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
#include <windows.h>
#include <threadpoolapiset.h>
#include "mozilla/AlreadyAddRefed.h"
#include "mozilla/Assertions.h"
#include "mozilla/Logging.h"
#include "mozilla/Mutex.h"
#include "mozilla/RefPtr.h"
#include "mozilla/ThreadSafety.h"
#include "mozilla/WinHandleWatcher.h"
#include "nsCOMPtr.h"
#include "nsIRunnable.h"
#include "nsISerialEventTarget.h"
#include "nsISupportsImpl.h"
#include "nsITargetShutdownTask.h"
#include "nsIWeakReferenceUtils.h"
#include "nsThreadUtils.h"
mozilla::LazyLogModule sHWLog("HandleWatcher");
namespace mozilla {
namespace details {
struct WaitHandleDeleter {
void operator()(PTP_WAIT waitHandle) {
MOZ_LOG(sHWLog, LogLevel::Debug, ("Closing PTP_WAIT %p", waitHandle));
::CloseThreadpoolWait(waitHandle);
}
};
} // namespace details
using WaitHandlePtr = UniquePtr<TP_WAIT, details::WaitHandleDeleter>;
// HandleWatcher::Impl
//
// The backing implementation of HandleWatcher is a PTP_WAIT, an OS-threadpool
// wait-object. Windows doesn't actually create a new thread per wait-object;
// OS-threadpool threads are assigned to wait-objects only when their associated
// handle become signaled -- although explicit documentation of this fact is
// somewhat obscurely placed. [1]
//
// Throughout this class, we use manual locking and unlocking guarded by Clang's
// thread-safety warnings, rather than scope-based lock-guards. See `Replace()`
// for an explanation and justification.
//
class HandleWatcher::Impl final : public nsITargetShutdownTask {
NS_DECL_THREADSAFE_ISUPPORTS
public:
Impl() = default;
private:
~Impl() { MOZ_ASSERT(IsStopped()); }
struct Data {
// The watched handle and its callback.
HANDLE handle;
RefPtr<nsIEventTarget> target;
nsCOMPtr<nsIRunnable> runnable;
// Handle to the threadpool wait-object.
WaitHandlePtr waitHandle;
// A pointer to ourselves, notionally owned by the wait-object.
RefPtr<Impl> self;
// (We can't actually do this because a) it has annoying consequences in
// C++20 thanks to P1008R1, and b) Clang just ignores it anyway.)
//
// ~Data() MOZ_EXCLUDES(mMutex) = default;
};
mozilla::Mutex mMutex{"HandleWatcher::Impl"};
Data mData MOZ_GUARDED_BY(mMutex) = {};
// Callback from OS threadpool wait-object.
static void CALLBACK WaitCallback(PTP_CALLBACK_INSTANCE, void* ctx,
PTP_WAIT aWaitHandle,
TP_WAIT_RESULT aResult) {
static_cast<Impl*>(ctx)->OnWaitCompleted(aWaitHandle, aResult);
}
void OnWaitCompleted(PTP_WAIT aWaitHandle, TP_WAIT_RESULT aResult)
MOZ_EXCLUDES(mMutex) {
MOZ_ASSERT(aResult == WAIT_OBJECT_0);
mMutex.Lock();
// If this callback is no longer the active callback, skip out.
// All cleanup is someone else's problem.
if (aWaitHandle != mData.waitHandle.get()) {
MOZ_LOG(sHWLog, LogLevel::Debug,
("Recv'd already-stopped callback: HW %p | PTP_WAIT %p", this,
aWaitHandle));
mMutex.Unlock();
return;
}
// Take our self-pointer so that we release it on exit.
RefPtr<Impl> self = std::move(mData.self);
MOZ_LOG(sHWLog, LogLevel::Info,
("Recv'd callback: HW %p | handle %p | target %p | PTP_WAIT %p",
this, mData.handle, mData.target.get(), aWaitHandle));
// This may fail if (for example) `mData.target` is being shut down, but we
// have not yet received the shutdown callback.
mData.target->Dispatch(mData.runnable.forget());
Replace(Data{});
}
public:
static RefPtr<Impl> Create(HANDLE aHandle, nsIEventTarget* aTarget,
already_AddRefed<nsIRunnable> aRunnable) {
auto impl = MakeRefPtr<Impl>();
bool const ok [[maybe_unused]] =
impl->Watch(aHandle, aTarget, std::move(aRunnable));
MOZ_ASSERT(ok);
return impl;
}
private:
bool Watch(HANDLE aHandle, nsIEventTarget* aTarget,
already_AddRefed<nsIRunnable> aRunnable) MOZ_EXCLUDES(mMutex) {
MOZ_ASSERT(aHandle);
MOZ_ASSERT(aTarget);
RefPtr<nsIEventTarget> target(aTarget);
WaitHandlePtr waitHandle{
::CreateThreadpoolWait(&WaitCallback, this, nullptr)};
if (!waitHandle) {
return false;
}
{
mMutex.Lock();
nsresult const ret = aTarget->RegisterShutdownTask(this);
if (NS_FAILED(ret)) {
mMutex.Unlock();
return false;
}
MOZ_LOG(sHWLog, LogLevel::Info,
("Setting callback: HW %p | handle %p | target %p | PTP_WAIT %p",
this, aHandle, aTarget, waitHandle.get()));
// returns `void`; presumably always succeeds given a successful
// `::CreateThreadpoolWait()`
::SetThreadpoolWait(waitHandle.get(), aHandle, nullptr);
// After this point, you must call `FlushWaitHandle(waitHandle.get())`
// before destroying the wait handle. (Note that this must be done while
// *not* holding `mMutex`!)
Replace(Data{.handle = aHandle,
.target = std::move(target),
.runnable = aRunnable,
.waitHandle = std::move(waitHandle),
.self = this});
}
return true;
}
void TargetShutdown() MOZ_EXCLUDES(mMutex) override final {
mMutex.Lock();
MOZ_LOG(sHWLog, LogLevel::Debug,
("Target shutdown: HW %p | handle %p | target %p | PTP_WAIT %p",
this, mData.handle, mData.target.get(), mData.waitHandle.get()));
// Clear mData.target, since there's no need to unregister the shutdown task
// anymore. Hold onto it until we release the mutex, though, to avoid any
// reentrancy issues.
//
// This is more for internal consistency than safety: someone has to be
// shutting `target` down, and that someone isn't us, so there's necessarily
// another reference out there. (Although decrementing the refcount might
// still have arbitrary effects if someone's been excessively clever with
// nsISupports::Release...)
auto const oldTarget = std::move(mData.target);
Replace(Data{});
// (Static-assert that the mutex has indeed been released.)
([&]() MOZ_EXCLUDES(mMutex) {})();
}
public:
void Stop() MOZ_EXCLUDES(mMutex) {
mMutex.Lock();
Replace(Data{});
}
bool IsStopped() MOZ_EXCLUDES(mMutex) {
mozilla::MutexAutoLock lock(mMutex);
return !mData.handle;
}
private:
// Throughout this class, we use manual locking and unlocking guarded by
// Clang's thread-safety warnings, rather than scope-based lock-guards. This
// is largely driven by `Replace()`, below, which performs both operations
// which require the mutex to be held and operations which require it to not
// be held, and therefore must explicitly sequence the mutex release.
//
// These explicit locks, unlocks, and annotations are both alien to C++ and
// offensively tedious; but they _are_ still checked for state consistency at
// scope boundaries. (The concerned reader is invited to test this by
// deliberately removing an `mMutex.Unlock()` call from anywhere in the class
// and viewing the resultant compiler diagnostics.)
//
// A more principled, or at least differently-principled, implementation might
// create a scope-based lock-guard and pass it to `Replace()` to dispose of at
// the proper time. Alas, it cannot be communicated to Clang's thread-safety
// checker that such a guard is associated with `mMutex`.
//
void Replace(Data&& aData) MOZ_CAPABILITY_RELEASE(mMutex) {
// either both handles are NULL, or neither is
MOZ_ASSERT(!!aData.handle == !!aData.waitHandle);
if (mData.handle) {
MOZ_LOG(sHWLog, LogLevel::Info,
("Stop callback: HW %p | handle %p | target %p | PTP_WAIT %p",
this, mData.handle, mData.target.get(), mData.waitHandle.get()));
}
if (mData.target) {
mData.target->UnregisterShutdownTask(this);
}
// Extract the old data and insert the new -- but hold onto the old data for
// now. (See [1] and [2], below.)
Data oldData = std::exchange(mData, std::move(aData));
////////////////////////////////////////////////////////////////////////////
// Release the mutex.
mMutex.Unlock();
////////////////////////////////////////////////////////////////////////////
// [1] `oldData.self` will be unset if the old callback already ran (or if
// there was no old callback in the first place). If it's set, though, we
// need to explicitly clear out the wait-object first.
if (oldData.self) {
MOZ_ASSERT(oldData.waitHandle);
FlushWaitHandle(oldData.waitHandle.get());
}
// [2] oldData also includes several other reference-counted pointers. It's
// possible that these may be the last pointer to something, so releasing
// them may have arbitrary side-effects -- like calling this->Stop(), which
// will try to reacquire the mutex.
//
// Now that we've released the mutex, we can (implicitly) release them all
// here.
}
// Either confirm as complete or cancel any callbacks on aWaitHandle. Block
// until this is done. (See documentation for ::CloseThreadpoolWait().)
void FlushWaitHandle(PTP_WAIT aWaitHandle) MOZ_EXCLUDES(mMutex) {
::SetThreadpoolWait(aWaitHandle, nullptr, nullptr);
// This might block on `OnWaitCompleted()`, so we can't hold `mMutex` here.
::WaitForThreadpoolWaitCallbacks(aWaitHandle, TRUE);
// ::CloseThreadpoolWait() itself is the caller's responsibility.
}
};
NS_IMPL_ISUPPORTS(HandleWatcher::Impl, nsITargetShutdownTask)
//////
// HandleWatcher member function implementations
HandleWatcher::HandleWatcher() : mImpl{} {}
HandleWatcher::~HandleWatcher() {
if (mImpl) {
MOZ_ASSERT(mImpl->IsStopped());
mImpl->Stop(); // just in case, in release
}
}
HandleWatcher::HandleWatcher(HandleWatcher&&) noexcept = default;
HandleWatcher& HandleWatcher::operator=(HandleWatcher&&) noexcept = default;
void HandleWatcher::Watch(HANDLE aHandle, nsIEventTarget* aTarget,
already_AddRefed<nsIRunnable> aRunnable) {
auto impl = Impl::Create(aHandle, aTarget, std::move(aRunnable));
MOZ_ASSERT(impl);
if (mImpl) {
mImpl->Stop();
}
mImpl = std::move(impl);
}
void HandleWatcher::Stop() {
if (mImpl) {
mImpl->Stop();
}
}
bool HandleWatcher::IsStopped() { return !mImpl || mImpl->IsStopped(); }
} // namespace mozilla