관리-도구
편집 파일: TThreadStateDestroy.cpp
/* -*- indent-tabs-mode: nil; tab-width: 4; -*- */ /** * Implementation of the ThreadState destructors. * * Format with: * clang-format -i --style=file src/greenlet/greenlet.c * * * Fix missing braces with: * clang-tidy src/greenlet/greenlet.c -fix -checks="readability-braces-around-statements" */ #ifndef T_THREADSTATE_DESTROY #define T_THREADSTATE_DESTROY #include "TGreenlet.hpp" #include "greenlet_thread_support.hpp" #include "greenlet_cpython_add_pending.hpp" #include "greenlet_compiler_compat.hpp" #include "TGreenletGlobals.cpp" #include "TThreadState.hpp" #include "TThreadStateCreator.hpp" namespace greenlet { extern "C" { struct ThreadState_DestroyNoGIL { /** This function uses the same lock that the PendingCallback does */ static void MarkGreenletDeadAndQueueCleanup(ThreadState* const state) { #if GREENLET_BROKEN_THREAD_LOCAL_CLEANUP_JUST_LEAK return; #endif // We are *NOT* holding the GIL. Our thread is in the middle // of its death throes and the Python thread state is already // gone so we can't use most Python APIs. One that is safe is // ``Py_AddPendingCall``, unless the interpreter itself has // been torn down. There is a limited number of calls that can // be queued: 32 (NPENDINGCALLS) in CPython 3.10, so we // coalesce these calls using our own queue. if (!MarkGreenletDeadIfNeeded(state)) { // No state, or no greenlet return; } // XXX: Because we don't have the GIL, this is a race condition. if (!PyInterpreterState_Head()) { // We have to leak the thread state, if the // interpreter has shut down when we're getting // deallocated, we can't run the cleanup code that // deleting it would imply. return; } AddToCleanupQueue(state); } private: // If the state has an allocated main greenlet: // - mark the greenlet as dead by disassociating it from the state; // - return 1 // Otherwise, return 0. static bool MarkGreenletDeadIfNeeded(ThreadState* const state) { if (state && state->has_main_greenlet()) { // mark the thread as dead ASAP. // this is racy! If we try to throw or switch to a // greenlet from this thread from some other thread before // we clear the state pointer, it won't realize the state // is dead which can crash the process. PyGreenlet* p(state->borrow_main_greenlet().borrow()); assert(p->pimpl->thread_state() == state || p->pimpl->thread_state() == nullptr); dynamic_cast<MainGreenlet*>(p->pimpl)->thread_state(nullptr); return true; } return false; } static void AddToCleanupQueue(ThreadState* const state) { assert(state && state->has_main_greenlet()); // NOTE: Because we're not holding the GIL here, some other // Python thread could run and call ``os.fork()``, which would // be bad if that happened while we are holding the cleanup // lock (it wouldn't function in the child process). // Make a best effort to try to keep the duration we hold the // lock short. // TODO: On platforms that support it, use ``pthread_atfork`` to // drop this lock. LockGuard cleanup_lock(*mod_globs->thread_states_to_destroy_lock); mod_globs->queue_to_destroy(state); if (mod_globs->thread_states_to_destroy.size() == 1) { // We added the first item to the queue. We need to schedule // the cleanup. // A size greater than 1 means that we have already added the pending call, // and in fact, it may be executing now. // If it is executing, our lock makes sure that it will see the item we just added // to the queue on its next iteration (after we release the lock) // // A size of 1 means there is no pending call, OR the pending call is // currently executing, has dropped the lock, and is deleting the last item // from the queue; its next iteration will go ahead and delete the item we just added. // And the pending call we schedule here will have no work to do. int result = AddPendingCall( PendingCallback_DestroyQueueWithGIL, nullptr); if (result < 0) { // Hmm, what can we do here? fprintf(stderr, "greenlet: WARNING: failed in call to Py_AddPendingCall; " "expect a memory leak.\n"); } } } static int PendingCallback_DestroyQueueWithGIL(void* UNUSED(arg)) { // We're holding the GIL here, so no Python code should be able to // run to call ``os.fork()``. while (1) { ThreadState* to_destroy; { LockGuard cleanup_lock(*mod_globs->thread_states_to_destroy_lock); if (mod_globs->thread_states_to_destroy.empty()) { break; } to_destroy = mod_globs->take_next_to_destroy(); } assert(to_destroy); assert(to_destroy->has_main_greenlet()); // Drop the lock while we do the actual deletion. // This allows other calls to MarkGreenletDeadAndQueueCleanup // to enter and add to our queue. DestroyOneWithGIL(to_destroy); } return 0; } static void DestroyOneWithGIL(const ThreadState* const state) { // Holding the GIL. // Passed a non-shared pointer to the actual thread state. // state -> main greenlet assert(state->has_main_greenlet()); PyGreenlet* main(state->borrow_main_greenlet()); // When we need to do cross-thread operations, we check this. // A NULL value means the thread died some time ago. // We do this here, rather than in a Python dealloc function // for the greenlet, in case there's still a reference out // there. dynamic_cast<MainGreenlet*>(main->pimpl)->thread_state(nullptr); delete state; // Deleting this runs the destructor, DECREFs the main greenlet. } // ensure this is actually defined. static_assert(GREENLET_BROKEN_PY_ADD_PENDING == 1 || GREENLET_BROKEN_PY_ADD_PENDING == 0, "GREENLET_BROKEN_PY_ADD_PENDING not defined correctly."); #if GREENLET_BROKEN_PY_ADD_PENDING static int _push_pending_call(struct _pending_calls *pending, int (*func)(void *), void *arg) { int i = pending->last; int j = (i + 1) % NPENDINGCALLS; if (j == pending->first) { return -1; /* Queue full */ } pending->calls[i].func = func; pending->calls[i].arg = arg; pending->last = j; return 0; } static int AddPendingCall(int (*func)(void *), void *arg) { _PyRuntimeState *runtime = &_PyRuntime; if (!runtime) { // obviously impossible return 0; } struct _pending_calls *pending = &runtime->ceval.pending; if (!pending->lock) { return 0; } int result = 0; PyThread_acquire_lock(pending->lock, WAIT_LOCK); if (!pending->finishing) { result = _push_pending_call(pending, func, arg); } PyThread_release_lock(pending->lock); SIGNAL_PENDING_CALLS(&runtime->ceval); return result; } #else // Python < 3.8 or >= 3.9 static int AddPendingCall(int (*func)(void*), void* arg) { // If the interpreter is in the middle of finalizing, we can't add a // pending call. Trying to do so will end up in a SIGSEGV, as // Py_AddPendingCall will not be able to get the interpreter and will // try to dereference a NULL pointer. It's possible this can still // segfault if we happen to get context switched, and maybe we should // just always implement our own AddPendingCall, but I'd like to see if // this works first #if GREENLET_PY313 if (Py_IsFinalizing()) { #else if (_Py_IsFinalizing()) { #endif #ifdef GREENLET_DEBUG // No need to log in the general case. Yes, we'll leak, // but we're shutting down so it should be ok. fprintf(stderr, "greenlet: WARNING: Interpreter is finalizing. Ignoring " "call to Py_AddPendingCall; \n"); #endif return 0; } return Py_AddPendingCall(func, arg); } #endif }; }; }; // namespace greenlet // The intent when GET_THREAD_STATE() is needed multiple times in a // function is to take a reference to its return value in a local // variable, to avoid the thread-local indirection. On some platforms // (macOS), accessing a thread-local involves a function call (plus an // initial function call in each function that uses a thread local); // in contrast, static volatile variables are at some pre-computed // offset. typedef greenlet::ThreadStateCreator<greenlet::ThreadState_DestroyNoGIL::MarkGreenletDeadAndQueueCleanup> ThreadStateCreator; static thread_local ThreadStateCreator g_thread_state_global; #define GET_THREAD_STATE() g_thread_state_global #endif //T_THREADSTATE_DESTROY