The Topen request at first looks weird. Here’s its signature
size Topen tag fid mode
It’s strange, isn’t it? For comparison, here’s the C signature for the open(2) system call:
int open(const char *path, int flags, ...);
Where is the path in the Topen call?
The description for the Topen request says:
The open request asks the file server to check permissions and prepare a fid for I/O with subsequent read and write messages.
Which implies that to write or read a fid you must open… an existing fid?
This morning (well, actually some days ago since this entry got published later) cage explained the mystery to me: Twalk.
When I first skimmed through the 9P documentation I thought that the walk request was basically a chdir(2), which it is, but also is not!
The Twalk requests allows one to navigate from a fid (usually representing the starting directory) through some path components and reach a destination file that will be associated with a new fid.
So, if I connect to a 9P file server and write something to ~/notes.txt (supposing that my home is on that file server) the 9P session could look like this:
# establish a connection and negotiate the version → Tversion <msize> 9P2000 ← Rversion <msize> 9P2000 # mount the home → Tattach 1 "op" "/home/op" ← Rattach <qid> # walk from fid #1 (/home/op) to “notes.txt” (a file!) # and associate to it fid #2 → Twalk 1 2 "notes.txt" ← Rwalk <qid...> # prepare fid 2 (/home/op/notes.txt) for # reading and writing → Topen 2 OWRITE|OREAD ← Ropen <qid> <iounit> # read/write fid 2… # close the fid 2 since it’s no longer used → Tclunk 2 ← Rclunk
(note that as always this is entirely my speculation from reading the documentation. I never used — sigh! — plan9 nor 9p)
So it actually makes sense for Topen to accept a fid and not a path, and Twalk is a general purpose request that can be used to implement various system calls (dup2, chdir, open…)
Then I started to think why it was like this. I mean, everything is finally starting to make sense in my head, but why the 9P people decided to implement Topen and Twalk this way?
Well, I can’t say for sure, but I’m starting to noticing that a fid is something more than a UNIX file descriptor: it’s both a file descriptor AND a path.
*my mind blows*
Which is actually a pretty clever solution. The client can get new fids by mean of Twalk, which then can be passed to Tremove (for removal), or Tstat/Twstat, or being opened and then written/closed. It’s also probably more efficient than passing string around on every request.
This has also some drawbacks probably. For one, it’s not clear at glance if a fid was prepared for I/O or not. On UNIX there is a clear distinction between a file descriptor (a number that references an object in the kernel) and a path (a mere sequence of bytes NUL-terminated.) But since this is an underlying mechanism, it seems pretty clever. It shouldn’t be too much difficult to map the usual UNIX syscalls on top of 9p.