How to test code, that depends on file system errors?

Question

I am developing a lib for a Linux project that stores files using boost::serialization. Because the files might be very large (1-50 GiB) and the program might fail writing them at some point (we are talking about thousands of them) I want to use a simple backup system to recover the files if they are not valid.

I have to test this using QtTests (the Qt unit test suite), but I could add additional libs if neccessary. The question is now how to test the backup system. In theory the test needs to let the ofstream ofs add invalid data after a given amount of data.

Here is the source code of the function in question. I am already testing the saving process itself, only the backup system could only be tested by hand.

The code (glued together and shortened to one method):

void DataStore::storeNodeData(TrajectoryDataSet data) {
   string path = getDataFileName(data);
   string oldPath = getDataFileName(data).append(".old");

   if (access(path.c_str(), F_OK) == 0 &&
      access(oldPath.c_str(), F_OK) != 0) {
      if (rename(path.c_str(), oldPath.c_str()) != 0) {
         throw FileSystemException("Unable to rename file to '.old'.");
      }
   }

   bool broken = false;
   ofstream ofs(path);
   boost::archive::text_oarchive oa(ofs);
   oa << data;
   if (!ofs.good())
      broken = true;
   ofs.close();

   //File verification
   TrajectoryDataSet fData;
   ifstream ifs(path);
   boost::archive::text_iarchive ia(ifs);
   ia >> fData;
   ifs.close();
   if (fData != data)
      broken = true;

   if (broken) {
      rename(oldPath.c_str(), path.c_str());
      throw FileSystemException("Unable write valid file.");
   } else if (access(oldPath.c_str(), F_OK) != -1 &&
      unlink(oldPath.c_str()) != 0)
      throw FileSystemException("Unable to delete backup file.");
}

Answer 1

How would you test it manually? Then wonder can I automate this later?

One scenario is that the disk it is writing to is unmounted for a short while, this probably result in some sort of error:

• Thread the test and in parallel unmount the file system on which the data gets written.

This will result in a end-2-end test, needing a filesystem, data-files, setup time and probably is slow.

Now I can come-up with more scenarios, but I think you want to be able to inject a mock of ofs and let its good() method return false just for this test, afterwards verify your recovery system worked. E.g. decoupling and unit-testing.