Commit fbbd37b3 authored by Chris Wilson's avatar Chris Wilson
Browse files

drm/i915: Move object release to a freelist + worker 



We want to hide the latency of releasing objects and their backing
storage from the submission, so we move the actual free to a worker.
This allows us to switch to struct_mutex freeing of the object in the
next patch.

Furthermore, if we know that the object we are dereferencing remains valid
for the duration of our access, we can forgo the usual synchronisation
barriers and atomic reference counting. To ensure this we defer freeing
an object til after an RCU grace period, such that any lookup of the
object within an RCU read critical section will remain valid until
after we exit that critical section. We also employ this delay for
rate-limiting the serialisation on reallocation - we have to slow down
object creation in order to prevent resource starvation (in particular,
files).

v2: Return early in i915_gem_tiling() ioctl to skip over superfluous
work on error.

Signed-off-by: default avatarChris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: default avatarJoonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20161028125858.23563-19-chris@chris-wilson.co.uk
parent 40e62d5d

drivers/gpu/drm/i915/i915_debugfs.c

+11 −4
Original line number Diff line number Diff line
@@ -4957,10 +4957,12 @@ DEFINE_SIMPLE_ATTRIBUTE(i915_ring_test_irq_fops, 
#define DROP_BOUND 0x2

#define DROP_RETIRE 0x4

#define DROP_ACTIVE 0x8

#define DROP_FREED 0x10

#define DROP_ALL (DROP_UNBOUND	| \

		  DROP_BOUND	| \

		  DROP_RETIRE	| \

		  DROP_ACTIVE)

		  DROP_ACTIVE	| \

		  DROP_FREED)

static int

i915_drop_caches_get(void *data, u64 *val)

{
@@ -5004,6 +5006,11 @@ i915_drop_caches_set(void *data, u64 val) 
unlock:

	mutex_unlock(&dev->struct_mutex);



	if (val & DROP_FREED) {

		synchronize_rcu();

		flush_work(&dev_priv->mm.free_work);

	}



	return ret;

}

drivers/gpu/drm/i915/i915_drv.c

+11 −8
Original line number Diff line number Diff line
@@ -537,14 +537,17 @@ static const struct vga_switcheroo_client_ops i915_switcheroo_ops = { 
	.can_switch = i915_switcheroo_can_switch,

};



static void i915_gem_fini(struct drm_device *dev)

static void i915_gem_fini(struct drm_i915_private *dev_priv)

{

	mutex_lock(&dev->struct_mutex);

	i915_gem_cleanup_engines(dev);

	i915_gem_context_fini(dev);

	mutex_unlock(&dev->struct_mutex);

	mutex_lock(&dev_priv->drm.struct_mutex);

	i915_gem_cleanup_engines(&dev_priv->drm);

	i915_gem_context_fini(&dev_priv->drm);

	mutex_unlock(&dev_priv->drm.struct_mutex);



	synchronize_rcu();

	flush_work(&dev_priv->mm.free_work);



	WARN_ON(!list_empty(&to_i915(dev)->context_list));

	WARN_ON(!list_empty(&dev_priv->context_list));

}



static int i915_load_modeset_init(struct drm_device *dev)
@@ -619,7 +622,7 @@ static int i915_load_modeset_init(struct drm_device *dev) 
cleanup_gem:

	if (i915_gem_suspend(dev))

		DRM_ERROR("failed to idle hardware; continuing to unload!\n");

	i915_gem_fini(dev);

	i915_gem_fini(dev_priv);

cleanup_irq:

	intel_guc_fini(dev);

	drm_irq_uninstall(dev);
@@ -1305,7 +1308,7 @@ void i915_driver_unload(struct drm_device *dev) 
	drain_workqueue(dev_priv->wq);



	intel_guc_fini(dev);

	i915_gem_fini(dev);

	i915_gem_fini(dev_priv);

	intel_fbc_cleanup_cfb(dev_priv);



	intel_power_domains_fini(dev_priv);

drivers/gpu/drm/i915/i915_drv.h

+41 −3
Original line number Diff line number Diff line
@@ -1365,8 +1365,8 @@ struct i915_gem_mm { 
	struct list_head bound_list;

	/**

	 * List of objects which are not bound to the GTT (thus

	 * are idle and not used by the GPU) but still have

	 * (presumably uncached) pages still attached.

	 * are idle and not used by the GPU). These objects may or may

	 * not actually have any pages attached.

	 */

	struct list_head unbound_list;
@@ -1375,6 +1375,12 @@ struct i915_gem_mm { 
	 */

	struct list_head userfault_list;



	/**

	 * List of objects which are pending destruction.

	 */

	struct llist_head free_list;

	struct work_struct free_work;



	/** Usable portion of the GTT for GEM */

	unsigned long stolen_base; /* limited to low memory (32-bit) */
@@ -2224,6 +2230,10 @@ struct drm_i915_gem_object { 
	/** Stolen memory for this object, instead of being backed by shmem. */

	struct drm_mm_node *stolen;

	struct list_head global_list;

	union {

		struct rcu_head rcu;

		struct llist_node freed;

	};



	/**

	 * Whether the object is currently in the GGTT mmap.
@@ -2341,10 +2351,38 @@ to_intel_bo(struct drm_gem_object *gem) 
	return container_of(gem, struct drm_i915_gem_object, base);

}



/**

 * i915_gem_object_lookup_rcu - look up a temporary GEM object from its handle

 * @filp: DRM file private date

 * @handle: userspace handle

 *

 * Returns:

 *

 * A pointer to the object named by the handle if such exists on @filp, NULL

 * otherwise. This object is only valid whilst under the RCU read lock, and

 * note carefully the object may be in the process of being destroyed.

 */

static inline struct drm_i915_gem_object *

i915_gem_object_lookup_rcu(struct drm_file *file, u32 handle)

{

#ifdef CONFIG_LOCKDEP

	WARN_ON(debug_locks && !lock_is_held(&rcu_lock_map));

#endif

	return idr_find(&file->object_idr, handle);

}



static inline struct drm_i915_gem_object *

i915_gem_object_lookup(struct drm_file *file, u32 handle)

{

	return to_intel_bo(drm_gem_object_lookup(file, handle));

	struct drm_i915_gem_object *obj;



	rcu_read_lock();

	obj = i915_gem_object_lookup_rcu(file, handle);

	if (obj && !kref_get_unless_zero(&obj->base.refcount))

		obj = NULL;

	rcu_read_unlock();



	return obj;

}



__deprecated

drivers/gpu/drm/i915/i915_gem.c

+116 −50
Original line number Diff line number Diff line
@@ -42,6 +42,7 @@ 
#include <linux/pci.h>

#include <linux/dma-buf.h>



static void i915_gem_flush_free_objects(struct drm_i915_private *i915);

static void i915_gem_object_flush_gtt_write_domain(struct drm_i915_gem_object *obj);

static void i915_gem_object_flush_cpu_write_domain(struct drm_i915_gem_object *obj);
@@ -648,6 +649,8 @@ i915_gem_create_ioctl(struct drm_device *dev, void *data, 
{

	struct drm_i915_gem_create *args = data;



	i915_gem_flush_free_objects(to_i915(dev));



	return i915_gem_create(file, dev,

			       args->size, &args->handle);

}
@@ -3524,10 +3527,14 @@ int i915_gem_get_caching_ioctl(struct drm_device *dev, void *data, 
{

	struct drm_i915_gem_caching *args = data;

	struct drm_i915_gem_object *obj;

	int err = 0;



	obj = i915_gem_object_lookup(file, args->handle);

	if (!obj)

		return -ENOENT;

	rcu_read_lock();

	obj = i915_gem_object_lookup_rcu(file, args->handle);

	if (!obj) {

		err = -ENOENT;

		goto out;

	}



	switch (obj->cache_level) {

	case I915_CACHE_LLC:
@@ -3543,9 +3550,9 @@ int i915_gem_get_caching_ioctl(struct drm_device *dev, void *data, 
		args->caching = I915_CACHING_NONE;

		break;

	}



	i915_gem_object_put_unlocked(obj);

	return 0;

out:

	rcu_read_unlock();

	return err;

}



int i915_gem_set_caching_ioctl(struct drm_device *dev, void *data,
@@ -4089,10 +4096,14 @@ i915_gem_busy_ioctl(struct drm_device *dev, void *data, 
	struct drm_i915_gem_busy *args = data;

	struct drm_i915_gem_object *obj;

	unsigned long active;

	int err;



	obj = i915_gem_object_lookup(file, args->handle);

	if (!obj)

		return -ENOENT;

	rcu_read_lock();

	obj = i915_gem_object_lookup_rcu(file, args->handle);

	if (!obj) {

		err = -ENOENT;

		goto out;

	}



	args->busy = 0;

	active = __I915_BO_ACTIVE(obj);
@@ -4122,7 +4133,6 @@ i915_gem_busy_ioctl(struct drm_device *dev, void *data, 
		 * are busy is not completely reliable - we only guarantee

		 * that the object was busy.

		 */

		rcu_read_lock();



		for_each_active(active, idx)

			args->busy |= busy_check_reader(&obj->last_read[idx]);
@@ -4140,12 +4150,11 @@ i915_gem_busy_ioctl(struct drm_device *dev, void *data, 
		 * the result.

		 */

		args->busy |= busy_check_writer(&obj->last_write);



		rcu_read_unlock();

	}



	i915_gem_object_put_unlocked(obj);

	return 0;

out:

	rcu_read_unlock();

	return err;

}



int
@@ -4308,7 +4317,6 @@ i915_gem_object_create(struct drm_device *dev, u64 size) 


fail:

	i915_gem_object_free(obj);



	return ERR_PTR(ret);

}
@@ -4336,60 +4344,115 @@ static bool discard_backing_storage(struct drm_i915_gem_object *obj) 
	return atomic_long_read(&obj->base.filp->f_count) == 1;

}



void i915_gem_free_object(struct drm_gem_object *gem_obj)

static void __i915_gem_free_objects(struct drm_i915_private *i915,

				    struct llist_node *freed)

{

	struct drm_i915_gem_object *obj = to_intel_bo(gem_obj);

	struct drm_device *dev = obj->base.dev;

	struct drm_i915_private *dev_priv = to_i915(dev);

	struct i915_vma *vma, *next;

	struct drm_i915_gem_object *obj, *on;



	intel_runtime_pm_get(dev_priv);

	mutex_lock(&i915->drm.struct_mutex);

	intel_runtime_pm_get(i915);

	llist_for_each_entry(obj, freed, freed) {

		struct i915_vma *vma, *vn;



		trace_i915_gem_object_destroy(obj);



	/* All file-owned VMA should have been released by this point through

	 * i915_gem_close_object(), or earlier by i915_gem_context_close().

	 * However, the object may also be bound into the global GTT (e.g.

	 * older GPUs without per-process support, or for direct access through

	 * the GTT either for the user or for scanout). Those VMA still need to

	 * unbound now.

	 */

	list_for_each_entry_safe(vma, next, &obj->vma_list, obj_link) {

		GEM_BUG_ON(i915_gem_object_is_active(obj));

		list_for_each_entry_safe(vma, vn,

					 &obj->vma_list, obj_link) {

			GEM_BUG_ON(!i915_vma_is_ggtt(vma));

			GEM_BUG_ON(i915_vma_is_active(vma));

			vma->flags &= ~I915_VMA_PIN_MASK;

			i915_vma_close(vma);

		}

	GEM_BUG_ON(obj->bind_count);



	WARN_ON(atomic_read(&obj->frontbuffer_bits));

		list_del(&obj->global_list);

	}

	intel_runtime_pm_put(i915);

	mutex_unlock(&i915->drm.struct_mutex);



	if (obj->mm.pages && obj->mm.madv == I915_MADV_WILLNEED &&

	    dev_priv->quirks & QUIRK_PIN_SWIZZLED_PAGES &&

	    i915_gem_object_is_tiled(obj))

		__i915_gem_object_unpin_pages(obj);

	llist_for_each_entry_safe(obj, on, freed, freed) {

		GEM_BUG_ON(obj->bind_count);

		GEM_BUG_ON(atomic_read(&obj->frontbuffer_bits));



		if (obj->ops->release)

			obj->ops->release(obj);



		if (WARN_ON(i915_gem_object_has_pinned_pages(obj)))

			atomic_set(&obj->mm.pages_pin_count, 0);

	if (discard_backing_storage(obj))

		obj->mm.madv = I915_MADV_DONTNEED;

		__i915_gem_object_put_pages(obj);



		GEM_BUG_ON(obj->mm.pages);



		if (obj->base.import_attach)

			drm_prime_gem_destroy(&obj->base, NULL);



		drm_gem_object_release(&obj->base);

	i915_gem_info_remove_obj(dev_priv, obj->base.size);

		i915_gem_info_remove_obj(i915, obj->base.size);



		kfree(obj->bit_17);

		i915_gem_object_free(obj);

	}

}



	intel_runtime_pm_put(dev_priv);

static void i915_gem_flush_free_objects(struct drm_i915_private *i915)

{

	struct llist_node *freed;



	freed = llist_del_all(&i915->mm.free_list);

	if (unlikely(freed))

		__i915_gem_free_objects(i915, freed);

}



static void __i915_gem_free_work(struct work_struct *work)

{

	struct drm_i915_private *i915 =

		container_of(work, struct drm_i915_private, mm.free_work);

	struct llist_node *freed;



	/* All file-owned VMA should have been released by this point through

	 * i915_gem_close_object(), or earlier by i915_gem_context_close().

	 * However, the object may also be bound into the global GTT (e.g.

	 * older GPUs without per-process support, or for direct access through

	 * the GTT either for the user or for scanout). Those VMA still need to

	 * unbound now.

	 */



	while ((freed = llist_del_all(&i915->mm.free_list)))

		__i915_gem_free_objects(i915, freed);

}



static void __i915_gem_free_object_rcu(struct rcu_head *head)

{

	struct drm_i915_gem_object *obj =

		container_of(head, typeof(*obj), rcu);

	struct drm_i915_private *i915 = to_i915(obj->base.dev);



	/* We can't simply use call_rcu() from i915_gem_free_object()

	 * as we need to block whilst unbinding, and the call_rcu

	 * task may be called from softirq context. So we take a

	 * detour through a worker.

	 */

	if (llist_add(&obj->freed, &i915->mm.free_list))

		schedule_work(&i915->mm.free_work);

}



void i915_gem_free_object(struct drm_gem_object *gem_obj)

{

	struct drm_i915_gem_object *obj = to_intel_bo(gem_obj);



	if (discard_backing_storage(obj))

		obj->mm.madv = I915_MADV_DONTNEED;



	if (obj->mm.pages && obj->mm.madv == I915_MADV_WILLNEED &&

	    to_i915(obj->base.dev)->quirks & QUIRK_PIN_SWIZZLED_PAGES &&

	    i915_gem_object_is_tiled(obj))

		__i915_gem_object_unpin_pages(obj);



	/* Before we free the object, make sure any pure RCU-only

	 * read-side critical sections are complete, e.g.

	 * i915_gem_busy_ioctl(). For the corresponding synchronized

	 * lookup see i915_gem_object_lookup_rcu().

	 */

	call_rcu(&obj->rcu, __i915_gem_free_object_rcu);

}



void __i915_gem_object_release_unless_active(struct drm_i915_gem_object *obj)
@@ -4438,6 +4501,7 @@ int i915_gem_suspend(struct drm_device *dev) 
	cancel_delayed_work_sync(&dev_priv->gpu_error.hangcheck_work);

	cancel_delayed_work_sync(&dev_priv->gt.retire_work);

	flush_delayed_work(&dev_priv->gt.idle_work);

	flush_work(&dev_priv->mm.free_work);



	/* Assert that we sucessfully flushed all the work and

	 * reset the GPU back to its idle, low power state.
@@ -4753,6 +4817,8 @@ i915_gem_load_init(struct drm_device *dev) 
				  NULL);



	INIT_LIST_HEAD(&dev_priv->context_list);

	INIT_WORK(&dev_priv->mm.free_work, __i915_gem_free_work);

	init_llist_head(&dev_priv->mm.free_list);

	INIT_LIST_HEAD(&dev_priv->mm.unbound_list);

	INIT_LIST_HEAD(&dev_priv->mm.bound_list);

	INIT_LIST_HEAD(&dev_priv->mm.fence_list);

drivers/gpu/drm/i915/i915_gem_shrinker.c

+10 −4
Original line number Diff line number Diff line
@@ -201,6 +201,10 @@ i915_gem_shrink(struct drm_i915_private *dev_priv, 
						       typeof(*obj),

						       global_list))) {

			list_move_tail(&obj->global_list, &still_in_list);

			if (!obj->mm.pages) {

				list_del_init(&obj->global_list);

				continue;

			}



			if (flags & I915_SHRINK_PURGEABLE &&

			    obj->mm.madv != I915_MADV_DONTNEED)
@@ -218,8 +222,6 @@ i915_gem_shrink(struct drm_i915_private *dev_priv, 
			if (!can_release_pages(obj))

				continue;



			i915_gem_object_get(obj);



			if (unsafe_drop_pages(obj)) {

				mutex_lock(&obj->mm.lock);

				if (!obj->mm.pages) {
@@ -228,8 +230,6 @@ i915_gem_shrink(struct drm_i915_private *dev_priv, 
				}

				mutex_unlock(&obj->mm.lock);

			}



			i915_gem_object_put(obj);

		}

		list_splice(&still_in_list, phase->list);

	}
@@ -396,12 +396,18 @@ i915_gem_shrinker_oom(struct notifier_block *nb, unsigned long event, void *ptr) 
	 */

	unbound = bound = unevictable = 0;

	list_for_each_entry(obj, &dev_priv->mm.unbound_list, global_list) {

		if (!obj->mm.pages)

			continue;



		if (!can_release_pages(obj))

			unevictable += obj->base.size >> PAGE_SHIFT;

		else

			unbound += obj->base.size >> PAGE_SHIFT;

	}

	list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {

		if (!obj->mm.pages)

			continue;



		if (!can_release_pages(obj))

			unevictable += obj->base.size >> PAGE_SHIFT;

		else

CRA Git | Maintained and supported by SUSTech CRA and CCSE