| ... | @@ -9,23 +9,795 @@ const net = std.Io.net; | ... | @@ -9,23 +9,795 @@ const net = std.Io.net; |
| 9 | const assert = std.debug.assert; | 9 | const assert = std.debug.assert; |
| 10 | const Allocator = std.mem.Allocator; | 10 | const Allocator = std.mem.Allocator; |
| 11 | const Alignment = std.mem.Alignment; | 11 | const Alignment = std.mem.Alignment; |
| 12 | const posix = std.posix; | | |
| 13 | const IpAddress = std.Io.net.IpAddress; | 12 | const IpAddress = std.Io.net.IpAddress; |
| 14 | const errnoBug = std.Io.Threaded.errnoBug; | 13 | const errnoBug = std.Io.Threaded.errnoBug; |
| | 14 | const posix = std.posix; |
| 15 | | 15 | |
| 16 | /// Must be a thread-safe allocator. | 16 | /// Must be a thread-safe allocator. |
| 17 | gpa: Allocator, | 17 | gpa: Allocator, |
| | 18 | mutex: std.Thread.Mutex, |
| | 19 | main_fiber_buffer: [@sizeOf(Fiber) + Fiber.max_result_size]u8 align(@alignOf(Fiber)), |
| | 20 | threads: Thread.List, |
| 18 | | 21 | |
| 19 | pub fn init(gpa: Allocator) Kqueue { | 22 | /// Empirically saw >128KB being used by the self-hosted backend to panic. |
| 20 | return .{ | 23 | const idle_stack_size = 256 * 1024; |
| | 24 | |
| | 25 | const max_idle_search = 4; |
| | 26 | const max_steal_ready_search = 4; |
| | 27 | |
| | 28 | const changes_buffer_len = 64; |
| | 29 | |
| | 30 | const Thread = struct { |
| | 31 | thread: std.Thread, |
| | 32 | idle_context: Context, |
| | 33 | current_context: *Context, |
| | 34 | ready_queue: ?*Fiber, |
| | 35 | kq_fd: posix.fd_t, |
| | 36 | idle_search_index: u32, |
| | 37 | steal_ready_search_index: u32, |
| | 38 | |
| | 39 | const canceling: ?*Thread = @ptrFromInt(@alignOf(Thread)); |
| | 40 | |
| | 41 | threadlocal var self: *Thread = undefined; |
| | 42 | |
| | 43 | fn current() *Thread { |
| | 44 | return self; |
| | 45 | } |
| | 46 | |
| | 47 | fn currentFiber(thread: *Thread) *Fiber { |
| | 48 | return @fieldParentPtr("context", thread.current_context); |
| | 49 | } |
| | 50 | |
| | 51 | const List = struct { |
| | 52 | allocated: []Thread, |
| | 53 | reserved: u32, |
| | 54 | active: u32, |
| | 55 | }; |
| | 56 | }; |
| | 57 | |
| | 58 | const Fiber = struct { |
| | 59 | required_align: void align(4), |
| | 60 | context: Context, |
| | 61 | awaiter: ?*Fiber, |
| | 62 | queue_next: ?*Fiber, |
| | 63 | cancel_thread: ?*Thread, |
| | 64 | awaiting_completions: std.StaticBitSet(3), |
| | 65 | |
| | 66 | const finished: ?*Fiber = @ptrFromInt(@alignOf(Thread)); |
| | 67 | |
| | 68 | const max_result_align: Alignment = .@"16"; |
| | 69 | const max_result_size = max_result_align.forward(64); |
| | 70 | /// This includes any stack realignments that need to happen, and also the |
| | 71 | /// initial frame return address slot and argument frame, depending on target. |
| | 72 | const min_stack_size = 4 * 1024 * 1024; |
| | 73 | const max_context_align: Alignment = .@"16"; |
| | 74 | const max_context_size = max_context_align.forward(1024); |
| | 75 | const max_closure_size: usize = @sizeOf(AsyncClosure); |
| | 76 | const max_closure_align: Alignment = .of(AsyncClosure); |
| | 77 | const allocation_size = std.mem.alignForward( |
| | 78 | usize, |
| | 79 | max_closure_align.max(max_context_align).forward( |
| | 80 | max_result_align.forward(@sizeOf(Fiber)) + max_result_size + min_stack_size, |
| | 81 | ) + max_closure_size + max_context_size, |
| | 82 | std.heap.page_size_max, |
| | 83 | ); |
| | 84 | |
| | 85 | fn allocate(k: *Kqueue) error{OutOfMemory}!*Fiber { |
| | 86 | return @ptrCast(try k.gpa.alignedAlloc(u8, .of(Fiber), allocation_size)); |
| | 87 | } |
| | 88 | |
| | 89 | fn allocatedSlice(f: *Fiber) []align(@alignOf(Fiber)) u8 { |
| | 90 | return @as([*]align(@alignOf(Fiber)) u8, @ptrCast(f))[0..allocation_size]; |
| | 91 | } |
| | 92 | |
| | 93 | fn allocatedEnd(f: *Fiber) [*]u8 { |
| | 94 | const allocated_slice = f.allocatedSlice(); |
| | 95 | return allocated_slice[allocated_slice.len..].ptr; |
| | 96 | } |
| | 97 | |
| | 98 | fn resultPointer(f: *Fiber, comptime Result: type) *Result { |
| | 99 | return @ptrCast(@alignCast(f.resultBytes(.of(Result)))); |
| | 100 | } |
| | 101 | |
| | 102 | fn resultBytes(f: *Fiber, alignment: Alignment) [*]u8 { |
| | 103 | return @ptrFromInt(alignment.forward(@intFromPtr(f) + @sizeOf(Fiber))); |
| | 104 | } |
| | 105 | |
| | 106 | fn enterCancelRegion(fiber: *Fiber, thread: *Thread) error{Canceled}!void { |
| | 107 | if (@cmpxchgStrong( |
| | 108 | ?*Thread, |
| | 109 | &fiber.cancel_thread, |
| | 110 | null, |
| | 111 | thread, |
| | 112 | .acq_rel, |
| | 113 | .acquire, |
| | 114 | )) |cancel_thread| { |
| | 115 | assert(cancel_thread == Thread.canceling); |
| | 116 | return error.Canceled; |
| | 117 | } |
| | 118 | } |
| | 119 | |
| | 120 | fn exitCancelRegion(fiber: *Fiber, thread: *Thread) void { |
| | 121 | if (@cmpxchgStrong( |
| | 122 | ?*Thread, |
| | 123 | &fiber.cancel_thread, |
| | 124 | thread, |
| | 125 | null, |
| | 126 | .acq_rel, |
| | 127 | .acquire, |
| | 128 | )) |cancel_thread| assert(cancel_thread == Thread.canceling); |
| | 129 | } |
| | 130 | |
| | 131 | const Queue = struct { head: *Fiber, tail: *Fiber }; |
| | 132 | }; |
| | 133 | |
| | 134 | fn recycle(k: *Kqueue, fiber: *Fiber) void { |
| | 135 | std.log.debug("recyling {*}", .{fiber}); |
| | 136 | assert(fiber.queue_next == null); |
| | 137 | k.gpa.free(fiber.allocatedSlice()); |
| | 138 | } |
| | 139 | |
| | 140 | pub fn init(k: *Kqueue, gpa: Allocator) !void { |
| | 141 | const threads_size = @max(std.Thread.getCpuCount() catch 1, 1) * @sizeOf(Thread); |
| | 142 | const idle_stack_end_offset = std.mem.alignForward(usize, threads_size + idle_stack_size, std.heap.page_size_max); |
| | 143 | const allocated_slice = try gpa.alignedAlloc(u8, .of(Thread), idle_stack_end_offset); |
| | 144 | errdefer gpa.free(allocated_slice); |
| | 145 | k.* = .{ |
| 21 | .gpa = gpa, | 146 | .gpa = gpa, |
| | 147 | .mutex = .{}, |
| | 148 | .main_fiber_buffer = undefined, |
| | 149 | .threads = .{ |
| | 150 | .allocated = @ptrCast(allocated_slice[0..threads_size]), |
| | 151 | .reserved = 1, |
| | 152 | .active = 1, |
| | 153 | }, |
| 22 | }; | 154 | }; |
| | 155 | const main_fiber: *Fiber = @ptrCast(&k.main_fiber_buffer); |
| | 156 | main_fiber.* = .{ |
| | 157 | .required_align = {}, |
| | 158 | .context = undefined, |
| | 159 | .awaiter = null, |
| | 160 | .queue_next = null, |
| | 161 | .cancel_thread = null, |
| | 162 | .awaiting_completions = .initEmpty(), |
| | 163 | }; |
| | 164 | const main_thread = &k.threads.allocated[0]; |
| | 165 | Thread.self = main_thread; |
| | 166 | const idle_stack_end: [*]align(16) usize = @ptrCast(@alignCast(allocated_slice[idle_stack_end_offset..].ptr)); |
| | 167 | (idle_stack_end - 1)[0..1].* = .{@intFromPtr(k)}; |
| | 168 | main_thread.* = .{ |
| | 169 | .thread = undefined, |
| | 170 | .idle_context = switch (builtin.cpu.arch) { |
| | 171 | .aarch64 => .{ |
| | 172 | .sp = @intFromPtr(idle_stack_end), |
| | 173 | .fp = 0, |
| | 174 | .pc = @intFromPtr(&mainIdleEntry), |
| | 175 | .x18 = asm ("" |
| | 176 | : [x18] "={x18}" (-> u64), |
| | 177 | ), |
| | 178 | }, |
| | 179 | .x86_64 => .{ |
| | 180 | .rsp = @intFromPtr(idle_stack_end - 1), |
| | 181 | .rbp = 0, |
| | 182 | .rip = @intFromPtr(&mainIdleEntry), |
| | 183 | }, |
| | 184 | else => @compileError("unimplemented architecture"), |
| | 185 | }, |
| | 186 | .current_context = &main_fiber.context, |
| | 187 | .ready_queue = null, |
| | 188 | .kq_fd = try posix.kqueue(), |
| | 189 | .idle_search_index = 1, |
| | 190 | .steal_ready_search_index = 1, |
| | 191 | }; |
| | 192 | errdefer std.posix.close(main_thread.kq_fd); |
| | 193 | std.log.debug("created main idle {*}", .{&main_thread.idle_context}); |
| | 194 | std.log.debug("created main {*}", .{main_fiber}); |
| 23 | } | 195 | } |
| 24 | | 196 | |
| 25 | pub fn deinit(k: *Kqueue) void { | 197 | pub fn deinit(k: *Kqueue) void { |
| | 198 | const active_threads = @atomicLoad(u32, &k.threads.active, .acquire); |
| | 199 | for (k.threads.allocated[0..active_threads]) |*thread| { |
| | 200 | const ready_fiber = @atomicLoad(?*Fiber, &thread.ready_queue, .monotonic); |
| | 201 | assert(ready_fiber == null or ready_fiber == Fiber.finished); // pending async |
| | 202 | } |
| | 203 | k.yield(null, .exit); |
| | 204 | const allocated_ptr: [*]align(@alignOf(Thread)) u8 = @ptrCast(@alignCast(k.threads.allocated.ptr)); |
| | 205 | const idle_stack_end_offset = std.mem.alignForward(usize, k.threads.allocated.len * @sizeOf(Thread) + idle_stack_size, std.heap.page_size_max); |
| | 206 | for (k.threads.allocated[1..active_threads]) |*thread| thread.thread.join(); |
| | 207 | k.gpa.free(allocated_ptr[0..idle_stack_end_offset]); |
| 26 | k.* = undefined; | 208 | k.* = undefined; |
| 27 | } | 209 | } |
| 28 | | 210 | |
| | 211 | fn findReadyFiber(k: *Kqueue, thread: *Thread) ?*Fiber { |
| | 212 | if (@atomicRmw(?*Fiber, &thread.ready_queue, .Xchg, Fiber.finished, .acquire)) |ready_fiber| { |
| | 213 | @atomicStore(?*Fiber, &thread.ready_queue, ready_fiber.queue_next, .release); |
| | 214 | ready_fiber.queue_next = null; |
| | 215 | return ready_fiber; |
| | 216 | } |
| | 217 | const active_threads = @atomicLoad(u32, &k.threads.active, .acquire); |
| | 218 | for (0..@min(max_steal_ready_search, active_threads)) |_| { |
| | 219 | defer thread.steal_ready_search_index += 1; |
| | 220 | if (thread.steal_ready_search_index == active_threads) thread.steal_ready_search_index = 0; |
| | 221 | const steal_ready_search_thread = &k.threads.allocated[0..active_threads][thread.steal_ready_search_index]; |
| | 222 | if (steal_ready_search_thread == thread) continue; |
| | 223 | const ready_fiber = @atomicLoad(?*Fiber, &steal_ready_search_thread.ready_queue, .acquire) orelse continue; |
| | 224 | if (ready_fiber == Fiber.finished) continue; |
| | 225 | if (@cmpxchgWeak( |
| | 226 | ?*Fiber, |
| | 227 | &steal_ready_search_thread.ready_queue, |
| | 228 | ready_fiber, |
| | 229 | null, |
| | 230 | .acquire, |
| | 231 | .monotonic, |
| | 232 | )) |_| continue; |
| | 233 | @atomicStore(?*Fiber, &thread.ready_queue, ready_fiber.queue_next, .release); |
| | 234 | ready_fiber.queue_next = null; |
| | 235 | return ready_fiber; |
| | 236 | } |
| | 237 | // couldn't find anything to do, so we are now open for business |
| | 238 | @atomicStore(?*Fiber, &thread.ready_queue, null, .monotonic); |
| | 239 | return null; |
| | 240 | } |
| | 241 | |
| | 242 | fn yield(k: *Kqueue, maybe_ready_fiber: ?*Fiber, pending_task: SwitchMessage.PendingTask) void { |
| | 243 | const thread: *Thread = .current(); |
| | 244 | const ready_context = if (maybe_ready_fiber orelse k.findReadyFiber(thread)) |ready_fiber| |
| | 245 | &ready_fiber.context |
| | 246 | else |
| | 247 | &thread.idle_context; |
| | 248 | const message: SwitchMessage = .{ |
| | 249 | .contexts = .{ |
| | 250 | .prev = thread.current_context, |
| | 251 | .ready = ready_context, |
| | 252 | }, |
| | 253 | .pending_task = pending_task, |
| | 254 | }; |
| | 255 | std.log.debug("switching from {*} to {*}", .{ message.contexts.prev, message.contexts.ready }); |
| | 256 | contextSwitch(&message).handle(k); |
| | 257 | } |
| | 258 | |
| | 259 | fn schedule(k: *Kqueue, thread: *Thread, ready_queue: Fiber.Queue) void { |
| | 260 | { |
| | 261 | var fiber = ready_queue.head; |
| | 262 | while (true) { |
| | 263 | std.log.debug("scheduling {*}", .{fiber}); |
| | 264 | fiber = fiber.queue_next orelse break; |
| | 265 | } |
| | 266 | assert(fiber == ready_queue.tail); |
| | 267 | } |
| | 268 | // shared fields of previous `Thread` must be initialized before later ones are marked as active |
| | 269 | const new_thread_index = @atomicLoad(u32, &k.threads.active, .acquire); |
| | 270 | for (0..@min(max_idle_search, new_thread_index)) |_| { |
| | 271 | defer thread.idle_search_index += 1; |
| | 272 | if (thread.idle_search_index == new_thread_index) thread.idle_search_index = 0; |
| | 273 | const idle_search_thread = &k.threads.allocated[0..new_thread_index][thread.idle_search_index]; |
| | 274 | if (idle_search_thread == thread) continue; |
| | 275 | if (@cmpxchgWeak( |
| | 276 | ?*Fiber, |
| | 277 | &idle_search_thread.ready_queue, |
| | 278 | null, |
| | 279 | ready_queue.head, |
| | 280 | .release, |
| | 281 | .monotonic, |
| | 282 | )) |_| continue; |
| | 283 | const changes = [_]posix.Kevent{ |
| | 284 | .{ |
| | 285 | .ident = 0, |
| | 286 | .filter = std.c.EVFILT.USER, |
| | 287 | .flags = std.c.EV.ADD | std.c.EV.ONESHOT, |
| | 288 | .fflags = std.c.NOTE.TRIGGER, |
| | 289 | .data = 0, |
| | 290 | .udata = @intFromEnum(Completion.UserData.wakeup), |
| | 291 | }, |
| | 292 | }; |
| | 293 | // If an error occurs it only pessimises scheduling. |
| | 294 | _ = posix.kevent(idle_search_thread.kq_fd, &changes, &.{}, null) catch {}; |
| | 295 | return; |
| | 296 | } |
| | 297 | spawn_thread: { |
| | 298 | // previous failed reservations must have completed before retrying |
| | 299 | if (new_thread_index == k.threads.allocated.len or @cmpxchgWeak( |
| | 300 | u32, |
| | 301 | &k.threads.reserved, |
| | 302 | new_thread_index, |
| | 303 | new_thread_index + 1, |
| | 304 | .acquire, |
| | 305 | .monotonic, |
| | 306 | ) != null) break :spawn_thread; |
| | 307 | const new_thread = &k.threads.allocated[new_thread_index]; |
| | 308 | const next_thread_index = new_thread_index + 1; |
| | 309 | new_thread.* = .{ |
| | 310 | .thread = undefined, |
| | 311 | .idle_context = undefined, |
| | 312 | .current_context = &new_thread.idle_context, |
| | 313 | .ready_queue = ready_queue.head, |
| | 314 | .kq_fd = posix.kqueue() catch |err| { |
| | 315 | @atomicStore(u32, &k.threads.reserved, new_thread_index, .release); |
| | 316 | // no more access to `thread` after giving up reservation |
| | 317 | std.log.warn("unable to create worker thread due to kqueue init failure: {t}", .{err}); |
| | 318 | break :spawn_thread; |
| | 319 | }, |
| | 320 | .idle_search_index = 0, |
| | 321 | .steal_ready_search_index = 0, |
| | 322 | }; |
| | 323 | new_thread.thread = std.Thread.spawn(.{ |
| | 324 | .stack_size = idle_stack_size, |
| | 325 | .allocator = k.gpa, |
| | 326 | }, threadEntry, .{ k, new_thread_index }) catch |err| { |
| | 327 | posix.close(new_thread.kq_fd); |
| | 328 | @atomicStore(u32, &k.threads.reserved, new_thread_index, .release); |
| | 329 | // no more access to `thread` after giving up reservation |
| | 330 | std.log.warn("unable to create worker thread due spawn failure: {s}", .{@errorName(err)}); |
| | 331 | break :spawn_thread; |
| | 332 | }; |
| | 333 | // shared fields of `Thread` must be initialized before being marked active |
| | 334 | @atomicStore(u32, &k.threads.active, next_thread_index, .release); |
| | 335 | return; |
| | 336 | } |
| | 337 | // nobody wanted it, so just queue it on ourselves |
| | 338 | while (@cmpxchgWeak( |
| | 339 | ?*Fiber, |
| | 340 | &thread.ready_queue, |
| | 341 | ready_queue.tail.queue_next, |
| | 342 | ready_queue.head, |
| | 343 | .acq_rel, |
| | 344 | .acquire, |
| | 345 | )) |old_head| ready_queue.tail.queue_next = old_head; |
| | 346 | } |
| | 347 | |
| | 348 | fn mainIdle(k: *Kqueue, message: *const SwitchMessage) callconv(.withStackAlign(.c, @max(@alignOf(Thread), @alignOf(Context)))) noreturn { |
| | 349 | message.handle(k); |
| | 350 | k.idle(&k.threads.allocated[0]); |
| | 351 | k.yield(@ptrCast(&k.main_fiber_buffer), .nothing); |
| | 352 | unreachable; // switched to dead fiber |
| | 353 | } |
| | 354 | |
| | 355 | fn threadEntry(k: *Kqueue, index: u32) void { |
| | 356 | const thread: *Thread = &k.threads.allocated[index]; |
| | 357 | Thread.self = thread; |
| | 358 | std.log.debug("created thread idle {*}", .{&thread.idle_context}); |
| | 359 | k.idle(thread); |
| | 360 | } |
| | 361 | |
| | 362 | const Completion = struct { |
| | 363 | const UserData = enum(usize) { |
| | 364 | unused, |
| | 365 | wakeup, |
| | 366 | cleanup, |
| | 367 | exit, |
| | 368 | /// *Fiber |
| | 369 | _, |
| | 370 | }; |
| | 371 | /// Corresponds to Kevent field. |
| | 372 | flags: u16, |
| | 373 | /// Corresponds to Kevent field. |
| | 374 | fflags: u32, |
| | 375 | /// Corresponds to Kevent field. |
| | 376 | data: isize, |
| | 377 | }; |
| | 378 | |
| | 379 | fn idle(k: *Kqueue, thread: *Thread) void { |
| | 380 | var events_buffer: [changes_buffer_len]posix.Kevent = undefined; |
| | 381 | var maybe_ready_fiber: ?*Fiber = null; |
| | 382 | while (true) { |
| | 383 | while (maybe_ready_fiber orelse k.findReadyFiber(thread)) |ready_fiber| { |
| | 384 | k.yield(ready_fiber, .nothing); |
| | 385 | maybe_ready_fiber = null; |
| | 386 | } |
| | 387 | const n = posix.kevent(thread.kq_fd, &.{}, &events_buffer, null) catch |err| { |
| | 388 | // TODO handle EINTR for cancellation purposes |
| | 389 | @panic(@errorName(err)); |
| | 390 | }; |
| | 391 | var maybe_ready_queue: ?Fiber.Queue = null; |
| | 392 | for (events_buffer[0..n]) |event| switch (@as(Completion.UserData, @enumFromInt(event.udata))) { |
| | 393 | .unused => unreachable, // bad submission queued? |
| | 394 | .wakeup => {}, |
| | 395 | .cleanup => @panic("failed to notify other threads that we are exiting"), |
| | 396 | .exit => { |
| | 397 | assert(maybe_ready_fiber == null and maybe_ready_queue == null); // pending async |
| | 398 | return; |
| | 399 | }, |
| | 400 | _ => { |
| | 401 | const fiber: *Fiber = @ptrFromInt(event.udata); |
| | 402 | assert(fiber.queue_next == null); |
| | 403 | fiber.resultPointer(Completion).* = .{ |
| | 404 | .flags = event.flags, |
| | 405 | .fflags = event.fflags, |
| | 406 | .data = event.data, |
| | 407 | }; |
| | 408 | if (maybe_ready_fiber == null) maybe_ready_fiber = fiber else if (maybe_ready_queue) |*ready_queue| { |
| | 409 | ready_queue.tail.queue_next = fiber; |
| | 410 | ready_queue.tail = fiber; |
| | 411 | } else maybe_ready_queue = .{ .head = fiber, .tail = fiber }; |
| | 412 | }, |
| | 413 | }; |
| | 414 | if (maybe_ready_queue) |ready_queue| k.schedule(thread, ready_queue); |
| | 415 | } |
| | 416 | } |
| | 417 | |
| | 418 | const SwitchMessage = struct { |
| | 419 | contexts: extern struct { |
| | 420 | prev: *Context, |
| | 421 | ready: *Context, |
| | 422 | }, |
| | 423 | pending_task: PendingTask, |
| | 424 | |
| | 425 | const PendingTask = union(enum) { |
| | 426 | nothing, |
| | 427 | reschedule, |
| | 428 | recycle: *Fiber, |
| | 429 | register_awaiter: *?*Fiber, |
| | 430 | register_select: []const *Io.AnyFuture, |
| | 431 | mutex_lock: struct { |
| | 432 | prev_state: Io.Mutex.State, |
| | 433 | mutex: *Io.Mutex, |
| | 434 | }, |
| | 435 | condition_wait: struct { |
| | 436 | cond: *Io.Condition, |
| | 437 | mutex: *Io.Mutex, |
| | 438 | }, |
| | 439 | exit, |
| | 440 | }; |
| | 441 | |
| | 442 | fn handle(message: *const SwitchMessage, k: *Kqueue) void { |
| | 443 | const thread: *Thread = .current(); |
| | 444 | thread.current_context = message.contexts.ready; |
| | 445 | switch (message.pending_task) { |
| | 446 | .nothing => {}, |
| | 447 | .reschedule => if (message.contexts.prev != &thread.idle_context) { |
| | 448 | const prev_fiber: *Fiber = @alignCast(@fieldParentPtr("context", message.contexts.prev)); |
| | 449 | assert(prev_fiber.queue_next == null); |
| | 450 | k.schedule(thread, .{ .head = prev_fiber, .tail = prev_fiber }); |
| | 451 | }, |
| | 452 | .recycle => |fiber| { |
| | 453 | k.recycle(fiber); |
| | 454 | }, |
| | 455 | .register_awaiter => |awaiter| { |
| | 456 | const prev_fiber: *Fiber = @alignCast(@fieldParentPtr("context", message.contexts.prev)); |
| | 457 | assert(prev_fiber.queue_next == null); |
| | 458 | if (@atomicRmw(?*Fiber, awaiter, .Xchg, prev_fiber, .acq_rel) == Fiber.finished) |
| | 459 | k.schedule(thread, .{ .head = prev_fiber, .tail = prev_fiber }); |
| | 460 | }, |
| | 461 | .register_select => |futures| { |
| | 462 | const prev_fiber: *Fiber = @alignCast(@fieldParentPtr("context", message.contexts.prev)); |
| | 463 | assert(prev_fiber.queue_next == null); |
| | 464 | for (futures) |any_future| { |
| | 465 | const future_fiber: *Fiber = @ptrCast(@alignCast(any_future)); |
| | 466 | if (@atomicRmw(?*Fiber, &future_fiber.awaiter, .Xchg, prev_fiber, .acq_rel) == Fiber.finished) { |
| | 467 | const closure: *AsyncClosure = .fromFiber(future_fiber); |
| | 468 | if (!@atomicRmw(bool, &closure.already_awaited, .Xchg, true, .seq_cst)) { |
| | 469 | k.schedule(thread, .{ .head = prev_fiber, .tail = prev_fiber }); |
| | 470 | } |
| | 471 | } |
| | 472 | } |
| | 473 | }, |
| | 474 | .mutex_lock => |mutex_lock| { |
| | 475 | const prev_fiber: *Fiber = @alignCast(@fieldParentPtr("context", message.contexts.prev)); |
| | 476 | assert(prev_fiber.queue_next == null); |
| | 477 | var prev_state = mutex_lock.prev_state; |
| | 478 | while (switch (prev_state) { |
| | 479 | else => next_state: { |
| | 480 | prev_fiber.queue_next = @ptrFromInt(@intFromEnum(prev_state)); |
| | 481 | break :next_state @cmpxchgWeak( |
| | 482 | Io.Mutex.State, |
| | 483 | &mutex_lock.mutex.state, |
| | 484 | prev_state, |
| | 485 | @enumFromInt(@intFromPtr(prev_fiber)), |
| | 486 | .release, |
| | 487 | .acquire, |
| | 488 | ); |
| | 489 | }, |
| | 490 | .unlocked => @cmpxchgWeak( |
| | 491 | Io.Mutex.State, |
| | 492 | &mutex_lock.mutex.state, |
| | 493 | .unlocked, |
| | 494 | .locked_once, |
| | 495 | .acquire, |
| | 496 | .acquire, |
| | 497 | ) orelse { |
| | 498 | prev_fiber.queue_next = null; |
| | 499 | k.schedule(thread, .{ .head = prev_fiber, .tail = prev_fiber }); |
| | 500 | return; |
| | 501 | }, |
| | 502 | }) |next_state| prev_state = next_state; |
| | 503 | }, |
| | 504 | .condition_wait => |condition_wait| { |
| | 505 | const prev_fiber: *Fiber = @alignCast(@fieldParentPtr("context", message.contexts.prev)); |
| | 506 | assert(prev_fiber.queue_next == null); |
| | 507 | const cond_impl = prev_fiber.resultPointer(Condition); |
| | 508 | cond_impl.* = .{ |
| | 509 | .tail = prev_fiber, |
| | 510 | .event = .queued, |
| | 511 | }; |
| | 512 | if (@cmpxchgStrong( |
| | 513 | ?*Fiber, |
| | 514 | @as(*?*Fiber, @ptrCast(&condition_wait.cond.state)), |
| | 515 | null, |
| | 516 | prev_fiber, |
| | 517 | .release, |
| | 518 | .acquire, |
| | 519 | )) |waiting_fiber| { |
| | 520 | const waiting_cond_impl = waiting_fiber.?.resultPointer(Condition); |
| | 521 | assert(waiting_cond_impl.tail.queue_next == null); |
| | 522 | waiting_cond_impl.tail.queue_next = prev_fiber; |
| | 523 | waiting_cond_impl.tail = prev_fiber; |
| | 524 | } |
| | 525 | condition_wait.mutex.unlock(k.io()); |
| | 526 | }, |
| | 527 | .exit => for (k.threads.allocated[0..@atomicLoad(u32, &k.threads.active, .acquire)]) |*each_thread| { |
| | 528 | const changes = [_]posix.Kevent{ |
| | 529 | .{ |
| | 530 | .ident = 0, |
| | 531 | .filter = std.c.EVFILT.USER, |
| | 532 | .flags = std.c.EV.ADD | std.c.EV.ONESHOT, |
| | 533 | .fflags = std.c.NOTE.TRIGGER, |
| | 534 | .data = 0, |
| | 535 | .udata = @intFromEnum(Completion.UserData.exit), |
| | 536 | }, |
| | 537 | }; |
| | 538 | _ = posix.kevent(each_thread.kq_fd, &changes, &.{}, null) catch |err| { |
| | 539 | @panic(@errorName(err)); |
| | 540 | }; |
| | 541 | }, |
| | 542 | } |
| | 543 | } |
| | 544 | }; |
| | 545 | |
| | 546 | const Context = switch (builtin.cpu.arch) { |
| | 547 | .aarch64 => extern struct { |
| | 548 | sp: u64, |
| | 549 | fp: u64, |
| | 550 | pc: u64, |
| | 551 | x18: u64, |
| | 552 | }, |
| | 553 | .x86_64 => extern struct { |
| | 554 | rsp: u64, |
| | 555 | rbp: u64, |
| | 556 | rip: u64, |
| | 557 | }, |
| | 558 | else => |arch| @compileError("unimplemented architecture: " ++ @tagName(arch)), |
| | 559 | }; |
| | 560 | |
| | 561 | inline fn contextSwitch(message: *const SwitchMessage) *const SwitchMessage { |
| | 562 | return @fieldParentPtr("contexts", switch (builtin.cpu.arch) { |
| | 563 | .aarch64 => asm volatile ( |
| | 564 | \\ ldp x0, x2, [x1] |
| | 565 | \\ ldp x3, x18, [x2, #16] |
| | 566 | \\ mov x4, sp |
| | 567 | \\ stp x4, fp, [x0] |
| | 568 | \\ adr x5, 0f |
| | 569 | \\ ldp x4, fp, [x2] |
| | 570 | \\ stp x5, x18, [x0, #16] |
| | 571 | \\ mov sp, x4 |
| | 572 | \\ br x3 |
| | 573 | \\0: |
| | 574 | : [received_message] "={x1}" (-> *const @FieldType(SwitchMessage, "contexts")), |
| | 575 | : [message_to_send] "{x1}" (&message.contexts), |
| | 576 | : .{ |
| | 577 | .x0 = true, |
| | 578 | .x1 = true, |
| | 579 | .x2 = true, |
| | 580 | .x3 = true, |
| | 581 | .x4 = true, |
| | 582 | .x5 = true, |
| | 583 | .x6 = true, |
| | 584 | .x7 = true, |
| | 585 | .x8 = true, |
| | 586 | .x9 = true, |
| | 587 | .x10 = true, |
| | 588 | .x11 = true, |
| | 589 | .x12 = true, |
| | 590 | .x13 = true, |
| | 591 | .x14 = true, |
| | 592 | .x15 = true, |
| | 593 | .x16 = true, |
| | 594 | .x17 = true, |
| | 595 | .x19 = true, |
| | 596 | .x20 = true, |
| | 597 | .x21 = true, |
| | 598 | .x22 = true, |
| | 599 | .x23 = true, |
| | 600 | .x24 = true, |
| | 601 | .x25 = true, |
| | 602 | .x26 = true, |
| | 603 | .x27 = true, |
| | 604 | .x28 = true, |
| | 605 | .x30 = true, |
| | 606 | .z0 = true, |
| | 607 | .z1 = true, |
| | 608 | .z2 = true, |
| | 609 | .z3 = true, |
| | 610 | .z4 = true, |
| | 611 | .z5 = true, |
| | 612 | .z6 = true, |
| | 613 | .z7 = true, |
| | 614 | .z8 = true, |
| | 615 | .z9 = true, |
| | 616 | .z10 = true, |
| | 617 | .z11 = true, |
| | 618 | .z12 = true, |
| | 619 | .z13 = true, |
| | 620 | .z14 = true, |
| | 621 | .z15 = true, |
| | 622 | .z16 = true, |
| | 623 | .z17 = true, |
| | 624 | .z18 = true, |
| | 625 | .z19 = true, |
| | 626 | .z20 = true, |
| | 627 | .z21 = true, |
| | 628 | .z22 = true, |
| | 629 | .z23 = true, |
| | 630 | .z24 = true, |
| | 631 | .z25 = true, |
| | 632 | .z26 = true, |
| | 633 | .z27 = true, |
| | 634 | .z28 = true, |
| | 635 | .z29 = true, |
| | 636 | .z30 = true, |
| | 637 | .z31 = true, |
| | 638 | .p0 = true, |
| | 639 | .p1 = true, |
| | 640 | .p2 = true, |
| | 641 | .p3 = true, |
| | 642 | .p4 = true, |
| | 643 | .p5 = true, |
| | 644 | .p6 = true, |
| | 645 | .p7 = true, |
| | 646 | .p8 = true, |
| | 647 | .p9 = true, |
| | 648 | .p10 = true, |
| | 649 | .p11 = true, |
| | 650 | .p12 = true, |
| | 651 | .p13 = true, |
| | 652 | .p14 = true, |
| | 653 | .p15 = true, |
| | 654 | .fpcr = true, |
| | 655 | .fpsr = true, |
| | 656 | .ffr = true, |
| | 657 | .memory = true, |
| | 658 | }), |
| | 659 | .x86_64 => asm volatile ( |
| | 660 | \\ movq 0(%%rsi), %%rax |
| | 661 | \\ movq 8(%%rsi), %%rcx |
| | 662 | \\ leaq 0f(%%rip), %%rdx |
| | 663 | \\ movq %%rsp, 0(%%rax) |
| | 664 | \\ movq %%rbp, 8(%%rax) |
| | 665 | \\ movq %%rdx, 16(%%rax) |
| | 666 | \\ movq 0(%%rcx), %%rsp |
| | 667 | \\ movq 8(%%rcx), %%rbp |
| | 668 | \\ jmpq *16(%%rcx) |
| | 669 | \\0: |
| | 670 | : [received_message] "={rsi}" (-> *const @FieldType(SwitchMessage, "contexts")), |
| | 671 | : [message_to_send] "{rsi}" (&message.contexts), |
| | 672 | : .{ |
| | 673 | .rax = true, |
| | 674 | .rcx = true, |
| | 675 | .rdx = true, |
| | 676 | .rbx = true, |
| | 677 | .rsi = true, |
| | 678 | .rdi = true, |
| | 679 | .r8 = true, |
| | 680 | .r9 = true, |
| | 681 | .r10 = true, |
| | 682 | .r11 = true, |
| | 683 | .r12 = true, |
| | 684 | .r13 = true, |
| | 685 | .r14 = true, |
| | 686 | .r15 = true, |
| | 687 | .mm0 = true, |
| | 688 | .mm1 = true, |
| | 689 | .mm2 = true, |
| | 690 | .mm3 = true, |
| | 691 | .mm4 = true, |
| | 692 | .mm5 = true, |
| | 693 | .mm6 = true, |
| | 694 | .mm7 = true, |
| | 695 | .zmm0 = true, |
| | 696 | .zmm1 = true, |
| | 697 | .zmm2 = true, |
| | 698 | .zmm3 = true, |
| | 699 | .zmm4 = true, |
| | 700 | .zmm5 = true, |
| | 701 | .zmm6 = true, |
| | 702 | .zmm7 = true, |
| | 703 | .zmm8 = true, |
| | 704 | .zmm9 = true, |
| | 705 | .zmm10 = true, |
| | 706 | .zmm11 = true, |
| | 707 | .zmm12 = true, |
| | 708 | .zmm13 = true, |
| | 709 | .zmm14 = true, |
| | 710 | .zmm15 = true, |
| | 711 | .zmm16 = true, |
| | 712 | .zmm17 = true, |
| | 713 | .zmm18 = true, |
| | 714 | .zmm19 = true, |
| | 715 | .zmm20 = true, |
| | 716 | .zmm21 = true, |
| | 717 | .zmm22 = true, |
| | 718 | .zmm23 = true, |
| | 719 | .zmm24 = true, |
| | 720 | .zmm25 = true, |
| | 721 | .zmm26 = true, |
| | 722 | .zmm27 = true, |
| | 723 | .zmm28 = true, |
| | 724 | .zmm29 = true, |
| | 725 | .zmm30 = true, |
| | 726 | .zmm31 = true, |
| | 727 | .fpsr = true, |
| | 728 | .fpcr = true, |
| | 729 | .mxcsr = true, |
| | 730 | .rflags = true, |
| | 731 | .dirflag = true, |
| | 732 | .memory = true, |
| | 733 | }), |
| | 734 | else => |arch| @compileError("unimplemented architecture: " ++ @tagName(arch)), |
| | 735 | }); |
| | 736 | } |
| | 737 | |
| | 738 | fn mainIdleEntry() callconv(.naked) void { |
| | 739 | switch (builtin.cpu.arch) { |
| | 740 | .x86_64 => asm volatile ( |
| | 741 | \\ movq (%%rsp), %%rdi |
| | 742 | \\ jmp %[mainIdle:P] |
| | 743 | : |
| | 744 | : [mainIdle] "X" (&mainIdle), |
| | 745 | ), |
| | 746 | .aarch64 => asm volatile ( |
| | 747 | \\ ldr x0, [sp, #-8] |
| | 748 | \\ b %[mainIdle] |
| | 749 | : |
| | 750 | : [mainIdle] "X" (&mainIdle), |
| | 751 | ), |
| | 752 | else => |arch| @compileError("unimplemented architecture: " ++ @tagName(arch)), |
| | 753 | } |
| | 754 | } |
| | 755 | |
| | 756 | fn fiberEntry() callconv(.naked) void { |
| | 757 | switch (builtin.cpu.arch) { |
| | 758 | .x86_64 => asm volatile ( |
| | 759 | \\ leaq 8(%%rsp), %%rdi |
| | 760 | \\ jmp %[AsyncClosure_call:P] |
| | 761 | : |
| | 762 | : [AsyncClosure_call] "X" (&AsyncClosure.call), |
| | 763 | ), |
| | 764 | else => |arch| @compileError("unimplemented architecture: " ++ @tagName(arch)), |
| | 765 | } |
| | 766 | } |
| | 767 | |
| | 768 | const AsyncClosure = struct { |
| | 769 | event_loop: *Kqueue, |
| | 770 | fiber: *Fiber, |
| | 771 | start: *const fn (context: *const anyopaque, result: *anyopaque) void, |
| | 772 | result_align: Alignment, |
| | 773 | already_awaited: bool, |
| | 774 | |
| | 775 | fn contextPointer(closure: *AsyncClosure) [*]align(Fiber.max_context_align.toByteUnits()) u8 { |
| | 776 | return @alignCast(@as([*]u8, @ptrCast(closure)) + @sizeOf(AsyncClosure)); |
| | 777 | } |
| | 778 | |
| | 779 | fn call(closure: *AsyncClosure, message: *const SwitchMessage) callconv(.withStackAlign(.c, @alignOf(AsyncClosure))) noreturn { |
| | 780 | message.handle(closure.event_loop); |
| | 781 | const fiber = closure.fiber; |
| | 782 | std.log.debug("{*} performing async", .{fiber}); |
| | 783 | closure.start(closure.contextPointer(), fiber.resultBytes(closure.result_align)); |
| | 784 | const awaiter = @atomicRmw(?*Fiber, &fiber.awaiter, .Xchg, Fiber.finished, .acq_rel); |
| | 785 | const ready_awaiter = r: { |
| | 786 | const a = awaiter orelse break :r null; |
| | 787 | if (@atomicRmw(bool, &closure.already_awaited, .Xchg, true, .acq_rel)) break :r null; |
| | 788 | break :r a; |
| | 789 | }; |
| | 790 | closure.event_loop.yield(ready_awaiter, .nothing); |
| | 791 | unreachable; // switched to dead fiber |
| | 792 | } |
| | 793 | |
| | 794 | fn fromFiber(fiber: *Fiber) *AsyncClosure { |
| | 795 | return @ptrFromInt(Fiber.max_context_align.max(.of(AsyncClosure)).backward( |
| | 796 | @intFromPtr(fiber.allocatedEnd()) - Fiber.max_context_size, |
| | 797 | ) - @sizeOf(AsyncClosure)); |
| | 798 | } |
| | 799 | }; |
| | 800 | |
| 29 | pub fn io(k: *Kqueue) Io { | 801 | pub fn io(k: *Kqueue) Io { |
| 30 | return .{ | 802 | return .{ |
| 31 | .userdata = k, | 803 | .userdata = k, |
| ... | @@ -229,11 +1001,33 @@ fn mutexUnlock(userdata: ?*anyopaque, prev_state: Io.Mutex.State, mutex: *Io.Mut | ... | @@ -229,11 +1001,33 @@ fn mutexUnlock(userdata: ?*anyopaque, prev_state: Io.Mutex.State, mutex: *Io.Mut |
| 229 | | 1001 | |
| 230 | fn conditionWait(userdata: ?*anyopaque, cond: *Io.Condition, mutex: *Io.Mutex) Io.Cancelable!void { | 1002 | fn conditionWait(userdata: ?*anyopaque, cond: *Io.Condition, mutex: *Io.Mutex) Io.Cancelable!void { |
| 231 | const k: *Kqueue = @ptrCast(@alignCast(userdata)); | 1003 | const k: *Kqueue = @ptrCast(@alignCast(userdata)); |
| 232 | _ = k; | 1004 | k.yield(null, .{ .condition_wait = .{ .cond = cond, .mutex = mutex } }); |
| 233 | _ = cond; | 1005 | const thread = Thread.current(); |
| 234 | _ = mutex; | 1006 | const fiber = thread.currentFiber(); |
| 235 | @panic("TODO"); | 1007 | const cond_impl = fiber.resultPointer(Condition); |
| | 1008 | try mutex.lock(k.io()); |
| | 1009 | switch (cond_impl.event) { |
| | 1010 | .queued => {}, |
| | 1011 | .wake => |wake| if (fiber.queue_next) |next_fiber| switch (wake) { |
| | 1012 | .one => if (@cmpxchgStrong( |
| | 1013 | ?*Fiber, |
| | 1014 | @as(*?*Fiber, @ptrCast(&cond.state)), |
| | 1015 | null, |
| | 1016 | next_fiber, |
| | 1017 | .release, |
| | 1018 | .acquire, |
| | 1019 | )) |old_fiber| { |
| | 1020 | const old_cond_impl = old_fiber.?.resultPointer(Condition); |
| | 1021 | assert(old_cond_impl.tail.queue_next == null); |
| | 1022 | old_cond_impl.tail.queue_next = next_fiber; |
| | 1023 | old_cond_impl.tail = cond_impl.tail; |
| | 1024 | }, |
| | 1025 | .all => k.schedule(thread, .{ .head = next_fiber, .tail = cond_impl.tail }), |
| | 1026 | }, |
| | 1027 | } |
| | 1028 | fiber.queue_next = null; |
| 236 | } | 1029 | } |
| | 1030 | |
| 237 | fn conditionWaitUncancelable(userdata: ?*anyopaque, cond: *Io.Condition, mutex: *Io.Mutex) void { | 1031 | fn conditionWaitUncancelable(userdata: ?*anyopaque, cond: *Io.Condition, mutex: *Io.Mutex) void { |
| 238 | const k: *Kqueue = @ptrCast(@alignCast(userdata)); | 1032 | const k: *Kqueue = @ptrCast(@alignCast(userdata)); |
| 239 | _ = k; | 1033 | _ = k; |
| ... | @@ -243,10 +1037,9 @@ fn conditionWaitUncancelable(userdata: ?*anyopaque, cond: *Io.Condition, mutex: | ... | @@ -243,10 +1037,9 @@ fn conditionWaitUncancelable(userdata: ?*anyopaque, cond: *Io.Condition, mutex: |
| 243 | } | 1037 | } |
| 244 | fn conditionWake(userdata: ?*anyopaque, cond: *Io.Condition, wake: Io.Condition.Wake) void { | 1038 | fn conditionWake(userdata: ?*anyopaque, cond: *Io.Condition, wake: Io.Condition.Wake) void { |
| 245 | const k: *Kqueue = @ptrCast(@alignCast(userdata)); | 1039 | const k: *Kqueue = @ptrCast(@alignCast(userdata)); |
| 246 | _ = k; | 1040 | const waiting_fiber = @atomicRmw(?*Fiber, @as(*?*Fiber, @ptrCast(&cond.state)), .Xchg, null, .acquire) orelse return; |
| 247 | _ = cond; | 1041 | waiting_fiber.resultPointer(Condition).event = .{ .wake = wake }; |
| 248 | _ = wake; | 1042 | k.yield(waiting_fiber, .reschedule); |
| 249 | @panic("TODO"); | | |
| 250 | } | 1043 | } |
| 251 | | 1044 | |
| 252 | fn dirMake(userdata: ?*anyopaque, dir: Dir, sub_path: []const u8, mode: Dir.Mode) Dir.MakeError!void { | 1045 | fn dirMake(userdata: ?*anyopaque, dir: Dir, sub_path: []const u8, mode: Dir.Mode) Dir.MakeError!void { |
| ... | @@ -426,7 +1219,7 @@ fn netBindIp( | ... | @@ -426,7 +1219,7 @@ fn netBindIp( |
| 426 | const k: *Kqueue = @ptrCast(@alignCast(userdata)); | 1219 | const k: *Kqueue = @ptrCast(@alignCast(userdata)); |
| 427 | const family = Io.Threaded.posixAddressFamily(address); | 1220 | const family = Io.Threaded.posixAddressFamily(address); |
| 428 | const socket_fd = try openSocketPosix(k, family, options); | 1221 | const socket_fd = try openSocketPosix(k, family, options); |
| 429 | errdefer posix.close(socket_fd); | 1222 | errdefer std.posix.close(socket_fd); |
| 430 | var storage: Io.Threaded.PosixAddress = undefined; | 1223 | var storage: Io.Threaded.PosixAddress = undefined; |
| 431 | var addr_len = Io.Threaded.addressToPosix(address, &storage); | 1224 | var addr_len = Io.Threaded.addressToPosix(address, &storage); |
| 432 | try posixBind(k, socket_fd, &storage.any, addr_len); | 1225 | try posixBind(k, socket_fd, &storage.any, addr_len); |
| ... | @@ -704,3 +1497,11 @@ fn setSocketOption(k: *Kqueue, fd: posix.fd_t, level: i32, opt_name: u32, option | ... | @@ -704,3 +1497,11 @@ fn setSocketOption(k: *Kqueue, fd: posix.fd_t, level: i32, opt_name: u32, option |
| 704 | fn checkCancel(k: *Kqueue) error{Canceled}!void { | 1497 | fn checkCancel(k: *Kqueue) error{Canceled}!void { |
| 705 | if (cancelRequested(k)) return error.Canceled; | 1498 | if (cancelRequested(k)) return error.Canceled; |
| 706 | } | 1499 | } |
| | 1500 | |
| | 1501 | const Condition = struct { |
| | 1502 | tail: *Fiber, |
| | 1503 | event: union(enum) { |
| | 1504 | queued, |
| | 1505 | wake: Io.Condition.Wake, |
| | 1506 | }, |
| | 1507 | }; |
Andrew Kelley <andrew@ziglang.org> 2025-10-23 03:21:32-07:00
lib