path: root/docs/fs.md

# Filesystem

Every native session serves 9P2000 on a Unix socket. Pane shells receive
`PARDES_PID` (the editor's process id), `PARDES_9P` (socket path) and
`PARDES_PANE` (pane serial). The socket is
`$XDG_RUNTIME_DIR/pardes-9p-<pid>.sock`, or lives under
`~/.local/state/pardes` when XDG_RUNTIME_DIR is unset. Detached sessions use
their session name; `--9p=<name>` overrides it.

A `pardes <file>` launched from a pane forwards Look to that pane over 9P.
`PARDES_PID` alone says the shell is inside pardes; `PARDES_9P` and
`PARDES_PANE` say how to reach it, and a launch that has the first without the
other two refuses rather than opening a second editor. `--nested` opens a
separate editor and withholds `PARDES_PID` from its pane shells, so a pardes
started in one of them runs a session of its own; its 9P service stays
available.

Look resolves the OS filesystem first, then the editor's virtual filesystem.
Explicit paths bypass that search:

| Editor path | Meaning | 9P server path |
|---|---|---|
| `/n/os/proc/self` | OS filesystem | `/os/proc/self` |
| `/n/self/pane/2/body` | pane 2's text | `/pane/2/body` |
| `/virtual/pane/2/body` | the same, in the editor's own spelling | `/pane/2/body` |
| `/virtual/src/pardes.zig` | source embedded in this build | `/src/pardes.zig` |
| `/n/peer/pane/2/body` | another session's text | peer's `/pane/2/body` |

The mount name `self` is reserved and maps to the server root, so `/n/self/X`
and `/virtual/X` both name the served `/X`.

`--mount=peer=work` mounts the named session `work`; the dial can also be an
absolute socket path, `unix!/path`, `tcp!IP!port`, or `quic!IP!port`.
At runtime, use `Mount peer dial` and
`Unmount peer`. There are eight named mounts; `os` and `self` are reserved.
Unmount refuses mounts still used by a pane, its working directory, or a
pending Save. Mounts are saved in dumps. Save uses the file's original mount.

`pardes --9p-tcp='tcp!127.0.0.1!5640'` adds a TCP listener alongside the Unix
socket. Build with `-Dquic=true` and system OpenSSL 3.6+ to enable QUIC;
`--9p-quic='quic!127.0.0.1!5641'` adds its listener. Both accept numeric
IPv4/IPv6 addresses, not DNS names. Listener port zero chooses a free port;
`/listeners` reports all active dial addresses.

All connections have session access, including `os`. TCP is unencrypted.
QUIC uses an ephemeral TLS identity without peer verification or login.
It carries 9P2000 on one bidirectional stream with ALPN `pardes-9p`.
Unix and TCP connections share four slots served by cloud9's `std.Io`
runner; QUIC has four of its own on the editor's poll loop. OpenSSL's
internal buffers are separate, dynamically allocated memory.

[Plan9port's client](https://9fans.github.io/plan9port/man/man1/9p.html) can
drive Unix or TCP without a kernel mount, and `9ns` mounts the tree in a
private namespace:

```sh
9p -n -a "unix!$PARDES_9P" read index
9p -n -a 'tcp!127.0.0.1!5640' ls pane/1
9ns --unix "$PARDES_9P" -- sh -c 'cat "$NINE_MOUNT/index"'
```

For [Linux v9fs](https://www.kernel.org/doc/html/latest/filesystems/9p.html),
use `version=9p2000,cache=none,access=any` and `trans=unix`, or `trans=tcp`
with `port=5640`. Set `uname`, `dfltuid`, and `dfltgid` for the local user.
Leave `aname` empty. The opt-in
[Linux v9fs experiment](v9fs.md) tests a kernel mount in a separate subprocess
namespace (`zig build v9fs-test`, requiring explicit mount authorization).
Neither 9P2000.u nor 9P2000.L is implemented.
Existing Plan9port/v9fs clients need a userspace bridge for QUIC.

## The served tree

```
/README      this guide, also src/fs-help.txt
/index       one line per pane: serial, kind (text|term|pdf|image), dirty flag, name
/status      pid, version and pane count
/look        write a line: a right click on it at the active pane; read: the serials it touched
/exec        write a line: a middle click; read the same serials
/log         recent events, one a line: new|del|rename|save <serial> <name>, msg <serial|-> <text>;
             write follow to that open to wait for more
/screen      rendered screen JSON; frozen per open handle
/listeners   the session's dial addresses
/pane/new    open it to make a pane; the read answers that pane's serial
/pane/<n>/   name body tag ctl addr dot limit data xdata sel dirty mark scroll
             errors event look exec, plus pty/{ctl,status,data} on terminals
/os/         the host filesystem
/src/        the editor's embedded sources, only when built with -Dembed-sources=true
```

A pane is made by **opening** `/pane/new`, and closed by Tremove on
`/pane/<n>` (`rmdir`), which is the only remove the tree serves; Tcreate is
refused everywhere, as it is in acme. Reading the open fid answers the serial
of the pane that open made, so `n=$(cat /pane/new)` makes one and names it in
a line. Each open makes another pane, and two reads of one fid answer the same
serial: the open acted, the read only observes. Closing the fid leaves the
pane.

This is `/net/tcp/clone`'s mechanism, not acme's `new`, and the difference is
deliberate. acme allocates during the *walk* and lets the walk land inside the
new window, so `/dev/new/body` works in one step (acme(4): "accessing any file
in `new` creates a new window"). acme can also afford to list `new`, because a
Plan 9 directory read carries the stat of every entry and nothing walks. A
kernel or FUSE mount is not so lucky: it walks and stats each name a listing
gave it, so an allocate-on-walk name would make a pane per `ls -l`. Allocating
on open instead keeps `new` listed and `ls` honest — a stat is not an open —
at the cost of acme's one-step `new/body`. Nothing in the tree is created by
list, stat, walk or read; only that one open. Every other name in `/pane` is a
serial.

A session can open its own tree through a mount: a Look at
`/mnt/9p/pardes/<me>/pane/2/body` from inside that very editor opens it, and
a Save of that pane writes back through the mount into pane 2. Requests on
the Unix and TCP listeners are answered on the 9P connection's own task, not
by the editor's loop, so the realpath, the stat and the read the editor makes
out through the mount come back while it waits for them. The one rule is
whose turn it is with the core (`pardes.turn`): the editor has it, and gives
it up while it waits for input and while a step of it is out in a syscall. A
step of a connection task's own -- a Look written to `look` -- goes out the
same way, and the editor waits for it to return before it takes a step of
its own. While any step is out, a request that would change a pane (a write,
a truncation, an rmdir) parks in the engine until none is; everything else,
opening `pane/new` and `screen` included, is answered at once, which is why a
Look at any path in the tree comes back. A write into the tree from the
editor itself only ever happens between steps (a Save), so nothing it waits
on out there is a request that has to park. QUIC is still served on the
editor's loop, so through QUIC the old hang remains. `/n/self/...` names the
same tree without leaving the process.

`/look` and `/exec` are the editor's two clicks, one per line of a write:

- a line written to `look` is a right click on it: a path opens a file,
  `file:12` jumps to a line, a directory opens a shell there, a URL opens in
  the browser.
- a line written to `exec` is a middle click: a command word from
  `src/builtins.zig` (`Save`, `Del`, `New`, `Newcol`, `Mount NAME DIAL`,
  `Unmount NAME`, `Dump`, `Restore`, `Msg TEXT`, `Find`, `Grep`, `Tty`, ...),
  or anything else, which runs in the pane's terminal.

The root's pair clicks at the active pane and `/pane/<n>/look` and
`/pane/<n>/exec` at that pane. Blank lines are skipped, and every other line
is checked before any of them runs, so a control character fails the whole
write with EINVAL; a command that fails inside the editor is reported on the
message row, not as a write error. Reading any of these files answers the
serials of the panes the last command created, or, when it created none, the
pane a look focused or the pane an exec acted on (even one it closed), one
per line.

`/pane/<n>/name` reads the pane's file name (a terminal's directory) and
writing it renames the buffer; a relative name resolves against the pane's
directory. `body` appends on write and replaces on truncating open. `sel`
reads the selected text and writing it replaces the selection. `errors`
appends to the directory's `+Errors` pane. Holding `event` open redirects the
pane's Look and Exec clicks to that client; writing a record back performs the
action. `ctl` reads acme's window status line — serial, tag length, body
length, a reserved zero, the dirty flag, the width in cells, the font and the
tab width — and takes `get`, which reloads the buffer from the name it
carries, and acme's `lock` and `unlock` (editors/acme/xfid.c:603-611), for an
edit of several writes to `addr` and `data` that another client must not
land in the middle of. As in acme the lock binds only the clients that take
it: a `lock` while another open holds it waits until that open writes
`unlock` or closes (or the pane does), and nothing else is refused for it --
not a write to any other file, not the person at the keyboard. It belongs to
the open that wrote it, so only that open's `unlock` is taken; a write on an
open that cannot write (or the editor's own, on none) cannot lock.

The three range files `addr`, `dot` and `limit` each read the pair of offsets
they also accept, so copying one onto another is all that acme's `addr=dot`,
`dot=addr` and `limit=addr` ever were. A write is either that pair or an
address expression (`#0,#5`, `/pattern/`, `2+1`); `addr` selects what `data`
and `xdata` read or replace, `dot` is the editor's own selection and moving it
scrolls the pane into view, and `limit` bounds a search and reads empty until
it is set. Truncating a range file empties it; truncating `limit` lifts it.
`addr` belongs to the pane rather than to a client and keeps what was written
until someone writes or truncates it, so writing an address and reading it
back evaluates it, which is what acme(4) promises of its own `addr`.

The three flag files `dirty`, `mark` and `scroll` read `0` or `1` and take
`0` or `1`: whether the buffer differs from its file, whether a write pushes
an undo point (writing `1` pushes one now), and whether a write scrolls the
pane. Truncating `tag` clears the part of the tag you may edit.

Stats report real lengths for `index`, `status`, `look`, `exec`, `listeners`,
`name`, `body`, `tag`, `sel`, `ctl`, the range files and the flag files, and
for `event` and `pty/data` the length of the record a read would
answer, which is zero when nothing is waiting; for `log`, the text an open
would freeze now. Modes are 0644/0666 (0444 for
read-only files, 0222 for write-only); mtime is the pane's last edit or the
process start. The qid version of `body`, `data` and `xdata` is the pane's
revision, so a stat sees an edit land without reading the text; every other
file leaves it zero rather than promise a version it cannot keep. Directory
entries carry no sizes, and neither does `/screen`, which has no length until
an open renders its frame; stat the entry.

`/log` is one ring (64 KiB) that records whether or not anyone reads it:
`new`, `del`, `rename` and `save <serial> <name>`, and `msg <serial|-> <text>`
for every line the editor says, repeats included (with `verbose` on, that
includes each builtin announcing itself as it runs). A `msg` said while a
pane is being made can precede that pane's `new`; panes present at boot are
recorded before anything else.
Control characters in a record become spaces, so a record is one line. An
open freezes the ring's text the way `/screen` freezes a frame: reads walk it
and end. Writing `follow` to that same open makes reads past it wait for the
next record, one per read; a follower the ring outran reads `lost N` first.
Closing the open is the only way back, as with rio's `consctl`.

A read with nothing to give yet -- a following `log`, `event`, `pty/data`, a
`pty/run` before its answer -- is held, the way factotum holds its log's reads
(security/auth/factotum/log.c) and acme an event read: the core keeps it, and
whoever next has news for it (a record, output, a run's answer, the pane
closing, which answers "No such file or directory") answers it on the connection it
came on as the turn is given up. Nothing else parked is retried for it. A
read the client flushed meanwhile is dropped unanswered, so no record is
spent on it. An open waits with one read at a time, as acme's window keeps
one `eventx`. QUIC connections still retry their parked reads each tick.

`pty/run` runs one line at a terminal's prompt and answers how it ended, on
the same open (factotum's `rpc` shape): write the line, then read `exit N`
once the command has ended and the shell is back at a prompt, followed by
what it printed. The output is what the screen showed between the command's
start and end marks: stderr interleaved, `\r` progress collapsed to its last
state, no colour, tabs as the spaces they drew, trailing spaces trimmed and
trailing blank lines dropped (`printf 'a\n\n\n'` answers `a`), leading
whitespace kept; a program on the alternate screen (vim, less, htop) leaves
none, and rows a program redrew above its start are missed. A bash job notice
printed before its PROMPT_COMMAND lands in the next run's output. Only its
last 64 KiB are kept, from a line start, and the header then reads `exit N
cut M` (M bytes left out). `exit N cut`, with no count, says the output's
start is not there to read: it scrolled out of the history, the command
erased it (`clear`, a reset), a start or end mark came on the alternate
screen, or the command printed more than 8192 rows, of which only the last
are read so that the answer costs the editor a bounded amount. `exit ?` is a
command whose end mark carried no status, which is not a success; `error out
of memory` is an answer that could not be made. The header is always the
whole first line. It reads
`busy` at once when a command is running, text is typed at the prompt or the
shell has not drawn its first prompt; `error not run` when the shell refused
the line without running it (a fish syntax error; the line is taken back off
the prompt); `error shell gone` when the pane closed or its shell was
replaced; `error no prompt marks` for a shell pardes could not instrument.
It relies on the OSC 133 marks pardes injects into bash and fish, tagged
`aid=pardes` so fish's own marks and a nested shell's are ignored. A second
line on an open whose command still runs fails the write. `exec zsh` or a
continuation prompt never reports an end; cancel the read. A read of `run`
waits in pardes, so through 9ns it needs 9ns's concurrent requests or it
holds up the rest of the mount. A record
longer than a read comes in pieces, so a shell's `read` loop works. `tail -f`
never writes `follow`, so it sees nothing new: use the follow open instead. `/screen` returns JSON with
`cols`, `rows`, `cursor`, a `styles` table, and row-major `cells` of
`[grapheme, style_index]`. Each open freezes one frame until close. A
terminal `body` freezes its history on the first read of each open handle;
`pty/data` streams live output. An open that holds something between open
and close -- a frozen screen, terminal body or log, a run, an `event` or
`pty/data` open -- takes one of 64 records (lib9p's per-fid aux, acme's
Fid), released on close or disconnect; past that such an open fails with
`ENFILE`. Other opens hold nothing and are not counted.

`-Dembed-sources=true` embeds the editor's sources and serves them under
`/src` (and `/shaders` on GUI builds). `EffectCode <effect>` lists the current
backend's implementation files under `/virtual`, which Look opens; without the
option the command reports the sources as unavailable. The esp32p4 build
enables the option by default, so the device can serve its own source.

This is a control filesystem, not a complete POSIX export. Native filenames
may contain up to 255 bytes. Existing regular OS files support read, write,
and truncation to zero; under `/os` protocol create, remove, rename and other
metadata changes are refused, as is every create in the control tree and every
remove in it but a pane directory's. Ownership and permissions under `/os` are
synthetic.
Zero-length truncation accepts the accompanying `mtime` hint sent by Linux
v9fs; the hint is not stored. Standalone timestamp changes remain refused.

The tree lives in `src/ninep/`: `tree.zig` (nodes, lookup, readdir, dispatch,
and the editor's reply payload over cloud9's backend contract), `pane.zig`
(pane files), `ctl.zig`, `addr.zig`, `pty.zig`, `events.zig` (event and log
streams), `screen.zig` and `sources.zig`. The protocol engine is cloud9's
`fs.Server`, configured in `src/9p.zig` (the editor's and the board's
capacities); the transports are `src/9p_io.zig` (cloud9's `serve.Runner`
for Unix and TCP, a poll loop for QUIC, and the 9P client for mounts);
`src/fs.zig` keeps host access, mounts, resolution, find and grep.

`zig build fs-test` drives real sessions using the independent Python client
in `test/ninep.py`; `zig build fs-discovery-test` checks that browsing
creates nothing, that an open of `/pane/new` and a remove work, and that
`look`, `exec`, `name`, `sel` and `log` behave. `zig build
9p-test` checks the two engine configurations' budgets (the engine's own
tests are cloud9's `zig build test`); `zig build fs-bench` measures
filesystem transactions in the core.
`zig build fs-test quic-test -Dquic=true` also exercises QUIC mounts and I/O.