Skip to contentSkip to conversation

drm/i915: Per-process stats work better when evaluated per-process

PATCH In review Rodrigo Vivi 5 participants 9 messages Mar 13, 2014 – Mar 20, 2014

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

PATCH 3/4 drm/i915: Per-process stats work better when evaluated per-process

From: Chris Wilson <[email protected]>

The idea of printing objects used by each process is to judge how each process is using them. This means that we need to evaluate whether the object is bound for that particular process, rather than just whether it is bound into the global GTT.

---
 drivers/gpu/drm/i915/i915_debugfs.c     | 34 ++++++++++++++++++++++++++-------
 drivers/gpu/drm/i915/i915_drv.h         |  2 ++
 drivers/gpu/drm/i915/i915_gem_context.c |  1 +
 3 files changed, 30 insertions(+), 7 deletions(-)
drivers/gpu/drm/i915/i915_debugfs.c +27 -7
index a90d31c..ed3965f 100644
@@ -299,28 +299,46 @@ static int i915_gem_stolen_list_info(struct seq_file *m, void *data)
 } while (0)
 
 struct file_stats {
+	struct drm_i915_file_private *file_priv;
 	int count;
-	size_t total, active, inactive, unbound;
+	size_t total, global, active, inactive, unbound;
 };
 
 static int per_file_stats(int id, void *ptr, void *data)
 {
 	struct drm_i915_gem_object *obj = ptr;
 	struct file_stats *stats = data;
+	struct i915_vma *vma;
 
 	stats->count++;
 	stats->total += obj->base.size;
 
-	if (i915_gem_obj_ggtt_bound(obj)) {
-		if (!list_empty(&obj->ring_list))
+	list_for_each_entry(vma, &obj->vma_list, vma_link) {
+		struct i915_hw_ppgtt *ppgtt;
+
+		if (!drm_mm_node_allocated(&vma->node))
+			continue;
+
+		ppgtt = container_of(vma->vm, struct i915_hw_ppgtt, base);
+		if (ppgtt->ctx == NULL) {
+			stats->global += obj->base.size;
+			continue;
+		}
+
+		if (ppgtt->ctx->file_priv != stats->file_priv)
+			continue;
+
+		if (obj->ring) /* XXX per-vma statistic */
 			stats->active += obj->base.size;
 		else
 			stats->inactive += obj->base.size;
-	} else {
-		if (!list_empty(&obj->global_list))
-			stats->unbound += obj->base.size;
+
+		return 0;
 	}
 
+	if (!list_empty(&obj->global_list))
+		stats->unbound += obj->base.size;
+
 	return 0;
 }
 
@@ -411,6 +429,7 @@ static int i915_gem_object_info(struct seq_file *m, void* data)
 		struct task_struct *task;
 
 		memset(&stats, 0, sizeof(stats));
+		stats.file_priv = file->driver_priv;
 		idr_for_each(&file->object_idr, per_file_stats, &stats);
 		/*
 		 * Although we have a valid reference on file->pid, that does
@@ -420,12 +439,13 @@ static int i915_gem_object_info(struct seq_file *m, void* data)
 		 */
 		rcu_read_lock();
 		task = pid_task(file->pid, PIDTYPE_PID);
-		seq_printf(m, "%s: %u objects, %zu bytes (%zu active, %zu inactive, %zu unbound)\n",
+		seq_printf(m, "%s: %u objects, %zu bytes (%zu active, %zu inactive, %zu global, %zu unbound)\n",
 			   task ? task->comm : "<unknown>",
 			   stats.count,
 			   stats.total,
 			   stats.active,
 			   stats.inactive,
+			   stats.global,
 			   stats.unbound);
 		rcu_read_unlock();
 	}
drivers/gpu/drm/i915/i915_drv.h +2 -0
index 2a319ba..b76c6de 100644
@@ -721,6 +721,8 @@ struct i915_hw_ppgtt {
 		dma_addr_t *gen8_pt_dma_addr[4];
 	};
 
+	struct i915_hw_context *ctx;
+
 	int (*enable)(struct i915_hw_ppgtt *ppgtt);
 	int (*switch_mm)(struct i915_hw_ppgtt *ppgtt,
 			 struct intel_ring_buffer *ring,
drivers/gpu/drm/i915/i915_gem_context.c +1 -0
index ce41cff..1a94b07 100644
@@ -215,6 +215,7 @@ create_vm_for_ctx(struct drm_device *dev, struct i915_hw_context *ctx)
 		return ERR_PTR(ret);
 	}
 
+	ppgtt->ctx = ctx;
 	return ppgtt;
 }
 
Signature · 1 line
1.8.5.3
#1 lore raw
Quoted 42 lines · Rodrigo Vivi

From: Chris Wilson <[email protected]>

The idea of printing objects used by each process is to judge how each process is using them. This means that we need to evaluate whether the object is bound for that particular process, rather than just whether it is bound into the global GTT.

---
 drivers/gpu/drm/i915/i915_debugfs.c     | 34 ++++++++++++++++++++++++++-------
 drivers/gpu/drm/i915/i915_drv.h         |  2 ++
 drivers/gpu/drm/i915/i915_gem_context.c |  1 +
 3 files changed, 30 insertions(+), 7 deletions(-)
drivers/gpu/drm/i915/i915_debugfs.c +4 -3
index a90d31c..ed3965f 100644
@@ -299,28 +299,46 @@ static int i915_gem_stolen_list_info(struct seq_file *m, void *data)
 } while (0)
 
 struct file_stats {
+	struct drm_i915_file_private *file_priv;
 	int count;
-	size_t total, active, inactive, unbound;
+	size_t total, global, active, inactive, unbound;
 };
 
 static int per_file_stats(int id, void *ptr, void *data)
 {
 	struct drm_i915_gem_object *obj = ptr;
 	struct file_stats *stats = data;
+	struct i915_vma *vma;
 
 	stats->count++;
 	stats->total += obj->base.size;
 
-	if (i915_gem_obj_ggtt_bound(obj)) {
-		if (!list_empty(&obj->ring_list))
+	list_for_each_entry(vma, &obj->vma_list, vma_link) {

drivers/gpu/drm/i915/i915_debugfs.c static int i915_gem_stolen_list_info(struct seq_file *m, void *data) · line 326

Quoted 10 lines
drivers/gpu/drm/i915/i915_debugfs.c +10 -0
+		struct i915_hw_ppgtt *ppgtt;
+
+		if (!drm_mm_node_allocated(&vma->node))
+			continue;
+
+		ppgtt = container_of(vma->vm, struct i915_hw_ppgtt, base);
+		if (ppgtt->ctx == NULL) {
+			stats->global += obj->base.size;
+			continue;
+		}

I'm not really clear how this is supposed to work for global. Can you make me happy and change it to:

if (i915_is_ggtt(vma->vm))

drivers/gpu/drm/i915/i915_debugfs.c static int i915_gem_stolen_list_info(struct seq_file *m, void *data) · line 332

Quoted 6 lines
drivers/gpu/drm/i915/i915_debugfs.c +5 -0
+
+		if (ppgtt->ctx->file_priv != stats->file_priv)
+			continue;
+
+		if (obj->ring) /* XXX per-vma statistic */
 			stats->active += obj->base.size;

Doesn't active get counted too many times if multiple VMAs exist for the same active object (not a new problem to this patch)?

Quoted 66 lines
drivers/gpu/drm/i915/i915_debugfs.c +8 -4
 		else
 			stats->inactive += obj->base.size;
-	} else {
-		if (!list_empty(&obj->global_list))
-			stats->unbound += obj->base.size;
+
+		return 0;
 	}
 
+	if (!list_empty(&obj->global_list))
+		stats->unbound += obj->base.size;
+
 	return 0;
 }
 
@@ -411,6 +429,7 @@ static int i915_gem_object_info(struct seq_file *m, void* data)
 		struct task_struct *task;
 
 		memset(&stats, 0, sizeof(stats));
+		stats.file_priv = file->driver_priv;
 		idr_for_each(&file->object_idr, per_file_stats, &stats);
 		/*
 		 * Although we have a valid reference on file->pid, that does
@@ -420,12 +439,13 @@ static int i915_gem_object_info(struct seq_file *m, void* data)
 		 */
 		rcu_read_lock();
 		task = pid_task(file->pid, PIDTYPE_PID);
-		seq_printf(m, "%s: %u objects, %zu bytes (%zu active, %zu inactive, %zu unbound)\n",
+		seq_printf(m, "%s: %u objects, %zu bytes (%zu active, %zu inactive, %zu global, %zu unbound)\n",
 			   task ? task->comm : "<unknown>",
 			   stats.count,
 			   stats.total,
 			   stats.active,
 			   stats.inactive,
+			   stats.global,
 			   stats.unbound);
 		rcu_read_unlock();
 	}
drivers/gpu/drm/i915/i915_drv.h +2 -0
index 2a319ba..b76c6de 100644
@@ -721,6 +721,8 @@ struct i915_hw_ppgtt {
 		dma_addr_t *gen8_pt_dma_addr[4];
 	};
 
+	struct i915_hw_context *ctx;
+
 	int (*enable)(struct i915_hw_ppgtt *ppgtt);
 	int (*switch_mm)(struct i915_hw_ppgtt *ppgtt,
 			 struct intel_ring_buffer *ring,
drivers/gpu/drm/i915/i915_gem_context.c +1 -0
index ce41cff..1a94b07 100644
@@ -215,6 +215,7 @@ create_vm_for_ctx(struct drm_device *dev, struct i915_hw_context *ctx)
 		return ERR_PTR(ret);
 	}
 
+	ppgtt->ctx = ctx;
 	return ppgtt;
 }
 
Signature · 1 line
1.8.5.3
Signature · 1 line
Ben Widawsky, Intel Open Source Technology Center
#2 lore raw
Quoted 57 lines · Ben Widawsky
Quoted 42 lines · Rodrigo Vivi

From: Chris Wilson <[email protected]>

The idea of printing objects used by each process is to judge how each process is using them. This means that we need to evaluate whether the object is bound for that particular process, rather than just whether it is bound into the global GTT.

---
 drivers/gpu/drm/i915/i915_debugfs.c     | 34 ++++++++++++++++++++++++++-------
 drivers/gpu/drm/i915/i915_drv.h         |  2 ++
 drivers/gpu/drm/i915/i915_gem_context.c |  1 +
 3 files changed, 30 insertions(+), 7 deletions(-)
drivers/gpu/drm/i915/i915_debugfs.c +4 -3
index a90d31c..ed3965f 100644
@@ -299,28 +299,46 @@ static int i915_gem_stolen_list_info(struct seq_file *m, void *data)
 } while (0)
 
 struct file_stats {
+	struct drm_i915_file_private *file_priv;
 	int count;
-	size_t total, active, inactive, unbound;
+	size_t total, global, active, inactive, unbound;
 };
 
 static int per_file_stats(int id, void *ptr, void *data)
 {
 	struct drm_i915_gem_object *obj = ptr;
 	struct file_stats *stats = data;
+	struct i915_vma *vma;
 
 	stats->count++;
 	stats->total += obj->base.size;
 
-	if (i915_gem_obj_ggtt_bound(obj)) {
-		if (!list_empty(&obj->ring_list))
+	list_for_each_entry(vma, &obj->vma_list, vma_link) {

drivers/gpu/drm/i915/i915_debugfs.c static int i915_gem_stolen_list_info(struct seq_file *m, void *data) · line 326

Quoted 10 lines
drivers/gpu/drm/i915/i915_debugfs.c +10 -0
+		struct i915_hw_ppgtt *ppgtt;
+
+		if (!drm_mm_node_allocated(&vma->node))
+			continue;
+
+		ppgtt = container_of(vma->vm, struct i915_hw_ppgtt, base);
+		if (ppgtt->ctx == NULL) {
+			stats->global += obj->base.size;
+			continue;
+		}

I'm not really clear how this is supposed to work for global. Can you make me happy and change it to:

Quoted 1 line

if (i915_is_ggtt(vma->vm))

That better reflects ggtt. How about
if (i915_is_ggtt(vma->vm)) { global++; continue; } if (ppgtt->ctx == NULL) { default++; continue; }

Quoted 9 lines

drivers/gpu/drm/i915/i915_debugfs.c static int i915_gem_stolen_list_info(struct seq_file *m, void *data) · line 332

Quoted 6 lines
drivers/gpu/drm/i915/i915_debugfs.c +5 -0
+
+		if (ppgtt->ctx->file_priv != stats->file_priv)
+			continue;
+
+		if (obj->ring) /* XXX per-vma statistic */
 			stats->active += obj->base.size;

Doesn't active get counted too many times if multiple VMAs exist for the same active object (not a new problem to this patch)?

Yes. It's a bridge I'm willing to cross in the future! -Chris

Signature · 1 line
Chris Wilson, Intel Open Source Technology Centre
#3 lore raw

PATCH 1/2 drm/i915: Per-process stats work better when evaluated per-process

The idea of printing objects used by each process is to judge how each process is using them. This means that we need to evaluate whether the object is bound for that particular process, rather than just whether it is bound into the global GTT.

v2: Restore the non-full-ppgtt path for simplicity as we may not even
create vma with older hardware.

v3: Tweak handling of global entries and default context entries.

---
 drivers/gpu/drm/i915/i915_debugfs.c     | 49 +++++++++++++++++++++++++++------
 drivers/gpu/drm/i915/i915_drv.h         |  2 ++
 drivers/gpu/drm/i915/i915_gem_context.c |  1 +
 3 files changed, 43 insertions(+), 9 deletions(-)
drivers/gpu/drm/i915/i915_debugfs.c +40 -9
index 3e0d1906ef8d..4e1787ee8f37 100644
@@ -299,28 +299,57 @@ static int i915_gem_stolen_list_info(struct seq_file *m, void *data)
 } while (0)
 
 struct file_stats {
+	struct drm_i915_file_private *file_priv;
 	int count;
-	size_t total, active, inactive, unbound;
+	size_t total, global, active, inactive, unbound;
 };
 
 static int per_file_stats(int id, void *ptr, void *data)
 {
 	struct drm_i915_gem_object *obj = ptr;
 	struct file_stats *stats = data;
+	struct i915_vma *vma;
 
 	stats->count++;
 	stats->total += obj->base.size;
 
-	if (i915_gem_obj_ggtt_bound(obj)) {
-		if (!list_empty(&obj->ring_list))
-			stats->active += obj->base.size;
-		else
-			stats->inactive += obj->base.size;
+	if (USES_FULL_PPGTT(obj->base.dev)) {
+		list_for_each_entry(vma, &obj->vma_list, vma_link) {
+			struct i915_hw_ppgtt *ppgtt;
+
+			if (!drm_mm_node_allocated(&vma->node))
+				continue;
+
+			if (i915_is_ggtt(vma->vm)) {
+				stats->global += obj->base.size;
+				continue;
+			}
+
+			ppgtt = container_of(vma->vm, struct i915_hw_ppgtt, base);
+			if (ppgtt->ctx && ppgtt->ctx->file_priv != stats->file_priv)
+				continue;
+
+			if (obj->ring) /* XXX per-vma statistic */
+				stats->active += obj->base.size;
+			else
+				stats->inactive += obj->base.size;
+
+			return 0;
+		}
 	} else {
-		if (!list_empty(&obj->global_list))
-			stats->unbound += obj->base.size;
+		if (i915_gem_obj_ggtt_bound(obj)) {
+			stats->global += obj->base.size;
+			if (obj->ring)
+				stats->active += obj->base.size;
+			else
+				stats->inactive += obj->base.size;
+			return 0;
+		}
 	}
 
+	if (!list_empty(&obj->global_list))
+		stats->unbound += obj->base.size;
+
 	return 0;
 }
 
@@ -411,6 +440,7 @@ static int i915_gem_object_info(struct seq_file *m, void* data)
 		struct task_struct *task;
 
 		memset(&stats, 0, sizeof(stats));
+		stats.file_priv = file->driver_priv;
 		idr_for_each(&file->object_idr, per_file_stats, &stats);
 		/*
 		 * Although we have a valid reference on file->pid, that does
@@ -420,12 +450,13 @@ static int i915_gem_object_info(struct seq_file *m, void* data)
 		 */
 		rcu_read_lock();
 		task = pid_task(file->pid, PIDTYPE_PID);
-		seq_printf(m, "%s: %u objects, %zu bytes (%zu active, %zu inactive, %zu unbound)\n",
+		seq_printf(m, "%s: %u objects, %zu bytes (%zu active, %zu inactive, %zu global, %zu unbound)\n",
 			   task ? task->comm : "<unknown>",
 			   stats.count,
 			   stats.total,
 			   stats.active,
 			   stats.inactive,
+			   stats.global,
 			   stats.unbound);
 		rcu_read_unlock();
 	}
drivers/gpu/drm/i915/i915_drv.h +2 -0
index c4ab6d6ab926..950c242e49d5 100644
@@ -730,6 +730,8 @@ struct i915_hw_ppgtt {
 		dma_addr_t *gen8_pt_dma_addr[4];
 	};
 
+	struct i915_hw_context *ctx;
+
 	int (*enable)(struct i915_hw_ppgtt *ppgtt);
 	int (*switch_mm)(struct i915_hw_ppgtt *ppgtt,
 			 struct intel_ring_buffer *ring,
drivers/gpu/drm/i915/i915_gem_context.c +1 -0
index 40f430612324..b57b48cc5087 100644
@@ -218,6 +218,7 @@ create_vm_for_ctx(struct drm_device *dev, struct i915_hw_context *ctx)
 		return ERR_PTR(ret);
 	}
 
+	ppgtt->ctx = ctx;
 	return ppgtt;
 }
 
Signature · 1 line
1.9.1
#4 lore raw

PATCH 2/2 drm/i915: Print how many objects are shared in per-process stats

---
 drivers/gpu/drm/i915/i915_debugfs.c | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/i915_debugfs.c +8 -2
index 4e1787ee8f37..9cc1c9360238 100644
@@ -301,7 +301,9 @@ static int i915_gem_stolen_list_info(struct seq_file *m, void *data)
 struct file_stats {
 	struct drm_i915_file_private *file_priv;
 	int count;
-	size_t total, global, active, inactive, unbound;
+	size_t total, unbound;
+	size_t global, shared;
+	size_t active, inactive;
 };
 
 static int per_file_stats(int id, void *ptr, void *data)
@@ -313,6 +315,9 @@ static int per_file_stats(int id, void *ptr, void *data)
 	stats->count++;
 	stats->total += obj->base.size;
 
+	if (obj->base.name || obj->base.dma_buf)
+		stats->shared += obj->base.size;
+
 	if (USES_FULL_PPGTT(obj->base.dev)) {
 		list_for_each_entry(vma, &obj->vma_list, vma_link) {
 			struct i915_hw_ppgtt *ppgtt;
@@ -450,13 +455,14 @@ static int i915_gem_object_info(struct seq_file *m, void* data)
 		 */
 		rcu_read_lock();
 		task = pid_task(file->pid, PIDTYPE_PID);
-		seq_printf(m, "%s: %u objects, %zu bytes (%zu active, %zu inactive, %zu global, %zu unbound)\n",
+		seq_printf(m, "%s: %u objects, %zu bytes (%zu active, %zu inactive, %zu global, %zu shared, %zu unbound)\n",
 			   task ? task->comm : "<unknown>",
 			   stats.count,
 			   stats.total,
 			   stats.active,
 			   stats.inactive,
 			   stats.global,
+			   stats.shared,
 			   stats.unbound);
 		rcu_read_unlock();
 	}
Signature · 1 line
1.9.1
Reviewed-by on v1 2/2
#5 lore raw

Re: [PATCH 2/2] drm/i915: Print how many objects are shared in per-process statsg

Quoted 1 line · Chris Wilson

Any clue how you intend to use this for a commit message (I'm actually curious)? Also, the subject is wrong, you're counting size, not quantity. Anyhoo, looks correct.

Quoted 52 lines
---
 drivers/gpu/drm/i915/i915_debugfs.c | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/i915_debugfs.c +8 -2
index 4e1787ee8f37..9cc1c9360238 100644
@@ -301,7 +301,9 @@ static int i915_gem_stolen_list_info(struct seq_file *m, void *data)
 struct file_stats {
 	struct drm_i915_file_private *file_priv;
 	int count;
-	size_t total, global, active, inactive, unbound;
+	size_t total, unbound;
+	size_t global, shared;
+	size_t active, inactive;
 };
 
 static int per_file_stats(int id, void *ptr, void *data)
@@ -313,6 +315,9 @@ static int per_file_stats(int id, void *ptr, void *data)
 	stats->count++;
 	stats->total += obj->base.size;
 
+	if (obj->base.name || obj->base.dma_buf)
+		stats->shared += obj->base.size;
+
 	if (USES_FULL_PPGTT(obj->base.dev)) {
 		list_for_each_entry(vma, &obj->vma_list, vma_link) {
 			struct i915_hw_ppgtt *ppgtt;
@@ -450,13 +455,14 @@ static int i915_gem_object_info(struct seq_file *m, void* data)
 		 */
 		rcu_read_lock();
 		task = pid_task(file->pid, PIDTYPE_PID);
-		seq_printf(m, "%s: %u objects, %zu bytes (%zu active, %zu inactive, %zu global, %zu unbound)\n",
+		seq_printf(m, "%s: %u objects, %zu bytes (%zu active, %zu inactive, %zu global, %zu shared, %zu unbound)\n",
 			   task ? task->comm : "<unknown>",
 			   stats.count,
 			   stats.total,
 			   stats.active,
 			   stats.inactive,
 			   stats.global,
+			   stats.shared,
 			   stats.unbound);
 		rcu_read_unlock();
 	}
Signature · 6 lines
1.9.1

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

Re: [PATCH 2/2] drm/i915: Print how many objects are shared in per-process statsg

Quoted 6 lines · Ben Widawsky
Quoted 1 line · Chris Wilson

Any clue how you intend to use this for a commit message (I'm actually curious)? Also, the subject is wrong, you're counting size, not quantity. Anyhoo, looks correct.

Knowing how much of the allocated objects are shared between processes helps gauge whether a process is leaking private objects, or if it simply a display server suffering memory pressure from lots of clients.

Quoted 1 line
Signature · 1 line
Chris Wilson, Intel Open Source Technology Centre
#7 lore raw

Re: [PATCH 2/2] drm/i915: Print how many objects are shared in per-process statsg

Quoted 11 lines · Chris Wilson
Quoted 6 lines · Ben Widawsky
Quoted 1 line · Chris Wilson

Any clue how you intend to use this for a commit message (I'm actually curious)? Also, the subject is wrong, you're counting size, not quantity. Anyhoo, looks correct.

Knowing how much of the allocated objects are shared between processes helps gauge whether a process is leaking private objects, or if it simply a display server suffering memory pressure from lots of clients.

I've added something along these lines here.

Quoted 2 lines
Quoted 2 lines

Both patches merged, thanks.
-Daniel

Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
Reviewed-by on v1 1/2
#8 lore raw
In reply to Chris Wilson
Quoted 5 lines · Chris Wilson

The idea of printing objects used by each process is to judge how each process is using them. This means that we need to evaluate whether the object is bound for that particular process, rather than just whether it is bound into the global GTT.

Quoted 7 lines

v2: Restore the non-full-ppgtt path for simplicity as we may not even
create vma with older hardware.

v3: Tweak handling of global entries and default context entries.

[snip]

Signature · 1 line
Ben Widawsky, Intel Open Source Technology Center