1 files changed, 37 insertions, 3 deletions

diff --git a/fs/btrfs/tree-log.c b/fs/btrfs/tree-log.c
index 3fc8d854d7fb..6c8297bcfeb7 100644
--- a/fs/btrfs/tree-log.c
+++ b/fs/btrfs/tree-log.c
@@ -3323,6 +3323,30 @@ int btrfs_free_log_root_tree(struct btrfs_trans_handle *trans,
 }
 
 /*
+ * Check if an inode was logged in the current transaction. We can't always rely
+ * on an inode's logged_trans value, because it's an in-memory only field and
+ * therefore not persisted. This means that its value is lost if the inode gets
+ * evicted and loaded again from disk (in which case it has a value of 0, and
+ * certainly it is smaller then any possible transaction ID), when that happens
+ * the full_sync flag is set in the inode's runtime flags, so on that case we
+ * assume eviction happened and ignore the logged_trans value, assuming the
+ * worst case, that the inode was logged before in the current transaction.
+ */
+static bool inode_logged(struct btrfs_trans_handle *trans,
+                         struct btrfs_inode *inode)
+{
+        if (inode->logged_trans == trans->transid)
+                return true;
+
+        if (inode->last_trans == trans->transid &&
+            test_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags) &&
+            !test_bit(BTRFS_FS_LOG_RECOVERING, &trans->fs_info->flags))
+                return true;
+
+        return false;
+}
+
+/*
  * If both a file and directory are logged, and unlinks or renames are
  * mixed in, we have a few interesting corners:
  *
@@ -3356,7 +3380,7 @@ int btrfs_del_dir_entries_in_log(struct btrfs_trans_handle *trans,
         int bytes_del = 0;
         u64 dir_ino = btrfs_ino(dir);
 
-        if (dir->logged_trans < trans->transid)
+        if (!inode_logged(trans, dir))
                 return 0;
 
         ret = join_running_log_trans(root);
@@ -3460,7 +3484,7 @@ int btrfs_del_inode_ref_in_log(struct btrfs_trans_handle *trans,
         u64 index;
         int ret;
 
-        if (inode->logged_trans < trans->transid)
+        if (!inode_logged(trans, inode))
                 return 0;
 
         ret = join_running_log_trans(root);
@@ -5420,9 +5444,19 @@ log_extents:
                 }
         }
 
+        /*
+         * Don't update last_log_commit if we logged that an inode exists after
+         * it was loaded to memory (full_sync bit set).
+         * This is to prevent data loss when we do a write to the inode, then
+         * the inode gets evicted after all delalloc was flushed, then we log
+         * it exists (due to a rename for example) and then fsync it. This last
+         * fsync would do nothing (not logging the extents previously written).
+         */
         spin_lock(&inode->lock);
         inode->logged_trans = trans->transid;
-        inode->last_log_commit = inode->last_sub_trans;
+        if (inode_only != LOG_INODE_EXISTS ||
+            !test_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags))
+                inode->last_log_commit = inode->last_sub_trans;
         spin_unlock(&inode->lock);
 out_unlock:
         mutex_unlock(&inode->log_mutex);