[Jobs API] Lockless CancelAndDetach.

Wasm wants to uses a CancelAndDetach + shared_ptr pattern that can cause
CancelAndDetach to be called within GetMaxConcurrency(). This causes
recursive lock errors. CancelAndDetach doesn't need a lock since state_
is atomic; making it lockless allows this use case.

The current wasm alternative is to hold a vector of JobHandle until the
renderer is teared down; this causes a growing 'live' leak.

Change-Id: I50e220fc87136739bd3641a2a6d5fad1fe376f9a
Reviewed-on: https://chromium-review.googlesource.com/c/chromium/src/+/2508335
Commit-Queue: Etienne Pierre-Doray <etiennep@chromium.org>
Reviewed-by: Gabriel Charette <gab@chromium.org>
Cr-Commit-Position: refs/heads/master@{#822870}

base/task/thread_pool/job_task_source.cc

@@ -149,11 +149,10 @@ bool JobTaskSource::RunJoinTask() {
 }
 
 void JobTaskSource::Cancel(TaskSource::Transaction* transaction) {
-  CheckedAutoLock auto_lock(worker_lock_);
   // Sets the kCanceledMask bit on |state_| so that further calls to
-  // WillRunTask() never succeed. std::memory_order_relaxed is sufficient
-  // because this task source never needs to be re-enqueued after Cancel().
-  state_.Cancel();
+  // WillRunTask() never succeed. std::memory_order_relaxed without a lock is
+  // safe because this task source never needs to be re-enqueued after Cancel().
+  TS_UNCHECKED_READ(state_).Cancel();
 }
 
 // EXCLUSIVE_LOCK_REQUIRED(worker_lock_)

base/task/thread_pool/job_task_source.h

@@ -192,7 +192,7 @@ class BASE_EXPORT JobTaskSource : public TaskSource {
   mutable CheckedLock worker_lock_{UniversalSuccessor()};
 
   // Current atomic state (atomic despite the lock to allow optimistic reads
-  // without the lock).
+  // and cancellation without the lock).
   State state_ GUARDED_BY(worker_lock_);
   // Normally, |join_flag_| is protected by |lock_|, except in ShouldYield()
   // hence the use of atomics.