From 1c87b21d9255d9b0a7af102afd7defec25ac2c8a Mon Sep 17 00:00:00 2001 From: Gabriel Schneider Date: Thu, 29 Oct 2026 16:55:47 -0300 Subject: rollback-bench and fork_stall.py: what a replica's rollback copy is (a walk of ~100 KB once file texts are pinned) and what a fork stalls the authority for, which settle open points 3 and 5 `zig build rollback-bench -Dplatform=tty -Doptimize=ReleaseFast` (test/rollback_bench.zig) builds a typical session (a shell with 1 MB of scrollback, three 60 KB files and a 3 MB one, 40 edits) and a big one (14 panes, 18 MB, 4x3 MB and 8 small files, 8 MB of scrollback, 200 edits) in process, then measures the state the digest walk visits, the part of it that is a file's text and its undo snapshots (immutable once made: an edit puts a new buffer in place, a snapshot is a copy, so a copy pins them), and the walk again with those taken out, plus a memcpy of what is left. Typical: 36.3 MB walked, 80 KB to copy, about 0.02 ms; big: 282.5 MB, 96 KB, about 0.045 ms. Then a fork of the big session (431 MB RSS: 3.4 to 8.3 ms) and 64 MB written by the parent under the child (~28 ms of copy-on-write). test/fork_stall.py does the fork alone, on any system with a Python (the FreeBSD box): N MB written, then fork() timed and 64 MB written under the child, with Linux's THP as set and with 4 KiB pages. 1 GB: 13.3 to 19.6 ms here, 20.4 ms on FreeBSD 15.1; the copy-on-write after it ~1.3 us a page on both. Co-Authored-By: Claude Opus 5.5 --- test/rollback_bench.zig | 246 ++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 246 insertions(+) create mode 100644 test/rollback_bench.zig (limited to 'test/rollback_bench.zig') diff --git a/test/rollback_bench.zig b/test/rollback_bench.zig new file mode 100644 index 00000000..721c6f13 --- /dev/null +++ b/test/rollback_bench.zig @@ -0,0 +1,246 @@ +//! What a replica's rollback copy would cost (docs/next-daemon.md, open +//! point 3), and what a fork of the authority stalls it for (open point 5). +//! A typical session and a big one are built in process; for each: +//! - the state the walker visits (src/walk.zig's digest walk, which is +//! what speculation can write: views, buffers, registers, counters, +//! terminal metadata), in bytes; +//! - the part of it that is pinned blobs: a file's text and its undo +//! snapshots, which an edit replaces rather than changes +//! (File.setContentSpan), so a copy keeps a reference; +//! - the walk's time, the blobs' share of it, and a memcpy of the rest: +//! a walker copy costs about the walk plus that write. +//! Then a fork of the big one, child gone at once, and the authority's +//! copy-on-write after it. +//! +//! zig build rollback-bench -Dplatform=tty -Doptimize=ReleaseFast +const std = @import("std"); +const libc = std.c; +const pardes = @import("pardes"); +const walk = pardes.walk; + +const gpa = std.heap.c_allocator; + +fn nowNs() u64 { + var ts: libc.timespec = undefined; + _ = libc.clock_gettime(.MONOTONIC, &ts); + return @as(u64, @intCast(ts.sec)) *| 1_000_000_000 +| @as(u64, @intCast(ts.nsec)); +} + +const File = struct { lines: usize, copies: usize = 1 }; + +const Session = struct { + name: []const u8, + files: []const File, + /// The shell's scrollback, fed through its terminal. + scrollback_kb: usize, + /// Insert sessions (`i x Esc`), each an undo step, spread over the files. + edits: usize, +}; + +// 60-column lines: 1k lines is 60 KB, 50k is 3 MB, 300k is 18 MB. +const sessions = [_]Session{ + .{ .name = "typical", .files = &.{ .{ .lines = 1_000, .copies = 3 }, .{ .lines = 50_000 } }, .scrollback_kb = 1024, .edits = 40 }, + .{ .name = "big", .files = &.{ .{ .lines = 1_000, .copies = 8 }, .{ .lines = 50_000, .copies = 4 }, .{ .lines = 300_000 } }, .scrollback_kb = 8 * 1024, .edits = 200 }, +}; + +fn pump(core: *pardes.Pardes) void { + while (core.nextEffect()) |_| {} +} + +fn key(core: *pardes.Pardes, cp: u21, text: []const u8) void { + core.update(.{ .key = .{ .cp = cp, .text = text } }); + pump(core); +} + +fn writeFile(path: []const u8, lines: usize) !void { + var out: std.Io.Writer.Allocating = .init(gpa); + defer out.deinit(); + for (0..lines) |n| try out.writer.print("const value_{d:0>7}: u32 = {d:0>10}; // a line of sixty\n", .{ n, n *% 2654435761 }); + var pathbuf: [4096]u8 = undefined; + const fd = libc.open(try std.fmt.bufPrintSentinel(&pathbuf, "{s}", .{path}, 0), .{ .ACCMODE = .WRONLY, .CREAT = true, .TRUNC = true }, @as(libc.mode_t, 0o644)); + if (fd < 0) return error.OpenFailed; + defer _ = libc.close(fd); + const text = out.written(); + var off: usize = 0; + while (off < text.len) { + const n = libc.write(fd, text.ptr + off, text.len - off); + if (n <= 0) return error.WriteFailed; + off += @intCast(n); + } +} + +fn build(s: Session, dir: []const u8) !*pardes.Pardes { + const core = try pardes.Pardes.init(gpa, .{ .tty_only = true }); + core.update(.{ .resize = .{ .cols = 200, .rows = 60 } }); + pump(core); + // The shell: pane 0 of a tty boot, its output fed as a pty's. + var chunk: std.Io.Writer.Allocating = .init(gpa); + defer chunk.deinit(); + var n: usize = 0; + while (chunk.written().len < s.scrollback_kb * 1024) : (n += 1) + try chunk.writer.print("\x1b[90m[{d:0>6}]\x1b[0m compiling module_{d} ... \x1b[32mok\x1b[0m\r\n", .{ n, n }); + var off: usize = 0; + while (off < chunk.written().len) { + const end = @min(off + 8192, chunk.written().len); + core.update(.{ .output = .{ .pane = 0, .bytes = chunk.written()[off..end] } }); + pump(core); + off = end; + } + var opened: [pardes.MAX_PANES]usize = undefined; + var nopened: usize = 0; + for (s.files, 0..) |f, fi| for (0..f.copies) |c| { + var buf: [512]u8 = undefined; + const path = try std.fmt.bufPrint(&buf, "{s}/f{d}-{d}.zig", .{ dir, fi, c }); + try writeFile(path, f.lines); + pardes.look.lookAt(core, 0, path); + pump(core); + for (core.panes, 0..) |slot, id| if (slot) |pane| if (pane.file) |file| if (std.mem.eql(u8, file.path, path)) { + opened[nopened] = id; + nopened += 1; + }; + }; + if (nopened == 0) return error.NothingOpened; + for (0..s.edits) |e| { + core.active = opened[e % nopened]; + key(core, 'i', "i"); + key(core, 'x', "x"); + key(core, pardes.Key.escape, ""); + } + var arena: std.heap.ArenaAllocator = .init(gpa); + defer arena.deinit(); + _ = try core.render(arena.allocator()); + return core; +} + +/// Where a file's text and its undo snapshots are held: what a copy pins. +fn blobs(core: *pardes.Pardes, out: *std.ArrayList(*[]u8)) !usize { + var total: usize = 0; + for (core.panes) |slot| if (slot) |pane| if (pane.file) |*f| { + try out.append(gpa, &f.content); + for (f.history.undo[0..f.history.undo_len]) |*snap| try out.append(gpa, &snap.content); + for (f.history.redo[0..f.history.redo_len]) |*snap| try out.append(gpa, &snap.content); + }; + for (out.items) |b| total += b.len; + return total; +} + +fn walkNs(core: *const pardes.Pardes) u64 { + var t: [reps]u64 = undefined; + var sink: u64 = 0; + for (&t) |*x| { + const t0 = nowNs(); + sink +%= walk.digest(pardes.Pardes, core); + x.* = nowNs() - t0; + } + std.mem.doNotOptimizeAway(sink); + return median(&t); +} + +fn walked(core: *const pardes.Pardes) usize { + var h = walk.Hasher.init(0); + walk.hash(pardes.Pardes, core, &h); + return h.total_len; +} + +fn median(samples: []u64) u64 { + std.mem.sort(u64, samples, {}, std.sort.asc(u64)); + return samples[samples.len / 2]; +} + +fn rssKb() usize { + var ru: libc.rusage = undefined; + _ = libc.getrusage(libc.rusage.SELF, &ru); + return @intCast(ru.maxrss); +} + +const reps = 15; + +pub fn main(init: std.process.Init) !void { + var dir_buf: [256]u8 = undefined; + const tmp = std.mem.span(libc.getenv("TMPDIR") orelse "/tmp"); + const dir = try std.fmt.bufPrint(&dir_buf, "{s}/rollback-bench-{d}", .{ tmp, @as(u32, @intCast(libc.getpid())) }); + try std.Io.Dir.createDirAbsolute(init.io, dir, .default_dir); + defer std.Io.Dir.cwd().deleteTree(init.io, dir) catch {}; + std.debug.print("session panes walked MB blobs MB copy KB walk ms walk ms, blobs out memcpy ms copy ms (est.)\n", .{}); + var big: ?*pardes.Pardes = null; + for (sessions) |s| { + const core = try build(s, dir); + var list: std.ArrayList(*[]u8) = .empty; + defer list.deinit(gpa); + const blob_bytes = try blobs(core, &list); + const all = walked(core); + const walk_ns = walkNs(core); + // The blobs taken out, as a copy pins them: the walk of what a copy + // writes, then put back. + const kept = try gpa.alloc([]u8, list.items.len); + defer gpa.free(kept); + for (list.items, kept) |b, *k| { + k.* = b.*; + b.* = &.{}; + } + const copy_bytes = walked(core); + const copy_walk_ns = walkNs(core); + for (list.items, kept) |b, k| b.* = k; + const src = try gpa.alloc(u8, @max(copy_bytes, 1)); + defer gpa.free(src); + const dst = try gpa.alloc(u8, @max(copy_bytes, 1)); + defer gpa.free(dst); + @memset(src, 1); + @memset(dst, 0); + var t: [reps]u64 = undefined; + for (&t) |*x| { + const t0 = nowNs(); + @memcpy(dst, src); + std.mem.doNotOptimizeAway(dst.ptr); + x.* = nowNs() - t0; + } + const memcpy_ns = median(&t); + var panes: usize = 0; + for (core.panes) |slot| panes += @intFromBool(slot != null); + const ms = 1e-6; + std.debug.print("{s:<8} {d:>5} {d:>9.1} {d:>8.1} {d:>7.1} {d:>7.3} {d:>18.3} {d:>9.3} {d:>14.3}\n", .{ + s.name, panes, + @as(f64, @floatFromInt(all)) / 1e6, @as(f64, @floatFromInt(blob_bytes)) / 1e6, + @as(f64, @floatFromInt(copy_bytes)) / 1e3, @as(f64, @floatFromInt(walk_ns)) * ms, + @as(f64, @floatFromInt(copy_walk_ns)) * ms, @as(f64, @floatFromInt(memcpy_ns)) * ms, + @as(f64, @floatFromInt(copy_walk_ns + memcpy_ns)) * ms, + }); + if (big) |b| b.deinit(); + big = core; + } + // The join writer as a fork child: what fork() holds the authority for, + // and its copy-on-write after, the child alive. The child waits on a + // pipe and exits. + const core = big.?; + defer core.deinit(); + var t: [reps]u64 = undefined; + var cow: [reps]u64 = undefined; + const touch = try gpa.alloc(u8, 64 << 20); + defer gpa.free(touch); + @memset(touch, 1); + for (&t, &cow) |*x, *c| { + var pipe: [2]c_int = undefined; + if (libc.pipe(&pipe) != 0) return error.Pipe; + const t0 = nowNs(); + const pid = libc.fork(); + if (pid == 0) { + var one: [1]u8 = undefined; + _ = libc.read(pipe[0], &one, 1); + libc._exit(0); + } + x.* = nowNs() - t0; + // 64 MB the authority writes while the child holds its pages. + const t1 = nowNs(); + var i: usize = 0; + while (i < touch.len) : (i += 4096) touch[i] +%= 1; + c.* = nowNs() - t1; + _ = libc.close(pipe[0]); + _ = libc.write(pipe[1], "x", 1); + _ = libc.close(pipe[1]); + var status: c_int = 0; + _ = libc.waitpid(pid, &status, 0); + } + std.debug.print("fork of the big session (max RSS {d} MB): {d:.3} ms; then 64 MB written under the child: {d:.3} ms\n", .{ + rssKb() / 1024, @as(f64, @floatFromInt(median(&t))) / 1e6, @as(f64, @floatFromInt(median(&cow))) / 1e6, + }); +} -- cgit v1.3 From 3b57a72becd6dd6700d996c9c5c000e1cb53e076 Mon Sep 17 00:00:00 2001 From: Gabriel Schneider Date: Thu, 29 Oct 2026 18:19:13 -0300 Subject: phase 5 part A: the speculation and rollback engine, in one process: a walker copy put back where it was, replicas that apply their own input at once and take the authority's tape as the truth, and a harness of typists over delayed links src/rollback.zig: the replica core's allocator (`Epoch`) knows every allocation, and while a copy is held puts off frees of what was there before and refuses to grow it in place, so what the core held stays where it was. The copy (`Copy`) is the digest's walk saving the bytes of every allocation it reaches, field by field (the derived caches and this process's own too: a safe build's free writes over what it frees, and the frame arena points at what the epoch then frees), but the queues a step drains and the frame statistics; a file's text, its line index and its undo snapshots are pinned instead (made anew at each edit, never written; now freed raw everywhere), checked unchanged in a safe build. Put back, each allocation's bytes go back over it, the epoch's allocations are freed and its put-off frees forgotten; let go, they are made. Tree-sitter's trees, the C heap's, are held the same way (syntax.held_trees). src/replica.zig: an `Authority` records its core's every input to a tape in memory and hands each client's input back as a unit (origin, sequence number, its tape); a `Replica` applies its own input at once, as the authority will (update, then pump), recorded by its own tape (tape.zig's speculation mode), and for each unit in order drops its pending head when the authority recorded the same (inputs, host calls and answers: the fast path), else puts the copy back, replays the units since and this one (every digest checked) and applies its pending inputs again. A host call or a syscall stops a speculation: it is answered harmlessly, undone when it returns, and the input waits for the authority's answer; a clock or environment read is a guess the comparison checks; an emulator is the authority's (terminal.zig's writers stop a speculation; a safe build asserts none changed). The view's events (a resize, the pointer) wait. Messages are bytes: no transport yet. Each client has a view of its own (src/views.zig): the authority goes into it before the client's input (`toView`, taped), the replica's speculation the same, and the copy walks the views not live too (`parked`, each pane's `pv`). test/replicas.zig (`zig build replicas`): one authority and N replicas, a view each (`--shared` puts them in one), each client typing monkey.zig's mix at 50-200 ms over its own link (half the RTT plus 10% jitter, in order), the authority's pty echoing 2 ms later; gates: no replica diverges, every replica's digest is the authority's at quiescence; reported per RTT: inputs applied at once, stopped, waiting, fast paths, rollbacks (mispredicted, foreign), and what a speculative apply, a copy and a rollback cost. 3 clients x 100 inputs at 30, 80 and 207 ms: no divergence over 12 Debug seeds and 20 ReleaseFast ones with a filled heap; ReleaseFast p99 a speculative apply under 0.2 ms, a rollback under 1.6 ms. `zig build rollback-bench` times the real copy now (take, release, an edit, restore). What the replicas found, fixed: a release build reads what the heap had where Debug's fill made two cores agree: `col_panes` and `restore_buf` digested undefined (zeroed now), `terminal.State.reply_dropped` never set at create, `Text.sels` and `PendingPipe.ranges` digested past their length (a walker `len` cuts a slice too). PARDES_REPLICA_FILL fills a replica's allocations to show what is left. Co-Authored-By: Claude Opus 5.5 --- build.zig | 10 + src/File.zig | 14 +- src/Pipe.zig | 2 + src/Text.zig | 6 +- src/memory.zig | 9 + src/pardes.zig | 11 +- src/replica.zig | 578 ++++++++++++++++++++++++++++++++++++++++++++++++ src/rollback.zig | 496 +++++++++++++++++++++++++++++++++++++++++ src/syntax.zig | 24 ++ src/tape.zig | 236 ++++++++++++++++++-- src/terminal.zig | 17 ++ src/walk.zig | 24 +- test/replicas.zig | 327 +++++++++++++++++++++++++++ test/rollback_bench.zig | 120 ++++------ 14 files changed, 1763 insertions(+), 111 deletions(-) create mode 100644 src/replica.zig create mode 100644 src/rollback.zig create mode 100644 test/replicas.zig (limited to 'test/rollback_bench.zig') diff --git a/build.zig b/build.zig index 13cb8ebc..1a5dd200 100644 --- a/build.zig +++ b/build.zig @@ -1504,6 +1504,16 @@ pub fn build(b: *std.Build) void { run_rollback_bench.has_side_effects = true; b.step("rollback-bench", "a replica's rollback copy (walked state, pinned blobs, walk time) for a typical and a big session, and a fork of the big one (-Doptimize=ReleaseFast)").dependOn(&run_rollback_bench.step); + const replicas = b.addExecutable(.{ + .name = "pardes-replicas", + .root_module = b.createModule(.{ .target = target, .optimize = optimize, .root_source_file = b.path("test/replicas.zig"), .link_libc = true }), + }); + replicas.root_module.addImport("pardes", core_module); + const run_replicas = b.addRunArtifact(replicas); + run_replicas.addPassthruArgs(); + run_replicas.has_side_effects = true; + b.step("replicas", "one authority and several replicas in process, each client typing over a delayed link: digests at quiescence, rollback rate and costs per RTT (-- [--seeds N] [--steps M] [--clients C] [--rtt 30,80,207])").dependOn(&run_replicas.step); + const locations_bench_step = b.step("locations-bench", "measure cold/warm result formatting (-Doptimize=ReleaseFast; CSV on stderr)"); if (optimize != .fast) { locations_bench_step.dependOn(&b.addFail("locations-bench requires -Doptimize=ReleaseFast").step); diff --git a/src/File.zig b/src/File.zig index dd6b2d34..84222191 100644 --- a/src/File.zig +++ b/src/File.zig @@ -1159,10 +1159,12 @@ pub fn deinit(p: *Pardes, pane: *Pane, file: *State) void { if (p.panes[id].?.file.?.line_starts.ptr == file.line_starts.ptr) break true; } else false; p.gpa.free(file.path); - if (ntwins == 0) p.gpa.free(file.content); + // Raw, as every text is freed (a replica's rollback copy pins them, + // src/rollback.zig, and a safe build's free writes over what it frees). + if (ntwins == 0) pardes.memory.freeUnfilled(p.gpa, file.content); if (file.listing) |listing| p.gpa.free(listing.names); if (file.mini) |*mini| mini.deinit(p.gpa); - if (file.line_starts.len > 0 and !index_shared) p.gpa.free(file.line_starts); + if (file.line_starts.len > 0 and !index_shared) pardes.memory.freeRaw(p.gpa, file.line_starts); if (file.highlights.len > 0) p.tree_sitter_gpa.free(file.highlights); p.gpa.free(file.pager_spans); locations.freeRows(p.gpa, file.location_rows); @@ -1264,7 +1266,7 @@ pub fn setContentSpan(p: *Pardes, f: *State, new: []u8, span_in: ?Span) void { }; // The clock the session that recorded read (src/tape.zig). f.mtime = pardes.tape.value(.mtime, pardes.ctlfs.events.now, .{}); - if (f.line_starts.len > 0) p.gpa.free(f.line_starts); + if (f.line_starts.len > 0) pardes.memory.freeRaw(p.gpa, f.line_starts); f.line_starts = starts; if (f.highlights.len > 0) p.tree_sitter_gpa.free(f.highlights); f.highlights = &.{}; @@ -1276,7 +1278,7 @@ pub fn setContentSpan(p: *Pardes, f: *State, new: []u8, span_in: ?Span) void { const tf = &t.file.?; // Its own index, built while the shared one was empty, goes; the // shared one, freed above, is replaced. - if (tf.line_starts.len > 0 and tf.line_starts.ptr != old_starts) p.gpa.free(tf.line_starts); + if (tf.line_starts.len > 0 and tf.line_starts.ptr != old_starts) pardes.memory.freeRaw(p.gpa, tf.line_starts); tf.content = f.content; tf.line_starts = f.line_starts; tf.revision = f.revision; @@ -1498,8 +1500,8 @@ pub fn relink(p: *Pardes, of: usize, id: usize, dot: ?[2]u32) void { sf.twin = p.next_twin; p.next_twin +%= 1; } - p.gpa.free(f.content); - if (f.line_starts.len > 0) p.gpa.free(f.line_starts); + pardes.memory.freeUnfilled(p.gpa, f.content); + if (f.line_starts.len > 0) pardes.memory.freeRaw(p.gpa, f.line_starts); f.history.destroy(p.gpa); f.content = sf.content; f.line_starts = sf.line_starts; diff --git a/src/Pipe.zig b/src/Pipe.zig index 850f9c64..239a94e4 100644 --- a/src/Pipe.zig +++ b/src/Pipe.zig @@ -52,6 +52,8 @@ pub fn pipeMarker(how: modal.Normal.PipeBehavior) []const u8 { } pub const PendingPipe = struct { + /// The digest's walk (src/walk.zig) of the ranges, as far as they go. + pub const replica = .{ .len = .{.{ "ranges", "nranges" }} }; id: u32, pane: usize, serial: u32, diff --git a/src/Text.zig b/src/Text.zig index 3f107542..da4be590 100644 --- a/src/Text.zig +++ b/src/Text.zig @@ -49,9 +49,9 @@ pub const SelRange = struct { sticky: i32 = -1, }; -/// The replay digest's walk (src/walk.zig) of the histories: as far as -/// they go. -pub const replica = .{ .len = .{ .{ "ed_undo", "ed_undo_len" }, .{ "ed_redo", "ed_redo_len" } } }; +/// The replay digest's walk (src/walk.zig) of the histories and the ranges: +/// as far as they go (the ranges' room is made whole and used `nsel` deep). +pub const replica = .{ .len = .{ .{ "ed_undo", "ed_undo_len" }, .{ "ed_redo", "ed_redo_len" }, .{ "sels", "nsel" } } }; what: What = .body, /// The characters of a text that holds its own (gpa-owned): everything but diff --git a/src/memory.zig b/src/memory.zig index 47bd755d..1d082964 100644 --- a/src/memory.zig +++ b/src/memory.zig @@ -421,6 +421,15 @@ pub fn freeUnfilled(allocator: Allocator, bytes: []const u8) void { if (bytes.len > 0) allocator.rawFree(@constCast(bytes), .of(u8), @returnAddress()); } +/// `freeUnfilled` for a slice of any element: a text's line index, made +/// anew at every edit too (and pinned by a replica's rollback copy, +/// src/rollback.zig, which a safe build's overwrite would spoil). +pub fn freeRaw(allocator: Allocator, items: anytype) void { + const E = @typeInfo(@TypeOf(items)).pointer.child; + const bytes = std.mem.sliceAsBytes(items); + if (bytes.len > 0) allocator.rawFree(@constCast(bytes), .of(E), @returnAddress()); +} + /// Gives free memory back to the system, the arenas' and libc's, inside its /// heap too: after a big result is formatted or a big pane torn down, the /// tree-sitter arena kept 73 MiB of emptied buckets resident and glibc what diff --git a/src/pardes.zig b/src/pardes.zig index 75e91ee9..5d889c86 100644 --- a/src/pardes.zig +++ b/src/pardes.zig @@ -65,6 +65,8 @@ pub const tape = @import("tape.zig"); pub const walk = @import("walk.zig"); pub const screen_state = if (terminal_panes) @import("screen_state.zig") else struct {}; pub const views = @import("views.zig"); +pub const rollback = @import("rollback.zig"); +pub const replica = @import("replica.zig"); pub const image = @import("image.zig"); pub const dump = @import("dump.zig"); pub const lsp = @import("lsp/lsp.zig"); @@ -474,6 +476,8 @@ pub const Loc = struct { }; test { + _ = @import("rollback.zig"); + _ = @import("replica.zig"); _ = @import("modal.zig"); _ = @import("tagline.zig"); _ = @import("hunks.zig"); @@ -4412,7 +4416,10 @@ pub const Pardes = struct { // layout: columns own x by weight; panes own y by vweight within a column. ncol: usize = 0, col_weight: [MAX_COLS]u64 = @splat(column_weight_unit), - col_panes: [MAX_COLS][MAX_PANES]usize = undefined, + // Zeroed, not undefined: a digest takes the whole of it, a release + // build's heap leaves its unused slots whatever was there before, and + // two replicas must agree on them (src/replica.zig). + col_panes: [MAX_COLS][MAX_PANES]usize = @splat(@splat(0)), col_n: [MAX_COLS]usize = @splat(0), /// Each column's tag and the workspace's: Texts like a pane's tag, with /// no prefix, showing a default until someone edits them. @@ -4698,7 +4705,7 @@ pub const Pardes = struct { /// a Restore builtin wants this dump loaded into the live session; the /// shell consumes it via takeRestore each frame (restore contents stay host-fed) restore_req: ?[]const u8 = null, - restore_buf: [1024]u8 = undefined, + restore_buf: [1024]u8 = @splat(0), surface: Surface = .{}, /// The view the fields above are of: whose layout, focus, pointer and diff --git a/src/replica.zig b/src/replica.zig new file mode 100644 index 00000000..42d94503 --- /dev/null +++ b/src/replica.zig @@ -0,0 +1,578 @@ +//! A session's replicas (docs/next-daemon.md, phase 5 part A): the authority +//! orders every input and records what its core consumed (src/tape.zig); +//! each replica runs a core of its own, applies its own inputs at once, and +//! takes the authority's stream as the truth. +//! +//! A replica's state is the confirmed position plus its own pending inputs, +//! applied on top (a speculation, recorded as the authority would record +//! it). For each unit of the authority's stream, in order: +//! - its own pending head, recorded the same (inputs, host calls and +//! answers): the head goes, nothing is redone (the fast path); +//! - else the rollback copy is put back (src/rollback.zig), the confirmed +//! units since it are replayed, then this one, and the pending inputs +//! are applied again. +//! A speculation that reaches a host call or a syscall (a Host function, +//! a `tape.site`) or an emulator (the authority's) stops: the copy is put +//! back and the input waits for the authority's answer, one RTT late but +//! never wrong. Clock and environment reads (`tape.value`) are made here, +//! and the comparison catches a guess the authority did not share. +//! +//! No transport: messages are bytes in and out. A client's input is its +//! id, a sequence number and the event as detached/wire.zig frames it; a +//! unit is the input's origin (0 for the authority's own: a pty's output, a +//! tick), its sequence number and the authority's tape of it. +const std = @import("std"); +const builtin = @import("builtin"); +const pardes = @import("pardes.zig"); +const tape = @import("tape.zig"); +const rollback = @import("rollback.zig"); +const walk = @import("walk.zig"); +const wire = @import("detached/wire.zig"); +const Allocator = std.mem.Allocator; +const Pardes = pardes.Pardes; + +const checked = builtin.mode == .debug or builtin.mode == .safe; +const head_len = 6; // origin or client: u16le, seq: u32le +/// Confirmed units fast-pathed under one copy before it is taken afresh: +/// a rollback replays them all. +const rebase_after = 32; + +fn header(buf: []u8, who: u16, seq: u32) void { + std.mem.writeInt(u16, buf[0..2], who, .little); + std.mem.writeInt(u32, buf[2..6], seq, .little); +} + +pub const Unit = struct { origin: u16, seq: u32, tape: []const u8 }; + +pub fn parseUnit(msg: []const u8) ?Unit { + if (msg.len < head_len) return null; + return .{ .origin = std.mem.readInt(u16, msg[0..2], .little), .seq = std.mem.readInt(u32, msg[2..6], .little), .tape = msg[head_len..] }; +} + +pub const Input = struct { client: u16, seq: u32, event: pardes.Event }; + +/// A client's input message; its event borrows `msg`. +pub fn parseInput(msg: []const u8) ?Input { + if (msg.len < head_len) return null; + const f = (wire.framed(msg[head_len..]) catch return null) orelse return null; + const ev = switch (wire.decodeClient(f.tag, f.payload) catch return null) { + .event => |ev| ev, + else => return null, + }; + return .{ .client = std.mem.readInt(u16, msg[0..2], .little), .seq = std.mem.readInt(u32, msg[2..6], .little), .event = ev }; +} + +/// The core whose tape is `t`, stepped: tape.on is process-wide, and so +/// are syntax.zig's held trees, so one process can step several. +const Stepping = struct { + tape: ?*tape.Tape, + trees: ?*pardes.syntax.HeldTrees, + + fn begin(t: *tape.Tape, epoch: ?*rollback.Epoch) Stepping { + const was: Stepping = .{ .tape = tape.on, .trees = pardes.syntax.held_trees }; + tape.on = t; + rollback.Epoch.select(epoch); + return was; + } + + fn end(was: Stepping) void { + tape.on = was.tape; + pardes.syntax.held_trees = was.trees; + } +}; + +fn nowNs() u64 { + return pardes.FrameStats.nowNs(); +} + +// ---- the authority ---- + +pub const Authority = struct { + gpa: Allocator, + t: tape.Tape, + core: *Pardes, + host: pardes.Host, + out: std.ArrayListUnmanaged(u8) = .empty, + seq: u32 = 0, + /// Each client's view (src/views.zig): its own layout, focus and + /// cursors over the session's text. The first client has the session's. + view_of: [64]u8 = @splat(0), + + /// A session booted with `opts`; `unit` is its first, which a replica + /// is made from. + pub fn init(gpa: Allocator, options: pardes.Options, host: pardes.Host) !*Authority { + // A replay gives its core every allocator (tape.zig), and a digest + // tells an allocator from none. + var opts = options; + opts.image_allocator = opts.image_allocator orelse gpa; + opts.pdf_allocator = opts.pdf_allocator orelse gpa; + opts.tree_sitter_allocator = opts.tree_sitter_allocator orelse gpa; + opts.frame_allocator = opts.frame_allocator orelse gpa; + const a = try gpa.create(Authority); + errdefer gpa.destroy(a); + a.* = .{ .gpa = gpa, .t = tape.memory(gpa), .core = undefined, .host = host }; + const was = Stepping.begin(&a.t, null); + defer was.end(); + a.core = try Pardes.init(gpa, opts); + a.core.host = host; + a.core.pump(host) catch {}; + return a; + } + + pub fn deinit(a: *Authority) void { + const was = Stepping.begin(&a.t, null); + a.core.deinit(); + was.end(); + tape.release(&a.t); + a.out.deinit(a.gpa); + a.gpa.destroy(a); + } + + /// What was recorded since the last unit, as one (the boot's, a setup's + /// done straight on the core). + pub fn unit(a: *Authority, origin: u16, seq: u32) ![]const u8 { + a.out.clearRetainingCapacity(); + try a.out.appendNTimes(a.gpa, 0, head_len); + header(a.out.items, origin, seq); + try a.out.appendSlice(a.gpa, tape.taken(&a.t)); + tape.clearTaken(&a.t); + return a.out.items; + } + + /// An input of the authority's own (a pty's output, a tick, a setup's + /// command), recorded into the unit being made. + pub fn apply(a: *Authority, ev: pardes.Event) void { + const was = Stepping.begin(&a.t, null); + a.core.update(ev); + a.core.pump(a.host) catch {}; + was.end(); + } + + /// One of the authority's own inputs, as its unit. + pub fn local(a: *Authority, ev: pardes.Event) ![]const u8 { + a.apply(ev); + a.seq += 1; + return a.unit(0, a.seq); + } + + /// A view of its own for `client`, as its unit; which it is. + pub fn openView(a: *Authority, client: u16) !struct { view: u8, unit: []const u8 } { + const was = Stepping.begin(&a.t, null); + const k = a.core.openView() orelse { + was.end(); + return error.TooManyViews; + }; + a.core.pump(a.host) catch {}; + was.end(); + a.view_of[client] = k; + a.seq += 1; + return .{ .view = k, .unit = try a.unit(0, a.seq) }; + } + + /// One client's input, applied in the order it came, in its view: the + /// unit every replica takes. + pub fn take(a: *Authority, msg: []const u8) ![]const u8 { + const in = parseInput(msg) orelse return error.BadInput; + const was = Stepping.begin(&a.t, null); + a.core.toView(a.view_of[in.client]); + a.core.update(in.event); + a.core.pump(a.host) catch {}; + was.end(); + return a.unit(in.client, in.seq); + } +}; + +// ---- a replica ---- + +pub const Stats = struct { + inputs: u32 = 0, + /// Applied at once and clean. + speculated: u32 = 0, + /// Stopped at a host call, a syscall or an emulator: waits one RTT. + stopped: u32 = 0, + /// Not tried: an event of the view's (a resize, the pointer), phase 4's. + unspeculated: u32 = 0, + /// The authority's unit was the pending head, recorded the same. + fast: u32 = 0, + /// The copy put back: the head recorded otherwise (a guess of the clock + /// the authority did not share), or another's unit first. + mispredicted: u32 = 0, + foreign: u32 = 0, + rebases: u32 = 0, + units: u32 = 0, + /// Nanoseconds: one speculative apply, a copy taken, a rollback (the + /// copy put back, the confirmed units replayed, the pending reapplied). + spec_ns: std.ArrayListUnmanaged(u64) = .empty, + copy_ns: std.ArrayListUnmanaged(u64) = .empty, + rollback_ns: std.ArrayListUnmanaged(u64) = .empty, + copy_bytes: u64 = 0, + copies: u32 = 0, + + pub fn deinit(s: *Stats, gpa: Allocator) void { + s.spec_ns.deinit(gpa); + s.copy_ns.deinit(gpa); + s.rollback_ns.deinit(gpa); + } +}; + +const Pending = struct { seq: u32, msg: []u8, spec: ?[]u8 = null }; + +pub const Replica = struct { + gpa: Allocator, + id: u16, + /// Its own view of the session (Authority.openView's). + view: u8 = 0, + /// The core's allocator; its tapes and the engine's own lists are not. + epoch: rollback.Epoch = .{}, + t: tape.Tape, + copy: rollback.Copy = .{}, + pending: std.ArrayListUnmanaged(Pending) = .empty, + /// pending[0..applied] are applied on top of the confirmed position. + applied: usize = 0, + /// The pending input that waits for the authority (it stopped at a host + /// call, or is the view's), and those after it with it. + blocked: ?u32 = null, + /// Confirmed units fast-pathed since the copy was taken. + since: std.ArrayListUnmanaged([]u8) = .empty, + /// Pending inputs to apply again (after a rollback). + dirty: bool = false, + next_seq: u32 = 1, + stats: Stats = .{}, + diverged: ?[]const u8 = null, + out: std.ArrayListUnmanaged(u8) = .empty, + /// Nanoseconds of the rollback being made, until its reapply. + recovering: ?u64 = null, + + /// A replica of the session whose first unit is `first`. + pub fn init(gpa: Allocator, id: u16, first: []const u8) !*Replica { + const r = try gpa.create(Replica); + r.* = .{ .gpa = gpa, .id = id, .t = tape.memory(gpa) }; + r.t.core_gpa = r.epoch.allocator(); + r.confirm(first); + if (r.t.core == null) { + r.deinit(); + return error.NoCore; + } + return r; + } + + pub fn deinit(r: *Replica) void { + if (r.t.core) |c| { + const was = Stepping.begin(&r.t, &r.epoch); + if (r.copy.held) r.copy.restore(); + c.deinit(); + was.end(); + } + r.copy.deinit(); + for (r.pending.items) |p| r.freePending(p); + r.pending.deinit(r.gpa); + for (r.since.items) |u| r.gpa.free(u); + r.since.deinit(r.gpa); + r.out.deinit(r.gpa); + r.stats.deinit(r.gpa); + tape.release(&r.t); + if (r.diverged) |why| r.gpa.free(why); + r.epoch.deinit(); + r.gpa.destroy(r); + } + + pub fn core(r: *const Replica) *Pardes { + return r.t.core.?; + } + + fn freePending(r: *Replica, p: Pending) void { + r.gpa.free(p.msg); + if (p.spec) |s| r.gpa.free(s); + } + + /// The replica's own input: applied at once when it can be, and the + /// message for the authority returned (valid until the next call). + pub fn input(r: *Replica, ev: pardes.Event) ![]const u8 { + r.stats.inputs += 1; + const seq = r.next_seq; + r.next_seq += 1; + const room = try r.gpa.alloc(u8, head_len + wire.clientBound(.{ .event = ev })); + defer r.gpa.free(room); + const n = (try wire.encodeClient(room[head_len..], .{ .event = ev })).len; + header(room, r.id, seq); + const msg = try r.gpa.dupe(u8, room[0 .. head_len + n]); + try r.pending.append(r.gpa, .{ .seq = seq, .msg = msg }); + r.dirty = true; + r.settle(); + r.out.clearRetainingCapacity(); + try r.out.appendSlice(r.gpa, msg); + return r.out.items; + } + + /// One unit of the authority's stream, in its order. + pub fn confirm(r: *Replica, msg: []const u8) void { + if (r.diverged != null) return; + const u = parseUnit(msg) orelse return r.fail("a unit too short to read"); + r.stats.units += 1; + const was = Stepping.begin(&r.t, &r.epoch); + defer was.end(); + if (u.origin == r.id and r.pending.items.len > 0 and r.pending.items[0].seq == u.seq and r.applied > 0 and + tape.sameRecords(r.pending.items[0].spec.?, u.tape)) + { + r.stats.fast += 1; + r.freePending(r.pending.orderedRemove(0)); + r.applied -= 1; + if (r.applied == 0) { + // Nothing speculative on top any more: the state is the + // confirmed one. + r.copy.release(); + rollback.Epoch.select(null); + r.clearSince(); + return; + } + const kept = r.gpa.dupe(u8, u.tape) catch return r.fail("out of memory"); + r.since.append(r.gpa, kept) catch return r.fail("out of memory"); + if (r.since.items.len >= rebase_after) { + r.stats.rebases += 1; + r.toConfirmed(); + r.dirty = true; + } + return; + } + if (r.applied > 0) { + if (u.origin == r.id) r.stats.mispredicted += 1 else r.stats.foreign += 1; + if (u.origin == r.id and std.c.getenv("PARDES_REPLICA_TRACE") != null) { + var w: std.Io.Writer.Allocating = .init(r.gpa); + defer w.deinit(); + w.writer.print("replica {d}: its input {d} was recorded otherwise; it guessed:\n", .{ r.id, u.seq }) catch {}; + tape.describe(r.pending.items[0].spec.?, &w.writer); + w.writer.print("and the authority recorded:\n", .{}) catch {}; + tape.describe(u.tape, &w.writer); + std.debug.print("{s}", .{w.written()}); + } + r.recovering = nowNs(); + r.toConfirmed(); + } + if (tape.replayUnit(&r.t, u.tape)) |why| { + if (std.c.getenv("PARDES_REPLICA_TRACE") != null) { + var w: std.Io.Writer.Allocating = .init(r.gpa); + defer w.deinit(); + w.writer.print("replica {d}: the unit of {d}/{d} diverged: {s}\n", .{ r.id, u.origin, u.seq, why }) catch {}; + tape.describe(u.tape, &w.writer); + std.debug.print("{s}", .{w.written()}); + } + return r.fail(why); + } + if (u.origin == r.id) { + if (r.pending.items.len == 0 or r.pending.items[0].seq != u.seq) return r.fail("the authority confirmed an input this replica did not send next"); + if (r.blocked == u.seq) r.blocked = null; + r.freePending(r.pending.orderedRemove(0)); + } + r.dirty = true; + } + + fn fail(r: *Replica, why: []const u8) void { + if (r.diverged == null) r.diverged = r.gpa.dupe(u8, why) catch return; + } + + fn clearSince(r: *Replica) void { + for (r.since.items) |u| r.gpa.free(u); + r.since.clearRetainingCapacity(); + } + + /// Back to the confirmed position: the copy put back, the units fast- + /// pathed since replayed. + fn toConfirmed(r: *Replica) void { + if (r.copy.held) { + r.copy.restore(); + rollback.Epoch.select(null); + } + r.applied = 0; + for (r.since.items) |u| if (tape.replayUnit(&r.t, u)) |why| return r.fail(why); + r.clearSince(); + for (r.pending.items) |*p| if (p.spec) |s| { + r.gpa.free(s); + p.spec = null; + }; + } + + /// The pending inputs applied again, as far as they go. + pub fn settle(r: *Replica) void { + if (!r.dirty or r.diverged != null) return; + r.dirty = false; + const was = Stepping.begin(&r.t, &r.epoch); + defer was.end(); + while (r.applied < r.pending.items.len and r.blocked != r.pending.items[r.applied].seq) if (!r.speculate()) break; + if (r.recovering) |t0| { + r.stats.rollback_ns.append(r.gpa, nowNs() - t0) catch {}; + r.recovering = null; + } + } + + /// pending[applied] applied on top: whether it was. + fn speculate(r: *Replica) bool { + const p = &r.pending.items[r.applied]; + const in = parseInput(p.msg) orelse return false; + switch (in.event) { + .key, .paste, .command => {}, + else => { + // The view's (a resize, the pointer): the authority's for + // now, phase 4's views to come. + r.stats.unspeculated += 1; + r.blocked = p.seq; + return false; + }, + } + const c = r.core(); + if (!r.copy.held) { + const t0 = nowNs(); + r.copy.take(&r.epoch, c) catch return false; + rollback.Epoch.select(&r.epoch); + r.stats.copy_ns.append(r.gpa, nowNs() - t0) catch {}; + r.stats.copy_bytes += r.copy.savedBytes(); + r.stats.copies += 1; + } + var terminals: [pardes.MAX_PANES]u64 = undefined; + if (checked) terminalDigests(c, &terminals); + const t0 = nowNs(); + tape.beginSpeculation(&r.t); + c.toView(r.view); + c.update(in.event); + c.pump(c.host) catch {}; + const clean = tape.endSpeculation(&r.t); + const dt = nowNs() - t0; + // Where a blob was written, should one be: the input that did. + if (checked and std.c.getenv("PARDES_REPLICA_TRACE") != null) r.copy.check(); + if (!clean) { + // Undone, and the earlier ones applied again on the confirmed + // position: this one waits for the authority. + r.stats.stopped += 1; + r.blocked = p.seq; + const upto = r.applied; + r.toConfirmed(); + while (r.applied < upto) if (!r.speculate()) break; + if (r.applied == 0 and r.copy.held) { + r.copy.release(); + rollback.Epoch.select(null); + } + return false; + } + if (checked) { + var after: [pardes.MAX_PANES]u64 = undefined; + terminalDigests(c, &after); + if (!std.mem.eql(u64, &terminals, &after)) @panic("replica: a speculation wrote a terminal's emulator"); + } + r.stats.speculated += 1; + r.stats.spec_ns.append(r.gpa, dt) catch {}; + p.spec = r.gpa.dupe(u8, tape.taken(&r.t)) catch return false; + tape.clearTaken(&r.t); + r.applied += 1; + return true; + } + + /// Nothing pending and nothing applied: the replica is the authority's + /// confirmed state, which `digest` says. + pub fn quiet(r: *const Replica) bool { + return r.pending.items.len == 0 and r.applied == 0 and !r.copy.held; + } + + pub fn digest(r: *Replica) u64 { + const was = Stepping.begin(&r.t, &r.epoch); + defer was.end(); + return tape.digestOf(r.core()); + } +}; + +fn terminalDigests(p: *const Pardes, out: *[pardes.MAX_PANES]u64) void { + for (p.panes, out) |slot, *d| { + d.* = 0; + const pane = slot orelse continue; + const state = pane.terminal orelse continue; + var h = walk.Hasher.init(0); + state.digest(&h); + d.* = h.final(); + } +} + +pub fn digestOf(a: *Authority) u64 { + const was = Stepping.begin(&a.t, null); + defer was.end(); + return tape.digestOf(a.core); +} + +const TestHost = struct { + fn watched(_: ?*anyopaque) bool { + return false; + } + const vtable: pardes.Host.VTable = .{ .watched = watched }; + fn host() pardes.Host { + return .{ .vtable = &vtable }; + } +}; + +fn key(r: *Replica, cp: u21, sent: *std.ArrayListUnmanaged([]u8)) !void { + const text = [_]u8{@intCast(cp & 0x7f)}; + const msg = try r.input(.{ .key = .{ .cp = cp, .text = if (cp < 0x80 and cp >= 0x20) &text else "" } }); + try sent.append(std.testing.allocator, try std.testing.allocator.dupe(u8, msg)); +} + +test "a replica's own keys apply at once, confirm on the fast path, and end at the authority's digest" { + const gpa = std.testing.allocator; + const a = try Authority.init(gpa, .{ .cols = 80, .rows = 24, .file = "mise.toml" }, TestHost.host()); + defer a.deinit(); + const r = try Replica.init(gpa, 1, try a.unit(0, 0)); + defer r.deinit(); + try std.testing.expectEqual(digestOf(a), r.digest()); + var sent: std.ArrayListUnmanaged([]u8) = .empty; + defer { + for (sent.items) |m| gpa.free(m); + sent.deinit(gpa); + } + for ("oabc") |k| try key(r, k, &sent); + try key(r, pardes.Key.escape, &sent); + try std.testing.expectEqual(@as(usize, 5), r.applied); + for (sent.items) |m| { + r.confirm(try a.take(m)); + r.settle(); + } + if (r.diverged) |why| std.debug.print("diverged: {s}\n", .{why}); + try std.testing.expect(r.diverged == null); + try std.testing.expectEqual(@as(u32, 5), r.stats.fast); + try std.testing.expect(r.quiet()); + try std.testing.expectEqual(digestOf(a), r.digest()); +} + +test "another client's input first: a replica rolls back, replays it and applies its own again" { + const gpa = std.testing.allocator; + const a = try Authority.init(gpa, .{ .cols = 80, .rows = 24, .file = "mise.toml" }, TestHost.host()); + defer a.deinit(); + const first = try gpa.dupe(u8, try a.unit(0, 0)); + defer gpa.free(first); + const r1 = try Replica.init(gpa, 1, first); + defer r1.deinit(); + const r2 = try Replica.init(gpa, 2, first); + defer r2.deinit(); + var s1: std.ArrayListUnmanaged([]u8) = .empty; + var s2: std.ArrayListUnmanaged([]u8) = .empty; + defer { + for (s1.items) |m| gpa.free(m); + for (s2.items) |m| gpa.free(m); + s1.deinit(gpa); + s2.deinit(gpa); + } + for ("oone") |k| try key(r1, k, &s1); + for ("Otwo") |k| try key(r2, k, &s2); + // The authority hears r2 first, then r1; both hear the units in that + // order. + for ([_]*std.ArrayListUnmanaged([]u8){ &s2, &s1 }) |list| for (list.items) |m| { + const u = try gpa.dupe(u8, try a.take(m)); + defer gpa.free(u); + r1.confirm(u); + r1.settle(); + r2.confirm(u); + r2.settle(); + }; + for ([_]*Replica{ r1, r2 }) |r| { + if (r.diverged) |why| std.debug.print("diverged: {s}\n", .{why}); + try std.testing.expect(r.diverged == null); + try std.testing.expect(r.quiet()); + try std.testing.expectEqual(digestOf(a), r.digest()); + } + try std.testing.expect(r1.stats.foreign >= 1); + try std.testing.expectEqual(@as(u32, 4), r2.stats.fast); +} diff --git a/src/rollback.zig b/src/rollback.zig new file mode 100644 index 00000000..9137b6f6 --- /dev/null +++ b/src/rollback.zig @@ -0,0 +1,496 @@ +//! A replica's rollback copy (docs/next-daemon.md, phase 5): the state a +//! speculation can write, saved before it writes, and put back where it was +//! when the authority's order differs. +//! +//! Two halves: +//! - `Epoch`, the replica core's allocator. It knows every allocation, and +//! while a copy is held it frees nothing that was there before (a free +//! is put off) and grows nothing of then in place. What is allocated +//! meanwhile is the epoch's. So whatever the core held when the copy was +//! taken is still where it was when the copy is put back. +//! - `Copy`, the walker's walk of the core's state (src/walk.zig: what a +//! digest covers, so views, buffers, registers, counters and terminal +//! metadata), saving the bytes of every allocation it reaches, the core +//! struct's own included. A file's text, its line index and its undo +//! snapshots are pinned instead: an edit puts a new buffer in their +//! place (File.setContentSpan) and never writes one, so the copy keeps +//! a reference and the put-off free keeps it alive. +//! +//! Putting the copy back writes each saved allocation's bytes over it, so +//! every pointer in them points where it did, at what was there; then the +//! epoch's allocations are freed and its put-off frees forgotten. Letting +//! the copy go (the speculation was right) makes the put-off frees. +//! +//! ponytail: an allocation reached only through a field the walk leaves +//! out (a derived cache's own heap, say) is not saved; that cache is the +//! epoch's as it was written, which is right for one remade from the text +//! it was made of and wrong for one changed in place. The replica harness +//! (test/replicas.zig) checks a digest at every confirmed step. +const std = @import("std"); +const builtin = @import("builtin"); +const pardes = @import("pardes.zig"); +const walk = @import("walk.zig"); +const Allocator = std.mem.Allocator; +const Alignment = std.mem.Alignment; +const Pardes = pardes.Pardes; + +/// Blobs this long or longer (`pinned`) are pinned, not copied. Checked +/// unchanged in a safe build when the copy goes. +const blob_min = 4096; +const checked = builtin.mode == .debug or builtin.mode == .safe; + +pub const Epoch = struct { + child: Allocator = std.heap.c_allocator, + /// Its own bookkeeping, never the epoch's. + meta: Allocator = std.heap.c_allocator, + live: std.AutoHashMapUnmanaged(usize, Entry) = .empty, + on: bool = false, + id: u32 = 0, + born: std.ArrayListUnmanaged(usize) = .empty, + deferred: std.ArrayListUnmanaged(usize) = .empty, + trees: pardes.syntax.HeldTrees = .{ .meta = std.heap.c_allocator }, + + const Entry = struct { len: usize, alignment: Alignment, epoch: u32, freed: bool = false }; + + pub fn allocator(e: *Epoch) Allocator { + return .{ .ptr = e, .vtable = &.{ .alloc = alloc, .resize = resize, .remap = remap, .free = free } }; + } + + pub fn deinit(e: *Epoch) void { + var it = e.live.iterator(); + while (it.next()) |kv| e.child.rawFree(@as([*]u8, @ptrFromInt(kv.key_ptr.*))[0..kv.value_ptr.len], kv.value_ptr.alignment, @returnAddress()); + e.live.deinit(e.meta); + e.born.deinit(e.meta); + e.deferred.deinit(e.meta); + e.trees.born.deinit(e.meta); + e.trees.freed.deinit(e.meta); + } + + fn mine(e: *const Epoch, entry: Entry) bool { + return e.on and entry.epoch == e.id; + } + + /// The live allocation that starts at `addr`. + pub fn owns(e: *const Epoch, addr: usize) ?Entry { + const entry = e.live.get(addr) orelse return null; + return if (entry.freed) null else entry; + } + + fn note(e: *Epoch, addr: usize, len: usize, alignment: Alignment) bool { + e.live.put(e.meta, addr, .{ .len = len, .alignment = alignment, .epoch = if (e.on) e.id else 0 }) catch return false; + if (e.on) e.born.append(e.meta, addr) catch {}; + return true; + } + + /// Every new allocation filled with this: a core reading memory it never + /// wrote shows as a replica's divergence (test/replicas.zig, a release + /// build's heap is otherwise whatever it was). + pub var fill: ?u8 = null; + + fn alloc(ctx: *anyopaque, len: usize, alignment: Alignment, ra: usize) ?[*]u8 { + const e: *Epoch = @ptrCast(@alignCast(ctx)); + const ptr = e.child.rawAlloc(len, alignment, ra) orelse return null; + if (fill) |b| @memset(ptr[0..len], b); + if (!e.note(@intFromPtr(ptr), len, alignment)) { + e.child.rawFree(ptr[0..len], alignment, ra); + return null; + } + return ptr; + } + + fn resize(ctx: *anyopaque, memory: []u8, alignment: Alignment, new_len: usize, ra: usize) bool { + const e: *Epoch = @ptrCast(@alignCast(ctx)); + const entry = e.live.getPtr(@intFromPtr(memory.ptr)) orelse return false; + // What the copy saved stays its size: the core moves instead. + if (e.on and !e.mine(entry.*)) return false; + if (!e.child.rawResize(memory, alignment, new_len, ra)) return false; + entry.len = new_len; + return true; + } + + fn remap(ctx: *anyopaque, memory: []u8, alignment: Alignment, new_len: usize, ra: usize) ?[*]u8 { + const e: *Epoch = @ptrCast(@alignCast(ctx)); + const old = @intFromPtr(memory.ptr); + const entry = e.live.get(old) orelse return null; + if (e.on and !e.mine(entry)) return null; + const ptr = e.child.rawRemap(memory, alignment, new_len, ra) orelse return null; + _ = e.live.remove(old); + e.live.put(e.meta, @intFromPtr(ptr), .{ .len = new_len, .alignment = alignment, .epoch = entry.epoch }) catch {}; + if (e.mine(entry) and @intFromPtr(ptr) != old) e.born.append(e.meta, @intFromPtr(ptr)) catch {}; + return ptr; + } + + fn free(ctx: *anyopaque, memory: []u8, alignment: Alignment, ra: usize) void { + const e: *Epoch = @ptrCast(@alignCast(ctx)); + const addr = @intFromPtr(memory.ptr); + const entry = e.live.getPtr(addr) orelse return e.child.rawFree(memory, alignment, ra); + if (e.on and !e.mine(entry.*)) { + // Held by the copy until it is put back or let go. + if (!entry.freed) e.deferred.append(e.meta, addr) catch {}; + entry.freed = true; + return; + } + e.child.rawFree(memory.ptr[0..entry.len], entry.alignment, ra); + _ = e.live.remove(addr); + } + + fn reallyFree(e: *Epoch, addr: usize) void { + const entry = e.live.get(addr) orelse return; + e.child.rawFree(@as([*]u8, @ptrFromInt(addr))[0..entry.len], entry.alignment, @returnAddress()); + _ = e.live.remove(addr); + } + + fn begin(e: *Epoch) void { + std.debug.assert(!e.on); + e.id +%= 1; + if (e.id == 0) e.id = 1; + e.on = true; + e.born.clearRetainingCapacity(); + e.deferred.clearRetainingCapacity(); + e.trees.born.clearRetainingCapacity(); + e.trees.freed.clearRetainingCapacity(); + pardes.syntax.held_trees = &e.trees; + } + + /// The copy put back: what the epoch made goes, what it freed stays. + fn discard(e: *Epoch) void { + for (e.born.items) |addr| if (e.live.get(addr)) |entry| if (e.mine(entry)) e.reallyFree(addr); + for (e.deferred.items) |addr| if (e.live.getPtr(addr)) |entry| { + entry.freed = false; + }; + for (e.trees.born.items) |t| pardes.syntax.destroyTree(t); + e.end(); + } + + /// The copy let go: what the epoch freed goes now. + fn keep(e: *Epoch) void { + for (e.deferred.items) |addr| e.reallyFree(addr); + for (e.trees.freed.items) |t| pardes.syntax.destroyTree(t); + e.end(); + } + + fn end(e: *Epoch) void { + e.on = false; + e.born.clearRetainingCapacity(); + e.deferred.clearRetainingCapacity(); + e.trees.born.clearRetainingCapacity(); + e.trees.freed.clearRetainingCapacity(); + if (pardes.syntax.held_trees == &e.trees) pardes.syntax.held_trees = null; + } + + /// The trees of whichever core is stepped next: one in-process harness + /// steps several (syntax.held_trees is process-wide). + pub fn select(e: ?*Epoch) void { + pardes.syntax.held_trees = if (e) |x| (if (x.on) &x.trees else null) else null; + } +}; + +pub const Copy = struct { + meta: Allocator = std.heap.c_allocator, + regions: std.ArrayListUnmanaged(Region) = .empty, + bytes: std.ArrayListUnmanaged(u8) = .empty, + seen: std.AutoHashMapUnmanaged(Seen, void) = .empty, + pins: std.ArrayListUnmanaged(Pin) = .empty, + held: bool = false, + epoch: ?*Epoch = null, + /// The field being walked, for a pin's report, and whether it is one + /// of the blobs: a file's text or an undo snapshot's. + where: []const u8 = "", + pinnable: bool = false, + + const Region = struct { addr: usize, off: usize, len: usize }; + const Pin = struct { addr: usize, len: usize, hash: u64, where: []const u8 }; + const Seen = struct { addr: usize, ty: usize }; + pub const Error = Allocator.Error; + + pub fn deinit(c: *Copy) void { + c.regions.deinit(c.meta); + c.bytes.deinit(c.meta); + c.seen.deinit(c.meta); + c.pins.deinit(c.meta); + } + + pub fn savedBytes(c: *const Copy) usize { + return c.bytes.items.len; + } + + pub fn pinnedBytes(c: *const Copy) usize { + var n: usize = 0; + for (c.pins.items) |pin| n += pin.len; + return n; + } + + /// The copy of `p`, whose allocator is `e`'s, taken: from now on what + /// it held stays where it is until the copy is put back or let go. + pub fn take(c: *Copy, e: *Epoch, p: *const Pardes) Error!void { + std.debug.assert(!c.held); + c.regions.clearRetainingCapacity(); + c.bytes.clearRetainingCapacity(); + c.seen.clearRetainingCapacity(); + c.pins.clearRetainingCapacity(); + c.epoch = e; + // A step drains these before it ends: between steps they hold + // nothing, and a copy is taken between steps. + std.debug.assert(p.effects_len == 0 and p.in_len == 0); + if (c.epoch.?.owns(@intFromPtr(p)) == null) @panic("rollback: the core is not its epoch's"); + try c.object(Pardes, p); + e.begin(); + c.held = true; + } + + /// The core as it was when the copy was taken. + pub fn restore(c: *Copy) void { + std.debug.assert(c.held); + c.check(); + for (c.regions.items) |r| @memcpy(@as([*]u8, @ptrFromInt(r.addr))[0..r.len], c.bytes.items[r.off..][0..r.len]); + c.epoch.?.discard(); + c.held = false; + } + + /// The speculation was right: the copy is not wanted. + pub fn release(c: *Copy) void { + std.debug.assert(c.held); + c.check(); + c.epoch.?.keep(); + c.held = false; + } + + pub fn check(c: *const Copy) void { + if (comptime !checked) return; + for (c.pins.items) |pin| if (std.hash.Wyhash.hash(0, @as([*]const u8, @ptrFromInt(pin.addr))[0..pin.len]) != pin.hash) { + const now = @as([*]const u8, @ptrFromInt(pin.addr))[0..pin.len]; + const entry = c.epoch.?.live.get(pin.addr); + std.debug.print("rollback: {s}, {d} bytes at {x}, pinned, was written in place: now {any}; entry {any}, epoch {d}\n", .{ pin.where, pin.len, pin.addr, now[0..32], entry, c.epoch.?.id }); + @panic("rollback: a pinned blob was written in place"); + }; + } + + fn save(c: *Copy, addr: usize, len: usize) Error!void { + if (len == 0) return; + const off = c.bytes.items.len; + try c.bytes.appendSlice(c.meta, @as([*]const u8, @ptrFromInt(addr))[0..len]); + try c.regions.append(c.meta, .{ .addr = addr, .off = off, .len = len }); + } + + /// Whether the value at `addr` of type `T` was reached before. + fn reached(c: *Copy, comptime T: type, addr: usize) Error!bool { + const key: Seen = .{ .addr = addr, .ty = @intFromPtr(@typeName(T).ptr) }; + const got = try c.seen.getOrPut(c.meta, key); + return got.found_existing; + } + + /// The allocation at `addr`, saved once, whatever reaches it. + fn allocation(c: *Copy, addr: usize) Error!?usize { + const entry = c.epoch.?.owns(addr) orelse return null; + if (try c.reached(void, addr)) return entry.len; + try c.save(addr, entry.len); + return entry.len; + } + + fn visit(c: *Copy, comptime T: type, v: *const T) Error!void { + if (comptime !heapIn(T, 0)) return; + switch (@typeInfo(T)) { + .optional => |o| if (v.*) |*payload| try c.visit(o.child, payload), + .array => |a| if (comptime heapIn(a.child, 0)) for (v) |*e| try c.visit(a.child, e), + .pointer => |p| switch (p.size) { + .one => try c.object(p.child, v.*), + .slice => try c.slice(p.child, v.*), + // No length to walk by: its keeper's (a C handle's buffer). + else => {}, + }, + .@"struct" => |s| { + if (comptime walk.machinery(T)) return; + // A terminal: the authority's (its emulator is never written + // by a speculation: terminal.zig's guards). + if (comptime @hasDecl(T, "digest") and @TypeOf(T.digest) == fn (*const T, *walk.Hasher) void) return; + if (s.layout == .@"packed") return; + if (comptime @hasField(T, "items") and @hasField(T, "capacity") and s.field_names.len == 2) { + if (v.capacity == 0) return; + _ = try c.allocation(@intFromPtr(v.items.ptr)); + const E = @typeInfo(@FieldType(T, "items")).pointer.child; + if (comptime heapIn(E, 0)) for (v.items) |*e| try c.visit(E, e); + return; + } + if (comptime @hasDecl(T, "KV") and @hasDecl(T, "Hash") and @hasField(T, "metadata")) { + const m = v.metadata orelse return; + // std's header sits before the metadata, at the start + // of the one allocation. + const header = std.mem.alignForward(usize, 2 * @sizeOf(usize) + @sizeOf(u32), @alignOf(usize)); + _ = try c.allocation(@intFromPtr(m) - header); + var it = v.iterator(); + while (it.next()) |entry| { + try c.visit(@TypeOf(entry.key_ptr.*), entry.key_ptr); + try c.visit(@TypeOf(entry.value_ptr.*), entry.value_ptr); + } + return; + } + if (comptime @hasField(T, "raw") and s.field_names.len == 1 and @hasDecl(T, "load")) return; + if (comptime @hasField(T, "any_future")) return; + if (comptime @hasField(T, "bit_length") and @hasField(T, "masks") and s.field_names.len == 2) { + if (v.bit_length == 0) return; + _ = try c.allocation(@intFromPtr(v.masks) - @sizeOf(@typeInfo(@FieldType(T, "masks")).pointer.child)); + return; + } + inline for (s.field_names, s.field_types, s.field_attrs) |name, F, attr| { + // What a digest covers, the caches made from it and this + // process's own (the views not live are among it): a + // safe build's free writes over what it frees, so what + // the speculation freed would come back as that. Not the + // queues a step drains (undefined between steps), nor the + // authority's. + if (attr.@"comptime" or comptime (annotated(T, "authority", name) or (annotated(T, "local", name) and unsaved(name)))) continue; + if (comptime opaqueTo(F)) continue; + c.where = @typeName(T) ++ "." ++ name; + c.pinnable = comptime pinned(T, name); + if (comptime walk.lengthOf(T, name) != null and @typeInfo(F) == .array) { + const len = comptime walk.lengthOf(T, name).?; + if (comptime heapIn(@typeInfo(F).array.child, 0)) for (@field(v, name)[0..walk.lengthIn(v, len)]) |*e| try c.visit(@typeInfo(F).array.child, e); + } else try c.visit(F, &@field(v, name)); + } + }, + .@"union" => switch (v.*) { + inline else => |*payload| try c.visit(@TypeOf(payload.*), payload), + }, + .error_union => |e| if (v.*) |*payload| try c.visit(e.payload, payload) else |_| {}, + else => {}, + } + } + + fn object(c: *Copy, comptime P: type, ptr: *const P) Error!void { + if (comptime walk.machinery(P)) return; + const addr = @intFromPtr(ptr); + if (try c.reached(P, addr)) return; + if (comptime @hasDecl(P, "digest") and @TypeOf(P.digest) == fn (*const P, *walk.Hasher) void) return; + if (c.epoch.?.owns(addr) != null) { + _ = try c.reached(void, addr); + try c.fields(P, ptr); + } + try c.visit(P, ptr); + } + + /// A struct's bytes, field by field where it says what its fields are: + /// all of them (this process's own too, which the epoch's frees would + /// leave pointing at nothing), but the queues a step drains and the + /// frame statistics (numbers, megabytes of them), and an array only as + /// far as its length says. + fn fields(c: *Copy, comptime T: type, v: *const T) Error!void { + if (comptime @typeInfo(T) != .@"struct" or !@hasDecl(T, "replica")) return c.save(@intFromPtr(v), @sizeOf(T)); + const s = @typeInfo(T).@"struct"; + inline for (s.field_names, s.field_types, s.field_attrs) |name, F, attr| { + if (attr.@"comptime") continue; + if (comptime annotated(T, "local", name) and unsaved(name)) continue; + const at = @intFromPtr(&@field(v, name)); + if (comptime walk.lengthOf(T, name) != null and @typeInfo(F) == .array) { + try c.save(at, walk.lengthIn(v, comptime walk.lengthOf(T, name).?) * @sizeOf(@typeInfo(F).array.child)); + } else if (comptime annotatedStruct(F)) { + try c.fields(F, &@field(v, name)); + } else if (comptime @typeInfo(F) == .array and annotatedStruct(@typeInfo(F).array.child)) { + for (&@field(v, name)) |*e| try c.fields(@typeInfo(F).array.child, e); + } else try c.save(at, @sizeOf(F)); + } + } + + fn slice(c: *Copy, comptime E: type, s: []const E) Error!void { + if (s.len == 0) return; + const addr = @intFromPtr(s.ptr); + if (try c.reached([]const E, addr)) return; + if (comptime !heapIn(E, 0)) if (c.pinnable and s.len * @sizeOf(E) >= blob_min and c.epoch.?.owns(addr) != null) { + const bytes = std.mem.sliceAsBytes(s); + try c.pins.append(c.meta, .{ .addr = addr, .len = bytes.len, .hash = if (checked) std.hash.Wyhash.hash(0, bytes) else 0, .where = c.where }); + return; + }; + _ = try c.allocation(addr); + if (comptime heapIn(E, 0)) for (s) |*e| try c.visit(E, e); + } +}; + +/// The blobs: what an edit puts in place of its predecessor, never writes +/// (File.setContentSpan), and frees raw (memory.freeUnfilled, freeRaw). +fn pinned(comptime T: type, comptime name: []const u8) bool { + const File = pardes.panes.File; + return (T == File.State and (std.mem.eql(u8, name, "content") or std.mem.eql(u8, name, "line_starts"))) or + (T == File.Snapshot and std.mem.eql(u8, name, "content")); +} + +fn annotatedStruct(comptime T: type) bool { + return @typeInfo(T) == .@"struct" and @hasDecl(T, "replica"); +} + +/// The queues every step drains (empty between steps, asserted when a copy +/// is taken: what is in their slots is nothing), and the frame statistics +/// (numbers, and a static name). +fn unsaved(comptime name: []const u8) bool { + return std.mem.eql(u8, name, "effects") or std.mem.eql(u8, name, "in_q") or std.mem.eql(u8, name, "frames"); +} + +fn annotated(comptime T: type, comptime kind: []const u8, comptime name: []const u8) bool { + @setEvalBranchQuota(100_000); + if (!@hasDecl(T, "replica") or !@hasField(@TypeOf(T.replica), kind)) return false; + inline for (@field(T.replica, kind)) |n| if (comptime std.mem.eql(u8, n, name)) return true; + return false; +} + +/// A field no walk sees into (a C handle, a function): its own keeper's. +fn opaqueTo(comptime F: type) bool { + return switch (@typeInfo(F)) { + .optional => |o| opaqueTo(o.child), + .pointer => |p| p.size == .many or p.size == .c or switch (@typeInfo(p.child)) { + .@"opaque", .@"fn" => true, + else => false, + }, + else => false, + }; +} + +/// Whether a `T` can hold a pointer to the heap at all (else a walk of it +/// finds nothing to save). Deep types count as yes. +fn heapIn(comptime T: type, comptime depth: u32) bool { + if (depth > 12) return true; + return switch (@typeInfo(T)) { + .void, .null, .undefined, .bool, .int, .float, .@"enum", .error_set, .comptime_int, .comptime_float, .enum_literal, .type => false, + .optional => |o| heapIn(o.child, depth + 1), + .array => |a| a.len > 0 and heapIn(a.child, depth + 1), + .pointer => true, + .error_union => |e| heapIn(e.payload, depth + 1), + .@"struct" => |s| blk: { + if (s.layout == .@"packed") break :blk false; + for (s.field_types, s.field_attrs) |F, attr| if (!attr.@"comptime" and heapIn(F, depth + 1)) break :blk true; + break :blk false; + }, + .@"union" => |u| blk: { + for (u.field_types) |F| if (heapIn(F, depth + 1)) break :blk true; + break :blk false; + }, + else => true, + }; +} + +test "a copy put back is the core as it was; one let go keeps what was done" { + var e: Epoch = .{ .child = std.testing.allocator }; + defer e.deinit(); + const gpa = e.allocator(); + const p = try Pardes.init(gpa, .{ .cols = 80, .rows = 24, .file = "mise.toml", .image_allocator = gpa, .pdf_allocator = gpa, .tree_sitter_allocator = gpa, .frame_allocator = gpa }); + p.pump(.{}) catch {}; + const before = walk.digest(Pardes, p); + var c: Copy = .{}; + defer c.deinit(); + const typed = "ox y z"; + try c.take(&e, p); + for (typed) |k| p.update(.{ .key = .{ .cp = k, .text = &.{k} } }); + p.update(.{ .key = .{ .cp = pardes.Key.escape } }); + p.pump(.{}) catch {}; + const edited = walk.digest(Pardes, p); + try std.testing.expect(edited != before); + c.restore(); + try std.testing.expectEqual(before, walk.digest(Pardes, p)); + // The same again, let go this time: what was typed stays. + try c.take(&e, p); + for (typed) |k| p.update(.{ .key = .{ .cp = k, .text = &.{k} } }); + p.update(.{ .key = .{ .cp = pardes.Key.escape } }); + p.pump(.{}) catch {}; + try std.testing.expectEqual(edited, walk.digest(Pardes, p)); + c.release(); + try std.testing.expectEqual(edited, walk.digest(Pardes, p)); + p.deinit(); + try std.testing.expectEqual(@as(usize, 0), e.live.count()); +} diff --git a/src/syntax.zig b/src/syntax.zig index 985a9888..858ef6a2 100644 --- a/src/syntax.zig +++ b/src/syntax.zig @@ -381,10 +381,34 @@ pub fn parseTree(path: []const u8, content: []const u8) ?*anyopaque { if (!enabled) return null; const selected = (forExt(std.fs.path.extension(path)) catch return null) orelse return null; const tree = selected.parser.parseString(content, null) orelse return null; + if (held_trees) |h| h.born.append(h.meta, @ptrCast(tree)) catch {}; return @ptrCast(tree); } +/// While a replica's rollback copy is held (src/rollback.zig) a tree parsed +/// is the speculation's, and one freed is put off: trees live on the C +/// heap, which the core's allocator cannot see. Null otherwise. +pub var held_trees: ?*HeldTrees = null; +pub const HeldTrees = struct { + meta: std.mem.Allocator, + born: std.ArrayListUnmanaged(*anyopaque) = .empty, + freed: std.ArrayListUnmanaged(*anyopaque) = .empty, +}; + pub fn freeTree(tree: *anyopaque) void { + if (!enabled) return; + if (held_trees) |h| { + for (h.born.items, 0..) |t, i| if (t == tree) { + _ = h.born.swapRemove(i); + return destroyTree(tree); + }; + h.freed.append(h.meta, tree) catch destroyTree(tree); + return; + } + destroyTree(tree); +} + +pub fn destroyTree(tree: *anyopaque) void { if (!enabled) return; const t: *ts.Tree = @ptrCast(@alignCast(tree)); t.destroy(); diff --git a/src/tape.zig b/src/tape.zig index ec38deac..70609858 100644 --- a/src/tape.zig +++ b/src/tape.zig @@ -127,19 +127,40 @@ pub const Tape = struct { /// PARDES_RECORD_EVERY: a digest after every input, a nested one's too /// (to find the step a divergence starts at), not only the host's last. every: bool = false, + /// A replica applying its own input before the authority has ordered it + /// (src/replica.zig): recorded as the authority would record it, so the + /// two can be compared, but nothing is performed. A host call or a + /// syscall stops it (`poisoned`): what the call returns is the + /// authority's to say, and the input waits for it. A read of the clock + /// or the environment (`value`) is made here, a guess the comparison + /// checks. + speculating: bool = false, + poisoned: bool = false, + /// The allocator of the core a replay makes (a replica's rollback epoch, + /// src/rollback.zig), when it is not the tape's own. + core_gpa: ?std.mem.Allocator = null, + /// Where a replay keeps what a core keeps of an input (its options), + /// for a tape that is let go; else the tape's allocator, for good. + kept_arena: ?std.heap.ArenaAllocator = null, + /// No digests recorded (a speculation's: a replica compares what it + /// was given and what it asked, from a state the authority shares). + no_digest: bool = false, fn record(t: *Tape, kind: Kind, comptime T: type, v: *const T) void { t.scratch.clearRetainingCapacity(); walk.encode(T, v, &t.scratch.writer) catch @panic("tape: out of memory"); const payload = t.scratch.written(); const len: u32 = @intCast(payload.len); - t.buf.append(t.gpa, @intFromEnum(kind)) catch @panic("tape: out of memory"); + t.buf.append(t.gpa, @backingInt(kind)) catch @panic("tape: out of memory"); t.buf.appendSlice(t.gpa, std.mem.asBytes(&len)) catch @panic("tape: out of memory"); t.buf.appendSlice(t.gpa, payload) catch @panic("tape: out of memory"); if (t.buf.items.len > 1 << 20) t.flush(); } fn flush(t: *Tape) void { + // A tape in memory (a replica's, an authority's stream) is taken + // by whoever holds it. + if (t.fd < 0) return; var at: usize = 0; while (at < t.buf.items.len) { const n = libc.write(t.fd, t.buf.items[at..].ptr, t.buf.items.len - at); @@ -222,7 +243,7 @@ pub const Tape = struct { t.wrap(core); }; t.record(.done, void, &{}); - if (t.depth == 0 or t.every) if (t.core) |core| { + if ((t.depth == 0 or t.every) and !t.no_digest) if (t.core) |core| { var parts: Parts = undefined; const d = digestParts(core, &parts); t.record(.digest, []const u64, &d); @@ -239,12 +260,12 @@ pub const Tape = struct { const head = t.in.take(5) catch return null; const len = std.mem.bytesToValue(u32, head[1..5]); const payload = t.in.take(len) catch return null; - return .{ .kind = @enumFromInt(head[0]), .payload = .{ .bytes = payload } }; + return .{ .kind = @fromBackingInt(@intCast(head[0])), .payload = .{ .bytes = payload } }; } fn peek(t: *const Tape) ?Kind { if (t.in.bytes.len - t.in.at < 5) return null; - return @enumFromInt(t.in.bytes[t.in.at]); + return @fromBackingInt(@intCast(t.in.bytes[t.in.at])); } fn diverge(t: *Tape, comptime why: []const u8, args: anytype) void { @@ -374,7 +395,8 @@ pub const Tape = struct { const tag = std.mem.bytesToValue(Tag, (r.take(@sizeOf(Tag)) catch return t.diverge("a short input", .{}))); r.at -= @sizeOf(Tag); const kept = tag == .init or tag == .init_from_dump or tag == .restore; - const input = walk.decode(Input, r, if (kept) t.gpa else t.arena.allocator()) catch return t.diverge("a short input", .{}); + const keep = if (t.kept_arena) |*k| k.allocator() else t.gpa; + const input = walk.decode(Input, r, if (kept) keep else t.arena.allocator()) catch return t.diverge("a short input", .{}); const outer = t.applying; t.applying = switch (input) { .update => |ev| @tagName(ev), @@ -390,13 +412,14 @@ pub const Tape = struct { .init_from_dump => |d| d.opts, else => unreachable, }; - opts.image_allocator = t.gpa; - opts.pdf_allocator = t.gpa; - opts.tree_sitter_allocator = t.gpa; - opts.frame_allocator = t.gpa; + const gpa = t.core_gpa orelse t.gpa; + opts.image_allocator = gpa; + opts.pdf_allocator = gpa; + opts.tree_sitter_allocator = gpa; + opts.frame_allocator = gpa; t.core = switch (input) { - .init => Pardes.init(t.gpa, opts), - .init_from_dump => |d| pardes.dump.initFromDump(t.gpa, opts, d.bytes), + .init => Pardes.init(gpa, opts), + .init_from_dump => |d| pardes.dump.initFromDump(gpa, opts, d.bytes), else => unreachable, } catch |err| return t.diverge("init failed: {t}", .{err}); t.core.?.host = t.hostOf(t.mask); @@ -608,8 +631,21 @@ fn hostCall(comptime i: usize, comptime R: type, ctx: ?*anyopaque, args: anytype } return walk.decode(R, &ret, t.arena.allocator()) catch std.mem.zeroes(R); } - const real = @field(t.real.vtable.*, name).?; const id: u16 = i; + if (t.speculating) { + // Nobody watches a replica's step: its frames are its own, drawn + // outside it (and the authority's host says so too, or the + // comparison fails). + if (comptime std.mem.eql(u8, name, "watched")) { + const r = false; + t.record(.host, u16, &id); + t.record(.ret, R, &r); + return r; + } + t.poisoned = true; + return poisonOf(R); + } + const real = @field(t.real.vtable.*, name).?; t.record(.host, u16, &id); t.seeFields(); const outer = !t.host_side; @@ -639,7 +675,7 @@ pub fn value(comptime s: Site, comptime f: anytype, args: anytype) @typeInfo(@Ty fn call(comptime s: Site, a: ?std.mem.Allocator, comptime f: anytype, args: anytype, comptime yields: bool) @typeInfo(@TypeOf(f)).@"fn".return_type.? { const R = @typeInfo(@TypeOf(f)).@"fn".return_type.?; const t = on orelse return if (yields) out(f, args) else @call(.auto, f, args); - const id: u16 = @as(u16, host_fns.len) + @intFromEnum(s); + const id: u16 = @as(u16, host_fns.len) + @backingInt(s); if (t.replay) { if (!t.expectHost(id, @tagName(s))) { std.debug.print("replay: {s}\n", .{t.diverged orelse "?"}); @@ -651,6 +687,12 @@ fn call(comptime s: Site, a: ?std.mem.Allocator, comptime f: anytype, args: anyt // The host's own (a shell's spawn asking the shell), or one inside a // site's syscall: the outer call is the record. if (t.host_side) return if (yields) out(f, args) else @call(.auto, f, args); + // A syscall a speculating replica must not make: its input waits for + // the authority's answer. + if (yields and t.speculating) { + t.poisoned = true; + return poisonOf(R); + } t.record(.host, u16, &id); t.seeFields(); t.host_side = true; @@ -679,7 +721,173 @@ fn out(comptime f: anytype, args: anytype) @typeInfo(@TypeOf(f)).@"fn".return_ty /// Asked by `turn.yield`: a syscall a replica must never make. pub fn replaying() bool { - return if (on) |t| t.replay else false; + return if (on) |t| t.replay or t.speculating else false; +} + +/// A replica speculating (src/replica.zig): state only the authority +/// changes (a terminal's emulator) is left alone, and the input waits. +pub fn speculating() bool { + return if (on) |t| t.speculating else false; +} + +/// What a speculation reached is the authority's: the input is undone and +/// waits for the authority's order. +pub fn poison() void { + if (on) |t| if (t.speculating) { + t.poisoned = true; + }; +} + +/// A harmless stand-in for what a host call or a syscall would have +/// returned to a speculation that is thrown away once it returns: an +/// error where one can be said, else zeroes. +fn poisonOf(comptime R: type) R { + return switch (@typeInfo(R)) { + .error_union => |e| if (@typeInfo(e.error_set).error_set.error_names) |errors| + (if (errors.len > 0) @field(e.error_set, errors[0]) else std.mem.zeroes(R)) + else + error.Unexpected, + .optional => null, + .void => {}, + .@"union" => |u| inline for (u.field_names, u.field_types) |name, F| { + if (comptime zeroable(F)) break @unionInit(R, name, std.mem.zeroes(F)); + } else @compileError("tape: no harmless value of " ++ @typeName(R)), + else => std.mem.zeroes(R), + }; +} + +fn zeroable(comptime T: type) bool { + return switch (@typeInfo(T)) { + .int, .float, .bool, .optional, .void => true, + .pointer => |p| p.size == .slice, + .array => |a| zeroable(a.child), + .@"struct" => |s| blk: { + for (s.field_types) |F| if (!zeroable(F)) break :blk false; + break :blk true; + }, + else => false, + }; +} + +// ---- a replica's tapes in memory (src/replica.zig) ---- + +/// A tape recorded in memory: an authority's stream, taken unit by unit. +pub fn memory(gpa: std.mem.Allocator) Tape { + return .{ .gpa = gpa, .replay = false, .scratch = .init(gpa), .arena = .init(gpa), .kept_arena = .init(gpa) }; +} + +/// What a tape in memory holds, given back (its core is gone already). +pub fn release(t: *Tape) void { + t.buf.deinit(t.gpa); + t.scratch.deinit(); + t.arena.deinit(); + if (t.kept_arena) |*k| k.deinit(); +} + +/// What was recorded since the last take; the tape's to reuse after. +pub fn taken(t: *Tape) []const u8 { + return t.buf.items; +} + +pub fn clearTaken(t: *Tape) void { + t.buf.clearRetainingCapacity(); +} + +/// The records of `bytes`, top-level inputs and all, applied to the +/// tape's core as the authority applied them, each digest checked; why +/// they diverged, or null. +pub fn replayUnit(t: *Tape, bytes: []const u8) ?[]const u8 { + t.replay = true; + t.speculating = false; + defer t.replay = false; + t.in = .{ .bytes = bytes }; + while (t.diverged == null) { + const kind = t.peek() orelse break; + var rec = t.next().?; + switch (kind) { + .input => { + t.apply(&rec.payload); + _ = t.arena.reset(.retain_capacity); + }, + .set => t.applySet(&rec.payload), + .digest => t.check(&rec.payload), + .host, .ret, .done => t.diverge("a {t} record at the top", .{kind}), + } + } + return t.diverged; +} + +/// Starts recording `t`'s core as a speculation: the host's watched +/// fields as they are now, nothing recorded yet. +pub fn beginSpeculation(t: *Tape) void { + t.replay = false; + t.speculating = true; + t.poisoned = false; + t.no_digest = true; + t.buf.clearRetainingCapacity(); + t.seeFields(); +} + +/// Whether it stayed clean (no host call, no syscall, no authority state). +pub fn endSpeculation(t: *Tape) bool { + t.speculating = false; + t.no_digest = false; + return !t.poisoned; +} + +/// Whether two tapes record the same inputs, host calls and answers, +/// their digests aside. +pub fn sameRecords(a: []const u8, b: []const u8) bool { + var x: walk.Reader = .{ .bytes = a }; + var y: walk.Reader = .{ .bytes = b }; + while (true) { + const ra = nextRecord(&x); + const rb = nextRecord(&y); + if (ra == null or rb == null) return ra == null and rb == null; + if (ra.?.kind != rb.?.kind or !std.mem.eql(u8, ra.?.payload, rb.?.payload)) return false; + } +} + +fn nextRecord(r: *walk.Reader) ?struct { kind: Kind, payload: []const u8 } { + while (true) { + const head = r.take(5) catch return null; + const payload = r.take(std.mem.bytesToValue(u32, head[1..5])) catch return null; + const kind: Kind = @fromBackingInt(@intCast(head[0])); + if (kind != .digest) return .{ .kind = kind, .payload = payload }; + } +} + +/// A tape's records, one a line (kind; an input's or a host call's name), +/// for a divergence's report. +pub fn describe(bytes: []const u8, w: *std.Io.Writer) void { + var r: walk.Reader = .{ .bytes = bytes }; + var depth: usize = 0; + while (true) { + const head = r.take(5) catch return; + const payload = r.take(std.mem.bytesToValue(u32, head[1..5])) catch return; + const kind: Kind = @fromBackingInt(@intCast(head[0])); + if (kind == .done or kind == .ret) depth -|= 1; + w.splatByteAll(' ', depth * 2) catch {}; + switch (kind) { + .input => { + const tag = std.mem.bytesToValue(std.meta.Tag(Input), payload[0..@sizeOf(std.meta.Tag(Input))]); + w.print("input {t}", .{tag}) catch {}; + if (tag == .update) w.print(" {t}", .{std.mem.bytesToValue(std.meta.Tag(pardes.Event), payload[@sizeOf(std.meta.Tag(Input))..][0..@sizeOf(std.meta.Tag(pardes.Event))])}) catch {}; + }, + .host => { + const id = std.mem.bytesToValue(u16, payload[0..2]); + w.print("host {s}", .{if (id < host_fns.len) host_fns[id] else @tagName(@as(Site, @fromBackingInt(@intCast(id - host_fns.len))))}) catch {}; + }, + else => w.print("{t} ({d} bytes)", .{ kind, payload.len }) catch {}, + } + w.writeByte('\n') catch {}; + if (kind == .input or kind == .host) depth += 1; + } +} + +/// The digest a replica and its authority must agree on. +pub fn digestOf(p: *const Pardes) u64 { + return walk.digest(Pardes, p); } /// PARDES_RECORD set: the tape is written from the session's first call. diff --git a/src/terminal.zig b/src/terminal.zig index e737a831..81fae5db 100644 --- a/src/terminal.zig +++ b/src/terminal.zig @@ -417,9 +417,18 @@ pub fn panFor(p: *const Pardes, pane: *const Pane, r: pardes.Rect) u16 { return @intCast(@min(pane.rows - shown, (at + 1) -| shown)); } +/// A replica's speculation leaves an emulator be: it is the authority's +/// (docs/next-daemon.md, phase 5), so the input waits for its order. +fn authorityOnly() bool { + if (!pardes.tape.speculating()) return false; + pardes.tape.poison(); + return true; +} + /// Move the emulator's viewport by `delta` shell rows (negative scrolls back). pub fn scrollGrid(pane: *Pane, delta: i32) void { if (comptime !enabled) return; + if (authorityOnly()) return; const state = pane.terminal orelse return; state.vt.screens.active.scroll(.{ .delta_row = delta }); } @@ -434,6 +443,7 @@ pub fn atBottom(pane: *const Pane) bool { /// Snap the viewport back onto live output. pub fn followOutput(pane: *Pane) void { if (comptime !enabled) return; + if (authorityOnly()) return; const state = pane.terminal orelse return; state.vt.screens.active.scroll(.active); } @@ -442,6 +452,7 @@ pub fn followOutput(pane: *Pane) void { /// a scrollback; the next resize retries with the same numbers. pub fn resizeGrid(pane: *Pane, gpa: std.mem.Allocator, cols: u16, rows: u16) void { if (comptime !enabled) return; + if (authorityOnly()) return; const state = pane.terminal orelse return; if (cols != pane.cols) for (0..pane.sel.len) |button| { pane.clearPointerSelection(button); @@ -643,6 +654,7 @@ pub fn screenTextLen(pane: *Pane) usize { /// Release the emulator's heap. The Pane allocation itself is the core's. pub fn deinitEmulator(pane: *Pane, gpa: std.mem.Allocator) void { if (comptime !enabled) return; + if (authorityOnly()) return; const state = pane.terminal orelse return; if (comptime images_enabled) state.images.deinit(gpa); state.stream.deinit(); @@ -654,6 +666,7 @@ pub fn deinitEmulator(pane: *Pane, gpa: std.mem.Allocator) void { /// Allocate the live emulator half of a terminal pane. Slot ownership, serial /// assignment, and spawn effects remain core lifecycle invariants. pub fn create(gpa: std.mem.Allocator, cols: u16, rows: u16) !*Pane { + if (authorityOnly()) return error.OutOfMemory; const pane = try createDoc(gpa, cols, rows); errdefer gpa.destroy(pane); if (comptime !enabled) return pane; @@ -674,6 +687,9 @@ pub fn create(gpa: std.mem.Allocator, cols: u16, rows: u16) !*Pane { }); state.stream = .init(.{ .handler = .{ .inner = state.vt.vtHandler() }, .allocator = gpa }); state.reply_len = 0; + // Every field its default (create leaves them what the heap had, which + // a digest then reads: two replicas must agree, src/replica.zig). + state.reply_dropped = 0; state.mouse_cell = null; state.images = if (images_enabled) .{} else {}; if (comptime images_enabled) TerminalImages.installPngDecoder(); @@ -900,6 +916,7 @@ pub fn feedOutput(p: *Pardes, pane: *Pane, bytes: []const u8) void { // There are no pty reads at all without an emulator to parse them into. if (comptime !enabled) return; if (pane.terminal == null) return; + if (authorityOnly()) return; const has_positions = pane.ovl != null or pane.body.cur_pinned or pane.body.vsel.active or pane.body.msel.active or pane.body.nsel > 0 or pane.body.last_edit != null or pane.look_at != null or pane.body.ed_undo_len > 0 or pane.body.ed_redo_len > 0 or pointerRowCount(pane) > 0; diff --git a/src/walk.zig b/src/walk.zig index dc1664cb..3e482e9a 100644 --- a/src/walk.zig +++ b/src/walk.zig @@ -42,7 +42,7 @@ pub fn fieldDigests(comptime T: type, v: *const T, out: []u64) []u64 { if (attr.@"comptime" or comptime excluded(T, name)) continue; var h = Hasher.init(0); if (comptime lengthOf(T, name)) |len| { - hashSlice(@typeInfo(F).array.child, @field(v, name)[0..lengthIn(v, len)], &h); + hashSlice(Element(F), @field(v, name)[0..lengthIn(v, len)], &h); } else hash(F, &@field(v, name), &h); out[n] = h.final(); n += 1; @@ -63,7 +63,7 @@ pub fn fieldNames(comptime T: type) []const []const u8 { } /// Types that are machinery, never state: an allocator, the Io, a lock. -fn machinery(comptime T: type) bool { +pub fn machinery(comptime T: type) bool { inline for (.{ std.mem.Allocator, std.Io, std.Io.Mutex, std.Io.Condition, std.Thread, std.heap.ArenaAllocator }) |M| if (T == M) return true; return false; } @@ -92,7 +92,7 @@ fn opaqueTo(comptime F: type) bool { }; } -fn excluded(comptime T: type, comptime name: []const u8) bool { +pub fn excluded(comptime T: type, comptime name: []const u8) bool { @setEvalBranchQuota(100_000); if (!@hasDecl(T, "replica")) return false; inline for (.{ "local", "derived", "authority" }) |kind| { @@ -104,12 +104,22 @@ fn excluded(comptime T: type, comptime name: []const u8) bool { } /// How far the field `len` names says its array is written (null: none). -fn lengthIn(v: anytype, comptime len: []const u8) usize { +/// What a `len`-cut field holds: a fixed array's, or a slice's room (a +/// selection's ranges, made once and used `nsel` deep). +fn Element(comptime F: type) type { + return switch (@typeInfo(F)) { + .array => |a| a.child, + .pointer => |p| p.child, + else => @compileError("walk: a `len` cuts an array or a slice, not " ++ @typeName(F)), + }; +} + +pub fn lengthIn(v: anytype, comptime len: []const u8) usize { const n = @field(v, len); return if (@typeInfo(@TypeOf(n)) == .optional) (if (n) |k| k else 0) else n; } -fn lengthOf(comptime T: type, comptime name: []const u8) ?[]const u8 { +pub fn lengthOf(comptime T: type, comptime name: []const u8) ?[]const u8 { @setEvalBranchQuota(100_000); if (!@hasDecl(T, "replica") or !@hasField(@TypeOf(T.replica), "len")) return null; inline for (T.replica.len) |pair| if (comptime std.mem.eql(u8, pair[0], name)) return pair[1]; @@ -134,7 +144,7 @@ pub fn hash(comptime T: type, v: *const T, h: *Hasher) void { .bool => h.update(&.{@intFromBool(v.*)}), .int => hashInt(T, v.*, h), .float => h.update(std.mem.asBytes(v)), - .@"enum" => |e| hashInt(e.tag_type, @intFromEnum(v.*), h), + .@"enum" => |e| hashInt(e.tag_type, @backingInt(v.*), h), .optional => |o| if (v.*) |*payload| { h.update(&.{1}); hash(o.child, payload, h); @@ -167,7 +177,7 @@ pub fn hash(comptime T: type, v: *const T, h: *Hasher) void { if (attr.@"comptime" or comptime excluded(T, name)) continue; if (comptime opaqueTo(F)) @compileError("walk: " ++ @typeName(T) ++ "." ++ name ++ " is " ++ @typeName(F) ++ ", which no walk sees into: say what it is (`replica`)"); if (comptime lengthOf(T, name)) |len| { - hashSlice(@typeInfo(F).array.child, @field(v, name)[0..lengthIn(v, len)], h); + hashSlice(Element(F), @field(v, name)[0..lengthIn(v, len)], h); } else hash(F, &@field(v, name), h); } }, diff --git a/test/replicas.zig b/test/replicas.zig new file mode 100644 index 00000000..cdb161e8 --- /dev/null +++ b/test/replicas.zig @@ -0,0 +1,327 @@ +//! pardes-replicas: one authority and several replicas of a session in one +//! process (src/replica.zig), each client typing at once over its own +//! delayed, jittered link, as monkey.zig's scripts type at a tty, each in a +//! view of its own (src/views.zig) over the text they share; `--shared` +//! puts them all in one, where they clash at every key: a stress of +//! correctness, not a picture of a session. +//! +//! zig build replicas -Dplatform=tty -- [--seeds N] [--steps M] [--clients C] [--rtt 30,80,207] [--first K] [--shared] +//! +//! Gates, per seed and RTT: no replica diverges (every confirmed unit's +//! digests agree as it is replayed), and at quiescence every replica's +//! digest is the authority's. Reported per RTT: how inputs went (applied at +//! once, stopped at a host call, left to the authority), how units were +//! taken (fast path, rollback for a misprediction or another's input), and +//! what a speculative apply, a copy and a rollback cost against a frame. +const std = @import("std"); +const libc = std.c; +const pardes = @import("pardes"); +const replica = pardes.replica; +const Key = pardes.Key; + +const gpa = std.heap.c_allocator; +var trace = false; +const frame_us = 16_667; + +/// The authority's host: no frames drawn (a replica draws its own), and a +/// pty that echoes what it is written, 2 ms later, as a shell's. +const Host = struct { + echo: std.ArrayListUnmanaged(struct { pane: u8, bytes: []u8 }) = .empty, + + fn watched(_: ?*anyopaque) bool { + return false; + } + + fn ptyWrite(ctx: ?*anyopaque, pane: u8, bytes: []const u8) void { + const h: *Host = @ptrCast(@alignCast(ctx.?)); + h.echo.append(gpa, .{ .pane = pane, .bytes = gpa.dupe(u8, bytes) catch return }) catch {}; + } + + fn ignore(_: ?*anyopaque, _: []const u8) void {} + + const vtable: pardes.Host.VTable = .{ .watched = watched, .pty_write = ptyWrite, .set_clipboard = ignore }; + + fn host(h: *Host) pardes.Host { + return .{ .ctx = h, .vtable = &vtable }; + } +}; + +const Kind = enum { input, to_authority, to_replica, echo }; + +const Msg = struct { + at: u64, + n: u64, + kind: Kind, + who: u16, + bytes: []u8, + + fn order(_: void, a: Msg, b: Msg) std.math.Order { + if (a.at != b.at) return std.math.order(a.at, b.at); + return std.math.order(a.n, b.n); + } +}; + +const Link = struct { + last: u64 = 0, + /// When a message sent at `now` arrives: half the RTT, jitter on top, + /// and never before the one sent ahead of it (a stream keeps order). + fn arrive(l: *Link, random: std.Random, now: u64, one_way: u64, jitter: u64) u64 { + const at = @max(l.last + 1, now + one_way + random.uintLessThan(u64, jitter + 1)); + l.last = at; + return at; + } +}; + +const Totals = struct { + inputs: u64 = 0, + speculated: u64 = 0, + stopped: u64 = 0, + unspeculated: u64 = 0, + units: u64 = 0, + fast: u64 = 0, + mispredicted: u64 = 0, + foreign: u64 = 0, + rebases: u64 = 0, + copies: u64 = 0, + copy_bytes: u64 = 0, + divergences: u64 = 0, + unequal: u64 = 0, + spec_ns: std.ArrayListUnmanaged(u64) = .empty, + copy_ns: std.ArrayListUnmanaged(u64) = .empty, + rollback_ns: std.ArrayListUnmanaged(u64) = .empty, + + fn add(t: *Totals, s: *const replica.Stats) !void { + t.inputs += s.inputs; + t.speculated += s.speculated; + t.stopped += s.stopped; + t.unspeculated += s.unspeculated; + t.units += s.units; + t.fast += s.fast; + t.mispredicted += s.mispredicted; + t.foreign += s.foreign; + t.rebases += s.rebases; + t.copies += s.copies; + t.copy_bytes += s.copy_bytes; + try t.spec_ns.appendSlice(gpa, s.spec_ns.items); + try t.copy_ns.appendSlice(gpa, s.copy_ns.items); + try t.rollback_ns.appendSlice(gpa, s.rollback_ns.items); + } +}; + +/// One client's next input: monkey.zig's mix, as events. +fn nextEvent(random: std.Random, buf: *[64]u8) pardes.Event { + const typed = "abcdefghijklmnopqrstuvwxyz0123456789 -_=+,:;()"; + const normal = "hjklwbe0$xpuyoOiaAIdcv.nN%sC"; + switch (random.uintLessThan(u8, 40)) { + 0...15 => { + buf[0] = typed[random.uintLessThan(usize, typed.len)]; + return .{ .key = .{ .cp = buf[0], .text = buf[0..1] } }; + }, + 16...27 => { + buf[0] = normal[random.uintLessThan(usize, normal.len)]; + return .{ .key = .{ .cp = buf[0], .text = buf[0..1] } }; + }, + 28...31 => return .{ .key = .{ .cp = Key.escape } }, + 32 => return .{ .key = .{ .cp = Key.enter } }, + 33 => return .{ .key = .{ .cp = Key.backspace } }, + 34 => return .{ .key = .{ .cp = switch (random.uintLessThan(u8, 4)) { + 0 => Key.up, + 1 => Key.down, + 2 => Key.left, + else => Key.right, + } } }, + 35 => { + const n = random.intRangeAtMost(usize, 1, 24); + for (buf[0..n]) |*c| c.* = typed[random.uintLessThan(usize, typed.len)]; + return .{ .paste = buf[0..n] }; + }, + 36 => return .{ .command = switch (random.uintLessThan(u8, 4)) { + 0 => "Newcol", + 1 => "Zerox", + 2 => "Undo", + else => "Redo", + } }, + 37 => return .{ .resize = .{ .cols = @intCast(random.intRangeAtMost(u16, 60, 120)), .rows = @intCast(random.intRangeAtMost(u16, 16, 40)) } }, + 38 => return .{ .mouse = .{ .button = .left, .kind = .press, .col = random.uintLessThan(u16, 80), .row = random.uintLessThan(u16, 24) } }, + else => return .{ .mouse = .{ .button = .left, .kind = .release, .col = random.uintLessThan(u16, 80), .row = random.uintLessThan(u16, 24) } }, + } +} + +fn writeFile(path: [:0]const u8, lines: usize) !void { + const fd = libc.open(path, .{ .ACCMODE = .WRONLY, .CREAT = true, .TRUNC = true }, @as(libc.mode_t, 0o644)); + if (fd < 0) return error.OpenFailed; + defer _ = libc.close(fd); + var line: [80]u8 = undefined; + for (0..lines) |n| { + const text = try std.fmt.bufPrint(&line, "const value_{d:0>4} = {d}; // a line to edit\n", .{ n, n *% 7919 }); + if (libc.write(fd, text.ptr, text.len) != @as(isize, @intCast(text.len))) return error.WriteFailed; + } +} + +const Run = struct { seed: u64, rtt_ms: u64, steps: usize, clients: u16, shared: bool }; + +fn run(r: Run, dir: []const u8, totals: *Totals) !void { + if (trace) std.debug.print("seed {d} rtt {d} ms\n", .{ r.seed, r.rtt_ms }); + var prng = std.Random.DefaultPrng.init(r.seed *% 0x9e3779b97f4a7c15 +% r.rtt_ms); + const random = prng.random(); + var host: Host = .{}; + defer { + for (host.echo.items) |e| gpa.free(e.bytes); + host.echo.deinit(gpa); + } + var paths: [3][256:0]u8 = undefined; + for (&paths, 0..) |*p, i| { + const path = try std.fmt.bufPrintSentinel(p, "{s}/f{d}.zig", .{ dir, i }, 0); + try writeFile(path, 60 + 40 * i); + } + // The tty's own boot (a raw shell, a text pane under it), then three + // files looked at: keys go to the last, a pane's shell echoes. + const a = try replica.Authority.init(gpa, .{ .cols = 100, .rows = 30, .boot = .tty }, host.host()); + defer a.deinit(); + const first = try gpa.dupe(u8, try a.unit(0, 0)); + defer gpa.free(first); + var reps: [8]*replica.Replica = undefined; + for (reps[0..r.clients], 1..) |*rp, id| rp.* = try replica.Replica.init(gpa, @intCast(id), first); + defer for (reps[0..r.clients]) |rp| rp.deinit(); + for (&paths) |*path| { + var line: [300]u8 = undefined; + const unit = try a.local(.{ .command = try std.fmt.bufPrint(&line, "Look {s}", .{std.mem.sliceTo(path, 0)}) }); + for (reps[0..r.clients]) |rp| rp.confirm(unit); + } + // Each client a view of its own (src/views.zig), the first the + // session's; or, --shared, every client in the one. + if (!r.shared) for (reps[1..r.clients], 2..) |rp, id| { + const opened = try a.openView(@intCast(id)); + rp.view = opened.view; + for (reps[0..r.clients]) |each| each.confirm(opened.unit); + }; + var queue: std.PriorityQueue(Msg, void, Msg.order) = .empty; + defer queue.deinit(gpa); + var n: u64 = 0; + const one_way = r.rtt_ms * 1000 / 2; + const jitter = one_way / 10; + var up: [8]Link = @splat(.{}); + var down: [8]Link = @splat(.{}); + var left: [8]usize = @splat(r.steps); + // Each client's first input, soon. + for (0..r.clients) |c| { + try queue.push(gpa, .{ .at = random.uintLessThan(u64, 200_000), .n = n, .kind = .input, .who = @intCast(c + 1), .bytes = &.{} }); + n += 1; + } + while (queue.pop()) |m| { + defer if (m.bytes.len > 0) gpa.free(m.bytes); + switch (m.kind) { + .input => { + const c = m.who - 1; + var buf: [64]u8 = undefined; + const ev = nextEvent(random, &buf); + const before = reps[c].stats; + const msg = try gpa.dupe(u8, try reps[c].input(ev)); + if (trace) std.debug.print("{d} ms client {d}: {t} {s} -> {s}\n", .{ m.at / 1000, m.who, ev, switch (ev) { + .key => |k| k.text, + .paste => |t| t, + .command => |t| t, + else => "", + }, if (reps[c].stats.speculated > before.speculated) "applied" else if (reps[c].stats.stopped > before.stopped) "STOPPED" else if (reps[c].stats.unspeculated > before.unspeculated) "view" else "waits" }); + try queue.push(gpa, .{ .at = up[c].arrive(random, m.at, one_way, jitter), .n = n, .kind = .to_authority, .who = m.who, .bytes = msg }); + n += 1; + left[c] -= 1; + if (left[c] > 0) { + // A typist: keys 50-200 ms apart, a pause now and then. + const gap: u64 = if (random.uintLessThan(u8, 12) == 0) random.intRangeAtMost(u64, 300_000, 1_500_000) else random.intRangeAtMost(u64, 50_000, 200_000); + try queue.push(gpa, .{ .at = m.at + gap, .n = n, .kind = .input, .who = m.who, .bytes = &.{} }); + n += 1; + } + }, + .to_authority, .echo => { + const unit = if (m.kind == .echo) + try a.local(.{ .output = .{ .pane = @intCast(m.who), .bytes = m.bytes } }) + else + try a.take(m.bytes); + for (0..r.clients) |c| { + const copy = try gpa.dupe(u8, unit); + try queue.push(gpa, .{ .at = down[c].arrive(random, m.at, one_way, jitter), .n = n, .kind = .to_replica, .who = @intCast(c + 1), .bytes = copy }); + n += 1; + } + for (host.echo.items) |e| { + try queue.push(gpa, .{ .at = m.at + 2_000, .n = n, .kind = .echo, .who = e.pane, .bytes = e.bytes }); + n += 1; + } + host.echo.clearRetainingCapacity(); + }, + .to_replica => { + const rp = reps[m.who - 1]; + rp.confirm(m.bytes); + rp.settle(); + }, + } + } + const want = replica.digestOf(a); + for (reps[0..r.clients], 1..) |rp, id| { + try totals.add(&rp.stats); + if (rp.diverged) |why| { + totals.divergences += 1; + std.debug.print("seed {d} rtt {d} ms: replica {d} diverged: {s}\n", .{ r.seed, r.rtt_ms, id, why }); + } else if (!rp.quiet() or rp.digest() != want) { + totals.unequal += 1; + std.debug.print("seed {d} rtt {d} ms: replica {d} at quiescence: {s}\n", .{ r.seed, r.rtt_ms, id, if (rp.quiet()) "another digest" else "inputs still pending" }); + } + } +} + +fn pct(samples: []u64, p: usize) u64 { + if (samples.len == 0) return 0; + std.mem.sort(u64, samples, {}, std.sort.asc(u64)); + return samples[@min(samples.len - 1, samples.len * p / 100)]; +} + +pub fn main(init: std.process.Init) !void { + const args = try init.minimal.args.toSlice(init.arena.allocator()); + var seeds: u64 = 4; + var first: u64 = 0; + var steps: usize = 120; + var clients: u16 = 3; + var rtts: []const u64 = &.{ 30, 80, 207 }; + var shared = false; + var rtt_buf: [8]u64 = undefined; + var i: usize = 1; + while (i < args.len and std.mem.eql(u8, args[i], "--shared")) : (i += 1) shared = true; + while (i + 1 < args.len) : (i += 2) { + while (i < args.len and std.mem.eql(u8, args[i], "--shared")) : (i += 1) shared = true; + if (i + 1 >= args.len) break; + const name = args[i]; + const value = args[i + 1]; + if (std.mem.eql(u8, name, "--seeds")) seeds = try std.fmt.parseInt(u64, value, 10) else if (std.mem.eql(u8, name, "--first")) first = try std.fmt.parseInt(u64, value, 10) else if (std.mem.eql(u8, name, "--steps")) steps = try std.fmt.parseInt(usize, value, 10) else if (std.mem.eql(u8, name, "--clients")) clients = @min(8, try std.fmt.parseInt(u16, value, 10)) else if (std.mem.eql(u8, name, "--rtt")) { + var it = std.mem.tokenizeScalar(u8, value, ','); + var k: usize = 0; + while (it.next()) |v| : (k += 1) rtt_buf[k] = try std.fmt.parseInt(u64, v, 10); + rtts = rtt_buf[0..k]; + } else return error.UnknownOption; + } + trace = libc.getenv("PARDES_REPLICA_TRACE") != null; + if (libc.getenv("PARDES_REPLICA_FILL")) |b| pardes.rollback.Epoch.fill = std.fmt.parseInt(u8, std.mem.span(b), 0) catch 0xff; + var dir_buf: [256]u8 = undefined; + const tmp = std.mem.span(libc.getenv("TMPDIR") orelse "/tmp"); + const dir = try std.fmt.bufPrint(&dir_buf, "{s}/replicas-{d}", .{ tmp, @as(u32, @intCast(libc.getpid())) }); + try std.Io.Dir.createDirAbsolute(init.io, dir, .default_dir); + defer std.Io.Dir.cwd().deleteTree(init.io, dir) catch {}; + var failed = false; + std.debug.print("{d} seeds x {d} inputs from each of {d} clients, {s}; costs in us (frame {d})\n", .{ seeds, steps, clients, if (shared) "sharing one view" else "a view each", frame_us }); + std.debug.print(" rtt inputs at-once stopped view units fast rollback(mis+foreign) rate rebase spec p50/p99 copy p50/p99 (KB) rollback p50/p99 diverged unequal\n", .{}); + for (rtts) |rtt| { + var t: Totals = .{}; + for (first..first + seeds) |seed| try run(.{ .seed = seed, .rtt_ms = rtt, .steps = steps, .clients = clients, .shared = shared }, dir, &t); + const rollbacks = t.mispredicted + t.foreign; + std.debug.print("{d:>4} {d:>6} {d:>7} {d:>7} {d:>4} {d:>6} {d:>6} {d:>6} ({d}+{d}) {d:>5.1}% {d:>6} {d:>5}/{d:<6} {d:>5}/{d:<5} ({d}) {d:>7}/{d:<7} {d:>8} {d:>7}\n", .{ + rtt, t.inputs, t.speculated, + t.stopped, t.unspeculated, t.units, + t.fast, rollbacks, t.mispredicted, + t.foreign, @as(f64, @floatFromInt(rollbacks)) * 100 / @as(f64, @floatFromInt(@max(t.inputs, 1))), t.rebases, + pct(t.spec_ns.items, 50) / 1000, pct(t.spec_ns.items, 99) / 1000, pct(t.copy_ns.items, 50) / 1000, + pct(t.copy_ns.items, 99) / 1000, t.copy_bytes / @max(t.copies, 1) / 1024, pct(t.rollback_ns.items, 50) / 1000, + pct(t.rollback_ns.items, 99) / 1000, t.divergences, t.unequal, + }); + if (t.divergences > 0 or t.unequal > 0) failed = true; + } + if (failed) std.process.exit(1); +} diff --git a/test/rollback_bench.zig b/test/rollback_bench.zig index 721c6f13..38fe011d 100644 --- a/test/rollback_bench.zig +++ b/test/rollback_bench.zig @@ -1,14 +1,11 @@ -//! What a replica's rollback copy would cost (docs/next-daemon.md, open -//! point 3), and what a fork of the authority stalls it for (open point 5). -//! A typical session and a big one are built in process; for each: -//! - the state the walker visits (src/walk.zig's digest walk, which is -//! what speculation can write: views, buffers, registers, counters, -//! terminal metadata), in bytes; -//! - the part of it that is pinned blobs: a file's text and its undo -//! snapshots, which an edit replaces rather than changes -//! (File.setContentSpan), so a copy keeps a reference; -//! - the walk's time, the blobs' share of it, and a memcpy of the rest: -//! a walker copy costs about the walk plus that write. +//! What a replica's rollback copy costs (docs/next-daemon.md, open point +//! 3, src/rollback.zig), and what a fork of the authority stalls it for +//! (open point 5). A typical session and a big one are built in process, on +//! a rollback epoch's allocator as a replica's core is; for each: +//! - the copy taken and let go: its time, the bytes it saves, the bytes +//! it pins (a file's text, its undo snapshots); +//! - a key that edits a file (`x`) applied on top, as a speculation, and +//! the copy put back: the apply's time and the restore's. //! Then a fork of the big one, child gone at once, and the authority's //! copy-on-write after it. //! @@ -69,8 +66,8 @@ fn writeFile(path: []const u8, lines: usize) !void { } } -fn build(s: Session, dir: []const u8) !*pardes.Pardes { - const core = try pardes.Pardes.init(gpa, .{ .tty_only = true }); +fn build(s: Session, dir: []const u8, a: std.mem.Allocator) !*pardes.Pardes { + const core = try pardes.Pardes.init(a, .{ .tty_only = true, .image_allocator = a, .pdf_allocator = a, .tree_sitter_allocator = a, .frame_allocator = a }); core.update(.{ .resize = .{ .cols = 200, .rows = 60 } }); pump(core); // The shell: pane 0 of a tty boot, its output fed as a pty's. @@ -112,36 +109,6 @@ fn build(s: Session, dir: []const u8) !*pardes.Pardes { return core; } -/// Where a file's text and its undo snapshots are held: what a copy pins. -fn blobs(core: *pardes.Pardes, out: *std.ArrayList(*[]u8)) !usize { - var total: usize = 0; - for (core.panes) |slot| if (slot) |pane| if (pane.file) |*f| { - try out.append(gpa, &f.content); - for (f.history.undo[0..f.history.undo_len]) |*snap| try out.append(gpa, &snap.content); - for (f.history.redo[0..f.history.redo_len]) |*snap| try out.append(gpa, &snap.content); - }; - for (out.items) |b| total += b.len; - return total; -} - -fn walkNs(core: *const pardes.Pardes) u64 { - var t: [reps]u64 = undefined; - var sink: u64 = 0; - for (&t) |*x| { - const t0 = nowNs(); - sink +%= walk.digest(pardes.Pardes, core); - x.* = nowNs() - t0; - } - std.mem.doNotOptimizeAway(sink); - return median(&t); -} - -fn walked(core: *const pardes.Pardes) usize { - var h = walk.Hasher.init(0); - walk.hash(pardes.Pardes, core, &h); - return h.total_len; -} - fn median(samples: []u64) u64 { std.mem.sort(u64, samples, {}, std.sort.asc(u64)); return samples[samples.len / 2]; @@ -161,49 +128,44 @@ pub fn main(init: std.process.Init) !void { const dir = try std.fmt.bufPrint(&dir_buf, "{s}/rollback-bench-{d}", .{ tmp, @as(u32, @intCast(libc.getpid())) }); try std.Io.Dir.createDirAbsolute(init.io, dir, .default_dir); defer std.Io.Dir.cwd().deleteTree(init.io, dir) catch {}; - std.debug.print("session panes walked MB blobs MB copy KB walk ms walk ms, blobs out memcpy ms copy ms (est.)\n", .{}); + std.debug.print("session panes copy KB pinned MB take ms release ms edit ms restore ms\n", .{}); var big: ?*pardes.Pardes = null; - for (sessions) |s| { - const core = try build(s, dir); - var list: std.ArrayList(*[]u8) = .empty; - defer list.deinit(gpa); - const blob_bytes = try blobs(core, &list); - const all = walked(core); - const walk_ns = walkNs(core); - // The blobs taken out, as a copy pins them: the walk of what a copy - // writes, then put back. - const kept = try gpa.alloc([]u8, list.items.len); - defer gpa.free(kept); - for (list.items, kept) |b, *k| { - k.* = b.*; - b.* = &.{}; + var epochs: [sessions.len]*pardes.rollback.Epoch = undefined; + for (sessions, &epochs) |s, *e| { + e.* = try gpa.create(pardes.rollback.Epoch); + e.*.* = .{}; + const core = try build(s, dir, e.*.allocator()); + var c: pardes.rollback.Copy = .{}; + defer c.deinit(); + var take: [reps]u64 = undefined; + var release: [reps]u64 = undefined; + for (&take, &release) |*x, *y| { + const t0 = nowNs(); + try c.take(e.*, core); + const t1 = nowNs(); + c.release(); + x.* = t1 - t0; + y.* = nowNs() - t1; } - const copy_bytes = walked(core); - const copy_walk_ns = walkNs(core); - for (list.items, kept) |b, k| b.* = k; - const src = try gpa.alloc(u8, @max(copy_bytes, 1)); - defer gpa.free(src); - const dst = try gpa.alloc(u8, @max(copy_bytes, 1)); - defer gpa.free(dst); - @memset(src, 1); - @memset(dst, 0); - var t: [reps]u64 = undefined; - for (&t) |*x| { + var edit: [reps]u64 = undefined; + var restore: [reps]u64 = undefined; + for (&edit, &restore) |*x, *y| { + try c.take(e.*, core); const t0 = nowNs(); - @memcpy(dst, src); - std.mem.doNotOptimizeAway(dst.ptr); - x.* = nowNs() - t0; + key(core, 'x', "x"); + const t1 = nowNs(); + c.restore(); + x.* = t1 - t0; + y.* = nowNs() - t1; } - const memcpy_ns = median(&t); var panes: usize = 0; for (core.panes) |slot| panes += @intFromBool(slot != null); const ms = 1e-6; - std.debug.print("{s:<8} {d:>5} {d:>9.1} {d:>8.1} {d:>7.1} {d:>7.3} {d:>18.3} {d:>9.3} {d:>14.3}\n", .{ - s.name, panes, - @as(f64, @floatFromInt(all)) / 1e6, @as(f64, @floatFromInt(blob_bytes)) / 1e6, - @as(f64, @floatFromInt(copy_bytes)) / 1e3, @as(f64, @floatFromInt(walk_ns)) * ms, - @as(f64, @floatFromInt(copy_walk_ns)) * ms, @as(f64, @floatFromInt(memcpy_ns)) * ms, - @as(f64, @floatFromInt(copy_walk_ns + memcpy_ns)) * ms, + std.debug.print("{s:<8} {d:>5} {d:>7.1} {d:>9.1} {d:>7.3} {d:>10.3} {d:>7.3} {d:>10.3}\n", .{ + s.name, panes, + @as(f64, @floatFromInt(c.savedBytes())) / 1e3, @as(f64, @floatFromInt(c.pinnedBytes())) / 1e6, + @as(f64, @floatFromInt(median(&take))) * ms, @as(f64, @floatFromInt(median(&release))) * ms, + @as(f64, @floatFromInt(median(&edit))) * ms, @as(f64, @floatFromInt(median(&restore))) * ms, }); if (big) |b| b.deinit(); big = core; -- cgit v1.3 From fdf74e199de7d0fc648f912875323f3dc920ffc2 Mon Sep 17 00:00:00 2001 From: Gabriel Schneider Date: Thu, 29 Oct 2026 20:56:03 -0300 Subject: replica clients over a socket: the session streams each its dump and every unit after it (replica.zig `Peers`, a writer and a reader thread a client, units at each step boundary, a client's input stepped there in its own view), and `pardes --attach=:` and `--attach=:` (through `ssh -tt -L`) are a client in the tty host, its keys applied at once by its own replica and drawn by its own process wire.zig version 15: welcome (the client's id and view, the dump's length, the session's UTC offset and digest), dump pieces, units (with the authority's clocks), terminal messages both ways, a check. A unit carries one digest, at its end (`unit_digests`). The client: tty.zig `Shell.remote`, `feed` (a terminal's input to the replica and the session); the session's clock guessed from the last unit's (tape.zig `Guess`); a pump woken only by a client's post reads no new time (`mail_only`), and a unit of another's ending where the last ended leaves the speculation as it is (`neutral`); the rollback copy taken while idle (`Replica.ready`); views a frontend here does not draw left undrawn. A lagging client never holds the session: its queue is cut off past 64 MiB (PARDES_REPLICA_LAG_KB). Detach in a client ends that client; a Restore lets every client go. daemon.zig: a replica the daemon hands over is the session's loop's (a note), not a stub thread's. A window session refuses one in words. test/daemon.py `--clients`: two clients (one through the delay relay) type into a detached session and agree with it, one killed outright, one detached, a third typing into a shell. Co-Authored-By: Claude Opus 5.5 --- src/daemon.zig | 61 +--- src/detached/wire.zig | 94 +++++- src/gui/gui.zig | 3 + src/main.zig | 101 ++++++ src/replica.zig | 883 ++++++++++++++++++++++++++++++++++++++++++++++-- src/rollback.zig | 23 +- src/tape.zig | 15 +- src/tty/tty.zig | 425 +++++++++++++++++++++-- test/daemon.py | 153 ++++++++- test/rollback_bench.zig | 18 + 10 files changed, 1659 insertions(+), 117 deletions(-) (limited to 'test/rollback_bench.zig') diff --git a/src/daemon.zig b/src/daemon.zig index 785f4afd..d2ae012c 100644 --- a/src/daemon.zig +++ b/src/daemon.zig @@ -357,19 +357,10 @@ pub fn adopt() void { } /// This session is about to exec another build (upgrade.zig): the daemon -/// is asked to move too, and every replica told why it is let go. -/// ponytail: a pong its thread writes at the same moment may split the -/// words; the client reconnects either way. +/// is asked to move too. Its replica clients were told why they are let go +/// (replica.zig `Peers.refuseAll`). pub fn upgraded() void { tell("upgrade", &.{}); - var out: [600]u8 = undefined; - const said = wire.encodeServer(&out, .{ .refused = "the session is upgrading to another build of pardes: connect again with that build" }) catch return; - for (&replica_fds) |*slot| { - const fd = slot.load(.acquire); - if (fd < 0) continue; - _ = writeAll(fd, said); - _ = libc.shutdown(fd, 2); - } } /// An upgraded session's link and stub as its old process left them: its @@ -388,6 +379,8 @@ pub const Note = union(enum) { /// stub's pipe; for a window session, both -1: open a window. attach: struct { tty: c_int, stub: c_int }, adopted: struct { pids: [max_fds]libc.pid_t, fds: [max_fds]c_int, n: usize }, + /// A replica's connection, its hello taken (src/replica.zig `Peers`). + replica: c_int, }; /// Hands every note the daemon sends on to `deliver`, on a thread of its @@ -416,9 +409,7 @@ fn readNotes(fd: c_int, ctx: *anyopaque, deliver: *const fn (ctx: *anyopaque, no } }); for (got.fds[@min(got.nfds, 2)..got.nfds]) |f| _ = libc.close(f); } else if (std.mem.eql(u8, verb, "replica")) { - if (got.nfds >= 1) { - if (std.Thread.spawn(.{}, serveReplica, .{got.fds[0]})) |t| t.detach() else |_| _ = libc.close(got.fds[0]); - } + if (got.nfds >= 1) deliver(ctx, .{ .replica = got.fds[0] }); for (got.fds[@min(got.nfds, 1)..got.nfds]) |f| _ = libc.close(f); } else if (std.mem.eql(u8, verb, "adopted")) { var note: Note = .{ .adopted = .{ .pids = undefined, .fds = undefined, .n = 0 } }; @@ -450,43 +441,11 @@ fn writeAll(fd: c_int, bytes: []const u8) bool { return true; } -var replica_fds: [max_replicas]std.atomic.Value(c_int) = @splat(.init(-1)); - -/// A replica the daemon handed over, its hello accepted: welcomed, its pings -/// answered, until it says bye or hangs up. ponytail: phase 5 proper makes it -/// a replica (the log, the dump, terminal state); this stub is what times a -/// round trip through the daemon's handoff and `ssh -L`. A message larger -/// than its buffer ends it. -fn serveReplica(fd: c_int) void { - defer _ = libc.close(fd); - // Where an upgrade finds it to say why it ends (`upgraded`). - const slot = for (&replica_fds) |*r| { - if (r.cmpxchgStrong(-1, fd, .acq_rel, .acquire) == null) break r; - } else null; - defer if (slot) |r| r.store(-1, .release); - log.info("a replica connected to '{s}'", .{name()}); - defer log.info("a replica of '{s}' left", .{name()}); - var out: [wire.hello_max]u8 = undefined; - const pid: u32 = @intCast(libc.getpid()); - if (!writeAll(fd, wire.encodeServer(&out, .{ .welcome = .{ .pid = pid, .session = name() } }) catch return)) return; - var buf: [msg_max]u8 = undefined; - var len: usize = 0; - while (true) { - while (wire.framed(buf[0..len]) catch return) |f| { - switch (wire.decodeClient(f.tag, f.payload) catch return) { - .ping => |v| if (!writeAll(fd, wire.encodeServer(&out, .{ .pong = v }) catch return)) return, - .bye, .hello => return, - .event => {}, - } - std.mem.copyForwards(u8, &buf, buf[f.total..len]); - len -= f.total; - } - if (len == buf.len) return; - const n = libc.read(fd, buf[len..].ptr, buf.len - len); - if (n < 0 and libc.errno(n) == .INTR) continue; - if (n <= 0) return; - len += @intCast(n); - } +/// A replica refused by a session that takes none (a window's, so far). +pub fn refuseReplica(fd: c_int, why: []const u8) void { + var out: [600]u8 = undefined; + if (wire.encodeServer(&out, .{ .refused = why })) |bytes| _ = writeAll(fd, bytes) else |_| {} + _ = libc.close(fd); } /// A session ended by a signal (TERM, HUP, INT): its shells go with it, as diff --git a/src/detached/wire.zig b/src/detached/wire.zig index 097cbee7..347d45d1 100644 --- a/src/detached/wire.zig +++ b/src/detached/wire.zig @@ -9,9 +9,11 @@ const builtin = @import("builtin"); const pardes = @import("../pardes.zig"); const config = @import("pardes_config"); -/// 14: the hello names a session and carries the semantics hash, and a -/// ping is answered (phase 5's groundwork); 13: the frames went (phase 3). -pub const version: u16 = 14; +/// 15: the session streams a replica its dump and units, and terminal +/// deltas both ways (phase 5 part B); 14: the hello names a session and +/// carries the semantics hash, and a ping is answered (phase 5's +/// groundwork); 13: the frames went (phase 3). +pub const version: u16 = 15; /// What makes two cores step differently, hashed: a replica's hello carries /// its own, and one of another build is refused. The protocol, the release @@ -52,6 +54,12 @@ pub const ClientTag = enum(u8) { hello = 0x01, bye = 0x02, ping = 0x03, + /// A replica's word about a terminal's state (src/screen_state.zig): the + /// delta it holds, or rows it asks for. Opaque here. + screen_ack = 0x04, + /// What is the session's digest, in stream order (a replica checks + /// itself against it once quiet). + check = 0x05, key = 0x10, mouse = 0x11, @@ -109,14 +117,65 @@ pub const ServerTag = enum(u8) { refused = 0x02, /// A ping's value back. pong = 0x03, + /// A piece of the join's dump (src/dump.zig): the welcome says how long + /// it is whole. + dump = 0x04, + /// A unit of the authority's stream (src/replica.zig): its origin, seq, + /// clocks and tape. + unit = 0x05, + /// A terminal's delta, or rows asked for (src/screen_state.zig), applied + /// before the next unit. Opaque here. + screen = 0x06, + /// The session's digest (u64le), after every unit before it. + check = 0x07, }; pub const ServerMsg = union(ServerTag) { - welcome: struct { pid: u32, session: []const u8 }, + /// Who the client is now (its id and view in the session), how long the + /// dump that follows is, and the session's offset from UTC (a replica + /// guesses the session's time of day). + welcome: struct { pid: u32, session: []const u8, client: u16 = 0, view: u8 = 0, dump_len: u64 = 0, gmtoff: i32 = 0, digest: u64 = 0 }, refused: []const u8, pong: u64, + dump: []const u8, + unit: []const u8, + screen: []const u8, + check: u64, }; +/// A message's header, for a payload sent on its own (a unit, a dump's +/// piece): no copy of it into a message buffer. +pub fn header(tag: u8, len: usize) [header_len]u8 { + var h: [header_len]u8 = undefined; + h[0] = tag; + std.mem.writeInt(u32, h[1..5], @intCast(len), .little); + return h; +} + +/// The other way: what a client is told, its slices borrowing `payload`. +pub fn decodeServer(tag: u8, payload: []const u8) Error!ServerMsg { + var r: std.Io.Reader = .fixed(payload); + const msg: ServerMsg = switch (std.enums.fromInt(ServerTag, tag) orelse return error.BadTag) { + .welcome => .{ .welcome = .{ + .pid = try r.takeInt(u32, .little), + .session = try r.take(try r.takeInt(u16, .little)), + .client = try r.takeInt(u16, .little), + .view = try r.takeByte(), + .dump_len = try r.takeInt(u64, .little), + .gmtoff = try r.takeInt(i32, .little), + .digest = try r.takeInt(u64, .little), + } }, + .refused => .{ .refused = try r.take(try r.takeInt(u16, .little)) }, + .pong => .{ .pong = try r.takeInt(u64, .little) }, + .dump => return .{ .dump = payload }, + .unit => return .{ .unit = payload }, + .screen => return .{ .screen = payload }, + .check => .{ .check = try r.takeInt(u64, .little) }, + }; + try end(&r); + return msg; +} + pub fn encodeServer(out: []u8, msg: ServerMsg) Error![]const u8 { var w: std.Io.Writer = .fixed(out); const at = try beginMessage(&w, @backingInt(std.meta.activeTag(msg))); @@ -124,9 +183,16 @@ pub fn encodeServer(out: []u8, msg: ServerMsg) Error![]const u8 { .welcome => |h| { try w.writeInt(u32, h.pid, .little); try putSlice16(&w, h.session); + try w.writeInt(u16, h.client, .little); + try w.writeByte(h.view); + try w.writeInt(u64, h.dump_len, .little); + try w.writeInt(i32, h.gmtoff, .little); + try w.writeInt(u64, h.digest, .little); }, .refused => |words| try putSlice16(&w, words), .pong => |v| try w.writeInt(u64, v, .little), + .dump, .unit, .screen => |bytes| try w.writeAll(bytes), + .check => |v| try w.writeInt(u64, v, .little), } try finishMessage(&w, at); return w.buffered(); @@ -157,6 +223,8 @@ pub const ClientMsg = union(enum) { bye, /// A round trip's probe, answered `pong` with the same value. ping: u64, + screen_ack: []const u8, + check, event: pardes.Event, }; @@ -366,6 +434,8 @@ pub fn encodeClient(out: []u8, msg: ClientMsg) Error![]const u8 { }, .bye => {}, .ping => |v| try w.writeInt(u64, v, .little), + .screen_ack => |bytes| try w.writeAll(bytes), + .check => {}, .event => |ev| switch (ev) { .key => |k| { try w.writeInt(u32, k.cp, .little); @@ -432,6 +502,8 @@ pub fn decodeClient(tag: u8, payload: []const u8) Error!ClientMsg { } }, .bye => .bye, .ping => .{ .ping = try r.takeInt(u64, .little) }, + .screen_ack => return .{ .screen_ack = payload }, + .check => .check, .key => blk: { const cp = try r.takeInt(u32, .little); // `Key.cp` is a u21, and the specials live in the private-use @@ -492,7 +564,8 @@ const msg_slack = header_len + 64; pub fn clientBound(msg: ClientMsg) usize { return msg_slack + switch (msg) { .hello => |h| h.session.len, - .bye, .ping => 0, + .bye, .ping, .check => 0, + .screen_ack => |bytes| bytes.len, .event => |ev| switch (ev) { .mouse, .resize, .pdf_scroll, .pinch, .touch_scroll, .pointer_leave, .focus => 0, .key => |k| k.text.len, @@ -547,7 +620,10 @@ test "detached wire: every Event variant round-trips" { const bytes = try encodeServer(&buf, .{ .welcome = .{ .pid = 42, .session = "work" } }); const f = (try framed(bytes)).?; try testing.expectEqual(@as(u8, 0x01), f.tag); - try testing.expectEqualSlices(u8, &.{ 42, 0, 0, 0, 4, 0, 'w', 'o', 'r', 'k' }, f.payload); + try testing.expectEqualSlices(u8, &.{ 42, 0, 0, 0, 4, 0, 'w', 'o', 'r', 'k', 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 }, f.payload); + const back = (try decodeServer(f.tag, f.payload)).welcome; + try testing.expectEqualStrings("work", back.session); + try testing.expectEqual(@as(u32, 42), back.pid); try testing.expectEqual(@as(u8, 0x03), (try framed(try encodeServer(&buf, .{ .pong = 7 }))).?.tag); } @@ -622,8 +698,10 @@ test "detached wire: a session is never told to do a frontend's remembering" { // them, and the machine that reports them is the one already holding the // core. const allowed = [_][]const u8{ - // Who this connection is, its round trip, and that it is finished. - "hello", "bye", "ping", + // Who this connection is, its round trip, and that it is finished, + // and which terminal state its replica holds (a protocol's word, + // as a ping is). + "hello", "bye", "ping", "screen_ack", "check", // ...and everything a person can do to a window. "key", "mouse", "resize", "paste", "command", "pdf_scroll", diff --git a/src/gui/gui.zig b/src/gui/gui.zig index e4d69281..a88a7335 100644 --- a/src/gui/gui.zig +++ b/src/gui/gui.zig @@ -4884,6 +4884,9 @@ fn deliverNote(ctx: *anyopaque, note: daemon.Note) void { for (a.pids[0..a.n], a.fds[0..a.n]) |pid, fd| q.push(.{ .adopted = .{ .pid = pid, .fd = fd } }); q.push(.{ .adopted = .{ .pid = 0, .fd = -1 } }); }, + // ponytail: a window session serves no replica yet; a terminal's + // does (tty.zig). + .replica => |fd| daemon.refuseReplica(fd, "a window session takes no replica yet: run the session in a terminal, or --detach"), } } diff --git a/src/main.zig b/src/main.zig index a6144475..ab5047ce 100644 --- a/src/main.zig +++ b/src/main.zig @@ -108,6 +108,13 @@ const help_text = \\ the session detaches or ends \\ --attach= ...the session called (its --9p or \\ --detach name, else its pid) + \\ --attach=: a client of session on another machine, + \\ over `ssh -L` to its pardes-replica.sock + \\ ($PARDES_SSH: the ssh command, words; and + \\ $PARDES_REMOTE_RUNTIME: its runtime directory, + \\ if not this one's path). Its keys show at once + \\ --attach=: ...through that socket (a path, with a '/'), + \\ one forwarded already, or this machine's \\ --geometry=WxH[+X+Y] the window's size in points, and where it goes \\ where the display lets a window say (X11) \\ -h, --help show this help and exit @@ -738,6 +745,10 @@ fn nativeMain(std_init: std.process.Init) !void { opts.config_dir = found.dir; opts.store_path = storePath(init.io, arena, init.environ_map); if (found.dir) |d| @import("crash.zig").setDir(d); + if (attach) |name| if (std.mem.lastIndexOfScalar(u8, name, ':')) |colon| { + if (comptime pardes.platform != .tty) badArgs(init.io, "a window is no replica client yet: run `pardes --attach={s}` in a terminal", .{name}); + remoteAttach(init, opts, name[0..colon], name[colon + 1 ..]); + }; if (attach) |name| if (!resuming) attachTo(init.io, name); if (detach) |name| { // Said plainly, never a Zig error trace: a name a live session's @@ -815,6 +826,96 @@ fn nativeMain(std_init: std.process.Init) !void { daemon.tellStubExit(0); } +extern "c" fn execvp(file: [*:0]const u8, argv: [*:null]const ?[*:0]const u8) c_int; + +/// `pardes --attach=:`: a replica client (tty.zig `remote`), +/// through the socket `where` names, or through `ssh -L` of host `where`'s +/// replica socket to one of this process's. +fn remoteAttach(init: std.process.Init, opts: pardes.Options, where: []const u8, name: []const u8) noreturn { + const io = init.io; + if (comptime pardes.platform != .tty) unreachable; + const Say = struct { + fn no(o: std.Io, comptime fmt: []const u8, args: anytype) noreturn { + var buf: [700]u8 = undefined; + std.Io.File.stderr().writeStreamingAll(o, std.fmt.bufPrint(&buf, "pardes: --attach: " ++ fmt ++ "\n", args) catch "pardes: --attach failed\n") catch {}; + std.process.exit(1); + } + }; + if (std.mem.indexOfScalar(u8, where, '/') != null) { + @import("tty/tty.zig").remote(init, opts, where, name) catch |err| Say.no(io, "{t}", .{err}); + std.process.exit(0); + } + // ssh -L : , until the socket here answers. + var here_buf: [ninep_io.sun_path_len:0]u8 = undefined; + const tmp = std.mem.span(std.c.getenv("TMPDIR") orelse "/tmp"); + const here = std.fmt.bufPrintSentinel(&here_buf, "{s}/pardes-ssh-{d}.sock", .{ tmp, std.c.getpid() }, 0) catch Say.no(io, "TMPDIR is too long for a socket's path", .{}); + var there_buf: [ninep_io.sun_path_len + 64:0]u8 = undefined; + var dir_buf: [ninep_io.sun_path_len:0]u8 = undefined; + const dir = if (std.c.getenv("PARDES_REMOTE_RUNTIME")) |d| std.mem.span(d) else ninep_io.socketDir(&dir_buf) orelse Say.no(io, "no runtime directory to name the remote socket by: set PARDES_REMOTE_RUNTIME", .{}); + var forward_buf: [2 * ninep_io.sun_path_len + 64]u8 = undefined; + const there = std.fmt.bufPrintSentinel(&there_buf, "{s}/pardes-replica.sock", .{dir}, 0) catch unreachable; + const forward = std.fmt.bufPrintSentinel(&forward_buf, "{s}:{s}", .{ here, there }, 0) catch unreachable; + var host_buf: [256]u8 = undefined; + const host = std.fmt.bufPrintSentinel(&host_buf, "{s}", .{where}, 0) catch Say.no(io, "that host's name is too long", .{}); + var argv: [48:null]?[*:0]const u8 = @splat(null); + var n: usize = 0; + var words_buf: [1024]u8 = undefined; + var words = std.mem.tokenizeScalar(u8, std.mem.span(std.c.getenv("PARDES_SSH") orelse "ssh"), ' '); + var at: usize = 0; + while (words.next()) |w| { + if (n >= 40 or at + w.len + 1 > words_buf.len) Say.no(io, "PARDES_SSH has too many words", .{}); + @memcpy(words_buf[at..][0..w.len], w); + words_buf[at + w.len] = 0; + argv[n] = @ptrCast(&words_buf[at]); + n += 1; + at += w.len + 1; + } + // A terminal asked for and a command that waits, not -N: only an + // interactive connection gets TCP_NODELAY from ssh and sshd, and with + // Nagle each key in a burst waited for the last one's ACK, a round trip. + for ([_][*:0]const u8{ "-tt", "-q", "-o", "ExitOnForwardFailure=yes", "-L", forward.ptr, host.ptr, "exec sleep 2147483647" }) |a| { + argv[n] = a; + n += 1; + } + _ = std.c.unlink(here); + const pid = std.c.fork(); + if (pid < 0) Say.no(io, "no fork for ssh", .{}); + if (pid == 0) { + const null_fd = std.c.open("/dev/null", .{ .ACCMODE = .RDONLY }, @as(std.c.mode_t, 0)); + if (null_fd >= 0) _ = std.c.dup2(null_fd, 0); + _ = execvp(argv[0].?, &argv); + std.c._exit(127); + } + const Ssh = struct { + var child: std.c.pid_t = 0; + var path: [:0]const u8 = ""; + fn end() void { + _ = std.c.kill(child, .TERM); + _ = std.c.waitpid(child, null, 0); + _ = std.c.unlink(path); + } + }; + Ssh.child = pid; + Ssh.path = here; + var waited: u32 = 0; + while (std.c.access(here, std.c.F_OK) != 0) : (waited += 1) { + var status: c_int = 0; + if (std.c.waitpid(pid, &status, std.c.W.NOHANG) == pid) Say.no(io, "ssh to {s} ended before its forward was up", .{where}); + if (waited > 3000) { + Ssh.end(); + Say.no(io, "ssh to {s} set up no forward in 30 s", .{where}); + } + const ts: std.c.timespec = .{ .sec = 0, .nsec = 10 * std.time.ns_per_ms }; + _ = std.c.nanosleep(&ts, null); + } + @import("tty/tty.zig").remote(init, opts, here, name) catch |err| { + Ssh.end(); + Say.no(io, "{t}", .{err}); + }; + Ssh.end(); + std.process.exit(0); +} + /// `pardes --attach[=name]` in a terminal: this terminal handed (through /// the daemon) to the detached session of that name, or the only one, and /// this process its stub until the session leaves it. `pardes-gui diff --git a/src/replica.zig b/src/replica.zig index 032e50af..5c965b2b 100644 --- a/src/replica.zig +++ b/src/replica.zig @@ -44,11 +44,34 @@ fn header(buf: []u8, who: u16, seq: u32) void { std.mem.writeInt(u32, buf[2..6], seq, .little); } -pub const Unit = struct { origin: u16, seq: u32, tape: []const u8 }; +/// A unit's head: its origin and seq, then the authority's clocks when it +/// was made (realtime and monotonic, ns), which a replica's guesses of the +/// session's clock start from. +const unit_head = head_len + 16; + +pub const Unit = struct { origin: u16, seq: u32, real_ns: i64 = 0, mono_ns: i64 = 0, tape: []const u8 }; pub fn parseUnit(msg: []const u8) ?Unit { - if (msg.len < head_len) return null; - return .{ .origin = std.mem.readInt(u16, msg[0..2], .little), .seq = std.mem.readInt(u32, msg[2..6], .little), .tape = msg[head_len..] }; + if (msg.len < unit_head) return null; + return .{ + .origin = std.mem.readInt(u16, msg[0..2], .little), + .seq = std.mem.readInt(u32, msg[2..6], .little), + .real_ns = std.mem.readInt(i64, msg[6..14], .little), + .mono_ns = std.mem.readInt(i64, msg[14..22], .little), + .tape = msg[unit_head..], + }; +} + +fn clockNs(id: std.c.clockid_t) i64 { + var ts: std.c.timespec = undefined; + if (std.c.clock_gettime(id, &ts) != 0) return 0; + return @as(i64, ts.sec) * std.time.ns_per_s + ts.nsec; +} + +fn unitHeader(buf: []u8, origin: u16, seq: u32) void { + header(buf, origin, seq); + std.mem.writeInt(i64, buf[6..14], clockNs(.REALTIME), .little); + std.mem.writeInt(i64, buf[14..22], clockNs(.MONOTONIC), .little); } pub const Input = struct { client: u16, seq: u32, event: pardes.Event }; @@ -87,6 +110,13 @@ fn nowNs() u64 { return pardes.FrameStats.nowNs(); } +/// The `p`th percentile of `samples` (sorted here); 100 is the worst. +fn pctOf(samples: []u64, p: usize) u64 { + if (samples.len == 0) return 0; + std.mem.sort(u64, samples, {}, std.sort.asc(u64)); + return samples[@min(samples.len - 1, samples.len * p / 100)]; +} + // ---- the authority ---- pub const Authority = struct { @@ -142,8 +172,8 @@ pub const Authority = struct { /// done straight on the core). pub fn unit(a: *Authority, origin: u16, seq: u32) ![]const u8 { a.out.clearRetainingCapacity(); - try a.out.appendNTimes(a.gpa, 0, head_len); - header(a.out.items, origin, seq); + try a.out.appendNTimes(a.gpa, 0, unit_head); + unitHeader(a.out.items, origin, seq); try a.out.appendSlice(a.gpa, tape.taken(&a.t)); tape.clearTaken(&a.t); return a.out.items; @@ -212,7 +242,7 @@ pub const Authority = struct { const was = Stepping.begin(&a.t, null); a.core.toView(a.view_of[in.client]); a.core.update(in.event); - a.core.pump(a.host) catch {}; + a.core.step(); was.end(); return a.unit(in.client, in.seq); } @@ -437,7 +467,7 @@ pub const Stats = struct { } }; -const Pending = struct { seq: u32, msg: []u8, spec: ?[]u8 = null }; +const Pending = struct { seq: u32, msg: []u8, spec: ?[]u8 = null, sent_ns: u64 = 0 }; pub const Replica = struct { gpa: Allocator, @@ -464,6 +494,26 @@ pub const Replica = struct { out: std.ArrayListUnmanaged(u8) = .empty, /// Nanoseconds of the rollback being made, until its reapply. recovering: ?u64 = null, + /// The digest the last confirmed unit ended on (its last record's + /// bytes): a unit of another's that ends on it again (the session's + /// pump that took this replica's input in, and moved nothing) leaves + /// the speculation on top as it was. + last_digest: std.ArrayListUnmanaged(u8) = .empty, + /// Units of another's taken so, no rollback made. + neutral: u32 = 0, + /// What it shows changed since the host last drew it: a unit replayed + /// on it, a rollback (a unit that confirmed what it applied, or moved + /// nothing, changes nothing to draw). + moved: bool = false, + /// A client of a session elsewhere (`initDump`): the session's clock is + /// guessed, never read here -- its last unit's reading, plus the time + /// since it came, plus a round trip (the input reaches the session that + /// much later). In process, both read one clock. + guess: bool = false, + stamp: struct { real_ns: i64 = 0, mono_ns: i64 = 0, at_ns: u64 = 0 } = .{}, + rtt_ns: u64 = 0, + gmtoff: i32 = 0, + /// Its terminals' screen state, as the authority's deltas made it. screens: [pardes.MAX_PANES]?ScreenOf = @splat(null), screen_out: std.ArrayListUnmanaged(u8) = .empty, @@ -497,6 +547,43 @@ pub const Replica = struct { return r; } + /// A replica of a session elsewhere, as its walker dump (src/dump.zig) + /// is: what a client of a socket joins with. `opts` are this process's + /// (a replica's own allocators and names); its core's allocator is the + /// rollback epoch's, but for images and PDFs, whose workers allocate on + /// threads of their own. ponytail: a picture's pixels freed while a copy + /// is held are not put off; a rollback over one shows an image a frame + /// late at worst... or reads freed pixels, which no test session has. + pub fn initDump(gpa: Allocator, id: u16, view: u8, bytes: []const u8, opts: pardes.Options) !*Replica { + const r = try gpa.create(Replica); + r.* = .{ .gpa = gpa, .id = id, .view = view, .t = tape.memory(gpa), .guess = true }; + r.t.headless = true; + errdefer { + tape.release(&r.t); + r.epoch.deinit(); + gpa.destroy(r); + } + const epoch = r.epoch.allocator(); + r.t.core_gpa = epoch; + var o = opts; + o.frame_allocator = epoch; + o.tree_sitter_allocator = epoch; + o.image_allocator = o.image_allocator orelse std.heap.c_allocator; + o.pdf_allocator = o.pdf_allocator orelse std.heap.c_allocator; + // The session's store and Recent are the authority's. + o.store_path = null; + const c = try pardes.dump.load(epoch, o, bytes, null); + // What loading queued (a watch, a theme asked for) is the + // authority's errand, done there already: the session was dumped + // between steps, with nothing queued. + c.effects_len = 0; + c.effects_head = 0; + c.in_len = 0; + c.in_head = 0; + tape.adoptCore(&r.t, c); + return r; + } + pub fn deinit(r: *Replica) void { for (&r.screens) |*slot| if (slot.*) |*sc| sc.rep.deinit(terminal.emulator_gpa); if (r.t.core) |c| { @@ -512,6 +599,7 @@ pub const Replica = struct { r.since.deinit(r.gpa); r.out.deinit(r.gpa); r.screen_out.deinit(r.gpa); + r.last_digest.deinit(r.gpa); r.stats.deinit(r.gpa); tape.release(&r.t); if (r.diverged) |why| r.gpa.free(why); @@ -539,7 +627,7 @@ pub const Replica = struct { const n = (try wire.encodeClient(room[head_len..], .{ .event = ev })).len; header(room, r.id, seq); const msg = try r.gpa.dupe(u8, room[0 .. head_len + n]); - try r.pending.append(r.gpa, .{ .seq = seq, .msg = msg }); + try r.pending.append(r.gpa, .{ .seq = seq, .msg = msg, .sent_ns = nowNs() }); r.dirty = true; r.settle(); if (r.t.screen_state and r.diverged == null) r.echo.typed(r.core(), r.view, ev, seq, r.applied == r.pending.items.len); @@ -554,6 +642,21 @@ pub const Replica = struct { const u = parseUnit(msg) orelse return r.fail("a unit too short to read"); r.stats.units += 1; if (u.origin == r.id) if (r.t.core) |c| r.echo.confirmed(c, u.seq); + // A copy taken ahead (`ready`), nothing on top: the unit moves the + // state it was of. + if (r.applied == 0 and r.copy.held) { + r.copy.release(); + rollback.Epoch.select(null); + r.clearSince(); + } + if (r.guess) { + r.stamp = .{ .real_ns = u.real_ns, .mono_ns = u.mono_ns, .at_ns = nowNs() }; + if (u.origin == r.id) for (r.pending.items) |p| if (p.seq == u.seq) { + // Sent to confirmed: what reaching the session takes, and back. + if (p.sent_ns != 0) r.rtt_ns = nowNs() -| p.sent_ns; + break; + }; + } const was = Stepping.begin(&r.t, &r.epoch); defer was.end(); if (u.origin == r.id and r.pending.items.len > 0 and r.pending.items[0].seq == u.seq and r.applied > 0 and @@ -579,6 +682,25 @@ pub const Replica = struct { } return; } + const ends = lastDigest(u.tape); + defer if (ends) |d| { + r.last_digest.clearRetainingCapacity(); + r.last_digest.appendSlice(r.gpa, d) catch {}; + }; + if (r.applied > 0 and u.origin != r.id and ends != null and std.mem.eql(u8, ends.?, r.last_digest.items)) { + // Nothing a digest covers moved: what was applied on top is + // what the session will apply on top of it. Replayed when the + // copy is next put back. + r.neutral += 1; + const kept = r.gpa.dupe(u8, u.tape) catch return r.fail("out of memory"); + r.since.append(r.gpa, kept) catch return r.fail("out of memory"); + if (r.since.items.len >= rebase_after) { + r.stats.rebases += 1; + r.toConfirmed(); + r.dirty = true; + } + return; + } if (r.applied > 0) { if (u.origin == r.id) r.stats.mispredicted += 1 else r.stats.foreign += 1; if (u.origin == r.id and std.c.getenv("PARDES_REPLICA_TRACE") != null) { @@ -593,6 +715,7 @@ pub const Replica = struct { r.recovering = nowNs(); r.toConfirmed(); } + r.moved = true; if (tape.replayUnit(&r.t, u.tape)) |why| { if (std.c.getenv("PARDES_REPLICA_TRACE") != null) { var w: std.Io.Writer.Allocating = .init(r.gpa); @@ -657,17 +780,10 @@ pub const Replica = struct { /// pending[applied] applied on top: whether it was. fn speculate(r: *Replica) bool { const p = &r.pending.items[r.applied]; + // Every kind, a view's own (a resize, the pointer) too: each client + // has its own view, and what reaches the authority's (a pty's size, + // the pointer over a sliding pane) stops it as any host call does. const in = parseInput(p.msg) orelse return false; - switch (in.event) { - .key, .paste, .command => {}, - else => { - // The view's (a resize, the pointer): the authority's for - // now, phase 4's views to come. - r.stats.unspeculated += 1; - r.blocked = p.seq; - return false; - }, - } const c = r.core(); if (!r.copy.held) { const t0 = nowNs(); @@ -683,6 +799,10 @@ pub const Replica = struct { if (checked) terminalDigests(c, &terminals); const t0 = nowNs(); tape.beginSpeculation(&r.t); + r.t.guess = if (r.guess) guess: { + const since: i64 = @intCast(nowNs() -| r.stamp.at_ns + r.rtt_ns); + break :guess .{ .real_ns = r.stamp.real_ns + since, .mono_ns = r.stamp.mono_ns + since, .gmtoff = r.gmtoff }; + } else null; c.toView(r.view); // Input into a terminal is the authority's: its step reads the // emulator (rows, cursor, modes), which here is the last delta's and @@ -690,7 +810,7 @@ pub const Replica = struct { // echo. if (r.t.screen_state) if (c.panes[c.active]) |pane| if (pane.isTerminal()) tape.poison(); c.update(in.event); - c.pump(c.host) catch {}; + c.step(); const clean = tape.endSpeculation(&r.t); const dt = nowNs() - t0; // Where a blob was written, should one be: the input that did. @@ -699,6 +819,7 @@ pub const Replica = struct { // Undone, and the earlier ones applied again on the confirmed // position: this one waits for the authority. r.stats.stopped += 1; + if (std.c.getenv("PARDES_REPLICA_TRACE") != null) std.debug.print("replica {d}: its input {d} ({t}) stopped at a host call: it waits\n", .{ r.id, p.seq, in.event }); r.blocked = p.seq; const upto = r.applied; r.toConfirmed(); @@ -732,9 +853,41 @@ pub const Replica = struct { } /// Nothing pending and nothing applied: the replica is the authority's - /// confirmed state, which `digest` says. + /// confirmed state, which `digest` says (a copy taken ahead, `ready`, + /// has nothing on top of it). pub fn quiet(r: *const Replica) bool { - return r.pending.items.len == 0 and r.applied == 0 and !r.copy.held; + return r.pending.items.len == 0 and r.applied == 0; + } + + /// The copy taken now, while nothing is pending (the host is idle, its + /// frame drawn), so that the next input of its own is applied at once + /// without taking it: a copy is a megabyte saved. Let go when a unit + /// comes first. + pub fn ready(r: *Replica) void { + if (r.copy.held or r.pending.items.len > 0 or r.diverged != null) return; + const c = r.core(); + const t0 = nowNs(); + rollback.Epoch.select(&r.epoch); + defer rollback.Epoch.select(null); + r.copy.take(&r.epoch, c) catch return; + r.stats.copy_ns.append(r.gpa, nowNs() - t0) catch {}; + r.stats.copy_bytes += r.copy.savedBytes(); + r.stats.copies += 1; + } + + /// Whether its core loaded from the dump is the session as it was + /// dumped, by the welcome's `digest`. + pub fn loaded(r: *Replica, want: u64) bool { + return tape.digestOf(r.core()) == want; + } + + /// Its own view drawn and presented through `present`, between steps + /// (views.drawOne): nothing the digest covers moves. The epoch's trees + /// held while a copy is (a render can make or drop a syntax tree). + pub fn draw(r: *Replica, present: pardes.views.Present) void { + rollback.Epoch.select(&r.epoch); + defer rollback.Epoch.select(null); + pardes.views.drawOne(r.core(), r.view, present) catch {}; } /// One of the authority's terminal messages (`Screens.flush`, or an @@ -827,6 +980,20 @@ pub const Replica = struct { } }; +/// The bytes of a unit's last digest record: the state it ended on; null +/// for one that says what the host wrote or which host it is (a replica's +/// core takes those only by replaying the unit). +fn lastDigest(bytes: []const u8) ?[]const u8 { + var r: walk.Reader = .{ .bytes = bytes }; + var last: ?[]const u8 = null; + while (true) { + const head = r.take(5) catch return last; + const payload = r.take(std.mem.bytesToValue(u32, head[1..5])) catch return last; + if (head[0] == tape.digest_kind) last = payload; + if (head[0] == tape.set_kind) return null; + } +} + fn terminalDigests(p: *const Pardes, out: *[pardes.MAX_PANES]u64) void { for (p.panes, out) |slot, *d| { d.* = 0; @@ -838,6 +1005,647 @@ fn terminalDigests(p: *const Pardes, out: *[pardes.MAX_PANES]u64) void { } } +// ---- over a socket (docs/next-daemon.md, "Phase 5 part B notes") ---- +// +// The session's half: each client handed over by the daemon (daemon.zig) is +// a Peer, with a view of its own, a thread reading its inputs and one +// writing what it is sent. The recorder is on while any is. At each step +// boundary of the session's loop (`Peers.boundary`, after a pump): +// - the pump's unit goes to every client; +// - a client that came is joined: its view opened, the session dumped +// (src/dump.zig) and sent ahead of every unit after; +// - each input that came is a unit of its own: its view, the input, a +// `step` (what the client's own core applied at once); +// - a client gone takes its view with it. +// A client's link only ever holds bytes queued for it: the session's step +// never waits on one. One that lets `lag_max` pile up is cut off. + +const libc = std.c; +const log = std.log.scoped(.replica); +const nosignal: u32 = if (builtin.os.tag == .linux) libc.MSG.NOSIGNAL else 0; +/// What a client may leave unread, past its join's dump, before it is cut +/// off: minutes of typing, a screenful of a flood's deltas many times over. +/// PARDES_REPLICA_LAG_KB says another (a test's). +var lag_max: usize = 64 << 20; +const dump_piece = 1 << 20; + +const Tm = extern struct { sec: c_int, min: c_int, hour: c_int, mday: c_int, mon: c_int, year: c_int, wday: c_int, yday: c_int, isdst: c_int, gmtoff: c_long, zone: ?[*:0]const u8 }; +extern "c" fn localtime_r(t: *const libc.time_t, tm: *Tm) ?*Tm; + +/// This zone's offset from UTC now, seconds. +fn gmtoff() i32 { + var ts: libc.timespec = undefined; + _ = libc.clock_gettime(.REALTIME, &ts); + const secs: libc.time_t = ts.sec; + var tm: Tm = undefined; + return if (localtime_r(&secs, &tm) != null) @intCast(tm.gmtoff) else 0; +} + +fn writeAll(fd: c_int, bytes: []const u8) bool { + var at: usize = 0; + while (at < bytes.len) { + const n = libc.send(fd, bytes[at..].ptr, bytes.len - at, nosignal); + if (n < 0 and libc.errno(n) == .INTR) continue; + if (n <= 0) return false; + at += @intCast(n); + } + return true; +} + +fn readSome(fd: c_int, buf: []u8) usize { + while (true) { + const n = libc.read(fd, buf.ptr, buf.len); + if (n < 0 and libc.errno(n) == .INTR) continue; + return if (n <= 0) 0 else @intCast(n); + } +} + +/// What a stream's reader holds of it, read and not yet taken: a dump's +/// reader leaves the units read with its last piece to the stream's. +pub const Framer = struct { + buf: std.ArrayListUnmanaged(u8) = .empty, + len: usize = 0, + + pub fn deinit(f: *Framer, gpa: Allocator) void { + f.buf.deinit(gpa); + } + + /// Framed messages off `fd` (detached/wire.zig) handed to `take` until + /// the stream ends, `take` says stop, or a message is malformed. + fn run(fr: *Framer, gpa: Allocator, fd: c_int, ctx: anytype, comptime take: fn (@TypeOf(ctx), u8, []const u8) bool) void { + while (true) { + var at: usize = 0; + var stop = false; + while (!stop) { + const f = (wire.framed(fr.buf.items[at..fr.len]) catch break) orelse break; + at += f.total; + stop = !take(ctx, f.tag, f.payload); + } + if (at > 0) { + std.mem.copyForwards(u8, fr.buf.items[0 .. fr.len - at], fr.buf.items[at..fr.len]); + fr.len -= at; + } + if (stop) return; + if (wire.framed(fr.buf.items[0..fr.len])) |_| {} else |_| return; + // Room for what the head says is coming, or 64 KiB more. + const want = if (fr.len >= wire.header_len) wire.header_len + @as(usize, std.mem.readInt(u32, fr.buf.items[1..5], .little)) else 0; + const room = @max(want, fr.len + (64 << 10)); + if (fr.buf.items.len < room) fr.buf.resize(gpa, room) catch return; + const n = readSome(fd, fr.buf.items[fr.len..]); + if (n == 0) return; + fr.len += n; + } + } +}; + +fn readFramed(gpa: Allocator, fd: c_int, ctx: anytype, comptime take: fn (@TypeOf(ctx), u8, []const u8) bool) void { + var fr: Framer = .{}; + defer fr.deinit(gpa); + fr.run(gpa, fd, ctx, take); +} + +pub const Peer = struct { + peers: *Peers, + fd: c_int, + id: u16, + view: u8, + /// Inputs read (each one's seq, as the client numbered them). + seq: u32 = 0, + reader: ?std.Thread = null, + writer: ?std.Thread = null, + lock: std.Io.Mutex = .init, + cond: std.Io.Condition = .init, + /// What it is sent, not written yet, and what the writer writes now. + out: std.ArrayListUnmanaged(u8) = .empty, + sending: std.ArrayListUnmanaged(u8) = .empty, + limit: usize = 64 << 20, + closing: bool = false, + cut: bool = false, + + /// `parts` queued for it, in order; never waits. A client whose queue + /// passes its limit is cut off: its link is shut down, and its reader + /// says it is gone. + fn send(peer: *Peer, parts: []const []const u8) void { + const io = peer.peers.io; + peer.lock.lockUncancelable(io); + defer peer.lock.unlock(io); + if (peer.cut or peer.closing) return; + var n: usize = peer.out.items.len; + for (parts) |p| n += p.len; + if (n > peer.limit) { + log.info("client {d} is {d} MB behind: cut off", .{ peer.id, n >> 20 }); + peer.peers.cut += 1; + peer.cut = true; + _ = libc.shutdown(peer.fd, 2); + peer.cond.signal(io); + return; + } + for (parts) |p| peer.out.appendSlice(peer.peers.gpa, p) catch { + peer.cut = true; + _ = libc.shutdown(peer.fd, 2); + break; + }; + peer.cond.signal(io); + } + + fn writeLoop(peer: *Peer) void { + const io = peer.peers.io; + while (true) { + peer.lock.lockUncancelable(io); + while (peer.out.items.len == 0 and !peer.closing and !peer.cut) peer.cond.waitUncancelable(io, &peer.lock); + if (peer.out.items.len == 0 or peer.cut) { + peer.lock.unlock(io); + return; + } + std.mem.swap(std.ArrayListUnmanaged(u8), &peer.out, &peer.sending); + // The limit was for the dump and what is queued behind it. + peer.limit = lag_max; + peer.lock.unlock(io); + const ok = writeAll(peer.fd, peer.sending.items); + peer.sending.clearRetainingCapacity(); + if (!ok) { + _ = libc.shutdown(peer.fd, 2); + return; + } + } + } + + fn take(peer: *Peer, tag: u8, payload: []const u8) bool { + const t = std.enums.fromInt(wire.ClientTag, tag) orelse return false; + switch (t) { + .hello, .bye => return false, + .ping => { + var out: [16]u8 = undefined; + const v = (wire.decodeClient(tag, payload) catch return false).ping; + const pong = wire.encodeServer(&out, .{ .pong = v }) catch return false; + peer.send(&.{pong}); + return true; + }, + // ponytail: rterm's acks (src/screen_state.zig) go to its + // session state once terminals cross as screens. + .screen_ack => return true, + .check => { + peer.peers.post(.{ .check = peer }); + return true; + }, + else => {}, + } + peer.seq += 1; + const msg = peer.peers.gpa.alloc(u8, head_len + wire.header_len + payload.len) catch return false; + header(msg, peer.id, peer.seq); + msg[head_len] = tag; + std.mem.writeInt(u32, msg[head_len + 1 ..][0..4], @intCast(payload.len), .little); + @memcpy(msg[head_len + wire.header_len ..], payload); + if (parseInput(msg) == null) { + peer.peers.gpa.free(msg); + return false; + } + peer.peers.post(.{ .input = .{ .peer = peer, .msg = msg } }); + return true; + } + + fn readLoop(peer: *Peer) void { + readFramed(peer.peers.gpa, peer.fd, peer, take); + peer.peers.post(.{ .gone = peer }); + } +}; + +pub const Peers = struct { + gpa: Allocator = std.heap.c_allocator, + io: std.Io = undefined, + /// This session's name and pid, for a welcome. + name: []const u8 = "", + t: tape.Tape = undefined, + recording: bool = false, + list: [pardes.views.max]?*Peer = @splat(null), + next_id: u16 = 1, + seq: u32 = 0, + lock: std.Io.Mutex = .init, + inbox: std.ArrayListUnmanaged(Note) = .empty, + taken: std.ArrayListUnmanaged(Note) = .empty, + /// A reader posted: the loop's one test a step while nothing is attached. + mail: std.atomic.Value(bool) = .init(false), + /// Wakes the session's loop (its wait for input) from a reader. + wake_ctx: ?*anyopaque = null, + wake: ?*const fn (?*anyopaque) void = null, + /// What the session would be with no client: unattended without a + /// terminal of its own (the host says). + alone_unattended: bool = false, + /// The worst a join cost the session's step, and the last one's dump. + join_ns: u64 = 0, + dump_bytes: usize = 0, + /// What each client input cost the session (its step, its unit sent), + /// and each boundary (its pump's unit sent, what came taken in), ns: + /// PARDES_REPLICA_STATS writes them at the session's end. + input_ns: std.ArrayListUnmanaged(u64) = .empty, + boundary_ns: std.ArrayListUnmanaged(u64) = .empty, + cut: u32 = 0, + head: [unit_head]u8 = undefined, + + pub const Note = union(enum) { join: c_int, input: struct { peer: *Peer, msg: []u8 }, gone: *Peer, check: *Peer }; + + pub fn count(ps: *const Peers) usize { + var n: usize = 0; + for (ps.list) |slot| n += @intFromBool(slot != null); + return n; + } + + fn post(ps: *Peers, note: Note) void { + ps.lock.lockUncancelable(ps.io); + ps.inbox.append(ps.gpa, note) catch { + ps.lock.unlock(ps.io); + switch (note) { + .join => |fd| _ = libc.close(fd), + .input => |in| ps.gpa.free(in.msg), + .gone, .check => {}, + } + return; + }; + ps.mail.store(true, .release); + ps.lock.unlock(ps.io); + if (ps.wake) |w| w(ps.wake_ctx); + } + + /// A replica's connection the daemon handed over (its hello taken + /// there): joined at the next step boundary. Any thread. + pub fn handed(ps: *Peers, fd: c_int) void { + ps.post(.{ .join = fd }); + } + + /// What was recorded since the last unit, to every client as one, its + /// digest at its end. + fn unit(ps: *Peers, origin: u16, seq: u32) void { + if (tape.taken(&ps.t).len == 0) return; + tape.recordDigest(&ps.t); + const bytes = tape.taken(&ps.t); + unitHeader(&ps.head, origin, seq); + if (std.c.getenv("PARDES_REPLICA_TRACE") != null) { + var w: std.Io.Writer.Allocating = .init(ps.gpa); + defer w.deinit(); + w.writer.print("session: unit {d}/{d}, {d} bytes, to {d} clients:\n", .{ origin, seq, bytes.len, ps.count() }) catch {}; + tape.describe(bytes, &w.writer); + std.debug.print("{s}", .{w.written()}); + } + const head = wire.header(@backingInt(wire.ServerTag.unit), unit_head + bytes.len); + for (ps.list) |slot| if (slot) |peer| peer.send(&.{ &head, &ps.head, bytes }); + tape.clearTaken(&ps.t); + } + + /// After each of the session's pumps: its unit sent, the clients that + /// came joined, their inputs taken in, the ones gone let go. + pub fn boundary(ps: *Peers, core: *Pardes) void { + if (!ps.recording and !ps.mail.load(.acquire)) return; + const t0 = nowNs(); + defer if (ps.recording) ps.boundary_ns.append(ps.gpa, nowNs() - t0) catch {}; + if (ps.recording) { + ps.seq += 1; + ps.unit(0, ps.seq); + } + if (!ps.mail.load(.acquire)) return; + ps.lock.lockUncancelable(ps.io); + std.mem.swap(std.ArrayListUnmanaged(Note), &ps.inbox, &ps.taken); + ps.mail.store(false, .release); + ps.lock.unlock(ps.io); + for (ps.taken.items) |note| switch (note) { + .join => |fd| ps.join(core, fd), + .input => |in| { + ps.input(core, in.peer, in.msg); + ps.gpa.free(in.msg); + }, + .gone => |peer| ps.leave(core, peer), + .check => |peer| { + var out: [16]u8 = undefined; + peer.send(&.{wire.encodeServer(&out, .{ .check = tape.digestOf(core) }) catch unreachable}); + }, + }; + ps.taken.clearRetainingCapacity(); + } + + fn input(ps: *Peers, core: *Pardes, peer: *Peer, msg: []const u8) void { + const in = parseInput(msg) orelse return; + const t0 = nowNs(); + defer ps.input_ns.append(ps.gpa, nowNs() - t0) catch {}; + core.toView(peer.view); + core.update(in.event); + core.step(); + ps.unit(peer.id, in.seq); + } + + fn refuse(fd: c_int, why: []const u8) void { + var out: [600]u8 = undefined; + if (wire.encodeServer(&out, .{ .refused = why })) |bytes| _ = writeAll(fd, bytes) else |_| {} + _ = libc.close(fd); + } + + fn join(ps: *Peers, core: *Pardes, fd: c_int) void { + const t0 = nowNs(); + defer ps.join_ns = @max(ps.join_ns, nowNs() - t0); + const slot = for (&ps.list) |*s| { + if (s.* == null) break s; + } else return refuse(fd, "that session has as many clients as it takes"); + if (!ps.recording) { + // A session recording to a file (PARDES_RECORD) has its tape. + if (tape.on != null) return refuse(fd, "that session records a tape of its own (PARDES_RECORD)"); + ps.t = tape.memory(ps.gpa); + ps.t.unit_digests = true; + tape.on = &ps.t; + ps.recording = true; + } + // Someone looks at it now: its messages go by keys, not the clock. + core.unattended = false; + const k = core.openView() orelse { + core.unattended = ps.count() == 0 and ps.alone_unattended; + if (ps.count() == 0) ps.stop(core); + return refuse(fd, "that session has as many views as it takes"); + }; + // Its own process draws it. + core.undrawn |= @as(u8, 1) << @intCast(k); + ps.seq += 1; + ps.unit(0, ps.seq); + const bytes = pardes.dump.dumpBytes(core) catch |err| { + core.undrawn &= ~(@as(u8, 1) << @intCast(k)); + core.closeView(k); + ps.unit(0, ps.seq); + if (ps.count() == 0) ps.stop(core); + var why: [96]u8 = undefined; + return refuse(fd, std.fmt.bufPrint(&why, "that session could not be dumped: {t}", .{err}) catch "that session could not be dumped"); + }; + // The core's, as a Dump's is. + defer core.gpa.free(bytes); + ps.dump_bytes = bytes.len; + // The joiner's core is the dump's: the host this tape wraps, and + // what the host wrote in the core, are said again at the next input. + tape.resync(&ps.t); + const peer = ps.gpa.create(Peer) catch return refuse(fd, "out of memory"); + if (std.c.getenv("PARDES_REPLICA_LAG_KB")) |kb| lag_max = (std.fmt.parseInt(usize, std.mem.span(kb), 10) catch 64 << 10) << 10; + peer.* = .{ .peers = ps, .fd = fd, .id = ps.next_id, .view = k, .limit = bytes.len + lag_max }; + ps.next_id +%= 1; + if (ps.next_id == 0) ps.next_id = 1; + var out: [wire.hello_max + 64]u8 = undefined; + // The session's digest as dumped, for the joiner to check its load by. + const digest = tape.digestOf(core); + const welcome = wire.encodeServer(&out, .{ .welcome = .{ .pid = @intCast(libc.getpid()), .session = ps.name[0..@min(ps.name.len, wire.max_name)], .client = peer.id, .view = k, .dump_len = bytes.len, .gmtoff = gmtoff(), .digest = digest } }) catch unreachable; + peer.send(&.{welcome}); + var at: usize = 0; + while (at < bytes.len) : (at += dump_piece) { + const piece = bytes[at..@min(bytes.len, at + dump_piece)]; + const head = wire.header(@backingInt(wire.ServerTag.dump), piece.len); + peer.send(&.{ &head, piece }); + } + peer.writer = std.Thread.spawn(.{}, Peer.writeLoop, .{peer}) catch null; + peer.reader = std.Thread.spawn(.{}, Peer.readLoop, .{peer}) catch null; + if (peer.writer == null or peer.reader == null) { + ps.letGo(peer); + core.undrawn &= ~(@as(u8, 1) << @intCast(k)); + core.closeView(k); + ps.unit(0, ps.seq); + return; + } + slot.* = peer; + log.info("client {d} joined '{s}' as view {d}: a {d} KB dump in {d} us", .{ peer.id, ps.name, k, bytes.len >> 10, (nowNs() - t0) / 1000 }); + if (std.c.getenv("PARDES_REPLICA_TRACE") != null) std.debug.print("session: client {d} joined after unit {d}\n", .{ peer.id, ps.seq }); + } + + fn leave(ps: *Peers, core: *Pardes, peer: *Peer) void { + for (&ps.list) |*slot| if (slot.* == peer) { + slot.* = null; + }; + log.info("client {d} left '{s}'", .{ peer.id, ps.name }); + core.undrawn &= ~(@as(u8, 1) << @intCast(peer.view)); + core.closeView(peer.view); + if (ps.count() == 0) core.unattended = ps.alone_unattended; + ps.seq += 1; + ps.unit(0, ps.seq); + ps.letGo(peer); + if (ps.count() == 0) ps.stop(core); + } + + /// Its threads ended and joined, its link closed. + fn letGo(ps: *Peers, peer: *Peer) void { + const io = ps.io; + peer.lock.lockUncancelable(io); + peer.closing = true; + peer.cond.signal(io); + peer.lock.unlock(io); + _ = libc.shutdown(peer.fd, 2); + if (peer.writer) |w| w.join(); + if (peer.reader) |r| r.join(); + _ = libc.close(peer.fd); + peer.out.deinit(ps.gpa); + peer.sending.deinit(ps.gpa); + // What its reader posted after it was let go (its end) is nobody's. + ps.lock.lockUncancelable(io); + var i: usize = 0; + while (i < ps.inbox.items.len) { + const mine = switch (ps.inbox.items[i]) { + .input => |in| in.peer == peer, + .gone, .check => |p| p == peer, + .join => false, + }; + if (!mine) { + i += 1; + continue; + } + if (ps.inbox.items[i] == .input) ps.gpa.free(ps.inbox.items[i].input.msg); + _ = ps.inbox.orderedRemove(i); + } + ps.lock.unlock(io); + ps.gpa.destroy(peer); + } + + /// Detach typed in client view `k`'s: its link ends (its reader says + /// it is gone, and the next boundary lets it go). + pub fn detach(ps: *Peers, k: u8) bool { + for (ps.list) |slot| if (slot) |peer| if (peer.view == k) { + _ = libc.shutdown(peer.fd, 2); + return true; + }; + return false; + } + + fn stop(ps: *Peers, core: *Pardes) void { + if (!ps.recording) return; + if (core.host.vtable == &ps.t.vtable) core.host = ps.t.real; + tape.on = null; + tape.release(&ps.t); + ps.recording = false; + } + + /// Every client let go (the session's end, a Restore): each sees its + /// link end. + pub fn dropAll(ps: *Peers, core: *Pardes) void { + for (&ps.list) |*slot| if (slot.*) |peer| { + slot.* = null; + core.undrawn &= ~(@as(u8, 1) << @intCast(peer.view)); + core.closeView(peer.view); + ps.letGo(peer); + }; + ps.stop(core); + ps.lock.lockUncancelable(ps.io); + for (ps.inbox.items) |note| switch (note) { + .join => |fd| _ = libc.close(fd), + .input => |in| ps.gpa.free(in.msg), + .gone, .check => {}, + }; + ps.inbox.clearRetainingCapacity(); + ps.lock.unlock(ps.io); + core.unattended = ps.alone_unattended; + } + + pub fn deinit(ps: *Peers) void { + if (std.c.getenv("PARDES_REPLICA_STATS")) |path| if (ps.input_ns.items.len > 0) ps.writeStats(std.mem.span(path)); + ps.inbox.deinit(ps.gpa); + ps.taken.deinit(ps.gpa); + ps.input_ns.deinit(ps.gpa); + ps.boundary_ns.deinit(ps.gpa); + } + + fn writeStats(ps: *Peers, path: []const u8) void { + var buf: [400]u8 = undefined; + const line = std.fmt.bufPrint(&buf, "session inputs {d} input_us {d}/{d}/{d} boundaries {d} boundary_us {d}/{d}/{d} join_us {d} dump_kb {d} cut {d}\n", .{ + ps.input_ns.items.len, pctOf(ps.input_ns.items, 50) / 1000, pctOf(ps.input_ns.items, 99) / 1000, pctOf(ps.input_ns.items, 100) / 1000, + ps.boundary_ns.items.len, pctOf(ps.boundary_ns.items, 50) / 1000, pctOf(ps.boundary_ns.items, 99) / 1000, pctOf(ps.boundary_ns.items, 100) / 1000, + ps.join_ns / 1000, ps.dump_bytes >> 10, ps.cut, + }) catch return; + var z: [4096]u8 = undefined; + const pz = std.fmt.bufPrintSentinel(&z, "{s}", .{path}, 0) catch return; + const fd = libc.open(pz, .{ .ACCMODE = .WRONLY, .CREAT = true, .APPEND = true, .CLOEXEC = true }, @as(libc.mode_t, 0o644)); + if (fd < 0) return; + _ = libc.write(fd, line.ptr, line.len); + _ = libc.close(fd); + } +}; + +// ---- the client's half ---- + +pub const Joined = struct { + fd: c_int, + pid: u32, + client: u16, + view: u8, + gmtoff: i32, + dump: []u8, + /// The session's digest as dumped (`Replica.loaded`). + digest: u64 = 0, + /// What was read past the dump: the stream's (`readStream`). + rest: Framer = .{}, +}; + +/// Connected to the replica socket at `path`, its hello said for +/// `session`, welcomed, and the dump read whole: what a client joins with. +/// Errors are said in `why`. +pub fn dial(gpa: Allocator, path: []const u8, session: []const u8, why: []u8) ?Joined { + const fd = libc.socket(libc.AF.UNIX, libc.SOCK.STREAM | libc.SOCK.CLOEXEC, 0); + if (fd < 0) return said(why, "no socket", .{}); + var addr: libc.sockaddr.un = .{ .path = @splat(0) }; + if (path.len >= addr.path.len) { + _ = libc.close(fd); + return said(why, "the socket's path is too long: {s}", .{path}); + } + @memcpy(addr.path[0..path.len], path); + if (libc.connect(fd, @ptrCast(&addr), @sizeOf(libc.sockaddr.un)) != 0) { + _ = libc.close(fd); + return said(why, "nothing answers at {s}", .{path}); + } + var out: [wire.hello_max]u8 = undefined; + const hello = wire.encodeClient(&out, .{ .hello = .{ .session = session } }) catch { + _ = libc.close(fd); + return said(why, "that session's name is too long", .{}); + }; + var j: Joined = .{ .fd = fd, .pid = 0, .client = 0, .view = 0, .gmtoff = 0, .dump = &.{} }; + const Read = struct { + j: *Joined, + gpa: Allocator, + why: []u8, + got: usize = 0, + failed: ?[]const u8 = null, + welcomed: bool = false, + fn take(r: *@This(), tag: u8, payload: []const u8) bool { + const msg = wire.decodeServer(tag, payload) catch { + r.failed = "the session's answer was not understood"; + return false; + }; + switch (msg) { + .welcome => |w| { + r.j.pid = w.pid; + r.j.client = w.client; + r.j.view = w.view; + r.j.gmtoff = w.gmtoff; + r.j.digest = w.digest; + r.j.dump = r.gpa.alloc(u8, @intCast(w.dump_len)) catch { + r.failed = "out of memory for the session's dump"; + return false; + }; + r.welcomed = true; + return w.dump_len > 0; + }, + .refused => |words| { + r.failed = std.fmt.bufPrint(r.why, "{s}\x00", .{words}) catch "refused"; + return false; + }, + .dump => |piece| { + if (!r.welcomed or r.got + piece.len > r.j.dump.len) { + r.failed = "the session sent more dump than it said"; + return false; + } + @memcpy(r.j.dump[r.got..][0..piece.len], piece); + r.got += piece.len; + return r.got < r.j.dump.len; + }, + else => { + r.failed = "the session sent a unit before its dump"; + return false; + }, + } + } + }; + if (!writeAll(fd, hello)) { + _ = libc.close(fd); + return said(why, "the session hung up", .{}); + } + var r: Read = .{ .j = &j, .gpa = gpa, .why = why }; + j.rest.run(gpa, fd, &r, Read.take); + if (r.failed != null or !r.welcomed or r.got < j.dump.len) { + j.rest.deinit(gpa); + if (j.dump.len > 0) gpa.free(j.dump); + _ = libc.close(fd); + const text = r.failed orelse "the session hung up before the dump was whole"; + if (text.ptr == why.ptr) return null; + return said(why, "{s}", .{text}); + } + return j; +} + +fn said(why: []u8, comptime fmt: []const u8, args: anytype) ?Joined { + _ = std.fmt.bufPrint(why, fmt ++ "\x00", args) catch {}; + return null; +} + +/// A message `Replica.input` made, to the session: the event's frame (its +/// client and seq are the connection's). +pub fn sendInput(fd: c_int, msg: []const u8) bool { + return writeAll(fd, msg[head_len..]); +} + +/// Asks the session's digest, answered in stream order. +pub fn askCheck(fd: c_int) bool { + var out: [16]u8 = undefined; + return writeAll(fd, wire.encodeClient(&out, .check) catch unreachable); +} + +/// The client's last word. +pub fn bye(fd: c_int) void { + var out: [16]u8 = undefined; + if (wire.encodeClient(&out, .bye)) |b| _ = writeAll(fd, b) else |_| {} +} + +/// What follows the dump on the client's connection, each message handed +/// to `take` (a unit, a terminal's delta) until the link ends; `rest` is +/// what the dump's reader read past it. +pub fn readStream(gpa: Allocator, fd: c_int, rest: *Framer, ctx: anytype, comptime take: fn (@TypeOf(ctx), u8, []const u8) bool) void { + rest.run(gpa, fd, ctx, take); +} + pub fn digestOf(a: *Authority) u64 { const was = Stepping.begin(&a.t, null); defer was.end(); @@ -1078,3 +1886,36 @@ test "the authority's units with a big file open: its blobs hashed once (PARDES_ std.debug.print("{d} MB file, blobs {s}: a unit p50 {d} us p99 {d} us; a digest p50 {d} us p99 {d} us\n", .{ mb, if (on) "hashed once" else "hashed whole", us[0] / 1000, us[1] / 1000, ds[0] / 1000, ds[1] / 1000 }); } } + +test "a replica joined from the session's dump replays the units after it" { + const gpa = std.testing.allocator; + const a = try Authority.init(gpa, .{ .cols = 80, .rows = 24, .file = "mise.toml" }, TestHost.host(), false); + defer a.deinit(); + _ = try a.local(.{ .key = .{ .cp = 'j', .text = "j" } }); + const was = Stepping.begin(&a.t, null); + const bytes = try pardes.dump.dumpBytes(a.core); + tape.resync(&a.t); + was.end(); + defer a.core.gpa.free(bytes); + const r = try Replica.initDump(gpa, 1, 0, bytes, .{ .cols = 80, .rows = 24 }); + defer r.deinit(); + try std.testing.expectEqual(digestOf(a), r.digest()); + r.confirm(try a.local(.{ .key = .{ .cp = 'j', .text = "j" } })); + if (r.diverged) |why| std.debug.print("diverged: {s}\n", .{why}); + try std.testing.expect(r.diverged == null); + var sent: std.ArrayListUnmanaged([]u8) = .empty; + defer { + for (sent.items) |m| gpa.free(m); + sent.deinit(gpa); + } + for ("oab") |k| try key(r, k, &sent); + try key(r, pardes.Key.escape, &sent); + for (sent.items) |m| { + r.confirm(try a.take(m)); + r.settle(); + } + if (r.diverged) |why| std.debug.print("diverged: {s}\n", .{why}); + try std.testing.expect(r.diverged == null); + try std.testing.expect(r.quiet()); + try std.testing.expectEqual(digestOf(a), r.digest()); +} diff --git a/src/rollback.zig b/src/rollback.zig index 7216664c..b0394d22 100644 --- a/src/rollback.zig +++ b/src/rollback.zig @@ -40,11 +40,22 @@ const Pardes = pardes.Pardes; const blob_min = 4096; const checked = builtin.mode == .debug or builtin.mode == .safe; +/// An address as a hash map's key: a multiply, not a Wyhash of its bytes +/// (every allocation of a replica's core is a lookup). +const Addr = struct { + pub fn hash(_: Addr, a: usize) u64 { + return @as(u64, a) *% 0x9e3779b97f4a7c15; + } + pub fn eql(_: Addr, a: usize, b: usize) bool { + return a == b; + } +}; + pub const Epoch = struct { child: Allocator = std.heap.c_allocator, /// Its own bookkeeping, never the epoch's. meta: Allocator = std.heap.c_allocator, - live: std.AutoHashMapUnmanaged(usize, Entry) = .empty, + live: std.HashMapUnmanaged(usize, Entry, Addr, std.hash_map.default_max_load_percentage) = .empty, on: bool = false, id: u32 = 0, born: std.ArrayListUnmanaged(usize) = .empty, @@ -190,7 +201,7 @@ pub const Copy = struct { meta: Allocator = std.heap.c_allocator, regions: std.ArrayListUnmanaged(Region) = .empty, bytes: std.ArrayListUnmanaged(u8) = .empty, - seen: std.AutoHashMapUnmanaged(Seen, void) = .empty, + seen: std.HashMapUnmanaged(Seen, void, SeenContext, std.hash_map.default_max_load_percentage) = .empty, pins: std.ArrayListUnmanaged(Pin) = .empty, held: bool = false, epoch: ?*Epoch = null, @@ -202,6 +213,14 @@ pub const Copy = struct { const Region = struct { addr: usize, off: usize, len: usize }; const Pin = struct { addr: usize, len: usize, hash: u64, where: []const u8 }; const Seen = struct { addr: usize, ty: usize }; + const SeenContext = struct { + pub fn hash(_: SeenContext, k: Seen) u64 { + return (@as(u64, k.addr) *% 0x9e3779b97f4a7c15) ^ (@as(u64, k.ty) *% 0xc2b2ae3d27d4eb4f); + } + pub fn eql(_: SeenContext, a: Seen, b: Seen) bool { + return a.addr == b.addr and a.ty == b.ty; + } + }; pub const Error = Allocator.Error; pub fn deinit(c: *Copy) void { diff --git a/src/tape.zig b/src/tape.zig index c2ddbb9d..b30096f5 100644 --- a/src/tape.zig +++ b/src/tape.zig @@ -167,6 +167,9 @@ pub const Tape = struct { /// output carries one, at the pty's read rate), not one a field; what /// differs is then not named. whole_digest: bool = false, + /// A digest only when asked (`recordDigest`): the replica stream's, + /// one a unit, not one after each of its inputs. + unit_digests: bool = false, /// A replica client's guess of the session's clocks for a speculation /// (src/replica.zig): its clock reads (a message's stamp, a file's /// mtime) answer this instead of the clocks here. @@ -275,7 +278,7 @@ pub const Tape = struct { t.wrap(core); }; t.record(.done, void, &{}); - if ((t.depth == 0 or t.every) and !t.no_digest) if (t.core) |core| { + if ((t.depth == 0 or t.every) and !t.no_digest and !t.unit_digests) if (t.core) |core| { var parts: Parts = undefined; const whole = [1]u64{digestOf(core)}; const d: []const u64 = if (t.whole_digest) &whole else digestParts(core, &parts); @@ -802,6 +805,16 @@ fn guessed(comptime R: type, g: Guess) R { @compileError("tape: no guess of a clock read as " ++ @typeName(R)); } +/// The session's digest recorded now (a unit's end: the replica stream's +/// one, `unit_digests`). +pub fn recordDigest(t: *Tape) void { + const core = t.core orelse return; + var parts: Parts = undefined; + const whole = [1]u64{digestOf(core)}; + const d: []const u64 = if (t.whole_digest) &whole else digestParts(core, &parts); + t.record(.digest, []const u64, &d); +} + /// A core made outside the tape (a replica client's, from a dump) is the /// one it replays into. pub fn adoptCore(t: *Tape, p: *Pardes) void { diff --git a/src/tty/tty.zig b/src/tty/tty.zig index 14dc8042..43994645 100644 --- a/src/tty/tty.zig +++ b/src/tty/tty.zig @@ -17,6 +17,8 @@ const panel_compositor = @import("panel_compositor.zig"); const daemon = @import("../daemon.zig"); const host_io = @import("../host_io.zig"); const upgrade = @import("../upgrade.zig"); +const replica = @import("../replica.zig"); +const wire = @import("../detached/wire.zig"); extern "c" fn setenv(name: [*:0]const u8, value: [*:0]const u8, overwrite: c_int) c_int; @@ -150,6 +152,15 @@ pub const Command = struct { search_rows: pardes.Stream.Batch, files_changed, fs_ready, + /// A replica's connection, handed over by the daemon (Peers). + replica: posix.fd_t, + /// A replica's reader posted: the step boundary takes it in. + replica_mail, + /// A replica client's: what the session streamed (replica.Client). + unit: []u8, + /// The session's digest, asked by PARDES_REPLICA_CHECK. + checked: u64, + link_gone, } = .nop; }; const Loop = vaxis.Loop(@TypeOf(Command.value)); @@ -838,14 +849,18 @@ fn terminalCellPixels(total_pixels: u16, cells: u16, fallback: u32) u32 { } fn updateCoreTerminalSize(core: *pardes.Pardes, cols: u16, rows: u16, pixel_w: u16, pixel_h: u16) void { - core.update(.{ .resize = .{ + core.update(.{ .resize = terminalSize(cols, rows, pixel_w, pixel_h) }); +} + +fn terminalSize(cols: u16, rows: u16, pixel_w: u16, pixel_h: u16) @FieldType(pardes.Event, "resize") { + return .{ .cols = cols, .rows = rows, .cell_pixels = .{ .w = @intCast(terminalCellPixels(pixel_w, cols, 8)), .h = @intCast(terminalCellPixels(pixel_h, rows, 16)), }, - } }); + }; } /// A terminal pane's image fragment (`native.exact`): its `src` into its @@ -1255,6 +1270,268 @@ pub fn run(init: std.process.Init, opts: pardes.Options, term: Term) !void { try localSession(init, opts, &screen, &loop, &paste_buf); } +/// A replica client's link to its session (`remote`). +const Remote = struct { + r: *replica.Replica, + fd: posix.fd_t, + quit: bool = false, + /// The session hung up. + gone: bool = false, + /// Something to draw: an input applied, a unit taken. + dirty: bool = true, + /// The last frame was held for the terminal's answer (FrameAck). + owed: bool = false, + /// PARDES_REPLICA_CHECK=: once quiet a while after a change, the + /// session's digest is asked, and whether this replica's agrees is + /// written there, a line a check. + check: ?[:0]const u8 = null, + changed: bool = false, + asked: bool = false, + /// How many of its own inputs the session recorded otherwise when it + /// last drew: a frame after one that writes to the terminal is a + /// rollback that may have been seen (counted so: an upper bound, the + /// frame shows the inputs after it too). + mispredicted: u32 = 0, + visible: u32 = 0, + + /// PARDES_REPLICA_STATS=: how its inputs and the session's units + /// went, a line at its end (test/daemon.py, the gates). + fn writeStats(c: *Remote, path: []const u8) void { + const st = &c.r.stats; + var buf: [600]u8 = undefined; + const line = std.fmt.bufPrint(&buf, "inputs {d} at_once {d} stopped {d} units {d} fast {d} neutral {d} mispredicted {d} foreign {d} visible {d} rtt_us {d} spec_us {d}/{d} copy_us {d}/{d} rollback_us {d}/{d}\n", .{ + st.inputs, st.speculated, st.stopped, st.units, st.fast, + c.r.neutral, st.mispredicted, st.foreign, c.visible, c.r.rtt_ns / 1000, pct(st.spec_ns.items, 50) / 1000, + pct(st.spec_ns.items, 99) / 1000, pct(st.copy_ns.items, 50) / 1000, pct(st.copy_ns.items, 99) / 1000, pct(st.rollback_ns.items, 50) / 1000, pct(st.rollback_ns.items, 99) / 1000, + }) catch return; + var z: [4096]u8 = undefined; + const pz = std.fmt.bufPrintSentinel(&z, "{s}", .{path}, 0) catch return; + const fd = libc.open(pz, .{ .ACCMODE = .WRONLY, .CREAT = true, .APPEND = true, .CLOEXEC = true }, @as(libc.mode_t, 0o644)); + if (fd < 0) return; + _ = libc.write(fd, line.ptr, line.len); + _ = libc.close(fd); + } + + fn pct(samples: []u64, p: usize) u64 { + if (samples.len == 0) return 0; + std.mem.sort(u64, samples, {}, std.sort.asc(u64)); + return samples[@min(samples.len - 1, samples.len * p / 100)]; + } + + /// The session's digest against this replica's, if it is quiet still. + fn checked(c: *Remote, theirs: u64) void { + c.asked = false; + const path = c.check orelse return; + if (!c.r.quiet()) return; + const mine = c.r.digest(); + var buf: [96]u8 = undefined; + const line = (if (mine == theirs) std.fmt.bufPrint(&buf, "agree {x:0>16}\n", .{mine}) else std.fmt.bufPrint(&buf, "differ {x:0>16} {x:0>16}\n", .{ mine, theirs })) catch return; + const fd = libc.open(path, .{ .ACCMODE = .WRONLY, .CREAT = true, .APPEND = true, .CLOEXEC = true }, @as(libc.mode_t, 0o644)); + if (fd < 0) return; + _ = libc.write(fd, line.ptr, line.len); + _ = libc.close(fd); + } + + /// The client's own view drawn on its terminal, outside any step: the + /// replica's digest stays the session's (views.drawOne). What the host + /// writes in the core as it presents (a frame skipped, one owed) is put + /// back, as the frame statistics are this process's. + fn draw(c: *Remote, s: *Shell) void { + const core = c.r.core(); + const skipped = core.present_skipped; + const needs = core.needs_frame; + core.frames.mark_start = pardes.FrameStats.nowNs(); + core.frames.mark_update = core.frames.mark_start; + c.r.draw(.{ .ctx = s, .f = Shell.present }); + core.frames.mark_render = core.frames.mark_update; + if (c.r.stats.mispredicted > c.mispredicted) { + c.mispredicted = c.r.stats.mispredicted; + if (core.frames.bytes > 0) c.visible += 1; + } + c.owed = core.present_skipped; + core.frames.record(pardes.FrameStats.nowNs(), false, c.owed); + core.present_skipped = skipped; + core.needs_frame = needs; + c.dirty = false; + } +}; + +/// The session's stream, onto the loop: its units, and its end. +fn linkReader(gpa: std.mem.Allocator, fd: posix.fd_t, rest: *replica.Framer, loop: *Loop, done: *std.atomic.Value(bool)) void { + defer done.store(true, .release); + const Take = struct { + fn take(l: *Loop, tag: u8, payload: []const u8) bool { + const t = std.enums.fromInt(wire.ServerTag, tag) orelse return false; + switch (t) { + .unit => { + const bytes = std.heap.c_allocator.dupe(u8, payload) catch return false; + l.postEvent(.{ .unit = bytes }) catch std.heap.c_allocator.free(bytes); + }, + .check => { + const d = (wire.decodeServer(tag, payload) catch return false).check; + l.postEvent(.{ .checked = d }) catch {}; + }, + .refused => return false, + // ponytail: a terminal's delta (rterm's, src/screen_state.zig) + // applies here once terminals cross as screens. + .screen, .pong, .welcome, .dump => {}, + } + return true; + } + }; + replica.readStream(gpa, fd, rest, loop, Take.take); + loop.postEvent(.link_gone) catch {}; +} + +/// `pardes --attach=:` (and `:`, through +/// ssh, main.zig): this terminal a replica client of a session elsewhere +/// (docs/next-daemon.md, "Phase 5 part B notes"). Its own core, loaded +/// from the session's dump, applies its keys at once (src/replica.zig); +/// the session's units are the truth it rolls back to. Its own view of the +/// session is drawn here, as the session draws its terminals'. +pub fn remote(init: std.process.Init, opts: pardes.Options, path: []const u8, session: []const u8) !void { + const io = init.io; + const gpa = init.gpa; + var why: [600]u8 = @splat(0); + var joined = replica.dial(gpa, path, session, &why) orelse { + var said: [700]u8 = undefined; + std.Io.File.stderr().writeStreamingAll(io, std.fmt.bufPrint(&said, "pardes: --attach: {s}\n", .{std.mem.sliceTo(&why, 0)}) catch "pardes: --attach failed\n") catch {}; + std.process.exit(1); + }; + defer _ = libc.close(joined.fd); + defer joined.rest.deinit(gpa); + + const allocs = pardes.memory.init(gpa); + defer pardes.memory.deinit(); + pardes.image.start(io, allocs.image); + if (comptime pardes.pdf_enabled) pardes.pdf.start(allocs.pdf); + pardes.syntax.start(allocs.tree_sitter); + defer { + pardes.image.stop(); + if (comptime pardes.pdf_enabled) pardes.pdf.stop(); + pardes.syntax.stop(); + } + var o = opts; + o.image_allocator = allocs.image; + o.pdf_allocator = allocs.pdf; + const r = replica.Replica.initDump(gpa, joined.client, joined.view, joined.dump, o) catch |err| { + gpa.free(joined.dump); + var said: [200]u8 = undefined; + std.Io.File.stderr().writeStreamingAll(io, std.fmt.bufPrint(&said, "pardes: --attach: the session's dump would not load: {t}\n", .{err}) catch "pardes: --attach: the dump would not load\n") catch {}; + std.process.exit(1); + }; + gpa.free(joined.dump); + defer r.deinit(); + r.gmtoff = joined.gmtoff; + if (!r.loaded(joined.digest)) { + std.Io.File.stderr().writeStreamingAll(io, "pardes: --attach: the session's dump loaded as another session (a build's dump is not complete)\n") catch {}; + std.process.exit(1); + } + + var winch_set = posix.sigemptyset(); + posix.sigaddset(&winch_set, posix.SIG.WINCH); + if (comptime builtin.os.tag == .freebsd) + posix.sigaction(.WINCH, &.{ .handler = .{ .handler = keepSignal }, .mask = posix.sigemptyset(), .flags = 0 }, null); + posix.sigprocmask(posix.SIG.BLOCK, &winch_set, null); + var tty_buf: [0x10000]u8 = undefined; + var screen: Screen = .{ .io = io, .gpa = gpa, .env = init.environ_map, .buf = &tty_buf, .kitty = .init(gpa), .primary = true }; + defer screen.kitty.deinit(); + try screen.enter(try vaxis.Tty.init(io, &tty_buf)); + defer screen.leave(); + defer screen.glide.deinit(gpa); + screen.feed.view = joined.view; + var paste_buf: std.Io.Writer.Allocating = .init(gpa); + defer paste_buf.deinit(); + var loop: Loop = .init(io, &screen.tty, &screen.vx); + var frame_arena: std.heap.ArenaAllocator = .init(allocs.frame); + defer frame_arena.deinit(); + var link: Remote = .{ .r = r, .fd = joined.fd }; + if (init.environ_map.get("PARDES_REPLICA_CHECK")) |file| link.check = try gpa.dupeSentinel(u8, file, 0); + defer if (link.check) |file| gpa.free(file); + var sh: Shell = .{ + .io = io, + .gpa = gpa, + .lsp_gpa = allocs.lsp, + .core = r.core(), + .prompt_rcs = undefined, + .loop = &loop, + .frame = &frame_arena, + .paste_buf = &paste_buf, + .inotify_fd = -1, + .from = joined.view, + .remote = &link, + }; + sh.terms[joined.view] = &screen; + + try startInput(&loop, &screen); + defer stopInput(io, &screen.thread); + (try std.Thread.spawn(.{}, winchWatch, .{&loop})).detach(); + var reader_done: std.atomic.Value(bool) = .init(false); + const reader = try std.Thread.spawn(.{}, linkReader, .{ gpa, joined.fd, &joined.rest, &loop, &reader_done }); + if (link.check != null) (try std.Thread.spawn(.{}, wakeWatch, .{ io, &loop })).detach(); + try screen.vx.queryTerminalSend(screen.tty.writer()); + // Its size is its view's first input. + sh.winched(); + while (!link.quit) { + if (link.dirty) link.draw(&sh); + const first = (loop.tryEvent() catch null) orelse idle: { + // Idle, its frame drawn: the next key needs no copy taken. + r.ready(); + break :idle loop.nextEvent() catch break; + }; + var ev = first; + var batch: usize = 1; + while (true) { + _ = sh.apply(ev); + if (link.quit or (!sh.in_paste and batch >= 64)) break; + ev = (loop.tryEvent() catch null) orelse break; + batch += 1; + } + r.settle(); + if (screen.caps_pending and screen.vx.queries_done.load(.unordered)) { + screen.caps_pending = false; + screen.vx.enableDetectedFeatures(screen.tty.writer()) catch {}; + screen.vx.queueRefresh(); + link.dirty = true; + } + if (link.owed) link.dirty = true; + if (link.check != null) { + // Quiet a while after a change: asked once (a wake, then the + // next quiet one). + if (first == .wake and link.changed and !link.asked and r.quiet()) { + link.changed = false; + link.asked = replica.askCheck(joined.fd); + } else if (link.changed and r.quiet()) requestWake(io, 300); + } + if (r.diverged) |reason| { + screen.leave(); + var said: [700]u8 = undefined; + std.Io.File.stderr().writeStreamingAll(io, std.fmt.bufPrint(&said, "pardes: --attach: this replica diverged from '{s}': {s}\n", .{ session, reason }) catch "pardes: --attach: diverged\n") catch {}; + std.process.exit(1); + } + } + if (init.environ_map.get("PARDES_REPLICA_STATS")) |file| link.writeStats(file); + if (!link.gone) replica.bye(joined.fd); + // The link's reader ends with the link, posting as it goes: what it + // and the terminal posted is let go until it has. + _ = libc.shutdown(joined.fd, 2); + while (true) { + while (loop.tryEvent() catch null) |ev| switch (ev) { + .unit => |b| std.heap.c_allocator.free(b), + .paste => |b| gpa.free(@constCast(b)), + else => {}, + }; + if (reader_done.load(.acquire)) break; + std.Thread.yield() catch {}; + } + reader.join(); + if (link.gone) { + screen.leave(); + var said: [300]u8 = undefined; + std.Io.File.stderr().writeStreamingAll(io, std.fmt.bufPrint(&said, "pardes: --attach: '{s}' let this client go (a Detach in it, or the session's end)\n", .{session}) catch "pardes: --attach: the session let this client go\n") catch {}; + } +} + /// A terminal a session draws on, while it has one: entered (raw, the /// alternate screen, mouse and paste reports) and left (all of it put back, /// the terminal closed) at the session's start and end, and at a detach and @@ -1462,6 +1739,10 @@ fn localSession( .test_clock = host_io.testClock(), }; sh.terms[0] = screen; + sh.peers = .{ .gpa = gpa, .io = io, .name = daemon.name(), .wake_ctx = loop, .wake = wakeForReplica }; + defer sh.peers.deinit(); + // Before the core goes: each client sees its link end. + defer sh.peers.dropAll(sh.core); pardes.lsp.setStatusSink(&sh, lspStatusSink); host_io.setExitWake(loop, wakeForExit); defer screen.glide.deinit(gpa); @@ -1524,7 +1805,7 @@ fn localSession( // An upgraded session's terminals, its own again. if (comptime !builtin.is_test) if (upgrade.resumed) |*rec| sh.resumeTerminals(rec, screen); // Nobody to see it, and nobody to size it: `size` sets it (ctl.zig). - core.unattended = !sh.attached(); + sh.setAttended(sh.attached()); while (core.nextEffect()) |effect| core.perform(effect); // ...and its shells, once a fresh session's own has a directory to // adopt them in. @@ -1564,6 +1845,8 @@ fn localSession( while (!core.quit) { pardes.turn.bump(&pardes.turn.restores); try core.pump(host); + // Its unit to each replica client, and theirs taken in. + sh.peers.boundary(core); upgrade.locate(); // Asked in the step (the word), done between steps: the session whole. if (sh.upgrade_asked) |pane| sh.upgradeNow(pane); @@ -1572,6 +1855,9 @@ fn localSession( // off a startup's way. daemon.watch(loop, deliverNote); if (core.takeRestore()) |rp| blk: { + // ponytail: a Restore replaces the core its clients replicate; + // they are let go, and attach again. + sh.peers.dropAll(core); const bytes = filesystem.readRestore(gpa, rp, core.settings.dump_dir.get()) catch |err| { core.reportError(core.active, "Restore", err); break :blk; @@ -1620,6 +1906,7 @@ fn localSession( if (live.on) clearNativeImages(&live.kitty, &live.vx, &live.tty); nc.native_images = live.on and live.vx.caps.kitty_graphics; nc.unattended = !live.on; + sh.peers.alone_unattended = !live.on; if (fs) |f| f.reset(nc); core.deinit(); core = nc; @@ -1664,9 +1951,21 @@ fn deliverNote(ctx: *anyopaque, note: daemon.Note) void { }; loop.postEvent(.{ .adopted = .{ .pid = 0, .fd = -1 } }) catch {}; }, + .replica => |fd| loop.postEvent(.{ .replica = fd }) catch { + _ = libc.close(fd); + }, } } +/// A replica reader's post: the session's wait ends, and its step +/// boundary takes the post in. +fn wakeForReplica(ctx: ?*anyopaque) void { + const loop: *Loop = @ptrCast(@alignCast(ctx.?)); + // Never waits: a full queue wakes the loop anyway, and a reader the + // loop is joining (a client let go) must not wait on it. + _ = loop.tryPostEvent(.replica_mail) catch {}; +} + const Shell = struct { io: std.Io, gpa: std.mem.Allocator, @@ -1718,9 +2017,42 @@ const Shell = struct { upgrade_asked: ?u8 = null, /// Its pty readers stopped for it: their end of file is not a shell's. upgrading: bool = false, + /// The session's replica clients (src/replica.zig), on sockets. + peers: replica.Peers = .{}, + /// This wait took only a replica's post (its input, its join): the + /// pump after it reads no new time, so it moves nothing the replicas' + /// speculations stand on (a replica takes such a unit as is). + mail_only: bool = false, + waited: bool = false, + /// This process a replica client of a session elsewhere (`remote`): + /// its inputs go to the replica, never straight to the core. + remote: ?*Remote = null, const Leaving = enum { detach, exit }; + /// A terminal's input, to the core, or to a replica client's engine + /// (applied at once there, and sent to the session). + fn feed(s: *Shell, ev: pardes.Event) void { + const c = s.remote orelse return s.core.update(ev); + const msg = c.r.input(ev) catch return; + if (!replica.sendInput(c.fd, msg)) c.quit = true; + c.dirty = true; + } + + /// Whether a terminal of this process's shows the session, or + /// nothing (a test) can. + fn local(s: *const Shell) bool { + for (s.terms) |slot| if (slot != null) return true; + return false; + } + + /// Whether anyone looks at the session: a terminal here, or a replica + /// client (Peers keeps it attended while one is there). + fn setAttended(s: *Shell, on: bool) void { + s.peers.alone_unattended = !on; + s.core.unattended = !on and s.peers.count() == 0; + } + fn of(ctx: ?*anyopaque) *Shell { return @ptrCast(@alignCast(ctx.?)); } @@ -1770,6 +2102,10 @@ const Shell = struct { fn clockNow(ctx: ?*anyopaque) u64 { const s = of(ctx); + if (s.waited) { + s.waited = false; + if (s.mail_only and s.peers.count() > 0 and s.core.clock_started) return s.core.now_ns; + } return s.test_clock orelse @intCast(std.Io.Clock.awake.now(s.io).nanoseconds); } @@ -1783,15 +2119,23 @@ const Shell = struct { return screen.on; } - /// Drawn only while a terminal shows it. + /// Drawn only while a terminal shows it: a replica client's view live + /// (its input the last), whichever terminal here shows another. fn watched(ctx: ?*anyopaque) bool { - return of(ctx).attached(); + const s = of(ctx); + if (s.term() == null and s.local()) { + for (s.terms) |slot| if (slot) |sc| if (sc.on) return true; + return false; + } + return s.attached(); } /// Detach: the terminal put back and let go, the stub told (it ends, /// and the shell has its terminal again), the session left running. fn detach(ctx: ?*anyopaque) void { const s = of(ctx); + // A replica client's: its link ends. + if (s.peers.detach(s.core.view)) return; if (!s.attached()) return s.core.reportFailure(s.core.active, "Detach: no terminal is attached"); // One of several: its view goes, and the session stays on the rest. if (pardes.views.count(s.core) > 1) return s.dropView(s.core.view, .detach); @@ -1835,7 +2179,7 @@ const Shell = struct { stopInput(s.io, &screen.thread); screen.leave(); s.core.native_images = false; - s.core.unattended = true; + s.setAttended(false); // The stub's stdin, stdout and stderr were its terminal too. const null_fd = libc.open("/dev/null", .{ .ACCMODE = .RDWR, .CLOEXEC = true }, @as(libc.mode_t, 0)); if (null_fd >= 0) { @@ -1883,7 +2227,7 @@ const Shell = struct { s.terms[k] = screen; } screen.setStub(stub); - s.core.unattended = false; + s.setAttended(true); startInput(s.loop, screen) catch {}; screen.vx.queryTerminalSend(screen.tty.writer()) catch {}; s.core.needs_frame = true; @@ -2124,6 +2468,8 @@ const Shell = struct { } s.input_seen = false; s.output_seen = false; + s.mail_only = true; + defer s.waited = true; // Paced, what is not an input (a frame's ack, a pty's output) is // taken in and the wait goes on to the tick: the acks alone ran the // loop at the terminal's round trip. @@ -2215,15 +2561,16 @@ const Shell = struct { else => {}, } if (s.input_seen) s.input_ns = @intCast(std.Io.Clock.awake.now(s.io).nanoseconds); + if (event != .replica_mail) s.mail_only = false; // A terminal's input goes into its view; a closed one's last is // dropped. switch (event) { .key_press, .mouse, .paste, .paste_start, .paste_end, .winsize, .focus_in, .focus_out, .quit => { - if (s.terms[s.from] == null and s.term() != null) { + if (s.terms[s.from] == null and s.local()) { if (event == .paste) s.gpa.free(@constCast(event.paste)); return false; } - if (s.term() != null) core.toView(s.from); + if (s.local() and s.remote == null) core.toView(s.from); }, else => {}, } @@ -2242,6 +2589,10 @@ const Shell = struct { virtual.* = @max(virtual.*, s.wake_to); }, .quit => { + if (s.remote) |c| { + c.quit = true; + return true; + } // One of several: its view goes with it. if (pardes.views.count(core) > 1) { s.dropView(core.view, .detach); @@ -2261,15 +2612,37 @@ const Shell = struct { }, .attach => |a| s.attachTerminal(a.tty, a.stub, null), .adopted => |a| s.adoptShell(a.pid, a.fd), + .replica => |fd| { + s.peers.handed(fd); + s.input_seen = true; + }, + .replica_mail => s.input_seen = true, + .unit => |bytes| { + if (s.remote) |c| { + c.r.confirm(bytes); + if (c.r.moved) c.dirty = true; + c.r.moved = false; + c.changed = true; + } + std.heap.c_allocator.free(bytes); + }, + .checked => |d| if (s.remote) |c| c.checked(d), + .link_gone => if (s.remote) |c| { + c.quit = true; + c.gone = true; + return true; + }, .focus_in => { - core.update(.{ .focus = true }); - upgrade.noticeNewer(core); + s.feed(.{ .focus = true }); + // A client's core is the session's: the notice is the session's to give. + // ponytail: not while clients replicate it (set outside a step). + if (s.remote == null and !s.peers.recording) upgrade.noticeNewer(core); }, .focus_out => { - core.update(.pointer_leave); - core.update(.{ .focus = false }); + s.feed(.pointer_leave); + s.feed(.{ .focus = false }); }, - .winsize => |ws| if (s.attached()) s.resized(ws), + .winsize => |ws| if (s.attached()) s.resized(s.terms[s.from], ws), .pty_read => |pr| { tookRead(s.io, pr.id); s.output_seen = true; @@ -2306,10 +2679,10 @@ const Shell = struct { const bytes = pasteBytes(key); const room = max_paste_bytes -| s.paste_buf.written().len; s.paste_buf.writer.writeAll(bytes[0..@min(bytes.len, room)]) catch {}; - } else core.update(vaxis_input.keyEvent(key)), + } else s.feed(vaxis_input.keyEvent(key)), .mouse => |m| { const pdf_before = nativePdfWheelTarget(core, m); - if (mouseEvent(m)) |ev| core.update(ev); + if (mouseEvent(m)) |ev| s.feed(ev); if (pdf_before) |before| { if (core.panes[before.pane]) |pane| { if (pane.pdfPage()) |page| { @@ -2319,7 +2692,7 @@ const Shell = struct { } }, .paste => |bytes| { - core.update(.{ .paste = bytes }); + s.feed(.{ .paste = bytes }); s.gpa.free(@constCast(bytes)); }, .paste_start => { @@ -2329,7 +2702,7 @@ const Shell = struct { .paste_end => { s.in_paste = false; const pasted = s.paste_buf.written(); - if (pasted.len > 0) core.update(.{ .paste = pasted }); + if (pasted.len > 0) s.feed(.{ .paste = pasted }); s.paste_buf.clearRetainingCapacity(); }, .files_changed => s.check_files = true, @@ -2366,17 +2739,13 @@ const Shell = struct { return false; } - /// The live view's terminal is `ws` big. - fn resized(s: *Shell, ws: vaxis.Winsize) void { - const core = s.core; - if (s.term()) |sc| { + /// Terminal `screen` (the live view's) is `ws` big. + fn resized(s: *Shell, screen: ?*Screen, ws: vaxis.Winsize) void { + if (screen) |sc| { sc.size = ws; sc.vx.resize(s.gpa, sc.tty.writer(), ws) catch {}; } - if (comptime pardes.pdf_enabled) - updateCoreTerminalSize(core, ws.cols, ws.rows, ws.x_pixel, ws.y_pixel) - else - core.update(.{ .resize = .{ .cols = ws.cols, .rows = ws.rows } }); + s.feed(.{ .resize = if (comptime pardes.pdf_enabled) terminalSize(ws.cols, ws.rows, ws.x_pixel, ws.y_pixel) else .{ .cols = ws.cols, .rows = ws.rows } }); } /// A SIGWINCH: each terminal whose size it changed is resized, in its @@ -2387,8 +2756,8 @@ const Shell = struct { if (!sc.on or sc.vx.state.in_band_resize) continue; const ws = sc.tty.getWinsize() catch continue; if (sc.size) |was| if (std.meta.eql(was, ws)) continue; - s.core.toView(@intCast(k)); - s.resized(ws); + if (s.remote == null) s.core.toView(@intCast(k)); + s.resized(sc, ws); } } diff --git a/test/daemon.py b/test/daemon.py index ce19b502..b54b2b54 100644 --- a/test/daemon.py +++ b/test/daemon.py @@ -7,8 +7,12 @@ private HOME and runtime directory: - a session killed outright: its shells live on, held by the daemon, and another session adopts them as terminal panes; - a replica's hello on the daemon's stream socket (phase 5): one of this - build handed to its session, which answers it and its pings; another - build's refused, and a session nobody has; + build handed to its session, which welcomes it, sends its dump and + answers its pings; another build's refused, and a session nobody has; +- replica clients (`pardes --attach=:`, phase 5 part B) of a + detached session, one through the delay relay: each one's keys in the + session's text, every client's digest the session's once quiet, one + killed outright while the other types on; - nothing left behind: no session, shell, daemon or socket. python3 -B test/daemon.py [cycles] @@ -17,6 +21,8 @@ private HOME and runtime directory: python3 -B test/daemon.py --ping [n] a hello of 's build and n pings on a replica socket (one `ssh -L` forwards, say), timed + python3 -B test/daemon.py --clients + the replica clients alone """ import fcntl import os @@ -170,7 +176,7 @@ def semantics(binary): class Replica: """A replica's connection (src/detached/wire.zig): length-prefixed messages, a tag byte and a little-endian u32 length.""" - WIRE = 14 + WIRE = 15 def __init__(self, path, timeout=10): self.s = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM) @@ -194,22 +200,34 @@ class Replica: return tag, self.take(n) def hello(self, session, build): - """('welcome', pid, name) or ('refused', why).""" + """('welcome', pid, name) or ('refused', why); a welcome's dump + is read, whole.""" name = session.encode() self.send(0x01, struct.pack(' after + + +def clients(binary): + """Replica clients of a detached session: one on the replica socket, + one through the delay relay; each one's keys in the session's text and + every client's digest the session's once quiet; one killed outright + while the other types on; nothing left behind.""" + scratch = Path(__file__).resolve().parent.parent.parent / '.scratch' / 'daemon' / 'dmn' + scratch.mkdir(parents=True, exist_ok=True) + root = Path(tempfile.mkdtemp(prefix='rc-', dir=scratch)) + env = {k: v for k, v in os.environ.items() if not k.startswith('PARDES_') and 'PAGER' not in k} + env.update(HOME=str(root), XDG_RUNTIME_DIR=str(root), XDG_CONFIG_HOME=str(root / 'config'), + TERM='xterm-256color', SHELL=shutil.which('bash') or '/bin/bash', PARDES_NOTIME='1') + (root / 'config' / 'pardes').mkdir(parents=True) + (root / 'config' / 'pardes' / 'init').write_text('Shell bash\nMessageLinger 0\n') + notes = root / 'notes.txt' + notes.write_text(''.join(f'line {i} of the notes\n' for i in range(300))) + started = [] + terms = [] + try: + sess = subprocess.Popen([binary, '--detach=files', str(notes)], env=env, cwd=root, stdin=subprocess.DEVNULL, + stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL, start_new_session=True) + started.append(sess.pid) + sock = root / 'pardes-replica.sock' + wait_for('the replica socket', sock.exists) + wait_for('the session\'s 9P socket', (root / 'pardes-9p-files.sock').exists) + nine = Client(root / 'pardes-9p-files.sock') + pane = next(int(row.split()[0]) for row in nine.read('/index').splitlines() if row.split()[1] == b'text') + alone = fds(sess.pid) + relay = subprocess.Popen([sys.executable, '-B', str(Path(__file__).parent / 'delay_relay.py'), + str(root / 'relay.sock'), str(sock), '--delay', '40'], + stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL) + started.append(relay.pid) + wait_for('the relay', (root / 'relay.sock').exists) + ta, tb = Term(30, 100), Term(24, 80) + terms += [ta, tb] + a = ta.spawn([binary, f'--attach={sock}:files'], dict(env, PARDES_REPLICA_CHECK=str(root / 'a.check'))) + b = tb.spawn([binary, f'--attach={root / "relay.sock"}:files'], dict(env, PARDES_REPLICA_CHECK=str(root / 'b.check'))) + started += [a.pid, b.pid] + for t in (ta, tb): + wait_for('a client\'s first frame', lambda: b'line 3 of the notes' in t.text(), timeout=20) + os.write(ta.master, b'Ofrom-A\x1b') + os.write(tb.master, b'jjofrom-B\x1b') + wait_for('both clients\' keys in the session\'s text', lambda: b'from-A' in nine.read(f'/pane/{pane}/body') + and b'from-B' in nine.read(f'/pane/{pane}/body')) + wait_for('client A agrees once quiet', lambda: agreed(root / 'a.check', 0)) + wait_for('client B agrees once quiet', lambda: agreed(root / 'b.check', 0)) + # Killed outright: the session and the other client carry on. + os.kill(b.pid, signal.SIGKILL) + b.wait(timeout=10) + seen = len(checks(root / 'a.check')) + os.write(ta.master, b'jjjAagain-A\x1b') + wait_for('client A\'s keys after B went', lambda: b'again-A' in nine.read(f'/pane/{pane}/body')) + wait_for('client A agrees again', lambda: agreed(root / 'a.check', seen)) + assert sess.poll() is None + # Detach typed in it: it alone goes. + os.write(ta.master, b' sD') + wait_for('client A let go', lambda: a.poll() is not None, timeout=10) + assert b'let this client go' in ta.text(), ta.text()[-300:] + # A third client, a shell: its keys wait for the session (a pty + # write), its output comes back in units, the same on both sides. + nine.write(f'/pane/{pane}/exec', b'Tty\n') + shell = wait_for('a shell pane', lambda: next((int(row.split()[0]) for row in nine.read('/index').splitlines() + if row.split()[1] == b'term'), None)) + wait_for('the shell\'s prompt', lambda: b'$' in nine.read(f'/pane/{shell}/body') or b'#' in nine.read(f'/pane/{shell}/body')) + alone = fds(sess.pid) + tc = Term(30, 100) + terms.append(tc) + c = tc.spawn([binary, f'--attach={sock}:files'], dict(env, PARDES_REPLICA_CHECK=str(root / 'c.check'))) + started.append(c.pid) + wait_for('client C\'s first frame', lambda: b'$' in tc.text(), timeout=20) + run(nine, shell, b'echo from-the-session\n') + os.write(tc.master, b'echo from-C\r') + wait_for('client C\'s keys in the shell', lambda: b'from-C' in nine.read(f'/pane/{shell}/body')) + wait_for('client C agrees with a shell\'s output', lambda: agreed(root / 'c.check', 0)) + # Its terminal hung up: it dies, and the session carries on. + c.send_signal(signal.SIGHUP) + c.wait(timeout=10) + relay.kill() + relay.wait() + wait_for('the session\'s fds level', lambda: fds(sess.pid) == alone, timeout=10) + print(f'daemon: two replica clients (one through a 40 ms relay) typed into a detached session; ' + f'digests agreed {len(checks(root / "a.check"))} and {len(checks(root / "b.check"))} times; ' + f'one killed, the other typed on and detached; a third typed into a shell, agreed ' + f'{len(checks(root / "c.check"))} times, and was hung up') + try: + nine.write(f'/pane/{pane}/exec', b'Exit\n') + except OSError as e: + # An edited file asks before it goes. + assert 'Exit again' in str(e), e + nine.write(f'/pane/{pane}/exec', b'Exit\n') + sess.wait(timeout=10) + nine.socket.close() + wait_for('nothing left', lambda: not [p for p in processes(root) if alive(p)], timeout=10) + print('daemon: replica clients\' session ended, nothing left behind') + finally: + for t in terms: + t.close() + for pid in started + processes(root): + try: + os.kill(pid, signal.SIGKILL) + except OSError: + pass + shutil.rmtree(root, ignore_errors=True) + + def terminals(client): return [int(row.split()[0]) for row in client.read('/index').splitlines() if row.split()[1] == b'term'] @@ -290,6 +428,7 @@ def main(binary, cycles): daemon = daemons[0] started.append(daemon) replicas(root, binary, session, daemon) + clients(binary) shell = terminals(client)[0] shell_pid = int(run(client, shell, b'echo $$\n').split(b'\n')[1]) started.append(shell_pid) @@ -856,6 +995,8 @@ if __name__ == '__main__': gui_upgrades(Path(sys.argv[3]).resolve(), Path(sys.argv[4]).resolve()) elif sys.argv[1:2] == ['--upgrade']: upgrades(Path(sys.argv[2]).resolve(), Path(sys.argv[3]).resolve()) + elif sys.argv[1:2] == ['--clients']: + clients(sys.argv[2]) elif sys.argv[1:2] == ['--ping']: ping(sys.argv[2], sys.argv[3], sys.argv[4], int(sys.argv[5]) if len(sys.argv) > 5 else 50) elif '--gui' in sys.argv: diff --git a/test/rollback_bench.zig b/test/rollback_bench.zig index 38fe011d..0f0c77a4 100644 --- a/test/rollback_bench.zig +++ b/test/rollback_bench.zig @@ -158,6 +158,24 @@ pub fn main(init: std.process.Init) !void { x.* = t1 - t0; y.* = nowNs() - t1; } + // A join (docs/next-daemon.md, Phase 5 part B notes): the session + // dumped at a step boundary, which is what the authority pays, and + // the joiner's load of it. + var dump: [5]u64 = undefined; + var load: [5]u64 = undefined; + var dump_len: usize = 0; + for (&dump, &load) |*x, *y| { + const t0 = nowNs(); + const bytes = try pardes.dump.dumpBytes(core); + const t1 = nowNs(); + dump_len = bytes.len; + const loaded = try pardes.dump.load(gpa, .{ .tty_only = true, .image_allocator = gpa, .pdf_allocator = gpa, .tree_sitter_allocator = gpa, .frame_allocator = gpa }, bytes, null); + y.* = nowNs() - t1; + x.* = t1 - t0; + loaded.deinit(); + core.gpa.free(bytes); + } + std.debug.print("{s:<8} join: dump {d:.1} MB in {d:.2} ms (the session's step), load {d:.2} ms (the joiner's)\n", .{ s.name, @as(f64, @floatFromInt(dump_len)) / 1e6, @as(f64, @floatFromInt(median(&dump))) * 1e-6, @as(f64, @floatFromInt(median(&load))) * 1e-6 }); var panes: usize = 0; for (core.panes) |slot| panes += @intFromBool(slot != null); const ms = 1e-6; -- cgit v1.3