| ... | @@ -194,42 +194,27 @@ const FutexImpl = struct { | ... | @@ -194,42 +194,27 @@ const FutexImpl = struct { |
| 194 | const signal_mask = 0xffff << 16; | 194 | const signal_mask = 0xffff << 16; |
| 195 | | 195 | |
| 196 | fn wait(self: *Impl, mutex: *Mutex, timeout: ?u64) error{Timeout}!void { | 196 | fn wait(self: *Impl, mutex: *Mutex, timeout: ?u64) error{Timeout}!void { |
| 197 | // Register that we're waiting on the state by incrementing the wait count. | 197 | // Observe the epoch, then check the state again to see if we should wake up. |
| 198 | // This assumes that there can be at most ((1<<16)-1) or 65,355 threads concurrently waiting on the same Condvar. | 198 | // The epoch must be observed before we check the state or we could potentially miss a wake() and deadlock: |
| 199 | // If this is hit in practice, then this condvar not working is the least of your concerns. | 199 | // |
| | 200 | // - T1: s = LOAD(&state) |
| | 201 | // - T2: UPDATE(&s, signal) |
| | 202 | // - T2: UPDATE(&epoch, 1) + FUTEX_WAKE(&epoch) |
| | 203 | // - T1: e = LOAD(&epoch) (was reordered after the state load) |
| | 204 | // - T1: s & signals == 0 -> FUTEX_WAIT(&epoch, e) (missed the state update + the epoch change) |
| | 205 | // |
| | 206 | // Acquire barrier to ensure the epoch load happens before the state load. |
| | 207 | var epoch = self.epoch.load(.Acquire); |
| 200 | var state = self.state.fetchAdd(one_waiter, .Monotonic); | 208 | var state = self.state.fetchAdd(one_waiter, .Monotonic); |
| 201 | assert(state & waiter_mask != waiter_mask); | 209 | assert(state & waiter_mask != waiter_mask); |
| 202 | state += one_waiter; | 210 | state += one_waiter; |
| 203 | | 211 | |
| 204 | // Temporarily release the mutex in order to block on the condition variable. | | |
| 205 | mutex.unlock(); | 212 | mutex.unlock(); |
| 206 | defer mutex.lock(); | 213 | defer mutex.lock(); |
| 207 | | 214 | |
| 208 | var futex_deadline = Futex.Deadline.init(timeout); | 215 | var futex_deadline = Futex.Deadline.init(timeout); |
| 209 | while (true) { | | |
| 210 | // Try to wake up by consuming a signal and decremented the waiter we added previously. | | |
| 211 | // Acquire barrier ensures code before the wake() which added the signal happens before we decrement it and return. | | |
| 212 | while (state & signal_mask != 0) { | | |
| 213 | const new_state = state - one_waiter - one_signal; | | |
| 214 | state = self.state.tryCompareAndSwap(state, new_state, .Acquire, .Monotonic) orelse return; | | |
| 215 | } | | |
| 216 | | | |
| 217 | // Observe the epoch, then check the state again to see if we should wake up. | | |
| 218 | // The epoch must be observed before we check the state or we could potentially miss a wake() and deadlock: | | |
| 219 | // | | |
| 220 | // - T1: s = LOAD(&state) | | |
| 221 | // - T2: UPDATE(&s, signal) | | |
| 222 | // - T2: UPDATE(&epoch, 1) + FUTEX_WAKE(&epoch) | | |
| 223 | // - T1: e = LOAD(&epoch) (was reordered after the state load) | | |
| 224 | // - T1: s & signals == 0 -> FUTEX_WAIT(&epoch, e) (missed the state update + the epoch change) | | |
| 225 | // | | |
| 226 | // Acquire barrier to ensure the epoch load happens before the state load. | | |
| 227 | const epoch = self.epoch.load(.Acquire); | | |
| 228 | state = self.state.load(.Monotonic); | | |
| 229 | if (state & signal_mask != 0) { | | |
| 230 | continue; | | |
| 231 | } | | |
| 232 | | 216 | |
| | 217 | while (true) { |
| 233 | futex_deadline.wait(&self.epoch, epoch) catch |err| switch (err) { | 218 | futex_deadline.wait(&self.epoch, epoch) catch |err| switch (err) { |
| 234 | // On timeout, we must decrement the waiter we added above. | 219 | // On timeout, we must decrement the waiter we added above. |
| 235 | error.Timeout => { | 220 | error.Timeout => { |
| ... | @@ -247,6 +232,16 @@ const FutexImpl = struct { | ... | @@ -247,6 +232,16 @@ const FutexImpl = struct { |
| 247 | } | 232 | } |
| 248 | }, | 233 | }, |
| 249 | }; | 234 | }; |
| | 235 | |
| | 236 | epoch = self.epoch.load(.Acquire); |
| | 237 | state = self.state.load(.Monotonic); |
| | 238 | |
| | 239 | // Try to wake up by consuming a signal and decremented the waiter we added previously. |
| | 240 | // Acquire barrier ensures code before the wake() which added the signal happens before we decrement it and return. |
| | 241 | while (state & signal_mask != 0) { |
| | 242 | const new_state = state - one_waiter - one_signal; |
| | 243 | state = self.state.tryCompareAndSwap(state, new_state, .Acquire, .Monotonic) orelse return; |
| | 244 | } |
| 250 | } | 245 | } |
| 251 | } | 246 | } |
| 252 | | 247 | |
| ... | @@ -536,3 +531,150 @@ test "Condition - broadcasting" { | ... | @@ -536,3 +531,150 @@ test "Condition - broadcasting" { |
| 536 | t.join(); | 531 | t.join(); |
| 537 | } | 532 | } |
| 538 | } | 533 | } |
| | 534 | |
| | 535 | test "Condition - broadcasting - wake all threads" { |
| | 536 | // Tests issue #12877 |
| | 537 | // This test requires spawning threads |
| | 538 | if (builtin.single_threaded) { |
| | 539 | return error.SkipZigTest; |
| | 540 | } |
| | 541 | |
| | 542 | var num_runs: usize = 1; |
| | 543 | const num_threads = 10; |
| | 544 | |
| | 545 | while (num_runs > 0) : (num_runs -= 1) { |
| | 546 | const BroadcastTest = struct { |
| | 547 | mutex: Mutex = .{}, |
| | 548 | cond: Condition = .{}, |
| | 549 | completed: Condition = .{}, |
| | 550 | count: usize = 0, |
| | 551 | thread_id_to_wake: usize = 0, |
| | 552 | threads: [num_threads]std.Thread = undefined, |
| | 553 | wakeups: usize = 0, |
| | 554 | |
| | 555 | fn run(self: *@This(), thread_id: usize) void { |
| | 556 | self.mutex.lock(); |
| | 557 | defer self.mutex.unlock(); |
| | 558 | |
| | 559 | // The last broadcast thread to start tells the main test thread it's completed. |
| | 560 | self.count += 1; |
| | 561 | if (self.count == num_threads) { |
| | 562 | self.completed.signal(); |
| | 563 | } |
| | 564 | |
| | 565 | while (self.thread_id_to_wake != thread_id) { |
| | 566 | self.cond.timedWait(&self.mutex, 1 * std.time.ns_per_s) catch std.debug.panic("thread_id {d} timeout {d}", .{ thread_id, self.thread_id_to_wake }); |
| | 567 | self.wakeups += 1; |
| | 568 | } |
| | 569 | if (self.thread_id_to_wake <= num_threads) { |
| | 570 | // Signal next thread to wake up. |
| | 571 | self.thread_id_to_wake += 1; |
| | 572 | self.cond.broadcast(); |
| | 573 | } |
| | 574 | } |
| | 575 | }; |
| | 576 | |
| | 577 | var broadcast_test = BroadcastTest{}; |
| | 578 | var thread_id: usize = 1; |
| | 579 | for (broadcast_test.threads) |*t| { |
| | 580 | t.* = try std.Thread.spawn(.{}, BroadcastTest.run, .{ &broadcast_test, thread_id }); |
| | 581 | thread_id += 1; |
| | 582 | } |
| | 583 | |
| | 584 | { |
| | 585 | broadcast_test.mutex.lock(); |
| | 586 | defer broadcast_test.mutex.unlock(); |
| | 587 | |
| | 588 | // Wait for all the broadcast threads to spawn. |
| | 589 | // timedWait() to detect any potential deadlocks. |
| | 590 | while (broadcast_test.count != num_threads) { |
| | 591 | try broadcast_test.completed.timedWait( |
| | 592 | &broadcast_test.mutex, |
| | 593 | 1 * std.time.ns_per_s, |
| | 594 | ); |
| | 595 | } |
| | 596 | |
| | 597 | // Signal thread 1 to wake up |
| | 598 | broadcast_test.thread_id_to_wake = 1; |
| | 599 | broadcast_test.cond.broadcast(); |
| | 600 | } |
| | 601 | |
| | 602 | for (broadcast_test.threads) |t| { |
| | 603 | t.join(); |
| | 604 | } |
| | 605 | } |
| | 606 | } |
| | 607 | |
| | 608 | test "Condition - signal wakes one" { |
| | 609 | // This test requires spawning threads |
| | 610 | if (builtin.single_threaded) { |
| | 611 | return error.SkipZigTest; |
| | 612 | } |
| | 613 | |
| | 614 | var num_runs: usize = 1; |
| | 615 | const num_threads = 3; |
| | 616 | const timeoutDelay = 10 * std.time.ns_per_ms; |
| | 617 | |
| | 618 | while (num_runs > 0) : (num_runs -= 1) { |
| | 619 | |
| | 620 | // Start multiple runner threads, wait for them to start and send the signal |
| | 621 | // then. Expect that one thread wake up and all other times out. |
| | 622 | // |
| | 623 | // Test depends on delay in timedWait! If too small all threads can timeout |
| | 624 | // before any one gets wake up. |
| | 625 | |
| | 626 | const Runner = struct { |
| | 627 | mutex: Mutex = .{}, |
| | 628 | cond: Condition = .{}, |
| | 629 | completed: Condition = .{}, |
| | 630 | count: usize = 0, |
| | 631 | threads: [num_threads]std.Thread = undefined, |
| | 632 | wakeups: usize = 0, |
| | 633 | timeouts: usize = 0, |
| | 634 | |
| | 635 | fn run(self: *@This()) void { |
| | 636 | self.mutex.lock(); |
| | 637 | defer self.mutex.unlock(); |
| | 638 | |
| | 639 | // The last started thread tells the main test thread it's completed. |
| | 640 | self.count += 1; |
| | 641 | if (self.count == num_threads) { |
| | 642 | self.completed.signal(); |
| | 643 | } |
| | 644 | |
| | 645 | self.cond.timedWait(&self.mutex, timeoutDelay) catch { |
| | 646 | self.timeouts += 1; |
| | 647 | return; |
| | 648 | }; |
| | 649 | self.wakeups += 1; |
| | 650 | } |
| | 651 | }; |
| | 652 | |
| | 653 | // Start threads |
| | 654 | var runner = Runner{}; |
| | 655 | for (runner.threads) |*t| { |
| | 656 | t.* = try std.Thread.spawn(.{}, Runner.run, .{&runner}); |
| | 657 | } |
| | 658 | |
| | 659 | { |
| | 660 | runner.mutex.lock(); |
| | 661 | defer runner.mutex.unlock(); |
| | 662 | |
| | 663 | // Wait for all the threads to spawn. |
| | 664 | // timedWait() to detect any potential deadlocks. |
| | 665 | while (runner.count != num_threads) { |
| | 666 | try runner.completed.timedWait(&runner.mutex, 1 * std.time.ns_per_s); |
| | 667 | } |
| | 668 | // Signal one thread, the others should get timeout. |
| | 669 | runner.cond.signal(); |
| | 670 | } |
| | 671 | |
| | 672 | for (runner.threads) |t| { |
| | 673 | t.join(); |
| | 674 | } |
| | 675 | |
| | 676 | // Expect that only one got singal |
| | 677 | try std.testing.expectEqual(runner.wakeups, 1); |
| | 678 | try std.testing.expectEqual(runner.timeouts, num_threads - 1); |
| | 679 | } |
| | 680 | } |
Andrew Kelley <andrew@ziglang.org> 2022-11-23 16:24:55-05:00
GitHub <noreply@github.com> 2022-11-23 16:24:55-05:00
Signed by PGP key 4AEE18F83AFDEB23
lib