Merge branch 'iommufd/for-rc' into for-next

The following selftest patch requires both the bug fixes and the
improvements of the selftest framework.

* iommufd/for-rc:
  iommufd: Do not corrupt the pfn list when doing batch carry
  iommufd: Fix unpinning of pages when an access is present
  iommufd: Check for uptr overflow
  Linux 6.3-rc5

Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>

1 files changed, 63 insertions, 22 deletions

diff --git a/fs/btrfs/backref.c b/fs/btrfs/backref.c
index 90e40d5ceccd..e54f0884802a 100644
--- a/fs/btrfs/backref.c
+++ b/fs/btrfs/backref.c
@@ -1921,8 +1921,7 @@ int btrfs_is_data_extent_shared(struct btrfs_inode *inode, u64 bytenr,
 	level = -1;
 	ULIST_ITER_INIT(&uiter);
 	while (1) {
-		bool is_shared;
-		bool cached;
+		const unsigned long prev_ref_count = ctx->refs.nnodes;
 
 		walk_ctx.bytenr = bytenr;
 		ret = find_parent_nodes(&walk_ctx, &shared);
@@ -1940,21 +1939,36 @@ int btrfs_is_data_extent_shared(struct btrfs_inode *inode, u64 bytenr,
 		ret = 0;
 
 		/*
-		 * If our data extent was not directly shared (without multiple
-		 * reference items), than it might have a single reference item
-		 * with a count > 1 for the same offset, which means there are 2
-		 * (or more) file extent items that point to the data extent -
-		 * this happens when a file extent item needs to be split and
-		 * then one item gets moved to another leaf due to a b+tree leaf
-		 * split when inserting some item. In this case the file extent
-		 * items may be located in different leaves and therefore some
-		 * of the leaves may be referenced through shared subtrees while
-		 * others are not. Since our extent buffer cache only works for
-		 * a single path (by far the most common case and simpler to
-		 * deal with), we can not use it if we have multiple leaves
-		 * (which implies multiple paths).
+		 * More than one extent buffer (bytenr) may have been added to
+		 * the ctx->refs ulist, in which case we have to check multiple
+		 * tree paths in case the first one is not shared, so we can not
+		 * use the path cache which is made for a single path. Multiple
+		 * extent buffers at the current level happen when:
+		 *
+		 * 1) level -1, the data extent: If our data extent was not
+		 *    directly shared (without multiple reference items), then
+		 *    it might have a single reference item with a count > 1 for
+		 *    the same offset, which means there are 2 (or more) file
+		 *    extent items that point to the data extent - this happens
+		 *    when a file extent item needs to be split and then one
+		 *    item gets moved to another leaf due to a b+tree leaf split
+		 *    when inserting some item. In this case the file extent
+		 *    items may be located in different leaves and therefore
+		 *    some of the leaves may be referenced through shared
+		 *    subtrees while others are not. Since our extent buffer
+		 *    cache only works for a single path (by far the most common
+		 *    case and simpler to deal with), we can not use it if we
+		 *    have multiple leaves (which implies multiple paths).
+		 *
+		 * 2) level >= 0, a tree node/leaf: We can have a mix of direct
+		 *    and indirect references on a b+tree node/leaf, so we have
+		 *    to check multiple paths, and the extent buffer (the
+		 *    current bytenr) may be shared or not. One example is
+		 *    during relocation as we may get a shared tree block ref
+		 *    (direct ref) and a non-shared tree block ref (indirect
+		 *    ref) for the same node/leaf.
 		 */
-		if (level == -1 && ctx->refs.nnodes > 1)
+		if ((ctx->refs.nnodes - prev_ref_count) > 1)
 			ctx->use_path_cache = false;
 
 		if (level >= 0)
@@ -1964,12 +1978,17 @@ int btrfs_is_data_extent_shared(struct btrfs_inode *inode, u64 bytenr,
 		if (!node)
 			break;
 		bytenr = node->val;
-		level++;
-		cached = lookup_backref_shared_cache(ctx, root, bytenr, level,
-						     &is_shared);
-		if (cached) {
-			ret = (is_shared ? 1 : 0);
-			break;
+		if (ctx->use_path_cache) {
+			bool is_shared;
+			bool cached;
+
+			level++;
+			cached = lookup_backref_shared_cache(ctx, root, bytenr,
+							     level, &is_shared);
+			if (cached) {
+				ret = (is_shared ? 1 : 0);
+				break;
+			}
 		}
 		shared.share_count = 0;
 		shared.have_delayed_delete_refs = false;
@@ -1977,6 +1996,28 @@ int btrfs_is_data_extent_shared(struct btrfs_inode *inode, u64 bytenr,
 	}
 
 	/*
+	 * If the path cache is disabled, then it means at some tree level we
+	 * got multiple parents due to a mix of direct and indirect backrefs or
+	 * multiple leaves with file extent items pointing to the same data
+	 * extent. We have to invalidate the cache and cache only the sharedness
+	 * result for the levels where we got only one node/reference.
+	 */
+	if (!ctx->use_path_cache) {
+		int i = 0;
+
+		level--;
+		if (ret >= 0 && level >= 0) {
+			bytenr = ctx->path_cache_entries[level].bytenr;
+			ctx->use_path_cache = true;
+			store_backref_shared_cache(ctx, root, bytenr, level, ret);
+			i = level + 1;
+		}
+
+		for ( ; i < BTRFS_MAX_LEVEL; i++)
+			ctx->path_cache_entries[i].bytenr = 0;
+	}
+
+	/*
 	 * Cache the sharedness result for the data extent if we know our inode
 	 * has more than 1 file extent item that refers to the data extent.
 	 */