1 files changed, 254 insertions, 139 deletions

diff --git a/drivers/gpu/drm/xe/xe_bo_evict.c b/drivers/gpu/drm/xe/xe_bo_evict.c
index 6a40eedd9db1..ed3746d32b27 100644
--- a/drivers/gpu/drm/xe/xe_bo_evict.c
+++ b/drivers/gpu/drm/xe/xe_bo_evict.c
@@ -10,28 +10,103 @@
 #include "xe_ggtt.h"
 #include "xe_tile.h"
 
+typedef int (*xe_pinned_fn)(struct xe_bo *bo);
+
+static int xe_bo_apply_to_pinned(struct xe_device *xe,
+				 struct list_head *pinned_list,
+				 struct list_head *new_list,
+				 const xe_pinned_fn pinned_fn)
+{
+	LIST_HEAD(still_in_list);
+	struct xe_bo *bo;
+	int ret = 0;
+
+	spin_lock(&xe->pinned.lock);
+	while (!ret) {
+		bo = list_first_entry_or_null(pinned_list, typeof(*bo),
+					      pinned_link);
+		if (!bo)
+			break;
+		xe_bo_get(bo);
+		list_move_tail(&bo->pinned_link, &still_in_list);
+		spin_unlock(&xe->pinned.lock);
+
+		ret = pinned_fn(bo);
+		if (ret && pinned_list != new_list) {
+			spin_lock(&xe->pinned.lock);
+			/*
+			 * We might no longer be pinned, since PM notifier can
+			 * call this. If the pinned link is now empty, keep it
+			 * that way.
+			 */
+			if (!list_empty(&bo->pinned_link))
+				list_move(&bo->pinned_link, pinned_list);
+			spin_unlock(&xe->pinned.lock);
+		}
+		xe_bo_put(bo);
+		spin_lock(&xe->pinned.lock);
+	}
+	list_splice_tail(&still_in_list, new_list);
+	spin_unlock(&xe->pinned.lock);
+
+	return ret;
+}
+
 /**
- * xe_bo_evict_all - evict all BOs from VRAM
+ * xe_bo_notifier_prepare_all_pinned() - Pre-allocate the backing pages for all
+ * pinned VRAM objects which need to be saved.
+ * @xe: xe device
+ *
+ * Should be called from PM notifier when preparing for s3/s4.
  *
+ * Return: 0 on success, negative error code on error.
+ */
+int xe_bo_notifier_prepare_all_pinned(struct xe_device *xe)
+{
+	int ret;
+
+	ret = xe_bo_apply_to_pinned(xe, &xe->pinned.early.kernel_bo_present,
+				    &xe->pinned.early.kernel_bo_present,
+				    xe_bo_notifier_prepare_pinned);
+	if (!ret)
+		ret = xe_bo_apply_to_pinned(xe, &xe->pinned.late.kernel_bo_present,
+					    &xe->pinned.late.kernel_bo_present,
+					    xe_bo_notifier_prepare_pinned);
+
+	return ret;
+}
+
+/**
+ * xe_bo_notifier_unprepare_all_pinned() - Remove the backing pages for all
+ * pinned VRAM objects which have been restored.
  * @xe: xe device
  *
- * Evict non-pinned user BOs first (via GPU), evict pinned external BOs next
- * (via GPU), wait for evictions, and finally evict pinned kernel BOs via CPU.
- * All eviction magic done via TTM calls.
+ * Should be called from PM notifier after exiting s3/s4 (either on success or
+ * failure).
+ */
+void xe_bo_notifier_unprepare_all_pinned(struct xe_device *xe)
+{
+	(void)xe_bo_apply_to_pinned(xe, &xe->pinned.early.kernel_bo_present,
+				    &xe->pinned.early.kernel_bo_present,
+				    xe_bo_notifier_unprepare_pinned);
+
+	(void)xe_bo_apply_to_pinned(xe, &xe->pinned.late.kernel_bo_present,
+				    &xe->pinned.late.kernel_bo_present,
+				    xe_bo_notifier_unprepare_pinned);
+}
+
+/**
+ * xe_bo_evict_all_user - evict all non-pinned user BOs from VRAM
+ * @xe: xe device
  *
- * Evict == move VRAM BOs to temporary (typically system) memory.
+ * Evict non-pinned user BOs (via GPU).
  *
- * This function should be called before the device goes into a suspend state
- * where the VRAM loses power.
+ * Evict == move VRAM BOs to temporary (typically system) memory.
  */
-int xe_bo_evict_all(struct xe_device *xe)
+int xe_bo_evict_all_user(struct xe_device *xe)
 {
 	struct ttm_device *bdev = &xe->ttm;
-	struct xe_bo *bo;
-	struct xe_tile *tile;
-	struct list_head still_in_list;
 	u32 mem_type;
-	u8 id;
 	int ret;
 
 	/* User memory */
@@ -57,34 +132,38 @@ int xe_bo_evict_all(struct xe_device *xe)
 		}
 	}
 
-	/* Pinned user memory in VRAM */
-	INIT_LIST_HEAD(&still_in_list);
-	spin_lock(&xe->pinned.lock);
-	for (;;) {
-		bo = list_first_entry_or_null(&xe->pinned.external_vram,
-					      typeof(*bo), pinned_link);
-		if (!bo)
-			break;
-		xe_bo_get(bo);
-		list_move_tail(&bo->pinned_link, &still_in_list);
-		spin_unlock(&xe->pinned.lock);
+	return 0;
+}
 
-		xe_bo_lock(bo, false);
-		ret = xe_bo_evict_pinned(bo);
-		xe_bo_unlock(bo);
-		xe_bo_put(bo);
-		if (ret) {
-			spin_lock(&xe->pinned.lock);
-			list_splice_tail(&still_in_list,
-					 &xe->pinned.external_vram);
-			spin_unlock(&xe->pinned.lock);
-			return ret;
-		}
+/**
+ * xe_bo_evict_all - evict all BOs from VRAM
+ * @xe: xe device
+ *
+ * Evict non-pinned user BOs first (via GPU), evict pinned external BOs next
+ * (via GPU), wait for evictions, and finally evict pinned kernel BOs via CPU.
+ * All eviction magic done via TTM calls.
+ *
+ * Evict == move VRAM BOs to temporary (typically system) memory.
+ *
+ * This function should be called before the device goes into a suspend state
+ * where the VRAM loses power.
+ */
+int xe_bo_evict_all(struct xe_device *xe)
+{
+	struct xe_tile *tile;
+	u8 id;
+	int ret;
 
-		spin_lock(&xe->pinned.lock);
-	}
-	list_splice_tail(&still_in_list, &xe->pinned.external_vram);
-	spin_unlock(&xe->pinned.lock);
+	ret = xe_bo_evict_all_user(xe);
+	if (ret)
+		return ret;
+
+	ret = xe_bo_apply_to_pinned(xe, &xe->pinned.late.kernel_bo_present,
+				    &xe->pinned.late.evicted, xe_bo_evict_pinned);
+
+	if (!ret)
+		ret = xe_bo_apply_to_pinned(xe, &xe->pinned.late.kernel_bo_present,
+					    &xe->pinned.late.evicted, xe_bo_evict_pinned);
 
 	/*
 	 * Wait for all user BO to be evicted as those evictions depend on the
@@ -93,32 +172,49 @@ int xe_bo_evict_all(struct xe_device *xe)
 	for_each_tile(tile, xe, id)
 		xe_tile_migrate_wait(tile);
 
-	spin_lock(&xe->pinned.lock);
-	for (;;) {
-		bo = list_first_entry_or_null(&xe->pinned.kernel_bo_present,
-					      typeof(*bo), pinned_link);
-		if (!bo)
-			break;
-		xe_bo_get(bo);
-		list_move_tail(&bo->pinned_link, &xe->pinned.evicted);
-		spin_unlock(&xe->pinned.lock);
+	if (ret)
+		return ret;
 
-		xe_bo_lock(bo, false);
-		ret = xe_bo_evict_pinned(bo);
-		xe_bo_unlock(bo);
-		xe_bo_put(bo);
-		if (ret)
-			return ret;
+	return xe_bo_apply_to_pinned(xe, &xe->pinned.early.kernel_bo_present,
+				     &xe->pinned.early.evicted,
+				     xe_bo_evict_pinned);
+}
 
-		spin_lock(&xe->pinned.lock);
+static int xe_bo_restore_and_map_ggtt(struct xe_bo *bo)
+{
+	struct xe_device *xe = xe_bo_device(bo);
+	int ret;
+
+	ret = xe_bo_restore_pinned(bo);
+	if (ret)
+		return ret;
+
+	if (bo->flags & XE_BO_FLAG_GGTT) {
+		struct xe_tile *tile;
+		u8 id;
+
+		for_each_tile(tile, xe_bo_device(bo), id) {
+			if (tile != bo->tile && !(bo->flags & XE_BO_FLAG_GGTTx(tile)))
+				continue;
+
+			mutex_lock(&tile->mem.ggtt->lock);
+			xe_ggtt_map_bo(tile->mem.ggtt, bo);
+			mutex_unlock(&tile->mem.ggtt->lock);
+		}
 	}
-	spin_unlock(&xe->pinned.lock);
+
+	/*
+	 * We expect validate to trigger a move VRAM and our move code
+	 * should setup the iosys map.
+	 */
+	xe_assert(xe, !(bo->flags & XE_BO_FLAG_PINNED_LATE_RESTORE) ||
+		  !iosys_map_is_null(&bo->vmap));
 
 	return 0;
 }
 
 /**
- * xe_bo_restore_kernel - restore kernel BOs to VRAM
+ * xe_bo_restore_early - restore early phase kernel BOs to VRAM
  *
  * @xe: xe device
  *
@@ -128,111 +224,130 @@ int xe_bo_evict_all(struct xe_device *xe)
  * This function should be called early, before trying to init the GT, on device
  * resume.
  */
-int xe_bo_restore_kernel(struct xe_device *xe)
+int xe_bo_restore_early(struct xe_device *xe)
 {
-	struct xe_bo *bo;
-	int ret;
+	return xe_bo_apply_to_pinned(xe, &xe->pinned.early.evicted,
+				     &xe->pinned.early.kernel_bo_present,
+				     xe_bo_restore_and_map_ggtt);
+}
 
-	spin_lock(&xe->pinned.lock);
-	for (;;) {
-		bo = list_first_entry_or_null(&xe->pinned.evicted,
-					      typeof(*bo), pinned_link);
-		if (!bo)
-			break;
-		xe_bo_get(bo);
-		list_move_tail(&bo->pinned_link, &xe->pinned.kernel_bo_present);
-		spin_unlock(&xe->pinned.lock);
+/**
+ * xe_bo_restore_late - restore pinned late phase BOs
+ *
+ * @xe: xe device
+ *
+ * Move pinned user and kernel BOs which can use blitter from temporary
+ * (typically system) memory to VRAM. All moves done via TTM calls.
+ *
+ * This function should be called late, after GT init, on device resume.
+ */
+int xe_bo_restore_late(struct xe_device *xe)
+{
+	struct xe_tile *tile;
+	int ret, id;
 
-		xe_bo_lock(bo, false);
-		ret = xe_bo_restore_pinned(bo);
-		xe_bo_unlock(bo);
-		if (ret) {
-			xe_bo_put(bo);
-			return ret;
-		}
+	ret = xe_bo_apply_to_pinned(xe, &xe->pinned.late.evicted,
+				    &xe->pinned.late.kernel_bo_present,
+				    xe_bo_restore_and_map_ggtt);
 
-		if (bo->flags & XE_BO_FLAG_GGTT) {
-			struct xe_tile *tile;
-			u8 id;
+	for_each_tile(tile, xe, id)
+		xe_tile_migrate_wait(tile);
 
-			for_each_tile(tile, xe, id) {
-				if (tile != bo->tile && !(bo->flags & XE_BO_FLAG_GGTTx(tile)))
-					continue;
+	if (ret)
+		return ret;
 
-				mutex_lock(&tile->mem.ggtt->lock);
-				xe_ggtt_map_bo(tile->mem.ggtt, bo);
-				mutex_unlock(&tile->mem.ggtt->lock);
-			}
-		}
+	if (!IS_DGFX(xe))
+		return 0;
 
-		/*
-		 * We expect validate to trigger a move VRAM and our move code
-		 * should setup the iosys map.
-		 */
-		xe_assert(xe, !iosys_map_is_null(&bo->vmap));
+	/* Pinned user memory in VRAM should be validated on resume */
+	ret = xe_bo_apply_to_pinned(xe, &xe->pinned.late.external,
+				    &xe->pinned.late.external,
+				    xe_bo_restore_pinned);
 
-		xe_bo_put(bo);
+	/* Wait for restore to complete */
+	for_each_tile(tile, xe, id)
+		xe_tile_migrate_wait(tile);
 
-		spin_lock(&xe->pinned.lock);
-	}
-	spin_unlock(&xe->pinned.lock);
+	return ret;
+}
 
-	return 0;
+static void xe_bo_pci_dev_remove_pinned(struct xe_device *xe)
+{
+	struct xe_tile *tile;
+	unsigned int id;
+
+	(void)xe_bo_apply_to_pinned(xe, &xe->pinned.late.external,
+				    &xe->pinned.late.external,
+				    xe_bo_dma_unmap_pinned);
+	for_each_tile(tile, xe, id)
+		xe_tile_migrate_wait(tile);
 }
 
 /**
- * xe_bo_restore_user - restore pinned user BOs to VRAM
- *
- * @xe: xe device
+ * xe_bo_pci_dev_remove_all() - Handle bos when the pci_device is about to be removed
+ * @xe: The xe device.
  *
- * Move pinned user BOs from temporary (typically system) memory to VRAM via
- * CPU. All moves done via TTM calls.
+ * On pci_device removal we need to drop all dma mappings and move
+ * the data of exported bos out to system. This includes SVM bos and
+ * exported dma-buf bos. This is done by evicting all bos, but
+ * the evict placement in xe_evict_flags() is chosen such that all
+ * bos except those mentioned are purged, and thus their memory
+ * is released.
  *
- * This function should be called late, after GT init, on device resume.
+ * For pinned bos, we're unmapping dma.
  */
-int xe_bo_restore_user(struct xe_device *xe)
+void xe_bo_pci_dev_remove_all(struct xe_device *xe)
 {
-	struct xe_bo *bo;
-	struct xe_tile *tile;
-	struct list_head still_in_list;
-	u8 id;
-	int ret;
+	unsigned int mem_type;
 
-	if (!IS_DGFX(xe))
-		return 0;
+	/*
+	 * Move pagemap bos and exported dma-buf to system, and
+	 * purge everything else.
+	 */
+	for (mem_type = XE_PL_VRAM1; mem_type >= XE_PL_TT; --mem_type) {
+		struct ttm_resource_manager *man =
+			ttm_manager_type(&xe->ttm, mem_type);
 
-	/* Pinned user memory in VRAM should be validated on resume */
-	INIT_LIST_HEAD(&still_in_list);
-	spin_lock(&xe->pinned.lock);
-	for (;;) {
-		bo = list_first_entry_or_null(&xe->pinned.external_vram,
-					      typeof(*bo), pinned_link);
-		if (!bo)
-			break;
-		list_move_tail(&bo->pinned_link, &still_in_list);
-		xe_bo_get(bo);
-		spin_unlock(&xe->pinned.lock);
+		if (man) {
+			int ret = ttm_resource_manager_evict_all(&xe->ttm, man);
 
-		xe_bo_lock(bo, false);
-		ret = xe_bo_restore_pinned(bo);
-		xe_bo_unlock(bo);
-		xe_bo_put(bo);
-		if (ret) {
-			spin_lock(&xe->pinned.lock);
-			list_splice_tail(&still_in_list,
-					 &xe->pinned.external_vram);
-			spin_unlock(&xe->pinned.lock);
-			return ret;
+			drm_WARN_ON(&xe->drm, ret);
 		}
-
-		spin_lock(&xe->pinned.lock);
 	}
-	list_splice_tail(&still_in_list, &xe->pinned.external_vram);
-	spin_unlock(&xe->pinned.lock);
 
-	/* Wait for restore to complete */
-	for_each_tile(tile, xe, id)
-		xe_tile_migrate_wait(tile);
+	xe_bo_pci_dev_remove_pinned(xe);
+}
 
-	return 0;
+static void xe_bo_pinned_fini(void *arg)
+{
+	struct xe_device *xe = arg;
+
+	(void)xe_bo_apply_to_pinned(xe, &xe->pinned.late.kernel_bo_present,
+				    &xe->pinned.late.kernel_bo_present,
+				    xe_bo_dma_unmap_pinned);
+	(void)xe_bo_apply_to_pinned(xe, &xe->pinned.early.kernel_bo_present,
+				    &xe->pinned.early.kernel_bo_present,
+				    xe_bo_dma_unmap_pinned);
+}
+
+/**
+ * xe_bo_pinned_init() - Initialize pinned bo tracking
+ * @xe: The xe device.
+ *
+ * Initializes the lists and locks required for pinned bo
+ * tracking and registers a callback to dma-unmap
+ * any remaining pinned bos on pci device removal.
+ *
+ * Return: %0 on success, negative error code on error.
+ */
+int xe_bo_pinned_init(struct xe_device *xe)
+{
+	spin_lock_init(&xe->pinned.lock);
+	INIT_LIST_HEAD(&xe->pinned.early.kernel_bo_present);
+	INIT_LIST_HEAD(&xe->pinned.early.evicted);
+	INIT_LIST_HEAD(&xe->pinned.late.kernel_bo_present);
+	INIT_LIST_HEAD(&xe->pinned.late.evicted);
+	INIT_LIST_HEAD(&xe->pinned.late.external);
+
+	return devm_add_action_or_reset(xe->drm.dev, xe_bo_pinned_fini, xe);
 }