Re: DogCows or Polymorphism in the Hurd

[Marcus Brinkmann]

At Mon, 06 Feb 2006 14:03:28 +0100,
address@hidden (Ludovic Courtès) wrote:
> On GNU/Linux with, say, an ext[23] filesystem, a directory does not
> implement the file interface, only the directory interface.  Thus, for
> instance, `cat DIR' returns `EISDIR'.
> 
> On some other systems (e.g., Tru64 with AdvFS IIRC), directories do
> implement both interfaces.  So there, one can `cat' the contents of a
> directory and it just shows its ``binary content''.
> 
> This difference doesn't preclude good behavior of `grep', `find' and
> friends.  This is because those programs call `stat ()' before actually
> accessing the node, and look at the value of `S_ISDIR ()' _before_
> looking at the value of `S_ISREG ()'.  In OO terms: they check for the
> most specialized type first.
> 
> IOW, the `st_mode' field can be viewed as a poor man's type system.  ;-)

Unfortunately, the actual behaviour of these utilities seems to be
much more complex than you describe.

For example, "rm -fR foo" first may try to simply unlink() foo,
without a stat.  Then, when it recurses, it may look at the directory
entry's "d_type" field.  As it happens, on the Hurd, d_type contains
the type of the _underlying_ node, not of the translated node, d'oh!

In the presence of hybrid or translated nodes, these tools can behave
in a difficult to predict way.

Thanks,
Marcus