[9fans] rewriting paths [was: mv vs cp]

[Richard Uhtenwoldt <greon@best.com>]

I was thinking about this in the bathtub just now.  One could interpose
a filesystem between the user and kfs --or wherever one's files are
really stored-- called renamefs, which rewrites paths before forwarding
them to kfs.  9P requests do not really contain paths, but you get the
idea: the new file server, renamefs, maintains a rewrite table
containing the paths specified by past move commands (2 paths per move
command to be precise).  This state is used to rewrite certain 9P
requests (eg, some walk requests) before relaying them to the 
server that really stores files.

To increase the usefulness of renamefs, we allow it to store 2 types of
records.  Specifically, a record can take the form

  D<path>

, which means that the string <path> "no longer lives here"
(ie, renamefs will act as if <path> does not exist), or it can
take the form

  A<path><path2>

, which means that <path2> will get rewritten as <path> and does not
carry the implication that <path> no longer lives here.  (Yes, yes, I
know: between <path> and <path2> you need some lexeme that does not
occur in paths.)

Then,

  Mv <path> <path2> 

can be implemented by adding two records, A<path><path2> and D<path>, and
something resembling Unix's

  Ln -s <path> <path2>

is implementable by adding one record, A<path><path2>, to the rewrite
table inside renamefs.  The main socioeconomic purpose of these features
is to make Unix users feel more at home by providing means to move
directories, means to do inter-directory moves and symlinks.  Moreover,

  Del <path>

which just adds D<path> to the rewrite table, is a delete that supports
undelete (via removal of the D record).  And resource forks can be
implemented via ... --Just kidding.

No need to modify 9P of course: just have renamefs listen on a specific
file like /foo/rename where foo might be mnt (sorry for my unfamiliarity
with Plan9 file naming conventions) and have mv write requests
(candidates already in the form of strings for entry into the rewrite
table) to /foo/rename (rather than having mv send rename or copy and
delete requests like it does now).

When everyone is sleeping, or when the user says so, renamefs can empty
the rewrite table by actually copying and deleting files in the
underlying filesystem.  Call this operation "update".

There are multiple ways I'm sure to create pathological situations with
renamefs, and I do not know enough about Plan 9 to address them, except
to point out that one way to recover from pathology --assuming the
update operation has not run yet-- is simply to delete the rewrite table
via, eg, echo clear >>/foo/rename or, if the namespace is so screwed up
that that impossible, to reboot without mounting renamefs.

Advantages of my design:

Users who do not need the functionality do not pay any cost: the code
for renamefs and it's clients (eg, the new mv command) lie dormant on
the hard drive where they cannot cause problems.  No modification
of extant code is required.

Makes Plan 9 look more like Unix for those who want want that sort
of thing, and yes I hear loud and clear that many here do not.