Identify the hung thread using the Wait Chain Traversal API

chrome/chrome_watcher/wait_chain_util_win.cc

Index: chrome/chrome_watcher/wait_chain_util_win.cc
diff --git a/chrome/chrome_watcher/wait_chain_util_win.cc b/chrome/chrome_watcher/wait_chain_util_win.cc
new file mode 100644
index 0000000000000000000000000000000000000000..22ef56f138ac20da534d27f1529907297ceb2f27
--- /dev/null
+++ b/chrome/chrome_watcher/wait_chain_util_win.cc
@@ -0,0 +1,115 @@
+// Copyright 2016 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#include "chrome/chrome_watcher/wait_chain_util_win.h"
+
+#include <memory>
+
+#include "base/logging.h"
+#include "base/strings/stringprintf.h"
+
+namespace {
+
+// Helper deleter to hold a HWCT into a unique_ptr.
+struct WaitChainSessionDeleter {
+  using pointer = HWCT;
+  void operator()(HWCT session_handle) {
+    ::CloseThreadWaitChainSession(session_handle);
+  }
+};
+
+const wchar_t* WctObjectTypeToString(WCT_OBJECT_TYPE type) {
+  switch (type) {
+    case WctCriticalSectionType:
+      return L"CriticalSection";
+    case WctSendMessageType:
+      return L"SendMessage";
+    case WctMutexType:
+      return L"Mutex";
+    case WctAlpcType:
+      return L"Alpc";
+    case WctComType:
+      return L"Com";
+    case WctThreadWaitType:
+      return L"ThreadWait";
+    case WctProcessWaitType:
+      return L"ProcessWait";
+    case WctThreadType:
+      return L"Thread";
+    case WctComActivationType:
+      return L"ComActivation";
+    case WctUnknownType:
+      return L"Unknown";
+    default:
+      NOTREACHED();
+      return L"";
+  }
+}
+
+const wchar_t* WctObjectStatusToString(WCT_OBJECT_STATUS status) {
+  switch (status) {
+    case WctStatusNoAccess:
+      return L"NoAccess";
+    case WctStatusRunning:
+      return L"Running";
+    case WctStatusBlocked:
+      return L"Blocked";
+    case WctStatusPidOnly:
+      return L"PidOnly";
+    case WctStatusPidOnlyRpcss:
+      return L"PidOnlyRpcss";
+    case WctStatusOwned:
+      return L"Owned";
+    case WctStatusNotOwned:
+      return L"NotOwned";
+    case WctStatusAbandoned:
+      return L"Abandoned";
+    case WctStatusUnknown:
+      return L"Unknown";
+    case WctStatusError:
+      return L"Error";
+    default:

[Sigurður Ásgeirsson, 2016/03/24 19:30:18]

I seem to remember there's a way to get compile-time failures if ever the enum
changes (it's clearly intended to be extensible). Offhand I don't remember how
that works, but one way is to compile_assert on the value of the kMax value?

[Patrick Monette, 2016/03/24 23:21:03]

I'm not sure I understand your suggestion. The enum doesn't have a kMax value.

+      NOTREACHED();
+      return L"";
+  }
+}
+
+}  // namespace
+
+bool GetThreadWaitChain(DWORD thread_id,
+                        std::vector<WAITCHAIN_NODE_INFO>* wait_chain,
+                        bool* is_deadlock) {
+  DCHECK(wait_chain);
+  DCHECK(is_deadlock);
+
+  std::unique_ptr<HWCT, WaitChainSessionDeleter> session_handle(

[Sigurður Ásgeirsson, 2016/03/24 19:30:18]

nit: more readable to typedef a ScopedWaitChainSessionHandle type or something
above and use here?

[manzagop (departed), 2016/03/24 19:36:36]

Nit: comment to mention it's a synchronous session?

[Patrick Monette, 2016/03/24 23:21:03]

Done.

[Patrick Monette, 2016/03/24 23:21:03]

I like it. Done.

+      ::OpenThreadWaitChainSession(0, nullptr));
+  if (!session_handle) {
+    return false;

[manzagop (departed), 2016/03/24 19:36:36]

Should you log the error message?

[Patrick Monette, 2016/03/24 23:21:03]

Done.

[Sigurður Ásgeirsson, 2016/03/29 16:57:32]

Not to be too contrary, but verbose logging is generally discouraged in Chrome.
As most of the time noone is listening, debug logging is more appropriate than
to bloat release code with useless log statements (is the rationale).

+  }
+
+  DWORD nb_nodes = WCT_MAX_NODE_COUNT;
+  WAITCHAIN_NODE_INFO nodes[WCT_MAX_NODE_COUNT];

[manzagop (departed), 2016/03/24 19:36:36]

I guess you could put it directly in the vector?

[Patrick Monette, 2016/03/24 23:21:03]

Done.

+  BOOL is_cycle;
+  if (!::GetThreadWaitChain(session_handle.get(), NULL, WCTP_GETINFO_ALL_FLAGS,

[manzagop (departed), 2016/03/24 19:36:36]

We should double check this is the right flag.

[Patrick Monette, 2016/03/24 23:21:03]

I removed it. They aren't doing anything.

+                            thread_id, &nb_nodes, nodes, &is_cycle))
+    return false;

[manzagop (departed), 2016/03/24 19:36:36]

Again, log an error message?
I'm guessing we shouldn't hit the ERROR_MORE_DATA?

[Patrick Monette, 2016/03/24 23:21:03]

Probably not but logging the error message is a good idea.

+
+  *is_deadlock = !!is_cycle;

[Sigurður Ásgeirsson, 2016/03/24 19:30:18]

we use the !! notion for coercing to bool?

[Patrick Monette, 2016/03/24 23:21:03]

Hem.. yeah I don't think so. What should I then?

is_deadlock = is_cycle ? true : false; ?

[Sigurður Ásgeirsson, 2016/03/29 16:57:32]

I dunno.

+  wait_chain->insert(wait_chain->end(), &nodes[0], &nodes[nb_nodes]);
+
+  return true;
+}
+
+std::wstring WaitChainNodeToString(const WAITCHAIN_NODE_INFO& node) {
+  if (node.ObjectType == WctThreadType) {
+    return base::StringPrintf(L"Thread #%d with status %ls",
+                              node.ThreadObject.ThreadId,
+                              WctObjectStatusToString(node.ObjectStatus));
+  } else {
+    return base::StringPrintf(L"Lock of type %ls with status %ls",
+                              WctObjectTypeToString(node.ObjectType),
+                              WctObjectStatusToString(node.ObjectStatus));
+  }
+}