Source code

Revision control

Copy as Markdown

Other Tools

/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
/* vim: set ts=8 sts=2 et sw=2 tw=80: */
/* This Source Code Form is subject to the terms of the Mozilla Public
* License, v. 2.0. If a copy of the MPL was not distributed with this
* file, You can obtain one at http://mozilla.org/MPL/2.0/. */
#ifndef XPCOM_THREADS_STATEWATCHING_H_
#define XPCOM_THREADS_STATEWATCHING_H_
#include <cstddef>
#include <new>
#include <utility>
#include "mozilla/AbstractThread.h"
#include "mozilla/Assertions.h"
#include "mozilla/Logging.h"
#include "mozilla/RefPtr.h"
#include "nsCycleCollectionNoteChild.h"
#include "nsISupports.h"
#include "nsTArray.h"
#include "nsThreadUtils.h"
/*
* The state-watching machinery automates the process of responding to changes
* in various pieces of state.
*
* A standard programming pattern is as follows:
*
* mFoo = ...;
* NotifyStuffChanged();
* ...
* mBar = ...;
* NotifyStuffChanged();
*
* This pattern is error-prone and difficult to audit because it requires the
* programmer to manually trigger the update routine. This can be especially
* problematic when the update routine depends on numerous pieces of state, and
* when that state is modified across a variety of helper methods. In these
* cases the responsibility for invoking the routine is often unclear, causing
* developers to scatter calls to it like pixie dust. This can result in
* duplicate invocations (which is wasteful) and missing invocations in corner-
* cases (which is a source of bugs).
*
* This file provides a set of primitives that automatically handle updates and
* allow the programmers to explicitly construct a graph of state dependencies.
* When used correctly, it eliminates the guess-work and wasted cycles described
* above.
*
* There are two basic pieces:
* (1) Objects that can be watched for updates. These inherit WatchTarget.
* (2) Objects that receive objects and trigger processing. These inherit
* AbstractWatcher. In the current machinery, these exist only internally
* within the WatchManager, though that could change.
*
* Note that none of this machinery is thread-safe - it must all happen on the
* same owning thread. To solve multi-threaded use-cases, use state mirroring
* and watch the mirrored value.
*
* Given that semantics may change and comments tend to go out of date, we
* deliberately don't provide usage examples here. Grep around to find them.
*/
namespace mozilla {
extern LazyLogModule gStateWatchingLog;
#define WATCH_LOG(x, ...) \
MOZ_LOG(gStateWatchingLog, LogLevel::Debug, (x, ##__VA_ARGS__))
/*
* AbstractWatcher is a superclass from which all watchers must inherit.
*/
class AbstractWatcher {
public:
NS_INLINE_DECL_THREADSAFE_REFCOUNTING(AbstractWatcher)
AbstractWatcher() : mDestroyed(false) {}
bool IsDestroyed() { return mDestroyed; }
virtual void Notify() = 0;
protected:
virtual ~AbstractWatcher() { MOZ_ASSERT(mDestroyed); }
bool mDestroyed;
};
/*
* WatchTarget is a superclass from which all watchable things must inherit.
* Unlike AbstractWatcher, it is a fully-implemented Mix-in, and the subclass
* needs only to invoke NotifyWatchers when something changes.
*
* The functionality that this class provides is not threadsafe, and should only
* be used on the thread that owns that WatchTarget.
*/
class WatchTarget {
public:
explicit WatchTarget(const char* aName) : mName(aName) {}
void AddWatcher(AbstractWatcher* aWatcher) {
MOZ_ASSERT(!mWatchers.Contains(aWatcher));
mWatchers.AppendElement(aWatcher);
}
void RemoveWatcher(AbstractWatcher* aWatcher) {
MOZ_ASSERT(mWatchers.Contains(aWatcher));
mWatchers.RemoveElement(aWatcher);
}
protected:
void NotifyWatchers() {
WATCH_LOG("%s[%p] notifying watchers\n", mName, this);
PruneWatchers();
for (size_t i = 0; i < mWatchers.Length(); ++i) {
mWatchers[i]->Notify();
}
}
private:
// We don't have Watchers explicitly unregister themselves when they die,
// because then they'd need back-references to all the WatchTargets they're
// subscribed to, and WatchTargets aren't reference-counted. So instead we
// just prune dead ones at appropriate times, which works just fine.
void PruneWatchers() {
mWatchers.RemoveElementsBy(
[](const auto& watcher) { return watcher->IsDestroyed(); });
}
nsTArray<RefPtr<AbstractWatcher>> mWatchers;
protected:
const char* mName;
};
/*
* Watchable is a wrapper class that turns any primitive into a WatchTarget.
*/
template <typename T>
class Watchable : public WatchTarget {
public:
Watchable(const T& aInitialValue, const char* aName)
: WatchTarget(aName), mValue(aInitialValue) {}
const T& Ref() const { return mValue; }
operator const T&() const { return Ref(); }
template <typename U>
Watchable& operator=(U&& aNewValue) {
if (aNewValue != mValue) {
mValue = std::forward<U>(aNewValue);
NotifyWatchers();
}
return *this;
}
private:
Watchable(const Watchable& aOther) = delete;
Watchable& operator=(const Watchable& aOther) = delete;
T mValue;
};
template <typename T>
inline void ImplCycleCollectionUnlink(Watchable<T>& aField) {
ImplCycleCollectionUnlink<T>(aField);
}
template <typename T>
inline void ImplCycleCollectionTraverse(
nsCycleCollectionTraversalCallback& aCallback, const Watchable<T>& aField,
const char* aName, uint32_t aFlags = 0) {
ImplCycleCollectionTraverse(aCallback, aField.Ref(), aName, aFlags);
}
// Manager class for state-watching. Declare one of these in any class for which
// you want to invoke method callbacks.
//
// Internally, WatchManager maintains one AbstractWatcher per callback method.
// Consumers invoke Watch/Unwatch on a particular (WatchTarget, Callback) tuple.
// This causes an AbstractWatcher for |Callback| to be instantiated if it
// doesn't already exist, and registers it with |WatchTarget|.
//
// Using Direct Tasks on the TailDispatcher, WatchManager ensures that we fire
// watch callbacks no more than once per task, once all other operations for
// that task have been completed.
//
// WatchManager<OwnerType> is intended to be declared as a member of |OwnerType|
// objects. Given that, it and its owned objects can't hold permanent strong
// refs to the owner, since that would keep the owner alive indefinitely.
// Instead, it _only_ holds strong refs while waiting for Direct Tasks to fire.
// This ensures that everything is kept alive just long enough.
template <typename OwnerType>
class WatchManager {
public:
typedef void (OwnerType::*CallbackMethod)();
explicit WatchManager(OwnerType* aOwner, AbstractThread* aOwnerThread)
: mOwner(aOwner), mOwnerThread(aOwnerThread) {}
~WatchManager() {
if (!IsShutdown()) {
Shutdown();
}
}
bool IsShutdown() const { return !mOwner; }
// Shutdown needs to happen on mOwnerThread. If the WatchManager will be
// destroyed on a different thread, Shutdown() must be called manually.
void Shutdown() {
MOZ_ASSERT(mOwnerThread->IsCurrentThreadIn());
for (auto& watcher : mWatchers) {
watcher->Destroy();
}
mWatchers.Clear();
mOwner = nullptr;
}
void Watch(WatchTarget& aTarget, CallbackMethod aMethod) {
MOZ_ASSERT(mOwnerThread->IsCurrentThreadIn());
aTarget.AddWatcher(&EnsureWatcher(aMethod));
}
void Unwatch(WatchTarget& aTarget, CallbackMethod aMethod) {
MOZ_ASSERT(mOwnerThread->IsCurrentThreadIn());
PerCallbackWatcher* watcher = GetWatcher(aMethod);
MOZ_ASSERT(watcher);
aTarget.RemoveWatcher(watcher);
}
void ManualNotify(CallbackMethod aMethod) {
MOZ_ASSERT(mOwnerThread->IsCurrentThreadIn());
PerCallbackWatcher* watcher = GetWatcher(aMethod);
MOZ_ASSERT(watcher);
watcher->Notify();
}
private:
class PerCallbackWatcher : public AbstractWatcher {
public:
PerCallbackWatcher(OwnerType* aOwner, AbstractThread* aOwnerThread,
CallbackMethod aMethod)
: mOwner(aOwner),
mOwnerThread(aOwnerThread),
mCallbackMethod(aMethod) {}
void Destroy() {
MOZ_ASSERT(mOwnerThread->IsCurrentThreadIn());
mDestroyed = true;
mOwner = nullptr;
}
void Notify() override {
MOZ_ASSERT(mOwnerThread->IsCurrentThreadIn());
MOZ_DIAGNOSTIC_ASSERT(mOwner,
"mOwner is only null after destruction, "
"at which point we shouldn't be notified");
if (mNotificationPending) {
// We've already got a notification job in the pipe.
return;
}
mNotificationPending = true;
// Queue up our notification jobs to run in a stable state.
AbstractThread::DispatchDirectTask(
NS_NewRunnableFunction("WatchManager::PerCallbackWatcher::Notify",
[self = RefPtr<PerCallbackWatcher>(this),
owner = RefPtr<OwnerType>(mOwner)]() {
if (!self->mDestroyed) {
((*owner).*(self->mCallbackMethod))();
}
self->mNotificationPending = false;
}));
}
bool CallbackMethodIs(CallbackMethod aMethod) const {
return mCallbackMethod == aMethod;
}
private:
~PerCallbackWatcher() = default;
OwnerType* mOwner; // Never null.
bool mNotificationPending = false;
RefPtr<AbstractThread> mOwnerThread;
CallbackMethod mCallbackMethod;
};
PerCallbackWatcher* GetWatcher(CallbackMethod aMethod) {
MOZ_ASSERT(mOwnerThread->IsCurrentThreadIn());
for (auto& watcher : mWatchers) {
if (watcher->CallbackMethodIs(aMethod)) {
return watcher;
}
}
return nullptr;
}
PerCallbackWatcher& EnsureWatcher(CallbackMethod aMethod) {
MOZ_ASSERT(mOwnerThread->IsCurrentThreadIn());
PerCallbackWatcher* watcher = GetWatcher(aMethod);
if (watcher) {
return *watcher;
}
watcher = mWatchers
.AppendElement(MakeAndAddRef<PerCallbackWatcher>(
mOwner, mOwnerThread, aMethod))
->get();
return *watcher;
}
nsTArray<RefPtr<PerCallbackWatcher>> mWatchers;
OwnerType* mOwner;
RefPtr<AbstractThread> mOwnerThread;
};
#undef WATCH_LOG
} // namespace mozilla
#endif