perf: reduce allocations in the hot-path of the pty-proxy

Commit `4c2924c`

Eric Bower · 2026-06-19 23:07:19 -0400 EDT

parent 79169a0

perf: reduce allocations in the hot-path of the pty-proxy

3 files changed, +17, -6

M src/ipc.zig

M src/main.zig

M src/util.zig

M src/ipc.zig

+5, -3

......@@ -92,14 +92,16 @@ pub fn appendMessage(
    tag: Tag,
    data: []const u8,
) !void {
    std.log.info("sending ipc message tag={s}", .{@tagName(tag)});
    const header = Header{
        .tag = tag,
        .len = @intCast(data.len),
    };
    try list.appendSlice(alloc, std.mem.asBytes(&header));
    // Guarantee capacity for header + payload in one check to avoid
    // intermediate realloc between the two appends on the hot path.
    try list.ensureTotalCapacity(alloc, list.items.len + @sizeOf(Header) + data.len);
    list.appendSliceAssumeCapacity(std.mem.asBytes(&header));
    if (data.len > 0) {
        try list.appendSlice(alloc, data);
        list.appendSliceAssumeCapacity(data);
    }
}

M src/main.zig

+9, -2

......@@ -2558,7 +2558,11 @@ fn daemonLoop(daemon: *Daemon, server_sock_fd: i32, pty_fd: i32) !void {
                .read_buf = try ipc.SocketBuffer.init(daemon.alloc),
                .write_buf = undefined,
            };
            client.write_buf = try std.ArrayList(u8).initCapacity(client.alloc, 4096);
            // 64KB initial capacity lets ~15 broadcast cycles (N_TTY_BUF_SIZE reads
            // * header) accumulate before the first ArrayList growth. The write
            // buffer is userspace-only: it drains via POLLOUT to the client socket,
            // which has no corresponding kernel-imposed per-write limit.
            client.write_buf = try std.ArrayList(u8).initCapacity(client.alloc, 65536);
            try daemon.clients.append(daemon.alloc, client);
            std.log.info(
                "client connected fd={d} total={d}",

......@@ -2568,7 +2572,10 @@ fn daemonLoop(daemon: *Daemon, server_sock_fd: i32, pty_fd: i32) !void {

        const inp_flags = posix.POLL.IN | posix.POLL.HUP | posix.POLL.ERR | posix.POLL.NVAL;
        if (poll_fds.items[1].revents & inp_flags != 0) {
            // Read from PTY
            // Read from PTY. Buffer is sized to N_TTY_BUF_SIZE (4096): the hard
            // kernel limit for the N_TTY line discipline. A larger buffer doesn't
            // help: each read() from a PTY master returns at most 4096 bytes
            // regardless of the userspace buffer size.
            var buf: [4096]u8 = undefined;
            const n_opt: ?usize = posix.read(pty_fd, &buf) catch |err| blk: {
                if (err == error.WouldBlock) break :blk null;

M src/util.zig

+3, -1

......@@ -197,7 +197,9 @@ const OSC_133_A = "\x1b]133;A";
/// makes the prompt invisible.
/// See: https://github.com/neurosnap/zmx/issues/111
pub fn rewritePromptRedraw(alloc: std.mem.Allocator, data: []const u8) ?[]const u8 {
    // Quick scan: is there any OSC 133;A in this chunk?
    // Fast-path: most PTY output has no escape sequences at all. A scalar
    // byte scan for ESC is cheaper than the full string indexOf below.
    if (std.mem.indexOfScalar(u8, data, '\x1b') == null) return null;
    if (std.mem.indexOf(u8, data, OSC_133_A) == null) return null;

    var result = std.ArrayList(u8).initCapacity(alloc, data.len + 200) catch return null;

...	...		@@ -92,14 +92,16 @@ pub fn appendMessage(
92	92		tag: Tag,
93	93		data: []const u8,
94	94		) !void {
95		-	std.log.info("sending ipc message tag={s}", .{@tagName(tag)});
96	95		const header = Header{
97	96		.tag = tag,
98	97		.len = @intCast(data.len),
99	98		};
100		-	try list.appendSlice(alloc, std.mem.asBytes(&header));
	99	+	// Guarantee capacity for header + payload in one check to avoid
	100	+	// intermediate realloc between the two appends on the hot path.
	101	+	try list.ensureTotalCapacity(alloc, list.items.len + @sizeOf(Header) + data.len);
	102	+	list.appendSliceAssumeCapacity(std.mem.asBytes(&header));
101	103		if (data.len > 0) {
102		-	try list.appendSlice(alloc, data);
	104	+	list.appendSliceAssumeCapacity(data);
103	105		}
104	106		}
105	107

...	...		@@ -2558,7 +2558,11 @@ fn daemonLoop(daemon: *Daemon, server_sock_fd: i32, pty_fd: i32) !void {
2558	2558		.read_buf = try ipc.SocketBuffer.init(daemon.alloc),
2559	2559		.write_buf = undefined,
2560	2560		};
2561		-	client.write_buf = try std.ArrayList(u8).initCapacity(client.alloc, 4096);
	2561	+	// 64KB initial capacity lets ~15 broadcast cycles (N_TTY_BUF_SIZE reads
	2562	+	// * header) accumulate before the first ArrayList growth. The write
	2563	+	// buffer is userspace-only: it drains via POLLOUT to the client socket,
	2564	+	// which has no corresponding kernel-imposed per-write limit.
	2565	+	client.write_buf = try std.ArrayList(u8).initCapacity(client.alloc, 65536);
2562	2566		try daemon.clients.append(daemon.alloc, client);
2563	2567		std.log.info(
2564	2568		"client connected fd={d} total={d}",

...	...		@@ -2568,7 +2572,10 @@ fn daemonLoop(daemon: *Daemon, server_sock_fd: i32, pty_fd: i32) !void {
2568	2572
2569	2573		const inp_flags = posix.POLL.IN \| posix.POLL.HUP \| posix.POLL.ERR \| posix.POLL.NVAL;
2570	2574		if (poll_fds.items[1].revents & inp_flags != 0) {
2571		-	// Read from PTY
	2575	+	// Read from PTY. Buffer is sized to N_TTY_BUF_SIZE (4096): the hard
	2576	+	// kernel limit for the N_TTY line discipline. A larger buffer doesn't
	2577	+	// help: each read() from a PTY master returns at most 4096 bytes
	2578	+	// regardless of the userspace buffer size.
2572	2579		var buf: [4096]u8 = undefined;
2573	2580		const n_opt: ?usize = posix.read(pty_fd, &buf) catch \|err\| blk: {
2574	2581		if (err == error.WouldBlock) break :blk null;

...	...		@@ -197,7 +197,9 @@ const OSC_133_A = "\x1b]133;A";
197	197		/// makes the prompt invisible.
198	198		/// See: https://github.com/neurosnap/zmx/issues/111
199	199		pub fn rewritePromptRedraw(alloc: std.mem.Allocator, data: []const u8) ?[]const u8 {
200		-	// Quick scan: is there any OSC 133;A in this chunk?
	200	+	// Fast-path: most PTY output has no escape sequences at all. A scalar
	201	+	// byte scan for ESC is cheaper than the full string indexOf below.
	202	+	if (std.mem.indexOfScalar(u8, data, '\x1b') == null) return null;
201	203		if (std.mem.indexOf(u8, data, OSC_133_A) == null) return null;
202	204
203	205		var result = std.ArrayList(u8).initCapacity(alloc, data.len + 200) catch return null;

Commit 4c2924c

Commit `4c2924c`