[01/41] drm/i915: Move user fault tracking to a separate list

Submitted by Chris Wilson on Oct. 14, 2016, 12:17 p.m.

Details

Message ID 20161014121833.439-2-chris@chris-wilson.co.uk
State New
Headers show
Series "Series without cover letter" ( rev: 1 ) in Intel GFX

Browsing this patch as part of:
"Series without cover letter" rev 1 in Intel GFX
<< prev patch [1/41] next patch >>

Commit Message

Chris Wilson Oct. 14, 2016, 12:17 p.m.
We want to decouple RPM and struct_mutex, but currently RPM has to walk
the list of bound objects and remove userspace mmapping before we
suspend (otherwise userspace may continue to access the GTT whilst it is
powered down). This currently requires the struct_mutex to walk the
bound_list, but if we move that to a separate list and lock we can take
the first step towards removing the struct_mutex.

v2: Split runtime suspend unmapping vs regular unmapping, to make the
locking (and barriers) clearer. Add the object to the userfault_list
prior to inserting the first PTE, the race between add/revoke depends
upon struct_mutex for regular unmappings and rpm for runtime-suspend.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> #v1
---
 drivers/gpu/drm/i915/i915_debugfs.c   |  2 +-
 drivers/gpu/drm/i915/i915_drv.h       | 20 +++++++++++------
 drivers/gpu/drm/i915/i915_gem.c       | 42 +++++++++++++++++++++++++++--------
 drivers/gpu/drm/i915/i915_gem_evict.c |  2 +-
 drivers/gpu/drm/i915/i915_gem_fence.c |  2 +-
 5 files changed, 49 insertions(+), 19 deletions(-)

Patch hide | download patch | download mbox

diff --git a/drivers/gpu/drm/i915/i915_debugfs.c b/drivers/gpu/drm/i915/i915_debugfs.c
index 515c206ba653..c1e3bd785b71 100644
--- a/drivers/gpu/drm/i915/i915_debugfs.c
+++ b/drivers/gpu/drm/i915/i915_debugfs.c
@@ -107,7 +107,7 @@  static char get_tiling_flag(struct drm_i915_gem_object *obj)
 
 static char get_global_flag(struct drm_i915_gem_object *obj)
 {
-	return obj->fault_mappable ? 'g' : ' ';
+	return !list_empty(&obj->userfault_link) ? 'g' : ' ';
 }
 
 static char get_pin_mapped_flag(struct drm_i915_gem_object *obj)
diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
index fe875b27a6bf..8077d325aa78 100644
--- a/drivers/gpu/drm/i915/i915_drv.h
+++ b/drivers/gpu/drm/i915/i915_drv.h
@@ -1360,6 +1360,14 @@  struct i915_gem_mm {
 	 */
 	struct list_head unbound_list;
 
+	/** Protects access to the userfault_list */
+	spinlock_t userfault_lock;
+
+	/** List of all objects in gtt_space, currently mmaped by userspace.
+	 * All objects within this list must also be on bound_list.
+	 */
+	struct list_head userfault_list;
+
 	/** Usable portion of the GTT for GEM */
 	unsigned long stolen_base; /* limited to low memory (32-bit) */
 
@@ -2208,6 +2216,11 @@  struct drm_i915_gem_object {
 	struct drm_mm_node *stolen;
 	struct list_head global_list;
 
+	/**
+	 * Whether the object is currently in the GGTT mmap.
+	 */
+	struct list_head userfault_link;
+
 	/** Used in execbuf to temporarily hold a ref */
 	struct list_head obj_exec_link;
 
@@ -2235,13 +2248,6 @@  struct drm_i915_gem_object {
 	 */
 	unsigned int madv:2;
 
-	/**
-	 * Whether the current gtt mapping needs to be mappable (and isn't just
-	 * mappable by accident). Track pin and fault separate for a more
-	 * accurate mappable working set.
-	 */
-	unsigned int fault_mappable:1;
-
 	/*
 	 * Is the object to be mapped as read-only to the GPU
 	 * Only honoured if hardware has relevant pte bit
diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index fe92e28ea0a8..6da9514ceefa 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -1839,16 +1839,19 @@  int i915_gem_fault(struct vm_area_struct *area, struct vm_fault *vmf)
 	if (ret)
 		goto err_unpin;
 
+	/* Mark as being mmapped into userspace for later revocation */
+	spin_lock(&dev_priv->mm.userfault_lock);
+	if (list_empty(&obj->userfault_link))
+		list_add(&obj->userfault_link, &dev_priv->mm.userfault_list);
+	spin_unlock(&dev_priv->mm.userfault_lock);
+
 	/* Finally, remap it using the new GTT offset */
 	ret = remap_io_mapping(area,
 			       area->vm_start + (vma->ggtt_view.params.partial.offset << PAGE_SHIFT),
 			       (ggtt->mappable_base + vma->node.start) >> PAGE_SHIFT,
 			       min_t(u64, vma->size, area->vm_end - area->vm_start),
 			       &ggtt->mappable);
-	if (ret)
-		goto err_unpin;
 
-	obj->fault_mappable = true;
 err_unpin:
 	__i915_vma_unpin(vma);
 err_unlock:
@@ -1916,13 +1919,22 @@  err:
 void
 i915_gem_release_mmap(struct drm_i915_gem_object *obj)
 {
+	struct drm_i915_private *i915 = to_i915(obj->base.dev);
+	bool zap = false;
+
 	/* Serialisation between user GTT access and our code depends upon
 	 * revoking the CPU's PTE whilst the mutex is held. The next user
 	 * pagefault then has to wait until we release the mutex.
 	 */
-	lockdep_assert_held(&obj->base.dev->struct_mutex);
+	lockdep_assert_held(&i915->drm.struct_mutex);
 
-	if (!obj->fault_mappable)
+	spin_lock(&i915->mm.userfault_lock);
+	if (!list_empty(&obj->userfault_link)) {
+		list_del_init(&obj->userfault_link);
+		zap = true;
+	}
+	spin_unlock(&i915->mm.userfault_lock);
+	if (!zap)
 		return;
 
 	drm_vma_node_unmap(&obj->base.vma_node,
@@ -1936,8 +1948,6 @@  i915_gem_release_mmap(struct drm_i915_gem_object *obj)
 	 * memory writes before touching registers / GSM.
 	 */
 	wmb();
-
-	obj->fault_mappable = false;
 }
 
 void
@@ -1945,8 +1955,19 @@  i915_gem_release_all_mmaps(struct drm_i915_private *dev_priv)
 {
 	struct drm_i915_gem_object *obj;
 
-	list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list)
-		i915_gem_release_mmap(obj);
+	spin_lock(&dev_priv->mm.userfault_lock);
+	while ((obj = list_first_entry_or_null(&dev_priv->mm.userfault_list,
+					       struct drm_i915_gem_object,
+					       userfault_link))) {
+		list_del_init(&obj->userfault_link);
+		spin_unlock(&dev_priv->mm.userfault_lock);
+
+		drm_vma_node_unmap(&obj->base.vma_node,
+				   obj->base.dev->anon_inode->i_mapping);
+
+		spin_lock(&dev_priv->mm.userfault_lock);
+	}
+	spin_unlock(&dev_priv->mm.userfault_lock);
 }
 
 /**
@@ -4108,6 +4129,7 @@  void i915_gem_object_init(struct drm_i915_gem_object *obj,
 	int i;
 
 	INIT_LIST_HEAD(&obj->global_list);
+	INIT_LIST_HEAD(&obj->userfault_link);
 	for (i = 0; i < I915_NUM_ENGINES; i++)
 		init_request_active(&obj->last_read[i],
 				    i915_gem_object_retire__read);
@@ -4506,6 +4528,7 @@  int i915_gem_init(struct drm_device *dev)
 	int ret;
 
 	mutex_lock(&dev->struct_mutex);
+	spin_lock_init(&dev_priv->mm.userfault_lock);
 
 	if (!i915.enable_execlists) {
 		dev_priv->gt.resume = intel_legacy_submission_resume;
@@ -4625,6 +4648,7 @@  i915_gem_load_init(struct drm_device *dev)
 	INIT_LIST_HEAD(&dev_priv->mm.unbound_list);
 	INIT_LIST_HEAD(&dev_priv->mm.bound_list);
 	INIT_LIST_HEAD(&dev_priv->mm.fence_list);
+	INIT_LIST_HEAD(&dev_priv->mm.userfault_list);
 	INIT_DELAYED_WORK(&dev_priv->gt.retire_work,
 			  i915_gem_retire_work_handler);
 	INIT_DELAYED_WORK(&dev_priv->gt.idle_work,
diff --git a/drivers/gpu/drm/i915/i915_gem_evict.c b/drivers/gpu/drm/i915/i915_gem_evict.c
index b5e9e669f50f..a934f372c5ce 100644
--- a/drivers/gpu/drm/i915/i915_gem_evict.c
+++ b/drivers/gpu/drm/i915/i915_gem_evict.c
@@ -56,7 +56,7 @@  mark_free(struct i915_vma *vma, unsigned int flags, struct list_head *unwind)
 	if (WARN_ON(!list_empty(&vma->exec_list)))
 		return false;
 
-	if (flags & PIN_NONFAULT && vma->obj->fault_mappable)
+	if (flags & PIN_NONFAULT && !list_empty(&vma->obj->userfault_link))
 		return false;
 
 	list_add(&vma->exec_list, unwind);
diff --git a/drivers/gpu/drm/i915/i915_gem_fence.c b/drivers/gpu/drm/i915/i915_gem_fence.c
index a6daf2deab74..67013179b8ed 100644
--- a/drivers/gpu/drm/i915/i915_gem_fence.c
+++ b/drivers/gpu/drm/i915/i915_gem_fence.c
@@ -391,7 +391,7 @@  void i915_gem_restore_fences(struct drm_device *dev)
 		 */
 		if (vma && !i915_gem_object_is_tiled(vma->obj)) {
 			GEM_BUG_ON(!reg->dirty);
-			GEM_BUG_ON(vma->obj->fault_mappable);
+			GEM_BUG_ON(!list_empty(&vma->obj->userfault_link));
 
 			list_move(&reg->link, &dev_priv->mm.fence_list);
 			vma->fence = NULL;