btrfs: actively run the delayed refs while deleting large files

When we are deleting large files with large extents, we are building up
a huge set of delayed refs for processing.  Truncate isn't checking
often enough to see if we need to back off and process those, or let
a commit proceed.

The end result is long stalls after the rm, and very long commit times.
During the commits, other processes back up waiting to start new
transactions and we get into trouble.

Signed-off-by: Chris Mason <clm@fb.com>

4 files changed, 52 insertions, 5 deletions

diff --git a/fs/btrfs/free-space-cache.c b/fs/btrfs/free-space-cache.c
index 764528a4f6fd..c51482031edf 100644
--- a/fs/btrfs/free-space-cache.c
+++ b/fs/btrfs/free-space-cache.c
@@ -235,6 +235,8 @@ int btrfs_truncate_free_space_cache(struct btrfs_root *root,
 	/*
 	 * We don't need an orphan item because truncating the free space cache
 	 * will never be split across transactions.
+	 * We don't need to check for -EAGAIN because we're a free space
+	 * cache inode
 	 */
 	ret = btrfs_truncate_inode_items(trans, root, inode,
 					 0, BTRFS_EXTENT_DATA_KEY);
diff --git a/fs/btrfs/inode.c b/fs/btrfs/inode.c
index 97bc1ffc9c7b..e3fe137fb826 100644
--- a/fs/btrfs/inode.c
+++ b/fs/btrfs/inode.c
@@ -4197,10 +4197,20 @@ int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
 	int extent_type = -1;
 	int ret;
 	int err = 0;
+	int be_nice = 0;
 	u64 ino = btrfs_ino(inode);
+	u64 bytes_deleted = 0;
 
 	BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
 
+	/*
+	 * for non-free space inodes and ref cows, we want to back off from
+	 * time to time
+	 */
+	if (!btrfs_is_free_space_inode(inode) &&
+	    test_bit(BTRFS_ROOT_REF_COWS, &root->state))
+		be_nice = 1;
+
 	path = btrfs_alloc_path();
 	if (!path)
 		return -ENOMEM;
@@ -4230,6 +4240,19 @@ int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
 	key.type = (u8)-1;
 
 search_again:
+	/*
+	 * with a 16K leaf size and 128MB extents, you can actually queue
+	 * up a huge file in a single leaf.  Most of the time that
+	 * bytes_deleted is > 0, it will be huge by the time we get here
+	 */
+	if (be_nice && bytes_deleted > 32 * 1024 * 1024) {
+		if (btrfs_should_end_transaction(trans, root)) {
+			err = -EAGAIN;
+			goto error;
+		}
+	}
+
+
 	path->leave_spinning = 1;
 	ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 	if (ret < 0) {
@@ -4376,11 +4399,18 @@ delete:
 		    (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
 		     root == root->fs_info->tree_root)) {
 			btrfs_set_path_blocking(path);
+			bytes_deleted += extent_num_bytes;
 			ret = btrfs_free_extent(trans, root, extent_start,
 						extent_num_bytes, 0,
 						btrfs_header_owner(leaf),
 						ino, extent_offset, 0);
 			BUG_ON(ret);
+			if (be_nice && pending_del_nr &&
+			    (pending_del_nr % 16 == 0) &&
+			    bytes_deleted > 1024 * 1024) {
+				btrfs_async_run_delayed_refs(root,
+					trans->delayed_ref_updates * 2, 0);
+			}
 		}
 
 		if (found_type == BTRFS_INODE_ITEM_KEY)
@@ -4416,7 +4446,18 @@ error:
 	if (last_size != (u64)-1 &&
 	    root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
 		btrfs_ordered_update_i_size(inode, last_size, NULL);
+
 	btrfs_free_path(path);
+
+	if (be_nice && bytes_deleted > 32 * 1024 * 1024) {
+		unsigned long updates = trans->delayed_ref_updates;
+		if (updates) {
+			trans->delayed_ref_updates = 0;
+			ret = btrfs_run_delayed_refs(trans, root, updates * 2);
+			if (ret && !err)
+				err = ret;
+		}
+	}
 	return err;
 }
 
@@ -5013,7 +5054,7 @@ void btrfs_evict_inode(struct inode *inode)
 		trans->block_rsv = rsv;
 
 		ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
-		if (ret != -ENOSPC)
+		if (ret != -ENOSPC && ret != -EAGAIN)
 			break;
 
 		trans->block_rsv = &root->fs_info->trans_block_rsv;
@@ -8582,7 +8623,7 @@ static int btrfs_truncate(struct inode *inode)
 		ret = btrfs_truncate_inode_items(trans, root, inode,
 						 inode->i_size,
 						 BTRFS_EXTENT_DATA_KEY);
-		if (ret != -ENOSPC) {
+		if (ret != -ENOSPC && ret != -EAGAIN) {
 			err = ret;
 			break;
 		}
diff --git a/fs/btrfs/transaction.c b/fs/btrfs/transaction.c
index 91c303ac40b6..ba831ee41891 100644
--- a/fs/btrfs/transaction.c
+++ b/fs/btrfs/transaction.c
@@ -718,7 +718,7 @@ int btrfs_should_end_transaction(struct btrfs_trans_handle *trans,
 	updates = trans->delayed_ref_updates;
 	trans->delayed_ref_updates = 0;
 	if (updates) {
-		err = btrfs_run_delayed_refs(trans, root, updates);
+		err = btrfs_run_delayed_refs(trans, root, updates * 2);
 		if (err) /* Error code will also eval true */
 			return err;
 	}
diff --git a/fs/btrfs/tree-log.c b/fs/btrfs/tree-log.c
index 016c90fc85db..a089b5944efc 100644
--- a/fs/btrfs/tree-log.c
+++ b/fs/btrfs/tree-log.c
@@ -4251,8 +4251,12 @@ static int btrfs_log_inode(struct btrfs_trans_handle *trans,
 					  &BTRFS_I(inode)->runtime_flags);
 				clear_bit(BTRFS_INODE_COPY_EVERYTHING,
 					  &BTRFS_I(inode)->runtime_flags);
-				ret = btrfs_truncate_inode_items(trans, log,
-								 inode, 0, 0);
+				while(1) {
+					ret = btrfs_truncate_inode_items(trans,
+							 log, inode, 0, 0);
+					if (ret != -EAGAIN)
+						break;
+				}
 			}
 		} else if (test_and_clear_bit(BTRFS_INODE_COPY_EVERYTHING,
 					      &BTRFS_I(inode)->runtime_flags) ||