Version of MessagePumpForUI optimized for GPU process

base/message_loop/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_loop/message_pump_win.h"
 
 #include <math.h>
 #include <stdint.h>
 
 #include <limits>
 
+#include "base/memory/ptr_util.h"
 #include "base/message_loop/message_loop.h"
 #include "base/metrics/histogram.h"
 #include "base/strings/stringprintf.h"
 #include "base/trace_event/trace_event.h"
 #include "base/win/current_module.h"
 #include "base/win/wrapped_window_proc.h"
 
 namespace base {
 
 namespace {
@@ skipping 64 matching lines @@
     : atom_(0) {
   InitMessageWnd();
 }
 
 MessagePumpForUI::~MessagePumpForUI() {
   DestroyWindow(message_hwnd_);
   UnregisterClass(MAKEINTATOM(atom_), CURRENT_MODULE());
 }
 
 void MessagePumpForUI::ScheduleWork() {
-  if (InterlockedExchange(&have_work_, 1))
+  if (InterlockedExchange(&work_state_, HAVE_WORK) != READY)
     return;  // Someone else continued the pumping.
 
   // Make sure the MessagePump does some work for us.
   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., so we might not get a time slice in such. :-(
   // 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.  Nested loops are pretty transient (we think), so this will
   // probably be recoverable.
-  InterlockedExchange(&have_work_, 0);  // Clarify that we didn't really insert.
+
+  // Clarify that we didn't really insert.
+  InterlockedExchange(&work_state_, READY);
   UMA_HISTOGRAM_ENUMERATION("Chrome.MessageLoopProblem", MESSAGE_POST_ERROR,
                             MESSAGE_LOOP_PROBLEM_MAX);
 }
 
 void MessagePumpForUI::ScheduleDelayedWork(const TimeTicks& delayed_work_time) {
   delayed_work_time_ = delayed_work_time;
   RescheduleTimer();
 }
 
 //-----------------------------------------------------------------------------
@@ skipping 122 matching lines @@
 
   DCHECK_NE(WAIT_FAILED, result) << GetLastError();
 }
 
 void MessagePumpForUI::HandleWorkMessage() {
   // If we are being called outside of the context of Run, then don't try to do
   // any work.  This could correspond to a MessageBox call or something of that
   // sort.
   if (!state_) {
     // Since we handled a kMsgHaveWork message, we must still update this flag.
-    InterlockedExchange(&have_work_, 0);
+    InterlockedExchange(&work_state_, READY);
     return;
   }
 
   // Let whatever would have run had we not been putting messages in the queue
   // run now.  This is an attempt to make our dummy message not starve other
   // messages that may be in the Windows message queue.
   ProcessPumpReplacementMessage();
 
   // Now give the delegate a chance to do some work.  He'll let us know if he
   // needs to do more work.
@@ skipping 122 matching lines @@
     have_message = PeekMessage(&msg, NULL, WM_PAINT, WM_PAINT, PM_REMOVE) ||
                    PeekMessage(&msg, NULL, WM_TIMER, WM_TIMER, PM_REMOVE);
   } else {
     have_message = PeekMessage(&msg, NULL, 0, 0, PM_REMOVE) != FALSE;
   }
 
   DCHECK(!have_message || kMsgHaveWork != msg.message ||
          msg.hwnd != message_hwnd_);
 
   // Since we discarded a kMsgHaveWork message, we must update the flag.
-  int old_have_work = InterlockedExchange(&have_work_, 0);
-  DCHECK(old_have_work);
+  int old_work_state_ = InterlockedExchange(&work_state_, READY);
+  DCHECK_EQ(HAVE_WORK, old_work_state_);
 
   // We don't need a special time slice if we didn't have_message to process.
   if (!have_message)
     return false;
 
   // Guarantee we'll get another time slice in the case where we go into native
   // windows code.   This ScheduleWork() may hurt performance a tiny bit when
   // tasks appear very infrequently, but when the event queue is busy, the
   // kMsgHaveWork events get (percentage wise) rarer and rarer.
   ScheduleWork();
   return ProcessMessageHelper(msg);
 }
 
 //-----------------------------------------------------------------------------
+// MessagePumpForGpu public:
+
+MessagePumpForGpu::MessagePumpForGpu() : thread_id_(GetCurrentThreadId()) {
+  // Init the message queue.
+  MSG msg;
+  PeekMessage(&msg, nullptr, 0, 0, PM_NOREMOVE);

[iyengar, 2016/04/11 22:18:21]

Is this needed?

[stanisc, 2016/04/11 23:57:32]

Yes, this is needed to ensure that the message queue is created for the thread.
Otherwise PostThreadMessage would fail. There is a comment above.

+}
+
+MessagePumpForGpu::~MessagePumpForGpu() {}
+
+// static
+void MessagePumpForGpu::InitFactory() {
+  bool init_result = MessageLoop::InitMessagePumpForUIFactory(
+      &MessagePumpForGpu::CreateMessagePumpForGpu);
+  DCHECK(init_result);
+}
+
+// static
+std::unique_ptr<MessagePump> MessagePumpForGpu::CreateMessagePumpForGpu() {
+  return WrapUnique<MessagePump>(new MessagePumpForGpu);
+}
+
+void MessagePumpForGpu::ScheduleWork() {
+  if (InterlockedExchange(&work_state_, HAVE_WORK) != READY)
+    return;  // Someone else continued the pumping.
+
+  // Make sure the MessagePump does some work for us.
+  BOOL ret = PostThreadMessage(thread_id_, kMsgHaveWork, 0, 0);
+  if (ret)
+    return;  // There was room in the Window Message queue.
+
+  // See comment in MessagePumpForUI::ScheduleWork.
+  InterlockedExchange(&work_state_, READY);
+  UMA_HISTOGRAM_ENUMERATION("Chrome.MessageLoopProblem", MESSAGE_POST_ERROR,
+                            MESSAGE_LOOP_PROBLEM_MAX);
+}
+
+void MessagePumpForGpu::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;
+}
+
+//-----------------------------------------------------------------------------
+// MessagePumpForGpu private:
+
+void MessagePumpForGpu::DoRunLoop() {
+  while (!state_->should_quit) {
+    // Indicate that the loop is handling the work.
+    // If there is a race condition between switching to WORKING state here and
+    // the producer thread setting the HAVE_WORK state after exiting the wait,
+    // the event might remain in the signalled state. That might be less than
+    // optimal but wouldn't result in failing to handle the work.
+    InterlockedExchange(&work_state_, WORKING);
+
+    bool more_work_is_plausible = state_->delegate->DoWork();

[iyengar, 2016/04/11 22:18:21]

This function has a lot in common with the MessagePumpForUI::DoRunLoop function.
Perhaps put a TODO here to refactor it out?

[stanisc, 2016/04/11 23:57:32]

I tried an approach with a common base class - see the first iteration. It was
suggested to me to keep this completely separate from MessagePumpForUI for now.
I hope to refactor later.

+    if (state_->should_quit)
+      break;
+
+    more_work_is_plausible |=
+        state_->delegate->DoDelayedWork(&delayed_work_time_);
+    if (state_->should_quit)
+      break;
+
+    if (more_work_is_plausible)
+      continue;
+
+    more_work_is_plausible = state_->delegate->DoIdleWork();
+    if (state_->should_quit)
+      break;
+
+    if (more_work_is_plausible)
+      continue;
+
+    // Switch that working state to READY to indicate that the loop is
+    // waiting for accepting new work if it is still in WORKING state and hasn't
+    // been signalled. Otherwise if it is in HAVE_WORK state skip the wait
+    // and proceed to handing the work.
+    if (InterlockedCompareExchange(&work_state_, READY, WORKING) == HAVE_WORK)
+      continue;  // Skip wait, more work was requested.
+
+    WaitForWork();  // Wait (sleep) until we have work to do again.
+  }
+}
+
+void MessagePumpForGpu::WaitForWork() {
+  // Wait until a message is available, up to the time needed by the timer
+  // manager to fire the next set of timers.
+  int delay;
+
+  // The while loop handles the situation where on Windows 7 and later versions
+  // MsgWaitForMultipleObjectsEx might time out slightly earlier (less than one
+  // ms) than the specified |delay|. In that situation it is more optimal to
+  // just wait again rather than waste a DoRunLoop cycle.
+  while ((delay = GetCurrentDelay()) != 0) {
+    if (delay < 0)  // Negative value means no timers waiting.
+      delay = INFINITE;
+
+    DWORD result = MsgWaitForMultipleObjectsEx(0, NULL, delay, QS_ALLINPUT, 0);
+    if (result == WAIT_OBJECT_0) {
+      // A WM_* message is available.
+      if (ProcessMessages())
+        return;
+    }
+
+    DCHECK_NE(WAIT_FAILED, result) << GetLastError();
+  }
+}
+
+bool MessagePumpForGpu::ProcessMessages() {
+  MSG msg;
+  bool have_work = false;
+  while (PeekMessage(&msg, nullptr, 0, 0, PM_REMOVE)) {
+    if (msg.message == WM_QUIT) {

[iyengar, 2016/04/11 22:18:21]

We should be calling the message filters if any in this function. Perhaps just
call ProcessMessageHelper here after your check for  kMsgHaveWork?

[stanisc, 2016/04/11 23:57:32]

Do you think calling CallMsgFilter would do any good in GPU process?

From CallMsgFilter documentation:
Passes the specified message and hook code to the hook procedures associated
with the WH_SYSMSGFILTER and WH_MSGFILTER hooks. A WH_SYSMSGFILTER or
WH_MSGFILTER hook procedure is an application-defined callback function that
examines and, optionally, modifies messages for a dialog box, message box, menu,
or scroll bar.

[ananta, 2016/04/12 00:00:47]

You never know who could inject into your process. Perhaps a global hook, etc.
In any case the common convention is if you have a message loop you typically do
this. I think calling ProcessMessageHelper here might be better.

[stanisc, 2016/04/12 21:59:34]

OK, I've added a call to CallMsgFilter below. I didn't use ProcessMessageHelper
for two reasons:
1) It isn't shared with this class
2) The implementation is slightly different. msg.hwnd is expected to be nullptr
in this version vs. a specific window in MessagePumpForUI.

+      // Repost the QUIT message so that it will be retrieved by the primary
+      // GetMessage() loop.
+      state_->should_quit = true;
+      PostQuitMessage(static_cast<int>(msg.wParam));
+      return true;
+    }
+
+    if (msg.hwnd == nullptr && msg.message == kMsgHaveWork) {
+      have_work = true;
+    } else {
+      TranslateMessage(&msg);
+      DispatchMessage(&msg);
+    }
+  }
+
+  return have_work;
+}
+
+//-----------------------------------------------------------------------------
 // MessagePumpForIO public:
 
 MessagePumpForIO::MessagePumpForIO() {
   port_.Set(CreateIoCompletionPort(INVALID_HANDLE_VALUE, NULL, NULL, 1));
   DCHECK(port_.IsValid());
 }
 
 MessagePumpForIO::~MessagePumpForIO() {
 }
 
 void MessagePumpForIO::ScheduleWork() {
-  if (InterlockedExchange(&have_work_, 1))
+  if (InterlockedExchange(&work_state_, HAVE_WORK) != READY)
     return;  // Someone else continued the pumping.
 
   // Make sure the MessagePump does some work for us.
   BOOL ret = PostQueuedCompletionStatus(port_.Get(), 0,
                                         reinterpret_cast<ULONG_PTR>(this),
                                         reinterpret_cast<OVERLAPPED*>(this));
   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.
+  InterlockedExchange(&work_state_, READY);  // 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;
 }
@@ skipping 124 matching lines @@
   item->handler = KeyToHandler(key, &item->has_valid_io_context);
   item->context = reinterpret_cast<IOContext*>(overlapped);
   return true;
 }
 
 bool MessagePumpForIO::ProcessInternalIOItem(const IOItem& item) {
   if (reinterpret_cast<void*>(this) == reinterpret_cast<void*>(item.context) &&
       reinterpret_cast<void*>(this) == reinterpret_cast<void*>(item.handler)) {
     // This is our internal completion.
     DCHECK(!item.bytes_transfered);
-    InterlockedExchange(&have_work_, 0);
+    InterlockedExchange(&work_state_, READY);
     return true;
   }
   return false;
 }
 
 // Returns a completion item that was previously received.
 bool MessagePumpForIO::MatchCompletedIOItem(IOHandler* filter, IOItem* item) {
   DCHECK(!completed_io_.empty());
   for (std::list<IOItem>::iterator it = completed_io_.begin();
        it != completed_io_.end(); ++it) {
@@ skipping 40 matching lines @@
 
 // static
 MessagePumpForIO::IOHandler* MessagePumpForIO::KeyToHandler(
     ULONG_PTR key,
     bool* has_valid_io_context) {
   *has_valid_io_context = ((key & 1) == 0);
   return reinterpret_cast<IOHandler*>(key & ~static_cast<ULONG_PTR>(1));
 }
 
 }  // namespace base