xref: /aosp_15_r20/external/cronet/base/message_loop/message_pump_win.cc (revision 6777b5387eb2ff775bb5750e3f5d96f37fb7352b)
1 // Copyright 2012 The Chromium Authors
2 // Use of this source code is governed by a BSD-style license that can be
3 // found in the LICENSE file.
4 
5 #include "base/message_loop/message_pump_win.h"
6 
7 #include <winbase.h>
8 
9 #include <algorithm>
10 #include <atomic>
11 #include <cstdint>
12 #include <type_traits>
13 
14 #include "base/auto_reset.h"
15 #include "base/check.h"
16 #include "base/debug/alias.h"
17 #include "base/feature_list.h"
18 #include "base/functional/bind.h"
19 #include "base/memory/raw_ptr.h"
20 #include "base/metrics/histogram_macros.h"
21 #include "base/numerics/safe_conversions.h"
22 #include "base/task/task_features.h"
23 #include "base/trace_event/base_tracing.h"
24 #include "base/tracing_buildflags.h"
25 
26 #if BUILDFLAG(ENABLE_BASE_TRACING)
27 #include "third_party/perfetto/protos/perfetto/trace/track_event/chrome_message_pump.pbzero.h"
28 #endif  // BUILDFLAG(ENABLE_BASE_TRACING)
29 
30 namespace base {
31 
32 namespace {
33 
34 // Returns the number of milliseconds before |next_task_time|, clamped between
35 // zero and the biggest DWORD value (or INFINITE if |next_task_time.is_max()|).
36 // Optionally, a recent value of Now() may be passed in to avoid resampling it.
GetSleepTimeoutMs(TimeTicks next_task_time,TimeTicks recent_now=TimeTicks ())37 DWORD GetSleepTimeoutMs(TimeTicks next_task_time,
38                         TimeTicks recent_now = TimeTicks()) {
39   // Shouldn't need to sleep or install a timer when there's pending immediate
40   // work.
41   DCHECK(!next_task_time.is_null());
42 
43   if (next_task_time.is_max())
44     return INFINITE;
45 
46   auto now = recent_now.is_null() ? TimeTicks::Now() : recent_now;
47   auto timeout_ms = (next_task_time - now).InMillisecondsRoundedUp();
48 
49   // A saturated_cast with an unsigned destination automatically clamps negative
50   // values at zero.
51   static_assert(!std::is_signed_v<DWORD>, "DWORD is unexpectedly signed");
52   return saturated_cast<DWORD>(timeout_ms);
53 }
54 
55 bool g_ui_pump_improvements_win = false;
56 
57 }  // namespace
58 
59 // Message sent to get an additional time slice for pumping (processing) another
60 // task (a series of such messages creates a continuous task pump).
61 static const int kMsgHaveWork = WM_USER + 1;
62 
63 //-----------------------------------------------------------------------------
64 // MessagePumpWin public:
65 
66 MessagePumpWin::MessagePumpWin() = default;
67 MessagePumpWin::~MessagePumpWin() = default;
68 
69 // static
InitializeFeatures()70 void MessagePumpWin::InitializeFeatures() {
71   g_ui_pump_improvements_win = FeatureList::IsEnabled(kUIPumpImprovementsWin);
72 }
73 
Run(Delegate * delegate)74 void MessagePumpWin::Run(Delegate* delegate) {
75   DCHECK_CALLED_ON_VALID_THREAD(bound_thread_);
76 
77   RunState run_state(delegate);
78   if (run_state_)
79     run_state.is_nested = true;
80 
81   AutoReset<raw_ptr<RunState>> auto_reset_run_state(&run_state_, &run_state);
82   DoRunLoop();
83 }
84 
Quit()85 void MessagePumpWin::Quit() {
86   DCHECK_CALLED_ON_VALID_THREAD(bound_thread_);
87 
88   DCHECK(run_state_);
89   run_state_->should_quit = true;
90 }
91 
92 //-----------------------------------------------------------------------------
93 // MessagePumpForUI public:
94 
MessagePumpForUI()95 MessagePumpForUI::MessagePumpForUI() {
96   bool succeeded = message_window_.Create(
97       BindRepeating(&MessagePumpForUI::MessageCallback, Unretained(this)));
98   CHECK(succeeded);
99 }
100 
101 MessagePumpForUI::~MessagePumpForUI() = default;
102 
ScheduleWork()103 void MessagePumpForUI::ScheduleWork() {
104   // This is the only MessagePumpForUI method which can be called outside of
105   // |bound_thread_|.
106 
107   if (g_ui_pump_improvements_win &&
108       !in_nested_native_loop_with_application_tasks_) {
109     // The pump is running using `event_` as its chrome-side synchronization
110     // variable. In this case, no deduplication is done, since the event has its
111     // own state.
112     event_.Signal();
113     return;
114   }
115 
116   bool not_scheduled = false;
117   if (!native_msg_scheduled_.compare_exchange_strong(
118           not_scheduled, true, std::memory_order_relaxed)) {
119     return;  // Someone else continued the pumping.
120   }
121 
122   const BOOL ret = ::PostMessage(message_window_.hwnd(), kMsgHaveWork, 0, 0);
123   if (ret) {
124     return;  // There was room in the Window Message queue.
125   }
126 
127   // We have failed to insert a have-work message, so there is a chance that we
128   // will starve tasks/timers while sitting in a nested run loop. Nested loops
129   // only look at Windows Message queues, and don't look at *our* task queues,
130   // etc., so we might not get a time slice in such. :-(
131   // We could abort here, but the fear is that this failure mode is plausibly
132   // common (queue is full, of about 2000 messages), so we'll do a near-graceful
133   // recovery.  Nested loops are pretty transient (we think), so this will
134   // probably be recoverable.
135 
136   // Clarify that we didn't really insert.
137   native_msg_scheduled_.store(false, std::memory_order_relaxed);
138   TRACE_EVENT_INSTANT0("base", "Chrome.MessageLoopProblem.MESSAGE_POST_ERROR",
139                        TRACE_EVENT_SCOPE_THREAD);
140 }
141 
ScheduleDelayedWork(const Delegate::NextWorkInfo & next_work_info)142 void MessagePumpForUI::ScheduleDelayedWork(
143     const Delegate::NextWorkInfo& next_work_info) {
144   DCHECK_CALLED_ON_VALID_THREAD(bound_thread_);
145 
146   // Since this is always called from |bound_thread_|, there is almost always
147   // nothing to do as the loop is already running. When the loop becomes idle,
148   // it will typically WaitForWork() in DoRunLoop() with the timeout provided by
149   // DoWork(). The only alternative to this is entering a native nested loop
150   // (e.g. modal dialog) under a
151   // `ScopedAllowApplicationTasksInNativeNestedLoop`, in which case
152   // HandleWorkMessage() will be invoked when the system picks up kMsgHaveWork
153   // and it will ScheduleNativeTimer() if it's out of immediate work. However,
154   // in that alternate scenario : it's possible for a Windows native work item
155   // (e.g. https://docs.microsoft.com/en-us/windows/desktop/winmsg/using-hooks)
156   // to wake the native nested loop and PostDelayedTask() to the current thread
157   // from it. This is the only case where we must install/adjust the native
158   // timer from ScheduleDelayedWork() because if we don't, the native loop will
159   // go back to sleep, unaware of the new |delayed_work_time|.
160   // See MessageLoopTest.PostDelayedTaskFromSystemPump for an example.
161   // TODO(gab): This could potentially be replaced by a ForegroundIdleProc hook
162   // if Windows ends up being the only platform requiring ScheduleDelayedWork().
163   if (in_nested_native_loop_with_application_tasks_ &&
164       !native_msg_scheduled_.load(std::memory_order_relaxed)) {
165     ScheduleNativeTimer(next_work_info);
166   }
167 }
168 
HandleNestedNativeLoopWithApplicationTasks(bool application_tasks_desired)169 bool MessagePumpForUI::HandleNestedNativeLoopWithApplicationTasks(
170     bool application_tasks_desired) {
171   // It is here assumed that we will be in a native loop until either
172   // DoRunLoop() gets control back, or this is called with `false`, and thus the
173   // Windows event queue is to be used for synchronization. This is to prevent
174   // being unable to wake up for application tasks in the case of a nested loop.
175   in_nested_native_loop_with_application_tasks_ = application_tasks_desired;
176   if (application_tasks_desired) {
177     ScheduleWork();
178   }
179   return true;
180 }
181 
AddObserver(Observer * observer)182 void MessagePumpForUI::AddObserver(Observer* observer) {
183   DCHECK_CALLED_ON_VALID_THREAD(bound_thread_);
184   observers_.AddObserver(observer);
185 }
186 
RemoveObserver(Observer * observer)187 void MessagePumpForUI::RemoveObserver(Observer* observer) {
188   DCHECK_CALLED_ON_VALID_THREAD(bound_thread_);
189   observers_.RemoveObserver(observer);
190 }
191 
192 //-----------------------------------------------------------------------------
193 // MessagePumpForUI private:
194 
MessageCallback(UINT message,WPARAM wparam,LPARAM lparam,LRESULT * result)195 bool MessagePumpForUI::MessageCallback(
196     UINT message, WPARAM wparam, LPARAM lparam, LRESULT* result) {
197   DCHECK_CALLED_ON_VALID_THREAD(bound_thread_);
198   switch (message) {
199     case kMsgHaveWork:
200       HandleWorkMessage();
201       break;
202     case WM_TIMER:
203       if (wparam == reinterpret_cast<UINT_PTR>(this))
204         HandleTimerMessage();
205       break;
206   }
207   return false;
208 }
209 
DoRunLoop()210 void MessagePumpForUI::DoRunLoop() {
211   DCHECK_CALLED_ON_VALID_THREAD(bound_thread_);
212 
213   // IF this was just a simple PeekMessage() loop (servicing all possible work
214   // queues), then Windows would try to achieve the following order according
215   // to MSDN documentation about PeekMessage with no filter):
216   //    * Sent messages
217   //    * Posted messages
218   //    * Sent messages (again)
219   //    * WM_PAINT messages
220   //    * WM_TIMER messages
221   //
222   // Summary: none of the above classes is starved, and sent messages has twice
223   // the chance of being processed (i.e., reduced service time).
224 
225   wakeup_state_ = WakeupState::kRunning;
226 
227   for (;;) {
228     // If we do any work, we may create more messages etc., and more work may
229     // possibly be waiting in another task group.  When we (for example)
230     // ProcessNextWindowsMessage(), there is a good chance there are still more
231     // messages waiting.  On the other hand, when any of these methods return
232     // having done no work, then it is pretty unlikely that calling them again
233     // quickly will find any work to do. Finally, if they all say they had no
234     // work, then it is a good time to consider sleeping (waiting) for more
235     // work.
236 
237     in_nested_native_loop_with_application_tasks_ = false;
238     bool more_work_is_plausible = false;
239 
240     if (!g_ui_pump_improvements_win ||
241         wakeup_state_ != WakeupState::kApplicationTask) {
242       more_work_is_plausible |= ProcessNextWindowsMessage();
243       // We can end up in native loops which allow application tasks outside of
244       // DoWork() when Windows calls back a Win32 message window owned by some
245       // Chromium code.
246       in_nested_native_loop_with_application_tasks_ = false;
247       if (run_state_->should_quit) {
248         break;
249       }
250     }
251 
252     Delegate::NextWorkInfo next_work_info = run_state_->delegate->DoWork();
253     // Since nested native loops with application tasks are initiated by a
254     // scoper, they should always be cleared before exiting DoWork().
255     DCHECK(!in_nested_native_loop_with_application_tasks_);
256     wakeup_state_ = WakeupState::kRunning;
257     more_work_is_plausible |= next_work_info.is_immediate();
258 
259     if (run_state_->should_quit) {
260       break;
261     }
262 
263     if (installed_native_timer_) {
264       // As described in ScheduleNativeTimer(), the native timer is only
265       // installed and needed while in a nested native loop. If it is installed,
266       // it means the above work entered such a loop. Having now resumed, the
267       // native timer is no longer needed.
268       KillNativeTimer();
269     }
270 
271     if (more_work_is_plausible)
272       continue;
273 
274     more_work_is_plausible = run_state_->delegate->DoIdleWork();
275     // DoIdleWork() shouldn't end up in native nested loops, nor should it
276     // permit native nested loops, and thus shouldn't have any chance of
277     // reinstalling a native timer.
278     DCHECK(!in_nested_native_loop_with_application_tasks_);
279     DCHECK(!installed_native_timer_);
280     if (run_state_->should_quit) {
281       break;
282     }
283 
284     if (more_work_is_plausible)
285       continue;
286 
287     WaitForWork(next_work_info);
288   }
289 }
290 
WaitForWork(Delegate::NextWorkInfo next_work_info)291 void MessagePumpForUI::WaitForWork(Delegate::NextWorkInfo next_work_info) {
292   DCHECK_CALLED_ON_VALID_THREAD(bound_thread_);
293 
294   // Wait until a message is available, up to the time needed by the timer
295   // manager to fire the next set of timers.
296   DWORD wait_flags = MWMO_INPUTAVAILABLE;
297   bool last_wakeup_was_spurious = false;
298   for (DWORD delay = GetSleepTimeoutMs(next_work_info.delayed_run_time,
299                                        next_work_info.recent_now);
300        delay != 0; delay = GetSleepTimeoutMs(next_work_info.delayed_run_time)) {
301     if (!last_wakeup_was_spurious) {
302       run_state_->delegate->BeforeWait();
303     }
304     last_wakeup_was_spurious = false;
305 
306     // Tell the optimizer to retain these values to simplify analyzing hangs.
307     base::debug::Alias(&delay);
308     base::debug::Alias(&wait_flags);
309     DWORD result;
310     if (g_ui_pump_improvements_win) {
311       HANDLE event_handle = event_.handle();
312       result = MsgWaitForMultipleObjectsEx(1, &event_handle, delay, QS_ALLINPUT,
313                                            wait_flags);
314       DPCHECK(WAIT_FAILED != result);
315       if (result == WAIT_OBJECT_0) {
316         wakeup_state_ = WakeupState::kApplicationTask;
317       } else if (result == WAIT_OBJECT_0 + 1) {
318         wakeup_state_ = WakeupState::kNative;
319       } else {
320         wakeup_state_ = WakeupState::kInactive;
321       }
322     } else {
323       result = MsgWaitForMultipleObjectsEx(0, nullptr, delay, QS_ALLINPUT,
324                                            wait_flags);
325       DPCHECK(WAIT_FAILED != result);
326       if (result == WAIT_OBJECT_0) {
327         wakeup_state_ = WakeupState::kNative;
328       } else {
329         wakeup_state_ = WakeupState::kInactive;
330       }
331     }
332 
333     if (wakeup_state_ == WakeupState::kApplicationTask) {
334       // This can only be reached when the pump woke up via `event_`. In that
335       // case, tasks are prioritized over native.
336       return;
337     } else if (wakeup_state_ == WakeupState::kNative) {
338       // A WM_* message is available.
339       // If a parent child relationship exists between windows across threads
340       // then their thread inputs are implicitly attached.
341       // This causes the MsgWaitForMultipleObjectsEx API to return indicating
342       // that messages are ready for processing (Specifically, mouse messages
343       // intended for the child window may appear if the child window has
344       // capture).
345       // The subsequent PeekMessages call may fail to return any messages thus
346       // causing us to enter a tight loop at times.
347       // The code below is a workaround to give the child window
348       // some time to process its input messages by looping back to
349       // MsgWaitForMultipleObjectsEx above when there are no messages for the
350       // current thread.
351 
352       // As in ProcessNextWindowsMessage().
353       auto scoped_do_work_item = run_state_->delegate->BeginWorkItem();
354       {
355         TRACE_EVENT0(TRACE_DISABLED_BY_DEFAULT("base"),
356                      "MessagePumpForUI::WaitForWork GetQueueStatus");
357         if (HIWORD(::GetQueueStatus(QS_SENDMESSAGE)) & QS_SENDMESSAGE)
358           return;
359       }
360       {
361         MSG msg;
362         TRACE_EVENT0(TRACE_DISABLED_BY_DEFAULT("base"),
363                      "MessagePumpForUI::WaitForWork PeekMessage");
364         if (::PeekMessage(&msg, nullptr, 0, 0, PM_NOREMOVE)) {
365           return;
366         }
367       }
368 
369       // We know there are no more messages for this thread because PeekMessage
370       // has returned false. Reset |wait_flags| so that we wait for a *new*
371       // message.
372       wait_flags = 0;
373     } else {
374       DCHECK_EQ(wakeup_state_, WakeupState::kInactive);
375       last_wakeup_was_spurious = true;
376       TRACE_EVENT_INSTANT(
377           "base", "MessagePumpForUI::WaitForWork Spurious Wakeup",
378           [&](perfetto::EventContext ctx) {
379             ctx.event<perfetto::protos::pbzero::ChromeTrackEvent>()
380                 ->set_chrome_message_pump_for_ui()
381                 ->set_wait_for_object_result(result);
382           });
383     }
384   }
385 }
386 
HandleWorkMessage()387 void MessagePumpForUI::HandleWorkMessage() {
388   DCHECK_CALLED_ON_VALID_THREAD(bound_thread_);
389 
390   // If we are being called outside of the context of Run, then don't try to do
391   // any work.  This could correspond to a MessageBox call or something of that
392   // sort.
393   if (!run_state_) {
394     // Since we handled a kMsgHaveWork message, we must still update this flag.
395     native_msg_scheduled_.store(false, std::memory_order_relaxed);
396     return;
397   }
398 
399   // Let whatever would have run had we not been putting messages in the queue
400   // run now.  This is an attempt to make our dummy message not starve other
401   // messages that may be in the Windows message queue.
402   ProcessPumpReplacementMessage();
403 
404   Delegate::NextWorkInfo next_work_info = run_state_->delegate->DoWork();
405   if (next_work_info.is_immediate()) {
406     ScheduleWork();
407   } else {
408     run_state_->delegate->BeforeWait();
409     ScheduleNativeTimer(next_work_info);
410   }
411 }
412 
HandleTimerMessage()413 void MessagePumpForUI::HandleTimerMessage() {
414   DCHECK_CALLED_ON_VALID_THREAD(bound_thread_);
415 
416   // ::KillTimer doesn't remove pending WM_TIMER messages from the queue,
417   // explicitly ignore the last WM_TIMER message in that case to avoid handling
418   // work from here when DoRunLoop() is active (which could result in scheduling
419   // work from two places at once). Note: we're still fine in the event that a
420   // second native nested loop is entered before such a dead WM_TIMER message is
421   // discarded because ::SetTimer merely resets the timer if invoked twice with
422   // the same id.
423   if (!installed_native_timer_)
424     return;
425 
426   // We only need to fire once per specific delay, another timer may be
427   // scheduled below but we're done with this one.
428   KillNativeTimer();
429 
430   // If we are being called outside of the context of Run, then don't do
431   // anything.  This could correspond to a MessageBox call or something of
432   // that sort.
433   if (!run_state_)
434     return;
435 
436   Delegate::NextWorkInfo next_work_info = run_state_->delegate->DoWork();
437   if (next_work_info.is_immediate()) {
438     ScheduleWork();
439   } else {
440     run_state_->delegate->BeforeWait();
441     ScheduleNativeTimer(next_work_info);
442   }
443 }
444 
ScheduleNativeTimer(Delegate::NextWorkInfo next_work_info)445 void MessagePumpForUI::ScheduleNativeTimer(
446     Delegate::NextWorkInfo next_work_info) {
447   DCHECK(!next_work_info.is_immediate());
448   // We should only ScheduleNativeTimer() under the new pump implementation
449   // while nested with application tasks.
450   DCHECK(!g_ui_pump_improvements_win ||
451          in_nested_native_loop_with_application_tasks_);
452 
453   // Do not redundantly set the same native timer again if it was already set.
454   // This can happen when a nested native loop goes idle with pending delayed
455   // tasks, then gets woken up by an immediate task, and goes back to idle with
456   // the same pending delay. No need to kill the native timer if there is
457   // already one but the |delayed_run_time| has changed as ::SetTimer reuses the
458   // same id and will replace and reset the existing timer.
459   if (installed_native_timer_ &&
460       *installed_native_timer_ == next_work_info.delayed_run_time) {
461     return;
462   }
463 
464   if (next_work_info.delayed_run_time.is_max())
465     return;
466 
467   // We do not use native Windows timers in general as they have a poor, 10ms,
468   // granularity. Instead we rely on MsgWaitForMultipleObjectsEx's
469   // high-resolution timeout to sleep without timers in WaitForWork(). However,
470   // when entering a nested native ::GetMessage() loop (e.g. native modal
471   // windows) under a `ScopedAllowApplicationTasksInNativeNestedLoop`, we have
472   // to rely on a native timer when HandleWorkMessage() runs out of immediate
473   // work. Since `ScopedAllowApplicationTasksInNativeNestedLoop` invokes
474   // ScheduleWork() : we are guaranteed that HandleWorkMessage() will be called
475   // after entering a nested native loop that should process application
476   // tasks. But once HandleWorkMessage() is out of immediate work, ::SetTimer()
477   // is used to guarantee we are invoked again should the next delayed task
478   // expire before the nested native loop ends. The native timer being
479   // unnecessary once we return to our DoRunLoop(), we ::KillTimer when it
480   // resumes (nested native loops should be rare so we're not worried about
481   // ::SetTimer<=>::KillTimer churn).  TODO(gab): The long-standing legacy
482   // dependency on the behavior of
483   // `ScopedAllowApplicationTasksInNativeNestedLoop` is unfortunate, would be
484   // nice to make this a MessagePump concept (instead of requiring impls to
485   // invoke ScheduleWork() one-way and no-op DoWork() the other way).
486 
487   UINT delay_msec = strict_cast<UINT>(GetSleepTimeoutMs(
488       next_work_info.delayed_run_time, next_work_info.recent_now));
489   if (delay_msec == 0) {
490     ScheduleWork();
491   } else {
492     // TODO(gab): ::SetTimer()'s documentation claims it does this for us.
493     // Consider removing this safety net.
494     delay_msec = std::clamp(delay_msec, static_cast<UINT>(USER_TIMER_MINIMUM),
495                             static_cast<UINT>(USER_TIMER_MAXIMUM));
496 
497     // Tell the optimizer to retain the delay to simplify analyzing hangs.
498     base::debug::Alias(&delay_msec);
499     const UINT_PTR ret =
500         ::SetTimer(message_window_.hwnd(), reinterpret_cast<UINT_PTR>(this),
501                    delay_msec, nullptr);
502 
503     if (ret) {
504       installed_native_timer_ = next_work_info.delayed_run_time;
505       return;
506     }
507     // This error is likely similar to MESSAGE_POST_ERROR (i.e. native queue is
508     // full). Since we only use ScheduleNativeTimer() in native nested loops
509     // this likely means this pump will not be given a chance to run application
510     // tasks until the nested loop completes.
511     TRACE_EVENT_INSTANT0("base", "Chrome.MessageLoopProblem.SET_TIMER_ERROR",
512                          TRACE_EVENT_SCOPE_THREAD);
513   }
514 }
515 
KillNativeTimer()516 void MessagePumpForUI::KillNativeTimer() {
517   DCHECK(installed_native_timer_);
518   const bool success =
519       ::KillTimer(message_window_.hwnd(), reinterpret_cast<UINT_PTR>(this));
520   DPCHECK(success);
521   installed_native_timer_.reset();
522 }
523 
ProcessNextWindowsMessage()524 bool MessagePumpForUI::ProcessNextWindowsMessage() {
525   DCHECK_CALLED_ON_VALID_THREAD(bound_thread_);
526 
527   MSG msg;
528   bool has_msg = false;
529   bool more_work_is_plausible = false;
530   {
531     // ::PeekMessage() may process sent and/or internal messages (regardless of
532     // |had_messages| as ::GetQueueStatus() is an optimistic check that may
533     // racily have missed an incoming event -- it doesn't hurt to have empty
534     // internal units of work when ::PeekMessage turns out to be a no-op).
535     // Instantiate |scoped_do_work| ahead of GetQueueStatus() so that
536     // trace events it emits fully outscope GetQueueStatus' events
537     // (GetQueueStatus() itself not being expected to do work; it's fine to use
538     // only one ScopedDoWorkItem for both calls -- we trace them independently
539     // just in case internal work stalls).
540     auto scoped_do_work_item = run_state_->delegate->BeginWorkItem();
541 
542     {
543       // Individually trace ::GetQueueStatus and ::PeekMessage because sampling
544       // profiler is hinting that we're spending a surprising amount of time
545       // with these on top of the stack. Tracing will be able to tell us whether
546       // this is a bias of sampling profiler (e.g. kernel takes ::GetQueueStatus
547       // as an opportunity to swap threads and is more likely to schedule the
548       // sampling profiler's thread while the sampled thread is swapped out on
549       // this frame).
550       TRACE_EVENT0(
551           TRACE_DISABLED_BY_DEFAULT("base"),
552           "MessagePumpForUI::ProcessNextWindowsMessage GetQueueStatus");
553       DWORD queue_status = ::GetQueueStatus(QS_SENDMESSAGE);
554 
555       // If there are sent messages in the queue then PeekMessage internally
556       // dispatches the message and returns false. We return true in this case
557       // to ensure that the message loop peeks again instead of calling
558       // MsgWaitForMultipleObjectsEx.
559       if (HIWORD(queue_status) & QS_SENDMESSAGE)
560         more_work_is_plausible = true;
561     }
562 
563     {
564       // PeekMessage can run a message if there are sent messages, trace that
565       // and emit the boolean param to see if it ever janks independently (ref.
566       // comment on GetQueueStatus).
567       TRACE_EVENT(
568           TRACE_DISABLED_BY_DEFAULT("base"),
569           "MessagePumpForUI::ProcessNextWindowsMessage PeekMessage",
570           [&](perfetto::EventContext ctx) {
571             perfetto::protos::pbzero::ChromeMessagePump* msg_pump_data =
572                 ctx.event()->set_chrome_message_pump();
573             msg_pump_data->set_sent_messages_in_queue(more_work_is_plausible);
574           });
575       has_msg = ::PeekMessage(&msg, nullptr, 0, 0, PM_REMOVE) != FALSE;
576     }
577   }
578   if (has_msg)
579     more_work_is_plausible |= ProcessMessageHelper(msg);
580 
581   return more_work_is_plausible;
582 }
583 
ProcessMessageHelper(const MSG & msg)584 bool MessagePumpForUI::ProcessMessageHelper(const MSG& msg) {
585   DCHECK_CALLED_ON_VALID_THREAD(bound_thread_);
586 
587   if (msg.message == WM_QUIT) {
588     // WM_QUIT is the standard way to exit a ::GetMessage() loop. Our
589     // MessageLoop has its own quit mechanism, so WM_QUIT is generally
590     // unexpected.
591     TRACE_EVENT_INSTANT0("base",
592                          "Chrome.MessageLoopProblem.RECEIVED_WM_QUIT_ERROR",
593                          TRACE_EVENT_SCOPE_THREAD);
594     return true;
595   }
596 
597   // While running our main message pump, we discard kMsgHaveWork messages.
598   if (msg.message == kMsgHaveWork && msg.hwnd == message_window_.hwnd())
599     return ProcessPumpReplacementMessage();
600 
601   run_state_->delegate->BeginNativeWorkBeforeDoWork();
602   auto scoped_do_work_item = run_state_->delegate->BeginWorkItem();
603 
604   TRACE_EVENT("base,toplevel", "MessagePumpForUI DispatchMessage",
605               [&](perfetto::EventContext ctx) {
606                 ctx.event<perfetto::protos::pbzero::ChromeTrackEvent>()
607                     ->set_chrome_message_pump_for_ui()
608                     ->set_message_id(msg.message);
609               });
610 
611   for (Observer& observer : observers_)
612     observer.WillDispatchMSG(msg);
613   ::TranslateMessage(&msg);
614   ::DispatchMessage(&msg);
615   for (Observer& observer : observers_)
616     observer.DidDispatchMSG(msg);
617 
618   return true;
619 }
620 
ProcessPumpReplacementMessage()621 bool MessagePumpForUI::ProcessPumpReplacementMessage() {
622   DCHECK_CALLED_ON_VALID_THREAD(bound_thread_);
623 
624   // When we encounter a kMsgHaveWork message, this method is called to peek and
625   // process a replacement message. The goal is to make the kMsgHaveWork as non-
626   // intrusive as possible, even though a continuous stream of such messages are
627   // posted. This method carefully peeks a message while there is no chance for
628   // a kMsgHaveWork to be pending, then resets the |have_work_| flag (allowing a
629   // replacement kMsgHaveWork to possibly be posted), and finally dispatches
630   // that peeked replacement. Note that the re-post of kMsgHaveWork may be
631   // asynchronous to this thread!!
632 
633   MSG msg;
634   bool have_message = false;
635   {
636     // Note: Ideally this call wouldn't process sent-messages (as we already did
637     // that in the PeekMessage call that lead to receiving this kMsgHaveWork),
638     // but there's no way to specify this (omitting PM_QS_SENDMESSAGE as in
639     // crrev.com/791043 doesn't do anything). Hence this call must be considered
640     // as a potential work item.
641     auto scoped_do_work_item = run_state_->delegate->BeginWorkItem();
642     TRACE_EVENT0(TRACE_DISABLED_BY_DEFAULT("base"),
643                  "MessagePumpForUI::ProcessPumpReplacementMessage PeekMessage");
644     have_message = ::PeekMessage(&msg, nullptr, 0, 0, PM_REMOVE) != FALSE;
645   }
646 
647   // Expect no message or a message different than kMsgHaveWork.
648   DCHECK(!have_message || kMsgHaveWork != msg.message ||
649          msg.hwnd != message_window_.hwnd());
650 
651   // Since we discarded a kMsgHaveWork message, we must update the flag.
652   DCHECK(native_msg_scheduled_.load(std::memory_order_relaxed));
653   native_msg_scheduled_.store(false, std::memory_order_relaxed);
654 
655   // We don't need a special time slice if we didn't |have_message| to process.
656   if (!have_message)
657     return false;
658 
659   if (msg.message == WM_QUIT) {
660     // If we're in a nested ::GetMessage() loop then we must let that loop see
661     // the WM_QUIT in order for it to exit. If we're in DoRunLoop then the re-
662     // posted WM_QUIT will be either ignored, or handled, by
663     // ProcessMessageHelper() called directly from ProcessNextWindowsMessage().
664     ::PostQuitMessage(static_cast<int>(msg.wParam));
665     // Note: we *must not* ScheduleWork() here as WM_QUIT is a low-priority
666     // message on Windows (it is only returned by ::PeekMessage() when idle) :
667     // https://blogs.msdn.microsoft.com/oldnewthing/20051104-33/?p=33453. As
668     // such posting a kMsgHaveWork message via ScheduleWork() would cause an
669     // infinite loop (kMsgHaveWork message handled first means we end up here
670     // again and repost WM_QUIT+ScheduleWork() again, etc.). Not leaving a
671     // kMsgHaveWork message behind however is also problematic as unwinding
672     // multiple layers of nested ::GetMessage() loops can result in starving
673     // application tasks. TODO(https://crbug.com/890016) : Fix this.
674 
675     // The return value is mostly irrelevant but return true like we would after
676     // processing a QuitClosure() task.
677     return true;
678   } else if (msg.message == WM_TIMER &&
679              msg.wParam == reinterpret_cast<UINT_PTR>(this)) {
680     // This happens when a native nested loop invokes HandleWorkMessage() =>
681     // ProcessPumpReplacementMessage() which finds the WM_TIMER message
682     // installed by ScheduleNativeTimer(). That message needs to be handled
683     // directly as handing it off to ProcessMessageHelper() below would cause an
684     // unnecessary ScopedDoWorkItem which may incorrectly lead the Delegate's
685     // heuristics to conclude that the DoWork() in HandleTimerMessage() is
686     // nested inside a native work item. It's also safe to skip the below
687     // ScheduleWork() as it is not mandatory before invoking DoWork() and
688     // HandleTimerMessage() handles re-installing the necessary followup
689     // messages.
690     HandleTimerMessage();
691     return true;
692   }
693 
694   // Guarantee we'll get another time slice in the case where we go into native
695   // windows code. This ScheduleWork() may hurt performance a tiny bit when
696   // tasks appear very infrequently, but when the event queue is busy, the
697   // kMsgHaveWork events get (percentage wise) rarer and rarer.
698   ScheduleWork();
699   return ProcessMessageHelper(msg);
700 }
701 
702 //-----------------------------------------------------------------------------
703 // MessagePumpForIO public:
704 
IOContext()705 MessagePumpForIO::IOContext::IOContext() {
706   memset(&overlapped, 0, sizeof(overlapped));
707 }
708 
IOHandler(const Location & from_here)709 MessagePumpForIO::IOHandler::IOHandler(const Location& from_here)
710     : io_handler_location_(from_here) {}
711 
712 MessagePumpForIO::IOHandler::~IOHandler() = default;
713 
MessagePumpForIO()714 MessagePumpForIO::MessagePumpForIO() {
715   port_.Set(::CreateIoCompletionPort(INVALID_HANDLE_VALUE, nullptr,
716                                      reinterpret_cast<ULONG_PTR>(nullptr), 1));
717   DCHECK(port_.is_valid());
718 }
719 
720 MessagePumpForIO::~MessagePumpForIO() = default;
721 
ScheduleWork()722 void MessagePumpForIO::ScheduleWork() {
723   // This is the only MessagePumpForIO method which can be called outside of
724   // |bound_thread_|.
725 
726   bool not_scheduled = false;
727   if (!native_msg_scheduled_.compare_exchange_strong(
728           not_scheduled, true, std::memory_order_relaxed)) {
729     return;  // Work already scheduled.
730   }
731 
732   // Make sure the MessagePump does some work for us.
733   const BOOL ret = ::PostQueuedCompletionStatus(
734       port_.get(), 0, reinterpret_cast<ULONG_PTR>(this),
735       reinterpret_cast<OVERLAPPED*>(this));
736   if (ret)
737     return;  // Post worked perfectly.
738 
739   // See comment in MessagePumpForUI::ScheduleWork() for this error recovery.
740 
741   native_msg_scheduled_.store(
742       false, std::memory_order_relaxed);  // Clarify that we didn't succeed.
743   TRACE_EVENT_INSTANT0("base",
744                        "Chrome.MessageLoopProblem.COMPLETION_POST_ERROR",
745                        TRACE_EVENT_SCOPE_THREAD);
746 }
747 
ScheduleDelayedWork(const Delegate::NextWorkInfo & next_work_info)748 void MessagePumpForIO::ScheduleDelayedWork(
749     const Delegate::NextWorkInfo& next_work_info) {
750   DCHECK_CALLED_ON_VALID_THREAD(bound_thread_);
751 
752   // Since this is always called from |bound_thread_|, there is nothing to do as
753   // the loop is already running. It will WaitForWork() in
754   // DoRunLoop() with the correct timeout when it's out of immediate tasks.
755 }
756 
RegisterIOHandler(HANDLE file_handle,IOHandler * handler)757 HRESULT MessagePumpForIO::RegisterIOHandler(HANDLE file_handle,
758                                             IOHandler* handler) {
759   DCHECK_CALLED_ON_VALID_THREAD(bound_thread_);
760 
761   HANDLE port = ::CreateIoCompletionPort(
762       file_handle, port_.get(), reinterpret_cast<ULONG_PTR>(handler), 1);
763   return (port != nullptr) ? S_OK : HRESULT_FROM_WIN32(GetLastError());
764 }
765 
RegisterJobObject(HANDLE job_handle,IOHandler * handler)766 bool MessagePumpForIO::RegisterJobObject(HANDLE job_handle,
767                                          IOHandler* handler) {
768   DCHECK_CALLED_ON_VALID_THREAD(bound_thread_);
769 
770   JOBOBJECT_ASSOCIATE_COMPLETION_PORT info;
771   info.CompletionKey = handler;
772   info.CompletionPort = port_.get();
773   return ::SetInformationJobObject(job_handle,
774                                    JobObjectAssociateCompletionPortInformation,
775                                    &info, sizeof(info)) != FALSE;
776 }
777 
778 //-----------------------------------------------------------------------------
779 // MessagePumpForIO private:
780 
DoRunLoop()781 void MessagePumpForIO::DoRunLoop() {
782   DCHECK_CALLED_ON_VALID_THREAD(bound_thread_);
783 
784   for (;;) {
785     // If we do any work, we may create more messages etc., and more work may
786     // possibly be waiting in another task group.  When we (for example)
787     // WaitForIOCompletion(), there is a good chance there are still more
788     // messages waiting.  On the other hand, when any of these methods return
789     // having done no work, then it is pretty unlikely that calling them
790     // again quickly will find any work to do.  Finally, if they all say they
791     // had no work, then it is a good time to consider sleeping (waiting) for
792     // more work.
793 
794     Delegate::NextWorkInfo next_work_info = run_state_->delegate->DoWork();
795     bool more_work_is_plausible = next_work_info.is_immediate();
796     if (run_state_->should_quit)
797       break;
798 
799     more_work_is_plausible |= WaitForIOCompletion(0);
800     if (run_state_->should_quit)
801       break;
802 
803     if (more_work_is_plausible)
804       continue;
805 
806     more_work_is_plausible = run_state_->delegate->DoIdleWork();
807     if (run_state_->should_quit)
808       break;
809 
810     if (more_work_is_plausible)
811       continue;
812 
813     run_state_->delegate->BeforeWait();
814     WaitForWork(next_work_info);
815   }
816 }
817 
818 // Wait until IO completes, up to the time needed by the timer manager to fire
819 // the next set of timers.
WaitForWork(Delegate::NextWorkInfo next_work_info)820 void MessagePumpForIO::WaitForWork(Delegate::NextWorkInfo next_work_info) {
821   DCHECK_CALLED_ON_VALID_THREAD(bound_thread_);
822 
823   // We do not support nested IO message loops. This is to avoid messy
824   // recursion problems.
825   DCHECK(!run_state_->is_nested) << "Cannot nest an IO message loop!";
826 
827   DWORD timeout = GetSleepTimeoutMs(next_work_info.delayed_run_time,
828                                     next_work_info.recent_now);
829 
830   // Tell the optimizer to retain these values to simplify analyzing hangs.
831   base::debug::Alias(&timeout);
832   WaitForIOCompletion(timeout);
833 }
834 
WaitForIOCompletion(DWORD timeout)835 bool MessagePumpForIO::WaitForIOCompletion(DWORD timeout) {
836   DCHECK_CALLED_ON_VALID_THREAD(bound_thread_);
837 
838   IOItem item;
839   if (!GetIOItem(timeout, &item))
840     return false;
841 
842   if (ProcessInternalIOItem(item))
843     return true;
844 
845   run_state_->delegate->BeginNativeWorkBeforeDoWork();
846   auto scoped_do_work_item = run_state_->delegate->BeginWorkItem();
847 
848   TRACE_EVENT(
849       "base,toplevel", "IOHandler::OnIOCompleted",
850       [&](perfetto::EventContext ctx) {
851         ctx.event()->set_chrome_message_pump()->set_io_handler_location_iid(
852             base::trace_event::InternedSourceLocation::Get(
853                 &ctx, base::trace_event::TraceSourceLocation(
854                           item.handler->io_handler_location())));
855       });
856 
857   item.handler.ExtractAsDangling()->OnIOCompleted(
858       item.context.ExtractAsDangling(), item.bytes_transfered, item.error);
859 
860   return true;
861 }
862 
863 // Asks the OS for another IO completion result.
GetIOItem(DWORD timeout,IOItem * item)864 bool MessagePumpForIO::GetIOItem(DWORD timeout, IOItem* item) {
865   DCHECK_CALLED_ON_VALID_THREAD(bound_thread_);
866 
867   memset(item, 0, sizeof(*item));
868   ULONG_PTR key = reinterpret_cast<ULONG_PTR>(nullptr);
869   OVERLAPPED* overlapped = nullptr;
870   if (!::GetQueuedCompletionStatus(port_.get(), &item->bytes_transfered, &key,
871                                    &overlapped, timeout)) {
872     if (!overlapped)
873       return false;  // Nothing in the queue.
874     item->error = GetLastError();
875     item->bytes_transfered = 0;
876   }
877 
878   item->handler = reinterpret_cast<IOHandler*>(key);
879   item->context = reinterpret_cast<IOContext*>(overlapped);
880   return true;
881 }
882 
ProcessInternalIOItem(const IOItem & item)883 bool MessagePumpForIO::ProcessInternalIOItem(const IOItem& item) {
884   DCHECK_CALLED_ON_VALID_THREAD(bound_thread_);
885 
886   if (reinterpret_cast<void*>(this) ==
887           reinterpret_cast<void*>(item.context.get()) &&
888       reinterpret_cast<void*>(this) ==
889           reinterpret_cast<void*>(item.handler.get())) {
890     // This is our internal completion.
891     DCHECK(!item.bytes_transfered);
892     native_msg_scheduled_.store(false, std::memory_order_relaxed);
893     return true;
894   }
895   return false;
896 }
897 
898 }  // namespace base
899