Compare commits

...

39 Commits

Author SHA1 Message Date
Pratik Mankawde
bacfb38233 Merge branch 'pratik/std-coro/migrate-test-code' into pratik/std-coro/cleanup-boost-coroutine 2026-08-19 17:47:42 +01:00
Pratik Mankawde
bec82ee704 Merge branch 'pratik/std-coro/migrate-entry-points' into pratik/std-coro/migrate-test-code 2026-08-19 17:47:41 +01:00
Pratik Mankawde
c6224c48ed Merge branch 'pratik/std-coro/add-coroutine-primitives' into pratik/std-coro/migrate-entry-points
Bring in develop via the upstream branch. Two classes of collision needed
manual resolution:

Namespace case refactor (develop #7933) renamed RPC -> rpc and
xrpl::RPC::Tuning -> xrpl::rpc::tuning. Conflicts were resolved by taking
develop's naming together with this stack's coroutine API (postCoroTask,
CoroTask<void>, no Context::coro):
  - src/test/app/Path_test.cpp
  - src/test/app/PathMPT_test.cpp
  - src/test/jtx/impl/TestHelpers.cpp
  - src/xrpld/rpc/handlers/orderbook/RipplePathFind.cpp (no textual
    conflict; stale spellings carried forward on stack-authored lines)

New old-API caller: develop added postCoro and Context::coro uses in
src/test/app/PayChan_test.cpp, which this stack does not otherwise touch,
so it merged without conflict but would not compile once Context::coro is
removed here. Both call sites are migrated to postCoroTask and the
.coro = {} designated initialisers dropped. The kFeeHeavyBurdenRpc
assertions that motivated develop's change are preserved unchanged.
2026-08-19 17:47:30 +01:00
Pratik Mankawde
5fb1429e3a Merge remote-tracking branch 'origin/develop' into pratik/std-coro/add-coroutine-primitives
# Conflicts:
#	.cspell.config.yaml
2026-08-19 17:41:56 +01:00
Pratik Mankawde
b54c500bb2 Merge branch 'pratik/std-coro/migrate-test-code' into pratik/std-coro/cleanup-boost-coroutine 2026-07-27 13:48:08 +01:00
Pratik Mankawde
b6526522fe Merge branch 'pratik/std-coro/migrate-entry-points' into pratik/std-coro/migrate-test-code 2026-07-27 13:48:08 +01:00
Pratik Mankawde
61041155d2 Merge branch 'pratik/std-coro/add-coroutine-primitives' into pratik/std-coro/migrate-entry-points 2026-07-27 13:48:08 +01:00
Pratik Mankawde
5ea72b45f1 Merge remote-tracking branch 'origin/develop' into pratik/std-coro/add-coroutine-primitives 2026-07-27 13:47:54 +01:00
Pratik Mankawde
7b95a8ea8b Merge branch 'pratik/std-coro/migrate-test-code' into pratik/std-coro/cleanup-boost-coroutine 2026-07-27 13:06:43 +01:00
Pratik Mankawde
7385004952 Move coroutine-capture NOLINTNEXTLINE comments next to the lambdas
clang-format placed the suppression comments two lines above the lambda
expressions (before the postCoroTask call), so NOLINTNEXTLINE suppressed
the wrong line and clang-tidy still reported
cppcoreguidelines-avoid-capturing-lambda-coroutines at the lambda. Move
the comments inside the argument list, immediately before each lambda,
matching the pattern already used in threadSpecificStorage.
2026-07-27 13:06:13 +01:00
Pratik Mankawde
fce702d5c3 Merge branch 'pratik/std-coro/migrate-test-code' into pratik/std-coro/cleanup-boost-coroutine 2026-07-27 12:36:46 +01:00
Pratik Mankawde
a6fea1227f Suppress clang-tidy coroutine-capture warnings in core tests
The JobQueue_test and Coroutine_test coroutine lambdas capture pointers
to test-scope locals, but each test drives the coroutine to completion
(or, for the stopped-queue case, guarantees it never starts) before the
locals go out of scope. Add NOLINTNEXTLINE for
cppcoreguidelines-avoid-capturing-lambda-coroutines with comments
explaining the lifetime guarantee, and drop the unused <memory> include
flagged by misc-include-cleaner.
2026-07-27 12:36:10 +01:00
Pratik Mankawde
968f9e1bf3 Merge branch 'pratik/std-coro/migrate-entry-points' into pratik/std-coro/migrate-test-code 2026-07-27 12:34:16 +01:00
Pratik Mankawde
70813b336a Suppress clang-tidy coroutine-capture warnings in pathfinding tests
The Path_test and TestHelpers coroutine lambdas capture locals by
reference, but the caller blocks on Gate::waitFor() until the coroutine
signals completion, so the captures cannot dangle. Add NOLINTNEXTLINE
for cppcoreguidelines-avoid-capturing-lambda-coroutines with a comment
explaining the lifetime guarantee.
2026-07-27 12:33:49 +01:00
Pratik Mankawde
47d84d2e21 Merge branch 'pratik/std-coro/migrate-test-code' into pratik/std-coro/cleanup-boost-coroutine 2026-07-27 12:06:41 +01:00
Pratik Mankawde
ab8350b558 Merge branch 'pratik/std-coro/migrate-entry-points' into pratik/std-coro/migrate-test-code 2026-07-27 12:06:41 +01:00
Pratik Mankawde
072e11ecd1 Address clang-tidy findings in entry-point migration
- Suppress cppcoreguidelines-avoid-capturing-lambda-coroutines with
  NOLINT where lifetime is guaranteed: GRPCServer (thisShared keeps
  CallData alive), ServerHandler RPC/WS clients (captures by value,
  handler outlives JobQueue jobs), and PathMPT_test (test blocks on
  Gate::waitFor until the coroutine completes).
- Change ServerHandler::processRequest to take Output const& to fix
  cppcoreguidelines-rvalue-reference-param-not-moved.
- Replace std::lock_guard with std::scoped_lock in RipplePathFind.
- Remove redundant includes and default member initializer in
  Context.h, GRPCServer.h, RipplePathFind.cpp.
2026-07-27 12:06:25 +01:00
Pratik Mankawde
3ac48e297f Merge branch 'pratik/std-coro/migrate-test-code' into pratik/std-coro/cleanup-boost-coroutine 2026-07-27 11:30:34 +01:00
Pratik Mankawde
f669b70451 Merge branch 'pratik/std-coro/migrate-entry-points' into pratik/std-coro/migrate-test-code 2026-07-27 11:30:34 +01:00
Pratik Mankawde
62839b1531 Merge branch 'pratik/std-coro/add-coroutine-primitives' into pratik/std-coro/migrate-entry-points 2026-07-27 11:30:34 +01:00
Pratik Mankawde
9bddc54722 Fix clang-tidy violations in coroutine primitives
- NOLINT the compiler-mandated coroutine protocol names (promise_type,
  await_ready, ...) that conflict with readability-identifier-naming and
  readability-convert-member-functions-to-static
- Add [[nodiscard]] to handle(), done(), await_ready()
- Replace std::lock_guard with std::scoped_lock const
- Pass CoroTaskRunner name by value and std::move it; default-init
  runCount_ in-class
- Drop unused <coroutine> include from JobQueue.h; include
  instrumentation.h directly in JobQueueAwaiter.h
- CoroTask_test: kN constant naming, const locals, file-level NOLINT for
  cppcoreguidelines-avoid-capturing-lambda-coroutines (lifetimes are
  gated and joined)
2026-07-27 11:30:16 +01:00
Pratik Mankawde
598bda0ebc Merge branch 'pratik/std-coro/migrate-test-code' into pratik/std-coro/cleanup-boost-coroutine 2026-07-27 10:30:16 +01:00
Pratik Mankawde
25b9b4d27d Fix JobQueue_test resume() contract violation and Coroutine_test timeout hazards
- PostCoroTest2 called runner->resume() directly from the test thread,
  violating the documented precondition on CoroTaskRunner::resume() (a
  post() must precede every resume) and driving runCount_ negative.
  Rewrite the loop as post()+join(); the non-atomic yieldCount now also
  verifies the happens-before edge join() provides via mutexRun_.
- PostCoroTest3 wrote false into an already-false flag, so it could
  not detect the coroutine running after stop(). Write true and assert
  the flag stays false.
- Coroutine_test: early-return when a Gate waitFor() times out instead
  of falling through to c->join()/a[i]->join(), which would deref a
  null shared_ptr (correctOrder, threadSpecificStorage first loop) or
  block indefinitely (second loop).
2026-07-27 10:29:57 +01:00
Pratik Mankawde
22ed85c3ad Merge branch 'pratik/std-coro/migrate-entry-points' into pratik/std-coro/migrate-test-code 2026-07-27 10:26:00 +01:00
Pratik Mankawde
8c7a27d721 Name the ripple_path_find wait timeout and fire completion on exception
- Replace the magic 30s in doRipplePathFind with
  RPC::Tuning::kPathfindCompletionTimeout.
- In PathRequestManager::updateAll, invoke updateComplete() via a
  ScopeExit guard on the one-shot (hasCompletion) path so the blocked
  RPC handler is released immediately even if doUpdate throws, instead
  of waiting out the full timeout.
2026-07-25 15:05:08 +01:00
Pratik Mankawde
9266c23ef6 Bound concurrent blocking ripple_path_find calls to prevent worker starvation
The no-ledger branch of doRipplePathFind parks its JobQueue worker on a
condition_variable for up to 30s, waiting for a completion that is fired
from a JtUpdatePf job -- which itself needs a free worker to run. With
unbounded JtClientRpc concurrency, enough simultaneous ripple_path_find
calls could park every worker, stalling the entire JobQueue until the
30s timeouts expired. The old Boost.Coroutine implementation suspended
and released the worker, so it did not have this failure mode.

Reuse the RPC::LegacyPathFind guard (already applied on the
ledger-specified branch) to admit at most kMaxPathfindsInProgress
non-admin blocking requests, returning rpcTOO_BUSY beyond that. The
guard stays in scope across the wait.
2026-07-25 12:39:04 +01:00
Pratik Mankawde
5e9ce16de6 Merge branch 'pratik/std-coro/add-coroutine-primitives' into pratik/std-coro/migrate-entry-points 2026-07-25 12:37:14 +01:00
Pratik Mankawde
8d4ea00453 Strengthen CoroTask_test assertions and timeout handling
- testExceptionPropagation now co_awaits an inner CoroTask<void> that
  throws and asserts the rethrown message, covering the
  CoroTask<void>::await_resume rethrow path; the old version could not
  distinguish a throw from a normal return.
- testJobQueueAwaiter now actually awaits the JobQueueAwaiter struct
  (single use per coroutine, per the GCC-12 note) and asserts the full
  ordered step sequence instead of only the terminal value.
- testValueException asserts the caught exception's message.
- All waitFor() timeouts early-return on failure instead of falling
  through to join()/state reads (null-deref and TSAN-race hazards on
  timeout, plus the nSuspend_ assert in ~Env).
- Remove dead shared_ptr locals in testCorrectOrder/testMultipleYields.
- Add missing includes (<array>, <stdexcept>, <string>, <vector>,
  LocalValue.h, CoroTask.h).
2026-07-25 12:36:30 +01:00
Pratik Mankawde
bdcf094171 Guard resume() against completed coroutines and log unhandled exceptions
- resume() now skips the handle resume when the task is null or done
  (duplicate external post() after completion), matching the old
  Coro::resume() 'if (coro_)' guard instead of invoking UB in release
  builds. The runCount_ bookkeeping still runs to balance post().
- Exceptions escaping a top-level coroutine body were captured by
  unhandled_exception() and destroyed unobserved with the frame; they
  are now logged at error level before the frame is released.
- Document that join() may return via the finished_ disjunct while the
  final resume() is still completing its bookkeeping.
2026-07-25 12:33:51 +01:00
Pratik Mankawde
a42e8174d9 Fix check-rename CI failure
- Remove BoostToStdCoroutineSwitchPlan.md and BoostToStdCoroutineTaskList.md
  working documents from the repo root; the rename script rewrites their
  'rippled' references and the job fails on the resulting dirty tree.
- Reword two CoroTask.h comments to use 'xrpld'.
- Drop cspell words (cppcoro, gantt, Pratik, Mankawde) that existed only
  for the removed documents.
2026-07-25 12:30:38 +01:00
Pratik Mankawde
791c061a98 Merge branch 'pratik/std-coro/migrate-test-code' into pratik/std-coro/cleanup-boost-coroutine
Forward-merge the test-code migration (carrying the latest develop) into the
Boost.Coroutine cleanup.

Conflict resolutions: keep this branch's removal of JobQueue::Coro,
CoroCreateT, postCoro, Coro.ipp and the Boost.Coroutine2 includes, while
taking develop's include set, Doxygen comment style and formatting for the
retained CoroTaskRunner / postCoroTask declarations.
2026-07-24 21:02:25 +01:00
Pratik Mankawde
91004209e9 Merge branch 'pratik/std-coro/migrate-entry-points' into pratik/std-coro/migrate-test-code
Forward-merge the entry-point migration (carrying the latest develop) into
the test-code migration.

Conflict resolutions in Coroutine_test.cpp and JobQueue_test.cpp: keep the
postCoroTask / CoroTaskRunner form from this branch, with develop's renames
(JtClient, Gate::waitFor, kN) and develop's initialization of the
unprotected flag.
2026-07-24 20:57:01 +01:00
Pratik Mankawde
4c18dd867c Merge branch 'pratik/std-coro/add-coroutine-primitives' into pratik/std-coro/migrate-entry-points
Forward-merge the updated coroutine primitives (which themselves carry the
latest develop) into the entry-point migration.

Conflict resolutions:
- .cspell.config.yaml: keep develop's dotfile name, merge word lists.
- Context.h: drop the coro member; infoSub gets a default initializer.
- Application.cpp, ServerHandler.cpp, TestHelpers.cpp, Path_test.cpp,
  PathMPT_test.cpp: adopt develop's designated-initializer JsonContext and
  renamed identifiers (JtClient, kApiVersionIfUnspecified, kMaxSrcCur, Gate).
- GRPCServer: process(coro) becomes processRequest(), posted with
  postCoroTask.
- RipplePathFind.cpp: follow develop's move to handlers/orderbook/ and
  PathRequestManager; the completion state shared with the path-finding
  continuation is now heap-allocated so it outlives an early return.
- AMMTest.cpp: develop moved find_paths_request into TestHelpers.cpp, so the
  local copy is dropped and the migration applied there instead.
2026-07-24 20:52:28 +01:00
Pratik Mankawde
fb6ece56f2 Merge remote-tracking branch 'origin/develop' into pratik/std-coro/add-coroutine-primitives
Resolves the develop rename of cspell.config.yaml and the JobQueue.h
conflicts, and aligns the new coroutine primitives with develop naming
(CreateT, mutex_/mutexRun_, JtClient, forceMultiThread,
beast::unit_test::Suite).

Also addresses two review findings:

- postCoroTask now holds a jobCounter_ reservation for the whole
  function. JobQueue::stop() joins jobCounter_ before asserting
  nSuspend_ == 0, so the reservation closes the window between the
  ++nSuspend_ and the balancing post()/expectEarlyExit(), and doubles
  as the shutdown check.

- YieldPostAwaiter::await_suspend returns a coroutine_handle<>
  (symmetric transfer) instead of resuming inline. This keeps a
  yield loop against a stopping JobQueue from growing the stack
  without bound and avoids touching a frame that resume() may have
  already destroyed.
2026-07-24 20:39:41 +01:00
Pratik Mankawde
b53df32334 feat: Remove Boost.Coroutine dependency and old Coro API
Remove the legacy Boost.Coroutine infrastructure now that all callers
have been migrated to C++20 std::coroutine:

- Delete Coro.ipp: Remove the Boost.Coroutine-based Coro class and
  its suspend/resume/post/runOnJobQueue helpers.
- Remove Coro API from JobQueue.h: Remove postCoro(), Coro class
  declaration, and all Boost.Coroutine-related methods.
- Remove Boost.Coroutine from build: Drop coroutine and context from
  Boost components in cmake and conanfile.py.
- Remove BOOST_COROUTINES_NO_DEPRECATION_WARNING compile definition.
2026-03-25 15:48:17 +00:00
Pratik Mankawde
42cced50fb feat: Migrate coroutine tests from Boost.Coroutine to C++20 coroutines
Migrate Coroutine_test and JobQueue_test from Boost.Coroutine to
C++20 std::coroutine using CoroTask/CoroTaskRunner:

- Coroutine_test: Replace Coro-based coroutine tests with CoroTask
  equivalents using co_await runner->yieldAndPost().
- JobQueue_test: Replace Coro suspend/resume patterns with CoroTask
  equivalents, use pointer-by-value captures in coroutine lambdas
  to avoid dangling reference issues.
2026-03-25 15:48:17 +00:00
Pratik Mankawde
0e815aa1ac feat: Migrate production entry points from Boost.Coroutine to C++20 coroutines
Migrate all production coroutine entry points from Boost.Coroutine
to C++20 std::coroutine using the CoroTask/CoroTaskRunner primitives:

- RipplePathFind: Replace Coro suspend/resume with co_await pattern,
  add cv timeout for graceful shutdown.
- ServerHandler: Replace Coro-based processRequest with CoroTask,
  simplify coroutine lifecycle management.
- GRPCServer: Replace Coro with CoroTask for streaming RPC handlers.
- Remove Coro usage from Context.h aggregate initialization.
- Add exception handling in coroutine bodies to prevent unhandled
  exceptions from escaping the coroutine frame.
2026-03-25 15:48:10 +00:00
Pratik Mankawde
21149a81e3 feat: Add C++20 coroutine primitives: CoroTask, CoroTaskRunner, JobQueueAwaiter
Add C++20 std::coroutine based task primitives for the JobQueue:

- CoroTask<T>: A coroutine return type with RAII ownership semantics
  and symmetric transfer for efficient resumption.
- CoroTaskRunner: Manages coroutine lifecycle on the JobQueue with
  suspend/resume tracking, LocalValue preservation, and graceful
  shutdown support.
- JobQueueAwaiter: External awaiter combining yield+post atomically.
- yieldAndPost(): Inline awaiter workaround for GCC-12 codegen bug
  where external awaiters at multiple co_await points corrupt the
  coroutine state machine resume index.
- CoroTask_test: Comprehensive test suite covering task lifecycle,
  suspend/resume, shutdown, and value-returning coroutines.
- BoostToStdCoroutineSwitchPlan.md: Migration plan documentation.
2026-03-25 15:46:44 +00:00
Pratik Mankawde
b78202a99a docs: Add Boost to C++20 coroutine migration plan
Comprehensive migration plan documenting the switch from
Boost.Coroutine2 to C++20 standard coroutines in rippled, including
research analysis, implementation phases, risk assessment, and
testing strategy.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-03-25 15:44:19 +00:00
24 changed files with 2723 additions and 512 deletions

View File

@@ -256,6 +256,8 @@ words:
- replayer
- repodata
- repomd
- repost
- reposts
- rerandomize
- rerandomization
- rerandomized
@@ -303,6 +305,7 @@ words:
- sponsees
- SRPMS
- sslws
- stackful
- statsd
- STATSDCOLLECTOR
- stissue

View File

@@ -39,21 +39,12 @@ if(Boost_COMPILER)
target_link_libraries(xrpl_boost INTERFACE Boost::disable_autolinking)
endif()
# GCC 14+ has a false positive -Wuninitialized warning in Boost.Coroutine2's
# state.hpp when compiled with -O3. This is due to GCC's intentional behavior
# change (Bug #98871, #119388) where warnings from inlined system header code
# are no longer suppressed by -isystem. The warning occurs in operator|= in
# boost/coroutine2/detail/state.hpp when inlined from push_control_block::destroy().
# See: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119388
if(is_gcc AND CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 14)
target_compile_options(xrpl_boost INTERFACE -Wno-uninitialized)
endif()
# Boost.Context's ucontext backend has ASAN fiber-switching annotations
# (start/finish_switch_fiber) that are compiled in when BOOST_USE_ASAN is defined.
# This tells ASAN about coroutine stack switches, preventing false positive
# stack-use-after-scope errors. BOOST_USE_UCONTEXT ensures the ucontext backend
# is selected (fcontext does not support ASAN annotations).
# This tells ASAN about fiber stack switches used by boost::asio::spawn,
# preventing false positive stack-use-after-scope errors.
# BOOST_USE_UCONTEXT ensures the ucontext backend is selected (fcontext does
# not support ASAN annotations).
# These defines must match what Boost was compiled with (see conan/profiles/sanitizers).
if(enable_asan)
target_compile_definitions(

View File

@@ -62,7 +62,7 @@ class Xrpl(ConanFile):
"boost/*:without_cobalt": True,
"boost/*:without_context": False,
"boost/*:without_coroutine": True,
"boost/*:without_coroutine2": False,
"boost/*:without_coroutine2": True,
"date/*:header_only": True,
"ed25519/*:shared": False,
"grpc/*:shared": False,

View File

@@ -1,136 +0,0 @@
#pragma once
#include <utility>
namespace xrpl {
/**
* Coroutine stack size (1.5 MB). Increased from 1 MB because
* ASAN-instrumented deep call stacks exceeded the original limit.
*/
constexpr std::size_t kCoroStackSize = 1536 * 1024;
template <class F>
JobQueue::Coro::Coro(CoroCreateT, JobQueue& jq, JobType type, std::string name, F&& f)
: jq_(jq)
, type_(type)
, name_(std::move(name))
, coro_(
boost::context::protected_fixedsize_stack(kCoroStackSize),
[this, fn = std::forward<F>(f)](boost::coroutines2::coroutine<void>::push_type& doYield) {
yield_ = &doYield;
yield();
fn(shared_from_this());
#ifndef NDEBUG
finished_ = true;
#endif
})
{
}
inline JobQueue::Coro::~Coro()
{
#ifndef NDEBUG
XRPL_ASSERT(finished_, "xrpl::JobQueue::Coro::~Coro : is finished");
#endif
}
inline void
JobQueue::Coro::yield() const
{
{
std::scoped_lock const lock(jq_.mutex_);
++jq_.nSuspend_;
}
(*yield_)();
}
inline bool
JobQueue::Coro::post()
{
{
std::scoped_lock const lk(mutexRun_);
running_ = true;
}
// sp keeps 'this' alive
if (jq_.addJob(type_, name_, [this, sp = shared_from_this()]() { resume(); }))
{
return true;
}
// The coroutine will not run. Clean up running_.
std::scoped_lock const lk(mutexRun_);
running_ = false;
cv_.notify_all();
return false;
}
inline void
JobQueue::Coro::resume()
{
{
std::scoped_lock const lk(mutexRun_);
running_ = true;
}
{
std::scoped_lock const lk(jq_.mutex_);
--jq_.nSuspend_;
}
auto saved = detail::getLocalValues().release();
detail::getLocalValues().reset(&lvs_);
std::scoped_lock const lock(mutex_);
// A late resume() can arrive after the coroutine has already completed.
// This is an expected (if rare) outcome of the race condition documented
// in JobQueue.h:354-377 where post() schedules a resume job before the
// coroutine yields — the mutex serializes access, but by the time this
// resume() acquires the lock the coroutine may have already run to
// completion. Calling operator() on a completed boost::coroutine2 is
// undefined behavior, so we must check and skip invoking the coroutine
// body if it has already completed.
if (coro_)
{
coro_();
}
detail::getLocalValues().release();
detail::getLocalValues().reset(saved);
std::scoped_lock const lk(mutexRun_);
running_ = false;
cv_.notify_all();
}
inline bool
JobQueue::Coro::runnable() const
{
return static_cast<bool>(coro_);
}
inline void
JobQueue::Coro::expectEarlyExit()
{
#ifndef NDEBUG
if (!finished_)
#endif
{
// expectEarlyExit() must only ever be called from outside the
// Coro's stack. It you're inside the stack you can simply return
// and be done.
//
// That said, since we're outside the Coro's stack, we need to
// decrement the nSuspend that the Coro's call to yield caused.
std::scoped_lock const lock(jq_.mutex_);
--jq_.nSuspend_;
#ifndef NDEBUG
finished_ = true;
#endif
}
}
inline void
JobQueue::Coro::join()
{
std::unique_lock<std::mutex> lk(mutexRun_);
cv_.wait(lk, [this]() { return !running_; });
}
} // namespace xrpl

View File

@@ -0,0 +1,713 @@
#pragma once
#include <xrpl/beast/utility/instrumentation.h>
#include <coroutine>
#include <exception>
#include <type_traits>
#include <utility>
#include <variant>
namespace xrpl {
template <typename T = void>
class CoroTask;
/**
* CoroTask<void> -- coroutine return type for void-returning coroutines.
*
* Class / Dependency Diagram
* ==========================
*
* CoroTask<void>
* +-----------------------------------------------+
* | - handle_ : Handle (coroutine_handle<promise>) |
* +-----------------------------------------------+
* | + handle(), done() |
* | + await_ready/suspend/resume (Awaiter iface) |
* +-----------------------------------------------+
* | owns
* v
* promise_type
* +-----------------------------------------------+
* | - exception_ : std::exception_ptr |
* | - continuation_ : std::coroutine_handle<> |
* +-----------------------------------------------+
* | + get_return_object() -> CoroTask |
* | + initial_suspend() -> suspend_always (lazy) |
* | + final_suspend() -> FinalAwaiter |
* | + return_void() |
* | + unhandled_exception() |
* +-----------------------------------------------+
* | returns at final_suspend
* v
* FinalAwaiter
* +-----------------------------------------------+
* | await_suspend(h): |
* | if continuation_ set -> symmetric transfer |
* | else -> noop_coroutine |
* +-----------------------------------------------+
*
* Design Notes
* ------------
* - Lazy start: initial_suspend returns suspend_always, so the coroutine
* body does not execute until the handle is explicitly resumed.
* - Symmetric transfer: await_suspend returns a coroutine_handle instead
* of void/bool, allowing the scheduler to jump directly to the next
* coroutine without growing the call stack.
* - Continuation chaining: when one CoroTask is co_await-ed inside
* another, the caller's handle is stored as continuation_ so
* FinalAwaiter can resume it when this task finishes.
* - Move-only: the handle is exclusively owned; copy is deleted.
*
* Usage Examples
* ==============
*
* 1. Basic void coroutine (the most common case in xrpld):
*
* CoroTask<void> doWork(std::shared_ptr<CoroTaskRunner> runner) {
* // do something
* co_await runner->suspend(); // yield control
* // resumed later via runner->post() or runner->resume()
* co_return;
* }
*
* 2. co_await-ing one CoroTask<void> from another (chaining):
*
* CoroTask<void> inner() {
* // ...
* co_return;
* }
* CoroTask<void> outer() {
* co_await inner(); // continuation_ links outer -> inner
* co_return; // FinalAwaiter resumes outer
* }
*
* 3. Exceptions propagate through co_await:
*
* CoroTask<void> failing() {
* throw std::runtime_error("oops");
* co_return;
* }
* CoroTask<void> caller() {
* try { co_await failing(); }
* catch (std::runtime_error const&) { // caught here }
* }
*
* Caveats / Pitfalls
* ==================
*
* BUG-RISK: Dangling references in coroutine parameters.
* Coroutine parameters are copied into the frame, but references
* are NOT -- they are stored as-is. If the referent goes out of scope
* before the coroutine finishes, you get use-after-free.
*
* // BROKEN -- local dies before coroutine runs:
* CoroTask<void> bad(int& ref) { co_return; }
* void launch() {
* int local = 42;
* auto task = bad(local); // frame stores &local
* } // local destroyed; frame holds dangling ref
*
* // FIX -- pass by value, or ensure lifetime via shared_ptr.
*
* BUG-RISK: GCC 14 corrupts reference captures in coroutine lambdas.
* When a lambda that returns CoroTask captures by reference ([&]),
* GCC 14 may generate a corrupted coroutine frame. Always capture
* by explicit pointer-to-value instead:
*
* // BROKEN on GCC 14:
* jq.postCoroTask(t, n, [&](auto) -> CoroTask<void> { ... });
*
* // FIX -- capture pointers explicitly:
* jq.postCoroTask(t, n, [ptr = &val](auto) -> CoroTask<void> { ... });
*
* BUG-RISK: Resuming a destroyed or completed CoroTask.
* Calling handle().resume() after the coroutine has already run to
* completion (done() == true) is undefined behavior. The CoroTaskRunner
* guards against this with an XRPL_ASSERT, but standalone usage of
* CoroTask must check done() before resuming.
*
* BUG-RISK: Moving a CoroTask that is being awaited.
* If task A is co_await-ed by task B (so A.continuation_ == B), moving
* or destroying A will invalidate the continuation link. Never move
* or reassign a CoroTask while it is mid-execution or being awaited.
*
* LIMITATION: CoroTask is fire-and-forget for the top-level owner.
* There is no built-in notification when the coroutine finishes.
* The caller must use external synchronization (e.g. CoroTaskRunner::join
* or a gate/condition_variable) to know when it is done.
*
* LIMITATION: No cancellation token.
* There is no way to cancel a suspended CoroTask from outside. The
* coroutine body must cooperatively check a flag (e.g. jq_.isStopping())
* after each co_await and co_return early if needed.
*
* LIMITATION: Stackless -- cannot suspend from nested non-coroutine calls.
* If a coroutine calls a regular function that wants to "yield", it
* cannot. Only the immediate coroutine body can use co_await.
* This is acceptable for xrpld because all yield() sites are shallow.
*/
template <>
class CoroTask<void>
{
public:
// The C++ coroutine protocol mandates these names (promise_type,
// initial_suspend, await_ready, ...) and instance-callable awaiter
// methods, which conflict with the project naming/static conventions.
// NOLINTBEGIN(readability-identifier-naming, readability-convert-member-functions-to-static)
struct promise_type;
using Handle = std::coroutine_handle<promise_type>;
/**
* Coroutine promise. Compiler uses this to manage coroutine state.
* Stores the exception (if any) and the continuation handle for
* symmetric transfer back to the awaiting coroutine.
*/
struct promise_type
{
// Captured exception from the coroutine body, rethrown in
// await_resume() when this task is co_await-ed by a caller.
std::exception_ptr exception_;
// Handle to the coroutine that is co_await-ing this task.
// Set by await_suspend(). FinalAwaiter uses it for symmetric
// transfer back to the caller. Null if this is a top-level task.
std::coroutine_handle<> continuation_;
/**
* Create the CoroTask return object.
* Called by the compiler at coroutine creation.
*/
CoroTask
get_return_object()
{
return CoroTask{Handle::from_promise(*this)};
}
/**
* Lazy start. The coroutine body does not execute until the
* handle is explicitly resumed (e.g. by CoroTaskRunner::resume).
*/
std::suspend_always
initial_suspend() noexcept
{
return {};
}
/**
* Awaiter returned by final_suspend(). Uses symmetric transfer:
* if a continuation exists, transfers control directly to it
* (tail-call, no stack growth). Otherwise returns noop_coroutine
* so the coroutine frame stays alive for the owner to destroy.
*/
struct FinalAwaiter
{
/**
* Always false. We need await_suspend to run for
* symmetric transfer.
*/
bool
await_ready() noexcept
{
return false;
}
/**
* Symmetric transfer: returns the continuation handle so
* the compiler emits a tail-call instead of a nested resume.
* If no continuation is set, returns noop_coroutine to
* suspend at final_suspend without destroying the frame.
*
* @param h Handle to this completing coroutine
*
* @return Continuation handle, or noop_coroutine
*/
std::coroutine_handle<>
await_suspend(Handle h) noexcept
{
if (auto cont = h.promise().continuation_)
return cont;
return std::noop_coroutine();
}
void
await_resume() noexcept
{
}
};
/**
* Returns FinalAwaiter for symmetric transfer at coroutine end.
*/
FinalAwaiter
final_suspend() noexcept
{
return {};
}
/**
* Called by the compiler for `co_return;` (void coroutine).
*/
void
return_void()
{
}
/**
* Called by the compiler when an exception escapes the coroutine
* body. Captures it for later rethrowing in await_resume().
*/
void
unhandled_exception()
{
exception_ = std::current_exception();
}
};
// NOLINTEND(readability-identifier-naming, readability-convert-member-functions-to-static)
/**
* Default constructor. Creates an empty (null handle) task.
*/
CoroTask() = default;
/**
* Takes ownership of a compiler-generated coroutine handle.
*
* @param h Coroutine handle to own
*/
explicit CoroTask(Handle h) : handle_(h)
{
}
/**
* Destroys the coroutine frame if this task owns one.
*/
~CoroTask()
{
if (handle_)
handle_.destroy();
}
/**
* Move constructor. Transfers handle ownership, leaves other empty.
*/
CoroTask(CoroTask&& other) noexcept : handle_(std::exchange(other.handle_, {}))
{
}
/**
* Move assignment. Destroys current frame (if any), takes other's.
*/
CoroTask&
operator=(CoroTask&& other) noexcept
{
if (this != &other)
{
if (handle_)
handle_.destroy();
handle_ = std::exchange(other.handle_, {});
}
return *this;
}
CoroTask(CoroTask const&) = delete;
CoroTask&
operator=(CoroTask const&) = delete;
/**
* @return The underlying coroutine_handle
*/
[[nodiscard]] Handle
handle() const
{
return handle_;
}
/**
* @return true if the coroutine has run to completion (or thrown)
*/
[[nodiscard]] bool
done() const
{
return handle_ && handle_.done();
}
// -- Awaiter interface: allows `co_await someCoroTask;` --
/**
* Always false. This task is lazy, so co_await always suspends
* the caller to set up the continuation link.
*/
// NOLINTBEGIN(readability-identifier-naming, readability-convert-member-functions-to-static)
[[nodiscard]] bool
await_ready() const noexcept
{
return false;
}
/**
* Stores the caller's handle as our continuation, then returns
* our handle for symmetric transfer (caller suspends, we resume).
*
* @param caller Handle of the coroutine doing co_await on us
*
* @return Our handle for symmetric transfer
*/
std::coroutine_handle<>
await_suspend(std::coroutine_handle<> caller) noexcept
{
XRPL_ASSERT(handle_, "xrpl::CoroTask<void>::await_suspend : handle is valid");
handle_.promise().continuation_ = caller;
return handle_; // Symmetric transfer
}
/**
* Called in the awaiting coroutine's context after this task
* completes. Rethrows any exception captured by
* unhandled_exception().
*/
void
await_resume()
{
XRPL_ASSERT(handle_, "xrpl::CoroTask<void>::await_resume : handle is valid");
if (auto& ep = handle_.promise().exception_)
std::rethrow_exception(ep);
}
// NOLINTEND(readability-identifier-naming, readability-convert-member-functions-to-static)
private:
// Exclusively-owned coroutine handle. Null after move or default
// construction. Destroyed in the destructor.
Handle handle_;
};
/**
* CoroTask<T> -- coroutine return type for value-returning coroutines.
*
* Class / Dependency Diagram
* ==========================
*
* CoroTask<T>
* +-----------------------------------------------+
* | - handle_ : Handle (coroutine_handle<promise>) |
* +-----------------------------------------------+
* | + handle(), done() |
* | + await_ready/suspend/resume (Awaiter iface) |
* +-----------------------------------------------+
* | owns
* v
* promise_type
* +-----------------------------------------------+
* | - result_ : variant<monostate, T, |
* | exception_ptr> |
* | - continuation_ : std::coroutine_handle<> |
* +-----------------------------------------------+
* | + get_return_object() -> CoroTask |
* | + initial_suspend() -> suspend_always (lazy) |
* | + final_suspend() -> FinalAwaiter |
* | + return_value(T) -> stores in result_[1] |
* | + unhandled_exception -> stores in result_[2] |
* +-----------------------------------------------+
* | returns at final_suspend
* v
* FinalAwaiter (same symmetric-transfer pattern as CoroTask<void>)
*
* Value Extraction
* ----------------
* await_resume() inspects the variant:
* - index 2 (exception_ptr) -> rethrow
* - index 1 (T) -> return value via move
*
* Usage Examples
* ==============
*
* 1. Simple value return:
*
* CoroTask<int> computeAnswer() { co_return 42; }
*
* CoroTask<void> caller() {
* int v = co_await computeAnswer(); // v == 42
* }
*
* 2. Chaining value-returning coroutines:
*
* CoroTask<int> add(int a, int b) { co_return a + b; }
* CoroTask<int> doubleSum(int a, int b) {
* int s = co_await add(a, b);
* co_return s * 2;
* }
*
* 3. Exception propagation from inner to outer:
*
* CoroTask<int> failing() {
* throw std::runtime_error("bad");
* co_return 0; // never reached
* }
* CoroTask<void> caller() {
* try {
* int v = co_await failing(); // throws here
* } catch (std::runtime_error const& e) {
* // e.what() == "bad"
* }
* }
*
* Caveats / Pitfalls (in addition to CoroTask<void> caveats above)
* ================================================================
*
* BUG-RISK: await_resume() moves the value out of the variant.
* Calling co_await on the same CoroTask<T> instance twice is undefined
* behavior -- the second call will see a moved-from T. CoroTask is
* single-shot: one co_return, one co_await.
*
* BUG-RISK: T must be move-constructible.
* return_value(T) takes by value and moves into the variant.
* Types that are not movable cannot be used as T.
*
* LIMITATION: No co_yield support.
* CoroTask<T> only supports a single co_return. It does not implement
* yield_value(), so using co_yield inside a CoroTask<T> coroutine is a
* compile error. For streaming values, a different return type
* (e.g. Generator<T>) would be needed.
*
* LIMITATION: Result is only accessible via co_await.
* There is no .get() or .result() method. The value can only be
* extracted by co_await-ing the CoroTask<T> from inside another
* coroutine. For extracting results in non-coroutine code, pass a
* pointer to the caller and write through it (as the tests do).
*/
template <typename T>
class CoroTask
{
static_assert(
std::is_move_constructible_v<T>,
"CoroTask<T> requires T to be move-constructible");
public:
// The C++ coroutine protocol mandates these names (promise_type,
// initial_suspend, await_ready, ...) and instance-callable awaiter
// methods, which conflict with the project naming/static conventions.
// NOLINTBEGIN(readability-identifier-naming, readability-convert-member-functions-to-static)
struct promise_type;
using Handle = std::coroutine_handle<promise_type>;
/**
* Coroutine promise for value-returning coroutines.
* Stores the result as a variant: monostate (not yet set),
* T (co_return value), or exception_ptr (unhandled exception).
*/
struct promise_type
{
// Tri-state result:
// index 0 (monostate) -- coroutine has not yet completed
// index 1 (T) -- co_return value stored here
// index 2 (exception) -- unhandled exception captured here
std::variant<std::monostate, T, std::exception_ptr> result_;
// Handle to the coroutine co_await-ing this task. Used by
// FinalAwaiter for symmetric transfer. Null for top-level tasks.
std::coroutine_handle<> continuation_;
/**
* Create the CoroTask return object.
* Called by the compiler at coroutine creation.
*/
CoroTask
get_return_object()
{
return CoroTask{Handle::from_promise(*this)};
}
/**
* Lazy start. Coroutine body does not run until explicitly resumed.
*/
std::suspend_always
initial_suspend() noexcept
{
return {};
}
/**
* Symmetric-transfer awaiter at coroutine completion.
* Same pattern as CoroTask<void>::FinalAwaiter.
*/
struct FinalAwaiter
{
bool
await_ready() noexcept
{
return false;
}
/**
* Returns continuation for symmetric transfer, or
* noop_coroutine if this is a top-level task.
*
* @param h Handle to this completing coroutine
*
* @return Continuation handle, or noop_coroutine
*/
std::coroutine_handle<>
await_suspend(Handle h) noexcept
{
if (auto cont = h.promise().continuation_)
return cont;
return std::noop_coroutine();
}
void
await_resume() noexcept
{
}
};
FinalAwaiter
final_suspend() noexcept
{
return {};
}
/**
* Called by the compiler for `co_return value;`.
* Moves the value into result_ at index 1.
*
* @param value The value to store
*/
void
return_value(T value)
{
result_.template emplace<1>(std::move(value));
}
/**
* Captures unhandled exceptions at index 2 of result_.
* Rethrown later in await_resume().
*/
void
unhandled_exception()
{
result_.template emplace<2>(std::current_exception());
}
};
// NOLINTEND(readability-identifier-naming, readability-convert-member-functions-to-static)
/**
* Default constructor. Creates an empty (null handle) task.
*/
CoroTask() = default;
/**
* Takes ownership of a compiler-generated coroutine handle.
*
* @param h Coroutine handle to own
*/
explicit CoroTask(Handle h) : handle_(h)
{
}
/**
* Destroys the coroutine frame if this task owns one.
*/
~CoroTask()
{
if (handle_)
handle_.destroy();
}
/**
* Move constructor. Transfers handle ownership, leaves other empty.
*/
CoroTask(CoroTask&& other) noexcept : handle_(std::exchange(other.handle_, {}))
{
}
/**
* Move assignment. Destroys current frame (if any), takes other's.
*/
CoroTask&
operator=(CoroTask&& other) noexcept
{
if (this != &other)
{
if (handle_)
handle_.destroy();
handle_ = std::exchange(other.handle_, {});
}
return *this;
}
CoroTask(CoroTask const&) = delete;
CoroTask&
operator=(CoroTask const&) = delete;
/**
* @return The underlying coroutine_handle
*/
[[nodiscard]] Handle
handle() const
{
return handle_;
}
/**
* @return true if the coroutine has run to completion (or thrown)
*/
[[nodiscard]] bool
done() const
{
return handle_ && handle_.done();
}
// -- Awaiter interface: allows `T val = co_await someCoroTask;` --
/**
* Always false. co_await always suspends to set up continuation.
*/
// NOLINTBEGIN(readability-identifier-naming, readability-convert-member-functions-to-static)
[[nodiscard]] bool
await_ready() const noexcept
{
return false;
}
/**
* Stores caller as continuation, returns our handle for
* symmetric transfer.
*
* @param caller Handle of the coroutine doing co_await on us
*
* @return Our handle for symmetric transfer
*/
std::coroutine_handle<>
await_suspend(std::coroutine_handle<> caller) noexcept
{
XRPL_ASSERT(handle_, "xrpl::CoroTask<T>::await_suspend : handle is valid");
handle_.promise().continuation_ = caller;
return handle_;
}
/**
* Extracts the result: rethrows if exception, otherwise moves
* the T value out of the variant. Single-shot: calling twice
* on the same task is undefined (moved-from T).
*
* @return The co_return-ed value
*/
T
await_resume()
{
XRPL_ASSERT(handle_, "xrpl::CoroTask<T>::await_resume : handle is valid");
auto& result = handle_.promise().result_;
if (auto* ep = std::get_if<2>(&result))
std::rethrow_exception(*ep);
return std::get<1>(std::move(result));
}
// NOLINTEND(readability-identifier-naming, readability-convert-member-functions-to-static)
private:
// Exclusively-owned coroutine handle. Null after move or default
// construction. Destroyed in the destructor.
Handle handle_;
};
} // namespace xrpl

View File

@@ -0,0 +1,448 @@
#pragma once
/**
* @file CoroTaskRunner.ipp
*
* CoroTaskRunner inline implementation.
*
* This file contains the business logic for managing C++20 coroutines
* on the JobQueue. It is included at the bottom of JobQueue.h.
*
* Data Flow: suspend / post / resume cycle
* =========================================
*
* coroutine body CoroTaskRunner JobQueue
* -------------- -------------- --------
* |
* co_await runner->suspend()
* |
* +--- await_suspend ------> onSuspend()
* | ++nSuspend_ ------------> nSuspend_
* | [coroutine is now suspended]
* |
* . (externally or by yieldAndPost())
* .
* +--- (caller calls) -----> post()
* | ++runCount_
* | addJob(resume) ----------> job enqueued
* | |
* | [worker picks up]
* | |
* +--- <----- resume() <-----------------------------------+
* | --nSuspend_ ------> nSuspend_
* | swap in LocalValues (lvs_)
* | task_.handle().resume()
* | |
* | [coroutine body continues here]
* | |
* | swap out LocalValues
* | --runCount_
* | cv_.notify_all()
* v
*
* Thread Safety
* =============
* - mutex_ : guards task_.handle().resume() so that post()-before-suspend
* races cannot resume the coroutine while it is still running.
* (See the race condition discussion in JobQueue.h)
* - mutexRun_ : guards runCount_ counter; used by join() to wait until
* all in-flight resume operations complete.
* - jq_.mutex_: guards nSuspend_ increments/decrements.
*
* Common Mistakes When Modifying This File
* =========================================
*
* 1. Changing lock ordering.
* resume() acquires locks sequentially (never held simultaneously):
* jq_.mutex_ (released immediately), then mutex_ (held across resume),
* then mutexRun_ (released after decrement). post() acquires only
* mutexRun_. Any new code path must follow the same order.
*
* 2. Removing the shared_from_this() capture in post().
* The lambda passed to addJob captures [this, sp = shared_from_this()].
* If you remove sp, 'this' can be destroyed before the job runs,
* causing use-after-free. The sp capture is load-bearing.
*
* 3. Forgetting to decrement nSuspend_ on a new code path.
* Every ++nSuspend_ must have a matching --nSuspend_. If you add a new
* suspension path (e.g. a new awaiter) and forget to decrement on resume
* or on failure, JobQueue::stop() will hang.
*
* 4. Calling task_.handle().resume() without holding mutex_.
* This allows a race where the coroutine runs on two threads
* simultaneously. Always hold mutex_ around resume().
*
* 5. Swapping LocalValues outside of the mutex_ critical section.
* The swap-in and swap-out of LocalValues must bracket the resume()
* call. If you move the swap-out before the lock_guard(mutex_) is
* released, you break LocalValue isolation for any code that runs
* after the coroutine suspends but before the lock is dropped.
*/
namespace xrpl {
/**
* Construct a CoroTaskRunner. Sets runCount_ to 0; does not
* create the coroutine. Call init() afterwards.
*
* @param jq The JobQueue this coroutine will run on
* @param type Job type for scheduling priority
* @param name Human-readable name for logging
*/
inline JobQueue::CoroTaskRunner::CoroTaskRunner(
CreateT,
JobQueue& jq,
JobType type,
std::string name)
: jq_(jq), type_(type), name_(std::move(name))
{
}
/**
* Initialize with a coroutine-returning callable.
* Stores the callable on the heap (FuncStore) so it outlives the
* coroutine frame. Coroutine frames store a reference to the
* callable's implicit object parameter (the lambda). If the callable
* is a temporary, that reference dangles after the caller returns.
* Keeping the callable alive here ensures the coroutine's captures
* remain valid.
*
* @param f Callable: CoroTask<void>(shared_ptr<CoroTaskRunner>)
*/
template <class F>
void
JobQueue::CoroTaskRunner::init(F&& f)
{
using Fn = std::decay_t<F>;
auto store = std::make_unique<FuncStore<Fn>>(std::forward<F>(f));
task_ = store->func(shared_from_this());
storedFunc_ = std::move(store);
}
/**
* Destructor. Waits for any in-flight resume() to complete, then
* asserts (debug) that the coroutine has finished or
* expectEarlyExit() was called.
*
* The join() call is necessary because with async dispatch the
* coroutine runs on a worker thread. The gate signal (which wakes
* the test thread) can arrive before resume() has set finished_.
* join() synchronizes via mutexRun_, establishing a happens-before
* edge: finished_ = true -> unlock(mutexRun_) in resume() ->
* lock(mutexRun_) in join() -> read finished_.
*/
inline JobQueue::CoroTaskRunner::~CoroTaskRunner()
{
#ifndef NDEBUG
join();
XRPL_ASSERT(finished_, "xrpl::JobQueue::CoroTaskRunner::~CoroTaskRunner : is finished");
#endif
}
/**
* Increment the JobQueue's suspended-coroutine count (nSuspend_).
*/
inline void
JobQueue::CoroTaskRunner::onSuspend()
{
std::scoped_lock const lock(jq_.mutex_);
++jq_.nSuspend_;
}
/**
* Decrement nSuspend_ without resuming.
*/
inline void
JobQueue::CoroTaskRunner::onUndoSuspend()
{
std::scoped_lock const lock(jq_.mutex_);
--jq_.nSuspend_;
}
/**
* Return a SuspendAwaiter whose await_suspend() increments nSuspend_
* before the coroutine actually suspends. The caller must later call
* post() or resume() to continue execution.
*
* @return Awaiter for use with `co_await runner->suspend()`
*/
inline auto
JobQueue::CoroTaskRunner::suspend()
{
/**
* Custom awaiter for suspend(). Always suspends (await_ready
* returns false) and increments nSuspend_ in await_suspend().
*/
// The C++ coroutine protocol mandates these awaiter names and
// instance-callable methods, which conflict with the project
// naming/static conventions.
// NOLINTBEGIN(readability-identifier-naming, readability-convert-member-functions-to-static)
struct SuspendAwaiter
{
CoroTaskRunner& runner_; // The runner that owns this coroutine.
/**
* Always returns false so the coroutine suspends.
*/
[[nodiscard]] bool
await_ready() const noexcept
{
return false;
}
/**
* Called when the coroutine suspends. Increments nSuspend_
* so the JobQueue knows a coroutine is waiting.
*/
void
await_suspend(std::coroutine_handle<>) const
{
runner_.onSuspend();
}
void
await_resume() const noexcept
{
}
};
// NOLINTEND(readability-identifier-naming, readability-convert-member-functions-to-static)
return SuspendAwaiter{*this};
}
/**
* Suspend and immediately repost on the JobQueue. Equivalent to
* `co_await JobQueueAwaiter{runner}` but uses an inline struct
* to work around a GCC-12 codegen bug (see declaration in JobQueue.h).
*
* If the JobQueue is stopping (post fails), the suspend count is
* undone and the coroutine continues immediately via symmetric
* transfer back to its own handle.
*
* @return An inline YieldPostAwaiter
*/
inline auto
JobQueue::CoroTaskRunner::yieldAndPost()
{
// The C++ coroutine protocol mandates these awaiter names and
// instance-callable methods, which conflict with the project
// naming/static conventions.
// NOLINTBEGIN(readability-identifier-naming, readability-convert-member-functions-to-static)
struct YieldPostAwaiter
{
CoroTaskRunner& runner_;
[[nodiscard]] bool
await_ready() const noexcept
{
return false;
}
/**
* Returns a coroutine_handle<> (symmetric transfer) rather than
* void + h.resume(). Two reasons:
*
* 1. h.resume() runs the coroutine nested inside this frame. A
* coroutine that yields in a loop against a stopping JobQueue
* fails post() every iteration, so the stack grows without
* bound. Symmetric transfer is a tail call and does not nest.
*
* 2. After h.resume() returns, the coroutine may have completed
* and destroyed its frame -- the frame this awaiter lives in.
* Returning from await_suspend would then touch freed memory.
*
* A bool return would also avoid nesting, but GCC-12 miscompiles
* bool-returning await_suspend (see JobQueueAwaiter.h).
*
* @return noop_coroutine() to stay suspended (job posted);
* the caller's handle to continue now (JQ stopping)
*/
std::coroutine_handle<>
await_suspend(std::coroutine_handle<> h)
{
runner_.onSuspend();
if (!runner_.post())
{
runner_.onUndoSuspend();
return h;
}
return std::noop_coroutine();
}
void
await_resume() const noexcept
{
}
};
// NOLINTEND(readability-identifier-naming, readability-convert-member-functions-to-static)
return YieldPostAwaiter{*this};
}
/**
* Schedule coroutine resumption as a job on the JobQueue.
* A shared_ptr capture (sp) prevents this CoroTaskRunner from being
* destroyed while the job is queued but not yet executed.
*
* @return false if the JobQueue rejected the job (shutting down)
*/
inline bool
JobQueue::CoroTaskRunner::post()
{
{
std::scoped_lock const lk(mutexRun_);
++runCount_;
}
// sp prevents 'this' from being destroyed while the job is pending
if (jq_.addJob(type_, name_, [this, sp = shared_from_this()]() { resume(); }))
{
return true;
}
// The coroutine will not run. Undo the runCount_ increment.
std::scoped_lock const lk(mutexRun_);
--runCount_;
cv_.notify_all();
return false;
}
/**
* Resume the coroutine on the current thread.
*
* Steps:
* 1. Decrement nSuspend_ (under jq_.mutex_)
* 2. Swap in this coroutine's LocalValues for thread-local isolation
* 3. Resume the coroutine handle (under mutex_)
* 4. Swap out LocalValues, restoring the thread's previous state
* 5. Decrement runCount_ and notify join() waiters
*
* @pre post() must have been called before resume(). Direct calls
* without a prior post() will corrupt runCount_ and break join().
* Note: runCount_ is NOT incremented here — post() already did that.
* This ensures join() stays blocked for the entire post->resume lifetime.
*/
inline void
JobQueue::CoroTaskRunner::resume()
{
{
std::scoped_lock const lock(jq_.mutex_);
--jq_.nSuspend_;
}
auto saved = detail::getLocalValues().release();
detail::getLocalValues().reset(&lvs_);
std::scoped_lock const lock(mutex_);
XRPL_ASSERT(
task_.handle() && !task_.done(),
"xrpl::JobQueue::CoroTaskRunner::resume : task handle is valid and not done");
if (task_.handle() && !task_.done())
{
task_.handle().resume();
}
else
{
// A resume() with no coroutine to run (e.g. a duplicate external
// post() after completion). Resuming a null or finished handle is
// undefined behavior, so skip it -- this matches the old
// Coro::resume() `if (coro_)` guard. The bookkeeping below still
// runs to balance the ++runCount_ done by the post() that
// scheduled this call.
JLOG(jq_.journal_.warn())
<< "CoroTaskRunner::resume called for coroutine '" << name_
<< "' with no runnable coroutine (duplicate post or already completed)";
}
detail::getLocalValues().release();
detail::getLocalValues().reset(saved);
if (task_.done())
{
finished_ = true;
// An exception that escapes a top-level coroutine body is captured
// by promise_type::unhandled_exception() but has no awaiter to
// rethrow it, so it would vanish with the frame. Surface it in the
// log. (The old Boost path propagated it out of resume() instead.)
if (auto const& ep = task_.handle().promise().exception_)
{
try
{
std::rethrow_exception(ep);
}
catch (std::exception const& e)
{
JLOG(jq_.journal_.error())
<< "Unhandled exception in coroutine '" << name_ << "': " << e.what();
}
catch (...)
{
JLOG(jq_.journal_.error())
<< "Unhandled non-standard exception in coroutine '" << name_ << "'";
}
}
// Break the shared_ptr cycle: frame -> shared_ptr<runner> -> this.
// Use std::move (not task_ = {}) so task_.handle_ is null BEFORE the
// frame is destroyed. operator= would destroy the frame while handle_
// still holds the old value -- a re-entrancy hazard on GCC-12 if
// frame destruction triggers runner cleanup.
[[maybe_unused]] auto completed = std::move(task_);
}
std::scoped_lock const lk(mutexRun_);
--runCount_;
cv_.notify_all();
}
/**
* @return true if the coroutine has not yet run to completion
*/
inline bool
JobQueue::CoroTaskRunner::runnable() const
{
// After normal completion, task_ is reset to break the shared_ptr cycle
// (handle_ becomes null). A null handle means the coroutine is done.
return task_.handle() && !task_.done();
}
/**
* Handle early termination when the coroutine never ran (e.g. JobQueue
* is stopping). Decrements nSuspend_ and destroys the coroutine frame
* to break the shared_ptr cycle: frame -> lambda -> runner -> frame.
*/
inline void
JobQueue::CoroTaskRunner::expectEarlyExit()
{
if (!finished_)
{
std::scoped_lock const lock(jq_.mutex_);
--jq_.nSuspend_;
finished_ = true;
}
// Break the shared_ptr cycle: frame -> shared_ptr<runner> -> this.
// The coroutine is at initial_suspend and never ran user code, so
// destroying it is safe. Use std::move (not task_ = {}) so
// task_.handle_ is null before the frame is destroyed.
{
[[maybe_unused]] auto completed = std::move(task_);
}
storedFunc_.reset();
}
/**
* Block until all pending/active resume operations complete.
* Uses cv_ + mutexRun_ to wait until runCount_ reaches 0 or
* finished_ becomes true. The finished_ check handles the case
* where resume() is called directly (without post()), which
* decrements runCount_ below zero. In that scenario runCount_
* never returns to 0, but finished_ becoming true guarantees
* the coroutine is done and no more resumes will occur.
*
* Note: when join() returns via the finished_ disjunct, the final
* resume() call may still be executing its post-completion
* bookkeeping (the --runCount_ / notify after finished_ is set).
* That is safe -- the coroutine body has fully completed and the
* runner is kept alive by the resume job's shared_ptr -- but
* callers must not assume resume() itself has returned.
*/
inline void
JobQueue::CoroTaskRunner::join()
{
std::unique_lock<std::mutex> lk(mutexRun_);
cv_.wait(lk, [this]() { return runCount_ == 0 || finished_; });
}
} // namespace xrpl

View File

@@ -1,25 +1,17 @@
#pragma once
#include <xrpl/basics/LocalValue.h>
#include <xrpl/core/ClosureCounter.h>
#include <xrpl/core/JobTypeData.h>
#include <xrpl/core/detail/Workers.h>
#include <xrpl/json/json_value.h>
// Include only the specific Boost.Coroutine2 headers actually used here.
// Avoid `boost/coroutine2/all.hpp` because it transitively pulls in
// `boost/context/pooled_fixedsize_stack.hpp`, whose `.malloc()` / `.free()`
// member calls on `boost::pool` collide with MSVC's `_CRTDBG_MAP_ALLOC` macros
// in Debug builds (see cmake/XrplCompiler.cmake).
#include <xrpl/beast/insight/Collector.h>
#include <xrpl/beast/insight/Gauge.h>
#include <xrpl/beast/insight/Hook.h>
#include <xrpl/beast/utility/Journal.h>
#include <xrpl/core/ClosureCounter.h>
#include <xrpl/core/CoroTask.h>
#include <xrpl/core/Job.h>
#include <xrpl/core/JobTypeData.h>
#include <xrpl/core/LoadEvent.h>
#include <boost/context/protected_fixedsize_stack.hpp>
#include <boost/coroutine2/coroutine.hpp>
#include <xrpl/core/detail/Workers.h>
#include <xrpl/json/json_value.h>
#include <atomic>
#include <chrono>
@@ -40,10 +32,6 @@ class PerfLog;
} // namespace perf
class Logs;
struct CoroCreateT
{
explicit CoroCreateT() = default;
};
/**
* A pool of threads to perform work.
@@ -60,93 +48,414 @@ class JobQueue : private Workers::Callback
{
public:
/**
* Coroutines must run to completion.
* C++20 coroutine lifecycle manager.
*
* Class / Inheritance / Dependency Diagram
* =========================================
*
* std::enable_shared_from_this<CoroTaskRunner>
* ^
* | (public inheritance)
* |
* CoroTaskRunner
* +---------------------------------------------------+
* | - lvs_ : detail::LocalValues |
* | - jq_ : JobQueue& |
* | - type_ : JobType |
* | - name_ : std::string |
* | - runCount_ : int (in-flight resumes) |
* | - mutex_ : std::mutex (coroutine guard) |
* | - mutexRun_ : std::mutex (join guard) |
* | - cv_ : condition_variable |
* | - task_ : CoroTask<void> |
* | - storedFunc_ : unique_ptr<FuncBase> (type-erased)|
* +---------------------------------------------------+
* | + init(F&&) : set up coroutine callable |
* | + onSuspend() : ++jq_.nSuspend_ |
* | + onUndoSuspend() : --jq_.nSuspend_ |
* | + suspend() : returns SuspendAwaiter |
* | + post() : schedule resume on JobQueue |
* | + resume() : resume coroutine on caller |
* | + runnable() : !task_.done() |
* | + expectEarlyExit() : teardown for failed post |
* | + join() : block until not running |
* +---------------------------------------------------+
* | |
* | owns | references
* v v
* CoroTask<void> JobQueue
* (coroutine frame) (thread pool + nSuspend_)
*
* FuncBase / FuncStore<F> (type-erased heap storage
* for the coroutine lambda)
*
* Coroutine Lifecycle (Control Flow)
* ===================================
*
* Caller thread JobQueue worker thread
* ------------- ----------------------
* postCoroTask(f)
* |
* +-- reserve a jobCounter_ slot (reject if JQ shutting down)
* +-- ++nSuspend_ (lazy start counts as suspended)
* +-- make_shared<CoroTaskRunner>
* +-- init(f)
* | +-- store lambda on heap (FuncStore)
* | +-- task_ = f(shared_from_this())
* | [coroutine created, suspended at initial_suspend]
* +-- post()
* | +-- ++runCount_
* | +-- addJob(type_, [resume]{})
* | resume()
* | |
* | +-- --nSuspend_
* | +-- swap in LocalValues
* | +-- task_.handle().resume()
* | | [coroutine body runs]
* | | ...
* | | co_await suspend()
* | | +-- ++nSuspend_
* | | [coroutine suspends]
* | +-- swap out LocalValues
* | +-- --runCount_
* | +-- cv_.notify_all()
* |
* post() <-- called externally or by yieldAndPost()
* +-- ++runCount_
* +-- addJob(type_, [resume]{})
* resume()
* |
* +-- [coroutine body continues]
* +-- co_return
* +-- --runCount_
* +-- cv_.notify_all()
* join()
* +-- cv_.wait([]{runCount_ == 0})
* +-- [done]
*
* Usage Examples
* ==============
*
* 1. Fire-and-forget coroutine (most common pattern):
*
* jq.postCoroTask(JtClient, "MyWork",
* [](auto runner) -> CoroTask<void> {
* doSomeWork();
* co_await runner->suspend(); // yield to other jobs
* doMoreWork();
* co_return;
* });
*
* 2. Manually controlling suspend / resume (external trigger):
*
* auto runner = jq.postCoroTask(JtClient, "ExtTrigger",
* [&result](auto runner) -> CoroTask<void> {
* startAsyncOperation(callback);
* co_await runner->suspend();
* // callback called runner->post() to get here
* result = collectResult();
* co_return;
* });
* // ... later, from the callback:
* runner->post(); // reschedule the coroutine on the JobQueue
*
* 3. Using yieldAndPost() for automatic suspend + repost:
*
* jq.postCoroTask(JtClient, "AutoRepost",
* [](auto runner) -> CoroTask<void> {
* step1();
* co_await runner->yieldAndPost(); // yield + auto-repost
* step2();
* co_await runner->yieldAndPost();
* step3();
* co_return;
* });
*
* 4. Checking shutdown after co_await (cooperative cancellation):
*
* jq.postCoroTask(JtClient, "Cancellable",
* [&jq](auto runner) -> CoroTask<void> {
* while (moreWork()) {
* co_await runner->yieldAndPost();
* if (jq.isStopping())
* co_return; // bail out cleanly
* processNextItem();
* }
* co_return;
* });
*
* Caveats / Pitfalls
* ==================
*
* BUG-RISK: Calling suspend() without a matching post()/resume().
* After co_await runner->suspend(), the coroutine is parked and
* nSuspend_ is incremented. If nothing ever calls post() or
* resume(), the coroutine is leaked and JobQueue::stop() will
* hang forever waiting for nSuspend_ to reach zero.
*
* BUG-RISK: Calling post() on an already-running coroutine.
* post() schedules a resume() job. If the coroutine has not
* actually suspended yet (no co_await executed), the resume job
* will try to call handle().resume() while the coroutine is still
* running on another thread. This is UB. The mutex_ prevents
* data corruption but the logic is wrong — always co_await
* suspend() before calling post(). (The test incorrect_order()
* shows this works only because mutex_ serializes the calls.)
*
* BUG-RISK: Dropping the shared_ptr<CoroTaskRunner> before join().
* The CoroTaskRunner destructor asserts that finished_ is true
* (the coroutine completed). If you let the last shared_ptr die
* while the coroutine is still running or suspended, you get an
* assertion failure in debug and UB in release. Always call
* join() or expectEarlyExit() first.
*
* BUG-RISK: Lambda captures outliving the coroutine frame.
* The lambda passed to postCoroTask is heap-allocated (FuncStore)
* to prevent dangling. But objects captured by pointer still need
* their own lifetime management. If you capture a raw pointer to
* a stack variable, and the stack frame exits before the coroutine
* finishes, the pointer dangles. Use shared_ptr or ensure the
* pointed-to object outlives the coroutine.
*
* BUG-RISK: Forgetting co_return in a void coroutine.
* If the coroutine body falls off the end without co_return,
* the compiler may silently treat it as co_return (per standard),
* but some compilers warn. Always write explicit co_return.
*
* LIMITATION: CoroTaskRunner only supports CoroTask<void>.
* The task_ member is CoroTask<void>. To return values from
* the top-level coroutine, write through a captured pointer
* (as the tests demonstrate), or co_await inner CoroTask<T>
* coroutines that return values.
*
* LIMITATION: One coroutine per CoroTaskRunner.
* init() must be called exactly once. You cannot reuse a
* CoroTaskRunner to run a second coroutine. Create a new one
* via postCoroTask() instead.
*
* LIMITATION: No timeout on join().
* join() blocks indefinitely. If the coroutine is suspended
* and never posted, join() will deadlock. Use timed waits
* on the gate pattern (condition_variable + wait_for) in tests.
*/
class Coro : public std::enable_shared_from_this<Coro>
class CoroTaskRunner : public std::enable_shared_from_this<CoroTaskRunner>
{
private:
// Per-coroutine thread-local storage. Swapped in before resume()
// and swapped out after, so each coroutine sees its own LocalValue
// state regardless of which worker thread executes it.
detail::LocalValues lvs_;
// Back-reference to the owning JobQueue. Used to post jobs,
// increment/decrement nSuspend_, and acquire jq_.mutex_.
JobQueue& jq_;
// Job type passed to addJob() when posting this coroutine.
JobType type_;
// Human-readable name for this coroutine job (for logging).
std::string name_;
bool running_{false};
// Number of in-flight resume operations (pending + active).
// Incremented by post(), decremented when resume() finishes.
// Guarded by mutexRun_. join() blocks until this reaches 0.
//
// A counter (not a bool) is needed because post() can be called
// from within the coroutine body (e.g. via yieldAndPost()),
// enqueuing a second resume while the first is still running.
// A bool would be clobbered: R2.post() sets true, then R1's
// cleanup sets false — losing the fact that R2 is still pending.
int runCount_ = 0;
// Serializes all coroutine resume() calls, preventing concurrent
// execution of the coroutine body on multiple threads. Handles the
// race where post() enqueues a resume before the coroutine has
// actually suspended (post-before-suspend pattern).
std::mutex mutex_;
// Guards runCount_. Used with cv_ for join() to wait
// until all pending/active resume operations complete.
std::mutex mutexRun_;
// Notified when runCount_ reaches zero, allowing
// join() waiters to wake up.
std::condition_variable cv_;
boost::coroutines2::coroutine<void>::push_type* yield_{};
boost::coroutines2::coroutine<void>::pull_type coro_;
#ifndef NDEBUG
// The coroutine handle wrapper. Owns the coroutine frame.
// Set by init(). Reset to empty in resume() upon coroutine
// completion (to break the shared_ptr cycle) or in
// expectEarlyExit() on early termination.
CoroTask<void> task_;
/**
* Type-erased base for heap-stored callables.
* Prevents the coroutine lambda from being destroyed before
* the coroutine frame is done with it.
*
* @see FuncStore
*/
struct FuncBase
{
virtual ~FuncBase() = default;
};
/**
* Concrete type-erased storage for a callable of type F.
* The coroutine frame stores a reference to the lambda's implicit
* object parameter. If the lambda is a temporary, that reference
* dangles after the call returns. FuncStore keeps it alive on
* the heap for the lifetime of the CoroTaskRunner.
*/
template <class F>
struct FuncStore : FuncBase
{
F func; // The stored callable (coroutine lambda).
explicit FuncStore(F&& f) : func(std::move(f))
{
}
};
// Heap-allocated callable storage. Set by init(), ensures the
// lambda outlives the coroutine frame that references it.
std::unique_ptr<FuncBase> storedFunc_;
// True once the coroutine has completed or expectEarlyExit() was
// called. Asserted in the destructor (debug) to catch leaked
// runners. Available in all builds to guard expectEarlyExit()
// against double-decrementing nSuspend_.
bool finished_ = false;
#endif
public:
template <class F>
Coro(CoroCreateT, JobQueue&, JobType, std::string, F&&);
// Not copy-constructible or assignable
Coro(Coro const&) = delete;
Coro&
operator=(Coro const&) = delete;
~Coro();
/**
* Tag type for private construction. Prevents external code
* from constructing CoroTaskRunner directly. Use postCoroTask().
*/
struct CreateT
{
explicit CreateT() = default;
};
/**
* Suspend coroutine execution.
* Effects:
* The coroutine's stack is saved.
* The associated Job thread is released.
* Note:
* The associated Job function returns.
* Undefined behavior if called consecutively without a corresponding
* post.
* Construct a CoroTaskRunner. Private by convention (CreateT tag).
*
* @param jq The JobQueue this coroutine will run on
* @param type Job type for scheduling priority
* @param name Human-readable name for logging
*/
CoroTaskRunner(CreateT, JobQueue&, JobType, std::string);
CoroTaskRunner(CoroTaskRunner const&) = delete;
CoroTaskRunner&
operator=(CoroTaskRunner const&) = delete;
/**
* Destructor. Asserts (debug) that the coroutine has finished
* or expectEarlyExit() was called.
*/
~CoroTaskRunner();
/**
* Initialize with a coroutine-returning callable.
* Must be called exactly once, after the object is managed by
* shared_ptr (because init uses shared_from_this internally).
* This is handled automatically by postCoroTask().
*
* @param f Callable: CoroTask<void>(shared_ptr<CoroTaskRunner>)
*/
template <class F>
void
init(F&& f);
/**
* Increment the JobQueue's suspended-coroutine count (nSuspend_).
* Called when the coroutine is about to suspend. Every call
* must be balanced by a corresponding decrement (via resume()
* or onUndoSuspend()), or JobQueue::stop() will hang.
*/
void
yield() const;
onSuspend();
/**
* Schedule coroutine execution.
* Effects:
* Returns immediately.
* A new job is scheduled to resume the execution of the coroutine.
* When the job runs, the coroutine's stack is restored and execution
* continues at the beginning of coroutine function or the
* statement after the previous call to yield. Undefined behavior if
* called after the coroutine has completed with a return (as opposed to
* a yield()). Undefined behavior if post() or resume() called
* consecutively without a corresponding yield.
* Decrement nSuspend_ without resuming.
* Used to undo onSuspend() when a scheduled post() fails
* (e.g. JobQueue is stopping).
*/
void
onUndoSuspend();
/**
* Suspend the coroutine.
* The awaiter's await_suspend() increments nSuspend_ before the
* coroutine actually suspends. The caller must later call post()
* or resume() to continue execution.
*
* @return true if the Coro's job is added to the JobQueue.
* @return An awaiter for use with `co_await runner->suspend()`
*/
auto
suspend();
/**
* Suspend the coroutine and immediately repost it on the
* JobQueue. Combines suspend() + post() atomically inside
* await_suspend, so there is no window where an external
* event could race between the two.
*
* Equivalent to JobQueueAwaiter but defined as an inline
* awaiter returned from a member function. This avoids a
* GCC-12 coroutine codegen bug where an external awaiter
* struct (JobQueueAwaiter) used at multiple co_await points
* corrupts the coroutine state machine's resume index,
* causing the coroutine to hang on the third resumption.
*
* @return An awaiter for use with `co_await runner->yieldAndPost()`
*/
auto
yieldAndPost();
/**
* Schedule coroutine resumption as a job on the JobQueue.
* Captures shared_from_this() to prevent this runner from being
* destroyed while the job is queued.
*
* @return true if the job was accepted; false if the JobQueue
* is stopping (caller must handle cleanup)
*/
bool
post();
/**
* Resume coroutine execution.
* Effects:
* The coroutine continues execution from where it last left off
* using this same thread.
* If the coroutine has already completed, returns immediately
* (handles the documented post-before-yield race condition).
* Undefined behavior if resume() or post() called consecutively
* without a corresponding yield.
* Resume the coroutine on the current thread.
* Decrements nSuspend_, swaps in LocalValues, resumes the
* coroutine handle, swaps out LocalValues, and notifies join()
* waiters. Lock ordering (sequential, non-overlapping):
* jq_.mutex_ -> mutex_ -> mutexRun_.
*
* @pre post() must have been called before resume(). Direct
* calls without a prior post() will corrupt runCount_
* and break join().
*/
void
resume();
/**
* Returns true if the Coro is still runnable (has not returned).
* @return true if the coroutine has not yet run to completion
*/
[[nodiscard]] bool
bool
runnable() const;
/**
* Once called, the Coro allows early exit without an assert.
* Handle early termination when the coroutine never ran.
* Decrements nSuspend_ and destroys the coroutine frame to
* break the shared_ptr cycle (frame -> lambda -> runner -> frame).
* Called by postCoroTask() when post() fails.
*/
void
expectEarlyExit();
/**
* Waits until coroutine returns from the user function.
* Block until all pending/active resume operations complete.
* Uses cv_ + mutexRun_ to wait until runCount_ reaches 0.
* Warning: deadlocks if the coroutine is suspended and never posted.
*/
void
join();
@@ -184,18 +493,18 @@ public:
}
/**
* Creates a coroutine and adds a job to the queue which will run it.
* Creates a C++20 coroutine and adds a job to the queue to run it.
*
* @param t The type of job.
* @param name Name of the job.
* @param f Has a signature of void(std::shared_ptr<Coro>). Called when the
* job executes.
* @param f Callable with signature
* CoroTask<void>(std::shared_ptr<CoroTaskRunner>).
*
* @return shared_ptr to posted Coro. nullptr if post was not successful.
* @return shared_ptr to posted CoroTaskRunner. nullptr if not successful.
*/
template <class F>
std::shared_ptr<Coro>
postCoro(JobType t, std::string const& name, F&& f);
std::shared_ptr<CoroTaskRunner>
postCoroTask(JobType t, std::string const& name, F&& f);
/**
* Jobs waiting at this priority.
@@ -256,8 +565,6 @@ public:
isStopped() const;
private:
friend class Coro;
using JobDataMap = std::map<JobType, JobTypeData>;
beast::Journal journal_;
@@ -358,88 +665,89 @@ private:
getJobLimit(JobType type);
};
/*
An RPC command is received and is handled via ServerHandler(HTTP) or
Handler(websocket), depending on the connection type. The handler then calls
the JobQueue::postCoro() method to create a coroutine and run it at a later
point. This frees up the handler thread and allows it to continue handling
other requests while the RPC command completes its work asynchronously.
postCoro() creates a Coro object. When the Coro ctor is called, and its
coro_ member is initialized (a boost::coroutines::pull_type), execution
automatically passes to the coroutine, which we don't want at this point,
since we are still in the handler thread context. It's important to note
here that construction of a boost pull_type automatically passes execution to
the coroutine. A pull_type object automatically generates a push_type that is
passed as a parameter (do_yield) in the signature of the function the
pull_type was created with. This function is immediately called during coro_
construction and within it, Coro::yield_ is assigned the push_type
parameter (do_yield) address and called (yield()) so we can return execution
back to the caller's stack.
postCoro() then calls Coro::post(), which schedules a job on the job
queue to continue execution of the coroutine in a JobQueue worker thread at
some later time. When the job runs, we lock on the Coro::mutex_ and call
coro_ which continues where we had left off. Since we the last thing we did
in coro_ was call yield(), the next thing we continue with is calling the
function param f, that was passed into Coro ctor. It is within this
function body that the caller specifies what he would like to do while
running in the coroutine and allow them to suspend and resume execution.
A task that relies on other events to complete, such as path finding, calls
Coro::yield() to suspend its execution while waiting on those events to
complete and continue when signaled via the Coro::post() method.
There is a potential race condition that exists here where post() can get
called before yield() after f is called. Technically the problem only occurs
if the job that post() scheduled is executed before yield() is called.
If the post() job were to be executed before yield(), undefined behavior
would occur. The lock ensures that coro_ is not called again until we exit
the coroutine. At which point a scheduled resume() job waiting on the lock
would gain entry. resume() checks if the coroutine has already completed
(coro_ converts to false) and, if so, skips invoking operator() since
calling operator() on a completed boost::coroutine2 pull_type is undefined
behavior.
The race condition occurs as follows:
1- The coroutine is running.
2- The coroutine is about to suspend, but before it can do so, it must
arrange for some event to wake it up.
3- The coroutine arranges for some event to wake it up.
4- Before the coroutine can suspend, that event occurs and the
resumption of the coroutine is scheduled on the job queue. 5- Again, before
the coroutine can suspend, the resumption of the coroutine is dispatched. 6-
Again, before the coroutine can suspend, the resumption code runs the
coroutine.
The coroutine is now running in two threads.
The lock prevents this from happening as step 6 will block until the
lock is released which only happens after the coroutine completes.
*/
} // namespace xrpl
#include <xrpl/core/Coro.ipp> // IWYU pragma: keep
#include <xrpl/core/CoroTaskRunner.ipp> // IWYU pragma: keep
namespace xrpl {
// postCoroTask — entry point for launching a C++20 coroutine on the JobQueue.
//
// Control Flow
// ============
//
// postCoroTask(t, name, f)
// |
// +-- 1. Reserve a jobCounter_ slot — reject if JQ shutting down
// |
// +-- 2. ++nSuspend_ (the coroutine uses lazy-start, so it is
// | "suspended" from the JQ's perspective before its first resume.
// | This keeps the JQ shutdown logic correct — it waits for
// | nSuspend_ to reach 0).
// |
// +-- 3. Create CoroTaskRunner (shared_ptr, ref-counted)
// |
// +-- 4. runner->init(f)
// | +-- Heap-allocate the lambda (FuncStore) to prevent
// | | dangling captures in the coroutine frame
// | +-- task_ = f(shared_from_this())
// | [coroutine created but NOT started — lazy initial_suspend]
// |
// +-- 5. runner->post()
// | +-- addJob(type_, [resume]{}) → resume on worker thread
// | +-- failure (JQ stopping):
// | +-- runner->expectEarlyExit()
// | | --nSuspend_, destroy coroutine frame
// | +-- return nullptr
// |
// +-- 6. Release the jobCounter_ slot (on return)
//
// Why reserve a jobCounter_ slot?
// ===============================
// JobQueue::stop() joins jobCounter_ before it acquires mutex_ and asserts
// nSuspend_ == 0. Without a reservation, stop() could run to completion in
// the window between the ++nSuspend_ in step 2 and the balancing post() or
// expectEarlyExit() in step 5, tripping that assert. Holding a slot blocks
// stop()'s join() for the whole function, closing the window. wrap() also
// returns nullopt once the counter is joined, so it doubles as the shutdown
// check — a plain stopping_ read cannot do this, because the read and the
// ++nSuspend_ are not a single atomic step.
//
// Why async post() instead of synchronous resume()?
// ==================================================
// The initial dispatch MUST use async post() so the coroutine body runs on
// a JobQueue worker thread, not the caller's thread. resume() swaps the
// caller's thread-local LocalValues with the coroutine's private copy.
// If the coroutine mutates LocalValues (e.g. thread_specific_storage test),
// those mutations bleed back into the caller's thread-local state after the
// swap-out, corrupting subsequent tests that share the same thread pool.
// Async post() avoids this by running the coroutine on a worker thread whose
// LocalValues are managed by the thread pool, not by the caller.
//
template <class F>
std::shared_ptr<JobQueue::Coro>
JobQueue::postCoro(JobType t, std::string const& name, F&& f)
std::shared_ptr<JobQueue::CoroTaskRunner>
JobQueue::postCoroTask(JobType t, std::string const& name, F&& f)
{
/* First param is a detail type to make construction private.
Last param is the function the coroutine runs. Signature of
void(std::shared_ptr<Coro>).
*/
auto coro = std::make_shared<Coro>(CoroCreateT{}, *this, t, name, std::forward<F>(f));
if (!coro->post())
// Held until this function returns. Null once jobCounter_ is joined,
// which is how a shutting-down JobQueue rejects new coroutines.
auto const shutdownGuard = jobCounter_.wrap([]() {});
if (!shutdownGuard)
return nullptr;
// Account for the initial suspension (CoroTask uses lazy start).
{
// The Coro was not successfully posted. Disable it so it's destructor
// can run with no negative side effects. Then destroy it.
coro->expectEarlyExit();
coro.reset();
std::scoped_lock const lock(mutex_);
++nSuspend_;
}
return coro;
auto runner = std::make_shared<CoroTaskRunner>(CoroTaskRunner::CreateT{}, *this, t, name);
runner->init(std::forward<F>(f));
if (!runner->post())
{
runner->expectEarlyExit();
runner.reset();
}
return runner;
}
} // namespace xrpl

View File

@@ -0,0 +1,212 @@
#pragma once
#include <xrpl/beast/utility/instrumentation.h>
#include <xrpl/core/JobQueue.h>
#include <coroutine>
#include <memory>
namespace xrpl {
/**
* Awaiter that suspends and immediately reschedules on the JobQueue.
* Equivalent to calling yield() followed by post() in the old Coro API.
*
* Usage:
* co_await JobQueueAwaiter{runner};
*
* What it waits for: The coroutine is re-queued as a job and resumes
* when a worker thread picks it up.
*
* Which thread resumes: A JobQueue worker thread.
*
* What await_resume() returns: void.
*
* Dependency Diagram
* ==================
*
* JobQueueAwaiter
* +----------------------------------------------+
* | + runner : shared_ptr<CoroTaskRunner> |
* +----------------------------------------------+
* | + await_ready() -> false (always suspend) |
* | + await_suspend() -> bool (suspend or cancel) |
* | + await_resume() -> void |
* +----------------------------------------------+
* | |
* | uses | uses
* v v
* CoroTaskRunner JobQueue
* .onSuspend() (via runner->post() -> addJob)
* .onUndoSuspend()
* .post()
*
* Control Flow (await_suspend)
* ============================
*
* co_await JobQueueAwaiter{runner}
* |
* +-- await_ready() -> false
* +-- await_suspend(handle)
* |
* +-- runner->onSuspend() // ++nSuspend_
* +-- runner->post() // addJob to JobQueue
* | |
* | +-- success? return noop_coroutine()
* | | // coroutine stays suspended;
* | | // worker thread will call resume()
* | +-- failure? (JQ stopping)
* | +-- runner->onUndoSuspend() // --nSuspend_
* | +-- return handle // symmetric transfer back
* | // coroutine continues immediately
* | // so it can clean up and co_return
*
* DEPRECATED — prefer `co_await runner->yieldAndPost()`
* =====================================================
*
* GCC-12 has a coroutine codegen bug where using this external awaiter
* struct at multiple co_await points in the same coroutine corrupts the
* state machine's resume index. After the second co_await, the third
* resumption enters handle().resume() but never reaches await_resume()
* or any subsequent user code — the coroutine hangs indefinitely.
*
* The fix is `co_await runner->yieldAndPost()`, which defines the
* awaiter as an inline struct inside a CoroTaskRunner member function.
* GCC-12 handles inline awaiters correctly at multiple co_await points.
*
* This struct is retained for single-use scenarios and documentation
* purposes. For any code that may use co_await in a loop or at
* multiple points, always use `runner->yieldAndPost()`.
*
* Usage Examples
* ==============
*
* 1. Yield and auto-repost (preferred — works on all compilers):
*
* CoroTask<void> handler(auto runner) {
* doPartA();
* co_await runner->yieldAndPost(); // yield + repost
* doPartB(); // runs on a worker thread
* co_return;
* }
*
* 2. Multiple yield points in a loop:
*
* CoroTask<void> batchProcessor(auto runner) {
* for (auto& item : items) {
* process(item);
* co_await runner->yieldAndPost(); // let other jobs run
* }
* co_return;
* }
*
* 3. Graceful shutdown — checking after resume:
*
* CoroTask<void> longTask(auto runner, JobQueue& jq) {
* while (hasWork()) {
* co_await runner->yieldAndPost();
* // If JQ is stopping, await_suspend resumes the coroutine
* // immediately without re-queuing. Always check
* // isStopping() to decide whether to proceed:
* if (jq.isStopping())
* co_return;
* doNextChunk();
* }
* co_return;
* }
*
* Caveats / Pitfalls
* ==================
*
* BUG-RISK: Using a stale or null runner.
* The runner shared_ptr must be valid and point to the CoroTaskRunner
* that owns the coroutine currently executing. Passing a runner from
* a different coroutine, or a default-constructed shared_ptr, is UB.
*
* BUG-RISK: Assuming resume happens on the same thread.
* After co_await, the coroutine resumes on whatever worker thread
* picks up the job. Do not rely on thread-local state unless it is
* managed through LocalValue (which CoroTaskRunner automatically
* swaps in/out).
*
* BUG-RISK: Ignoring the shutdown path.
* When the JobQueue is stopping, post() fails and await_suspend()
* resumes the coroutine immediately (symmetric transfer back to h).
* The coroutine body continues on the same thread. If your code
* after co_await assumes it was re-queued and is running on a worker
* thread, that assumption breaks during shutdown. Always handle the
* "JQ is stopping" case, either by checking jq.isStopping() or by
* letting the coroutine fall through to co_return naturally.
*
* DIFFERENCE from runner->suspend() + runner->post():
* Both JobQueueAwaiter and yieldAndPost() combine suspend + post
* in one atomic operation. With the manual suspend()/post() pattern,
* there is a window between the two calls where an external event
* could race. The atomic awaiters remove that window — onSuspend()
* and post() happen within the same await_suspend() call while the
* coroutine is guaranteed to be suspended. Use yieldAndPost() unless
* you need an external party to decide *when* to call post().
*/
struct JobQueueAwaiter
{
// The CoroTaskRunner that owns the currently executing coroutine.
std::shared_ptr<JobQueue::CoroTaskRunner> runner;
/**
* Always returns false so the coroutine suspends.
*/
// The C++ coroutine protocol mandates these awaiter names and
// instance-callable methods, which conflict with the project
// naming/static conventions.
// NOLINTBEGIN(readability-identifier-naming, readability-convert-member-functions-to-static)
[[nodiscard]] bool
await_ready() const noexcept
{
return false;
}
/**
* Increment nSuspend (equivalent to yield()) and schedule resume
* on the JobQueue (equivalent to post()). If the JobQueue is
* stopping, undoes the suspend count and transfers back to the
* coroutine so it can clean up and co_return.
*
* Returns a coroutine_handle<> (symmetric transfer) instead of
* bool to work around a GCC-12 codegen bug where bool-returning
* await_suspend leaves the coroutine in an invalid state —
* neither properly suspended nor resumed — causing a hang.
*
* WARNING: GCC-12 has an additional codegen bug where using this
* external awaiter struct at multiple co_await points in the same
* coroutine corrupts the state machine's resume index, causing the
* coroutine to hang on the third resumption. Prefer
* `co_await runner->yieldAndPost()` which uses an inline awaiter
* that GCC-12 handles correctly.
*
* @return noop_coroutine() to stay suspended (job posted);
* the caller's handle to resume immediately (JQ stopping)
*/
std::coroutine_handle<>
await_suspend(std::coroutine_handle<> h)
{
XRPL_ASSERT(runner, "xrpl::JobQueueAwaiter::await_suspend : runner is valid");
runner->onSuspend();
if (!runner->post())
{
// JobQueue is stopping. Undo the suspend count and
// transfer back to the coroutine so it can clean up
// and co_return.
runner->onUndoSuspend();
return h;
}
return std::noop_coroutine();
}
void
await_resume() const noexcept
{
}
// NOLINTEND(readability-identifier-naming, readability-convert-member-functions-to-static)
};
} // namespace xrpl

View File

@@ -19,6 +19,7 @@
#include <xrpl/basics/base_uint.h>
#include <xrpl/beast/unit_test/suite.h>
#include <xrpl/core/CoroTask.h>
#include <xrpl/core/Job.h>
#include <xrpl/core/JobQueue.h>
#include <xrpl/json/json_value.h>
@@ -123,7 +124,6 @@ public:
.ledgerMaster = app.getLedgerMaster(),
.consumer = c,
.role = Role::USER,
.coro = {},
.infoSub = {},
.apiVersion = rpc::kApiVersionIfUnspecified},
{},
@@ -135,22 +135,28 @@ public:
numSrc.reserve(rpc::tuning::kMaxSrcCur);
for (std::uint8_t i = 0; i < rpc::tuning::kMaxSrcCur; ++i)
numSrc.push_back(makeMptID(i, bob));
app.getJobQueue().postCoro(JtClient, "RPC-Client", [&](auto const& coro) {
// Safe capture: the test blocks on g.waitFor() until the coroutine
// completes, so the captured locals outlive the coroutine.
// NOLINTNEXTLINE(cppcoreguidelines-avoid-capturing-lambda-coroutines)
app.getJobQueue().postCoroTask(JtClient, "RPC-Client", [&](auto) -> CoroTask<void> {
context.params = xrpl::test::detail::rpf(alice, bob, usd, numSrc);
context.coro = coro;
rpc::doCommand(context, result);
g.signal();
co_return;
});
BEAST_EXPECT(g.waitFor(5s));
BEAST_EXPECT(!result.isMember(jss::error));
// Test more than rpc::tuning::max_src_cur source currencies.
numSrc.push_back(makeMptID(rpc::tuning::kMaxSrcCur, bob));
app.getJobQueue().postCoro(JtClient, "RPC-Client", [&](auto const& coro) {
// Safe capture: the test blocks on g.waitFor() until the coroutine
// completes, so the captured locals outlive the coroutine.
// NOLINTNEXTLINE(cppcoreguidelines-avoid-capturing-lambda-coroutines)
app.getJobQueue().postCoroTask(JtClient, "RPC-Client", [&](auto) -> CoroTask<void> {
context.params = xrpl::test::detail::rpf(alice, bob, usd, numSrc);
context.coro = coro;
rpc::doCommand(context, result);
g.signal();
co_return;
});
BEAST_EXPECT(g.waitFor(5s));
BEAST_EXPECT(result.isMember(jss::error));
@@ -162,22 +168,28 @@ public:
auto curm = MPTTester({.env = env, .issuer = alice, .holders = {bob}});
numSrc.push_back(curm.issuanceID());
}
app.getJobQueue().postCoro(JtClient, "RPC-Client", [&](auto const& coro) {
// Safe capture: the test blocks on g.waitFor() until the coroutine
// completes, so the captured locals outlive the coroutine.
// NOLINTNEXTLINE(cppcoreguidelines-avoid-capturing-lambda-coroutines)
app.getJobQueue().postCoroTask(JtClient, "RPC-Client", [&](auto) -> CoroTask<void> {
context.params = xrpl::test::detail::rpf(alice, bob, usd, {});
context.coro = coro;
rpc::doCommand(context, result);
g.signal();
co_return;
});
BEAST_EXPECT(g.waitFor(5s));
BEAST_EXPECT(!result.isMember(jss::error));
// Test more than rpc::tuning::max_auto_src_cur source currencies.
auto curm = MPTTester({.env = env, .issuer = alice, .holders = {bob}});
app.getJobQueue().postCoro(JtClient, "RPC-Client", [&](auto const& coro) {
// Safe capture: the test blocks on g.waitFor() until the coroutine
// completes, so the captured locals outlive the coroutine.
// NOLINTNEXTLINE(cppcoreguidelines-avoid-capturing-lambda-coroutines)
app.getJobQueue().postCoroTask(JtClient, "RPC-Client", [&](auto) -> CoroTask<void> {
context.params = xrpl::test::detail::rpf(alice, bob, usd, {});
context.coro = coro;
rpc::doCommand(context, result);
g.signal();
co_return;
});
BEAST_EXPECT(g.waitFor(5s));
BEAST_EXPECT(result.isMember(jss::error));

View File

@@ -26,6 +26,7 @@
#include <xrpl/basics/base_uint.h>
#include <xrpl/beast/unit_test/suite.h>
#include <xrpl/core/CoroTask.h>
#include <xrpl/core/Job.h>
#include <xrpl/core/JobQueue.h>
#include <xrpl/json/json_reader.h>
@@ -163,7 +164,6 @@ public:
.ledgerMaster = app.getLedgerMaster(),
.consumer = c,
.role = Role::USER,
.coro = {},
.infoSub = {},
.apiVersion = rpc::kApiVersionIfUnspecified},
{},
@@ -188,11 +188,14 @@ public:
json::Value result;
Gate g;
app.getJobQueue().postCoro(JtClient, "RPC-Client", [&](auto const& coro) {
// Safe capture: the test blocks on g.waitFor() until the coroutine
// completes, so the captured locals outlive the coroutine.
// NOLINTNEXTLINE(cppcoreguidelines-avoid-capturing-lambda-coroutines)
app.getJobQueue().postCoroTask(JtClient, "RPC-Client", [&](auto) -> CoroTask<void> {
context.params = std::move(params);
context.coro = coro;
rpc::doCommand(context, result);
g.signal();
co_return;
});
using namespace std::chrono_literals;
@@ -274,7 +277,6 @@ public:
.ledgerMaster = app.getLedgerMaster(),
.consumer = c,
.role = Role::USER,
.coro = {},
.infoSub = {},
.apiVersion = rpc::kApiVersionIfUnspecified},
{},
@@ -282,21 +284,27 @@ public:
json::Value result;
Gate g;
// Test rpc::tuning::max_src_cur source currencies.
app.getJobQueue().postCoro(JtClient, "RPC-Client", [&](auto const& coro) {
// Safe capture: the test blocks on g.waitFor() until the coroutine
// completes, so the captured locals outlive the coroutine.
// NOLINTNEXTLINE(cppcoreguidelines-avoid-capturing-lambda-coroutines)
app.getJobQueue().postCoroTask(JtClient, "RPC-Client", [&](auto) -> CoroTask<void> {
context.params = rpf(Account("alice"), Account("bob"), rpc::tuning::kMaxSrcCur);
context.coro = coro;
rpc::doCommand(context, result);
g.signal();
co_return;
});
BEAST_EXPECT(g.waitFor(5s));
BEAST_EXPECT(!result.isMember(jss::error));
// Test more than rpc::tuning::max_src_cur source currencies.
app.getJobQueue().postCoro(JtClient, "RPC-Client", [&](auto const& coro) {
// Safe capture: the test blocks on g.waitFor() until the coroutine
// completes, so the captured locals outlive the coroutine.
// NOLINTNEXTLINE(cppcoreguidelines-avoid-capturing-lambda-coroutines)
app.getJobQueue().postCoroTask(JtClient, "RPC-Client", [&](auto) -> CoroTask<void> {
context.params = rpf(Account("alice"), Account("bob"), rpc::tuning::kMaxSrcCur + 1);
context.coro = coro;
rpc::doCommand(context, result);
g.signal();
co_return;
});
BEAST_EXPECT(g.waitFor(5s));
BEAST_EXPECT(result.isMember(jss::error));
@@ -304,22 +312,28 @@ public:
// Test rpc::tuning::max_auto_src_cur source currencies.
for (auto i = 0; i < (rpc::tuning::kMaxAutoSrcCur - 1); ++i)
env.trust(Account("alice")[std::to_string(i + 100)](100), "bob");
app.getJobQueue().postCoro(JtClient, "RPC-Client", [&](auto const& coro) {
// Safe capture: the test blocks on g.waitFor() until the coroutine
// completes, so the captured locals outlive the coroutine.
// NOLINTNEXTLINE(cppcoreguidelines-avoid-capturing-lambda-coroutines)
app.getJobQueue().postCoroTask(JtClient, "RPC-Client", [&](auto) -> CoroTask<void> {
context.params = rpf(Account("alice"), Account("bob"), 0);
context.coro = coro;
rpc::doCommand(context, result);
g.signal();
co_return;
});
BEAST_EXPECT(g.waitFor(5s));
BEAST_EXPECT(!result.isMember(jss::error));
// Test more than rpc::tuning::max_auto_src_cur source currencies.
env.trust(Account("alice")["AUD"](100), "bob");
app.getJobQueue().postCoro(JtClient, "RPC-Client", [&](auto const& coro) {
// Safe capture: the test blocks on g.waitFor() until the coroutine
// completes, so the captured locals outlive the coroutine.
// NOLINTNEXTLINE(cppcoreguidelines-avoid-capturing-lambda-coroutines)
app.getJobQueue().postCoroTask(JtClient, "RPC-Client", [&](auto) -> CoroTask<void> {
context.params = rpf(Account("alice"), Account("bob"), 0);
context.coro = coro;
rpc::doCommand(context, result);
g.signal();
co_return;
});
BEAST_EXPECT(g.waitFor(5s));
BEAST_EXPECT(result.isMember(jss::error));

View File

@@ -24,7 +24,9 @@
#include <xrpl/basics/strHex.h>
#include <xrpl/beast/unit_test/suite.h>
#include <xrpl/beast/utility/Zero.h>
#include <xrpl/core/CoroTask.h>
#include <xrpl/core/Job.h>
#include <xrpl/core/JobQueue.h>
#include <xrpl/core/ServiceRegistry.h>
#include <xrpl/json/json_value.h>
#include <xrpl/json/to_string.h>
@@ -1638,7 +1640,6 @@ struct PayChan_test : public beast::unit_test::Suite
.ledgerMaster = app.getLedgerMaster(),
.consumer = c,
.role = Role::USER,
.coro = {},
.infoSub = {},
.apiVersion = rpc::kApiVersionIfUnspecified},
{},
@@ -1654,10 +1655,13 @@ struct PayChan_test : public beast::unit_test::Suite
BEAST_EXPECT(context.loadType == resource::kFeeReferenceRpc);
json::Value result;
Gate g;
app.getJobQueue().postCoro(JtClient, "RPC-Client", [&](auto const& coro) {
context.coro = coro;
// Safe capture: the test blocks on g.waitFor() until the coroutine
// completes, so the captured locals outlive the coroutine.
// NOLINTNEXTLINE(cppcoreguidelines-avoid-capturing-lambda-coroutines)
app.getJobQueue().postCoroTask(JtClient, "RPC-Client", [&](auto) -> CoroTask<void> {
result = doChannelVerify(context);
g.signal();
co_return;
});
using namespace std::chrono_literals;
@@ -1705,7 +1709,6 @@ struct PayChan_test : public beast::unit_test::Suite
.ledgerMaster = app.getLedgerMaster(),
.consumer = c,
.role = Role::ADMIN, // channel_authorize requires ADMIN or canSign()
.coro = {},
.infoSub = {},
.apiVersion = rpc::kApiVersionIfUnspecified},
{},
@@ -1720,10 +1723,13 @@ struct PayChan_test : public beast::unit_test::Suite
BEAST_EXPECT(context.loadType == resource::kFeeReferenceRpc);
json::Value result;
Gate g;
app.getJobQueue().postCoro(JtClient, "RPC-Client", [&](auto const& coro) {
context.coro = coro;
// Safe capture: the test blocks on g.waitFor() until the coroutine
// completes, so the captured locals outlive the coroutine.
// NOLINTNEXTLINE(cppcoreguidelines-avoid-capturing-lambda-coroutines)
app.getJobQueue().postCoroTask(JtClient, "RPC-Client", [&](auto) -> CoroTask<void> {
result = doChannelAuthorize(context);
g.signal();
co_return;
});
using namespace std::chrono_literals;

View File

@@ -0,0 +1,581 @@
#include <test/jtx/Env.h>
#include <test/jtx/envconfig.h>
#include <xrpld/core/Config.h>
#include <xrpl/basics/LocalValue.h>
#include <xrpl/beast/unit_test/suite.h>
#include <xrpl/core/CoroTask.h>
#include <xrpl/core/Job.h>
#include <xrpl/core/JobQueue.h>
#include <xrpl/core/JobQueueAwaiter.h>
#include <array>
#include <chrono>
#include <condition_variable>
#include <memory>
#include <mutex>
#include <stdexcept>
#include <string>
#include <vector>
// Tests intentionally capture state in coroutine lambdas; lifetimes are
// controlled by Gate synchronization and join() before scope exit.
// NOLINTBEGIN(cppcoreguidelines-avoid-capturing-lambda-coroutines)
namespace xrpl::test {
/**
* Test suite for the C++20 coroutine primitives: CoroTask, CoroTaskRunner,
* and JobQueueAwaiter.
*
* Dependency Diagram
* ==================
*
* CoroTask_test
* +-------------------------------------------------+
* | + Gate (inner class) : condition_variable helper |
* +-------------------------------------------------+
* | uses
* v
* jtx::Env --> JobQueue::postCoroTask()
* |
* +-- CoroTaskRunner (suspend / post / resume)
* +-- CoroTask<void> / CoroTask<T>
* +-- JobQueueAwaiter
*
* Test Coverage Matrix
* ====================
*
* Test | Primitives exercised
* --------------------------+----------------------------------------------
* testVoidCompletion | CoroTask<void> basic lifecycle
* testCorrectOrder | suspend() -> join() -> post() -> complete
* testIncorrectOrder | post() before suspend() (race-safe path)
* testJobQueueAwaiter | JobQueueAwaiter + yieldAndPost suspend/repost
* testThreadSpecificStorage | LocalValue isolation across coroutines
* testExceptionPropagation | CoroTask<void> exception via co_await
* testMultipleYields | N sequential suspend/resume cycles
* testValueReturn | CoroTask<T> co_return value
* testValueException | CoroTask<T> exception via co_await
* testValueChaining | nested CoroTask<T> -> CoroTask<T>
* testShutdownRejection | postCoroTask returns nullptr when stopping
*/
class CoroTask_test : public beast::unit_test::Suite
{
public:
/**
* Simple one-shot gate for synchronizing between test thread
* and coroutine worker threads. signal() sets the flag;
* waitFor() blocks until signaled or timeout.
*/
class Gate
{
private:
std::condition_variable cv_;
std::mutex mutex_;
bool signaled_ = false;
public:
/**
* Block until signaled or timeout expires.
*
* @param relTime Maximum duration to wait
*
* @return true if signaled before timeout
*/
template <class Rep, class Period>
bool
waitFor(std::chrono::duration<Rep, Period> const& relTime)
{
std::unique_lock<std::mutex> lk(mutex_);
auto b = cv_.wait_for(lk, relTime, [this] { return signaled_; });
signaled_ = false;
return b;
}
/**
* Signal the gate, waking any waiting thread.
*/
void
signal()
{
std::scoped_lock const lk(mutex_);
signaled_ = true;
cv_.notify_all();
}
};
// NOTE: All coroutine lambdas passed to postCoroTask use explicit
// pointer-by-value captures instead of [&] to work around a GCC 14
// bug where reference captures in coroutine lambdas are corrupted
// in the coroutine frame.
/**
* CoroTask<void> runs to completion and runner becomes non-runnable.
*/
void
testVoidCompletion()
{
using namespace std::chrono_literals;
using namespace jtx;
testcase("void completion");
Env env(*this, envconfig([](std::unique_ptr<Config> cfg) {
cfg->forceMultiThread = true;
return cfg;
}));
Gate g;
auto runner = env.app().getJobQueue().postCoroTask(
JtClient, "CoroTaskTest", [gp = &g](auto) -> CoroTask<void> {
gp->signal();
co_return;
});
BEAST_EXPECT(runner);
if (!BEAST_EXPECT(g.waitFor(5s)))
return;
runner->join();
BEAST_EXPECT(!runner->runnable());
}
/**
* Correct order: suspend, join, post, complete.
* Mirrors existing Coroutine_test::correct_order.
*/
void
testCorrectOrder()
{
using namespace std::chrono_literals;
using namespace jtx;
testcase("correct order");
Env env(*this, envconfig([](std::unique_ptr<Config> cfg) {
cfg->forceMultiThread = true;
return cfg;
}));
Gate g1, g2;
auto runner = env.app().getJobQueue().postCoroTask(
JtClient, "CoroTaskTest", [g1p = &g1, g2p = &g2](auto runner) -> CoroTask<void> {
g1p->signal();
co_await runner->suspend();
g2p->signal();
co_return;
});
BEAST_EXPECT(runner);
if (!BEAST_EXPECT(g1.waitFor(5s)))
return;
runner->join();
runner->post();
if (!BEAST_EXPECT(g2.waitFor(5s)))
return;
runner->join();
}
/**
* Incorrect order: post() before suspend(). Verifies the
* race-safe path. Mirrors Coroutine_test::incorrect_order.
*/
void
testIncorrectOrder()
{
using namespace std::chrono_literals;
using namespace jtx;
testcase("incorrect order");
Env env(*this, envconfig([](std::unique_ptr<Config> cfg) {
cfg->forceMultiThread = true;
return cfg;
}));
Gate g;
env.app().getJobQueue().postCoroTask(
JtClient, "CoroTaskTest", [gp = &g](auto runner) -> CoroTask<void> {
runner->post();
co_await runner->suspend();
gp->signal();
co_return;
});
BEAST_EXPECT(g.waitFor(5s));
}
/**
* Suspend + auto-repost across multiple yield points, using the
* external JobQueueAwaiter struct for the first suspension and the
* inline yieldAndPost() awaiter for the second. JobQueueAwaiter is
* used at only one co_await point per coroutine (see the GCC-12
* multi-use warning in JobQueueAwaiter.h).
*/
void
testJobQueueAwaiter()
{
using namespace std::chrono_literals;
using namespace jtx;
testcase("JobQueueAwaiter");
Env env(*this, envconfig([](std::unique_ptr<Config> cfg) {
cfg->forceMultiThread = true;
return cfg;
}));
Gate g;
std::vector<int> steps;
auto runner = env.app().getJobQueue().postCoroTask(
JtClient, "CoroTaskTest", [sp = &steps, gp = &g](auto runner) -> CoroTask<void> {
sp->push_back(1);
co_await JobQueueAwaiter{runner};
sp->push_back(2);
co_await runner->yieldAndPost();
sp->push_back(3);
gp->signal();
co_return;
});
BEAST_EXPECT(runner);
if (!BEAST_EXPECT(g.waitFor(5s)))
return;
runner->join();
BEAST_EXPECT(steps == std::vector<int>({1, 2, 3}));
}
/**
* Per-coroutine LocalValue isolation. Each coroutine sees its own
* copy of thread-local state. Mirrors Coroutine_test::thread_specific_storage.
*/
void
testThreadSpecificStorage()
{
using namespace std::chrono_literals;
using namespace jtx;
testcase("thread specific storage");
Env env(*this);
auto& jq = env.app().getJobQueue();
static constexpr int kN = 4;
std::array<std::shared_ptr<JobQueue::CoroTaskRunner>, kN> a;
LocalValue<int> lv(-1);
BEAST_EXPECT(*lv == -1);
Gate g;
jq.addJob(JtClient, "LocalValTest", [&]() {
this->BEAST_EXPECT(*lv == -1);
*lv = -2;
this->BEAST_EXPECT(*lv == -2);
g.signal();
});
if (!BEAST_EXPECT(g.waitFor(5s)))
return;
BEAST_EXPECT(*lv == -1);
for (int i = 0; i < kN; ++i)
{
jq.postCoroTask(
JtClient,
"CoroTaskTest",
[this, ap = &a, gp = &g, lvp = &lv, id = i](auto runner) -> CoroTask<void> {
(*ap)[id] = runner;
gp->signal();
co_await runner->suspend();
this->BEAST_EXPECT(**lvp == -1);
**lvp = id;
this->BEAST_EXPECT(**lvp == id);
gp->signal();
co_await runner->suspend();
this->BEAST_EXPECT(**lvp == id);
co_return;
});
if (!BEAST_EXPECT(g.waitFor(5s)))
return;
a[i]->join();
}
for (auto const& r : a)
{
r->post();
if (!BEAST_EXPECT(g.waitFor(5s)))
return;
r->join();
}
for (auto const& r : a)
{
r->post();
r->join();
}
jq.addJob(JtClient, "LocalValTest", [&]() {
this->BEAST_EXPECT(*lv == -2);
g.signal();
});
if (!BEAST_EXPECT(g.waitFor(5s)))
return;
BEAST_EXPECT(*lv == -1);
}
/**
* An exception thrown in an awaited CoroTask<void> is rethrown into
* the awaiting coroutine by await_resume(), with the original
* message intact. (An exception escaping the top-level body has no
* awaiter to rethrow it; it is captured by unhandled_exception()
* and logged by CoroTaskRunner::resume().)
*/
void
testExceptionPropagation()
{
using namespace std::chrono_literals;
using namespace jtx;
testcase("exception propagation");
Env env(*this, envconfig([](std::unique_ptr<Config> cfg) {
cfg->forceMultiThread = true;
return cfg;
}));
Gate g;
std::string what;
auto runner = env.app().getJobQueue().postCoroTask(
JtClient, "CoroTaskTest", [wp = &what, gp = &g](auto) -> CoroTask<void> {
auto inner = []() -> CoroTask<void> {
throw std::runtime_error("test exception");
co_return;
};
try
{
co_await inner();
}
catch (std::runtime_error const& e)
{
*wp = e.what();
}
gp->signal();
co_return;
});
BEAST_EXPECT(runner);
if (!BEAST_EXPECT(g.waitFor(5s)))
return;
runner->join();
BEAST_EXPECT(what == "test exception");
BEAST_EXPECT(!runner->runnable());
}
/**
* Multiple sequential suspend/resume cycles via co_await.
*/
void
testMultipleYields()
{
using namespace std::chrono_literals;
using namespace jtx;
testcase("multiple yields");
Env env(*this, envconfig([](std::unique_ptr<Config> cfg) {
cfg->forceMultiThread = true;
return cfg;
}));
Gate g;
int counter = 0;
auto runner = env.app().getJobQueue().postCoroTask(
JtClient, "CoroTaskTest", [cp = &counter, gp = &g](auto runner) -> CoroTask<void> {
++(*cp);
gp->signal();
co_await runner->suspend();
++(*cp);
gp->signal();
co_await runner->suspend();
++(*cp);
gp->signal();
co_return;
});
BEAST_EXPECT(runner);
if (!BEAST_EXPECT(g.waitFor(5s)))
return;
BEAST_EXPECT(counter == 1);
runner->join();
runner->post();
if (!BEAST_EXPECT(g.waitFor(5s)))
return;
BEAST_EXPECT(counter == 2);
runner->join();
runner->post();
if (!BEAST_EXPECT(g.waitFor(5s)))
return;
BEAST_EXPECT(counter == 3);
runner->join();
BEAST_EXPECT(!runner->runnable());
}
/**
* CoroTask<T> returns a value via co_return. Outer coroutine
* extracts it with co_await.
*/
void
testValueReturn()
{
using namespace std::chrono_literals;
using namespace jtx;
testcase("value return");
Env env(*this, envconfig([](std::unique_ptr<Config> cfg) {
cfg->forceMultiThread = true;
return cfg;
}));
Gate g;
int result = 0;
auto runner = env.app().getJobQueue().postCoroTask(
JtClient, "CoroTaskTest", [rp = &result, gp = &g](auto) -> CoroTask<void> {
auto inner = []() -> CoroTask<int> { co_return 42; };
*rp = co_await inner();
gp->signal();
co_return;
});
BEAST_EXPECT(runner);
if (!BEAST_EXPECT(g.waitFor(5s)))
return;
runner->join();
BEAST_EXPECT(result == 42);
BEAST_EXPECT(!runner->runnable());
}
/**
* CoroTask<T> propagates exceptions from inner coroutines.
* Outer coroutine catches via try/catch around co_await.
*/
void
testValueException()
{
using namespace std::chrono_literals;
using namespace jtx;
testcase("value exception");
Env env(*this, envconfig([](std::unique_ptr<Config> cfg) {
cfg->forceMultiThread = true;
return cfg;
}));
Gate g;
std::string what;
auto runner = env.app().getJobQueue().postCoroTask(
JtClient, "CoroTaskTest", [wp = &what, gp = &g](auto) -> CoroTask<void> {
auto inner = []() -> CoroTask<int> {
throw std::runtime_error("inner error");
co_return 0;
};
try
{
co_await inner();
}
catch (std::runtime_error const& e)
{
*wp = e.what();
}
gp->signal();
co_return;
});
BEAST_EXPECT(runner);
if (!BEAST_EXPECT(g.waitFor(5s)))
return;
runner->join();
BEAST_EXPECT(what == "inner error");
BEAST_EXPECT(!runner->runnable());
}
/**
* CoroTask<T> chaining. Nested value-returning coroutines
* compose via co_await.
*/
void
testValueChaining()
{
using namespace std::chrono_literals;
using namespace jtx;
testcase("value chaining");
Env env(*this, envconfig([](std::unique_ptr<Config> cfg) {
cfg->forceMultiThread = true;
return cfg;
}));
Gate g;
int result = 0;
auto runner = env.app().getJobQueue().postCoroTask(
JtClient, "CoroTaskTest", [rp = &result, gp = &g](auto) -> CoroTask<void> {
auto add = [](int a, int b) -> CoroTask<int> { co_return a + b; };
auto mul = [add](int a, int b) -> CoroTask<int> {
int const sum = co_await add(a, b);
co_return sum * 2;
};
*rp = co_await mul(3, 4);
gp->signal();
co_return;
});
BEAST_EXPECT(runner);
if (!BEAST_EXPECT(g.waitFor(5s)))
return;
runner->join();
BEAST_EXPECT(result == 14); // (3 + 4) * 2
BEAST_EXPECT(!runner->runnable());
}
/**
* postCoroTask returns nullptr when JobQueue is stopping.
*/
void
testShutdownRejection()
{
using namespace std::chrono_literals;
using namespace jtx;
testcase("shutdown rejection");
Env env(*this, envconfig([](std::unique_ptr<Config> cfg) {
cfg->forceMultiThread = true;
return cfg;
}));
// Stop the JobQueue
env.app().getJobQueue().stop();
auto runner = env.app().getJobQueue().postCoroTask(
JtClient, "CoroTaskTest", [](auto) -> CoroTask<void> { co_return; });
BEAST_EXPECT(!runner);
}
void
run() override
{
testVoidCompletion();
testCorrectOrder();
testIncorrectOrder();
testJobQueueAwaiter();
testThreadSpecificStorage();
testExceptionPropagation();
testMultipleYields();
testValueReturn();
testValueException();
testValueChaining();
testShutdownRejection();
}
};
BEAST_DEFINE_TESTSUITE(CoroTask, core, xrpl);
} // namespace xrpl::test
// NOLINTEND(cppcoreguidelines-avoid-capturing-lambda-coroutines)

View File

@@ -6,6 +6,7 @@
#include <xrpl/basics/LocalValue.h>
#include <xrpl/beast/unit_test/suite.h>
#include <xrpl/core/CoroTask.h>
#include <xrpl/core/Job.h>
#include <xrpl/core/JobQueue.h>
@@ -49,6 +50,11 @@ public:
}
};
// NOTE: All coroutine lambdas passed to postCoroTask use explicit
// pointer-by-value captures instead of [&] to work around a GCC 14
// bug where reference captures in coroutine lambdas are corrupted
// in the coroutine frame.
void
correctOrder()
{
@@ -63,14 +69,22 @@ public:
}));
Gate g1, g2;
std::shared_ptr<JobQueue::Coro> c;
env.app().getJobQueue().postCoro(JtClient, "CoroTest", [&](auto const& cr) {
c = cr;
g1.signal();
c->yield();
g2.signal();
});
BEAST_EXPECT(g1.waitFor(5s));
std::shared_ptr<JobQueue::CoroTaskRunner> c;
env.app().getJobQueue().postCoroTask(
JtClient,
"CoroTest",
// Safe capture: the test blocks on the Gates until the coroutine
// completes, so the captured pointers outlive the coroutine.
// NOLINTNEXTLINE(cppcoreguidelines-avoid-capturing-lambda-coroutines)
[cp = &c, g1p = &g1, g2p = &g2](auto runner) -> CoroTask<void> {
*cp = runner;
g1p->signal();
co_await runner->suspend();
g2p->signal();
co_return;
});
if (!BEAST_EXPECT(g1.waitFor(5s)))
return;
c->join();
c->post();
BEAST_EXPECT(g2.waitFor(5s));
@@ -90,11 +104,22 @@ public:
}));
Gate g;
env.app().getJobQueue().postCoro(JtClient, "CoroTest", [&](auto const& c) {
c->post();
c->yield();
g.signal();
});
env.app().getJobQueue().postCoroTask(
JtClient,
"CoroTest",
// Safe capture: the test blocks on the Gate until the coroutine
// completes, so the captured pointer outlives the coroutine.
// NOLINTNEXTLINE(cppcoreguidelines-avoid-capturing-lambda-coroutines)
[gp = &g](auto runner) -> CoroTask<void> {
// Schedule a resume before suspending. The posted job
// cannot actually call resume() until the current resume()
// releases CoroTaskRunner::mutex_, which only happens after
// the coroutine suspends at co_await.
runner->post();
co_await runner->suspend();
gp->signal();
co_return;
});
BEAST_EXPECT(g.waitFor(5s));
}
@@ -110,7 +135,7 @@ public:
auto& jq = env.app().getJobQueue();
static int const kN = 4;
std::array<std::shared_ptr<JobQueue::Coro>, kN> a;
std::array<std::shared_ptr<JobQueue::CoroTaskRunner>, kN> a;
LocalValue<int> lv(-1);
BEAST_EXPECT(*lv == -1);
@@ -127,26 +152,36 @@ public:
for (int i = 0; i < kN; ++i)
{
jq.postCoro(JtClient, "CoroTest", [&, id = i](auto const& c) {
a[id] = c;
g.signal();
c->yield();
jq.postCoroTask(
JtClient,
"CoroTest",
// Safe capture: the test drives every coroutine to completion
// via the Gate/post()/join() sequences below, so the captured
// pointers outlive the coroutines.
// NOLINTNEXTLINE(cppcoreguidelines-avoid-capturing-lambda-coroutines)
[this, ap = &a, gp = &g, lvp = &lv, id = i](auto runner) -> CoroTask<void> {
(*ap)[id] = runner;
gp->signal();
co_await runner->suspend();
this->BEAST_EXPECT(*lv == -1);
*lv = id;
this->BEAST_EXPECT(*lv == id);
g.signal();
c->yield();
this->BEAST_EXPECT(**lvp == -1);
**lvp = id;
this->BEAST_EXPECT(**lvp == id);
gp->signal();
co_await runner->suspend();
this->BEAST_EXPECT(*lv == id);
});
BEAST_EXPECT(g.waitFor(5s));
this->BEAST_EXPECT(**lvp == id);
co_return;
});
if (!BEAST_EXPECT(g.waitFor(5s)))
return;
a[i]->join();
}
for (auto const& c : a)
{
c->post();
BEAST_EXPECT(g.waitFor(5s));
if (!BEAST_EXPECT(g.waitFor(5s)))
return;
c->join();
}
for (auto const& c : a)

View File

@@ -1,11 +1,11 @@
#include <test/jtx/Env.h>
#include <xrpl/beast/unit_test/suite.h>
#include <xrpl/core/CoroTask.h>
#include <xrpl/core/Job.h>
#include <xrpl/core/JobQueue.h>
#include <atomic>
#include <memory>
namespace xrpl::test {
@@ -46,87 +46,114 @@ class JobQueue_test : public beast::unit_test::Suite
}
}
// NOTE: All coroutine lambdas passed to postCoroTask use explicit
// pointer-by-value captures instead of [&] to work around a GCC 14
// bug where reference captures in coroutine lambdas are corrupted
// in the coroutine frame.
void
testPostCoro()
testPostCoroTask()
{
jtx::Env env{*this};
JobQueue& jQueue = env.app().getJobQueue();
{
// Test repeated post()s until the Coro completes.
// Test repeated post()s until the coroutine completes.
std::atomic<int> yieldCount{0};
auto const coro = jQueue.postCoro(
auto const runner = jQueue.postCoroTask(
JtClient,
"PostCoroTest1",
[&yieldCount](std::shared_ptr<JobQueue::Coro> const& coroCopy) {
while (++yieldCount < 4)
coroCopy->yield();
// Safe capture: the test blocks below until the coroutine
// completes, so the captured pointer outlives the coroutine.
// NOLINTNEXTLINE(cppcoreguidelines-avoid-capturing-lambda-coroutines)
[ycp = &yieldCount](auto runner) -> CoroTask<void> {
while (++(*ycp) < 4)
co_await runner->suspend();
co_return;
});
BEAST_EXPECT(coro != nullptr);
BEAST_EXPECT(runner != nullptr);
// Wait for the Job to run and yield.
while (yieldCount == 0)
;
// Now re-post until the Coro says it is done.
// Now re-post until the CoroTaskRunner says it is done.
int old = yieldCount;
while (coro->runnable())
while (runner->runnable())
{
BEAST_EXPECT(coro->post());
BEAST_EXPECT(runner->post());
while (old == yieldCount)
{
}
coro->join();
runner->join();
BEAST_EXPECT(++old == yieldCount);
}
BEAST_EXPECT(yieldCount == 4);
}
{
// Test repeated resume()s until the Coro completes.
// Test repeated post()+join()s until the coroutine completes.
int yieldCount{0};
auto const coro = jQueue.postCoro(
auto const runner = jQueue.postCoroTask(
JtClient,
"PostCoroTest2",
[&yieldCount](std::shared_ptr<JobQueue::Coro> const& coroCopy) {
while (++yieldCount < 4)
coroCopy->yield();
// Safe capture: the test blocks below until the coroutine
// completes, so the captured pointer outlives the coroutine.
// NOLINTNEXTLINE(cppcoreguidelines-avoid-capturing-lambda-coroutines)
[ycp = &yieldCount](auto runner) -> CoroTask<void> {
while (++(*ycp) < 4)
co_await runner->suspend();
co_return;
});
if (!coro)
if (!runner)
{
// There's no good reason we should not get a Coro, but we
// There's no good reason we should not get a runner, but we
// can't continue without one.
BEAST_EXPECT(false);
return;
}
// Wait for the Job to run and yield.
coro->join();
runner->join();
// Now resume until the Coro says it is done.
// Now post()+join() until the CoroTaskRunner says it is done.
// resume() requires a prior post() (see the precondition on
// CoroTaskRunner::resume()), so the posted job performs the
// resume and join() blocks until it completes. yieldCount is
// deliberately not atomic: the mutexRun_ handoff inside join()
// must provide the happens-before edge that makes the
// increment visible to this thread.
int old = yieldCount;
while (coro->runnable())
while (runner->runnable())
{
coro->resume(); // Resume runs synchronously on this thread.
BEAST_EXPECT(runner->post());
runner->join();
BEAST_EXPECT(++old == yieldCount);
}
BEAST_EXPECT(yieldCount == 4);
}
{
// If the JobQueue is stopped, we should no
// longer be able to add a Coro (and calling postCoro() should
// return false).
// longer be able to post a coroutine (and calling postCoroTask()
// should return nullptr).
using namespace std::chrono_literals;
jQueue.stop();
// The Coro should never run, so having the Coro access this
// The coroutine should never run, so having it access this
// unprotected variable on the stack should be completely safe.
// Not recommended for the faint of heart...
bool unprotected = false;
auto const coro = jQueue.postCoro(
JtClient, "PostCoroTest3", [&unprotected](std::shared_ptr<JobQueue::Coro> const&) {
unprotected = false;
auto const runner = jQueue.postCoroTask(
JtClient,
"PostCoroTest3",
// Safe capture: the JobQueue is stopped, so the coroutine is
// never started and the captured pointer is never dereferenced.
// NOLINTNEXTLINE(cppcoreguidelines-avoid-capturing-lambda-coroutines)
[up = &unprotected](auto) -> CoroTask<void> {
*up = true;
co_return;
});
BEAST_EXPECT(coro == nullptr);
BEAST_EXPECT(runner == nullptr);
BEAST_EXPECT(unprotected == false);
}
}
@@ -135,7 +162,7 @@ public:
run() override
{
testAddJob();
testPostCoro();
testPostCoroTask();
}
};

View File

@@ -23,7 +23,9 @@
#include <xrpl/basics/strHex.h>
#include <xrpl/beast/unit_test/suite.h>
#include <xrpl/beast/utility/instrumentation.h>
#include <xrpl/core/CoroTask.h>
#include <xrpl/core/Job.h>
#include <xrpl/core/JobQueue.h>
#include <xrpl/core/ServiceRegistry.h>
#include <xrpl/json/json_value.h>
#include <xrpl/json/to_string.h>
@@ -223,7 +225,6 @@ findPathsRequest(
.ledgerMaster = app.getLedgerMaster(),
.consumer = c,
.role = Role::USER,
.coro = {},
.infoSub = {},
.apiVersion = rpc::kApiVersionIfUnspecified},
{},
@@ -250,11 +251,14 @@ findPathsRequest(
json::Value result;
Gate g;
app.getJobQueue().postCoro(JtClient, "RPC-Client", [&](auto const& coro) {
// Safe capture: the caller blocks on g.waitFor() until the coroutine
// completes, so the captured locals outlive the coroutine.
// NOLINTNEXTLINE(cppcoreguidelines-avoid-capturing-lambda-coroutines)
app.getJobQueue().postCoroTask(JtClient, "RPC-Client", [&](auto) -> CoroTask<void> {
context.params = std::move(params);
context.coro = coro;
rpc::doCommand(context, result);
g.signal();
co_return;
});
using namespace std::chrono_literals;

View File

@@ -1483,7 +1483,6 @@ ApplicationImp::setup(boost::program_options::variables_map const& cmdline)
.ledgerMaster = getLedgerMaster(),
.consumer = c,
.role = Role::ADMIN,
.coro = {},
.infoSub = {},
.apiVersion = rpc::kApiMaximumSupportedVersion},
jvCommand};

View File

@@ -17,6 +17,7 @@
#include <xrpl/beast/utility/instrumentation.h>
#include <xrpl/config/BasicConfig.h>
#include <xrpl/config/Constants.h>
#include <xrpl/core/CoroTask.h>
#include <xrpl/core/Job.h>
#include <xrpl/core/JobQueue.h>
#include <xrpl/protocol/ErrorCodes.h>
@@ -147,13 +148,19 @@ GRPCServerImpl::CallData<Request, Response>::process()
// ensures that finished is always true when this CallData object
// is returned as a tag in handleRpcs(), after sending the response
finished_ = true;
auto coro = app_.getJobQueue().postCoro(
JobType::JtRpc, "gRPC-Client", [thisShared](std::shared_ptr<JobQueue::Coro> coro) {
thisShared->process(coro);
auto runner = app_.getJobQueue().postCoroTask(
JobType::JtRpc,
"gRPC-Client",
// Safe capture: postCoroTask heap-allocates the lambda (FuncStore)
// and thisShared keeps the CallData alive until the coroutine ends.
// NOLINTNEXTLINE(cppcoreguidelines-avoid-capturing-lambda-coroutines)
[thisShared](auto) -> CoroTask<void> {
thisShared->processRequest();
co_return;
});
// If coro is null, then the JobQueue has already been shutdown
if (!coro)
// If runner is null, then the JobQueue has already been shutdown
if (!runner)
{
grpc::Status const status{grpc::StatusCode::INTERNAL, "Job Queue is already stopped"};
responder_.FinishWithError(status, this);
@@ -162,7 +169,7 @@ GRPCServerImpl::CallData<Request, Response>::process()
template <class Request, class Response>
void
GRPCServerImpl::CallData<Request, Response>::process(std::shared_ptr<JobQueue::Coro> coro)
GRPCServerImpl::CallData<Request, Response>::processRequest()
{
try
{
@@ -204,7 +211,6 @@ GRPCServerImpl::CallData<Request, Response>::process(std::shared_ptr<JobQueue::C
app_.getLedgerMaster(),
usage,
role,
coro,
InfoSub::pointer(),
kApiVersion},
request_};

View File

@@ -6,7 +6,6 @@
#include <xrpld/rpc/detail/Handler.h>
#include <xrpl/beast/utility/Journal.h>
#include <xrpl/core/JobQueue.h>
#include <xrpl/proto/org/xrpl/rpc/v1/xrp_ledger.grpc.pb.h>
#include <xrpl/resource/Charge.h>
#include <xrpl/resource/Consumer.h>
@@ -228,9 +227,12 @@ private:
clone() override;
private:
// process the request. Called inside the coroutine passed to JobQueue
/**
* Process the gRPC request. Called inside the CoroTask lambda
* posted to the JobQueue by process().
*/
void
process(std::shared_ptr<JobQueue::Coro> coro);
processRequest();
// return load type of this RPC
resource::Charge

View File

@@ -3,13 +3,11 @@
#include <xrpld/rpc/Role.h>
#include <xrpl/beast/utility/Journal.h>
#include <xrpl/core/JobQueue.h>
#include <xrpl/json/json_value.h>
#include <xrpl/resource/Charge.h>
#include <xrpl/resource/Consumer.h>
#include <xrpl/server/InfoSub.h>
#include <memory>
#include <string_view>
namespace xrpl {
@@ -32,7 +30,6 @@ struct Context
LedgerMaster& ledgerMaster;
resource::Consumer& consumer;
Role role;
std::shared_ptr<JobQueue::Coro> coro;
InfoSub::pointer infoSub;
unsigned int apiVersion;
};

View File

@@ -184,13 +184,10 @@ public:
private:
json::Value
processSession(
std::shared_ptr<WSSession> const& session,
std::shared_ptr<JobQueue::Coro> const& coro,
json::Value const& jv);
processSession(std::shared_ptr<WSSession> const& session, json::Value const& jv);
void
processSession(std::shared_ptr<Session> const&, std::shared_ptr<JobQueue::Coro> coro);
processSession(std::shared_ptr<Session> const&);
void
processRequest(
@@ -198,7 +195,6 @@ private:
std::string const& request,
beast::ip::Endpoint const& remoteIPAddress,
Output const&,
std::shared_ptr<JobQueue::Coro> coro,
std::string_view forwardedFor,
std::string_view user);

View File

@@ -6,6 +6,7 @@
#include <xrpld/rpc/detail/PathRequest.h>
#include <xrpl/basics/Log.h>
#include <xrpl/basics/scope.h>
#include <xrpl/core/Job.h>
#include <xrpl/core/JobQueue.h>
#include <xrpl/json/json_value.h>
@@ -137,9 +138,12 @@ PathRequestManager::updateAll(std::shared_ptr<ReadView const> const& inLedger)
}
else if (request->hasCompletion())
{
// One-shot request with completion function
// One-shot request with completion function. Fire the
// completion even if doUpdate throws, so the RPC
// handler blocked in doRipplePathFind is released
// immediately instead of waiting out its timeout.
ScopeExit const complete{[&request] { request->updateComplete(); }};
request->doUpdate(cache, false);
request->updateComplete();
++processed;
}
}

View File

@@ -17,6 +17,7 @@
#include <xrpl/beast/rfc2616.h>
#include <xrpl/beast/utility/Journal.h>
#include <xrpl/config/Constants.h>
#include <xrpl/core/CoroTask.h>
#include <xrpl/core/Job.h>
#include <xrpl/core/JobQueue.h>
#include <xrpl/json/Output.h>
@@ -317,9 +318,23 @@ ServerHandler::onRequest(Session& session)
}
std::shared_ptr<Session> const detachedSession = session.detach();
auto const postResult = jobQueue_.postCoro(
JtClientRpc, "RPC-Client", [this, detachedSession](std::shared_ptr<JobQueue::Coro> coro) {
processSession(detachedSession, coro);
auto const postResult = jobQueue_.postCoroTask(
JtClientRpc,
"RPC-Client",
// Safe capture: postCoroTask heap-allocates the lambda (FuncStore),
// detachedSession is captured by value, and this (the ServerHandler)
// outlives the JobQueue jobs.
// NOLINTNEXTLINE(cppcoreguidelines-avoid-capturing-lambda-coroutines)
[this, detachedSession](auto) -> CoroTask<void> {
try
{
processSession(detachedSession);
}
catch (std::exception const& e)
{
JLOG(journal_.error()) << "RPC-Client coroutine exception: " << e.what();
}
co_return;
});
if (postResult == nullptr)
{
@@ -355,17 +370,30 @@ ServerHandler::onWSMessage(
JLOG(journal_.trace()) << "Websocket received '" << jv << "'";
auto const postResult = jobQueue_.postCoro(
auto const postResult = jobQueue_.postCoroTask(
JtClientWebsocket,
"WS-Client",
[this, session, jv = std::move(jv)](std::shared_ptr<JobQueue::Coro> const& coro) {
auto const jr = this->processSession(session, coro, jv);
auto const s = to_string(jr);
auto const n = s.length();
boost::beast::multi_buffer sb(n);
sb.commit(boost::asio::buffer_copy(sb.prepare(n), boost::asio::buffer(s.c_str(), n)));
session->send(std::make_shared<StreambufWSMsg<decltype(sb)>>(std::move(sb)));
session->complete();
// Safe capture: postCoroTask heap-allocates the lambda (FuncStore),
// session and jv are captured by value, and this (the ServerHandler)
// outlives the JobQueue jobs.
// NOLINTNEXTLINE(cppcoreguidelines-avoid-capturing-lambda-coroutines)
[this, session, jv = std::move(jv)](auto) -> CoroTask<void> {
try
{
auto const jr = this->processSession(session, jv);
auto const s = to_string(jr);
auto const n = s.length();
boost::beast::multi_buffer sb(n);
sb.commit(
boost::asio::buffer_copy(sb.prepare(n), boost::asio::buffer(s.c_str(), n)));
session->send(std::make_shared<StreambufWSMsg<decltype(sb)>>(std::move(sb)));
session->complete();
}
catch (std::exception const& e)
{
JLOG(journal_.error()) << "WS-Client coroutine exception: " << e.what();
}
co_return;
});
if (postResult == nullptr)
{
@@ -410,10 +438,7 @@ logDuration(json::Value const& request, T const& duration, beast::Journal& journ
}
json::Value
ServerHandler::processSession(
std::shared_ptr<WSSession> const& session,
std::shared_ptr<JobQueue::Coro> const& coro,
json::Value const& jv)
ServerHandler::processSession(std::shared_ptr<WSSession> const& session, json::Value const& jv)
{
auto is = std::static_pointer_cast<WSInfoSub>(session->appDefined);
if (is->getConsumer().disconnect(journal_))
@@ -480,7 +505,6 @@ ServerHandler::processSession(
.ledgerMaster = app_.getLedgerMaster(),
.consumer = is->getConsumer(),
.role = role,
.coro = coro,
.infoSub = is,
.apiVersion = apiVersion},
jv,
@@ -551,18 +575,14 @@ ServerHandler::processSession(
return jr;
}
// Run as a coroutine.
void
ServerHandler::processSession(
std::shared_ptr<Session> const& session,
std::shared_ptr<JobQueue::Coro> coro)
ServerHandler::processSession(std::shared_ptr<Session> const& session)
{
processRequest(
session->port(),
buffersToString(session->request().body().data()),
session->remoteAddress().atPort(0),
makeOutput(*session),
coro,
forwardedFor(session->request()),
[&] -> std::string_view {
auto const iter = session->request().find("X-User");
@@ -603,7 +623,6 @@ ServerHandler::processRequest(
std::string const& request,
beast::ip::Endpoint const& remoteIPAddress,
Output const& output,
std::shared_ptr<JobQueue::Coro> coro,
std::string_view forwardedFor,
std::string_view user)
{
@@ -863,7 +882,6 @@ ServerHandler::processRequest(
.ledgerMaster = app_.getLedgerMaster(),
.consumer = usage,
.role = role,
.coro = coro,
.infoSub = InfoSub::pointer(),
.apiVersion = apiVersion},
params,

View File

@@ -65,6 +65,12 @@ static constexpr int kDefaultAutoFillFeeMultiplier = 10;
static constexpr int kDefaultAutoFillFeeDivisor = 1;
static constexpr int kMaxPathfindsInProgress = 2;
static constexpr int kMaxPathfindJobCount = 50;
/**
* How long a blocking ripple_path_find call waits for the path-finding
* continuation before giving up and returning an internal error.
*/
constexpr auto kPathfindCompletionTimeout = std::chrono::seconds{30};
static constexpr int kMaxJobQueueClients = 500;
constexpr auto kMaxValidatedLedgerAge = std::chrono::minutes{2};
static constexpr int kMaxRequestSize = 1000000;

View File

@@ -7,14 +7,15 @@
#include <xrpld/rpc/detail/RPCLedgerHelpers.h>
#include <xrpld/rpc/detail/Tuning.h>
#include <xrpl/core/JobQueue.h>
#include <xrpl/json/json_value.h>
#include <xrpl/protocol/ErrorCodes.h>
#include <xrpl/protocol/RPCErr.h>
#include <xrpl/protocol/jss.h>
#include <xrpl/resource/Fees.h>
#include <condition_variable>
#include <memory>
#include <mutex>
#include <utility>
namespace xrpl {
@@ -47,98 +48,62 @@ doRipplePathFind(rpc::JsonContext& context)
PathRequest::pointer request;
lpLedger = context.ledgerMaster.getClosedLedger();
// It doesn't look like there's much odd happening here, but you should
// be aware this code runs in a JobQueue::Coro, which is a coroutine.
// And we may be flipping around between threads. Here's an overview:
// The wait below parks this JobQueue worker thread until the
// path-finding continuation fires. The continuation is fired by a
// JtUpdatePf job, which itself needs a free worker to run. Bound
// the number of concurrently parked workers (LegacyPathFind admits
// at most kMaxPathfindsInProgress non-admin requests) so that
// concurrent ripple_path_find calls cannot occupy every worker and
// stall the whole JobQueue. The guard must stay in scope until the
// wait completes. (The old Boost.Coroutine implementation did not
// need this: it suspended and released the worker instead of
// blocking it.)
rpc::LegacyPathFind const lpf(isUnlimited(context.role), context.app);
if (!lpf.isOk())
return rpcError(RpcTooBusy);
// makeLegacyPathRequest enqueues a path-finding job that runs
// asynchronously. We block this thread with a condition_variable
// until the path-finding continuation signals completion.
// If makeLegacyPathRequest cannot schedule the job (e.g. during
// shutdown), it returns an empty request and we skip the wait.
// Replaces the old Coro yield/resume pattern with synchronous
// blocking, eliminating shutdown race conditions.
//
// 1. We're running doRipplePathFind() due to a call to
// ripple_path_find. doRipplePathFind() is currently running
// inside of a JobQueue::Coro using a JobQueue thread.
//
// 2. doRipplePathFind's call to makeLegacyPathRequest() enqueues the
// path-finding request. That request will (probably) run at some
// indeterminate future time on a (probably different) JobQueue
// thread.
//
// 3. As a continuation from that path-finding JobQueue thread, the
// coroutine we're currently running in (!) is posted to the
// JobQueue. Because it is a continuation, that post won't
// happen until the path-finding request completes.
//
// 4. Once the continuation is enqueued, and we have reason to think
// the path-finding job is likely to run, then the coroutine we're
// running in yield()s. That means it surrenders its thread in
// the JobQueue. The coroutine is suspended, but ready to run,
// because it is kept resident by a shared_ptr in the
// path-finding continuation.
//
// 5. If all goes well then path-finding runs on a JobQueue thread
// and executes its continuation. The continuation posts this
// same coroutine (!) to the JobQueue.
//
// 6. When the JobQueue calls this coroutine, this coroutine resumes
// from the line below the coro->yield() and returns the
// path-finding result.
//
// With so many moving parts, what could go wrong?
//
// Just in terms of the JobQueue refusing to add jobs at shutdown
// there are two specific things that can go wrong.
//
// 1. The path-finding Job queued by makeLegacyPathRequest() might be
// rejected (because we're shutting down).
//
// Fortunately this problem can be addressed by looking at the
// return value of makeLegacyPathRequest(). If
// makeLegacyPathRequest() cannot get a thread to run the path-find
// on, then it returns an empty request.
//
// 2. The path-finding job might run, but the Coro::post() might be
// rejected by the JobQueue (because we're shutting down).
//
// We handle this case by resuming (not posting) the Coro.
// By resuming the Coro, we allow the Coro to run to completion
// on the current thread instead of requiring that it run on a
// new thread from the JobQueue.
//
// Both of these failure modes are hard to recreate in a unit test
// because they are so dependent on inter-thread timing. However
// the failure modes can be observed by synchronously (inside the
// xrpld source code) shutting down the application. The code to
// do so looks like this:
//
// context.app.signalStop();
// while (! context.app.getJobQueue().jobCounter().joined()) { }
//
// The first line starts the process of shutting down the app.
// The second line waits until no more jobs can be added to the
// JobQueue before letting the thread continue.
//
// May 2017
// The state is shared with the continuation so that it stays alive
// even if we stop waiting before the continuation runs.
struct PathDone
{
std::mutex mtx;
std::condition_variable cv;
bool done = false;
};
auto const state = std::make_shared<PathDone>();
jvResult = context.app.getPathRequestManager().makeLegacyPathRequest(
request,
[&context]() {
// Copying the shared_ptr keeps the coroutine alive up
// through the return. Otherwise the storage under the
// captured reference could evaporate when we return from
// coroCopy->resume(). This is not strictly necessary, but
// will make maintenance easier.
std::shared_ptr<JobQueue::Coro> const coroCopy{context.coro};
if (!coroCopy->post())
[state]() {
{
// The post() failed, so we won't get a thread to let
// the Coro finish. We'll call Coro::resume() so the
// Coro can finish on our thread. Otherwise the
// application will hang on shutdown.
coroCopy->resume();
std::scoped_lock const lk(state->mtx);
state->done = true;
}
state->cv.notify_one();
},
context.consumer,
lpLedger,
context.params);
if (request)
{
context.coro->yield();
std::unique_lock lk(state->mtx);
if (!state->cv.wait_for(
lk, rpc::tuning::kPathfindCompletionTimeout, [&state] { return state->done; }))
{
// Path-finding continuation never fired (e.g. shutdown
// race or unexpected failure). Return an internal error
// rather than blocking the RPC thread indefinitely.
return rpcError(RpcInternal);
}
lk.unlock();
jvResult = request->doStatus(context.params);
}