From: Jonathan Turkanis (technews_at_[hidden])
Date: 2004-03-16 20:05:17
"Jeremy Maitin-Shepard" <jbms_at_[hidden]> wrote in message
news:878yi09w29.fsf_at_jbms.ath.cx...
> Making it handle symbolic links properly is certainly not trivial,
and
> cannot be done efficiently with the current version of the
filesystem
> library (assuming no additional platform-specific code is added).
>
> As has been discussed already in relation to the implementation of
the
> `equivalent' function which Beman is adding, the common
implementation
> is to store the inode numbers of every directory with a link count
> greater than one in a hash set, and then make sure that they are
> traversed only once. This implementation is not perfect, however,
> because most platforms, including Windows and Linux, do not
guarantee
> that inode numbers are unique for files for which there is not an
open
> handle. Thus, the way to remedy this would be to keep an open
handle to
> each directory whose inode number is being stored; but this creates
the
> problem of the traversal being limited by the number of open file
> handles/descriptors a single process can have.
I understand the reluctance to add a feature which raises such thorny
issues. However, I think the fact that recursive directory iteration
is hard to implement properly is an excellent reason to put it in a
library. I would imagine that a lot of existing code doesn't handle
all these issues correctly. Limitations of the implementation can be
explained in the documentation.
Jonathan