Re: [boost] Re: [filesystem] create_directories not thread-safe?

From: Beman Dawes (bdawes_at_[hidden])
Date: 2004-03-09 11:36:52

At 01:55 AM 3/9/2004, Vladimir Prus wrote:

 >After more though, I wonder if create_directories should not swallow
 >exceptions itself. E.g.
 >
 >
 >   try {
 >        create_directory(ph);
 >   }
 >   catch(const filesystem_error& e) {
 >         if (e.error() == filesystem::already_exists_error)
 >               ; // do nothing
 >         else
 >               throw;
 >   }

Note that the actual implementation would not use a try/catch, but would 
have the same effect as your code.

 >After all, caller only cares if directory exists, if it was created by
 >another thread it's ok.

That's an interesting thought. Of course, it might not be another thread 
but rather another process that created the directory, and that could be a 
sign of real trouble.

But such a change would allow simpler user code. What is now written:

    if ( !exists( foo ) ) create_directory( foo );

becomes:

    create_directory( foo );

What happens when the directory existing really does constitute and error, 
and create_directory() doesn't throw? Will applications fail in difficult 
to diagnose ways? Can applications protect themselves against silent 
failure?

It seems to me some applications will expect directory foo to be empty, and 
might otherwise fail, but that this isn't a commonplace problem. 
Applications can detect that a directory is empty, although because of 
possible race conditions such a test could be misleading. But that's 
nothing new.

Overall, it seems the change would be reasonably safe. What do others 
think?

What happens if the path exists but is not a directory? Presumably that 
case must still throw.

Thanks,

--Beman