1 files changed, 148 insertions, 7 deletions

diff --git a/src/detached/server.zig b/src/detached/server.zig
index 8e6b92d4..2d7a5238 100644
--- a/src/detached/server.zig
+++ b/src/detached/server.zig
@@ -174,6 +174,10 @@ const file_watch = @import("../file_watch.zig");
 /// drain, at the same point in the frame that tty.zig drains.
 const fuse = @import("../fuse.zig");
 const fs_service = @import("../fs_service.zig");
+/// ...and the SECOND transport onto that same tree, on a unix socket of its
+/// own. `Source.ninep_listener` and `Source.ninep` are its arms; `pollFrame`
+/// is its drain, beside the mount's, at the same point in the frame.
+const fs9_service = @import("../fs9_service.zig");
 
 /// Host-lifetime storage for the OSC 133 rc files a forked shell sources, held
 /// by `Session` because a Session is exactly one host's lifetime.
@@ -250,9 +254,11 @@ const read_chunk = 16 * 1024;
 const pty_chunk = 64 * 1024;
 
 /// Descriptors in the ONE poll this process runs: the listener, every frontend,
-/// every pane's pty master, the inotify descriptor behind every watch, and
-/// `/dev/fuse`. 51 on a full house, and one syscall covers all of them.
-const poll_slots = 1 + max_clients + pardes.MAX_PANES + 2;
+/// every pane's pty master, the inotify descriptor behind every watch,
+/// `/dev/fuse`, the 9P listener, and every 9P connection. That is
+/// 1 + 32 + 16 + 1 + 1 + 1 + 4 = 56 on a full house, and one syscall covers
+/// all of them.
+const poll_slots = 1 + max_clients + pardes.MAX_PANES + 2 + 1 + fs9_service.max_conns;
 
 /// How many times `reloadWatched` will honour `file_watch.reloadChanged`'s
 /// request for another pass within one round. See `reloadWatched`.
@@ -411,6 +417,16 @@ const Source = union(enum) {
     /// one, and the `fd < 0` guard cannot see it because the fd is valid,
     /// merely different. Same hazard as the client slots, same reason.
     fuse,
+    /// The 9P listener, and one arm per connection on it. Same shape and same
+    /// hazard as `fuse` above, and stated separately only because the hazard
+    /// is WORSE here: a 9P `Twrite` to `ctl` reaches `core.perform` exactly as
+    /// a FUSE write does, so it can `push_spawn` and close a pane's master
+    /// under a later `.pty` entry in this same pass. So `dispatch` moves BYTES
+    /// — accept, read into `push`, drain `output` — and never serves a
+    /// request; `pollFrame` does the serving, after the snapshot is done with.
+    ninep_listener,
+    /// A connection's index in `Listener.conns`.
+    ninep: u8,
 };
 
 pub const Session = struct {
@@ -463,6 +479,31 @@ pub const Session = struct {
     /// Same role as `check_files`: nothing else will wake us, because no
     /// acknowledgement has gone back, so the next round must not sleep.
     fs_pending: bool = false,
+    /// The same tree on a unix socket, or null when `--fs9` was not asked for
+    /// or the bind failed. Owned here, like `fs`, so that `deinit` closes the
+    /// socket and unlinks its path on every way out.
+    ninep: ?*fs9_service.Listener = null,
+    /// A 9P drain stopped with work still owed. `fs_pending`'s twin, and it
+    /// needs its own field rather than sharing that one: they are cleared by
+    /// different drains, and or'ing them into one flag would make a busy 9P
+    /// script keep the FUSE drain's `pending` set forever.
+    ninep_pending: bool = false,
+    /// WHICH TRANSPORT THE REQUEST BEING SERVED CAME FROM, for the whole of
+    /// one `fs_service.drain` and never outside one.
+    ///
+    /// A filesystem reply reaches its transport through `push_fs_reply`, which
+    /// is a HOST method: the core answers a `fs_req` and does not know, and
+    /// must not know, that this process has two filesystems on one tree. This
+    /// field is the routing origin docs/9p.typ's layering table says a second
+    /// listener costs, and it is one pointer set by the drain that already
+    /// knows the answer.
+    ///
+    /// Null outside a drain, and `fsReply` then falls back to the mount. That
+    /// fallback is for a reply the core produced with no request outstanding,
+    /// which `Fs.reply` drops on a slot lookup; it is NOT a routing guess, and
+    /// a 9P reply cannot reach it — every 9P request is answered inside the
+    /// `step` that made it, which is inside the drain that set this.
+    fs_origin: ?fs_service.Transport = null,
     /// Which directory mark belongs to which pane, and the generation the core
     /// has already accepted from each. file_watch.zig owns the shape and the
     /// transaction; this host owns only the descriptor and the wake.
@@ -513,6 +554,15 @@ pub const Session = struct {
             f.deinit();
             s.fs = null;
         }
+        // ...and the 9P socket, for the mount's reason above: a script blocked
+        // on `event` over 9P is woken by the EOF its own connection closing
+        // produces, while its shell is still alive to run its exit path. The
+        // orphaned fids are not pumped — the core they would report releases to
+        // is going with them.
+        if (s.ninep) |l| {
+            l.deinit(s.gpa);
+            s.ninep = null;
+        }
         // Tell everyone the session is over before the socket disappears, so a
         // frontend exits on a `quit` rather than on a read error whose meaning
         // it has to guess. Best effort by construction: these descriptors are
@@ -960,9 +1010,17 @@ pub const Session = struct {
     /// transport holding it. `bytes` was resolved by `pardes.fsPayload` inside
     /// `perform` and is borrowed only for this call, so a body read is a window
     /// onto the pane's live text and copies nothing. `.again` needs no case:
-    /// `Fs.reply` reads the status and re-parks the request itself.
+    /// both transports read the status and re-park the request themselves.
+    ///
+    /// WHICH transport is `fs_origin`, set by the drain that asked. This used
+    /// to be `s.fs` unconditionally, which was right while a mount was the only
+    /// answer there was and became a silent misroute the moment `--fs9` gave
+    /// the session a second one: `Fs.reply` looks the tag up in ITS park table,
+    /// finds nothing, and returns — so a 9P `Tattach` was answered into the
+    /// void and its client waited forever. Measured against plan9port's `9p`.
     fn fsReply(ctx: ?*anyopaque, reply: *const pardes.acmefs.Reply, bytes: []const u8) void {
         const s = of(ctx);
+        if (s.fs_origin) |t| return t.reply(reply, bytes);
         if (s.fs) |f| f.reply(reply, bytes);
     }
 
@@ -985,7 +1043,29 @@ pub const Session = struct {
         // in the surface this frame composes, not the next one. The flag is
         // read by `waitInput`, which must not sleep while the kernel still has
         // requests we have not acknowledged.
-        if (s.fs) |f| s.fs_pending = fs_service.drain(f.transport(), s.core).pending;
+        if (s.fs) |f| {
+            s.fs_origin = f.transport();
+            s.fs_pending = fs_service.drain(s.fs_origin.?, s.core).pending;
+            s.fs_origin = null;
+        }
+        // ...and the 9P connections, in the same breath and for the same
+        // reason. One connection is one `Transport`, so this is
+        // `fs_service.drain` per connection per frame, with `fs_origin` naming
+        // the one being served so its replies come back to it (see `fsReply`).
+        // HERE rather than in `dispatch` for `Source.ninep`'s reason: a
+        // `Twrite` to `ctl` can `push_spawn`, and `dispatch` is mid-iteration
+        // over a descriptor snapshot when it runs.
+        if (s.ninep) |l| {
+            var pending = false;
+            for (0..fs9_service.max_conns) |i| {
+                const t = l.transport(@intCast(i)) orelse continue;
+                s.fs_origin = t;
+                const d = fs_service.drain(t, s.core);
+                s.fs_origin = null;
+                if (l.settle(@intCast(i), d)) pending = true;
+            }
+            s.ninep_pending = pending;
+        }
         for (&s.ptys, 0..) |*pt, pane| {
             if (pt.fd < 0) continue;
             var lbuf: [1024]u8 = undefined;
@@ -1345,6 +1425,28 @@ pub const Session = struct {
             src[n] = .fuse;
             n += 1;
         };
+        // ...and the 9P socket, when there is one: the listener, plus one
+        // descriptor per live connection. POLLOUT only while a connection owes
+        // bytes, which is the same rule and the same reason as a client's and
+        // a pty's — asking for it unconditionally makes every idle socket a
+        // ready descriptor and turns the poll into a spin.
+        if (s.ninep) |l| {
+            if (l.fd >= 0) {
+                fds[n] = .{ .fd = l.fd, .events = poll_in, .revents = 0 };
+                src[n] = .ninep_listener;
+                n += 1;
+            }
+            for (0..fs9_service.max_conns) |i| {
+                if (!l.live(@intCast(i))) continue;
+                fds[n] = .{
+                    .fd = l.conns[i].fd,
+                    .events = if (l.owes(@intCast(i))) poll_in | poll_out else poll_in,
+                    .revents = 0,
+                };
+                src[n] = .{ .ninep = @intCast(i) };
+                n += 1;
+            }
+        }
         // A session with no listener, no clients, no shells and no watches has
         // no event source at all. Returning immediately would spin the outer
         // `while (!core.quit)` at full speed, so sleep the interval the core
@@ -1355,7 +1457,13 @@ pub const Session = struct {
         // `reconcile` posts a resize only when a client is ATTACHED, which means
         // its socket is in the set; and `fs_pending` is only ever set by a drain,
         // which runs only when `fs` is live, which puts `/dev/fuse` in the set.
-        // So `n == 0` implies `!regridded` and `!fs_pending` too.
+        // So `n == 0` implies `!regridded` and `!fs_pending` too. `--fs9` is
+        // the ONE exception, and it is why `ninep_pending` is named on the
+        // `timeout = 0` line below rather than here: a hung-up 9P connection
+        // still owing the core its orphaned fids has no descriptor at all, so
+        // it can be the only work left with `n == 0`. It is also finite — one
+        // release per fid — so the 16 ms nap that path takes costs it a couple
+        // of frames and never a stall.
         if (n == 0) return nap(if (timeout_ms == 0) 16 else timeout_ms);
         // Zero is the core's word for "sleep until something happens" (see
         // pardes.zig `pump`: it passes a frame interval only while an animation
@@ -1373,7 +1481,7 @@ pub const Session = struct {
         // during `perform`, outside this function) and a regrid this round has
         // already performed. Both are consumed before this function returns, so
         // the round must not sleep before reaching them.
-        if (s.check_files or regridded or s.fs_pending) timeout = 0;
+        if (s.check_files or regridded or s.fs_pending or s.ninep_pending) timeout = 0;
         const ready = libc.poll(&fds, @intCast(n), timeout);
         // A timeout is an ordinary frame boundary and EINTR is a signal we do not
         // handle here. Neither skips anything below any more: what used to be an
@@ -1455,6 +1563,26 @@ pub const Session = struct {
             .fuse => if (pfd.revents & (poll_hup | poll_err | poll_nval) != 0) {
                 if (s.fs) |f| f.dead = true;
             },
+            .ninep_listener => if (pfd.revents != 0) {
+                if (s.ninep) |l| l.accept();
+            },
+            // BYTES ONLY. The requests those bytes decode into are served by
+            // `pollFrame`, after this snapshot is done with — see
+            // `Source.ninep`. POLLOUT before POLLIN so a reply the last frame
+            // could not finish writing goes before more work arrives, and
+            // POLLIN before the hangup because a script that wrote a `Tclunk`
+            // and closed has bytes still worth reading; the `live` guard is the
+            // client slots' `c.fd < 0`, for its reason.
+            .ninep => |i| if (s.ninep) |l| {
+                if (!l.live(i)) continue;
+                if (pfd.revents & poll_out != 0) l.flush(i);
+                if (!l.live(i)) continue;
+                if (pfd.revents & poll_in != 0) {
+                    l.fill(i);
+                } else if (pfd.revents & (poll_hup | poll_err | poll_nval) != 0) {
+                    l.drop(i);
+                }
+            },
         };
     }
 
@@ -1898,6 +2026,19 @@ pub fn run(init: std.process.Init, opts: pardes.Options, name: []const u8) !void
     // process polls `/dev/fuse` itself, in the same syscall as everything else.
     // See `Source.fuse`.
     session.fs = fs_service.start(gpa, core);
+    // ...and the same tree on a unix socket, independently: `--fs` and `--fs9`
+    // are two transports and neither is the other's prerequisite, so a daemon
+    // may serve one, both or neither. Null on every failure, for
+    // `fs_service.start`'s reason — a transport that will not bind must cost
+    // the operator their scripting, never their session — and NOT a
+    // `return error` the way the frontend socket above is, because a session
+    // whose frontend socket did not bind is one nobody can ever find, while
+    // this one is merely one nobody can script over 9P.
+    //
+    // A bare `--fs9` is named by the SESSION rather than by the pid: the
+    // operator typed that name to find the daemon again, and having to look up
+    // a pid to reach its filesystem would undo it. `--fs9=<name>` wins.
+    if (opts.fs9) |named| session.ninep = fs9_service.open(gpa, named, name);
 
     const h = session.host();
     core.host = h;