Handle rare conditions, like message queue exhaustion

base/message_pump_win.cc

 // Copyright (c) 2012 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 "base/message_pump_win.h"
 
 #include <math.h>
 
 #include "base/message_loop.h"
 #include "base/metrics/histogram.h"
 #include "base/win/wrapped_window_proc.h"
 
+namespace {
+
+enum MessageLoopProblems {
+  MESSAGE_QUEUE_FULL,
+  COMPLETION_POST_ERROR,
+  MESSAGE_LOOP_PROBLEM_MAX,
+};
+
+}  // namespace
+
 namespace base {
 
 static const wchar_t kWndClass[] = L"Chrome_MessagePumpWindow";
 
 // Message sent to get an additional time slice for pumping (processing) another
 // task (a series of such messages creates a continuous task pump).
 static const int kMsgHaveWork = WM_USER + 1;
 
 //-----------------------------------------------------------------------------
 // MessagePumpWin public:
@@ skipping 66 matching lines @@
 MessagePumpForUI::~MessagePumpForUI() {
   DestroyWindow(message_hwnd_);
   UnregisterClass(kWndClass, GetModuleHandle(NULL));
 }
 
 void MessagePumpForUI::ScheduleWork() {
   if (InterlockedExchange(&have_work_, 1))
     return;  // Someone else continued the pumping.
 
   // Make sure the MessagePump does some work for us.
-  PostMessage(message_hwnd_, kMsgHaveWork, reinterpret_cast<WPARAM>(this), 0);
+  BOOL ret = PostMessage(message_hwnd_, kMsgHaveWork,
+                         reinterpret_cast<WPARAM>(this), 0);
+  if (ret)
+    return;  // There was room in the Window Message queue.
+
+  // We have failed to insert a have-work message, so there is a chance that we
+  // will starve tasks/timers while sitting in a nested message loop.  Nested
+  // loops only look at Windows Message queues, and don't look at *our* task
+  // queues, etc.  Without the have-work message, we might not get a time slice
+  // in such a nested message loop. :-(
+  // We could abort here, but the fear is that this failure mode is plausibly
+  // common (queue is full, of about 2000 messages), so we'll do a near-graceful
+  // recovery.  We'll put the have_work_ flag back in the state to indicate that
+  // there is no have-work message, and the next time someone posts a task to
+  // our Chromium task queue (or other processing that calls for
+  // ScheduleWork()), we will get back-on-track and be able to avoid starvation
+  // in nested loops.  Nested loops are pretty transient (we think), so this
+  // will probably be workable.

[rvargas (doing something else), 2012/05/19 00:19:11]

I cannot say don't comment the code... but the other extreme is not good either.
For instance everything after "we'll put the have_work_ flag" is actually
redundant with the code... and the comment on that line of code!

[jar (doing other things), 2012/05/19 01:01:20]

I shortened the more redundant section..

On 2012/05/19 00:19:11, rvargas wrote:

+  InterlockedExchange(&have_work_, 0);  // Clarify that we didn't really insert.
+  UMA_HISTOGRAM_ENUMERATION("Chrome.MessageLoopProblem", MESSAGE_QUEUE_FULL,
+                            MESSAGE_LOOP_PROBLEM_MAX);
 }
 
 void MessagePumpForUI::ScheduleDelayedWork(const TimeTicks& delayed_work_time) {
   //
   // We would *like* to provide high resolution timers.  Windows timers using
   // SetTimer() have a 10ms granularity.  We have to use WM_TIMER as a wakeup
   // mechanism because the application can enter modal windows loops where it
   // is not running our MessageLoop; the only way to have our timers fire in
   // these cases is to post messages there.
   //
@@ skipping 12 matching lines @@
   //
   delayed_work_time_ = delayed_work_time;
 
   int delay_msec = GetCurrentDelay();
   DCHECK_GE(delay_msec, 0);
   if (delay_msec < USER_TIMER_MINIMUM)
     delay_msec = USER_TIMER_MINIMUM;
 
   // Create a WM_TIMER event that will wake us up to check for any pending
   // timers (in case we are running within a nested, external sub-pump).
-  SetTimer(message_hwnd_, reinterpret_cast<UINT_PTR>(this), delay_msec, NULL);
+  BOOL ret = SetTimer(message_hwnd_, reinterpret_cast<UINT_PTR>(this),
+                      delay_msec, NULL);
+  CHECK(ret);  // If we can't set timers, we are in big trouble.

[rvargas (doing something else), 2012/05/19 00:19:11]

I'd say do a GetLastError before crashing... and be prepared for an increased
crash rate.

[jar (doing other things), 2012/05/19 01:01:20]

I switched to histogramming the presence of this error, rather than crashing. 
IF we don't see the histogram tallying, we can then try a CHECK() instead.

On 2012/05/19 00:19:11, rvargas wrote:

 }
 
 void MessagePumpForUI::PumpOutPendingPaintMessages() {
   // If we are being called outside of the context of Run, then don't try to do
   // any work.
   if (!state_)
     return;
 
   // Create a mini-message-pump to force immediate processing of only Windows
   // WM_PAINT messages.  Don't provide an infinite loop, but do enough peeking
@@ skipping 270 matching lines @@
 }
 
 void MessagePumpForIO::ScheduleWork() {
   if (InterlockedExchange(&have_work_, 1))
     return;  // Someone else continued the pumping.
 
   // Make sure the MessagePump does some work for us.
   BOOL ret = PostQueuedCompletionStatus(port_, 0,
                                         reinterpret_cast<ULONG_PTR>(this),
                                         reinterpret_cast<OVERLAPPED*>(this));
-  DCHECK(ret);
+  if (ret)
+    return;  // Post worked perfectly.
+
+  // See comment in MessagePumpForUI::ScheduleWork() for this error recovery.
+  InterlockedExchange(&have_work_, 0);  // Clarify that we didn't succeed.
+  UMA_HISTOGRAM_ENUMERATION("Chrome.MessageLoopProblem", COMPLETION_POST_ERROR,
+                            MESSAGE_LOOP_PROBLEM_MAX);
 }
 
 void MessagePumpForIO::ScheduleDelayedWork(const TimeTicks& delayed_work_time) {
   // We know that we can't be blocked right now since this method can only be
   // called on the same thread as Run, so we only need to update our record of
   // how long to sleep when we do sleep.
   delayed_work_time_ = delayed_work_time;
 }
 
 void MessagePumpForIO::RegisterIOHandler(HANDLE file_handle,
@@ skipping 140 matching lines @@
 
 void MessagePumpForIO::WillProcessIOEvent() {
   FOR_EACH_OBSERVER(IOObserver, io_observers_, WillProcessIOEvent());
 }
 
 void MessagePumpForIO::DidProcessIOEvent() {
   FOR_EACH_OBSERVER(IOObserver, io_observers_, DidProcessIOEvent());
 }
 
 }  // namespace base