Re: Monolith: Facts, Failures, Fallacies, Falsehoods and Furphies
Posted: Dec 6, 2006 4:40 PM
> > I do not like these examples:
> > ::stlsoft::scoped_handle<int> h1(::open("file.ext"),
> > ::close);
> > ::stlsoft::scoped_handle<FILE*> h2(file, ::fclose);
> > because the resulting code will not check the return
> > of
> > can easily lead to unrecognized data loss.
> > Closing files is a bit problematic with RAII. Imagine
> > error occurs while the destructor is running. There is
> > return value, we can throw no exception, and there is
> > object left that we might check.
> This is a very good point, and one which is more involved
> than either of us want to get into here.
> I'm going to note this down for the book, and make sure to
> discuss the problems you've raised.
> For now, I'll just amend the examples in the doco to
> pretend like the issue doesn't exist. ;-) (I'll mention
> the issue in the docs as well, don't worry.)
> It's also worth noting that in many (most??) cases,
> closing a file in this way, i.e. without checking, is a
> reasonable tactic, and that something such as
> scoped_handle is a very helpful tool in that regard.
> Clearly, what's not reasonable is to not mention that it's
> not always reasonable: sometimes RAII is not the
> appropriate approach.
I have thought about this a little more, and I found the following: there are actually two different resources involved, the file handle and the file content. It is the the design of
fclose() that is problematic, because it mixes two different operations: buffer flushing and file handle deallocation. If we perform fsync()/fflush()/shutdown() so that the the only task that is left is releasing the file handle, the problem disappears, and
scoped_handle can be used without regret.
So maybe we can extrapolate from this case and get the following rule of thumb:
Think twice about using RAII if the deallocation function perfoms more than one task at once.
Finally, if data consistency is really important, using a database may be simpler :-)