Base: Fix another deadlock during shutdown of SequencedWorkerPool.

This CL is very similar to https://codereview.chromium.org/956583004/. There is
yet another code path that causes SequencedWorkerPool to destroy some tasks
without releasing its global lock.

BUG=465271

Review URL: https://codereview.chromium.org/994483005

Cr-Commit-Position: refs/heads/master@{#320600}

2 files changed, 45 insertions, 0 deletions

diff --git a/base/threading/sequenced_worker_pool.cc b/base/threading/sequenced_worker_pool.cc
index f20d997..52b178b 100644
--- a/base/threading/sequenced_worker_pool.cc
+++ b/base/threading/sequenced_worker_pool.cc
@@ -803,6 +803,20 @@ void SequencedWorkerPool::Inner::ThreadLoop(Worker* this_worker) {
           delete_these_outside_lock.clear();
           break;
         }
+
+        // No work was found, but there are tasks that need deletion. The
+        // deletion must happen outside of the lock.
+        if (delete_these_outside_lock.size()) {
+          AutoUnlock unlock(lock_);
+          delete_these_outside_lock.clear();
+
+          // Since the lock has been released, |status| may no longer be
+          // accurate. It might read GET_WORK_WAIT even if there are tasks
+          // ready to perform work. Jump to the top of the loop to recalculate
+          // |status|.
+          continue;
+        }
+
         waiting_thread_count_++;
 
         switch (status) {
diff --git a/base/threading/sequenced_worker_pool_unittest.cc b/base/threading/sequenced_worker_pool_unittest.cc
index c132731..5d0880c 100644
--- a/base/threading/sequenced_worker_pool_unittest.cc
+++ b/base/threading/sequenced_worker_pool_unittest.cc
@@ -148,6 +148,17 @@ class TestTracker : public base::RefCountedThreadSafe<TestTracker> {
         FROM_HERE, reposting_task, SequencedWorkerPool::SKIP_ON_SHUTDOWN);
   }
 
+  // This task reposts itself back onto the SequencedWorkerPool before it
+  // finishes running.
+  void PostRepostingBlockingTask(
+      const scoped_refptr<SequencedWorkerPool>& pool,
+      const SequencedWorkerPool::SequenceToken& token) {
+    Closure reposting_task =
+        base::Bind(&TestTracker::PostRepostingBlockingTask, this, pool, token);
+    pool->PostSequencedWorkerTaskWithShutdownBehavior(token,
+        FROM_HERE, reposting_task, SequencedWorkerPool::BLOCK_SHUTDOWN);
+  }
+
   // Waits until the given number of tasks have started executing.
   void WaitUntilTasksBlocked(size_t count) {
     {
@@ -795,6 +806,26 @@ TEST_F(SequencedWorkerPoolTest, AvoidsDeadlockOnShutdown) {
   pool()->Shutdown();
 }
 
+// Similar to the test AvoidsDeadlockOnShutdown, but there are now also
+// sequenced, blocking tasks in the queue during shutdown.
+TEST_F(SequencedWorkerPoolTest,
+       AvoidsDeadlockOnShutdownWithSequencedBlockingTasks) {
+  const std::string sequence_token_name("name");
+  for (int i = 0; i < 4; ++i) {
+    scoped_refptr<DestructionDeadlockChecker> checker(
+        new DestructionDeadlockChecker(pool()));
+    tracker()->PostRepostingTask(pool(), checker);
+
+    SequencedWorkerPool::SequenceToken token1 =
+        pool()->GetNamedSequenceToken(sequence_token_name);
+    tracker()->PostRepostingBlockingTask(pool(), token1);
+  }
+
+  // Shutting down the pool should destroy the DestructionDeadlockCheckers,
+  // which in turn should not deadlock in their destructors.
+  pool()->Shutdown();
+}
+
 // Verify that FlushForTesting works as intended.
 TEST_F(SequencedWorkerPoolTest, FlushForTesting) {
   // Should be fine to call on a new instance.