- file length (to avoid having to forward reads to detect end of file)

- virtual directories, reading a directory wipes out the
 virtual directory. virtual property in the inode.

- inode contains the qid, inuse, pointers, and virtual
 property.  one pointer each for 3 levels.

? reserve version == -1 to mean a virtual file? makes
 some things easier.

- first inode is always the root.

- adding any virtual entry to a directory makes the
 whole directory virtual.

? how do we resolve access to missing parts when running
 remotely?

? what are default properties of files (ones we haven't
 stat'd)?

- local walk won't move fid remotely: must later walk fid if it's needed at the server,
   or hold a local clone (perhaps in special fid space) representing the result of the walk

- why not use exportfs's scheme (turn them back into system calls as needed)?
    after all, currently the clients actually use exportfs connection
	- auth?
- why not just add cache to exportfs?
	- need shared cache

- QTEXCL should pass through (ok for cfs because opens already do)
- permission checks on open?

- shared writable files?

- cfs assumes T -> R in order

- could re-use tag of incoming op (probably assumes in order, or put client id in high-order bits)

x String hash table

- auth interaction, and attach

- fid management different for input from clients and output to server

- if any client has fid to file, server has fid to file
- if no client has fid to file, server might have fid to file (LRU)
- if any client has fid open, server has fid open (BUT: modes, auth)
- if no client has fid open, server will not have fid open (for now)
	- OREAD, OWRITE, ORDWR, OEXEC(!)
	- OTRUNC [if not QTAPPEND]
	- OEXCL
	- ORCLOSE
	- OCEXEC (not in protocol)
- initially it's probably adequate to cache only OREAD

fid walk
- Path has optional SFid
- root has SFid from Tattach
	- must not be lost, in case of later attaches
- all local attach fids are bound to root.sfid
- if not localwalk, allocate SFid and walk that
	- if walks to Path with SFid, clunk and substitute
	(in non-concurrent system, can't happen? because localwalk would have returned it)

attach and auth handling
- attach can't be done until successful Tauth
- Tauth just introduces a fid for Tread/Twrite[/Tclunk]
	- need to assume that existing Attach with similar parameters is adequate authentication
	(obviously wrong, because another incoming connection saying Tauth might not represent the
	same principal. with one connection, it hardly matters, because if it's shared,
	spoofing could occur by fid substitution if the connection can't be trusted.
	also, in this particular application, it's usually safe to assume that every connection is either the
	same principal or "none". trust lies in the authentication of the connection, and the associated "speaks for"
	relationship.)

- private sfid in ropen/rcreate for isdirect files; must clunk

- Path's SFid is only walkable; Fid must have separate path to (possibly shared) SFid for IO

- I/O fids should be shared (unless impossible)
	- can't share QTEXCL, or QTDIR [offset clashes]
	- for QTDIR, could read whole directory and cache that, with purely local fid,
		or read and accumulate contents incrementally
	- need a way to find the shared fid
		- stored in File? as at present [one for each possible mode]
		- map [hash] qid.path x (mode&3) x rclose -> SFid
		- first open in given mode -> clone Path's SFid (as at present)
	- have Fid.open: SFid -> shared fid once located?
		- also private SFid when needed (eg, isdirect, ORCLOSE)
	- Qid changes on open (eg, clone)
		- might detect change of qid.path on first open and mark File as volatile
		    (enough to discard File reference from Path)