9ns --mntgen: registry subdirectories are mount points too

A registry entry that is a directory is now served the way the root is:
a synthetic directory listing the real one, dialing the sockets inside
it on walk and recursing into further directories, to max_synth_depth
(8) levels across max_synth_dirs (64) synthetic nodes. That is the
plan9port mntgen shape and the layout zmx now posts under, so a live
session reads at /mnt/9p/zmx/<name>. Before this a directory in the
registry was dialed like a socket and answered EIO for good.

post gains the two entry points the traversal needs: postedDir (the
registry scan, against any directory) and dialPath (a dial by composed
path, no name validation).

Hardening, each from an attack that broke the code:

- BATCH_FORGET carries entries for many owners and puts 0 in the header
  nodeid, so routing it by the header dropped all of them: 32 of 64
  synthetic slots leaked in one close burst and the subdirectories that
  held them answered EIO forever. distributeForgets unpacks the body and
  hands each entry to its owner.
- probe() and connectBlocking() copied a caller's path into the kernel
  address with no bound: a path past sun_path overran the 110-byte stack
  sockaddr (a panic in Debug, silent corruption in ReleaseFast). Both
  refuse it now, probe as `.live` so a claim never deletes what it could
  not inspect.
- That bound then caught 9proc's own listener, which handed probe() the
  whole 108-byte sun_path array instead of the path inside it. The probe
  reads `.live` for anything it cannot ask about, so every stale socket
  became AlreadyListening and no server could ever take a dead
  predecessor's name back. It passes the path now.

Suites: 87/87 root (+7 post/serve attack regressions), 48/48 9ns,
51+88 9ns integration (+4 traversal and slot-recycling checks), 213/0
9ns adversarial, 60/60 9proc plus its adversarial suites with a new
stale-socket takeover check, freestanding green.

1 files changed, 365 insertions, 7 deletions

diff --git a/9ns/src/bridge.zig b/9ns/src/bridge.zig
index 76327c3..457c883 100644
--- a/9ns/src/bridge.zig
+++ b/9ns/src/bridge.zig
@@ -855,6 +855,15 @@ pub const max_mounts: usize = 4096;
 pub const max_root_dirs: usize = 64;
 /// The staged buffer handed to `post.posted` for one root listing.
 pub const stage_len: usize = 8192;
+/// Registry entries that are directories are served like the root itself
+/// (a synthetic directory mirroring the real one, dialing the sockets
+/// found inside); this bounds the synthetic directories one 9ns serves.
+pub const max_synth_dirs: usize = 64;
+/// How deep those registry subdirectories nest.
+pub const max_synth_depth: u8 = 8;
+/// The node-id index reserved for synthetic registry subdirectories;
+/// mounts use ordinals below 4096, so this never collides with one.
+pub const synth_index: u32 = 0xFFFF_FFFF;
 
 /// The node id a server's subtree lives under: `index` in the top bits,
 /// `local` (1 = that server's 9P root) below.
@@ -891,6 +900,31 @@ pub const MntgenOptions = struct {
     msize: u32 = 131072,
 };
 
+/// One registry subdirectory served as a synthetic directory: the real
+/// path it mirrors, the registry-relative key its children dial under,
+/// its slot (which names its node id) and its depth from the registry.
+const SynthDir = struct {
+    slot: usize = 0,
+    parent: u64 = 0,
+    depth: u8 = 0,
+    path_buf: [post.sun_path_len]u8 = undefined,
+    path_len: u16 = 0,
+    rel_buf: [post.sun_path_len]u8 = undefined,
+    rel_len: u16 = 0,
+
+    fn path(sd: *const SynthDir) [:0]const u8 {
+        return sd.path_buf[0..sd.path_len :0];
+    }
+
+    fn rel(sd: *const SynthDir) []const u8 {
+        return sd.rel_buf[0..sd.rel_len];
+    }
+
+    fn nodeid(sd: *const SynthDir) u64 {
+        return mountNode(synth_index, sd.slot + 1);
+    }
+};
+
 /// One dialed server: its 9P session, its bridge state and its worker
 /// thread, plus the queue the dispatcher feeds requests through.
 const Mount = struct {
@@ -998,6 +1032,8 @@ const Mntgen = struct {
     next_index: u32 = 1,
     /// Open directory handles of the synthetic root (dispatcher-owned).
     root_dirs: [max_root_dirs]?*DirList = @splat(null),
+    /// Synthetic registry subdirectories (dispatcher-owned), by slot.
+    synths: [max_synth_dirs]?*SynthDir = @splat(null),
 
     fn deinit(mg: *Mntgen) void {
         // Wake every worker, then join: at this point the child is gone (or
@@ -1031,6 +1067,12 @@ const Mntgen = struct {
                 slot.* = null;
             }
         }
+        for (&mg.synths) |*slot| {
+            if (slot.*) |sd| {
+                mg.gpa.destroy(sd);
+                slot.* = null;
+            }
+        }
         if (mg.req_buf.len != 0) mg.gpa.free(mg.req_buf);
         if (mg.data_buf.len != 0) mg.gpa.free(mg.data_buf);
         if (mg.stage.len != 0) mg.gpa.free(mg.stage);
@@ -1079,12 +1121,23 @@ const Mntgen = struct {
                 mg.routeInterrupt(in.unique);
                 return true;
             },
+            // A BATCH_FORGET carries entries for many owners at once and
+            // cannot be routed by its header nodeid (the kernel sends 0);
+            // see `distributeForgets`.
+            .batch_forget => {
+                mg.distributeForgets(req);
+                return true;
+            },
             else => {},
         }
         if (h.nodeid == fuse.root_id) {
             try mg.handleRoot(req);
             return true;
         }
+        if (mountIndex(h.nodeid) == synth_index) {
+            try mg.handleSynthDir(req);
+            return true;
+        }
         const wants_reply = switch (op) {
             .forget, .batch_forget => false,
             else => true,
@@ -1133,6 +1186,68 @@ const Mntgen = struct {
         }
     }
 
+    /// FUSE_BATCH_FORGET carries (nodeid, nlookup) entries for many owners
+    /// at once — synthetic-root, synthetic-subdirectory and per-mount
+    /// nodeids can all appear in one batch, and the kernel puts 0 in the
+    /// header nodeid. Routing such a batch like an ordinary request would
+    /// hand every entry to one wrong owner and drop the rest: a mount's
+    /// bridge would leak the fid behind each dropped entry forever, and
+    /// a dropped synthetic-subdirectory entry would leak its slot until
+    /// every subdirectory lookup answers EIO. So the dispatcher keeps
+    /// what it owns — the root holds nothing, a synth slot is freed here —
+    /// and forwards each mount-owned entry to its worker as a plain
+    /// single FUSE_FORGET.
+    fn distributeForgets(mg: *Mntgen, req: fuse.Request) void {
+        const in = fuse.body(fuse.BatchForgetIn, req) catch return;
+        const rest = req.body[@sizeOf(fuse.BatchForgetIn)..];
+        const count: usize = in.count;
+        if (rest.len < count * @sizeOf(fuse.ForgetOne)) return;
+        for (0..count) |i| {
+            const one = std.mem.bytesToValue(fuse.ForgetOne, rest[i * @sizeOf(fuse.ForgetOne) ..][0..@sizeOf(fuse.ForgetOne)]);
+            mg.forgetOne(one.nodeid, one.nlookup, req.header.unique);
+        }
+    }
+
+    /// Forgets one node by id, whatever owns it: the root holds nothing,
+    /// a synthetic subdirectory's slot goes back to the pool, and a
+    /// mount-owned node is forwarded to its worker (which clunks the fid
+    /// behind it). Unknown or dead mounts drop the entry, like a single
+    /// FORGET routed by `route`.
+    fn forgetOne(mg: *Mntgen, nodeid: u64, nlookup: u64, unique: u64) void {
+        if (nodeid == fuse.root_id) return;
+        const idx = mountIndex(nodeid);
+        if (idx == synth_index) {
+            const local = nodeid & mount_node_mask;
+            if (local == 0 or local > max_synth_dirs) return;
+            const slot: usize = @intCast(local - 1);
+            if (mg.synths[slot]) |sd| {
+                mg.trace("   synthetic directory slot {d} forgotten", .{slot});
+                mg.gpa.destroy(sd);
+                mg.synths[slot] = null;
+            }
+            return;
+        }
+        const m = if (idx < max_mounts) mg.mounts[idx] else null;
+        if (m == null or m.?.dead.load(.seq_cst)) return;
+        // A synthesized single FORGET (no reply is expected for one, so
+        // the unique is only bookkeeping).
+        var buf: [@sizeOf(fuse.InHeader) + @sizeOf(fuse.ForgetIn)]u8 = undefined;
+        const hdr = fuse.InHeader{
+            .len = @sizeOf(fuse.InHeader) + @sizeOf(fuse.ForgetIn),
+            .opcode = @intFromEnum(fuse.Opcode.forget),
+            .unique = unique,
+            .nodeid = nodeid,
+            .uid = 0,
+            .gid = 0,
+            .pid = 0,
+            .total_extlen = 0,
+            .padding = 0,
+        };
+        @memcpy(buf[0..@sizeOf(fuse.InHeader)], std.mem.asBytes(&hdr));
+        @memcpy(buf[@sizeOf(fuse.InHeader)..], std.mem.asBytes(&fuse.ForgetIn{ .nlookup = nlookup }));
+        mg.enqueue(m.?, &buf);
+    }
+
     /// A FUSE_INTERRUPT names the request it wants cancelled; FUSE uniques
     /// are unique across the whole connection, so the mount whose bridge is
     /// currently serving that unique gets the packet and its session turns
@@ -1257,11 +1372,34 @@ const Mntgen = struct {
             const out = mg.rootEntryOut(m.root_node, m.root_attr);
             return mg.reply(u, &.{std.mem.asBytes(&out)});
         }
-        const m = mg.dialMount(name) catch |e| switch (e) {
-            error.NotPosted => {
-                mg.trace("   lookup '{s}': nothing posted under that name", .{name});
-                return mg.replyError(u, .NOENT);
+        // What the entry is decides what a walk into it becomes: a socket
+        // dials (the original behavior), a directory is served like the
+        // root itself (its sockets dial on walk, its directories recurse),
+        // anything else answers EIO.
+        var path_buf: [post.sun_path_len]u8 = undefined;
+        const entry_path = post.registryPath(mg.mo.env, name, &path_buf) catch
+            return mg.replyError(u, .NOENT);
+        const st = std.Io.Dir.statFile(.cwd(), mg.mo.io, entry_path, .{}) catch {
+            mg.trace("   lookup '{s}': nothing posted under that name", .{name});
+            return mg.replyError(u, .NOENT);
+        };
+        switch (st.kind) {
+            .directory => {
+                const sd = mg.newSynth(entry_path, name, 1, fuse.root_id) catch |e| {
+                    mg.trace("   lookup '{s}': no synthetic slot: {t}", .{ name, e });
+                    return mg.replyError(u, .IO);
+                };
+                const node = sd.nodeid();
+                const out = mg.rootEntryOut(node, mg.synthAttr(node));
+                return mg.reply(u, &.{std.mem.asBytes(&out)});
             },
+            .unix_domain_socket => {},
+            else => {
+                mg.trace("   lookup '{s}': registry entry is not a socket", .{name});
+                return mg.replyError(u, .IO);
+            },
+        }
+        const m = mg.dialMount(name) catch |e| switch (e) {
             error.Stale => {
                 mg.trace("   lookup '{s}': registry entry is stale (no server behind it)", .{name});
                 return mg.replyError(u, .IO);
@@ -1297,6 +1435,204 @@ const Mntgen = struct {
         return list;
     }
 
+    // -- synthetic registry subdirectories ------------------------------------
+
+    /// A directory entry in the registry (or in one of its
+    /// subdirectories) is served like the root: a synthetic directory
+    /// listing the real one, whose sockets dial on walk and whose
+    /// directories recurse. Served on the dispatcher thread, like the
+    /// root.
+    fn handleSynthDir(mg: *Mntgen, req: fuse.Request) error{FuseIo}!void {
+        const u = req.header.unique;
+        const local = req.header.nodeid & mount_node_mask;
+        if (local == 0 or local > max_synth_dirs) return mg.replyError(u, .IO);
+        const slot: usize = @intCast(local - 1);
+        const sd = mg.synths[slot] orelse return mg.replyError(u, .IO);
+        switch (req.header.op()) {
+            .forget, .batch_forget => {
+                // The kernel dropped the dentry; the slot goes with it.
+                mg.trace("   synthetic directory slot {d} forgotten", .{slot});
+                mg.gpa.destroy(sd);
+                mg.synths[slot] = null;
+            },
+            .getattr => {
+                const out = fuse.AttrOut{ .attr = mg.synthAttr(sd.nodeid()) };
+                try mg.reply(u, &.{std.mem.asBytes(&out)});
+            },
+            .lookup => try mg.synthLookup(sd, req),
+            .opendir => {
+                var fh: ?usize = null;
+                for (&mg.root_dirs, 0..) |*dir_slot, i| {
+                    if (dir_slot.* == null) {
+                        fh = i;
+                        break;
+                    }
+                }
+                const dir_slot = fh orelse return mg.replyError(u, .MFILE);
+                const list = mg.synthListing(sd) catch {
+                    return mg.replyError(u, .IO);
+                };
+                mg.root_dirs[dir_slot] = list;
+                const out = fuse.OpenOut{ .fh = dir_slot };
+                try mg.reply(u, &.{std.mem.asBytes(&out)});
+            },
+            .readdir => {
+                const in = fuse.body(fuse.ReadIn, req) catch return mg.replyError(u, .BADF);
+                if (in.fh >= max_root_dirs) return mg.replyError(u, .BADF);
+                const list = mg.root_dirs[@intCast(in.fh)] orelse return mg.replyError(u, .BADF);
+                const size: usize = @min(in.size, max_write);
+                const used = packDirents(list.entries.items, in.offset, mg.data_buf[0..size]);
+                try mg.reply(u, &.{mg.data_buf[0..used]});
+            },
+            .release, .releasedir => {
+                const in = fuse.body(fuse.ReleaseIn, req) catch return mg.replyError(u, .BADF);
+                if (in.fh < max_root_dirs) {
+                    if (mg.root_dirs[@intCast(in.fh)]) |list| {
+                        list.deinit(mg.gpa);
+                        mg.gpa.destroy(list);
+                        mg.root_dirs[@intCast(in.fh)] = null;
+                    }
+                }
+                try mg.reply(u, &.{});
+            },
+            .statfs => {
+                const out = fuse.StatfsOut{ .st = .{ .bsize = 4096, .namelen = 255, .frsize = 4096 } };
+                try mg.reply(u, &.{std.mem.asBytes(&out)});
+            },
+            .flush, .fsync, .fsyncdir => try mg.reply(u, &.{}),
+            // Capability probes read as "not supported", like the root's.
+            .access, .setxattr, .getxattr, .listxattr, .removexattr, .statx => try mg.replyError(u, .NOSYS),
+            else => try mg.replyError(u, .PERM),
+        }
+    }
+
+    fn synthAttr(mg: *const Mntgen, nodeid: u64) fuse.Attr {
+        // Read-only like the root and like /srv.
+        return .{
+            .ino = nodeid,
+            .mode = fuse.S_IFDIR | 0o555,
+            .nlink = 2,
+            .uid = mg.opts.uid,
+            .gid = mg.opts.gid,
+            .blksize = 4096,
+        };
+    }
+
+    fn synthLookup(mg: *Mntgen, sd: *SynthDir, req: fuse.Request) error{FuseIo}!void {
+        const u = req.header.unique;
+        const name = fuse.nameAfter(void, req) catch return mg.replyError(u, .INVAL);
+        if (std.mem.eql(u8, name, ".")) {
+            const out = mg.rootEntryOut(sd.nodeid(), mg.synthAttr(sd.nodeid()));
+            return mg.reply(u, &.{std.mem.asBytes(&out)});
+        }
+        // The kernel resolves ".." from its own dentry tree and a LOOKUP of
+        // it has never been observed, but it must not alias the directory
+        // onto itself either: answer with the parent's node id (the root's
+        // for a top-level subdirectory — its attr is the same shape).
+        if (std.mem.eql(u8, name, "..")) {
+            const out = mg.rootEntryOut(sd.parent, mg.synthAttr(sd.parent));
+            return mg.reply(u, &.{std.mem.asBytes(&out)});
+        }
+        if (!post.legalName(name)) return mg.replyError(u, .NOENT);
+        // The registry-relative key this child dials under (a mount's
+        // name, for findMount).
+        var key_buf: [post.sun_path_len]u8 = undefined;
+        const key = std.fmt.bufPrint(&key_buf, "{s}/{s}", .{ sd.rel(), name }) catch
+            return mg.replyError(u, .NOTNAM);
+        var path_buf: [post.sun_path_len]u8 = undefined;
+        const child = std.fmt.bufPrintSentinel(&path_buf, "{s}/{s}", .{ sd.path(), name }, 0) catch
+            return mg.replyError(u, .NOTNAM);
+        if (findMount(mg.mounts, key)) |m| {
+            mg.trace("   lookup '{s}': mount {d} already live", .{ key, m.index });
+            const out = mg.rootEntryOut(m.root_node, m.root_attr);
+            return mg.reply(u, &.{std.mem.asBytes(&out)});
+        }
+        const st = std.Io.Dir.statFile(.cwd(), mg.mo.io, child, .{}) catch |e| switch (e) {
+            error.FileNotFound => {
+                mg.trace("   lookup '{s}': no entry", .{key});
+                return mg.replyError(u, .NOENT);
+            },
+            else => {
+                mg.trace("   lookup '{s}': stat failed: {t}", .{ key, e });
+                return mg.replyError(u, .IO);
+            },
+        };
+        switch (st.kind) {
+            .directory => {
+                const child_sd = mg.newSynth(child, key, sd.depth + 1, sd.nodeid()) catch |e| {
+                    mg.trace("   lookup '{s}': no synthetic slot: {t}", .{ key, e });
+                    return mg.replyError(u, .IO);
+                };
+                const node = child_sd.nodeid();
+                const out = mg.rootEntryOut(node, mg.synthAttr(node));
+                return mg.reply(u, &.{std.mem.asBytes(&out)});
+            },
+            .unix_domain_socket => {
+                const m = mg.dialMountAt(child, key) catch |e| {
+                    mg.trace("   lookup '{s}': dial failed: {t}", .{ key, e });
+                    return mg.replyError(u, .IO);
+                };
+                mg.trace("   lookup '{s}': dialed as mount {d}", .{ key, m.index });
+                const out = mg.rootEntryOut(m.root_node, m.root_attr);
+                try mg.reply(u, &.{std.mem.asBytes(&out)});
+            },
+            // Not a service and not a directory: the entry answers EIO on
+            // walk, like a plain file in the registry itself.
+            else => {
+                mg.trace("   lookup '{s}': entry is not a socket or directory", .{key});
+                return mg.replyError(u, .IO);
+            },
+        }
+    }
+
+    /// One OPENDIR of a synthetic subdirectory: `.` and `..` plus the
+    /// real directory's entries, snapshotted for the life of the handle
+    /// (a fresh OPENDIR sees fresh entries), like the root.
+    fn synthListing(mg: *Mntgen, sd: *SynthDir) !*DirList {
+        const list = try mg.gpa.create(DirList);
+        errdefer mg.gpa.destroy(list);
+        list.* = .{};
+        errdefer list.deinit(mg.gpa);
+        const node = sd.nodeid();
+        try list.entries.append(mg.gpa, .{ .name = try mg.gpa.dupe(u8, "."), .ino = node, .dtype = fuse.DT_DIR });
+        try list.entries.append(mg.gpa, .{ .name = try mg.gpa.dupe(u8, ".."), .ino = sd.parent, .dtype = fuse.DT_DIR });
+        var names = post.postedDir(mg.mo.io, sd.path(), mg.stage) catch |e| {
+            mg.trace("   directory listing failed: {t}", .{e});
+            return error.Registry;
+        };
+        while (names.next()) |name| {
+            if (!validDirentName(name)) continue;
+            try list.entries.append(mg.gpa, .{ .name = try mg.gpa.dupe(u8, name), .ino = nameIno(name), .dtype = fuse.DT_DIR });
+        }
+        return list;
+    }
+
+    /// Allocates a synthetic directory node mirroring `path`, keyed by
+    /// the registry-relative `key`, at `depth` under `parent`'s node id.
+    fn newSynth(mg: *Mntgen, path: [:0]const u8, key: []const u8, depth: u8, parent: u64) !*SynthDir {
+        if (depth > max_synth_depth) return error.TooDeep;
+        var slot: ?usize = null;
+        for (&mg.synths, 0..) |*s, i| {
+            if (s.* == null) {
+                slot = i;
+                break;
+            }
+        }
+        const i = slot orelse return error.TooMany;
+        const sd = try mg.gpa.create(SynthDir);
+        errdefer mg.gpa.destroy(sd);
+        sd.* = .{ .slot = i, .parent = parent, .depth = depth };
+        if (path.len + 1 > sd.path_buf.len) return error.NameTooLong;
+        @memcpy(sd.path_buf[0..path.len], path);
+        sd.path_buf[path.len] = 0;
+        sd.path_len = @intCast(path.len);
+        if (key.len > sd.rel_buf.len) return error.NameTooLong;
+        @memcpy(sd.rel_buf[0..key.len], key);
+        sd.rel_len = @intCast(key.len);
+        mg.synths[i] = sd;
+        return sd;
+    }
+
     // -- dialing ---------------------------------------------------------------
 
     /// Dials `name` out of the registry, attaches, stats the server root and
@@ -1304,10 +1640,22 @@ const Mntgen = struct {
     /// of the LOOKUP that triggered it (so a hung server delays that walk,
     /// like it would delay any 9P client).
     fn dialMount(mg: *Mntgen, name: []const u8) !*Mount {
+        const stream = try post.dial(mg.mo.io, mg.mo.env, name);
+        return mg.mountStream(stream, name);
+    }
+
+    /// Dials the socket at `path` (a registry subdirectory entry) and
+    /// mounts it under `key`, the registry-relative path — the
+    /// subdirectory analogue of `dialMount`.
+    fn dialMountAt(mg: *Mntgen, path: [:0]const u8, key: []const u8) !*Mount {
+        const stream = try post.dialPath(mg.mo.io, path);
+        return mg.mountStream(stream, key);
+    }
+
+    /// The shared dial tail: session, attach, stat, bridge and worker.
+    fn mountStream(mg: *Mntgen, stream: std.Io.net.Stream, key: []const u8) !*Mount {
         if (mg.next_index >= max_mounts) return error.TooManyMounts;
         const index: u32 = mg.next_index;
-
-        const stream = try post.dial(mg.mo.io, mg.mo.env, name);
         const fd: i32 = @intCast(stream.socket.handle);
         // The session below does blocking I/O: make sure a dial that left
         // the descriptor nonblocking cannot spin its read loop on EAGAIN,
@@ -1369,7 +1717,7 @@ const Mntgen = struct {
             .gpa = mg.gpa,
             .io = mg.io,
             .debug = mg.opts.debug,
-            .name = try mg.gpa.dupe(u8, name),
+            .name = try mg.gpa.dupe(u8, key),
             .index = index,
             .root_node = root_node,
             .root_attr = attrFromStat(st, inoFromPath(st.qid.path) ^ b.ino_xor, mg.opts.uid, mg.opts.gid),
@@ -1848,6 +2196,16 @@ test "mntgen node id layout: index in the top bits, local ids below" {
     try testing.expect(mount_node_mask == (1 << 32) - 1);
 }
 
+test "mntgen synthetic subdirectory node ids route apart from mounts" {
+    for (1..max_synth_dirs + 1) |i| {
+        const node = mountNode(synth_index, i);
+        try testing.expectEqual(synth_index, mountIndex(node));
+        try testing.expectEqual(i, node & mount_node_mask);
+    }
+    // The reserved index can never be a mount ordinal.
+    try testing.expect(synth_index >= max_mounts);
+}
+
 test "mntgen synthetic-root inos are deterministic and name-derived" {
     const a = nameIno("alpha");
     try testing.expectEqual(a, nameIno("alpha"));