+/*
+ * this is used while deleting old snapshots, and it drops the refs
+ * on a whole subtree starting from a level 1 node.
+ *
+ * The idea is to sort all the leaf pointers, and then drop the
+ * ref on all the leaves in order. Most of the time the leaves
+ * will have ref cache entries, so no leaf IOs will be required to
+ * find the extents they have references on.
+ *
+ * For each leaf, any references it has are also dropped in order
+ *
+ * This ends up dropping the references in something close to optimal
+ * order for reading and modifying the extent allocation tree.
+ */
+static noinline int drop_level_one_refs(struct btrfs_trans_handle *trans,
+ struct btrfs_root *root,
+ struct btrfs_path *path)
+{
+ u64 bytenr;
+ u64 root_owner;
+ u64 root_gen;
+ struct extent_buffer *eb = path->nodes[1];
+ struct extent_buffer *leaf;
+ struct btrfs_leaf_ref *ref;
+ struct refsort *sorted = NULL;
+ int nritems = btrfs_header_nritems(eb);
+ int ret;
+ int i;
+ int refi = 0;
+ int slot = path->slots[1];
+ u32 blocksize = btrfs_level_size(root, 0);
+ u32 refs;
+
+ if (nritems == 0)
+ goto out;
+
+ root_owner = btrfs_header_owner(eb);
+ root_gen = btrfs_header_generation(eb);
+ sorted = kmalloc(sizeof(*sorted) * nritems, GFP_NOFS);
+
+ /*
+ * step one, sort all the leaf pointers so we don't scribble
+ * randomly into the extent allocation tree
+ */
+ for (i = slot; i < nritems; i++) {
+ sorted[refi].bytenr = btrfs_node_blockptr(eb, i);
+ sorted[refi].slot = i;
+ refi++;
+ }
+
+ /*
+ * nritems won't be zero, but if we're picking up drop_snapshot
+ * after a crash, slot might be > 0, so double check things
+ * just in case.
+ */
+ if (refi == 0)
+ goto out;
+
+ sort(sorted, refi, sizeof(struct refsort), refsort_cmp, NULL);
+
+ /*
+ * the first loop frees everything the leaves point to
+ */
+ for (i = 0; i < refi; i++) {
+ u64 ptr_gen;
+
+ bytenr = sorted[i].bytenr;
+
+ /*
+ * check the reference count on this leaf. If it is > 1
+ * we just decrement it below and don't update any
+ * of the refs the leaf points to.
+ */
+ ret = drop_snap_lookup_refcount(root, bytenr, blocksize, &refs);
+ BUG_ON(ret);
+ if (refs != 1)
+ continue;
+
+ ptr_gen = btrfs_node_ptr_generation(eb, sorted[i].slot);
+
+ /*
+ * the leaf only had one reference, which means the
+ * only thing pointing to this leaf is the snapshot
+ * we're deleting. It isn't possible for the reference
+ * count to increase again later
+ *
+ * The reference cache is checked for the leaf,
+ * and if found we'll be able to drop any refs held by
+ * the leaf without needing to read it in.
+ */
+ ref = btrfs_lookup_leaf_ref(root, bytenr);
+ if (ref && ref->generation != ptr_gen) {
+ btrfs_free_leaf_ref(root, ref);
+ ref = NULL;
+ }
+ if (ref) {
+ ret = cache_drop_leaf_ref(trans, root, ref);
+ BUG_ON(ret);
+ btrfs_remove_leaf_ref(root, ref);
+ btrfs_free_leaf_ref(root, ref);
+ } else {
+ /*
+ * the leaf wasn't in the reference cache, so
+ * we have to read it.
+ */
+ leaf = read_tree_block(root, bytenr, blocksize,
+ ptr_gen);
+ ret = btrfs_drop_leaf_ref(trans, root, leaf);
+ BUG_ON(ret);
+ free_extent_buffer(leaf);
+ }
+ atomic_inc(&root->fs_info->throttle_gen);
+ wake_up(&root->fs_info->transaction_throttle);
+ cond_resched();
+ }
+
+ /*
+ * run through the loop again to free the refs on the leaves.
+ * This is faster than doing it in the loop above because
+ * the leaves are likely to be clustered together. We end up
+ * working in nice chunks on the extent allocation tree.
+ */
+ for (i = 0; i < refi; i++) {
+ bytenr = sorted[i].bytenr;
+ ret = __btrfs_free_extent(trans, root, bytenr,
+ blocksize, eb->start,
+ root_owner, root_gen, 0, 1);
+ BUG_ON(ret);
+
+ atomic_inc(&root->fs_info->throttle_gen);
+ wake_up(&root->fs_info->transaction_throttle);
+ cond_resched();
+ }
+out:
+ kfree(sorted);
+
+ /*
+ * update the path to show we've processed the entire level 1
+ * node. This will get saved into the root's drop_snapshot_progress
+ * field so these drops are not repeated again if this transaction
+ * commits.
+ */
+ path->slots[1] = nritems;
+ return 0;
+}
+