Fix leak of sequenced worker pool in unit tests when service_ is replaced.

net/ssl/server_bound_cert_service_unittest.cc

Index: net/ssl/server_bound_cert_service_unittest.cc
diff --git a/net/ssl/server_bound_cert_service_unittest.cc b/net/ssl/server_bound_cert_service_unittest.cc
index 0063f496036854c4c3a1ae2be4cbc1dea394ae66..f1d45d03e0d082abba4b9c89f6e2523b7c09cae2 100644
--- a/net/ssl/server_bound_cert_service_unittest.cc
+++ b/net/ssl/server_bound_cert_service_unittest.cc
@@ -536,7 +536,16 @@ TEST_F(ServerBoundCertServiceTest, AsyncStoreGetNoCertsInStore) {
   EXPECT_FALSE(private_key_info.empty());
   EXPECT_FALSE(der_cert.empty());
   EXPECT_FALSE(request_handle.is_active());
-}
+
+  // Wait for any posted ServerBoundCertServiceWorkers to finish, in order to
+  // ensure that the workers are cleared by the time service_'s destructor is
+  // called. This prevents a leak of the SequencedWorkerPool, which by default
+  // continues on shutdown.  (By calling
+  // CallGetServerBoundCertCallbackWithResult, callbacks are called directly,
+  // short-circuiting ServerBoundCertService's usual bookkeeping, so existing
+  // tasks must be both cleared and waited for to ensure proper memory freeing.)
+  sequenced_worker_pool_->FlushForTesting();

[Ryan Sleevi, 2013/08/09 00:45:10]

I can't help but keep reading this comment and code and not understanding what's
expected/correct.

It seems like your Mock is broken/violating expectations. Should you be
PostTasking from CallGetServerBoundCertCallbackWithResult to match the semantics
of the ServerBoundCertStore, so that you then use FlushForTesting &
RunUntilIdle?

This definitely seems to match the interface you're mocking, since you're
returning ERR_IO_PENDING (which we always handle via PostTask yielding)

[juanlang, 2013/08/09 04:48:52]

Yeah, that big comment should have been a clue to me that I still hadn't
understood entirely. These tests are similar to the other tests. I repeat the
pattern of the other tests here.
I thought so too, but upon further reading I believe my Mock is behaving
substantially like the real store.

[juanlang, 2013/08/09 16:25:28]

Yep, PostTask is the right thing here. Thanks. I don't need to FlushForTesting
and RunUntilIdle when I use it, either, because by PostTasking and then calling
callback.WaitForResult, the workers correctly terminate already.

In other words, done.

+  base::MessageLoop::current()->RunUntilIdle(); }
 
 TEST_F(ServerBoundCertServiceTest, AsyncStoreGetOneCertInStore) {
   MockServerBoundCertStoreWithAsyncGet* mock_store =
@@ -560,7 +569,6 @@ TEST_F(ServerBoundCertServiceTest, AsyncStoreGetOneCertInStore) {
 
   mock_store->CallGetServerBoundCertCallbackWithResult(
       OK, base::Time(), "ab", "cd");
-
   error = callback.WaitForResult();
   EXPECT_EQ(OK, error);
   EXPECT_EQ(1, service_->cert_count());
@@ -572,6 +580,16 @@ TEST_F(ServerBoundCertServiceTest, AsyncStoreGetOneCertInStore) {
   EXPECT_STREQ("ab", private_key_info.c_str());
   EXPECT_STREQ("cd", der_cert.c_str());
   EXPECT_FALSE(request_handle.is_active());
+
+  // Wait for any posted ServerBoundCertServiceWorkers to finish, in order to
+  // ensure that the workers are cleared by the time service_'s destructor is
+  // called. This prevents a leak of the SequencedWorkerPool, which by default
+  // continues on shutdown.  (By calling
+  // CallGetServerBoundCertCallbackWithResult, callbacks are called directly,
+  // short-circuiting ServerBoundCertService's usual bookkeeping, so existing
+  // tasks must be both cleared and waited for to ensure proper memory freeing.)
+  sequenced_worker_pool_->FlushForTesting();
+  base::MessageLoop::current()->RunUntilIdle(); }
 }
 
 #endif  // !defined(USE_OPENSSL)