Skip to content
Draft
Show file tree
Hide file tree
Changes from 8 commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
329 changes: 279 additions & 50 deletions src/windows/wslc/core/AsyncExecution.h
Original file line number Diff line number Diff line change
Expand Up @@ -9,92 +9,321 @@ Module Name:
Abstract:

Provides ForEachAsync, a generic helper for executing a work callback
over a collection concurrently in bounded batches using std::async.
over a collection concurrently using the Windows thread pool with bounded
concurrency and cooperative cancellation.

--*/
#pragma once

#include <algorithm>
#include <future>
#include <chrono>
#include <memory>
#include <optional>
#include <utility>
#include <vector>
#include <wil/resource.h>
#include <wil/result_macros.h>

namespace wsl::windows::wslc {

// Invokes onWork for each element in items concurrently, in batches of batchSize.
// Results are delivered serially to onSuccess. Errors are delivered serially to onError.
//
// This keeps wall time proportional to ceil(N / batchSize) rather than N for operations
// that have inherent per-item latency (e.g. network or IPC calls).
//
// Note: worker threads have no guaranteed per-thread initialization (e.g. COM). Callers
// whose onWork requires per-thread setup (such as CoInitializeEx) are responsible for
// performing it at the start of the onWork lambda.
//
// TWork : TItem -> TResult (called concurrently)
// TSuccess: TResult -> void (called serially)
// TError : (TItem, wil::ResultException) -> void (called serially)
template <typename TItem, typename TWork, typename TSuccess, typename TError>
void ForEachAsync(const std::vector<TItem>& items, TWork onWork, TSuccess onSuccess, TError onError, size_t batchSize = 10)
{
WI_ASSERT(batchSize > 0);
THROW_HR_IF(E_INVALIDARG, batchSize == 0);

using TResult = decltype(onWork(std::declval<TItem>()));
namespace detail {

struct BatchResult
template <typename TItem, typename TResult>
struct WorkerResult
{
explicit BatchResult(TItem capturedItem) : item(std::move(capturedItem))
WorkerResult() = default;

explicit WorkerResult(const TItem& item_) : item(item_)
{
}

TItem item;
std::optional<TItem> item;
std::optional<TResult> result;
wil::ResultException error{S_OK};
Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

To simplify things, I would recommend storing the error as a std::exception_ptr (null if no error was thrown).

This will have the benefit of allowing us to rethrow non-wil exceptions easily

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

(and we can get rid of hasError)

bool hasError{false};
};

for (size_t batchStart = 0; batchStart < items.size(); batchStart += batchSize)
// SharedContext holds state that must remain valid for the full lifetime of any running
// callback, including after WorkerPool is destroyed on the timeout path. Owned via
// shared_ptr and referenced by every SharedWorker.
template <typename TWork>
struct SharedContext
{
NON_COPYABLE(SharedContext);
NON_MOVABLE(SharedContext);

TWork onWork;
wil::unique_event cancelEvent;
Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I don't think this cancel event should be in the threadpool structure, since the threadpool logic doesn't actually look at it.

Callers can capture a cancel event in their work callback in they need to


explicit SharedContext(TWork onWork_) : onWork(std::move(onWork_))
{
cancelEvent.create(wil::EventOptions::ManualReset);
}
};

// Holds per-worker state. Each Launch heap-allocates a shared_ptr<SharedWorker> as the
// thread pool callback context, giving the callback shared ownership and ensuring this
// memory is not freed while a callback is still running.
template <typename TWork, typename TItem, typename TResult>
struct SharedWorker
{
WorkerResult<TItem, TResult> workerResult;
Comment thread
dkbennett marked this conversation as resolved.
std::shared_ptr<SharedContext<TWork>> context;
wil::unique_event done;
wil::unique_threadpool_work work;
};

// Manages a fixed pool of SharedWorkers.
template <typename TItem, typename TWork, typename TSuccess, typename TError>
struct WorkerPool
{
const size_t batchEnd = std::min(batchStart + batchSize, items.size());
NON_COPYABLE(WorkerPool);
NON_MOVABLE(WorkerPool);

using TResult = decltype(std::declval<TWork>()(std::declval<TItem>(), std::declval<HANDLE>()));
using TSharedWorker = SharedWorker<TWork, TItem, TResult>;
using TSharedContext = SharedContext<TWork>;

std::vector<std::shared_ptr<TSharedWorker>> workers;
std::vector<HANDLE> doneHandles;
Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

If we remove the "timeout on cancel" logic, we can get rid of those

std::shared_ptr<TSharedContext> context;
std::chrono::milliseconds timeout;
DWORD timeoutMs{};
Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

timeoutMs is unused

DWORD cancelDrainMs{};

WorkerPool(size_t workerCount, TWork onWork, std::chrono::milliseconds timeout_, std::chrono::milliseconds cancelDrainTimeout) :
context(std::make_shared<TSharedContext>(std::move(onWork))),
Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Assuming that we don't leak threads, context doesn't need to be a pointer here (could just be a regular class field)

timeout(timeout_),
timeoutMs(timeout_ == std::chrono::milliseconds::max() ? INFINITE : static_cast<DWORD>(timeout_.count())),
cancelDrainMs(static_cast<DWORD>(cancelDrainTimeout.count()))
Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Given that we have a Launch() method, I'd recommend only creating the workers there. That way this threadpool can dynamically resize itself as needed (which will allow us to reuse it in other places)

{
workers.reserve(workerCount);
doneHandles.reserve(workerCount);

for (size_t i = 0; i < workerCount; ++i)
{
auto worker = std::make_shared<TSharedWorker>();
worker->done.create(wil::EventOptions::ManualReset);

// Start signaled so that unstarted worker slots never block a
// WaitForMultipleObjects over the whole pool (e.g. in CancelAndDrainInFlight).
// Launch resets the event before submitting, ThreadPoolCallback sets it on exit.
worker->done.SetEvent();

worker->context = context;

doneHandles.push_back(worker->done.get());
workers.push_back(std::move(worker));
}
}

void Launch(size_t workerIndex, const TItem& item)
{
auto& worker = workers[workerIndex];

// Reset first so the slot appears busy. Re-signal on any failure so that
// CancelAndDrainInFlight does not block on a slot that was never submitted.
worker->done.ResetEvent();
auto signalOnFail = wil::scope_exit([&worker] { worker->done.SetEvent(); });

worker->workerResult.item.emplace(item);
worker->workerResult.result.reset();
worker->workerResult.error = wil::ResultException{S_OK};
worker->workerResult.hasError = false;

// Heap-allocate a shared_ptr as the callback context. The callback takes ownership,
// keeping the worker and its SharedContext alive for the full duration of the callback
// regardless of WorkerPool lifetime. Re-create the work item each launch so the
// context pointer is fresh.
auto* ctx = new std::shared_ptr<TSharedWorker>(worker);
Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Could we merge SharedContext and SharedWorker into one structure that the threadpool owns, and pass a pointer to that structure directly as the context ? This would simplify things a lot, and get rid of the shared_ptr pointer.

If we store them as an std::list, the pointers never get invalidated

worker->work.reset(::CreateThreadpoolWork(ThreadPoolCallback, ctx, nullptr));
if (!worker->work)
{
delete ctx;
THROW_LAST_ERROR();
}

::SubmitThreadpoolWork(worker->work.get());
signalOnFail.release();
}

void Drain(size_t workerIndex, TSuccess& onSuccess, TError& onError)
{
auto& worker = workers[workerIndex];

// Ensure the callback has fully returned before reading results.
::WaitForThreadpoolWorkCallbacks(worker->work.get(), FALSE);

if (worker->workerResult.hasError)
{
onError(*worker->workerResult.item, worker->workerResult.error);
}
else if (worker->workerResult.result.has_value())
{
onSuccess(*worker->workerResult.result);
}
}

std::vector<std::future<BatchResult>> futures;
futures.reserve(batchEnd - batchStart);
// Signals cancellation and waits up to cancelDrainMs for all in-flight workers to exit.
// Unstarted worker slots have their done events pre-signaled so they never block.
// Called noexcept from the scope_exit error guard in ForEachAsync.
void CancelAndDrainInFlight() noexcept
{
context->cancelEvent.SetEvent();
::WaitForMultipleObjects(static_cast<DWORD>(doneHandles.size()), doneHandles.data(), TRUE, cancelDrainMs);
Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I don't recommend having a "cancel timeout" here. If this timeout ever hits, this will leak threads which will most likely to undefined behavior / crashes

}

for (size_t i = batchStart; i < batchEnd; ++i)
// Signals cancellation, waits up to cancelDrainMs for workers to exit, then throws ERROR_TIMEOUT.
// Workers that do not exit within cancelDrainMs are abandoned. Each running callback holds a
// shared_ptr to its SharedWorker and SharedContext, so neither is freed while the callback runs.
// onWork implementations must check the cancel event at natural checkpoints and exit promptly.
//
// Note: TerminateThread() is not used - it skips C++ destructors, leaves user-mode locks
// permanently held (causing deadlocks), and corrupts COM apartment state.
[[noreturn]] void CancelAndThrow(size_t remainingItems)
{
const auto& item = items[i];
futures.push_back(std::async(std::launch::async, [&onWork, item]() -> BatchResult {
BatchResult result{item};
try
{
result.result = onWork(item);
}
catch (const wil::ResultException& ex)
{
result.hasError = true;
result.error = ex;
}
return result;
}));
context->cancelEvent.SetEvent();

::WaitForMultipleObjects(static_cast<DWORD>(doneHandles.size()), doneHandles.data(), TRUE, cancelDrainMs);
Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

nit: We should check the return code of WaitForMultipleObjects() here


THROW_HR_MSG(
HRESULT_FROM_WIN32(ERROR_TIMEOUT),
"ForEachAsync: worker exceeded timeout of %lld ms (%zu items remaining).",
static_cast<long long>(timeout.count()),
remainingItems);
}

for (auto& future : futures)
// Thread pool callback - invoked on a pool thread for each submitted work item.
// Takes ownership of the heap-allocated shared_ptr<TSharedWorker> passed as context,
// ensuring the worker and its SharedContext remain alive for the duration of this call.
static void CALLBACK ThreadPoolCallback(PTP_CALLBACK_INSTANCE, void* context, PTP_WORK) noexcept
{
auto batchResult = future.get();
const std::unique_ptr<std::shared_ptr<TSharedWorker>> owner(static_cast<std::shared_ptr<TSharedWorker>*>(context));
auto& worker = **owner;

if (batchResult.hasError)
try
{
worker.workerResult.result.emplace(worker.context->onWork(*worker.workerResult.item, worker.context->cancelEvent.get()));
}
catch (const wil::ResultException& ex)
{
onError(batchResult.item, batchResult.error);
worker.workerResult.hasError = true;
worker.workerResult.error = ex;
}
Comment thread
dkbennett marked this conversation as resolved.
else if (batchResult.result.has_value())
catch (...)
{
onSuccess(*batchResult.result);
worker.workerResult.hasError = true;
worker.workerResult.error = wil::ResultException{wil::ResultFromCaughtException()};
Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This catch should cover both catch() cases

}

worker.done.SetEvent();
}
};

} // namespace detail

// Invokes onWork for each element in items concurrently using the Windows thread pool,
// with concurrency bounded to poolSize. Results are delivered serially to onSuccess.
// Errors are delivered serially to onError.
//
// onWork receives a HANDLE to a cancellation event and should check it at natural
// checkpoints using WaitForSingleObject(cancel, 0), returning early if it is set.
// On timeout, the event is signalled and ForEachAsync waits up to cancelDrainTimeout
// for workers to exit before throwing HRESULT_FROM_WIN32(ERROR_TIMEOUT).
//
// If onSuccess, onError, or Launch throw, the cancel event is signalled and ForEachAsync
// waits up to cancelDrainTimeout for any in-flight workers to exit before rethrowing.
// This guarantees no background thread pool callbacks outlive the ForEachAsync call.
Comment thread
dkbennett marked this conversation as resolved.
Outdated
//
// poolSize must not exceed MAXIMUM_WAIT_OBJECTS (64).
//
// The timeout is a safety net against indefinite hangs, not a strict per-worker limit.
// A worker that hangs while other workers are still completing will be caught in the
// final wait at most one full timeout after all other work has finished.
//
// Note: thread pool threads have no guaranteed per-thread initialization. Callers
// whose onWork requires per-thread setup (e.g. CoInitializeEx) must perform it at
// the start of the onWork lambda.
//
// TWork : (TItem, HANDLE cancelEvent) -> TResult (called concurrently)
// TSuccess: TResult -> void (called serially)
// TError : (TItem, wil::ResultException) -> void (called serially)
template <typename TItem, typename TWork, typename TSuccess, typename TError>
void ForEachAsync(
const std::vector<TItem>& items,
TWork onWork,
TSuccess onSuccess,
TError onError,
size_t poolSize = 10,
std::chrono::milliseconds timeout = std::chrono::milliseconds::max(),
std::chrono::milliseconds cancelDrainTimeout = std::chrono::seconds(5))
{
THROW_HR_IF(E_INVALIDARG, poolSize == 0);
THROW_HR_IF(E_INVALIDARG, poolSize > MAXIMUM_WAIT_OBJECTS);
Comment thread
dkbennett marked this conversation as resolved.
THROW_HR_IF(E_INVALIDARG, timeout.count() < 0);
THROW_HR_IF(E_INVALIDARG, cancelDrainTimeout.count() < 0);

// INFINITE (0xFFFFFFFF) is reserved as the Win32 sentinel; values at or above it
// would either alias INFINITE or wrap, producing the opposite timeout behavior.
constexpr long long c_maxWaitMs = 0xFFFFFFFELL; // INFINITE - 1
THROW_HR_IF(E_INVALIDARG, timeout != std::chrono::milliseconds::max() && timeout.count() > c_maxWaitMs);
THROW_HR_IF(E_INVALIDARG, cancelDrainTimeout.count() > c_maxWaitMs);

if (items.empty())
{
return;
}

const size_t workerCount = std::min(poolSize, items.size());

detail::WorkerPool<TItem, TWork, TSuccess, TError> pool{workerCount, std::move(onWork), timeout, cancelDrainTimeout};

// On any exception from Launch, Drain, or the user callbacks (onSuccess/onError),
// signal cancellation and wait for in-flight workers before rethrowing. This guarantees
// no background thread pool callbacks outlive the ForEachAsync call.
auto cancelOnError = wil::scope_exit([&pool] { pool.CancelAndDrainInFlight(); });
Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I recommend doing this in the WorkerPool's destuctor so there's a strong guarantee of no leaked threads


// Fill the pool - submit one item per worker to saturate all workers immediately.
size_t nextItem = 0;
for (; nextItem < workerCount; ++nextItem)
{
pool.Launch(nextItem, items[nextItem]);
}

// Keep the pool full - as each worker completes, drain its result and immediately
// assign it the next pending item. WaitForMultipleObjects(FALSE) wakes on the first
// completion, so no worker idles while work remains.
while (nextItem < items.size())
{
const DWORD waitResult = ::WaitForMultipleObjects(static_cast<DWORD>(workerCount), pool.doneHandles.data(), FALSE, pool.timeoutMs);

if (waitResult == WAIT_TIMEOUT)
{
pool.CancelAndThrow(items.size() - nextItem);
Comment thread
dkbennett marked this conversation as resolved.
}

THROW_LAST_ERROR_IF(waitResult == WAIT_FAILED);

const size_t workerIndex = waitResult - WAIT_OBJECT_0;
pool.Drain(workerIndex, onSuccess, onError);
pool.Launch(workerIndex, items[nextItem++]);
Comment thread
dkbennett marked this conversation as resolved.
}

const DWORD finalWait = ::WaitForMultipleObjects(static_cast<DWORD>(workerCount), pool.doneHandles.data(), TRUE, pool.timeoutMs);

if (finalWait == WAIT_TIMEOUT)
{
pool.CancelAndThrow(0);
}

THROW_LAST_ERROR_IF(finalWait == WAIT_FAILED);

for (size_t i = 0; i < workerCount; ++i)
{
pool.Drain(i, onSuccess, onError);
}

cancelOnError.release();
}

} // namespace wsl::windows::wslc
Loading