1 files changed, 465 insertions, 105 deletions

diff --git a/drivers/gpu/drm/i915/i915_gem_evict.c b/drivers/gpu/drm/i915/i915_gem_evict.c
index c86d5d9356fd..c02ebd6900ae 100644
--- a/drivers/gpu/drm/i915/i915_gem_evict.c
+++ b/drivers/gpu/drm/i915/i915_gem_evict.c
@@ -26,160 +26,520 @@
  *
  */
 
-#include <drm/drmP.h>
+#include "gem/i915_gem_context.h"
+#include "gt/intel_gt.h"
+#include "gt/intel_gt_requests.h"
+
 #include "i915_drv.h"
-#include <drm/i915_drm.h>
+#include "i915_gem_evict.h"
 #include "i915_trace.h"
 
+I915_SELFTEST_DECLARE(static struct igt_evict_ctl {
+	bool fail_if_busy:1;
+} igt_evict_ctl;)
+
+static bool dying_vma(struct i915_vma *vma)
+{
+	return !kref_read(&vma->obj->base.refcount);
+}
+
+static int ggtt_flush(struct i915_address_space *vm)
+{
+	struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
+	struct intel_gt *gt;
+	int ret = 0;
+
+	list_for_each_entry(gt, &ggtt->gt_list, ggtt_link) {
+		/*
+		 * Not everything in the GGTT is tracked via vma (otherwise we
+		 * could evict as required with minimal stalling) so we are forced
+		 * to idle the GPU and explicitly retire outstanding requests in
+		 * the hopes that we can then remove contexts and the like only
+		 * bound by their active reference.
+		 */
+		ret = intel_gt_wait_for_idle(gt, MAX_SCHEDULE_TIMEOUT);
+		if (ret)
+			return ret;
+	}
+	return ret;
+}
+
+static bool grab_vma(struct i915_vma *vma, struct i915_gem_ww_ctx *ww)
+{
+	/*
+	 * We add the extra refcount so the object doesn't drop to zero until
+	 * after ungrab_vma(), this way trylock is always paired with unlock.
+	 */
+	if (i915_gem_object_get_rcu(vma->obj)) {
+		if (!i915_gem_object_trylock(vma->obj, ww)) {
+			i915_gem_object_put(vma->obj);
+			return false;
+		}
+	} else {
+		/* Dead objects don't need pins */
+		atomic_and(~I915_VMA_PIN_MASK, &vma->flags);
+	}
+
+	return true;
+}
+
+static void ungrab_vma(struct i915_vma *vma)
+{
+	if (dying_vma(vma))
+		return;
+
+	i915_gem_object_unlock(vma->obj);
+	i915_gem_object_put(vma->obj);
+}
+
 static bool
-mark_free(struct drm_i915_gem_object *obj, struct list_head *unwind)
+mark_free(struct drm_mm_scan *scan,
+	  struct i915_gem_ww_ctx *ww,
+	  struct i915_vma *vma,
+	  unsigned int flags,
+	  struct list_head *unwind)
 {
-	if (obj->pin_count)
+	if (i915_vma_is_pinned(vma))
 		return false;
 
-	list_add(&obj->exec_list, unwind);
-	return drm_mm_scan_add_block(obj->gtt_space);
+	if (!grab_vma(vma, ww))
+		return false;
+
+	list_add(&vma->evict_link, unwind);
+	return drm_mm_scan_add_block(scan, &vma->node);
+}
+
+static bool defer_evict(struct i915_vma *vma)
+{
+	if (i915_vma_is_active(vma))
+		return true;
+
+	if (i915_vma_is_scanout(vma))
+		return true;
+
+	return false;
 }
 
+/**
+ * i915_gem_evict_something - Evict vmas to make room for binding a new one
+ * @vm: address space to evict from
+ * @ww: An optional struct i915_gem_ww_ctx.
+ * @min_size: size of the desired free space
+ * @alignment: alignment constraint of the desired free space
+ * @color: color for the desired space
+ * @start: start (inclusive) of the range from which to evict objects
+ * @end: end (exclusive) of the range from which to evict objects
+ * @flags: additional flags to control the eviction algorithm
+ *
+ * This function will try to evict vmas until a free space satisfying the
+ * requirements is found. Callers must check first whether any such hole exists
+ * already before calling this function.
+ *
+ * This function is used by the object/vma binding code.
+ *
+ * Since this function is only used to free up virtual address space it only
+ * ignores pinned vmas, and not object where the backing storage itself is
+ * pinned. Hence obj->pages_pin_count does not protect against eviction.
+ *
+ * To clarify: This is for freeing up virtual address space, not for freeing
+ * memory in e.g. the shrinker.
+ */
 int
-i915_gem_evict_something(struct drm_device *dev, int min_size,
-			 unsigned alignment, unsigned cache_level,
-			 bool mappable, bool nonblocking)
+i915_gem_evict_something(struct i915_address_space *vm,
+			 struct i915_gem_ww_ctx *ww,
+			 u64 min_size, u64 alignment,
+			 unsigned long color,
+			 u64 start, u64 end,
+			 unsigned flags)
 {
-	drm_i915_private_t *dev_priv = dev->dev_private;
-	struct list_head eviction_list, unwind_list;
-	struct drm_i915_gem_object *obj;
-	int ret = 0;
+	struct drm_mm_scan scan;
+	struct list_head eviction_list;
+	struct i915_vma *vma, *next;
+	struct drm_mm_node *node;
+	enum drm_mm_insert_mode mode;
+	struct i915_vma *active;
+	struct intel_gt *gt;
+	int ret;
 
-	trace_i915_gem_evict(dev, min_size, alignment, mappable);
+	lockdep_assert_held(&vm->mutex);
+	trace_i915_gem_evict(vm, min_size, alignment, flags);
 
 	/*
-	 * The goal is to evict objects and amalgamate space in LRU order.
-	 * The oldest idle objects reside on the inactive list, which is in
-	 * retirement order. The next objects to retire are those on the (per
-	 * ring) active list that do not have an outstanding flush. Once the
-	 * hardware reports completion (the seqno is updated after the
-	 * batchbuffer has been finished) the clean buffer objects would
-	 * be retired to the inactive list. Any dirty objects would be added
-	 * to the tail of the flushing list. So after processing the clean
-	 * active objects we need to emit a MI_FLUSH to retire the flushing
-	 * list, hence the retirement order of the flushing list is in
-	 * advance of the dirty objects on the active lists.
+	 * The goal is to evict objects and amalgamate space in rough LRU order.
+	 * Since both active and inactive objects reside on the same list,
+	 * in a mix of creation and last scanned order, as we process the list
+	 * we sort it into inactive/active, which keeps the active portion
+	 * in a rough MRU order.
 	 *
 	 * The retirement sequence is thus:
-	 *   1. Inactive objects (already retired)
-	 *   2. Clean active objects
-	 *   3. Flushing list
-	 *   4. Dirty active objects.
-	 *
-	 * On each list, the oldest objects lie at the HEAD with the freshest
-	 * object on the TAIL.
+	 *   1. Inactive objects (already retired, random order)
+	 *   2. Active objects (will stall on unbinding, oldest scanned first)
 	 */
+	mode = DRM_MM_INSERT_BEST;
+	if (flags & PIN_HIGH)
+		mode = DRM_MM_INSERT_HIGH;
+	if (flags & PIN_MAPPABLE)
+		mode = DRM_MM_INSERT_LOW;
+	drm_mm_scan_init_with_range(&scan, &vm->mm,
+				    min_size, alignment, color,
+				    start, end, mode);
 
-	INIT_LIST_HEAD(&unwind_list);
-	if (mappable)
-		drm_mm_init_scan_with_range(&dev_priv->mm.gtt_space,
-					    min_size, alignment, cache_level,
-					    0, dev_priv->gtt.mappable_end);
-	else
-		drm_mm_init_scan(&dev_priv->mm.gtt_space,
-				 min_size, alignment, cache_level);
-
-	/* First see if there is a large enough contiguous idle region... */
-	list_for_each_entry(obj, &dev_priv->mm.inactive_list, mm_list) {
-		if (mark_free(obj, &unwind_list))
-			goto found;
+	if (i915_is_ggtt(vm)) {
+		struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
+
+		list_for_each_entry(gt, &ggtt->gt_list, ggtt_link)
+			intel_gt_retire_requests(gt);
+	} else {
+		intel_gt_retire_requests(vm->gt);
 	}
 
-	if (nonblocking)
-		goto none;
+search_again:
+	active = NULL;
+	INIT_LIST_HEAD(&eviction_list);
+	list_for_each_entry_safe(vma, next, &vm->bound_list, vm_link) {
+		if (vma == active) { /* now seen this vma twice */
+			if (flags & PIN_NONBLOCK)
+				break;
+
+			active = ERR_PTR(-EAGAIN);
+		}
+
+		/*
+		 * We keep this list in a rough least-recently scanned order
+		 * of active elements (inactive elements are cheap to reap).
+		 * New entries are added to the end, and we move anything we
+		 * scan to the end. The assumption is that the working set
+		 * of applications is either steady state (and thanks to the
+		 * userspace bo cache it almost always is) or volatile and
+		 * frequently replaced after a frame, which are self-evicting!
+		 * Given that assumption, the MRU order of the scan list is
+		 * fairly static, and keeping it in least-recently scan order
+		 * is suitable.
+		 *
+		 * To notice when we complete one full cycle, we record the
+		 * first active element seen, before moving it to the tail.
+		 */
+		if (active != ERR_PTR(-EAGAIN) && defer_evict(vma)) {
+			if (!active)
+				active = vma;
+
+			list_move_tail(&vma->vm_link, &vm->bound_list);
+			continue;
+		}
 
-	/* Now merge in the soon-to-be-expired objects... */
-	list_for_each_entry(obj, &dev_priv->mm.active_list, mm_list) {
-		if (mark_free(obj, &unwind_list))
+		if (mark_free(&scan, ww, vma, flags, &eviction_list))
 			goto found;
 	}
 
-none:
 	/* Nothing found, clean up and bail out! */
-	while (!list_empty(&unwind_list)) {
-		obj = list_first_entry(&unwind_list,
-				       struct drm_i915_gem_object,
-				       exec_list);
-
-		ret = drm_mm_scan_remove_block(obj->gtt_space);
+	list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
+		ret = drm_mm_scan_remove_block(&scan, &vma->node);
 		BUG_ON(ret);
-
-		list_del_init(&obj->exec_list);
+		ungrab_vma(vma);
 	}
 
-	/* We expect the caller to unpin, evict all and try again, or give up.
-	 * So calling i915_gem_evict_everything() is unnecessary.
+	/*
+	 * Can we unpin some objects such as idle hw contents,
+	 * or pending flips? But since only the GGTT has global entries
+	 * such as scanouts, rinbuffers and contexts, we can skip the
+	 * purge when inspecting per-process local address spaces.
+	 */
+	if (!i915_is_ggtt(vm) || flags & PIN_NONBLOCK)
+		return -ENOSPC;
+
+	/*
+	 * Not everything in the GGTT is tracked via VMA using
+	 * i915_vma_move_to_active(), otherwise we could evict as required
+	 * with minimal stalling. Instead we are forced to idle the GPU and
+	 * explicitly retire outstanding requests which will then remove
+	 * the pinning for active objects such as contexts and ring,
+	 * enabling us to evict them on the next iteration.
+	 *
+	 * To ensure that all user contexts are evictable, we perform
+	 * a switch to the perma-pinned kernel context. This all also gives
+	 * us a termination condition, when the last retired context is
+	 * the kernel's there is no more we can evict.
 	 */
-	return -ENOSPC;
+	if (I915_SELFTEST_ONLY(igt_evict_ctl.fail_if_busy))
+		return -EBUSY;
+
+	ret = ggtt_flush(vm);
+	if (ret)
+		return ret;
+
+	cond_resched();
+
+	flags |= PIN_NONBLOCK;
+	goto search_again;
 
 found:
 	/* drm_mm doesn't allow any other other operations while
-	 * scanning, therefore store to be evicted objects on a
-	 * temporary list. */
-	INIT_LIST_HEAD(&eviction_list);
-	while (!list_empty(&unwind_list)) {
-		obj = list_first_entry(&unwind_list,
-				       struct drm_i915_gem_object,
-				       exec_list);
-		if (drm_mm_scan_remove_block(obj->gtt_space)) {
-			list_move(&obj->exec_list, &eviction_list);
-			drm_gem_object_reference(&obj->base);
-			continue;
+	 * scanning, therefore store to-be-evicted objects on a
+	 * temporary list and take a reference for all before
+	 * calling unbind (which may remove the active reference
+	 * of any of our objects, thus corrupting the list).
+	 */
+	list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
+		if (drm_mm_scan_remove_block(&scan, &vma->node)) {
+			__i915_vma_pin(vma);
+		} else {
+			list_del(&vma->evict_link);
+			ungrab_vma(vma);
 		}
-		list_del_init(&obj->exec_list);
 	}
 
 	/* Unbinding will emit any required flushes */
-	while (!list_empty(&eviction_list)) {
-		obj = list_first_entry(&eviction_list,
-				       struct drm_i915_gem_object,
-				       exec_list);
+	ret = 0;
+	list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
+		__i915_vma_unpin(vma);
 		if (ret == 0)
-			ret = i915_gem_object_unbind(obj);
+			ret = __i915_vma_unbind(vma);
+		ungrab_vma(vma);
+	}
 
-		list_del_init(&obj->exec_list);
-		drm_gem_object_unreference(&obj->base);
+	while (ret == 0 && (node = drm_mm_scan_color_evict(&scan))) {
+		vma = container_of(node, struct i915_vma, node);
+
+		/* If we find any non-objects (!vma), we cannot evict them */
+		if (vma->node.color != I915_COLOR_UNEVICTABLE &&
+		    grab_vma(vma, ww)) {
+			ret = __i915_vma_unbind(vma);
+			ungrab_vma(vma);
+		} else {
+			ret = -ENOSPC;
+		}
 	}
 
 	return ret;
 }
 
-int
-i915_gem_evict_everything(struct drm_device *dev)
+/**
+ * i915_gem_evict_for_node - Evict vmas to make room for binding a new one
+ * @vm: address space to evict from
+ * @ww: An optional struct i915_gem_ww_ctx.
+ * @target: range (and color) to evict for
+ * @flags: additional flags to control the eviction algorithm
+ *
+ * This function will try to evict vmas that overlap the target node.
+ *
+ * To clarify: This is for freeing up virtual address space, not for freeing
+ * memory in e.g. the shrinker.
+ */
+int i915_gem_evict_for_node(struct i915_address_space *vm,
+			    struct i915_gem_ww_ctx *ww,
+			    struct drm_mm_node *target,
+			    unsigned int flags)
 {
-	drm_i915_private_t *dev_priv = dev->dev_private;
-	struct drm_i915_gem_object *obj, *next;
-	bool lists_empty;
-	int ret;
+	LIST_HEAD(eviction_list);
+	struct drm_mm_node *node;
+	u64 start = target->start;
+	u64 end = start + target->size;
+	struct i915_vma *vma, *next;
+	int ret = 0;
 
-	lists_empty = (list_empty(&dev_priv->mm.inactive_list) &&
-		       list_empty(&dev_priv->mm.active_list));
-	if (lists_empty)
-		return -ENOSPC;
+	lockdep_assert_held(&vm->mutex);
+	GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
+	GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
 
-	trace_i915_gem_evict_everything(dev);
+	trace_i915_gem_evict_node(vm, target, flags);
 
-	/* The gpu_idle will flush everything in the write domain to the
-	 * active list. Then we must move everything off the active list
-	 * with retire requests.
+	/*
+	 * Retire before we search the active list. Although we have
+	 * reasonable accuracy in our retirement lists, we may have
+	 * a stray pin (preventing eviction) that can only be resolved by
+	 * retiring.
 	 */
-	ret = i915_gpu_idle(dev);
-	if (ret)
-		return ret;
+	if (i915_is_ggtt(vm)) {
+		struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
+		struct intel_gt *gt;
+
+		list_for_each_entry(gt, &ggtt->gt_list, ggtt_link)
+			intel_gt_retire_requests(gt);
+	} else {
+		intel_gt_retire_requests(vm->gt);
+	}
+
+	if (i915_vm_has_cache_coloring(vm)) {
+		/* Expand search to cover neighbouring guard pages (or lack!) */
+		if (start)
+			start -= I915_GTT_PAGE_SIZE;
+
+		/* Always look at the page afterwards to avoid the end-of-GTT */
+		end += I915_GTT_PAGE_SIZE;
+	}
+	GEM_BUG_ON(start >= end);
+
+	drm_mm_for_each_node_in_range(node, &vm->mm, start, end) {
+		/* If we find any non-objects (!vma), we cannot evict them */
+		if (node->color == I915_COLOR_UNEVICTABLE) {
+			ret = -ENOSPC;
+			break;
+		}
+
+		GEM_BUG_ON(!drm_mm_node_allocated(node));
+		vma = container_of(node, typeof(*vma), node);
+
+		/*
+		 * If we are using coloring to insert guard pages between
+		 * different cache domains within the address space, we have
+		 * to check whether the objects on either side of our range
+		 * abutt and conflict. If they are in conflict, then we evict
+		 * those as well to make room for our guard pages.
+		 */
+		if (i915_vm_has_cache_coloring(vm)) {
+			if (node->start + node->size == target->start) {
+				if (node->color == target->color)
+					continue;
+			}
+			if (node->start == target->start + target->size) {
+				if (node->color == target->color)
+					continue;
+			}
+		}
+
+		if (i915_vma_is_pinned(vma)) {
+			ret = -ENOSPC;
+			break;
+		}
+
+		if (flags & PIN_NONBLOCK && i915_vma_is_active(vma)) {
+			ret = -ENOSPC;
+			break;
+		}
+
+		if (!grab_vma(vma, ww)) {
+			ret = -ENOSPC;
+			break;
+		}
+
+		/*
+		 * Never show fear in the face of dragons!
+		 *
+		 * We cannot directly remove this node from within this
+		 * iterator and as with i915_gem_evict_something() we employ
+		 * the vma pin_count in order to prevent the action of
+		 * unbinding one vma from freeing (by dropping its active
+		 * reference) another in our eviction list.
+		 */
+		__i915_vma_pin(vma);
+		list_add(&vma->evict_link, &eviction_list);
+	}
 
-	i915_gem_retire_requests(dev);
+	list_for_each_entry_safe(vma, next, &eviction_list, evict_link) {
+		__i915_vma_unpin(vma);
+		if (ret == 0)
+			ret = __i915_vma_unbind(vma);
 
-	/* Having flushed everything, unbind() should never raise an error */
-	list_for_each_entry_safe(obj, next,
-				 &dev_priv->mm.inactive_list, mm_list)
-		if (obj->pin_count == 0)
-			WARN_ON(i915_gem_object_unbind(obj));
+		ungrab_vma(vma);
+	}
 
-	return 0;
+	return ret;
 }
+
+/**
+ * i915_gem_evict_vm - Evict all idle vmas from a vm
+ * @vm: Address space to cleanse
+ * @ww: An optional struct i915_gem_ww_ctx. If not NULL, i915_gem_evict_vm
+ * will be able to evict vma's locked by the ww as well.
+ * @busy_bo: Optional pointer to struct drm_i915_gem_object. If not NULL, then
+ * in the event i915_gem_evict_vm() is unable to trylock an object for eviction,
+ * then @busy_bo will point to it. -EBUSY is also returned. The caller must drop
+ * the vm->mutex, before trying again to acquire the contended lock. The caller
+ * also owns a reference to the object.
+ *
+ * This function evicts all vmas from a vm.
+ *
+ * This is used by the execbuf code as a last-ditch effort to defragment the
+ * address space.
+ *
+ * To clarify: This is for freeing up virtual address space, not for freeing
+ * memory in e.g. the shrinker.
+ */
+int i915_gem_evict_vm(struct i915_address_space *vm, struct i915_gem_ww_ctx *ww,
+		      struct drm_i915_gem_object **busy_bo)
+{
+	int ret = 0;
+
+	lockdep_assert_held(&vm->mutex);
+	trace_i915_gem_evict_vm(vm);
+
+	/* Switch back to the default context in order to unpin
+	 * the existing context objects. However, such objects only
+	 * pin themselves inside the global GTT and performing the
+	 * switch otherwise is ineffective.
+	 */
+	if (i915_is_ggtt(vm)) {
+		ret = ggtt_flush(vm);
+		if (ret)
+			return ret;
+	}
+
+	do {
+		struct i915_vma *vma, *vn;
+		LIST_HEAD(eviction_list);
+		LIST_HEAD(locked_eviction_list);
+
+		list_for_each_entry(vma, &vm->bound_list, vm_link) {
+			if (i915_vma_is_pinned(vma))
+				continue;
+
+			/*
+			 * If we already own the lock, trylock fails. In case
+			 * the resv is shared among multiple objects, we still
+			 * need the object ref.
+			 */
+			if (!i915_gem_object_get_rcu(vma->obj) ||
+			    (ww && (dma_resv_locking_ctx(vma->obj->base.resv) == &ww->ctx))) {
+				__i915_vma_pin(vma);
+				list_add(&vma->evict_link, &locked_eviction_list);
+				continue;
+			}
+
+			if (!i915_gem_object_trylock(vma->obj, ww)) {
+				if (busy_bo) {
+					*busy_bo = vma->obj; /* holds ref */
+					ret = -EBUSY;
+					break;
+				}
+				i915_gem_object_put(vma->obj);
+				continue;
+			}
+
+			__i915_vma_pin(vma);
+			list_add(&vma->evict_link, &eviction_list);
+		}
+		if (list_empty(&eviction_list) && list_empty(&locked_eviction_list))
+			break;
+
+		/* Unbind locked objects first, before unlocking the eviction_list */
+		list_for_each_entry_safe(vma, vn, &locked_eviction_list, evict_link) {
+			__i915_vma_unpin(vma);
+
+			if (ret == 0) {
+				ret = __i915_vma_unbind(vma);
+				if (ret != -EINTR) /* "Get me out of here!" */
+					ret = 0;
+			}
+			if (!dying_vma(vma))
+				i915_gem_object_put(vma->obj);
+		}
+
+		list_for_each_entry_safe(vma, vn, &eviction_list, evict_link) {
+			__i915_vma_unpin(vma);
+			if (ret == 0) {
+				ret = __i915_vma_unbind(vma);
+				if (ret != -EINTR) /* "Get me out of here!" */
+					ret = 0;
+			}
+
+			i915_gem_object_unlock(vma->obj);
+			i915_gem_object_put(vma->obj);
+		}
+	} while (ret == 0);
+
+	return ret;
+}
+
+#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
+#include "selftests/i915_gem_evict.c"
+#endif