Re: [PATCH] fs: Introducing Lanyard Filesystem

[Al Viro <viro@ZenIV.linux.org.uk>]

On Sun, Aug 19, 2012 at 06:53:37PM +0200, Dan Luedtke wrote:
> > 	* minor point, but endianness-flipping in place is *the* way to get
> > hard-to-catch endianness bugs.  foo = cpu_to_le64(foo) is a bloody bad idea;
> > either use object for host-endian all along, or use it only for (in your
> > case) little-endian.
> I am not sure I understood this right.
> At what point should I convert e.g. the file size (little endian 64bit
> value stored on disk) to host endianess? When filling the inode?
> Is inode->i_size = le64_to_cpu(size) bad, too?

Conversions *in* *place* are bad.  The above is fine if you have 'size'
variable used only for little-endian values (and declare it __le64, then).
The thing is, thinking of endianness conversions as architecture-conditional
byteswaps is wrong.  Treat them as you would treat dealing with some
encoding; that's the reason why we have separate le64_to_cpu() and
cpu_to_le64(), even though they are identical as functions - on all
architectures we support.  And don't use the same variable for both the
host- and fixed-endian values; it's far too easy to get confused on
code modifications and get the situation when two code paths lead to
the same point, one with host-endian value in that variable and another
with fixed-endian.  Real fun to debug, especially when one of the codepaths
is rarely taken...