Technology Center
#4 lore raw

PATCH 2/5 drm/i915: Reclaim GTT space for failed PPGTT

In reply to Ben Widawsky

When the PPGTT init fails, we may as well reuse the space that we were reserving for the PPGTT PDEs.

This also fixes an extraneous mutex_unlock.

---
 drivers/gpu/drm/i915/i915_gem_gtt.c | 33 ++++++++++++++++-----------------
 1 file changed, 16 insertions(+), 17 deletions(-)
drivers/gpu/drm/i915/i915_gem_gtt.c +16 -17
index 462dec9..31c490e 100644
@@ -531,12 +531,20 @@ static void i915_gtt_color_adjust(struct drm_mm_node *node,
 			*end -= 4096;
 	}
 }
-
 void i915_gem_setup_global_gtt(struct drm_device *dev,
 			       unsigned long start,
 			       unsigned long mappable_end,
 			       unsigned long end)
 {
+	/* Let GEM Manage all of the aperture.
+	 *
+	 * However, leave one page at the end still bound to the scratch page.
+	 * There are a number of places where the hardware apparently prefetches
+	 * past the end of the object, and we've seen multiple hangs with the
+	 * GPU head pointer stuck in a batchbuffer bound at the last page of the
+	 * aperture.  One page should be enough to keep any prefetching inside
+	 * of the aperture.
+	 */
 	drm_i915_private_t *dev_priv = dev->dev_private;
 	struct drm_mm_node *entry;
 	struct drm_i915_gem_object *obj;
@@ -598,12 +606,12 @@ void i915_gem_init_global_gtt(struct drm_device *dev)
 {
 	struct drm_i915_private *dev_priv = dev->dev_private;
 	unsigned long gtt_size, mappable_size;
-	int ret;
 
 	gtt_size = dev_priv->mm.gtt->gtt_total_entries << PAGE_SHIFT;
 	mappable_size = dev_priv->gtt.mappable_end;
 
 	if (intel_enable_ppgtt(dev) && HAS_ALIASING_PPGTT(dev)) {
+		int ret;
 		/* PPGTT pdes are stolen from global gtt ptes, so shrink the
 		 * aperture accordingly when using aliasing ppgtt. */
 		gtt_size -= I915_PPGTT_PD_ENTRIES*PAGE_SIZE;
@@ -611,23 +619,14 @@ void i915_gem_init_global_gtt(struct drm_device *dev)
 		i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
 
 		ret = i915_gem_init_aliasing_ppgtt(dev);
-		if (ret) {
-			mutex_unlock(&dev->struct_mutex);
+		if (!ret)
 			return;
-		}
-	} else {
-		/* Let GEM Manage all of the aperture.
-		 *
-		 * However, leave one page at the end still bound to the scratch
-		 * page.  There are a number of places where the hardware
-		 * apparently prefetches past the end of the object, and we've
-		 * seen multiple hangs with the GPU head pointer stuck in a
-		 * batchbuffer bound at the last page of the aperture.  One page
-		 * should be enough to keep any prefetching inside of the
-		 * aperture.
-		 */
-		i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
+
+		DRM_ERROR("Aliased PPGTT setup failed %d\n", ret);
+		drm_mm_takedown(&dev_priv->mm.gtt_space);
+		gtt_size += I915_PPGTT_PD_ENTRIES*PAGE_SIZE;
 	}
+	i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
 }
 
 static int setup_scratch_page(struct drm_device *dev)
Signature · 1 line
1.8.1.1
#5 lore raw

Re: [PATCH 2/5] drm/i915: Reclaim GTT space for failed PPGTT

Quoted 7 lines · Ben Widawsky

When the PPGTT init fails, we may as well reuse the space that we were reserving for the PPGTT PDEs.

This also fixes an extraneous mutex_unlock.

Queued for -next, thanks for the patch.
-Daniel

Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
#6 lore raw

PATCH 3/5 drm/i915: Extract gen6 aliasing ppgtt code

In reply to Ben Widawsky

This refactor clearly identifies the GEN specific portion of the aliased ppgtt code. Aside from some of the gen specific parts being extracted, it also preps us for an upcoming patch that pulls out the size, which has some nice benefits.

v2: Fix wonky error handling (Chris)

---
 drivers/gpu/drm/i915/i915_gem_gtt.c | 42 ++++++++++++++++++++++++-------------
 1 file changed, 28 insertions(+), 14 deletions(-)
drivers/gpu/drm/i915/i915_gem_gtt.c +28 -14
index 31c490e..f79f427 100644
@@ -108,10 +108,10 @@ static void i915_ppgtt_clear_range(struct i915_hw_ppgtt *ppgtt,
 	}
 }
 
-static int i915_gem_init_aliasing_ppgtt(struct drm_device *dev)
+static int gen6_init_aliasing_ppgtt(struct i915_hw_ppgtt *ppgtt)
 {
+	struct drm_device *dev = ppgtt->dev;
 	struct drm_i915_private *dev_priv = dev->dev_private;
-	struct i915_hw_ppgtt *ppgtt;
 	unsigned first_pd_entry_in_global_pt;
 	int i;
 	int ret = -ENOMEM;
@@ -121,16 +121,12 @@ static int i915_gem_init_aliasing_ppgtt(struct drm_device *dev)
 	 * now. */
 	first_pd_entry_in_global_pt = dev_priv->mm.gtt->gtt_total_entries - I915_PPGTT_PD_ENTRIES;
 
-	ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
-	if (!ppgtt)
-		return ret;
-
-	ppgtt->dev = dev;
 	ppgtt->num_pd_entries = I915_PPGTT_PD_ENTRIES;
 	ppgtt->pt_pages = kzalloc(sizeof(struct page *)*ppgtt->num_pd_entries,
 				  GFP_KERNEL);
+
 	if (!ppgtt->pt_pages)
-		goto err_ppgtt;
+		return -ENOMEM;
 
 	for (i = 0; i < ppgtt->num_pd_entries; i++) {
 		ppgtt->pt_pages[i] = alloc_page(GFP_KERNEL);
@@ -157,15 +153,11 @@ static int i915_gem_init_aliasing_ppgtt(struct drm_device *dev)
 		ppgtt->pt_dma_addr[i] = pt_addr;
 	}
 
-	ppgtt->scratch_page_dma_addr = dev_priv->gtt.scratch_page_dma;
-
 	i915_ppgtt_clear_range(ppgtt, 0,
 			       ppgtt->num_pd_entries*I915_PPGTT_PT_ENTRIES);
 
 	ppgtt->pd_offset = (first_pd_entry_in_global_pt)*sizeof(gtt_pte_t);
 
-	dev_priv->mm.aliasing_ppgtt = ppgtt;
-
 	return 0;
 
 err_pd_pin:
@@ -180,9 +172,31 @@ err_pt_alloc:
 		if (ppgtt->pt_pages[i])
 			__free_page(ppgtt->pt_pages[i]);
 	}
+
 	kfree(ppgtt->pt_pages);
-err_ppgtt:
-	kfree(ppgtt);
+
+	return ret;
+}
+
+
+static int i915_gem_init_aliasing_ppgtt(struct drm_device *dev)
+{
+	struct drm_i915_private *dev_priv = dev->dev_private;
+	struct i915_hw_ppgtt *ppgtt;
+	int ret = -ENOMEM;
+
+	ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
+	if (!ppgtt)
+		return ret;
+
+	ppgtt->dev = dev;
+	ppgtt->scratch_page_dma_addr = dev_priv->gtt.scratch_page_dma;
+
+	ret = gen6_init_aliasing_ppgtt(ppgtt);
+	if (ret)
+		kfree(ppgtt);
+	else
+		dev_priv->mm.aliasing_ppgtt = ppgtt;
 
 	return ret;
 }
Signature · 1 line
1.8.1.1
#7 lore raw

Re: [PATCH 3/5] drm/i915: Extract gen6 aliasing ppgtt code

Quoted 8 lines · Ben Widawsky

This refactor clearly identifies the GEN specific portion of the aliased ppgtt code. Aside from some of the gen specific parts being extracted, it also preps us for an upcoming patch that pulls out the size, which has some nice benefits.

v2: Fix wonky error handling (Chris)

Functionally we seem to have this patch already with

commit e9ccfd7b94f1ca9b46b992a19d2e8c5af70e4c69

Author: Daniel Vetter <[email protected]>
Date:   Thu Jan 24 13:49:56 2013 -0800

drm/i915: extract hw ppgtt setup/cleanup code

Yeah, you're allowed to hate me with the heat of a thousand suns now ;-) -Daniel

Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
#8 lore raw

Re: [PATCH 3/5] drm/i915: Extract gen6 aliasing ppgtt code

Quoted 12 lines · Daniel Vetter
Quoted 8 lines · Ben Widawsky

This refactor clearly identifies the GEN specific portion of the aliased ppgtt code. Aside from some of the gen specific parts being extracted, it also preps us for an upcoming patch that pulls out the size, which has some nice benefits.

v2: Fix wonky error handling (Chris)

Functionally we seem to have this patch already with

Quoted 8 lines

commit e9ccfd7b94f1ca9b46b992a19d2e8c5af70e4c69

Author: Daniel Vetter <[email protected]>
Date:   Thu Jan 24 13:49:56 2013 -0800

drm/i915: extract hw ppgtt setup/cleanup code

Yeah, you're allowed to hate me with the heat of a thousand suns now ;-) -Daniel

I already did. I guess when I rebased the series, you hadn't merged those yet.

Signature · 1 line
Ben Widawsky, Intel Open Source Technology Center
#9 lore raw

PATCH 4/5 drm/i915: Aliased PPGTT size abstraction

In reply to Ben Widawsky

Aside from being a nice prep for some work I'm doing in PPGTT, this also leaves us with a more accurate size in our gtt_total_entries member. Since we can't actually use the PPGTT reserved part of the GGTT.

This will help some of the existing assertions find issues which would overwrite the PPGTT PDEs.

Another benefit is for platforms which don't use the entire 2GB GTT, this will save GTT space. For example if a platform has 1 1GB GTT, this patch should save 1MB of GTT space (512MB has an even greater savings). I don't have a platform to easily test this however.

If/when we ever move to real PPGTT, I think allocating the PDEs as objects will make the most sense.

---
 drivers/gpu/drm/i915/i915_debugfs.c |  2 ++
 drivers/gpu/drm/i915/i915_drv.h     |  3 ++-
 drivers/gpu/drm/i915/i915_gem_gtt.c | 35 ++++++++++++++++++++++++++---------
 3 files changed, 30 insertions(+), 10 deletions(-)
drivers/gpu/drm/i915/i915_debugfs.c +2 -0
index 384f193..4545357 100644
@@ -1605,6 +1605,8 @@ static int i915_ppgtt_info(struct seq_file *m, void *data)
 
 		seq_printf(m, "aliasing PPGTT:\n");
 		seq_printf(m, "pd gtt offset: 0x%08x\n", ppgtt->pd_offset);
+		seq_printf(m, "mapped size: %lldM\n",
+			   ppgtt->mapped_size>>20);
 	}
 	seq_printf(m, "ECOCHK: 0x%08x\n", I915_READ(GAM_ECOCHK));
 	mutex_unlock(&dev->struct_mutex);
drivers/gpu/drm/i915/i915_drv.h +2 -1
index 1af8e73..60ba4ab 100644
@@ -388,10 +388,11 @@ struct i915_gtt {
 	struct page *scratch_page;
 };
 
-#define I915_PPGTT_PD_ENTRIES 512
+#define GEN6_MAX_PD_ENTRIES 512
 #define I915_PPGTT_PT_ENTRIES 1024
 struct i915_hw_ppgtt {
 	struct drm_device *dev;
+	uint64_t mapped_size;
 	unsigned num_pd_entries;
 	struct page **pt_pages;
 	uint32_t pd_offset;
drivers/gpu/drm/i915/i915_gem_gtt.c +26 -9
index f79f427..d9af815 100644
@@ -108,7 +108,7 @@ static void i915_ppgtt_clear_range(struct i915_hw_ppgtt *ppgtt,
 	}
 }
 
-static int gen6_init_aliasing_ppgtt(struct i915_hw_ppgtt *ppgtt)
+static int gen6_init_ppgtt(struct i915_hw_ppgtt *ppgtt, uint64_t size)
 {
 	struct drm_device *dev = ppgtt->dev;
 	struct drm_i915_private *dev_priv = dev->dev_private;
@@ -116,15 +116,22 @@ static int gen6_init_aliasing_ppgtt(struct i915_hw_ppgtt *ppgtt)
 	int i;
 	int ret = -ENOMEM;
 
+	if (dev_priv->mm.aliasing_ppgtt)
+		return -EBUSY;
+
+	/* Each PDE maps 4MB (1k PAGE_SIZE pages)*/
+	ppgtt->mapped_size = round_up(size, 1<<22);
+	ppgtt->num_pd_entries = ppgtt->mapped_size >> 22;
+	BUG_ON(ppgtt->num_pd_entries > GEN6_MAX_PD_ENTRIES);
+
 	/* ppgtt PDEs reside in the global gtt pagetable, which has 512*1024
 	 * entries. For aliasing ppgtt support we just steal them at the end for
 	 * now. */
-	first_pd_entry_in_global_pt = dev_priv->mm.gtt->gtt_total_entries - I915_PPGTT_PD_ENTRIES;
+	first_pd_entry_in_global_pt = dev_priv->mm.gtt->gtt_total_entries -
+		ppgtt->num_pd_entries;
 
-	ppgtt->num_pd_entries = I915_PPGTT_PD_ENTRIES;
 	ppgtt->pt_pages = kzalloc(sizeof(struct page *)*ppgtt->num_pd_entries,
 				  GFP_KERNEL);
-
 	if (!ppgtt->pt_pages)
 		return -ENOMEM;
 
@@ -156,7 +163,10 @@ static int gen6_init_aliasing_ppgtt(struct i915_hw_ppgtt *ppgtt)
 	i915_ppgtt_clear_range(ppgtt, 0,
 			       ppgtt->num_pd_entries*I915_PPGTT_PT_ENTRIES);
 
-	ppgtt->pd_offset = (first_pd_entry_in_global_pt)*sizeof(gtt_pte_t);
+	ppgtt->pd_offset = first_pd_entry_in_global_pt * sizeof(gtt_pte_t);
+
+	DRM_INFO("Allocated %d PDEs for PPGTT\n", ppgtt->num_pd_entries);
+	DRM_DEBUG_DRIVER("First PDE: %d\n", first_pd_entry_in_global_pt);
 
 	return 0;
 
@@ -192,7 +202,7 @@ static int i915_gem_init_aliasing_ppgtt(struct drm_device *dev)
 	ppgtt->dev = dev;
 	ppgtt->scratch_page_dma_addr = dev_priv->gtt.scratch_page_dma;
 
-	ret = gen6_init_aliasing_ppgtt(ppgtt);
+	ret = gen6_init_ppgtt(ppgtt, dev_priv->gtt.total);
 	if (ret)
 		kfree(ppgtt);
 	else
@@ -625,20 +635,27 @@ void i915_gem_init_global_gtt(struct drm_device *dev)
 	mappable_size = dev_priv->gtt.mappable_end;
 
 	if (intel_enable_ppgtt(dev) && HAS_ALIASING_PPGTT(dev)) {
+		/* Each PDE represents 1k PAGE_SIZE objects */
+		const int reserve_pd_space = gtt_size >> 10;
 		int ret;
+
+		BUG_ON(!reserve_pd_space);
+
 		/* PPGTT pdes are stolen from global gtt ptes, so shrink the
 		 * aperture accordingly when using aliasing ppgtt. */
-		gtt_size -= I915_PPGTT_PD_ENTRIES*PAGE_SIZE;
+		gtt_size -= reserve_pd_space;
 
 		i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
 
 		ret = i915_gem_init_aliasing_ppgtt(dev);
-		if (!ret)
+		if (!ret) {
+			dev_priv->mm.gtt->gtt_total_entries = gtt_size >> PAGE_SHIFT;
 			return;
+		}
 
 		DRM_ERROR("Aliased PPGTT setup failed %d\n", ret);
 		drm_mm_takedown(&dev_priv->mm.gtt_space);
-		gtt_size += I915_PPGTT_PD_ENTRIES*PAGE_SIZE;
+		gtt_size += reserve_pd_space;
 	}
 	i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
 }
Signature · 1 line
1.8.1.1
maintainer
#10 lore raw

Re: [PATCH 4/5] drm/i915: Aliased PPGTT size abstraction

Quoted 113 lines · Ben Widawsky

Aside from being a nice prep for some work I'm doing in PPGTT, this also leaves us with a more accurate size in our gtt_total_entries member. Since we can't actually use the PPGTT reserved part of the GGTT.

This will help some of the existing assertions find issues which would overwrite the PPGTT PDEs.

Another benefit is for platforms which don't use the entire 2GB GTT, this will save GTT space. For example if a platform has 1 1GB GTT, this patch should save 1MB of GTT space (512MB has an even greater savings). I don't have a platform to easily test this however.

If/when we ever move to real PPGTT, I think allocating the PDEs as objects will make the most sense.

---
 drivers/gpu/drm/i915/i915_debugfs.c |  2 ++
 drivers/gpu/drm/i915/i915_drv.h     |  3 ++-
 drivers/gpu/drm/i915/i915_gem_gtt.c | 35 ++++++++++++++++++++++++++---------
 3 files changed, 30 insertions(+), 10 deletions(-)
drivers/gpu/drm/i915/i915_debugfs.c +2 -0
index 384f193..4545357 100644
@@ -1605,6 +1605,8 @@ static int i915_ppgtt_info(struct seq_file *m, void *data)
 
 		seq_printf(m, "aliasing PPGTT:\n");
 		seq_printf(m, "pd gtt offset: 0x%08x\n", ppgtt->pd_offset);
+		seq_printf(m, "mapped size: %lldM\n",
+			   ppgtt->mapped_size>>20);
 	}
 	seq_printf(m, "ECOCHK: 0x%08x\n", I915_READ(GAM_ECOCHK));
 	mutex_unlock(&dev->struct_mutex);
drivers/gpu/drm/i915/i915_drv.h +2 -1
index 1af8e73..60ba4ab 100644
@@ -388,10 +388,11 @@ struct i915_gtt {
 	struct page *scratch_page;
 };
 
-#define I915_PPGTT_PD_ENTRIES 512
+#define GEN6_MAX_PD_ENTRIES 512
 #define I915_PPGTT_PT_ENTRIES 1024
 struct i915_hw_ppgtt {
 	struct drm_device *dev;
+	uint64_t mapped_size;
 	unsigned num_pd_entries;
 	struct page **pt_pages;
 	uint32_t pd_offset;
drivers/gpu/drm/i915/i915_gem_gtt.c +16 -6
index f79f427..d9af815 100644
@@ -108,7 +108,7 @@ static void i915_ppgtt_clear_range(struct i915_hw_ppgtt *ppgtt,
 	}
 }
 
-static int gen6_init_aliasing_ppgtt(struct i915_hw_ppgtt *ppgtt)
+static int gen6_init_ppgtt(struct i915_hw_ppgtt *ppgtt, uint64_t size)
 {
 	struct drm_device *dev = ppgtt->dev;
 	struct drm_i915_private *dev_priv = dev->dev_private;
@@ -116,15 +116,22 @@ static int gen6_init_aliasing_ppgtt(struct i915_hw_ppgtt *ppgtt)
 	int i;
 	int ret = -ENOMEM;
 
+	if (dev_priv->mm.aliasing_ppgtt)
+		return -EBUSY;
+
+	/* Each PDE maps 4MB (1k PAGE_SIZE pages)*/
+	ppgtt->mapped_size = round_up(size, 1<<22);
+	ppgtt->num_pd_entries = ppgtt->mapped_size >> 22;
+	BUG_ON(ppgtt->num_pd_entries > GEN6_MAX_PD_ENTRIES);
+
 	/* ppgtt PDEs reside in the global gtt pagetable, which has 512*1024
 	 * entries. For aliasing ppgtt support we just steal them at the end for
 	 * now. */
-	first_pd_entry_in_global_pt = dev_priv->mm.gtt->gtt_total_entries - I915_PPGTT_PD_ENTRIES;
+	first_pd_entry_in_global_pt = dev_priv->mm.gtt->gtt_total_entries -
+		ppgtt->num_pd_entries;
 
-	ppgtt->num_pd_entries = I915_PPGTT_PD_ENTRIES;
 	ppgtt->pt_pages = kzalloc(sizeof(struct page *)*ppgtt->num_pd_entries,
 				  GFP_KERNEL);
-
 	if (!ppgtt->pt_pages)
 		return -ENOMEM;
 
@@ -156,7 +163,10 @@ static int gen6_init_aliasing_ppgtt(struct i915_hw_ppgtt *ppgtt)
 	i915_ppgtt_clear_range(ppgtt, 0,
 			       ppgtt->num_pd_entries*I915_PPGTT_PT_ENTRIES);
 
-	ppgtt->pd_offset = (first_pd_entry_in_global_pt)*sizeof(gtt_pte_t);
+	ppgtt->pd_offset = first_pd_entry_in_global_pt * sizeof(gtt_pte_t);
+
+	DRM_INFO("Allocated %d PDEs for PPGTT\n", ppgtt->num_pd_entries);
+	DRM_DEBUG_DRIVER("First PDE: %d\n", first_pd_entry_in_global_pt);
 
 	return 0;
 
@@ -192,7 +202,7 @@ static int i915_gem_init_aliasing_ppgtt(struct drm_device *dev)
 	ppgtt->dev = dev;
 	ppgtt->scratch_page_dma_addr = dev_priv->gtt.scratch_page_dma;
 
-	ret = gen6_init_aliasing_ppgtt(ppgtt);
+	ret = gen6_init_ppgtt(ppgtt, dev_priv->gtt.total);
 	if (ret)
 		kfree(ppgtt);
 	else

drivers/gpu/drm/i915/i915_gem_gtt.c void i915_gem_init_global_gtt(struct drm_device *dev) · line 639

Quoted 5 lines
drivers/gpu/drm/i915/i915_gem_gtt.c +2 -0
@@ -625,20 +635,27 @@ void i915_gem_init_global_gtt(struct drm_device *dev)
 	mappable_size = dev_priv->gtt.mappable_end;
 
 	if (intel_enable_ppgtt(dev) && HAS_ALIASING_PPGTT(dev)) {
+		/* Each PDE represents 1k PAGE_SIZE objects */
+		const int reserve_pd_space = gtt_size >> 10;

Maybe (likely) I'm being dense, but care to explain this shift a bit? Should it match ppgtt->num_pd_entries * PAGE_SIZE after i915_gem_init_aliasing_ppgtt()?

BR,
Jani.

Quoted 32 lines
drivers/gpu/drm/i915/i915_gem_gtt.c +8 -3
 		int ret;
+
+		BUG_ON(!reserve_pd_space);
+
 		/* PPGTT pdes are stolen from global gtt ptes, so shrink the
 		 * aperture accordingly when using aliasing ppgtt. */
-		gtt_size -= I915_PPGTT_PD_ENTRIES*PAGE_SIZE;
+		gtt_size -= reserve_pd_space;
 
 		i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
 
 		ret = i915_gem_init_aliasing_ppgtt(dev);
-		if (!ret)
+		if (!ret) {
+			dev_priv->mm.gtt->gtt_total_entries = gtt_size >> PAGE_SHIFT;
 			return;
+		}
 
 		DRM_ERROR("Aliased PPGTT setup failed %d\n", ret);
 		drm_mm_takedown(&dev_priv->mm.gtt_space);
-		gtt_size += I915_PPGTT_PD_ENTRIES*PAGE_SIZE;
+		gtt_size += reserve_pd_space;
 	}
 	i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
 }
Signature · 6 lines
1.8.1.1

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

Re: [PATCH 4/5] drm/i915: Aliased PPGTT size abstraction

Quoted 124 lines · Jani Nikula
Quoted 113 lines · Ben Widawsky

Aside from being a nice prep for some work I'm doing in PPGTT, this also leaves us with a more accurate size in our gtt_total_entries member. Since we can't actually use the PPGTT reserved part of the GGTT.

This will help some of the existing assertions find issues which would overwrite the PPGTT PDEs.

Another benefit is for platforms which don't use the entire 2GB GTT, this will save GTT space. For example if a platform has 1 1GB GTT, this patch should save 1MB of GTT space (512MB has an even greater savings). I don't have a platform to easily test this however.

If/when we ever move to real PPGTT, I think allocating the PDEs as objects will make the most sense.

---
 drivers/gpu/drm/i915/i915_debugfs.c |  2 ++
 drivers/gpu/drm/i915/i915_drv.h     |  3 ++-
 drivers/gpu/drm/i915/i915_gem_gtt.c | 35 ++++++++++++++++++++++++++---------
 3 files changed, 30 insertions(+), 10 deletions(-)
drivers/gpu/drm/i915/i915_debugfs.c +2 -0
index 384f193..4545357 100644
@@ -1605,6 +1605,8 @@ static int i915_ppgtt_info(struct seq_file *m, void *data)
 
 		seq_printf(m, "aliasing PPGTT:\n");
 		seq_printf(m, "pd gtt offset: 0x%08x\n", ppgtt->pd_offset);
+		seq_printf(m, "mapped size: %lldM\n",
+			   ppgtt->mapped_size>>20);
 	}
 	seq_printf(m, "ECOCHK: 0x%08x\n", I915_READ(GAM_ECOCHK));
 	mutex_unlock(&dev->struct_mutex);
drivers/gpu/drm/i915/i915_drv.h +2 -1
index 1af8e73..60ba4ab 100644
@@ -388,10 +388,11 @@ struct i915_gtt {
 	struct page *scratch_page;
 };
 
-#define I915_PPGTT_PD_ENTRIES 512
+#define GEN6_MAX_PD_ENTRIES 512
 #define I915_PPGTT_PT_ENTRIES 1024
 struct i915_hw_ppgtt {
 	struct drm_device *dev;
+	uint64_t mapped_size;
 	unsigned num_pd_entries;
 	struct page **pt_pages;
 	uint32_t pd_offset;
drivers/gpu/drm/i915/i915_gem_gtt.c +16 -6
index f79f427..d9af815 100644
@@ -108,7 +108,7 @@ static void i915_ppgtt_clear_range(struct i915_hw_ppgtt *ppgtt,
 	}
 }
 
-static int gen6_init_aliasing_ppgtt(struct i915_hw_ppgtt *ppgtt)
+static int gen6_init_ppgtt(struct i915_hw_ppgtt *ppgtt, uint64_t size)
 {
 	struct drm_device *dev = ppgtt->dev;
 	struct drm_i915_private *dev_priv = dev->dev_private;
@@ -116,15 +116,22 @@ static int gen6_init_aliasing_ppgtt(struct i915_hw_ppgtt *ppgtt)
 	int i;
 	int ret = -ENOMEM;
 
+	if (dev_priv->mm.aliasing_ppgtt)
+		return -EBUSY;
+
+	/* Each PDE maps 4MB (1k PAGE_SIZE pages)*/
+	ppgtt->mapped_size = round_up(size, 1<<22);
+	ppgtt->num_pd_entries = ppgtt->mapped_size >> 22;
+	BUG_ON(ppgtt->num_pd_entries > GEN6_MAX_PD_ENTRIES);
+
 	/* ppgtt PDEs reside in the global gtt pagetable, which has 512*1024
 	 * entries. For aliasing ppgtt support we just steal them at the end for
 	 * now. */
-	first_pd_entry_in_global_pt = dev_priv->mm.gtt->gtt_total_entries - I915_PPGTT_PD_ENTRIES;
+	first_pd_entry_in_global_pt = dev_priv->mm.gtt->gtt_total_entries -
+		ppgtt->num_pd_entries;
 
-	ppgtt->num_pd_entries = I915_PPGTT_PD_ENTRIES;
 	ppgtt->pt_pages = kzalloc(sizeof(struct page *)*ppgtt->num_pd_entries,
 				  GFP_KERNEL);
-
 	if (!ppgtt->pt_pages)
 		return -ENOMEM;
 
@@ -156,7 +163,10 @@ static int gen6_init_aliasing_ppgtt(struct i915_hw_ppgtt *ppgtt)
 	i915_ppgtt_clear_range(ppgtt, 0,
 			       ppgtt->num_pd_entries*I915_PPGTT_PT_ENTRIES);
 
-	ppgtt->pd_offset = (first_pd_entry_in_global_pt)*sizeof(gtt_pte_t);
+	ppgtt->pd_offset = first_pd_entry_in_global_pt * sizeof(gtt_pte_t);
+
+	DRM_INFO("Allocated %d PDEs for PPGTT\n", ppgtt->num_pd_entries);
+	DRM_DEBUG_DRIVER("First PDE: %d\n", first_pd_entry_in_global_pt);
 
 	return 0;
 
@@ -192,7 +202,7 @@ static int i915_gem_init_aliasing_ppgtt(struct drm_device *dev)
 	ppgtt->dev = dev;
 	ppgtt->scratch_page_dma_addr = dev_priv->gtt.scratch_page_dma;
 
-	ret = gen6_init_aliasing_ppgtt(ppgtt);
+	ret = gen6_init_ppgtt(ppgtt, dev_priv->gtt.total);
 	if (ret)
 		kfree(ppgtt);
 	else

drivers/gpu/drm/i915/i915_gem_gtt.c void i915_gem_init_global_gtt(struct drm_device *dev) · line 639

Quoted 5 lines
drivers/gpu/drm/i915/i915_gem_gtt.c +2 -0
@@ -625,20 +635,27 @@ void i915_gem_init_global_gtt(struct drm_device *dev)
 	mappable_size = dev_priv->gtt.mappable_end;
 
 	if (intel_enable_ppgtt(dev) && HAS_ALIASING_PPGTT(dev)) {
+		/* Each PDE represents 1k PAGE_SIZE objects */
+		const int reserve_pd_space = gtt_size >> 10;

Maybe (likely) I'm being dense, but care to explain this shift a bit? Should it match ppgtt->num_pd_entries * PAGE_SIZE after i915_gem_init_aliasing_ppgtt()?

Quoted 3 lines

BR,
Jani.

yeah, it's horribly confusing until you actually need to write code yourself. I actually think Daniel should have reviewed this patch, but I guess he is lazy.

In this context, gtt_size refers to the amount of GPU mappable memory. (I actually had another patch which renamed all of this stuff, but whatever).

We want to be able to map the entire space the gtt maps with the ppgtt.

Each page directory holds 1024 PTEs, which each map a 4k pfn. Therefore, each PDEs is effecitvely mapping (1 << (10 + 12)). Since the PDs are stored in the GGTT (suck), we're trying to carve out the necessary memory from the GGTT. The PDEs are special, see the DCLV stuff.

The number of PDEs required is:
gtt_size >> 22

But as I said above, we really need to calculate how much PD space we need. Each PD has PAGE_SIZE of space required
gtt_size >> 22 >> PAGE_SIZE = gtt_size >> 10

In hindsight, maybe doing that calculation in the code would have been better.

Anyway, to answer the question, I think yes, it should match ppgtt->num_pd_entries * PAGE_SIZE

Quoted 32 lines
Quoted 32 lines
drivers/gpu/drm/i915/i915_gem_gtt.c +8 -3
 		int ret;
+
+		BUG_ON(!reserve_pd_space);
+
 		/* PPGTT pdes are stolen from global gtt ptes, so shrink the
 		 * aperture accordingly when using aliasing ppgtt. */
-		gtt_size -= I915_PPGTT_PD_ENTRIES*PAGE_SIZE;
+		gtt_size -= reserve_pd_space;
 
 		i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
 
 		ret = i915_gem_init_aliasing_ppgtt(dev);
-		if (!ret)
+		if (!ret) {
+			dev_priv->mm.gtt->gtt_total_entries = gtt_size >> PAGE_SHIFT;
 			return;
+		}
 
 		DRM_ERROR("Aliased PPGTT setup failed %d\n", ret);
 		drm_mm_takedown(&dev_priv->mm.gtt_space);
-		gtt_size += I915_PPGTT_PD_ENTRIES*PAGE_SIZE;
+		gtt_size += reserve_pd_space;
 	}
 	i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
 }
Signature · 6 lines
1.8.1.1

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Signature · 1 line
Ben Widawsky, Intel Open Source Technology Center
#12 lore raw

Re: [PATCH 4/5] drm/i915: Aliased PPGTT size abstraction

Quoted 3 lines · Ben Widawsky

But as I said above, we really need to calculate how much PD space we need. Each PD has PAGE_SIZE of space required
gtt_size >> 22 >> PAGE_SIZE = gtt_size >> 10

gtt_size >> 22 * PAGE_SIZE

Quoted 5 lines

In hindsight, maybe doing that calculation in the code would have been better.

Anyway, to answer the question, I think yes, it should match ppgtt->num_pd_entries * PAGE_SIZE

That looks much clearer imo and spells out what happens much better (we still a pte = full page in the address space for each ppgtt pd). Can you please apply this bikeshed to the patch?
-Daniel

Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
#13 lore raw

PATCH v2 4/5 drm/i915: Aliased PPGTT size abstraction

In reply to Ben Widawsky

Aside from being a nice prep for some work I'm doing in PPGTT, this also leaves us with a more accurate size in our gtt_total_entries member. Since we can't actually use the PPGTT reserved part of the GGTT.

This will help some of the existing assertions find issues which would overwrite the PPGTT PDEs.

Another benefit is for platforms which don't use the entire 2GB GTT, this will save GTT space. For example if a platform has 1 1GB GTT, this patch should save 1MB of GTT space (512MB has an even greater savings). I don't have a platform to easily test this however.

If/when we ever move to real PPGTT, I think allocating the PDEs as objects will make the most sense.

v2: Change the size calculation for the PDs carved out of the ggtt to be more explicit. (Daniel)

---
 drivers/gpu/drm/i915/i915_debugfs.c |  2 ++
 drivers/gpu/drm/i915/i915_drv.h     |  3 ++-
 drivers/gpu/drm/i915/i915_gem_gtt.c | 35 ++++++++++++++++++++++++++---------
 3 files changed, 30 insertions(+), 10 deletions(-)
drivers/gpu/drm/i915/i915_debugfs.c +2 -0
index 384f193..4545357 100644
@@ -1605,6 +1605,8 @@ static int i915_ppgtt_info(struct seq_file *m, void *data)
 
 		seq_printf(m, "aliasing PPGTT:\n");
 		seq_printf(m, "pd gtt offset: 0x%08x\n", ppgtt->pd_offset);
+		seq_printf(m, "mapped size: %lldM\n",
+			   ppgtt->mapped_size>>20);
 	}
 	seq_printf(m, "ECOCHK: 0x%08x\n", I915_READ(GAM_ECOCHK));
 	mutex_unlock(&dev->struct_mutex);
drivers/gpu/drm/i915/i915_drv.h +2 -1
index 1af8e73..60ba4ab 100644
@@ -388,10 +388,11 @@ struct i915_gtt {
 	struct page *scratch_page;
 };
 
-#define I915_PPGTT_PD_ENTRIES 512
+#define GEN6_MAX_PD_ENTRIES 512
 #define I915_PPGTT_PT_ENTRIES 1024
 struct i915_hw_ppgtt {
 	struct drm_device *dev;
+	uint64_t mapped_size;
 	unsigned num_pd_entries;
 	struct page **pt_pages;
 	uint32_t pd_offset;
drivers/gpu/drm/i915/i915_gem_gtt.c +26 -9
index f79f427..7d8c2e5 100644
@@ -108,7 +108,7 @@ static void i915_ppgtt_clear_range(struct i915_hw_ppgtt *ppgtt,
 	}
 }
 
-static int gen6_init_aliasing_ppgtt(struct i915_hw_ppgtt *ppgtt)
+static int gen6_init_ppgtt(struct i915_hw_ppgtt *ppgtt, uint64_t size)
 {
 	struct drm_device *dev = ppgtt->dev;
 	struct drm_i915_private *dev_priv = dev->dev_private;
@@ -116,15 +116,22 @@ static int gen6_init_aliasing_ppgtt(struct i915_hw_ppgtt *ppgtt)
 	int i;
 	int ret = -ENOMEM;
 
+	if (dev_priv->mm.aliasing_ppgtt)
+		return -EBUSY;
+
+	/* Each PDE maps 4MB (1k PAGE_SIZE pages)*/
+	ppgtt->mapped_size = round_up(size, 1<<22);
+	ppgtt->num_pd_entries = ppgtt->mapped_size >> 22;
+	BUG_ON(ppgtt->num_pd_entries > GEN6_MAX_PD_ENTRIES);
+
 	/* ppgtt PDEs reside in the global gtt pagetable, which has 512*1024
 	 * entries. For aliasing ppgtt support we just steal them at the end for
 	 * now. */
-	first_pd_entry_in_global_pt = dev_priv->mm.gtt->gtt_total_entries - I915_PPGTT_PD_ENTRIES;
+	first_pd_entry_in_global_pt = dev_priv->mm.gtt->gtt_total_entries -
+		ppgtt->num_pd_entries;
 
-	ppgtt->num_pd_entries = I915_PPGTT_PD_ENTRIES;
 	ppgtt->pt_pages = kzalloc(sizeof(struct page *)*ppgtt->num_pd_entries,
 				  GFP_KERNEL);
-
 	if (!ppgtt->pt_pages)
 		return -ENOMEM;
 
@@ -156,7 +163,10 @@ static int gen6_init_aliasing_ppgtt(struct i915_hw_ppgtt *ppgtt)
 	i915_ppgtt_clear_range(ppgtt, 0,
 			       ppgtt->num_pd_entries*I915_PPGTT_PT_ENTRIES);
 
-	ppgtt->pd_offset = (first_pd_entry_in_global_pt)*sizeof(gtt_pte_t);
+	ppgtt->pd_offset = first_pd_entry_in_global_pt * sizeof(gtt_pte_t);
+
+	DRM_INFO("Allocated %d PDEs for PPGTT\n", ppgtt->num_pd_entries);
+	DRM_DEBUG_DRIVER("First PDE: %d\n", first_pd_entry_in_global_pt);
 
 	return 0;
 
@@ -192,7 +202,7 @@ static int i915_gem_init_aliasing_ppgtt(struct drm_device *dev)
 	ppgtt->dev = dev;
 	ppgtt->scratch_page_dma_addr = dev_priv->gtt.scratch_page_dma;
 
-	ret = gen6_init_aliasing_ppgtt(ppgtt);
+	ret = gen6_init_ppgtt(ppgtt, dev_priv->gtt.total);
 	if (ret)
 		kfree(ppgtt);
 	else
@@ -625,20 +635,27 @@ void i915_gem_init_global_gtt(struct drm_device *dev)
 	mappable_size = dev_priv->gtt.mappable_end;
 
 	if (intel_enable_ppgtt(dev) && HAS_ALIASING_PPGTT(dev)) {
+		/* Each PDE represents 1k PAGE_SIZE objects */
+		const int reserve_pd_space = (gtt_size >> 22) * PAGE_SIZE;
 		int ret;
+
+		BUG_ON(!reserve_pd_space);
+
 		/* PPGTT pdes are stolen from global gtt ptes, so shrink the
 		 * aperture accordingly when using aliasing ppgtt. */
-		gtt_size -= I915_PPGTT_PD_ENTRIES*PAGE_SIZE;
+		gtt_size -= reserve_pd_space;
 
 		i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
 
 		ret = i915_gem_init_aliasing_ppgtt(dev);
-		if (!ret)
+		if (!ret) {
+			dev_priv->mm.gtt->gtt_total_entries = gtt_size >> PAGE_SHIFT;
 			return;
+		}
 
 		DRM_ERROR("Aliased PPGTT setup failed %d\n", ret);
 		drm_mm_takedown(&dev_priv->mm.gtt_space);
-		gtt_size += I915_PPGTT_PD_ENTRIES*PAGE_SIZE;
+		gtt_size += reserve_pd_space;
 	}
 	i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
 }
Signature · 1 line
1.8.1.1
#14 lore raw

Re: [PATCH 4/5 v2] drm/i915: Aliased PPGTT size abstraction

Quoted 15 lines · Ben Widawsky

Aside from being a nice prep for some work I'm doing in PPGTT, this also leaves us with a more accurate size in our gtt_total_entries member. Since we can't actually use the PPGTT reserved part of the GGTT.

This will help some of the existing assertions find issues which would overwrite the PPGTT PDEs.

Another benefit is for platforms which don't use the entire 2GB GTT, this will save GTT space. For example if a platform has 1 1GB GTT, this patch should save 1MB of GTT space (512MB has an even greater savings). I don't have a platform to easily test this however.

If/when we ever move to real PPGTT, I think allocating the PDEs as objects will make the most sense.

Quoted 6 lines

v2: Change the size calculation for the PDs carved out of the ggtt to be more explicit. (Daniel)

I'm not sold yet on whether this is the interface we want. Furthermore I have some fading nightmares that maybe the pdes should be naturally aligned in the GSM. So trying to safe a few of those sounds like a risky idea. I'll punt on this one for now.

For patch 5 I'd prefer the clearer math over adding a comment (comments tend to get stale, fast), otherwise all patches of this series merged (or nothing left after resolving conflicts).

Thanks, Daniel

Quoted 130 lines
---
 drivers/gpu/drm/i915/i915_debugfs.c |  2 ++
 drivers/gpu/drm/i915/i915_drv.h     |  3 ++-
 drivers/gpu/drm/i915/i915_gem_gtt.c | 35 ++++++++++++++++++++++++++---------
 3 files changed, 30 insertions(+), 10 deletions(-)
drivers/gpu/drm/i915/i915_debugfs.c +2 -0
index 384f193..4545357 100644
@@ -1605,6 +1605,8 @@ static int i915_ppgtt_info(struct seq_file *m, void *data)
 
 		seq_printf(m, "aliasing PPGTT:\n");
 		seq_printf(m, "pd gtt offset: 0x%08x\n", ppgtt->pd_offset);
+		seq_printf(m, "mapped size: %lldM\n",
+			   ppgtt->mapped_size>>20);
 	}
 	seq_printf(m, "ECOCHK: 0x%08x\n", I915_READ(GAM_ECOCHK));
 	mutex_unlock(&dev->struct_mutex);
drivers/gpu/drm/i915/i915_drv.h +2 -1
index 1af8e73..60ba4ab 100644
@@ -388,10 +388,11 @@ struct i915_gtt {
 	struct page *scratch_page;
 };
 
-#define I915_PPGTT_PD_ENTRIES 512
+#define GEN6_MAX_PD_ENTRIES 512
 #define I915_PPGTT_PT_ENTRIES 1024
 struct i915_hw_ppgtt {
 	struct drm_device *dev;
+	uint64_t mapped_size;
 	unsigned num_pd_entries;
 	struct page **pt_pages;
 	uint32_t pd_offset;
drivers/gpu/drm/i915/i915_gem_gtt.c +26 -9
index f79f427..7d8c2e5 100644
@@ -108,7 +108,7 @@ static void i915_ppgtt_clear_range(struct i915_hw_ppgtt *ppgtt,
 	}
 }
 
-static int gen6_init_aliasing_ppgtt(struct i915_hw_ppgtt *ppgtt)
+static int gen6_init_ppgtt(struct i915_hw_ppgtt *ppgtt, uint64_t size)
 {
 	struct drm_device *dev = ppgtt->dev;
 	struct drm_i915_private *dev_priv = dev->dev_private;
@@ -116,15 +116,22 @@ static int gen6_init_aliasing_ppgtt(struct i915_hw_ppgtt *ppgtt)
 	int i;
 	int ret = -ENOMEM;
 
+	if (dev_priv->mm.aliasing_ppgtt)
+		return -EBUSY;
+
+	/* Each PDE maps 4MB (1k PAGE_SIZE pages)*/
+	ppgtt->mapped_size = round_up(size, 1<<22);
+	ppgtt->num_pd_entries = ppgtt->mapped_size >> 22;
+	BUG_ON(ppgtt->num_pd_entries > GEN6_MAX_PD_ENTRIES);
+
 	/* ppgtt PDEs reside in the global gtt pagetable, which has 512*1024
 	 * entries. For aliasing ppgtt support we just steal them at the end for
 	 * now. */
-	first_pd_entry_in_global_pt = dev_priv->mm.gtt->gtt_total_entries - I915_PPGTT_PD_ENTRIES;
+	first_pd_entry_in_global_pt = dev_priv->mm.gtt->gtt_total_entries -
+		ppgtt->num_pd_entries;
 
-	ppgtt->num_pd_entries = I915_PPGTT_PD_ENTRIES;
 	ppgtt->pt_pages = kzalloc(sizeof(struct page *)*ppgtt->num_pd_entries,
 				  GFP_KERNEL);
-
 	if (!ppgtt->pt_pages)
 		return -ENOMEM;
 
@@ -156,7 +163,10 @@ static int gen6_init_aliasing_ppgtt(struct i915_hw_ppgtt *ppgtt)
 	i915_ppgtt_clear_range(ppgtt, 0,
 			       ppgtt->num_pd_entries*I915_PPGTT_PT_ENTRIES);
 
-	ppgtt->pd_offset = (first_pd_entry_in_global_pt)*sizeof(gtt_pte_t);
+	ppgtt->pd_offset = first_pd_entry_in_global_pt * sizeof(gtt_pte_t);
+
+	DRM_INFO("Allocated %d PDEs for PPGTT\n", ppgtt->num_pd_entries);
+	DRM_DEBUG_DRIVER("First PDE: %d\n", first_pd_entry_in_global_pt);
 
 	return 0;
 
@@ -192,7 +202,7 @@ static int i915_gem_init_aliasing_ppgtt(struct drm_device *dev)
 	ppgtt->dev = dev;
 	ppgtt->scratch_page_dma_addr = dev_priv->gtt.scratch_page_dma;
 
-	ret = gen6_init_aliasing_ppgtt(ppgtt);
+	ret = gen6_init_ppgtt(ppgtt, dev_priv->gtt.total);
 	if (ret)
 		kfree(ppgtt);
 	else
@@ -625,20 +635,27 @@ void i915_gem_init_global_gtt(struct drm_device *dev)
 	mappable_size = dev_priv->gtt.mappable_end;
 
 	if (intel_enable_ppgtt(dev) && HAS_ALIASING_PPGTT(dev)) {
+		/* Each PDE represents 1k PAGE_SIZE objects */
+		const int reserve_pd_space = (gtt_size >> 22) * PAGE_SIZE;
 		int ret;
+
+		BUG_ON(!reserve_pd_space);
+
 		/* PPGTT pdes are stolen from global gtt ptes, so shrink the
 		 * aperture accordingly when using aliasing ppgtt. */
-		gtt_size -= I915_PPGTT_PD_ENTRIES*PAGE_SIZE;
+		gtt_size -= reserve_pd_space;
 
 		i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
 
 		ret = i915_gem_init_aliasing_ppgtt(dev);
-		if (!ret)
+		if (!ret) {
+			dev_priv->mm.gtt->gtt_total_entries = gtt_size >> PAGE_SHIFT;
 			return;
+		}
 
 		DRM_ERROR("Aliased PPGTT setup failed %d\n", ret);
 		drm_mm_takedown(&dev_priv->mm.gtt_space);
-		gtt_size += I915_PPGTT_PD_ENTRIES*PAGE_SIZE;
+		gtt_size += reserve_pd_space;
 	}
 	i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
 }
Signature · 1 line
1.8.1.1
Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
#15 lore raw

PATCH 5/5 drm/i915: Dynamically calculate dclv

In reply to Ben Widawsky

Directory CacheLine Valid controls which PDEs are held in the directory cache. Each bit represents 16 PDEs (16 PDEs at 4 bytes per entry is 1 cacheline, and therefore, almost makes sense).

Since we can now have an aliasing PPGTT which isn't 2GB, theoretically, we should also modify the DCLV value to fit accordingly.

---
 drivers/gpu/drm/i915/i915_gem_gtt.c | 5 +++--
 drivers/gpu/drm/i915/i915_reg.h     | 1 -
 2 files changed, 3 insertions(+), 3 deletions(-)
drivers/gpu/drm/i915/i915_gem_gtt.c +3 -2
index d9af815..168583b 100644
@@ -344,11 +344,12 @@ void i915_gem_init_ppgtt(struct drm_device *dev)
 	}
 
 	for_each_ring(ring, dev_priv, i) {
+		uint64_t dclv = ppgtt->mapped_size >> 26;
+		dclv = (1ULL << dclv) - 1;
 		if (INTEL_INFO(dev)->gen >= 7)
 			I915_WRITE(RING_MODE_GEN7(ring),
 				   _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
-
-		I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
+		I915_WRITE(RING_PP_DIR_DCLV(ring), dclv);
 		I915_WRITE(RING_PP_DIR_BASE(ring), pd_offset);
 	}
 }
drivers/gpu/drm/i915/i915_reg.h +0 -1
index 94a08b9..780a715 100644
@@ -117,7 +117,6 @@
 #define RING_PP_DIR_BASE(ring)		((ring)->mmio_base+0x228)
 #define RING_PP_DIR_BASE_READ(ring)	((ring)->mmio_base+0x518)
 #define RING_PP_DIR_DCLV(ring)		((ring)->mmio_base+0x220)
-#define   PP_DIR_DCLV_2G		0xffffffff
 
 #define GAM_ECOCHK			0x4090
 #define   ECOCHK_SNB_BIT		(1<<10)
Signature · 1 line
1.8.1.1
maintainer
#16 lore raw

Re: [PATCH 5/5] drm/i915: Dynamically calculate dclv

Quoted 18 lines · Ben Widawsky

Directory CacheLine Valid controls which PDEs are held in the directory cache. Each bit represents 16 PDEs (16 PDEs at 4 bytes per entry is 1 cacheline, and therefore, almost makes sense).

Since we can now have an aliasing PPGTT which isn't 2GB, theoretically, we should also modify the DCLV value to fit accordingly.

---
 drivers/gpu/drm/i915/i915_gem_gtt.c | 5 +++--
 drivers/gpu/drm/i915/i915_reg.h     | 1 -
 2 files changed, 3 insertions(+), 3 deletions(-)
drivers/gpu/drm/i915/i915_gem_gtt.c +0 -0
index d9af815..168583b 100644

drivers/gpu/drm/i915/i915_gem_gtt.c void i915_gem_init_ppgtt(struct drm_device *dev) · line 348

Quoted 5 lines
drivers/gpu/drm/i915/i915_gem_gtt.c +2 -0
@@ -344,11 +344,12 @@ void i915_gem_init_ppgtt(struct drm_device *dev)
 	}
 
 	for_each_ring(ring, dev_priv, i) {
+		uint64_t dclv = ppgtt->mapped_size >> 26;
+		dclv = (1ULL << dclv) - 1;

Maybe add a comment with some of the info from the commit message here?

BR,
Jani.

Quoted 28 lines
drivers/gpu/drm/i915/i915_gem_gtt.c +1 -2
 		if (INTEL_INFO(dev)->gen >= 7)
 			I915_WRITE(RING_MODE_GEN7(ring),
 				   _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
-
-		I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
+		I915_WRITE(RING_PP_DIR_DCLV(ring), dclv);
 		I915_WRITE(RING_PP_DIR_BASE(ring), pd_offset);
 	}
 }
drivers/gpu/drm/i915/i915_reg.h +0 -1
index 94a08b9..780a715 100644
@@ -117,7 +117,6 @@
 #define RING_PP_DIR_BASE(ring)		((ring)->mmio_base+0x228)
 #define RING_PP_DIR_BASE_READ(ring)	((ring)->mmio_base+0x518)
 #define RING_PP_DIR_DCLV(ring)		((ring)->mmio_base+0x220)
-#define   PP_DIR_DCLV_2G		0xffffffff
 
 #define GAM_ECOCHK			0x4090
 #define   ECOCHK_SNB_BIT		(1<<10)
Signature · 6 lines
1.8.1.1

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

PATCH 5/5 drm/i915: Dynamically calculate dclv

In reply to Ben Widawsky

Directory CacheLine Valid controls which PDEs are held in the directory cache. Each bit represents 16 PDEs (16 PDEs at 4 bytes per entry is 1 cacheline, and therefore, almost makes sense).

Since we can now have an aliasing PPGTT which isn't 2GB, theoretically, we should also modify the DCLV value to fit accordingly.

v2: Add part of this commit message to the code as a comment (Jani)

---
 drivers/gpu/drm/i915/i915_gem_gtt.c | 7 +++++--
 drivers/gpu/drm/i915/i915_reg.h     | 1 -
 2 files changed, 5 insertions(+), 3 deletions(-)
drivers/gpu/drm/i915/i915_gem_gtt.c +5 -2
index 7d8c2e5..2fc866a 100644
@@ -344,11 +344,14 @@ void i915_gem_init_ppgtt(struct drm_device *dev)
 	}
 
 	for_each_ring(ring, dev_priv, i) {
+		/* Directory CacheLine Valid controls which PDEs are held in the
+		 * directory cache. Each bit represents 16 PDEs */
+		uint64_t dclv = ppgtt->mapped_size >> 26;
+		dclv = (1ULL << dclv) - 1;
 		if (INTEL_INFO(dev)->gen >= 7)
 			I915_WRITE(RING_MODE_GEN7(ring),
 				   _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
-
-		I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
+		I915_WRITE(RING_PP_DIR_DCLV(ring), dclv);
 		I915_WRITE(RING_PP_DIR_BASE(ring), pd_offset);
 	}
 }
drivers/gpu/drm/i915/i915_reg.h +0 -1
index 0c89cf5..c5ac375 100644
@@ -117,7 +117,6 @@
 #define RING_PP_DIR_BASE(ring)		((ring)->mmio_base+0x228)
 #define RING_PP_DIR_BASE_READ(ring)	((ring)->mmio_base+0x518)
 #define RING_PP_DIR_DCLV(ring)		((ring)->mmio_base+0x220)
-#define   PP_DIR_DCLV_2G		0xffffffff
 
 #define GAM_ECOCHK			0x4090
 #define   ECOCHK_SNB_BIT		(1<<10)
Signature · 1 line
1.8.1.1
#18 lore raw

Re: [PATCH 5/5] drm/i915: Dynamically calculate dclv

In reply to Ben Widawsky
Quoted 18 lines · Ben Widawsky

Directory CacheLine Valid controls which PDEs are held in the directory cache. Each bit represents 16 PDEs (16 PDEs at 4 bytes per entry is 1 cacheline, and therefore, almost makes sense).

Since we can now have an aliasing PPGTT which isn't 2GB, theoretically, we should also modify the DCLV value to fit accordingly.

---
 drivers/gpu/drm/i915/i915_gem_gtt.c | 5 +++--
 drivers/gpu/drm/i915/i915_reg.h     | 1 -
 2 files changed, 3 insertions(+), 3 deletions(-)
drivers/gpu/drm/i915/i915_gem_gtt.c +0 -0
index d9af815..168583b 100644

drivers/gpu/drm/i915/i915_gem_gtt.c void i915_gem_init_ppgtt(struct drm_device *dev) · line 347

Quoted 4 lines
drivers/gpu/drm/i915/i915_gem_gtt.c +1 -0
@@ -344,11 +344,12 @@ void i915_gem_init_ppgtt(struct drm_device *dev)
 	}
 
 	for_each_ring(ring, dev_priv, i) {
+		uint64_t dclv = ppgtt->mapped_size >> 26;

ppgtt->num_pd_entries / 16? Would be much more self-documenting than the almost magic >> 26 ...
-Daniel

Quoted 29 lines
drivers/gpu/drm/i915/i915_gem_gtt.c +2 -2
+		dclv = (1ULL << dclv) - 1;
 		if (INTEL_INFO(dev)->gen >= 7)
 			I915_WRITE(RING_MODE_GEN7(ring),
 				   _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
-
-		I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
+		I915_WRITE(RING_PP_DIR_DCLV(ring), dclv);
 		I915_WRITE(RING_PP_DIR_BASE(ring), pd_offset);
 	}
 }
drivers/gpu/drm/i915/i915_reg.h +0 -1
index 94a08b9..780a715 100644
@@ -117,7 +117,6 @@
 #define RING_PP_DIR_BASE(ring)		((ring)->mmio_base+0x228)
 #define RING_PP_DIR_BASE_READ(ring)	((ring)->mmio_base+0x518)
 #define RING_PP_DIR_DCLV(ring)		((ring)->mmio_base+0x220)
-#define   PP_DIR_DCLV_2G		0xffffffff
 
 #define GAM_ECOCHK			0x4090
 #define   ECOCHK_SNB_BIT		(1<<10)
Signature · 6 lines
1.8.1.1

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
maintainer Reviewed-by on the v1 series
#19 lore raw
In reply to Ben Widawsky
Quoted 9 lines · Ben Widawsky

This work sort of led me on the gtt vtable distraction which ended in Daniel doing his own version. I posted these originally about a month ago. I've rebased them on dinq. I'd like to see them get merged since they still help with some other work I'm doing on PPGTT.

Some of this work does conflict a bit with the gtt vtable stuff.

I did a decent amount of testing on the original post, less so on this rebased version.

On the series, with my limited exposure to (pp)gtt,

Quoted 21 lines

Ben Widawsky (5):
drm/i915: trivial: kill-agp collateral cleanups
drm/i915: Reclaim GTT space for failed PPGTT
drm/i915: Extract gen6 aliasing ppgtt code
drm/i915: Aliased PPGTT size abstraction
drm/i915: Dynamically calculate dclv

 drivers/gpu/drm/i915/i915_debugfs.c |   2 +
 drivers/gpu/drm/i915/i915_drv.h     |   6 +--
 drivers/gpu/drm/i915/i915_gem_gtt.c | 101 +++++++++++++++++++++++-------------
 drivers/gpu/drm/i915/i915_reg.h     |   1 -
 4 files changed, 71 insertions(+), 39 deletions(-)
Signature · 6 lines
1.8.1.1

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