ruby/thread_pthread.c

/* -*-c-*- */
/**********************************************************************

  thread_pthread.c -

  $Author$

  Copyright (C) 2004-2007 Koichi Sasada

**********************************************************************/

#ifdef THREAD_SYSTEM_DEPENDENT_IMPLEMENTATION

#include "internal/gc.h"
#include "internal/sanitizers.h"

#ifdef HAVE_SYS_RESOURCE_H
#include <sys/resource.h>
#endif
#ifdef HAVE_THR_STKSEGMENT
#include <thread.h>
#endif
#if defined(HAVE_FCNTL_H)
#include <fcntl.h>
#elif defined(HAVE_SYS_FCNTL_H)
#include <sys/fcntl.h>
#endif
#ifdef HAVE_SYS_PRCTL_H
#include <sys/prctl.h>
#endif
#if defined(HAVE_SYS_TIME_H)
#include <sys/time.h>
#endif
#if defined(__HAIKU__)
#include <kernel/OS.h>
#endif
#ifdef __linux__
#include <sys/syscall.h> /* for SYS_gettid */
#endif
#include <time.h>
#include <signal.h>

#include "probes.h"

#if defined __APPLE__
# include <AvailabilityMacros.h>
#endif

#if defined(HAVE_SYS_EVENTFD_H) && defined(HAVE_EVENTFD)
#  define USE_EVENTFD (1)
#  include <sys/eventfd.h>
#else
#  define USE_EVENTFD (0)
#endif

#if defined(HAVE_PTHREAD_CONDATTR_SETCLOCK) && \
    defined(CLOCK_REALTIME) && defined(CLOCK_MONOTONIC) && \
    defined(HAVE_CLOCK_GETTIME)
static pthread_condattr_t condattr_mono;
static pthread_condattr_t *condattr_monotonic = &condattr_mono;
#else
static const void *const condattr_monotonic = NULL;
#endif

/* A retiring shared native thread frees its own context while the threads it
 * parked are still suspended with that context as their target. */
#define COROUTINE_TARGET_MAY_BE_FREED 1

#include COROUTINE_H

#ifndef HAVE_SYS_EVENT_H
#define HAVE_SYS_EVENT_H 0
#endif

#ifndef HAVE_SYS_EPOLL_H
#define HAVE_SYS_EPOLL_H 0
#else
// force setting for debug
// #undef HAVE_SYS_EPOLL_H
// #define HAVE_SYS_EPOLL_H 0
#endif

#ifndef USE_MN_THREADS
  #if defined(__EMSCRIPTEN__) || defined(COROUTINE_PTHREAD_CONTEXT)
    // on __EMSCRIPTEN__ provides epoll* declarations, but no implementations.
    // on COROUTINE_PTHREAD_CONTEXT, it doesn't worth to use it.
    #define USE_MN_THREADS 0
  #elif HAVE_SYS_EPOLL_H
    #include <sys/epoll.h>
    #define USE_MN_THREADS 1
  #elif HAVE_SYS_EVENT_H
    #include <sys/event.h>
    #define USE_MN_THREADS 1
  #else
    #define USE_MN_THREADS 0
  #endif
#endif

#ifdef HAVE_SCHED_YIELD
#define native_thread_yield() (void)sched_yield()
#else
#define native_thread_yield() ((void)0)
#endif

// native thread wrappers

#define NATIVE_MUTEX_LOCK_DEBUG 0
#define NATIVE_MUTEX_LOCK_DEBUG_YIELD 0

static void
mutex_debug(const char *msg, void *lock)
{
    if (NATIVE_MUTEX_LOCK_DEBUG) {
        int r;
        static pthread_mutex_t dbglock = PTHREAD_MUTEX_INITIALIZER;

        if ((r = pthread_mutex_lock(&dbglock)) != 0) {exit(EXIT_FAILURE);}
        fprintf(stdout, "%s: %p\n", msg, lock);
        if ((r = pthread_mutex_unlock(&dbglock)) != 0) {exit(EXIT_FAILURE);}
    }
}

void
rb_native_mutex_lock(pthread_mutex_t *lock)
{
    int r;
#if NATIVE_MUTEX_LOCK_DEBUG_YIELD
    native_thread_yield();
#endif
    mutex_debug("lock", lock);
    if ((r = pthread_mutex_lock(lock)) != 0) {
        rb_bug_errno("pthread_mutex_lock", r);
    }
}

void
rb_native_mutex_unlock(pthread_mutex_t *lock)
{
    int r;
    mutex_debug("unlock", lock);
    if ((r = pthread_mutex_unlock(lock)) != 0) {
        rb_bug_errno("pthread_mutex_unlock", r);
    }
}

int
rb_native_mutex_trylock(pthread_mutex_t *lock)
{
    int r;
    mutex_debug("trylock", lock);
    if ((r = pthread_mutex_trylock(lock)) != 0) {
        if (r == EBUSY) {
            return EBUSY;
        }
        else {
            rb_bug_errno("pthread_mutex_trylock", r);
        }
    }
    return 0;
}

void
rb_native_mutex_initialize(pthread_mutex_t *lock)
{
    int r = pthread_mutex_init(lock, 0);
    mutex_debug("init", lock);
    if (r != 0) {
        rb_bug_errno("pthread_mutex_init", r);
    }
}

void
rb_native_mutex_destroy(pthread_mutex_t *lock)
{
    int r = pthread_mutex_destroy(lock);
    mutex_debug("destroy", lock);
    if (r != 0) {
        rb_bug_errno("pthread_mutex_destroy", r);
    }
}

void
rb_native_cond_initialize(rb_nativethread_cond_t *cond)
{
    int r = pthread_cond_init(cond, condattr_monotonic);
    if (r != 0) {
        rb_bug_errno("pthread_cond_init", r);
    }
}

void
rb_native_cond_destroy(rb_nativethread_cond_t *cond)
{
    int r = pthread_cond_destroy(cond);
    if (r != 0) {
        rb_bug_errno("pthread_cond_destroy", r);
    }
}

/*
 * In OS X 10.7 (Lion), pthread_cond_signal and pthread_cond_broadcast return
 * EAGAIN after retrying 8192 times.  You can see them in the following page:
 *
 * http://www.opensource.apple.com/source/Libc/Libc-763.11/pthreads/pthread_cond.c
 *
 * The following rb_native_cond_signal and rb_native_cond_broadcast functions
 * need to retrying until pthread functions don't return EAGAIN.
 */

void
rb_native_cond_signal(rb_nativethread_cond_t *cond)
{
    int r;
    do {
        r = pthread_cond_signal(cond);
    } while (r == EAGAIN);
    if (r != 0) {
        rb_bug_errno("pthread_cond_signal", r);
    }
}

void
rb_native_cond_broadcast(rb_nativethread_cond_t *cond)
{
    int r;
    do {
        r = pthread_cond_broadcast(cond);
    } while (r == EAGAIN);
    if (r != 0) {
        rb_bug_errno("rb_native_cond_broadcast", r);
    }
}

void
rb_native_cond_wait(rb_nativethread_cond_t *cond, pthread_mutex_t *mutex)
{
    int r = pthread_cond_wait(cond, mutex);
    if (r != 0) {
        rb_bug_errno("pthread_cond_wait", r);
    }
}

static int
native_cond_timedwait(rb_nativethread_cond_t *cond, pthread_mutex_t *mutex, const rb_hrtime_t *abs)
{
    int r;
    struct timespec ts;

    /*
     * An old Linux may return EINTR. Even though POSIX says
     *   "These functions shall not return an error code of [EINTR]".
     *   http://pubs.opengroup.org/onlinepubs/009695399/functions/pthread_cond_timedwait.html
     * Let's hide it from arch generic code.
     */
    do {
        rb_hrtime2timespec(&ts, abs);
        r = pthread_cond_timedwait(cond, mutex, &ts);
    } while (r == EINTR);

    if (r != 0 && r != ETIMEDOUT) {
        rb_bug_errno("pthread_cond_timedwait", r);
    }

    return r;
}

static rb_hrtime_t
native_cond_timeout(rb_nativethread_cond_t *cond, const rb_hrtime_t rel)
{
    if (condattr_monotonic) {
        return rb_hrtime_add(rb_hrtime_now(), rel);
    }
    else {
        struct timespec ts;

        rb_timespec_now(&ts);
        return rb_hrtime_add(rb_timespec2hrtime(&ts), rel);
    }
}

void
rb_native_cond_timedwait(rb_nativethread_cond_t *cond, pthread_mutex_t *mutex, unsigned long msec)
{
    rb_hrtime_t hrmsec = native_cond_timeout(cond, RB_HRTIME_PER_MSEC * msec);
    native_cond_timedwait(cond, mutex, &hrmsec);
}

// thread scheduling

static rb_internal_thread_event_hook_t *rb_internal_thread_event_hooks = NULL;
static void rb_thread_execute_hooks(rb_event_flag_t event, rb_thread_t *th);

#if 0
static const char *
event_name(rb_event_flag_t event)
{
    switch (event) {
      case RUBY_INTERNAL_THREAD_EVENT_STARTED:
        return "STARTED";
      case RUBY_INTERNAL_THREAD_EVENT_READY:
        return "READY";
      case RUBY_INTERNAL_THREAD_EVENT_RESUMED:
        return "RESUMED";
      case RUBY_INTERNAL_THREAD_EVENT_SUSPENDED:
        return "SUSPENDED";
      case RUBY_INTERNAL_THREAD_EVENT_EXITED:
        return "EXITED";
    }
    return "no-event";
}

#define RB_INTERNAL_THREAD_HOOK(event, th) \
    if (UNLIKELY(rb_internal_thread_event_hooks)) { \
        fprintf(stderr, "[thread=%"PRIxVALUE"] %s in %s (%s:%d)\n", th->self, event_name(event), __func__, __FILE__, __LINE__); \
        rb_thread_execute_hooks(event, th); \
    }
#else
#define RB_INTERNAL_THREAD_HOOK(event, th) if (UNLIKELY(rb_internal_thread_event_hooks)) { rb_thread_execute_hooks(event, th); }
#endif

static rb_serial_t current_fork_gen = 1; /* We can't use GET_VM()->fork_gen */

#if defined(SIGVTALRM) && !defined(__EMSCRIPTEN__)
#  define USE_UBF_LIST 1
#endif

static void threadptr_trap_interrupt(rb_thread_t *);

static void native_thread_dedicated_inc(rb_vm_t *vm, rb_ractor_t *cr, struct rb_native_thread *nt);
static void native_thread_dedicated_dec(rb_vm_t *vm, rb_ractor_t *cr, struct rb_native_thread *nt);
static void native_thread_assign(struct rb_native_thread *nt, rb_thread_t *th);

static void ractor_sched_enq(rb_vm_t *vm, rb_ractor_t *r);
static void timer_thread_wakeup(void);
static void timer_thread_wakeup_locked(rb_vm_t *vm);
static void timer_thread_wakeup_force(void);
static void timer_thread_wake_fence(struct rb_thread_struct *th);
static bool ractor_sched_timeout_arm(rb_thread_t *th, const rb_hrtime_t *rel);
static bool ractor_sched_timeout_disarm(rb_thread_t *th);
static void thread_sched_switch(rb_thread_t *cth, rb_thread_t *next_th);
static void ractor_sched_cancel_enq(rb_vm_t *vm, struct rb_thread_sched *sched);
#if USE_MN_THREADS
static void nt_machine_stack_atfork(void);

// A coroutine thread's execution context: the coroutine_context (first, so
// th->sched.context points at the whole block) plus what is needed to free
// it without the rb_thread_t. Owned by the execution: the dying thread marks
// it dead before its final transfer, and whichever context RESUMES from that
// transfer frees it -- the resume itself proves the transfer's register save
// into this block has completed, so no further synchronization is needed.
struct rb_thread_context {
    struct coroutine_context co; // must be first
    void *stack;                 // the coroutine machine stack (pool stack)
    struct rb_native_thread *nt; // final transfer target, stashed at termination
    bool dead;
};

static bool thread_sched_reclaim(struct coroutine_context *dead_co);
#endif
static void coroutine_transfer0(struct coroutine_context *transfer_from,
                                struct coroutine_context *transfer_to, bool to_dead);

#define thread_sched_dump(s) thread_sched_dump_(__FILE__, __LINE__, s)

static bool
th_has_dedicated_nt(const rb_thread_t *th)
{
    // TODO: th->has_dedicated_nt
    return th->nt->dedicated > 0;
}

RBIMPL_ATTR_MAYBE_UNUSED()
static void
thread_sched_dump_(const char *file, int line, struct rb_thread_sched *sched)
{
    fprintf(stderr, "@%s:%d running:%d\n", file, line, sched->running ? (int)sched->running->serial : -1);
    rb_thread_t *th;
    int i = 0;
    ccan_list_for_each(&sched->readyq, th, sched.node.readyq) {
        i++; if (i>10) rb_bug("too many");
        fprintf(stderr, "  ready:%d (%sNT:%d)\n", th->serial,
                th->nt ? (th->nt->dedicated ? "D" : "S") : "x",
                th->nt ? (int)th->nt->serial : -1);
    }
}

#define ractor_sched_dump(s) ractor_sched_dump_(__FILE__, __LINE__, s)

RBIMPL_ATTR_MAYBE_UNUSED()
static void
ractor_sched_dump_(const char *file, int line, rb_vm_t *vm)
{
    rb_ractor_t *r;

    fprintf(stderr, "ractor_sched_dump %s:%d\n", file, line);

    int i = 0;
    ccan_list_for_each(&vm->ractor.sched.grq, r, threads.sched.grq_node) {
        i++;
        if (i>10) rb_bug("!!");
        fprintf(stderr, "  %d ready:%d\n", i, rb_ractor_id(r));
    }
}

#define thread_sched_lock(a, b) thread_sched_lock_(a, b, __FILE__, __LINE__)
#define thread_sched_unlock(a, b) thread_sched_unlock_(a, b, __FILE__, __LINE__)

static void
thread_sched_set_locked(struct rb_thread_sched *sched, rb_thread_t *th)
{
#if VM_CHECK_MODE > 0
    VM_ASSERT(sched->lock_owner == NULL);

    sched->lock_owner = th;
#endif
}

static void
thread_sched_set_unlocked(struct rb_thread_sched *sched, rb_thread_t *th)
{
#if VM_CHECK_MODE > 0
    VM_ASSERT(sched->lock_owner == th);

    sched->lock_owner = NULL;
#endif
}

static void
thread_sched_lock_(struct rb_thread_sched *sched, rb_thread_t *th, const char *file, int line)
{
    rb_native_mutex_lock(&sched->lock_);

#if VM_CHECK_MODE
    RUBY_DEBUG_LOG2(file, line, "r:%d th:%u", th ? (int)rb_ractor_id(th->ractor) : -1, rb_th_serial(th));
#else
    RUBY_DEBUG_LOG2(file, line, "th:%u", rb_th_serial(th));
#endif

    thread_sched_set_locked(sched, th);
}

static void
thread_sched_unlock_(struct rb_thread_sched *sched, rb_thread_t *th, const char *file, int line)
{
    RUBY_DEBUG_LOG2(file, line, "th:%u", rb_th_serial(th));

    thread_sched_set_unlocked(sched, th);

    rb_native_mutex_unlock(&sched->lock_);
}

static void
ASSERT_thread_sched_locked(struct rb_thread_sched *sched, rb_thread_t *th)
{
    VM_ASSERT(rb_native_mutex_trylock(&sched->lock_) == EBUSY);

#if VM_CHECK_MODE
    if (th) {
        VM_ASSERT(sched->lock_owner == th);
    }
    else {
        VM_ASSERT(sched->lock_owner != NULL);
    }
#endif
}

#define ractor_sched_lock(a, b) ractor_sched_lock_(a, b, __FILE__, __LINE__)
#define ractor_sched_unlock(a, b) ractor_sched_unlock_(a, b, __FILE__, __LINE__)

RBIMPL_ATTR_MAYBE_UNUSED()
static unsigned int
rb_ractor_serial(const rb_ractor_t *r)
{
    if (r) {
        return rb_ractor_id(r);
    }
    else {
        return 0;
    }
}

static void
ractor_sched_set_locked(rb_vm_t *vm, rb_ractor_t *cr)
{
#if VM_CHECK_MODE > 0
    VM_ASSERT(vm->ractor.sched.lock_owner == NULL);
    VM_ASSERT(vm->ractor.sched.locked == false);

    vm->ractor.sched.lock_owner = cr;
    vm->ractor.sched.locked = true;
#endif
}

static void
ractor_sched_set_unlocked(rb_vm_t *vm, rb_ractor_t *cr)
{
#if VM_CHECK_MODE > 0
    VM_ASSERT(vm->ractor.sched.locked);
    VM_ASSERT(vm->ractor.sched.lock_owner == cr);

    vm->ractor.sched.locked = false;
    vm->ractor.sched.lock_owner = NULL;
#endif
}


static void
ractor_sched_lock_(rb_vm_t *vm, rb_ractor_t *cr, const char *file, int line)
{
    rb_native_mutex_lock(&vm->ractor.sched.lock);

#if VM_CHECK_MODE
    RUBY_DEBUG_LOG2(file, line, "cr:%u prev_owner:%u", rb_ractor_serial(cr), rb_ractor_serial(vm->ractor.sched.lock_owner));
#else
    RUBY_DEBUG_LOG2(file, line, "cr:%u", rb_ractor_serial(cr));
#endif

    ractor_sched_set_locked(vm, cr);
}

static void
ractor_sched_unlock_(rb_vm_t *vm, rb_ractor_t *cr, const char *file, int line)
{
    RUBY_DEBUG_LOG2(file, line, "cr:%u", rb_ractor_serial(cr));

    ractor_sched_set_unlocked(vm, cr);
    rb_native_mutex_unlock(&vm->ractor.sched.lock);
}

static void
ASSERT_ractor_sched_locked(rb_vm_t *vm, rb_ractor_t *cr)
{
    VM_ASSERT(rb_native_mutex_trylock(&vm->ractor.sched.lock) == EBUSY);
    VM_ASSERT(vm->ractor.sched.locked);
    VM_ASSERT(cr == NULL || vm->ractor.sched.lock_owner == cr);
}

static void ractor_sched_barrier_join_signal_locked(rb_vm_t *vm);

/* ntlist registration: a thread that executes Ruby code is always registered,
 * in its snt's nt->running_th or on running_dnts via its dedicated nt.  The
 * only unregistered execution is scheduler glue (parking, resuming), which
 * touches no Ruby heap, and the barrier wait below. */
static void
ntlist_add_running(rb_vm_t *vm, rb_thread_t *th)
{
    struct rb_native_thread *nt = th->nt;

    // a dedicated nt is not on the snts list the scans walk: running_dnts instead
    if (nt != NULL && nt->dedicated == 0) {
        rb_native_mutex_lock(&nt->running_th_lock);
        {
            VM_ASSERT(nt->running_th == NULL);
            nt->running_th = th;
        }
        rb_native_mutex_unlock(&nt->running_th_lock);
    }
    else {
        rb_native_mutex_lock(&vm->ractor.sched.ntlist.lock);
        {
            // an snt gone dedicated (rb_thread_lock_native_thread) has no
            // creation-time running_thread: the registration supplies it
            nt->running_thread = th;
            ccan_list_add(&vm->ractor.sched.ntlist.running_dnts, &nt->running_dnts_node);
        }
        rb_native_mutex_unlock(&vm->ractor.sched.ntlist.lock);
    }
}

// Returns whether the active barrier's walk had counted this registration:
// such a deregistration owes the snapshot count a decrement.  Read and
// cleared under the registration's own lock, so it pairs with the walk.
static bool
ntlist_del_running(rb_vm_t *vm, rb_thread_t *th)
{
    struct rb_native_thread *nt = th->nt;
    uint32_t serial;
    bool counted;
    bool in_running_th;

    // The registration itself says where it is: nt->running_th holds th, or
    // th's nt hangs on running_dnts.  barrier_serial is read inside the
    // registration's lock, ordered with the walk that stamped there.
    rb_native_mutex_lock(&nt->running_th_lock);
    {
        in_running_th = (nt->running_th == th);
        if (in_running_th) {
            nt->running_th = NULL;
            serial = vm->ractor.sched.barrier_serial;
            counted = (nt->barrier_counted_serial == serial);
            nt->barrier_counted_serial = serial - 1; // only once per barrier
        }
    }
    rb_native_mutex_unlock(&nt->running_th_lock);

    if (!in_running_th) {
        rb_native_mutex_lock(&vm->ractor.sched.ntlist.lock);
        {
            ccan_list_del_init(&nt->running_dnts_node);
            serial = vm->ractor.sched.barrier_serial;
            counted = (nt->barrier_counted_serial == serial);
            nt->barrier_counted_serial = serial - 1;
        }
        rb_native_mutex_unlock(&vm->ractor.sched.ntlist.lock);
    }
    return counted;
}

// Stamp a registration into the active barrier's snapshot unless the walk
// already counted it; returns whether it stamped.  Called under sched.lock,
// so it is serialized with the walk: the stamp says exactly whether the
// registration came first.
static bool
ntlist_stamp_if_uncounted(rb_vm_t *vm, rb_thread_t *th)
{
    struct rb_native_thread *nt = th->nt;
    uint32_t serial = vm->ractor.sched.barrier_serial; // sched.lock is held
    bool stamped;
    bool in_running_th;

    rb_native_mutex_lock(&nt->running_th_lock);
    {
        in_running_th = (nt->running_th == th);
        if (in_running_th) {
            stamped = (nt->barrier_counted_serial != serial);
            nt->barrier_counted_serial = serial;
        }
    }
    rb_native_mutex_unlock(&nt->running_th_lock);

    if (!in_running_th) {
        rb_native_mutex_lock(&vm->ractor.sched.ntlist.lock);
        {
            stamped = (nt->barrier_counted_serial != serial);
            nt->barrier_counted_serial = serial;
        }
        rb_native_mutex_unlock(&vm->ractor.sched.ntlist.lock);
    }
    return stamped;
}

// Record a thread entering/leaving the running set, with no global lock and
// no count: the records themselves are what the barrier counts.  Pairing:
// the barrier sets barrier_is_waiting and then walks the records under their
// locks; we move a record and then read the flag, so one side sees the other.
// List sched for the timer's timeslice ticks.  The caller holds sched->lock_
// with the readyq non-empty, so the timer cannot prune the entry meanwhile.
static void
timeslice_sched_link(rb_vm_t *vm, struct rb_thread_sched *sched)
{
    rb_native_mutex_lock(&vm->ractor.sched.timeslice.lock);
    {
        if (sched->timeslice_node.next == &sched->timeslice_node) {
            ccan_list_add_tail(&vm->ractor.sched.timeslice.scheds, &sched->timeslice_node);
        }
    }
    rb_native_mutex_unlock(&vm->ractor.sched.timeslice.lock);
}

static void
thread_sched_setup_running_threads(struct rb_thread_sched *sched, rb_ractor_t *cr, rb_vm_t *vm,
                                   rb_thread_t *add_th, rb_thread_t *del_th)
{
    RUBY_DEBUG_LOG("+:%u -:%u", rb_th_serial(add_th), rb_th_serial(del_th));

    if (del_th) {
        bool counted = ntlist_del_running(vm, del_th);
        sched->is_running = false;

        // The first load is only a filter; the one under sched.lock decides.
        // A missed flag means this deregistration preceded the barrier's walk.
        if (UNLIKELY(RUBY_ATOMIC_LOAD(vm->ractor.sched.barrier_is_waiting))) {
            ractor_sched_lock(vm, cr);
            {
                if (RUBY_ATOMIC_LOAD(vm->ractor.sched.barrier_is_waiting)) {
                    if (counted) {
                        VM_ASSERT(vm->ractor.sched.barrier_running_cnt > 0);
                        vm->ractor.sched.barrier_running_cnt--;
                    }
                    ractor_sched_barrier_join_signal_locked(vm);
                }
            }
            ractor_sched_unlock(vm, cr);
        }
    }

    if (add_th) {
        ntlist_add_running(vm, add_th);

        if (UNLIKELY(RUBY_ATOMIC_LOAD(vm->ractor.sched.barrier_is_waiting))) {
            // A stop-the-world section.  In its waiting phase sched.lock is
            // takable: join the snapshot count and take the interrupt (this
            // thread joins at its next check, like any walked runner).  In
            // the GC phase the barrier holds sched.lock to its end, so this
            // blocks here, as the old global-lock design did.
            ractor_sched_lock(vm, cr);
            {
                if (RUBY_ATOMIC_LOAD(vm->ractor.sched.barrier_is_waiting) &&
                    ntlist_stamp_if_uncounted(vm, add_th)) {
                    // the walk ran before this registration; count it in
                    RUBY_DEBUG_LOG("barrier_is_waiting");
                    vm->ractor.sched.barrier_running_cnt++;
                    RUBY_VM_SET_VM_BARRIER_INTERRUPT(add_th->ec);
                }
            }
            ractor_sched_unlock(vm, cr);
        }

        sched->is_running = true;

        // taking a turn with waiters already queued needs the timeslice ticks
        if (!ccan_list_empty(&sched->readyq)) {
            timeslice_sched_link(vm, sched);
            ractor_sched_lock(vm, cr);
            {
                if (vm->ractor.sched.timeslice_wait_inf) {
                    timer_thread_wakeup_locked(vm);
                }
            }
            ractor_sched_unlock(vm, cr);
        }
    }
}

static void
thread_sched_add_running_thread(struct rb_thread_sched *sched, rb_thread_t *th)
{
    ASSERT_thread_sched_locked(sched, th);
    VM_ASSERT(sched->running == th);

    rb_vm_t *vm = th->vm;
    thread_sched_setup_running_threads(sched, th->ractor, vm, th, NULL);
}

static void
thread_sched_del_running_thread(struct rb_thread_sched *sched, rb_thread_t *th)
{
    ASSERT_thread_sched_locked(sched, th);

    rb_vm_t *vm = th->vm;
    thread_sched_setup_running_threads(sched, th->ractor, vm, NULL, th);
}

void
rb_add_running_thread(rb_thread_t *th)
{
    struct rb_thread_sched *sched = TH_SCHED(th);

    thread_sched_lock(sched, th);
    {
        thread_sched_add_running_thread(sched, th);
    }
    thread_sched_unlock(sched, th);
}

void
rb_del_running_thread(rb_thread_t *th)
{
    struct rb_thread_sched *sched = TH_SCHED(th);

    thread_sched_lock(sched, th);
    {
        thread_sched_del_running_thread(sched, th);
    }
    thread_sched_unlock(sched, th);
}

// setup current or next running thread
// sched->running should be set only on this function.
//
// if th is NULL, there is no running threads.
static void
thread_sched_set_running(struct rb_thread_sched *sched, rb_thread_t *th)
{
    RUBY_DEBUG_LOG("th:%u->th:%u", rb_th_serial(sched->running), rb_th_serial(th));
    VM_ASSERT(sched->running != th);

    if (RUBY_DTRACE_RTS_SET_RUNNING_ENABLED()) {
        RUBY_DTRACE_RTS_SET_RUNNING(sched, sched->running, th);
    }

    sched->running = th;
}

RBIMPL_ATTR_MAYBE_UNUSED()
static bool
thread_sched_readyq_contain_p(struct rb_thread_sched *sched, rb_thread_t *th)
{
    rb_thread_t *rth;
    ccan_list_for_each(&sched->readyq, rth, sched.node.readyq) {
        if (rth == th) {
            VM_ASSERT(th->sched.node.is_ready);
            return true;
        }
    }
    VM_ASSERT(!th->sched.node.is_ready);
    return false;
}

// deque thread from the ready queue.
// if the ready queue is empty, return NULL.
//
// return deque'ed running thread (or NULL).
static rb_thread_t *
thread_sched_deq(struct rb_thread_sched *sched)
{
    ASSERT_thread_sched_locked(sched, NULL);
    rb_thread_t *next_th;

    VM_ASSERT(sched->running != NULL);

    if (ccan_list_empty(&sched->readyq)) {
        next_th = NULL;
    }
    else {
        next_th = ccan_list_pop(&sched->readyq, rb_thread_t, sched.node.readyq);
        VM_ASSERT(next_th->sched.node.is_ready);
        next_th->sched.node.is_ready = false;

        VM_ASSERT(sched->readyq_cnt > 0);
        sched->readyq_cnt--;
        ccan_list_node_init(&next_th->sched.node.readyq);
    }

    RUBY_DEBUG_LOG("next_th:%u readyq_cnt:%d", rb_th_serial(next_th), sched->readyq_cnt);

    return next_th;
}

// enqueue ready thread to the ready queue.
static void
thread_sched_enq(struct rb_thread_sched *sched, rb_thread_t *ready_th)
{
    ASSERT_thread_sched_locked(sched, NULL);
    RUBY_DEBUG_LOG("ready_th:%u readyq_cnt:%d", rb_th_serial(ready_th), sched->readyq_cnt);

    VM_ASSERT(sched->running != NULL);
    VM_ASSERT(!thread_sched_readyq_contain_p(sched, ready_th));

    bool timeslice_onset = sched->is_running && ccan_list_empty(&sched->readyq);

    ccan_list_add_tail(&sched->readyq, &ready_th->sched.node.readyq);
    ready_th->sched.node.is_ready = true;
    sched->readyq_cnt++;

    if (timeslice_onset) {
        // The running thread needs the timeslice ticks now.  Linked before
        // the check under sched.lock: either the timer's scan (same lock)
        // sees the sched, or this sees timeslice_wait_inf.
        rb_vm_t *vm = ready_th->vm;
        timeslice_sched_link(vm, sched);
        ractor_sched_lock(vm, NULL);
        {
            if (vm->ractor.sched.timeslice_wait_inf) {
                timer_thread_wakeup_locked(vm);
            }
        }
        ractor_sched_unlock(vm, NULL);
    }
}

// DNT: kick condvar
// SNT: TODO
static void
thread_sched_wakeup_running_thread(struct rb_thread_sched *sched, rb_thread_t *next_th, bool will_switch)
{
    ASSERT_thread_sched_locked(sched, NULL);
    VM_ASSERT(sched->running == next_th);

    if (next_th) {
        if (next_th->nt) {
            if (th_has_dedicated_nt(next_th)) {
                RUBY_DEBUG_LOG("pinning th:%u", next_th->serial);
                rb_native_cond_signal(&next_th->nt->readyq);
            }
            else {
                // TODO
                RUBY_DEBUG_LOG("th:%u is already running.", next_th->serial);
            }
        }
        else {
            if (will_switch) {
                RUBY_DEBUG_LOG("th:%u (do nothing)", rb_th_serial(next_th));
            }
            else {
                RUBY_DEBUG_LOG("th:%u (enq)", rb_th_serial(next_th));
                ractor_sched_enq(next_th->vm, next_th->ractor);
            }
        }
    }
    else {
        RUBY_DEBUG_LOG("no waiting threads%s", "");
    }
}

// waiting -> ready (locked)
static void
thread_sched_to_ready_common(struct rb_thread_sched *sched, rb_thread_t *th, bool wakeup, bool will_switch)
{
    RUBY_DEBUG_LOG("th:%u running:%u redyq_cnt:%d", rb_th_serial(th), rb_th_serial(sched->running), sched->readyq_cnt);

    VM_ASSERT(sched->running != th);
    VM_ASSERT(!thread_sched_readyq_contain_p(sched, th));
    RB_INTERNAL_THREAD_HOOK(RUBY_INTERNAL_THREAD_EVENT_READY, th);

    if (sched->running == NULL) {
        thread_sched_set_running(sched, th);
        if (wakeup) thread_sched_wakeup_running_thread(sched, th, will_switch);
    }
    else {
        thread_sched_enq(sched, th);
    }
}

// waiting -> ready
//
// `th` had became "waiting" state by `thread_sched_to_waiting`
// and `thread_sched_to_ready` enqueue `th` to the thread ready queue.
RBIMPL_ATTR_MAYBE_UNUSED()
static void
thread_sched_to_ready(struct rb_thread_sched *sched, rb_thread_t *th)
{
    RUBY_DEBUG_LOG("th:%u", rb_th_serial(th));

    thread_sched_lock(sched, th);
    {
        thread_sched_to_ready_common(sched, th, true, false);
    }
    thread_sched_unlock(sched, th);
}

// wait until sched->running is `th`.  `end` is an absolute deadline for a dedicated
static void
thread_sched_wait_running_turn(struct rb_thread_sched *sched, rb_thread_t *th, bool can_direct_transfer, const rb_hrtime_t *end)
{
    RUBY_DEBUG_LOG("th:%u", rb_th_serial(th));

    ASSERT_thread_sched_locked(sched, th);
    VM_ASSERT(th == rb_ec_thread_ptr(rb_current_ec_noinline()));

    bool timedout = false;

    if (th != sched->running) {
        // TODO: This optimization should also be made to work for MN_THREADS
        if (th->has_dedicated_nt && th == sched->runnable_hot_th && (sched->running == NULL || sched->running->has_dedicated_nt)) {
            RUBY_DEBUG_LOG("(nt) stealing: hot-th:%u.  running:%u", rb_th_serial(th), rb_th_serial(sched->running));

            // th serves itself on its own nt, displacing the enqueued
            // running thread back to the readyq: cancel the entry that was
            // posted for it (a later dequeue would find this Ractor served
            // and its next enqueue would double-list the node)
            ractor_sched_cancel_enq(th->vm, sched);

            // If there is a thread set to run, move it back to the front of the readyq
            if (sched->running != NULL) {
                rb_thread_t *running = sched->running;
                VM_ASSERT(!thread_sched_readyq_contain_p(sched, running));
                running->sched.node.is_ready = true;
                ccan_list_add(&sched->readyq, &running->sched.node.readyq);
                sched->readyq_cnt++;
            }

            // Pull off the ready queue and start running.
            if (th->sched.node.is_ready) {
                VM_ASSERT(thread_sched_readyq_contain_p(sched, th));
                ccan_list_del_init(&th->sched.node.readyq);
                th->sched.node.is_ready = false;
                sched->readyq_cnt--;
            }
            thread_sched_set_running(sched, th);
            rb_ractor_thread_switch(th->ractor, th, false);
        }
        else if (th == sched->runnable_hot_th) {
            // The hot thread cannot steal the control (e.g. the running thread
            // is an MN thread). It is going to sleep, so it is no longer spinning;
            // drop the hint so that other threads don't yield the lock to it.
            sched->runnable_hot_th = NULL;
            sched->runnable_hot_th_waiting = 0;
        }

        // already deleted from running threads


        // wait for execution right
        rb_thread_t *next_th;
        while((next_th = sched->running) != th) {
            if (th_has_dedicated_nt(th)) {
                RUBY_DEBUG_LOG("(nt) sleep th:%u running:%u", rb_th_serial(th), rb_th_serial(sched->running));

                thread_sched_set_unlocked(sched, th);
                {
                    RUBY_DEBUG_LOG("nt:%d cond:%p", th->nt->serial, &th->nt->readyq);
                    rb_nativethread_cond_t *cond = &th->nt->readyq;

                    // Once someone has queued this thread the deadline is spent: it
                    // is waiting for a turn, not for the time, and arming a kernel
                    // timer for every round of that costs more than the wait.
                    // Once someone has queued this thread the deadline is spent: it
                    // is waiting for a turn, not for the time, and arming a kernel
                    // timer for every round of that costs more than the wait.
                    if (end && !th->sched.node.is_ready) {
                        rb_hrtime_t abs = *end;

                        if (!condattr_monotonic) {
                            // the condvar counts in another clock: restate it there
                            rb_hrtime_t now = rb_hrtime_now();
                            abs = native_cond_timeout(cond, *end > now ? *end - now : 0);
                        }
                        timedout = native_cond_timedwait(cond, &sched->lock_, &abs) == ETIMEDOUT;
                    }
                    else {
                        rb_native_cond_wait(cond, &sched->lock_);
                    }
                }
                thread_sched_set_locked(sched, th);

                if (timedout &&
                    sched->running != th && !th->sched.node.is_ready) {
                    // the deadline passed and nobody woke this thread: get back in
                    // line for the running turn, then wait for it without a deadline
                    thread_sched_to_ready_common(sched, th, false, false);
                    end = NULL;
                }

                if (sched->runnable_hot_th != NULL && sched->runnable_hot_th_waiting) {
                    VM_ASSERT(sched->runnable_hot_th != th);
                    // Give the hot thread a chance to preempt, if it's actively spinning.
                    // On multicore, this reduces the rate of core-switching. On single-core it
                    // should mostly be a nop, since the other thread can't be concurrently spinning.
                    thread_sched_unlock(sched, th);
                    thread_sched_lock(sched, th);
                }

                RUBY_DEBUG_LOG("(nt) wakeup %s", sched->running == th ? "success" : "failed");
                if (th == sched->running) {
                    rb_ractor_thread_switch(th->ractor, th, false);
                }
            }
            else {
                // search another ready thread
                if (can_direct_transfer &&
                    (next_th = sched->running) != NULL &&
                    !next_th->nt // next_th is running or has dedicated nt
                    ) {

                    RUBY_DEBUG_LOG("th:%u->%u (direct)", rb_th_serial(th), rb_th_serial(next_th));

                    thread_sched_set_unlocked(sched, th);
                    {
                        rb_ractor_set_current_ec(th->ractor, NULL);
                        thread_sched_switch(th, next_th);
                    }
                    thread_sched_set_locked(sched, th);
                }
                else {
                    // search another ready ractor
                    struct rb_native_thread *nt = th->nt;
                    native_thread_assign(NULL, th);

                    RUBY_DEBUG_LOG("th:%u->%u (ractor scheduling)", rb_th_serial(th), rb_th_serial(next_th));

                    thread_sched_set_unlocked(sched, th);
                    {
                        rb_ractor_set_current_ec(th->ractor, NULL);
                        coroutine_transfer0(th->sched.context, nt->nt_context, false);
                    }
                    thread_sched_set_locked(sched, th);
                }

                VM_ASSERT(rb_current_ec_noinline() == th->ec);
            }
        }

        VM_ASSERT(th->nt != NULL);
        VM_ASSERT(rb_current_ec_noinline() == th->ec);
        VM_ASSERT(th->sched.waiting_reason.flags == thread_sched_waiting_none);

        // add th to running threads
        thread_sched_add_running_thread(sched, th);
    }

    // Control transfer to the current thread is now complete. The original thread
    // cannot steal control at this point.
    sched->runnable_hot_th = NULL;
    sched->runnable_hot_th_waiting = 0;


    RB_INTERNAL_THREAD_HOOK(RUBY_INTERNAL_THREAD_EVENT_RESUMED, th);
}

// waiting -> ready -> running (locked)
static void
thread_sched_to_running_common(struct rb_thread_sched *sched, rb_thread_t *th)
{
    RUBY_DEBUG_LOG("th:%u dedicated:%d", rb_th_serial(th), th_has_dedicated_nt(th));

    VM_ASSERT(sched->running != th);
    VM_ASSERT(th_has_dedicated_nt(th));
    VM_ASSERT(GET_THREAD() == th);

    native_thread_dedicated_dec(th->vm, th->ractor, th->nt);

    // waiting -> ready
    thread_sched_to_ready_common(sched, th, false, false);

    if (sched->running == th) {
        thread_sched_add_running_thread(sched, th);
    }

    // TODO: check SNT number
    thread_sched_wait_running_turn(sched, th, false, NULL);
}

// waiting -> ready -> running
//
// `th` had been waiting by `thread_sched_to_waiting()`
// and run a dedicated task (like waitpid and so on).
// After the dedicated task, this function is called
// to join a normal thread-scheduling.
static void
thread_sched_to_running(struct rb_thread_sched *sched, rb_thread_t *th)
{
    // We are reading and writing these sched fields without lock cover, but
    // there are no correctness issues resulting from stale cache or delayed writeback.
    // When it works, this causes the next-scheduled thread to yield the sched lock
    // briefly so that we can grab it if we're still spinning (not descheduled yet).
    if (sched->runnable_hot_th == th) {
        sched->runnable_hot_th_waiting = 1;
    }
    thread_sched_lock(sched, th);
    {
        thread_sched_to_running_common(sched, th);
    }
    thread_sched_unlock(sched, th);
}

// resume a next thread in the thread ready queue.
//
// deque next running thread from the ready thread queue and
// resume this thread if available.
//
// If the next therad has a dedicated native thraed, simply signal to resume.
// Otherwise, make the ractor ready and other nt will run the ractor and the thread.
static void
thread_sched_wakeup_next_thread(struct rb_thread_sched *sched, rb_thread_t *th, bool will_switch)
{
    ASSERT_thread_sched_locked(sched, th);

    VM_ASSERT(sched->running == th);
    VM_ASSERT(sched->running->nt != NULL);

    rb_thread_t *next_th = thread_sched_deq(sched);

    RUBY_DEBUG_LOG("next_th:%u", rb_th_serial(next_th));
    VM_ASSERT(th != next_th);

    thread_sched_set_running(sched, next_th);
    VM_ASSERT(next_th == sched->running);
    thread_sched_wakeup_running_thread(sched, next_th, will_switch);

    if (th != next_th) {
        thread_sched_del_running_thread(sched, th);
    }
}

// running -> dead (locked)
static void
thread_sched_to_dead_common(struct rb_thread_sched *sched, rb_thread_t *th)
{
    RUBY_DEBUG_LOG("th:%u DNT:%d", rb_th_serial(th), th->nt->dedicated);

    RB_INTERNAL_THREAD_HOOK(RUBY_INTERNAL_THREAD_EVENT_SUSPENDED, th);

    // A dying coroutine thread (will_switch=true here) does NOT wake the
    // next thread now: it is still winding down (co_start's epilogue), and
    // the same Ractor must not have two threads executing at once. The
    // epilogue enqueues the Ractor after its last rb_ractor_t access.
    thread_sched_wakeup_next_thread(sched, th, !th_has_dedicated_nt(th));

    RB_INTERNAL_THREAD_HOOK(RUBY_INTERNAL_THREAD_EVENT_EXITED, th);
}

// running -> dead
static void
thread_sched_to_dead(struct rb_thread_sched *sched, rb_thread_t *th)
{
    // wait out any pending wake here, while th's Ractor is still alive
    timer_thread_wake_fence(th);

    thread_sched_lock(sched, th);
    {
        thread_sched_to_dead_common(sched, th);
    }
    thread_sched_unlock(sched, th);
}

// running -> waiting (locked)
//
// This thread will run dedicated task (th->nt->dedicated++).
static void
thread_sched_to_waiting_common(struct rb_thread_sched *sched, rb_thread_t *th, bool yield_immediately)
{
    RUBY_DEBUG_LOG("th:%u DNT:%d", rb_th_serial(th), th->nt->dedicated);

    RB_INTERNAL_THREAD_HOOK(RUBY_INTERNAL_THREAD_EVENT_SUSPENDED, th);

    native_thread_dedicated_inc(th->vm, th->ractor, th->nt);
    if (!yield_immediately) {
        sched->runnable_hot_th = th;
        sched->runnable_hot_th_waiting = 0;
    }
    thread_sched_wakeup_next_thread(sched, th, false);
}

// running -> waiting
//
// This thread will run a dedicated task.
static void
thread_sched_to_waiting(struct rb_thread_sched *sched, rb_thread_t *th, bool yield_immediately)
{
    thread_sched_lock(sched, th);
    {
        thread_sched_to_waiting_common(sched, th, yield_immediately);
    }
    thread_sched_unlock(sched, th);
}

// mini utility func
// return true if any there are any interrupts
static bool
ubf_set(rb_thread_t *th, rb_unblock_function_t *func, void *arg, rb_atomic_t *event_serial)
{
    VM_ASSERT(func != NULL);

  retry:
    if (RUBY_VM_INTERRUPTED(th->ec)) {
        RUBY_DEBUG_LOG("interrupted:0x%x", th->ec->interrupt_flag);
        return true;
    }

    rb_native_mutex_lock(&th->interrupt_lock);
    {
        if (!th->ec->raised_flag && RUBY_VM_INTERRUPTED(th->ec)) {
            rb_native_mutex_unlock(&th->interrupt_lock);
            goto retry;
        }

        VM_ASSERT(th->unblock.func == NULL);
        th->unblock.func = func;
        th->unblock.arg  = arg;
        if (event_serial) {
            rb_atomic_t prev_serial = RUBY_ATOMIC_FETCH_ADD(th->unblock.event_serial, 1);
            *event_serial = prev_serial+1;
        }
    }
    rb_native_mutex_unlock(&th->interrupt_lock);

    return false;
}

static void
ubf_clear(rb_thread_t *th, bool clear_serial)
{
    rb_native_mutex_lock(&th->interrupt_lock);
    {
        th->unblock.func = NULL;
        th->unblock.arg  = NULL;
        if (clear_serial) {
            RUBY_ATOMIC_ADD(th->unblock.event_serial, 1);
        }
    }
    rb_native_mutex_unlock(&th->interrupt_lock);
}

static void
ubf_waiting(void *ptr)
{
    rb_thread_t *th = (rb_thread_t *)ptr;
    struct rb_thread_sched *sched = TH_SCHED(th);

    // only once. it is safe because th->interrupt_lock is already acquired.
    th->unblock.func = NULL;
    th->unblock.arg = NULL;

    RUBY_DEBUG_LOG("th:%u", rb_th_serial(th));

    thread_sched_lock(sched, th);
    {
        if (sched->running == th || th->sched.node.is_ready) {
            // not sleeping yet, or a deadline already put it back in line
        }
        else {
            thread_sched_to_ready_common(sched, th, true, false);

            // If the turn is taken, th stays parked until the running thread yields.
            // For a timed wait, wake it early anyway: it re-parks at once, but its
            // wakeup then runs on another core in parallel with the running thread,
            // off the handoff path.  An untimed wait has no post-wake bookkeeping
            // worth pipelining, so it skips the extra futex round.
            if (sched->running != th && th->sched.waiting_timed &&
                th->nt != NULL && th_has_dedicated_nt(th)) {
                rb_native_cond_signal(&th->nt->readyq);
            }
        }
    }
    thread_sched_unlock(sched, th);
}

// running -> waiting
//
// This thread will sleep until other thread wakeup the thread.  `end` is an
// absolute deadline, NULL to sleep until woken; only a dedicated native thread,
// which parks on its own condvar, can take one.
static void
thread_sched_to_waiting_until_wakeup(struct rb_thread_sched *sched, rb_thread_t *th, const rb_hrtime_t *end)
{
    RUBY_DEBUG_LOG("th:%u", rb_th_serial(th));

    VM_ASSERT(end == NULL || th_has_dedicated_nt(th));

    RB_VM_SAVE_MACHINE_CONTEXT(th);


    RB_INTERNAL_THREAD_HOOK(RUBY_INTERNAL_THREAD_EVENT_SUSPENDED, th);

    thread_sched_lock(sched, th);
    {
        // NOTE: there's a lock ordering inversion here with the ubf call, but it's benign.
        if (ubf_set(th, ubf_waiting, (void *)th, NULL)) {
            RUBY_DEBUG_LOG("th:%u interrupted", rb_th_serial(th));
        }
        else {
            bool can_direct_transfer = !th_has_dedicated_nt(th);
            th->sched.waiting_timed = (end != NULL); // never true here for M:N (end is NULL)
            // NOTE: th->status is set before and after this sleep outside of this function in `sleep_forever`
            thread_sched_wakeup_next_thread(sched, th, can_direct_transfer);
            thread_sched_wait_running_turn(sched, th, can_direct_transfer, end);
            th->sched.waiting_timed = false;
        }
    }
    thread_sched_unlock(sched, th);

    ubf_clear(th, false);
}

// run another thread in the ready queue.
// continue to run if there are no ready threads.
static void
thread_sched_yield(struct rb_thread_sched *sched, rb_thread_t *th)
{
    RUBY_DEBUG_LOG("th:%d sched->readyq_cnt:%d", (int)th->serial, sched->readyq_cnt);

    thread_sched_lock(sched, th);
    {
        if (!ccan_list_empty(&sched->readyq)) {
            RB_INTERNAL_THREAD_HOOK(RUBY_INTERNAL_THREAD_EVENT_SUSPENDED, th);
            thread_sched_wakeup_next_thread(sched, th, !th_has_dedicated_nt(th));
            bool can_direct_transfer = !th_has_dedicated_nt(th);
            thread_sched_to_ready_common(sched, th, false, can_direct_transfer);
            thread_sched_wait_running_turn(sched, th, can_direct_transfer, NULL);
            th->status = THREAD_RUNNABLE;
        }
        else {
            VM_ASSERT(sched->readyq_cnt == 0);
        }
    }
    thread_sched_unlock(sched, th);
}

void
rb_thread_sched_init(struct rb_thread_sched *sched, bool atfork)
{
    rb_native_mutex_initialize(&sched->lock_);

#if VM_CHECK_MODE
    sched->lock_owner = NULL;
#endif

    ccan_list_head_init(&sched->readyq);
    sched->readyq_cnt = 0;
    ccan_list_node_init(&sched->grq_node); // self-linked = not enqueued
    ccan_list_node_init(&sched->timeslice_node);

#if USE_MN_THREADS
    if (!atfork) sched->enable_mn_threads = true; // MN is enabled on Ractors
#endif
}

static void
coroutine_transfer0(struct coroutine_context *transfer_from, struct coroutine_context *transfer_to, bool to_dead)
{
#ifdef RUBY_ASAN_ENABLED
    void **fake_stack = to_dead ? NULL : &transfer_from->fake_stack;
    __sanitizer_start_switch_fiber(fake_stack, transfer_to->stack_base, transfer_to->stack_size);
#endif

#if defined(COROUTINE_SANITIZE_THREAD)
    /* Tell TSan we are switching to transfer_to's fiber before the stack
     * switch, so its per-thread shadow stack stays bound to the right
     * coroutine. */
    __tsan_switch_to_fiber(transfer_to->tsan_fiber, 0);
#endif

    RBIMPL_ATTR_MAYBE_UNUSED()
    struct coroutine_context *returning_from = coroutine_transfer(transfer_from, transfer_to);

    /* if to_dead was passed, the caller is promising that this coroutine is finished and it should
     * never be resumed! */
    VM_ASSERT(!to_dead);
#ifdef RUBY_ASAN_ENABLED
   __sanitizer_finish_switch_fiber(transfer_from->fake_stack,
                                   (const void**)&returning_from->stack_base, &returning_from->stack_size);
#endif
}

static void
thread_sched_switch0(struct coroutine_context *current_cont, rb_thread_t *next_th, struct rb_native_thread *nt, bool to_dead)
{
    VM_ASSERT(!nt->dedicated);
    VM_ASSERT(next_th->nt == NULL);

    RUBY_DEBUG_LOG("next_th:%u", rb_th_serial(next_th));

    // this direct transfer serves next_th without a dequeue; cancel its
    // Ractor's outstanding grq entry (no-op when nothing is enqueued)
    ractor_sched_cancel_enq(next_th->vm, TH_SCHED(next_th));

    ruby_thread_set_native(next_th);
    native_thread_assign(nt, next_th);

    coroutine_transfer0(current_cont, next_th->sched.context, to_dead);
}

static void
thread_sched_switch(rb_thread_t *cth, rb_thread_t *next_th)
{
    struct rb_native_thread *nt = cth->nt;
    native_thread_assign(NULL, cth);
    RUBY_DEBUG_LOG("th:%u->%u on nt:%d", rb_th_serial(cth), rb_th_serial(next_th), nt->serial);
    thread_sched_switch0(cth->sched.context, next_th, nt, cth->status == THREAD_KILLED);
}

#if VM_CHECK_MODE > 0
RBIMPL_ATTR_MAYBE_UNUSED()
static unsigned int
grq_size(rb_vm_t *vm, rb_ractor_t *cr)
{
    ASSERT_ractor_sched_locked(vm, cr);

    rb_ractor_t *r, *prev_r = NULL;
    unsigned int i = 0;

    ccan_list_for_each(&vm->ractor.sched.grq, r, threads.sched.grq_node) {
        i++;

        VM_ASSERT(r != prev_r);
        prev_r = r;
    }
    return i;
}
#endif

// A native thread enters/leaves an epilogue that outlives its Ractor: from
// the increment until the decrement, ruby_vm_destruct waits for it below.
// The increment must happen while the VM still counts the thread's Ractor,
// so that the two never look absent at the same time.
void
rb_thread_sched_winding_begin(rb_vm_t *vm)
{
    RUBY_ATOMIC_INC(vm->ractor.sched.winding_cnt);
}

void
rb_thread_sched_winding_end(rb_vm_t *vm)
{
    VM_ASSERT(RUBY_ATOMIC_LOAD(vm->ractor.sched.winding_cnt) > 0);
    RUBY_ATOMIC_DEC(vm->ractor.sched.winding_cnt);
}

// ruby_vm_destruct: wait until no native thread is between a coroutine
// epilogue and its reclaim -- past that point the reclaim frees through the
// (about to be destroyed) objspace and reads the (about to be unset) VM.
// Runs without the VM lock, which the epilogue needs to progress.
void
rb_thread_sched_wait_winding(rb_vm_t *vm)
{
    while (RUBY_ATOMIC_LOAD(vm->ractor.sched.winding_cnt) > 0) {
        native_thread_yield();
    }
}

// A direct service of a runnable thread (direct transfer or the hot-thread
// steal) bypasses the grq; cancel the Ractor's outstanding entry so that
// "enqueued <=> runnable and unserved" keeps holding. The caller holds the
// per-Ractor sched lock, so no concurrent enqueue can relink the node: a
// self-linked read needs no lock (the common case -- direct switches whose
// transition never enqueued). A linked read can race only with a dequeue,
// hence the recheck under the grq lock.
static void
ractor_sched_cancel_enq(rb_vm_t *vm, struct rb_thread_sched *sched)
{
    if (sched->grq_node.next != &sched->grq_node) {
        ractor_sched_lock(vm, NULL);
        {
            if (sched->grq_node.next != &sched->grq_node) {
                ccan_list_del_init(&sched->grq_node);
                VM_ASSERT(vm->ractor.sched.grq_cnt > 0);
                vm->ractor.sched.grq_cnt--;
            }
        }
        ractor_sched_unlock(vm, NULL);
    }
}

static void
ractor_sched_enq(rb_vm_t *vm, rb_ractor_t *r)
{
    struct rb_thread_sched *sched = &r->threads.sched;
    rb_ractor_t *cr = NULL; // timer thread can call this function

    VM_ASSERT(sched->running != NULL);
    VM_ASSERT(sched->running->nt == NULL);

    ractor_sched_lock(vm, cr);
    {
        // Precondition: not already enqueued (the grq_node is self-linked).
        // This holds because every service of a runnable-but-unserved thread
        // either dequeues the entry (the nt scheduling loop) or cancels it
        // (direct transfers / the hot-thread steal; see
        // ractor_sched_cancel_enq) -- re-adding a linked node would corrupt
        // the queue, so check unconditionally (a CHECK-mode-only assert
        // would miss it: the race needs timing that CHECK builds perturb).
        if (sched->grq_node.next != &sched->grq_node) {
            rb_bug("ractor_sched_enq: already enqueued");
        }
        ccan_list_add_tail(&vm->ractor.sched.grq, &sched->grq_node);
        vm->ractor.sched.grq_cnt++;
        VM_ASSERT(grq_size(vm, cr) == vm->ractor.sched.grq_cnt);

        RUBY_DEBUG_LOG("r:%u th:%u grq_cnt:%u", rb_ractor_id(r), rb_th_serial(sched->running), vm->ractor.sched.grq_cnt);

        rb_native_cond_signal(&vm->ractor.sched.cond);

        // The signal reaches a parked snt, and a running one revisits the
        // queue in ractor_sched_deq before it can wait (same lock as here).
        // With every snt dedicated or retired, only the timer thread's
        // timeout branch can serve the entry or widen the pool: wake it
        // (a no-op unless it sleeps untimed).
        if (RUBY_ATOMIC_LOAD(vm->ractor.sched.snt_cnt) == 0) {
            timer_thread_wakeup_locked(vm);
        }

        // ractor_sched_dump(vm);
    }
    ractor_sched_unlock(vm, cr);
}


#ifndef MINIMUM_SNT
// make at least MINIMUM_SNT snts for debug.
#define MINIMUM_SNT 0
#endif

/* A shared thread woken with nothing to run is one whose turn another thread
 * took first.  After this many in a row it gives itself back: the queue keeps
 * running dry, so the pool is wider than the work.  0 retires on the first one
 * and is too eager to be useful; a negative value keeps every thread. */
#ifndef SNT_IDLE_RETIRE
#define SNT_IDLE_RETIRE 3
#endif

/* Never give the last shared thread back.  With none left an enqueue has nobody
 * to signal, and the only code that makes one runs on the timer thread's
 * timeout branch, which is reached only once it has seen a backlog. */
#define SNT_KEEP_MINIMUM (MINIMUM_SNT > 1 ? MINIMUM_SNT : 1)

static rb_ractor_t *
ractor_sched_deq(rb_vm_t *vm, rb_ractor_t *cr)
{
    rb_ractor_t *r;
    int idle_streak = 0;   // consecutive pops that found the queue empty

    ractor_sched_lock(vm, cr);
    {
        RUBY_DEBUG_LOG("empty? %d", ccan_list_empty(&vm->ractor.sched.grq));
        // ractor_sched_dump(vm);

        VM_ASSERT(rb_current_execution_context(false) == NULL);
        VM_ASSERT(grq_size(vm, cr) == vm->ractor.sched.grq_cnt);

        while ((r = ccan_list_pop(&vm->ractor.sched.grq, rb_ractor_t, threads.sched.grq_node)) == NULL) {
            RUBY_DEBUG_LOG("wait grq_cnt:%d", (int)vm->ractor.sched.grq_cnt);

            if (SNT_IDLE_RETIRE >= 0 && ++idle_streak > SNT_IDLE_RETIRE &&
                (int)RUBY_ATOMIC_LOAD(vm->ractor.sched.snt_cnt) > SNT_KEEP_MINIMUM) {
                RUBY_ATOMIC_DEC(vm->ractor.sched.snt_cnt);
                RUBY_DEBUG_LOG("retire, snt_cnt:%d", (int)vm->ractor.sched.snt_cnt);
                break;   // returning NULL ends this nt; see the caller
            }

            ractor_sched_set_unlocked(vm, cr);
            rb_native_cond_wait(&vm->ractor.sched.cond, &vm->ractor.sched.lock);
            ractor_sched_set_locked(vm, cr);

            RUBY_DEBUG_LOG("wakeup grq_cnt:%d", (int)vm->ractor.sched.grq_cnt);
        }

        VM_ASSERT(rb_current_execution_context(false) == NULL);

        if (r) {
            ccan_list_node_init(&r->threads.sched.grq_node); // back to self-linked
            VM_ASSERT(vm->ractor.sched.grq_cnt > 0);
            vm->ractor.sched.grq_cnt--;
            RUBY_DEBUG_LOG("r:%d grq_cnt:%u", (int)rb_ractor_id(r), vm->ractor.sched.grq_cnt);
        }
        else {
            // the retire branch is the only way out of the loop without a ractor
            VM_ASSERT(idle_streak > SNT_IDLE_RETIRE);
        }
    }
    ractor_sched_unlock(vm, cr);

    return r;
}

void rb_ractor_lock_self(rb_ractor_t *r);
void rb_ractor_unlock_self(rb_ractor_t *r);

// The current thread for a ractor is put to "sleep" (descheduled in the STOPPED_FOREVER state) waiting for
// a ractor action to wake it up.
void
rb_ractor_sched_wait(rb_execution_context_t *ec, rb_ractor_t *cr, rb_unblock_function_t *ubf, void *ubf_arg)
{
    // ractor lock of cr is acquired

    RUBY_DEBUG_LOG("start%s", "");

    rb_thread_t * volatile th = rb_ec_thread_ptr(ec);
    struct rb_thread_sched *sched = TH_SCHED(th);
    struct ractor_waiter *waiter = (struct ractor_waiter*)ubf_arg;

    if (ubf_set(th, ubf, ubf_arg, &waiter->event_serial)) {
        // interrupted
        return;
    }

    thread_sched_lock(sched, th);
    rb_ractor_unlock_self(cr);
    {
        // A dedicated native thread takes the deadline on the very condvar a wakeup
        // signals.  An M:N thread has no condvar of its own, so its deadline goes to
        // the timer thread, which then wakes it the way rb_ractor_sched_wakeup() does.
        bool dedicated = th_has_dedicated_nt(th);
        const rb_hrtime_t *end_p = NULL;
        bool armed = false, expired = false;

        if (waiter->end) {
            if (dedicated) {
                end_p = waiter->end;
            }
            else {
                // the timer wheel takes a relative timeout
                rb_hrtime_t now = rb_hrtime_now();
                rb_hrtime_t rel = *waiter->end > now ? *waiter->end - now : 0;

                armed = ractor_sched_timeout_arm(th, &rel);
                expired = !armed;
            }
        }

        if (expired) {
            RUBY_DEBUG_LOG("expired before sleep%s", "");
        }
        else if (armed && th->sched.waiting_reason.flags == thread_sched_waiting_none) {
            // the timer thread already took this thread out of the wheel; bump the
            // serial so that it does not try to wake a thread that never slept
            th->sched.event_serial++;
        }
        else {
            // setup sleep
            bool can_direct_transfer = !dedicated;
            RB_VM_SAVE_MACHINE_CONTEXT(th);
            th->status = THREAD_STOPPED_FOREVER;
            th->sched.waiting_timed = (end_p != NULL); // never true here for M:N (end_p is NULL)
            RB_INTERNAL_THREAD_HOOK(RUBY_INTERNAL_THREAD_EVENT_SUSPENDED, th);
            thread_sched_wakeup_next_thread(sched, th, can_direct_transfer);
            // sleep
            thread_sched_wait_running_turn(sched, th, can_direct_transfer, end_p);
            th->sched.waiting_timed = false;
            th->status = THREAD_RUNNABLE;

            // whoever woke this thread took the timeout back first
            VM_ASSERT(th->sched.waiting_reason.flags == thread_sched_waiting_none);
        }
    }
    thread_sched_unlock(sched, th);
    rb_ractor_lock_self(cr);

    ubf_clear(th, true);

    RUBY_DEBUG_LOG("end%s", "");
}

void
rb_ractor_sched_wakeup(rb_ractor_t *r, rb_thread_t *r_th)
{
    // ractor lock of r acquired
    struct rb_thread_sched *sched = TH_SCHED(r_th);

    RUBY_DEBUG_LOG("r:%u th:%d", (unsigned int)rb_ractor_id(r), r_th->serial);

    thread_sched_lock(sched, r_th);
    {
        if (r_th->status == THREAD_STOPPED_FOREVER) {
            RUBY_ATOMIC_ADD(r_th->unblock.event_serial, 1);

            // r_th must not resume with a wheel entry left behind: take its timeout
            // back, as ubf_event_waiting() does.  Only r_th arms it, and it is
            // parked here, so reading the flags without the timer lock is safe.
            if (r_th->sched.waiting_reason.flags != thread_sched_waiting_none) {
                ractor_sched_timeout_disarm(r_th);
            }

            // a timeout that fired first may have made r_th runnable already: waking
            // it twice would put it on the readyq twice
            if (sched->running != r_th && !r_th->sched.node.is_ready) {
                r_th->sched.event_serial++; // a timeout still armed must not wake it again
                thread_sched_to_ready_common(sched, r_th, true, false);
            }
        }
    }
    thread_sched_unlock(sched, r_th);
}

static bool
ractor_sched_barrier_completed_p(rb_vm_t *vm)
{
    // The snapshot barrier_running_cnt is taken by the barrier's walk and
    // decremented by counted deregistrations; no rescan is needed here.
    RUBY_DEBUG_LOG("run:%u wait:%u", vm->ractor.sched.barrier_running_cnt, vm->ractor.sched.barrier_joined_cnt);
    VM_ASSERT(vm->ractor.sched.barrier_running_cnt - 1 >= vm->ractor.sched.barrier_joined_cnt);

    return (vm->ractor.sched.barrier_running_cnt - vm->ractor.sched.barrier_joined_cnt) == 1;
}

void
rb_ractor_sched_barrier_start(rb_vm_t *vm, rb_ractor_t *cr)
{
    VM_ASSERT(cr == GET_RACTOR());
    VM_ASSERT(vm->ractor.sync.lock_owner == cr); // VM is locked
    VM_ASSERT(!vm->ractor.sched.barrier_is_waiting);
    VM_ASSERT(vm->ractor.sched.barrier_joined_cnt == 0);
    VM_ASSERT(vm->ractor.sched.barrier_ractor == NULL);
    VM_ASSERT(vm->ractor.sched.barrier_lock_rec == 0);

    RUBY_DEBUG_LOG("start serial:%u", vm->ractor.sched.barrier_serial);

    unsigned int lock_rec;

    ractor_sched_lock(vm, cr);
    {
        RUBY_ATOMIC_SET(vm->ractor.sched.barrier_is_waiting, 1);
        vm->ractor.sched.barrier_ractor = cr;
        vm->ractor.sched.barrier_lock_rec = vm->ractor.sync.lock_rec;

        // release VM lock
        lock_rec = vm->ractor.sync.lock_rec;
        vm->ractor.sync.lock_rec = 0;
        vm->ractor.sync.lock_owner = NULL;
        rb_native_mutex_unlock(&vm->ractor.sync.lock);

        // Interrupt all running threads: running_dnts plus each snt's running_th.
        // A switch before this scan is visible to it; one after it sees
        // barrier_is_waiting (set above) and waits.
        // Interrupt and count every registered runner, stamping each nt so a
        // deregistration during this barrier knows it was counted.
        rb_thread_t *ith;
        unsigned int running_cnt = 0;
        uint32_t serial = vm->ractor.sched.barrier_serial;

        rb_native_mutex_lock(&vm->ractor.sched.ntlist.lock);
        {
            struct rb_native_thread *dnt;
            ccan_list_for_each(&vm->ractor.sched.ntlist.running_dnts, dnt, running_dnts_node) {
                ith = dnt->running_thread;
                dnt->barrier_counted_serial = serial;
                running_cnt++;
                if (ith->ractor != cr) {
                    RUBY_DEBUG_LOG("barrier request to th:%u", rb_th_serial(ith));
                    RUBY_VM_SET_VM_BARRIER_INTERRUPT(ith->ec);
                }
            }

            struct rb_native_thread *nt;
            ccan_list_for_each(&vm->ractor.sched.ntlist.snts, nt, snts_node) {
                rb_native_mutex_lock(&nt->running_th_lock);
                {
                    ith = nt->running_th;
                    if (ith != NULL) {
                        nt->barrier_counted_serial = serial;
                        running_cnt++;
                        if (ith->ractor != cr) {
                            RUBY_DEBUG_LOG("barrier request to th:%u", rb_th_serial(ith));
                            RUBY_VM_SET_VM_BARRIER_INTERRUPT(ith->ec);
                        }
                    }
                }
                rb_native_mutex_unlock(&nt->running_th_lock);
            }
        }
        rb_native_mutex_unlock(&vm->ractor.sched.ntlist.lock);

        vm->ractor.sched.barrier_running_cnt = running_cnt;

        // wait for other ractors
        while (!ractor_sched_barrier_completed_p(vm)) {
            ractor_sched_set_unlocked(vm, cr);
            rb_native_cond_wait(&vm->ractor.sched.barrier_complete_cond, &vm->ractor.sched.lock);
            ractor_sched_set_locked(vm, cr);
        }

        RUBY_DEBUG_LOG("completed seirial:%u", vm->ractor.sched.barrier_serial);

        // no other ractors are there
        vm->ractor.sched.barrier_serial++;
        vm->ractor.sched.barrier_joined_cnt = 0;
        rb_native_cond_broadcast(&vm->ractor.sched.barrier_release_cond);

        // acquire VM lock
        rb_native_mutex_lock(&vm->ractor.sync.lock);
        vm->ractor.sync.lock_rec = lock_rec;
        vm->ractor.sync.lock_owner = cr;
    }

    // do not release ractor_sched_lock and there is no newly added (resumed) thread
    // thread_sched_setup_running_threads
}

// called from vm_lock_leave if the vm_lock used for barrierred
void
rb_ractor_sched_barrier_end(rb_vm_t *vm, rb_ractor_t *cr)
{
    RUBY_DEBUG_LOG("serial:%u", (unsigned int)vm->ractor.sched.barrier_serial - 1);
    VM_ASSERT(vm->ractor.sched.barrier_is_waiting);
    VM_ASSERT(vm->ractor.sched.barrier_ractor);
    VM_ASSERT(vm->ractor.sched.barrier_lock_rec > 0);

    RUBY_ATOMIC_SET(vm->ractor.sched.barrier_is_waiting, 0);
    vm->ractor.sched.barrier_ractor = NULL;
    vm->ractor.sched.barrier_lock_rec = 0;
    ractor_sched_unlock(vm, cr);
}

static void
ractor_sched_barrier_join_signal_locked(rb_vm_t *vm)
{
    if (ractor_sched_barrier_completed_p(vm)) {
        rb_native_cond_signal(&vm->ractor.sched.barrier_complete_cond);
    }
}

static void
ractor_sched_barrier_join_wait_locked(rb_vm_t *vm, rb_thread_t *th)
{
    VM_ASSERT(vm->ractor.sched.barrier_is_waiting);

    unsigned int barrier_serial = vm->ractor.sched.barrier_serial;

    while (vm->ractor.sched.barrier_serial == barrier_serial) {
        RUBY_DEBUG_LOG("sleep serial:%u", barrier_serial);
        RB_VM_SAVE_MACHINE_CONTEXT(th);

        rb_ractor_t *cr = th->ractor;
        ractor_sched_set_unlocked(vm, cr);
        rb_native_cond_wait(&vm->ractor.sched.barrier_release_cond, &vm->ractor.sched.lock);
        ractor_sched_set_locked(vm, cr);

        RUBY_DEBUG_LOG("wakeup serial:%u", barrier_serial);
    }
}

void
rb_ractor_sched_barrier_join(rb_vm_t *vm, rb_ractor_t *cr)
{
    VM_ASSERT(cr->threads.sched.running != NULL); // running ractor
    VM_ASSERT(cr == GET_RACTOR());
    VM_ASSERT(vm->ractor.sync.lock_owner == NULL); // VM is locked, but owner == NULL
    VM_ASSERT(vm->ractor.sched.barrier_is_waiting);  // VM needs barrier sync

#if USE_RUBY_DEBUG_LOG || VM_CHECK_MODE > 0
    unsigned int barrier_serial = vm->ractor.sched.barrier_serial;
#endif

    RUBY_DEBUG_LOG("join");

    rb_native_mutex_unlock(&vm->ractor.sync.lock);
    {
        VM_ASSERT(vm->ractor.sched.barrier_is_waiting);  // VM needs barrier sync
        VM_ASSERT(vm->ractor.sched.barrier_serial == barrier_serial);

        ractor_sched_lock(vm, cr);
        {
            // running_cnt
            /* Every joiner is a member of the running set: a dying thread
             * leaves the living set before handing over its scheduler slot. */
            vm->ractor.sched.barrier_joined_cnt++;
            RUBY_DEBUG_LOG("waiting_cnt:%u serial:%u", vm->ractor.sched.barrier_joined_cnt, barrier_serial);

            ractor_sched_barrier_join_signal_locked(vm);
            ractor_sched_barrier_join_wait_locked(vm, cr->threads.sched.running);
        }
        ractor_sched_unlock(vm, cr);
    }

    rb_native_mutex_lock(&vm->ractor.sync.lock);
    // VM locked here
}

// Called when the ractor holding this sched is freed.  A drained sched can
// still be on timeslice.scheds (pruning is lazy); an unlisted node is
// self-linked (fork re-inits them all), making this del a no-op.
void
rb_thread_sched_destroy(struct rb_thread_sched *sched)
{
    rb_vm_t *vm = GET_VM();

    rb_native_mutex_lock(&vm->ractor.sched.timeslice.lock);
    {
        ccan_list_del_init(&sched->timeslice_node);
    }
    rb_native_mutex_unlock(&vm->ractor.sched.timeslice.lock);
}

#ifdef RB_THREAD_T_HAS_NATIVE_ID
static int
get_native_thread_id(void)
{
#ifdef __linux__
    return (int)syscall(SYS_gettid);
#elif defined(__FreeBSD__)
    return pthread_getthreadid_np();
#endif
}
#endif

#if defined(HAVE_WORKING_FORK)
static void rb_internal_thread_event_hooks_rw_lock_atfork(void);

static void
thread_sched_atfork(struct rb_thread_sched *sched)
{
    current_fork_gen++;
    rb_thread_sched_init(sched, true);
    rb_thread_t *th =  GET_THREAD();
    rb_vm_t *vm = GET_VM();

    if (th_has_dedicated_nt(th)) {
        vm->ractor.sched.snt_cnt = 0;
#if USE_RUBY_DEBUG_LOG
        vm->ractor.sched.dnt_cnt = 1;
#endif
    }
    else {
        vm->ractor.sched.snt_cnt = 1;
#if USE_RUBY_DEBUG_LOG
        vm->ractor.sched.dnt_cnt = 0;
#endif
    }

    rb_native_mutex_initialize(&vm->ractor.sched.lock);
#if VM_CHECK_MODE > 0
    vm->ractor.sched.lock_owner = NULL;
    vm->ractor.sched.locked = false;
#endif

    // rb_native_cond_destroy(&vm->ractor.sched.cond);
    rb_native_cond_initialize(&vm->ractor.sched.cond);
    rb_native_cond_initialize(&vm->ractor.sched.barrier_complete_cond);
    rb_native_cond_initialize(&vm->ractor.sched.barrier_release_cond);

    ccan_list_head_init(&vm->ractor.sched.grq);
    vm->ractor.sched.grq_cnt = 0; // the list was just emptied; reset the count with it
    // A fork during a VM barrier leaves the child with barrier state that can
    // never complete (the other ractors are gone); reset it like the rest.
    vm->ractor.sched.barrier_is_waiting = 0; // single-threaded child
    vm->ractor.sched.barrier_joined_cnt = 0;
    vm->ractor.sched.barrier_ractor = NULL;
    vm->ractor.sched.barrier_lock_rec = 0;
    // Threads that were winding down in the parent do not exist in the child;
    // without this reset the child's ruby_vm_destruct would wait for their
    // reclaim (which never comes) forever.
    vm->ractor.sched.winding_cnt = 0;
    rb_native_mutex_initialize(&vm->ractor.sched.ntlist.lock);
    ccan_list_head_init(&vm->ractor.sched.ntlist.running_dnts);
    ccan_list_head_init(&vm->ractor.sched.ntlist.snts); // those nts are gone
    rb_native_mutex_initialize(&vm->ractor.sched.timeslice.lock);
    ccan_list_head_init(&vm->ractor.sched.timeslice.scheds);
    rb_native_mutex_initialize(&th->nt->running_th_lock); // a scan could hold it at fork
    // Fork can copy nodes linked (or torn mid-link); re-init every sched's
    // node so rb_thread_sched_destroy's del_init stays a no-op for them.
    rb_ractor_t *r;
    ccan_list_for_each(&vm->ractor.set, r, vmlr_node) {
        ccan_list_node_init(&r->threads.sched.timeslice_node);
    }
    ccan_list_for_each(&vm->ractor.terminated_set, r, vmlr_node) {
        ccan_list_node_init(&r->threads.sched.timeslice_node);
    }
    // th re-records itself below; the parent's record did not survive the lists
    if (th->nt && th->nt->dedicated == 0) {
        // surviving on an snt: put that nt back on the (just emptied) snts
        // list, or the scans could not see this thread's record
        ccan_list_add(&vm->ractor.sched.ntlist.snts, &th->nt->snts_node);
    }

#if USE_MN_THREADS
    nt_machine_stack_atfork();
#endif
    rb_internal_thread_event_hooks_rw_lock_atfork();

    VM_ASSERT(sched->is_running);

    if (sched->running != th) {
        thread_sched_to_running(sched, th);
    }
    else {
        thread_sched_setup_running_threads(sched, th->ractor, vm, th, NULL);
    }

#ifdef RB_THREAD_T_HAS_NATIVE_ID
    if (th->nt) {
        th->nt->tid = get_native_thread_id();
    }
#endif
}

#endif

#ifdef RB_THREAD_LOCAL_SPECIFIER
static RB_THREAD_LOCAL_SPECIFIER rb_thread_t *ruby_native_thread;
#else
static pthread_key_t ruby_native_thread_key;
#endif

static void
null_func(int i)
{
    /* null */
    // This function can be called from signal handler
    // RUBY_DEBUG_LOG("i:%d", i);
}

rb_thread_t *
ruby_thread_from_native(void)
{
#ifdef RB_THREAD_LOCAL_SPECIFIER
    return ruby_native_thread;
#else
    return pthread_getspecific(ruby_native_thread_key);
#endif
}

int
ruby_thread_set_native(rb_thread_t *th)
{
    if (th) {
#ifdef USE_UBF_LIST
        ccan_list_node_init(&th->sched.node.ubf);
#endif
    }

    // setup TLS

    if (th && th->ec) {
        rb_ractor_set_current_ec(th->ractor, th->ec);
    }
#ifdef RB_THREAD_LOCAL_SPECIFIER
    ruby_native_thread = th;
    return 1;
#else
    return pthread_setspecific(ruby_native_thread_key, th) == 0;
#endif
}

static void native_thread_setup(struct rb_native_thread *nt);
static void native_thread_setup_on_thread(struct rb_native_thread *nt);

// Internal cache of page size:
static size_t RB_THREAD_PAGE_SIZE;

void
Init_native_thread(rb_thread_t *main_th)
{
    // Get the system page size for later use in stack allocation and stack overflow checks:
    RB_THREAD_PAGE_SIZE = sysconf(_SC_PAGESIZE);

#if defined(HAVE_PTHREAD_CONDATTR_SETCLOCK)
    if (condattr_monotonic) {
        int r = pthread_condattr_init(condattr_monotonic);
        if (r == 0) {
            r = pthread_condattr_setclock(condattr_monotonic, CLOCK_MONOTONIC);
        }
        if (r) condattr_monotonic = NULL;
    }
#endif

#ifndef RB_THREAD_LOCAL_SPECIFIER
    if (pthread_key_create(&ruby_native_thread_key, 0) == EAGAIN) {
        rb_bug("pthread_key_create failed (ruby_native_thread_key)");
    }
    if (pthread_key_create(&ruby_current_ec_key, 0) == EAGAIN) {
        rb_bug("pthread_key_create failed (ruby_current_ec_key)");
    }
#endif
    ruby_posix_signal(SIGVTALRM, null_func);

    // setup vm
    rb_vm_t *vm = main_th->vm;
    rb_native_mutex_initialize(&vm->ractor.sched.lock);
    rb_native_cond_initialize(&vm->ractor.sched.cond);
    rb_native_cond_initialize(&vm->ractor.sched.barrier_complete_cond);
    rb_native_cond_initialize(&vm->ractor.sched.barrier_release_cond);

    ccan_list_head_init(&vm->ractor.sched.grq);
    rb_native_mutex_initialize(&vm->ractor.sched.ntlist.lock);
    ccan_list_head_init(&vm->ractor.sched.ntlist.running_dnts);
    ccan_list_head_init(&vm->ractor.sched.ntlist.snts);
    rb_native_mutex_initialize(&vm->ractor.sched.timeslice.lock);
    ccan_list_head_init(&vm->ractor.sched.timeslice.scheds);

    // setup main thread
    main_th->nt->thread_id = pthread_self();
    main_th->nt->serial = 1;
#ifdef RUBY_NT_SERIAL
    ruby_nt_serial = 1;
#endif
    ruby_thread_set_native(main_th);
    native_thread_setup(main_th->nt);
    native_thread_setup_on_thread(main_th->nt);

    TH_SCHED(main_th)->running = main_th;
    main_th->has_dedicated_nt = 1;

    // setup main NT (before the record below: its kind decides where it goes)
    main_th->nt->dedicated = 1;
    main_th->nt->running_thread = main_th;
    main_th->nt->vm = vm;

    thread_sched_setup_running_threads(TH_SCHED(main_th), main_th->ractor, vm, main_th, NULL);

    // setup mn
#if USE_RUBY_DEBUG_LOG
    vm->ractor.sched.dnt_cnt = 1;
#endif
}

extern int ruby_mn_threads_enabled;

void
ruby_mn_threads_params(void)
{
    rb_vm_t *vm = GET_VM();
    rb_ractor_t *main_ractor = GET_RACTOR();

    const char *mn_threads_cstr = getenv("RUBY_MN_THREADS");
    bool enable_mn_threads = false;

    if (USE_MN_THREADS && mn_threads_cstr && (enable_mn_threads = atoi(mn_threads_cstr) > 0)) {
        // enabled
        ruby_mn_threads_enabled = 1;
    }
    main_ractor->threads.sched.enable_mn_threads = enable_mn_threads;

    const char *max_cpu_cstr = getenv("RUBY_MAX_CPU");
#if defined(HAVE_SYSCONF) && defined(_SC_NPROCESSORS_ONLN)
    long nprocessors = sysconf(_SC_NPROCESSORS_ONLN);
    const int default_max_cpu = (nprocessors > 0) ? (int)nprocessors : 8;
#else
    const int default_max_cpu = 8;
#endif
    int max_cpu = default_max_cpu;

    if (USE_MN_THREADS && max_cpu_cstr)  {
        int given_max_cpu = atoi(max_cpu_cstr);
        if (given_max_cpu > 0) {
            max_cpu = given_max_cpu;
        }
    }

    vm->ractor.sched.max_cpu = max_cpu;
}

static void
native_thread_dedicated_inc(rb_vm_t *vm, rb_ractor_t *cr, struct rb_native_thread *nt)
{
    RUBY_DEBUG_LOG("nt:%d %d->%d", nt->serial, nt->dedicated, nt->dedicated + 1);

    if (nt->dedicated == 0) {
        // Lock-free; pairs with ractor_sched_enq (enq: grq_cnt up then read
        // snt_cnt / here: snt_cnt down then read grq_cnt) against lost wakeups.
        if (RUBY_ATOMIC_FETCH_SUB(vm->ractor.sched.snt_cnt, 1) == 1) {
            // the last snt went dedicated; pending entries need the timer thread
            ractor_sched_lock(vm, cr);
            {
                if (vm->ractor.sched.grq_cnt > 0) {
                    timer_thread_wakeup_locked(vm);
                }
            }
            ractor_sched_unlock(vm, cr);
        }
#if USE_RUBY_DEBUG_LOG
        vm->ractor.sched.dnt_cnt++;
#endif
    }

    nt->dedicated++;
}

static void
native_thread_dedicated_dec(rb_vm_t *vm, rb_ractor_t *cr, struct rb_native_thread *nt)
{
    RUBY_DEBUG_LOG("nt:%d %d->%d", nt->serial, nt->dedicated, nt->dedicated - 1);
    VM_ASSERT(nt->dedicated > 0);
    nt->dedicated--;

    if (nt->dedicated == 0) {
        // Rejoin under the max_cpu cap; with no room this nt retires and
        // belongs to neither count until it ends.
        while (1) {
            rb_atomic_t snt = RUBY_ATOMIC_LOAD(vm->ractor.sched.snt_cnt);
            if (snt < vm->ractor.sched.max_cpu || (int)snt <= MINIMUM_SNT) {
                if (RUBY_ATOMIC_CAS(vm->ractor.sched.snt_cnt, snt, snt + 1) == snt) break;
            }
            else {
                nt->retiring = true;
                break;
            }
        }
#if USE_RUBY_DEBUG_LOG
        vm->ractor.sched.dnt_cnt--;
#endif
    }
}

static void
native_thread_assign(struct rb_native_thread *nt, rb_thread_t *th)
{
#if USE_RUBY_DEBUG_LOG
    if (nt) {
        if (th->nt) {
            RUBY_DEBUG_LOG("th:%d nt:%d->%d", (int)th->serial, (int)th->nt->serial, (int)nt->serial);
        }
        else {
            RUBY_DEBUG_LOG("th:%d nt:NULL->%d", (int)th->serial, (int)nt->serial);
        }
    }
    else {
        if (th->nt) {
            RUBY_DEBUG_LOG("th:%d nt:%d->NULL", (int)th->serial, (int)th->nt->serial);
        }
        else {
            RUBY_DEBUG_LOG("th:%d nt:NULL->NULL", (int)th->serial);
        }
    }
#endif

    th->nt = nt;
}

static void
native_thread_destroy_atfork(struct rb_native_thread *nt)
{
    if (nt) {
        /* We can't call rb_native_cond_destroy here because according to the
         * specs of pthread_cond_destroy:
         *
         *   Attempting to destroy a condition variable upon which other threads
         *   are currently blocked results in undefined behavior.
         *
         * Specifically, glibc's pthread_cond_destroy waits on all the other
         * listeners. Since after forking all the threads are dead, the condition
         * variable's listeners will never wake up, so it will hang forever.
         */

        RB_ALTSTACK_FREE(nt->altstack);
        SIZED_FREE(nt->nt_context);
        SIZED_FREE(nt);
    }
}

static void
native_thread_destroy(struct rb_native_thread *nt)
{
    if (nt) {
        rb_native_cond_destroy(&nt->readyq);
        rb_native_mutex_destroy(&nt->running_th_lock);

        native_thread_destroy_atfork(nt);
    }
}

// A retiring snt frees its own rb_native_thread as it exits.  Disarm the
// altstack registration first: a signal between the free and the thread's
// end must not run on the freed block.
static void
native_thread_destroy_self(struct rb_native_thread *nt)
{
#ifdef USE_SIGALTSTACK
    stack_t disable = {0};
    disable.ss_flags = SS_DISABLE;
    sigaltstack(&disable, NULL);
#endif
    native_thread_destroy(nt);
}

#if defined HAVE_PTHREAD_GETATTR_NP || defined HAVE_PTHREAD_ATTR_GET_NP
#define STACKADDR_AVAILABLE 1
#elif defined HAVE_PTHREAD_GET_STACKADDR_NP && defined HAVE_PTHREAD_GET_STACKSIZE_NP
#define STACKADDR_AVAILABLE 1
#undef MAINSTACKADDR_AVAILABLE
#define MAINSTACKADDR_AVAILABLE 1
void *pthread_get_stackaddr_np(pthread_t);
size_t pthread_get_stacksize_np(pthread_t);
#elif defined HAVE_THR_STKSEGMENT || defined HAVE_PTHREAD_STACKSEG_NP
#define STACKADDR_AVAILABLE 1
#elif defined HAVE_PTHREAD_GETTHRDS_NP
#define STACKADDR_AVAILABLE 1
#elif defined __HAIKU__
#define STACKADDR_AVAILABLE 1
#endif

#ifndef MAINSTACKADDR_AVAILABLE
# ifdef STACKADDR_AVAILABLE
#   define MAINSTACKADDR_AVAILABLE 1
# else
#   define MAINSTACKADDR_AVAILABLE 0
# endif
#endif
#if MAINSTACKADDR_AVAILABLE && !defined(get_main_stack)
# define get_main_stack(addr, size) get_stack(addr, size)
#endif

#ifdef STACKADDR_AVAILABLE
/*
 * Get the initial address and size of current thread's stack
 */
static int
get_stack(void **addr, size_t *size)
{
#define CHECK_ERR(expr)				\
    {int err = (expr); if (err) return err;}
#ifdef HAVE_PTHREAD_GETATTR_NP /* Linux */
    pthread_attr_t attr;
    size_t guard = 0;
    STACK_GROW_DIR_DETECTION;
    CHECK_ERR(pthread_getattr_np(pthread_self(), &attr));
# ifdef HAVE_PTHREAD_ATTR_GETSTACK
    CHECK_ERR(pthread_attr_getstack(&attr, addr, size));
    STACK_DIR_UPPER((void)0, (void)(*addr = (char *)*addr + *size));
# else
    CHECK_ERR(pthread_attr_getstackaddr(&attr, addr));
    CHECK_ERR(pthread_attr_getstacksize(&attr, size));
# endif
# ifdef HAVE_PTHREAD_ATTR_GETGUARDSIZE
    CHECK_ERR(pthread_attr_getguardsize(&attr, &guard));
# else
    guard = RB_THREAD_PAGE_SIZE;
# endif
    *size -= guard;
    pthread_attr_destroy(&attr);
#elif defined HAVE_PTHREAD_ATTR_GET_NP /* FreeBSD, DragonFly BSD, NetBSD */
    pthread_attr_t attr;
    CHECK_ERR(pthread_attr_init(&attr));
    CHECK_ERR(pthread_attr_get_np(pthread_self(), &attr));
# ifdef HAVE_PTHREAD_ATTR_GETSTACK
    CHECK_ERR(pthread_attr_getstack(&attr, addr, size));
# else
    CHECK_ERR(pthread_attr_getstackaddr(&attr, addr));
    CHECK_ERR(pthread_attr_getstacksize(&attr, size));
# endif
    STACK_DIR_UPPER((void)0, (void)(*addr = (char *)*addr + *size));
    pthread_attr_destroy(&attr);
#elif (defined HAVE_PTHREAD_GET_STACKADDR_NP && defined HAVE_PTHREAD_GET_STACKSIZE_NP) /* MacOS X */
    pthread_t th = pthread_self();
    *addr = pthread_get_stackaddr_np(th);
    *size = pthread_get_stacksize_np(th);
#elif defined HAVE_THR_STKSEGMENT || defined HAVE_PTHREAD_STACKSEG_NP
    stack_t stk;
# if defined HAVE_THR_STKSEGMENT /* Solaris */
    CHECK_ERR(thr_stksegment(&stk));
# else /* OpenBSD */
    CHECK_ERR(pthread_stackseg_np(pthread_self(), &stk));
# endif
    *addr = stk.ss_sp;
    *size = stk.ss_size;
#elif defined HAVE_PTHREAD_GETTHRDS_NP /* AIX */
    pthread_t th = pthread_self();
    struct __pthrdsinfo thinfo;
    char reg[256];
    int regsiz=sizeof(reg);
    CHECK_ERR(pthread_getthrds_np(&th, PTHRDSINFO_QUERY_ALL,
                                   &thinfo, sizeof(thinfo),
                                   &reg, &regsiz));
    *addr = thinfo.__pi_stackaddr;
    /* Must not use thinfo.__pi_stacksize for size.
       It is around 3KB smaller than the correct size
       calculated by thinfo.__pi_stackend - thinfo.__pi_stackaddr. */
    *size = thinfo.__pi_stackend - thinfo.__pi_stackaddr;
    STACK_DIR_UPPER((void)0, (void)(*addr = (char *)*addr + *size));
#elif defined __HAIKU__
    thread_info info;
    STACK_GROW_DIR_DETECTION;
    CHECK_ERR(get_thread_info(find_thread(NULL), &info));
    *addr = info.stack_base;
    *size = (uintptr_t)info.stack_end - (uintptr_t)info.stack_base;
    STACK_DIR_UPPER((void)0, (void)(*addr = (char *)*addr + *size));
#else
#error STACKADDR_AVAILABLE is defined but not implemented.
#endif
    return 0;
#undef CHECK_ERR
}
#endif

static struct {
    rb_nativethread_id_t id;
    size_t stack_maxsize;
    VALUE *stack_start;
} native_main_thread;

#ifdef STACK_END_ADDRESS
extern void *STACK_END_ADDRESS;
#endif

static void
native_thread_init_main_thread_stack(void *addr)
{
    native_main_thread.id = pthread_self();
#ifdef RUBY_ASAN_ENABLED
    addr = asan_get_real_stack_addr((void *)addr);
#endif

#if MAINSTACKADDR_AVAILABLE
    if (native_main_thread.stack_maxsize) return;
    {
        void* stackaddr;
        size_t size;
        if (get_main_stack(&stackaddr, &size) == 0) {
            native_main_thread.stack_maxsize = size;
            native_main_thread.stack_start = stackaddr;
            goto bound_check;
        }
    }
#endif
#ifdef STACK_END_ADDRESS
    native_main_thread.stack_start = STACK_END_ADDRESS;
#else
    if (!native_main_thread.stack_start ||
        STACK_UPPER((VALUE *)(void *)&addr,
                    native_main_thread.stack_start > (VALUE *)addr,
                    native_main_thread.stack_start < (VALUE *)addr)) {
        native_main_thread.stack_start = (VALUE *)addr;
    }
#endif
    {
#if defined(HAVE_GETRLIMIT)
#if defined(PTHREAD_STACK_DEFAULT)
        size_t size = PTHREAD_STACK_DEFAULT;
#else
        size_t size = RUBY_VM_THREAD_VM_STACK_SIZE;
#endif
        size_t space;
        struct rlimit rlim;
        STACK_GROW_DIR_DETECTION;
        if (getrlimit(RLIMIT_STACK, &rlim) == 0) {
            size = (size_t)rlim.rlim_cur;
        }
        addr = native_main_thread.stack_start;
        if (IS_STACK_DIR_UPPER()) {
            space = ((size_t)((char *)addr + size) / RB_THREAD_PAGE_SIZE) * RB_THREAD_PAGE_SIZE - (size_t)addr;
        }
        else {
            space = (size_t)addr - ((size_t)((char *)addr - size) / RB_THREAD_PAGE_SIZE + 1) * RB_THREAD_PAGE_SIZE;
        }
        native_main_thread.stack_maxsize = space;
#endif
    }

#if MAINSTACKADDR_AVAILABLE
  bound_check:
#endif
    /* If addr is out of range of main-thread stack range estimation,  */
    /* it should be on co-routine (alternative stack). [Feature #2294] */
    {
        void *start, *end;
        STACK_GROW_DIR_DETECTION;

        if (IS_STACK_DIR_UPPER()) {
            start = native_main_thread.stack_start;
            end = (char *)native_main_thread.stack_start + native_main_thread.stack_maxsize;
        }
        else {
            start = (char *)native_main_thread.stack_start - native_main_thread.stack_maxsize;
            end = native_main_thread.stack_start;
        }

        if ((void *)addr < start || (void *)addr > end) {
            /* out of range */
            native_main_thread.stack_start = (VALUE *)addr;
            native_main_thread.stack_maxsize = 0; /* unknown */
        }
    }
}

#define CHECK_ERR(expr) \
    {int err = (expr); if (err) {rb_bug_errno(#expr, err);}}

static int
native_thread_init_stack(rb_thread_t *th, void *local_in_parent_frame)
{
    rb_nativethread_id_t curr = pthread_self();
#ifdef RUBY_ASAN_ENABLED
    local_in_parent_frame = asan_get_real_stack_addr(local_in_parent_frame);
    th->ec->machine.asan_fake_stack_handle = asan_get_thread_fake_stack_handle();
#endif

    if (!native_main_thread.id) {
        /* This thread is the first thread, must be the main thread -
         * configure the native_main_thread object */
        native_thread_init_main_thread_stack(local_in_parent_frame);
    }

    if (pthread_equal(curr, native_main_thread.id)) {
        th->ec->machine.stack_start = native_main_thread.stack_start;
        th->ec->machine.stack_maxsize = native_main_thread.stack_maxsize;
    }
    else {
#ifdef STACKADDR_AVAILABLE
        if (th_has_dedicated_nt(th)) {
            void *start;
            size_t size;

            if (get_stack(&start, &size) == 0) {
                uintptr_t diff = (uintptr_t)start - (uintptr_t)local_in_parent_frame;
                th->ec->machine.stack_start = local_in_parent_frame;
                th->ec->machine.stack_maxsize = size - diff;
            }
        }
#else
        rb_raise(rb_eNotImpError, "ruby engine can initialize only in the main thread");
#endif
    }

    return 0;
}

struct nt_param {
    rb_vm_t *vm;
    struct rb_native_thread *nt;
};

static void *
nt_start(void *ptr);

static int
native_thread_create0(struct rb_native_thread *nt)
{
    int err = 0;
    pthread_attr_t attr;

    const size_t stack_size = nt->vm->default_params.thread_machine_stack_size;

#ifdef USE_SIGALTSTACK
    nt->altstack = rb_allocate_sigaltstack();
#endif

    CHECK_ERR(pthread_attr_init(&attr));

# ifdef PTHREAD_STACK_MIN
    RUBY_DEBUG_LOG("stack size: %lu", (unsigned long)stack_size);
    CHECK_ERR(pthread_attr_setstacksize(&attr, stack_size));
# endif

# ifdef HAVE_PTHREAD_ATTR_SETINHERITSCHED
    CHECK_ERR(pthread_attr_setinheritsched(&attr, PTHREAD_INHERIT_SCHED));
# endif
    CHECK_ERR(pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_DETACHED));

    err = pthread_create(&nt->thread_id, &attr, nt_start, nt);

    RUBY_DEBUG_LOG("nt:%d err:%d", (int)nt->serial, err);

    CHECK_ERR(pthread_attr_destroy(&attr));

    return err;
}

static void
native_thread_setup(struct rb_native_thread *nt)
{
    // init cond
    rb_native_cond_initialize(&nt->readyq);
    // also for the main thread's nt: a zero-filled mutex is not usable on macOS
    rb_native_mutex_initialize(&nt->running_th_lock);
}

static void
native_thread_setup_on_thread(struct rb_native_thread *nt)
{
    // init tid
#ifdef RB_THREAD_T_HAS_NATIVE_ID
    nt->tid = get_native_thread_id();
#endif

    // init signal handler
    RB_ALTSTACK_INIT(nt->altstack, nt->altstack);
}

static struct rb_native_thread *
native_thread_alloc(void)
{
    struct rb_native_thread *nt = ZALLOC(struct rb_native_thread);
    native_thread_setup(nt);

#if USE_MN_THREADS
    nt->nt_context = ruby_xmalloc(sizeof(struct coroutine_context));
#endif

#if USE_RUBY_DEBUG_LOG
    static rb_atomic_t nt_serial = 2;
    nt->serial = RUBY_ATOMIC_FETCH_ADD(nt_serial, 1);
#endif
    return nt;
}

static int
native_thread_create_dedicated(rb_thread_t *th)
{
    th->nt = native_thread_alloc();
    th->nt->vm = th->vm;
    th->nt->running_thread = th;
    th->nt->dedicated = 1;

    // vm stack
    size_t vm_stack_word_size = th->vm->default_params.thread_vm_stack_size / sizeof(VALUE);
    void *vm_stack = ruby_xmalloc(vm_stack_word_size * sizeof(VALUE));
    th->sched.malloc_stack = true;
    rb_ec_initialize_vm_stack(th->ec, vm_stack, vm_stack_word_size);
    th->sched.context_stack = vm_stack;
    th->sched.context_stack_size = vm_stack_word_size;

    int err = native_thread_create0(th->nt);
    if (!err) {
        // setup
        thread_sched_to_ready(TH_SCHED(th), th);
    }
    return err;
}

static void
call_thread_start_func_2(rb_thread_t *th)
{
    /* Capture the address of a local in this stack frame to mark the beginning of the
       machine stack for this thread. This is required even if we can tell the real
       stack beginning from the pthread API in native_thread_init_stack, because
       glibc stores some of its own data on the stack before calling into user code
       on a new thread, and replacing that data on fiber-switch would break it (see
       bug #13887) */
    VALUE stack_start = 0;
    VALUE *stack_start_addr = asan_get_real_stack_addr(&stack_start);

    native_thread_init_stack(th, stack_start_addr);
    thread_start_func_2(th, th->ec->machine.stack_start);
}

static void *
nt_start(void *ptr)
{
    struct rb_native_thread *nt = (struct rb_native_thread *)ptr;
    rb_vm_t *vm = nt->vm;

    native_thread_setup_on_thread(nt);

    // init tid
#ifdef RB_THREAD_T_HAS_NATIVE_ID
    nt->tid = get_native_thread_id();
#endif

#if USE_RUBY_DEBUG_LOG && defined(RUBY_NT_SERIAL)
    ruby_nt_serial = nt->serial;
#endif

    RUBY_DEBUG_LOG("nt:%u", nt->serial);

    bool in_snts = false;

    if (!nt->dedicated) {
        coroutine_initialize_main(nt->nt_context);

        // join the snt list that the barrier/timeslice scans walk
        rb_native_mutex_lock(&vm->ractor.sched.ntlist.lock);
        {
            ccan_list_add(&vm->ractor.sched.ntlist.snts, &nt->snts_node);
        }
        rb_native_mutex_unlock(&vm->ractor.sched.ntlist.lock);
        in_snts = true;
    }

    bool retired = false;

    while (1) {
        if (nt->dedicated) {
            // wait running turn
            rb_thread_t *th = nt->running_thread;
            struct rb_thread_sched *sched = TH_SCHED(th);

            RUBY_DEBUG_LOG("on dedicated th:%u", rb_th_serial(th));
            ruby_thread_set_native(th);

            thread_sched_lock(sched, th);
            {
                if (sched->running == th) {
                    thread_sched_add_running_thread(sched, th);
                }
                thread_sched_wait_running_turn(sched, th, false, NULL);
            }
            thread_sched_unlock(sched, th);

            // start threads
            call_thread_start_func_2(th);
            break; // TODO: allow to change to the SNT
        }
        else {
            RUBY_DEBUG_LOG("check next");
            if (nt->retiring) {   // came back with no room in the shared pool
                retired = true;
                break;
            }

            rb_ractor_t *r = ractor_sched_deq(vm, NULL);

            if (r) {
                struct rb_thread_sched *sched = &r->threads.sched;

                bool locked = true;

                thread_sched_lock(sched, NULL);
                {
                    rb_thread_t *next_th = sched->running;

                    if (next_th && next_th->nt == NULL) {
                        RUBY_DEBUG_LOG("nt:%d next_th:%d", (int)nt->serial, (int)next_th->serial);
#if USE_MN_THREADS
                        thread_sched_switch0(nt->nt_context, next_th, nt, false);

                        // If a coroutine terminated during the transfer, co_start
                        // recorded it in nt->dead_co (switch0's return value is
                        // backend-dependent, unusable; see thread_pthread.h).
                        struct coroutine_context *dead_co = nt->dead_co;
                        nt->dead_co = NULL;
                        if (thread_sched_reclaim(dead_co)) {
                            // it already released the sched lock before its
                            // transfer (its Ractor may be gone): leave sched be.
                            locked = false;
                        }
#else
                        thread_sched_switch0(nt->nt_context, next_th, nt, false);
#endif
                    }
                    else {
                        RUBY_DEBUG_LOG("no schedulable threads -- next_th:%p", next_th);
                    }
                }
                if (locked) {
                    thread_sched_unlock(sched, NULL);
                }
            }
            else {
                // ractor_sched_deq retired this nt.
                retired = true;
                break;
            }

            if (nt->dedicated) {
                // SNT becomes DNT while running
                break;
            }
        }
    }

    if (in_snts) {
        // Leaving the shared loop: every path back here deregistered first
        // (park and death both precede the transfer), so only the snts entry
        // is left to remove.
        VM_ASSERT(nt->running_th == NULL);
        rb_native_mutex_lock(&vm->ractor.sched.ntlist.lock);
        {
            ccan_list_del_init(&nt->snts_node);
        }
        rb_native_mutex_unlock(&vm->ractor.sched.ntlist.lock);
    }

    if (retired) {
        // The counts dropped this nt already; nothing can reference it now.
        RUBY_DEBUG_LOG("retired nt:%u", nt->serial);
        native_thread_destroy_self(nt);
    }

    return NULL;
}

static int native_thread_create_shared(rb_thread_t *th);

#if USE_MN_THREADS
static void nt_free_stack(void *mstack);


// Reclaim the context a coroutine thread recorded in nt->dead_co before its
// final transfer (co_start's epilogue). Our running here proves that transfer's
// register save into the block completed. Returns true when a thread did
// terminate -- it RELEASED the sched lock before transferring; NULL/false means
// a live yield, where the loop still owns the lock.
static bool
thread_sched_reclaim(struct coroutine_context *dead_co)
{
    struct rb_thread_context *tctx = (struct rb_thread_context *)dead_co;

    if (tctx != NULL && tctx->dead) {
        nt_free_stack(tctx->stack);
        SIZED_FREE(tctx);
        // pairs with the increment at the top of coroutine_thread_terminated:
        // a waiting VM destruct may proceed once this reclaim is done
        VM_ASSERT(RUBY_ATOMIC_LOAD(GET_VM()->ractor.sched.winding_cnt) > 0);
        RUBY_ATOMIC_DEC(GET_VM()->ractor.sched.winding_cnt);
        return true;
    }
    return false;
}
#endif

void
rb_thread_wake_fence(rb_thread_t *th)
{
    timer_thread_wake_fence(th);
}

void
rb_threadptr_sched_free(rb_thread_t *th)
{
    timer_thread_wake_fence(th);
#if USE_MN_THREADS
    if (th->sched.malloc_stack) {
        // has dedicated
        SIZED_FREE_N((VALUE *)th->sched.context_stack, th->sched.context_stack_size);
        native_thread_destroy(th->nt);
    }
    else if (th->sched.context != NULL) {
        // a coroutine thread that never reached its epilogue (never started);
        // a terminated one is reclaimed by whoever resumed from its final
        // transfer (thread_sched_reclaim), and cleared this pointer.
        struct rb_thread_context *tctx = (struct rb_thread_context *)th->sched.context;
        nt_free_stack(tctx->stack);
        SIZED_FREE(tctx);
        th->sched.context = NULL;
        // TODO: how to free nt and nt->altstack?
    }
#else
    SIZED_FREE_N((VALUE *)th->sched.context_stack, th->sched.context_stack_size);
    native_thread_destroy(th->nt);
#endif

    th->nt = NULL;
}


static int
native_thread_create(rb_thread_t *th)
{
    VM_ASSERT(th->nt == 0);
    RUBY_DEBUG_LOG("th:%d has_dnt:%d", th->serial, th->has_dedicated_nt);
    RB_INTERNAL_THREAD_HOOK(RUBY_INTERNAL_THREAD_EVENT_STARTED, th);

    if (!th->ractor->threads.sched.enable_mn_threads) {
        th->has_dedicated_nt = 1;
    }

    if (th->has_dedicated_nt) {
        return native_thread_create_dedicated(th);
    }
    else {
        return native_thread_create_shared(th);
    }
}

#if USE_NATIVE_THREAD_PRIORITY

static void
native_thread_apply_priority(rb_thread_t *th)
{
#if defined(_POSIX_PRIORITY_SCHEDULING) && (_POSIX_PRIORITY_SCHEDULING > 0)
    struct sched_param sp;
    int policy;
    int priority = 0 - th->priority;
    int max, min;
    pthread_getschedparam(th->nt->thread_id, &policy, &sp);
    max = sched_get_priority_max(policy);
    min = sched_get_priority_min(policy);

    if (min > priority) {
        priority = min;
    }
    else if (max < priority) {
        priority = max;
    }

    sp.sched_priority = priority;
    pthread_setschedparam(th->nt->thread_id, policy, &sp);
#else
    /* not touched */
#endif
}

#endif /* USE_NATIVE_THREAD_PRIORITY */

static int
native_fd_select(int n, rb_fdset_t *readfds, rb_fdset_t *writefds, rb_fdset_t *exceptfds, struct timeval *timeout, rb_thread_t *th)
{
    return rb_fd_select(n, readfds, writefds, exceptfds, timeout);
}

#ifdef USE_UBF_LIST
static CCAN_LIST_HEAD(ubf_list_head);
static rb_nativethread_lock_t ubf_list_lock = RB_NATIVETHREAD_LOCK_INIT;

static void
ubf_list_atfork(void)
{
    ccan_list_head_init(&ubf_list_head);
    rb_native_mutex_initialize(&ubf_list_lock);
}

RBIMPL_ATTR_MAYBE_UNUSED()
static bool
ubf_list_contain_p(rb_thread_t *th)
{
    rb_thread_t *list_th;
    ccan_list_for_each(&ubf_list_head, list_th, sched.node.ubf) {
        if (list_th == th) return true;
    }
    return false;
}

/* The thread 'th' is registered to be trying unblock. */
static void
register_ubf_list(rb_thread_t *th)
{
    RUBY_DEBUG_LOG("th:%u", rb_th_serial(th));
    struct ccan_list_node *node = &th->sched.node.ubf;

    VM_ASSERT(th->unblock.func != NULL);

    rb_native_mutex_lock(&ubf_list_lock);
    {
        // check not connected yet
        if (ccan_list_empty((struct ccan_list_head*)node)) {
            VM_ASSERT(!ubf_list_contain_p(th));
            ccan_list_add(&ubf_list_head, node);
        }
    }
    rb_native_mutex_unlock(&ubf_list_lock);

    timer_thread_wakeup();
}

/* The thread 'th' is unblocked. It no longer need to be registered. */
static void
unregister_ubf_list(rb_thread_t *th)
{
    RUBY_DEBUG_LOG("th:%u", rb_th_serial(th));
    struct ccan_list_node *node = &th->sched.node.ubf;

    /* we can't allow re-entry into ubf_list_head */
    VM_ASSERT(th->unblock.func == NULL);

    if (!ccan_list_empty((struct ccan_list_head*)node)) {
        rb_native_mutex_lock(&ubf_list_lock);
        {
            VM_ASSERT(ubf_list_contain_p(th));
            ccan_list_del_init(node);
        }
        rb_native_mutex_unlock(&ubf_list_lock);
    }
}

/*
 * send a signal to intent that a target thread return from blocking syscall.
 * Maybe any signal is ok, but we chose SIGVTALRM.
 */
static void
ubf_wakeup_thread(rb_thread_t *th)
{
    RUBY_DEBUG_LOG("th:%u thread_id:%p", rb_th_serial(th), (void *)th->nt->thread_id);

    pthread_kill(th->nt->thread_id, SIGVTALRM);
}

static void
ubf_select(void *ptr)
{
    rb_thread_t *th = (rb_thread_t *)ptr;
    RUBY_DEBUG_LOG("wakeup th:%u", rb_th_serial(th));
    ubf_wakeup_thread(th);
    register_ubf_list(th);
}

static bool
ubf_threads_empty(void)
{
    return ccan_list_empty(&ubf_list_head) != 0;
}

static void
ubf_wakeup_all_threads(void)
{
    rb_thread_t *th;
    rb_native_mutex_lock(&ubf_list_lock);
    {
        ccan_list_for_each(&ubf_list_head, th, sched.node.ubf) {
            ubf_wakeup_thread(th);
        }
    }
    rb_native_mutex_unlock(&ubf_list_lock);
}

#else /* USE_UBF_LIST */
#define register_ubf_list(th) (void)(th)
#define unregister_ubf_list(th) (void)(th)
#define ubf_select 0
static void ubf_wakeup_all_threads(void) { return; }
static bool ubf_threads_empty(void) { return true; }
#define ubf_list_atfork() do {} while (0)
#endif /* USE_UBF_LIST */

#define TT_DEBUG 0
#define WRITE_CONST(fd, str) (void)(write((fd),(str),sizeof(str)-1)<0)

void
rb_thread_wakeup_timer_thread(int sig)
{
    // This function can be called from signal handlers so that
    // pthread_mutex_lock() should not be used.

    // wakeup timer thread
    timer_thread_wakeup_force();

    // interrupt main thread if main thread is available
    if (RUBY_ATOMIC_LOAD(system_working)) {
        rb_vm_t *vm = GET_VM();
        rb_thread_t *main_th = vm->ractor.main_thread;

        if (main_th) {
            volatile rb_execution_context_t *main_th_ec = ACCESS_ONCE(rb_execution_context_t *, main_th->ec);

            if (main_th_ec) {
                RUBY_VM_SET_TRAP_INTERRUPT(main_th_ec);

                if (vm->ubf_async_safe && main_th->unblock.func) {
                    (main_th->unblock.func)(main_th->unblock.arg);
                }
            }
        }
    }
}

#define CLOSE_INVALIDATE_PAIR(expr) \
    close_invalidate_pair(expr,"close_invalidate: "#expr)
static void
close_invalidate(int *fdp, const char *msg)
{
    int fd = *fdp;

    *fdp = -1;
    if (close(fd) < 0) {
        async_bug_fd(msg, errno, fd);
    }
}

static void
close_invalidate_pair(int fds[2], const char *msg)
{
    if (USE_EVENTFD && fds[0] == fds[1]) {
        fds[1] = -1; // disable write port first
        close_invalidate(&fds[0], msg);
    }
    else {
        close_invalidate(&fds[1], msg);
        close_invalidate(&fds[0], msg);
    }
}

static void
set_nonblock(int fd)
{
    int oflags;
    int err;

    oflags = fcntl(fd, F_GETFL);
    if (oflags == -1)
        rb_sys_fail(0);
    oflags |= O_NONBLOCK;
    err = fcntl(fd, F_SETFL, oflags);
    if (err == -1)
        rb_sys_fail(0);
}

/* communication pipe with timer thread and signal handler */
static void
setup_communication_pipe_internal(int pipes[2])
{
    int err;

    if (pipes[0] > 0 || pipes[1] > 0) {
        VM_ASSERT(pipes[0] > 0);
        VM_ASSERT(pipes[1] > 0);
        return;
    }

    /*
     * Don't bother with eventfd on ancient Linux 2.6.22..2.6.26 which were
     * missing EFD_* flags, they can fall back to pipe
     */
#if USE_EVENTFD && defined(EFD_NONBLOCK) && defined(EFD_CLOEXEC)
    pipes[0] = pipes[1] = eventfd(0, EFD_NONBLOCK|EFD_CLOEXEC);

    if (pipes[0] >= 0) {
        rb_update_max_fd(pipes[0]);
        return;
    }
#endif

    err = rb_cloexec_pipe(pipes);
    if (err != 0) {
        rb_bug("can not create communication pipe");
    }
    rb_update_max_fd(pipes[0]);
    rb_update_max_fd(pipes[1]);
    set_nonblock(pipes[0]);
    set_nonblock(pipes[1]);
}

#if !defined(SET_CURRENT_THREAD_NAME) && defined(__linux__) && defined(PR_SET_NAME)
# define SET_CURRENT_THREAD_NAME(name) prctl(PR_SET_NAME, name)
#endif

enum {
    THREAD_NAME_MAX =
#if defined(__linux__)
    16
#elif defined(__APPLE__)
/* Undocumented, and main thread seems unlimited */
    64
#else
    16
#endif
};

static VALUE threadptr_invoke_proc_location(rb_thread_t *th);

static void
native_set_thread_name(rb_thread_t *th)
{
#ifdef SET_CURRENT_THREAD_NAME
    VALUE loc;
    if (!NIL_P(loc = th->name)) {
        SET_CURRENT_THREAD_NAME(RSTRING_PTR(loc));
    }
    else if ((loc = threadptr_invoke_proc_location(th)) != Qnil) {
        char *name, *p;
        char buf[THREAD_NAME_MAX];
        size_t len;
        int n;

        name = RSTRING_PTR(RARRAY_AREF(loc, 0));
        p = strrchr(name, '/'); /* show only the basename of the path. */
        if (p && p[1])
            name = p + 1;

        n = snprintf(buf, sizeof(buf), "%s:%d", name, NUM2INT(RARRAY_AREF(loc, 1)));
        RB_GC_GUARD(loc);

        len = (size_t)n;
        if (len >= sizeof(buf)) {
            buf[sizeof(buf)-2] = '*';
            buf[sizeof(buf)-1] = '\0';
        }
        SET_CURRENT_THREAD_NAME(buf);
    }
#endif
}

static void
native_set_another_thread_name(rb_nativethread_id_t thread_id, VALUE name)
{
#if defined SET_ANOTHER_THREAD_NAME || defined SET_CURRENT_THREAD_NAME
    char buf[THREAD_NAME_MAX];
    const char *s = "";
# if !defined SET_ANOTHER_THREAD_NAME
    if (!pthread_equal(pthread_self(), thread_id)) return;
# endif
    if (!NIL_P(name)) {
        long n;
        RSTRING_GETMEM(name, s, n);
        if (n >= (int)sizeof(buf)) {
            memcpy(buf, s, sizeof(buf)-1);
            buf[sizeof(buf)-1] = '\0';
            s = buf;
        }
    }
# if defined SET_ANOTHER_THREAD_NAME
    SET_ANOTHER_THREAD_NAME(thread_id, s);
# elif defined SET_CURRENT_THREAD_NAME
    SET_CURRENT_THREAD_NAME(s);
# endif
#endif
}

#if defined(RB_THREAD_T_HAS_NATIVE_ID) || defined(__APPLE__)
static VALUE
native_thread_native_thread_id(rb_thread_t *target_th)
{
    if (!target_th->nt) return Qnil;

#ifdef RB_THREAD_T_HAS_NATIVE_ID
    int tid = target_th->nt->tid;
    if (tid == 0) return Qnil;
    return INT2FIX(tid);
#elif defined(__APPLE__)
    uint64_t tid;
/* The first condition is needed because MAC_OS_X_VERSION_10_6
    is not defined on 10.5, and while __POWERPC__ takes care of ppc/ppc64,
    i386 will be broken without this. Note, 10.5 is supported with GCC upstream,
    so it has C++17 and everything needed to build modern Ruby. */
# if (!defined(MAC_OS_X_VERSION_10_6) || \
      (MAC_OS_X_VERSION_MAX_ALLOWED < MAC_OS_X_VERSION_10_6) || \
      defined(__POWERPC__) /* never defined for PowerPC platforms */)
    const bool no_pthread_threadid_np = true;
#   define NO_PTHREAD_MACH_THREAD_NP 1
# elif MAC_OS_X_VERSION_MIN_REQUIRED >= MAC_OS_X_VERSION_10_6
    const bool no_pthread_threadid_np = false;
# else
#   if !(defined(__has_attribute) && __has_attribute(availability))
    /* __API_AVAILABLE macro does nothing on gcc */
    __attribute__((weak)) int pthread_threadid_np(pthread_t, uint64_t*);
#   endif
    /* Check weakly linked symbol */
    const bool no_pthread_threadid_np = !&pthread_threadid_np;
# endif
    if (no_pthread_threadid_np) {
        return ULL2NUM(pthread_mach_thread_np(pthread_self()));
    }
# ifndef NO_PTHREAD_MACH_THREAD_NP
    int e = pthread_threadid_np(target_th->nt->thread_id, &tid);
    if (e != 0) rb_syserr_fail(e, "pthread_threadid_np");
    return ULL2NUM((unsigned long long)tid);
# endif
#endif
}
# define USE_NATIVE_THREAD_NATIVE_THREAD_ID 1
#else
# define USE_NATIVE_THREAD_NATIVE_THREAD_ID 0
#endif

static struct {
    rb_serial_t created_fork_gen;
    pthread_t pthread_id;

    int comm_fds[2]; // r, w

#if (HAVE_SYS_EPOLL_H || HAVE_SYS_EVENT_H) && USE_MN_THREADS
    int event_fd; // kernel event queue fd (epoll/kqueue)
#endif
#if HAVE_SYS_EPOLL_H && USE_MN_THREADS
#define EPOLL_EVENTS_MAX 0x10
    struct epoll_event finished_events[EPOLL_EVENTS_MAX];
#elif HAVE_SYS_EVENT_H && USE_MN_THREADS
#define KQUEUE_EVENTS_MAX 0x10
    struct kevent finished_events[KQUEUE_EVENTS_MAX];
#endif

#if USE_MN_THREADS
    /* Timed waiters, bucketed by deadline into a hierarchical timer wheel;
     * untimed (fd-only) waiters keep a plain list.  Both under waiting_lock.
     * The wheel operations all live in thread_pthread_mn.c (timer_wheel_*). */
#define TIMER_WHEEL_LEVELS 4
#define TIMER_WHEEL_SLOT_BITS 6
#define TIMER_WHEEL_SLOTS (1 << TIMER_WHEEL_SLOT_BITS)
    struct timer_wheel_level {
        uint64_t occupied; // bit s: slots[s] is non-empty
        struct ccan_list_head slots[TIMER_WHEEL_SLOTS];
    } wheel[TIMER_WHEEL_LEVELS];
    uint64_t wheel_cursor_tick; // slots for ticks <= this are drained
    rb_hrtime_t next_expiry;   // never later than the earliest deadline
    pthread_mutex_t waiting_lock; // the wheel and the flags of fd-less timed waits

    /* The fd map entries and the flags of io waits are guarded per fd by a
     * shard lock, so unrelated fds register and wake in parallel.  Whoever
     * clears an io wait's flags under its shard owns the wakeup.  Lock order:
     * shard -> waiting_lock -> wake_pending_lock, never the other way. */
#define IO_WAIT_SHARDS 16
    pthread_mutex_t fd_shard_locks[IO_WAIT_SHARDS];

    // signaled when wake_pending clears on a thread; see timer_thread_wake_fence
    pthread_mutex_t wake_pending_lock;
    rb_nativethread_cond_t wake_pending_cond;
#endif

#if (HAVE_SYS_EPOLL_H || HAVE_SYS_EVENT_H) && USE_MN_THREADS
    // fd -> struct rb_fd_waiters, in chunks so entries never move.  The chunk
    // table installs slots by CAS (chunks span shards); entry state is guarded
    // by the fd's shard lock.
#define FDMAP_MAX_CHUNKS 1024 // fds up to FDMAP_MAX_CHUNKS * FDMAP_CHUNK_SIZE
    struct rb_fd_waiters *fdmap_chunks[FDMAP_MAX_CHUNKS];
#endif
} timer_th = {
    .created_fork_gen = 0,
};

#define TIMER_THREAD_CREATED_P() (timer_th.created_fork_gen == current_fork_gen)

static void timer_thread_check_timeslice(rb_vm_t *vm);
static bool timeslice_scan(rb_vm_t *vm, bool interrupt);
static int timer_thread_set_timeout(rb_vm_t *vm);

#include "thread_pthread_mn.c"

static int
timer_thread_set_timeout(rb_vm_t *vm)
{
#if 0
    return 10; // ms
#else
    int timeout = -1;

    ractor_sched_lock(vm, NULL);
    {
        if (   timeslice_scan(vm, false)                             // (1-1) Provide time slice for active NTs
            || !ubf_threads_empty()                                  // (1-3) Periodic UBF
            || vm->ractor.sched.grq_cnt > 0                          // (1-4) Lazy GRQ deq start
            ) {

            RUBY_DEBUG_LOG("ubf:%d grq:%d",
                           !ubf_threads_empty(),
                           (vm->ractor.sched.grq_cnt > 0));

            timeout = 10; // ms
            vm->ractor.sched.timeslice_wait_inf = false;
        }
        else {
            vm->ractor.sched.timeslice_wait_inf = true;
        }
    }
    ractor_sched_unlock(vm, NULL);

    timeout = timer_wheel_timeout(timeout);

    RUBY_DEBUG_LOG("timeout:%d inf:%d", timeout, (int)vm->ractor.sched.timeslice_wait_inf);

    // fprintf(stderr, "timeout:%d\n", timeout);
    return timeout;
#endif
}

static void
timer_thread_check_signal(rb_vm_t *vm)
{
    // ruby_sigchld_handler(vm); TODO

    int signum = rb_signal_buff_size();
    if (UNLIKELY(signum > 0) && vm->ractor.main_thread) {
        RUBY_DEBUG_LOG("signum:%d", signum);
        threadptr_trap_interrupt(vm->ractor.main_thread);
    }
}

// Tick (with `interrupt`) each listed sched's running thread and prune scheds
// whose readyq drained; returns whether any sched still needs ticks.
static bool
timeslice_scan(rb_vm_t *vm, bool interrupt)
{
    bool found = false;
    struct rb_thread_sched *sched, *next;

    rb_native_mutex_lock(&vm->ractor.sched.timeslice.lock);
    {
        ccan_list_for_each_safe(&vm->ractor.sched.timeslice.scheds, sched, next, timeslice_node) {
            // trylock: timeslice_sched_link nests sched.lock -> timeslice.lock,
            // this scan holds the locks the other way around
            if (rb_native_mutex_trylock(&sched->lock_) == 0) {
                if (ccan_list_empty(&sched->readyq)) {
                    ccan_list_del_init(&sched->timeslice_node); // a later enq relinks it
                }
                else if (sched->is_running) {
                    VM_ASSERT(sched->running != NULL);
                    found = true;
                    if (interrupt) {
                        RUBY_DEBUG_LOG("timeslice th:%u", rb_th_serial(sched->running));
                        RUBY_VM_SET_TIMER_INTERRUPT(sched->running->ec);
                    }
                }
                // else: waiters behind a blocked runner need no ticks; the
                // add path wakes the timer when the sched runs again
                rb_native_mutex_unlock(&sched->lock_);
            }
            else {
                found = true; // busy switching; tick it on the next round
            }
        }
    }
    rb_native_mutex_unlock(&vm->ractor.sched.timeslice.lock);

    return found;
}

static void
timer_thread_check_timeslice(rb_vm_t *vm)
{
    // TODO: check time
    timeslice_scan(vm, true);
}

void
rb_assert_sig(void)
{
    sigset_t oldmask;
    pthread_sigmask(0, NULL, &oldmask);
    if (sigismember(&oldmask, SIGVTALRM)) {
        rb_bug("!!!");
    }
    else {
        RUBY_DEBUG_LOG("ok");
    }
}

static void *
timer_thread_func(void *ptr)
{
    rb_vm_t *vm = (rb_vm_t *)ptr;
#if defined(RUBY_NT_SERIAL)
    ruby_nt_serial = (rb_atomic_t)-1;
#endif

    RUBY_DEBUG_LOG("started%s", "");

    while (RUBY_ATOMIC_LOAD(system_working)) {
        timer_thread_check_signal(vm);
        timer_thread_check_timeout(vm);
        ubf_wakeup_all_threads();

        RUBY_DEBUG_LOG("system_working:%d", RUBY_ATOMIC_LOAD(system_working));
        timer_thread_polling(vm);
    }

    RUBY_DEBUG_LOG("terminated");
    return NULL;
}

/* only use signal-safe system calls here */
static void
signal_communication_pipe(int fd)
{
#if USE_EVENTFD
    const uint64_t buff = 1;
#else
    const char buff = '!';
#endif
    ssize_t result;

    /* already opened */
    if (fd >= 0) {
      retry:
        if ((result = write(fd, &buff, sizeof(buff))) <= 0) {
            int e = errno;
            switch (e) {
              case EINTR: goto retry;
              case EAGAIN:
#if defined(EWOULDBLOCK) && EWOULDBLOCK != EAGAIN
              case EWOULDBLOCK:
#endif
                break;
              default:
                async_bug_fd("rb_thread_wakeup_timer_thread: write", e, fd);
            }
        }
        if (TT_DEBUG) WRITE_CONST(2, "rb_thread_wakeup_timer_thread: write\n");
    }
    else {
        // ignore wakeup
    }
}

static void
timer_thread_wakeup_force(void)
{
    // should not use RUBY_DEBUG_LOG() because it can be called within signal handlers.
    signal_communication_pipe(timer_th.comm_fds[1]);
}

static void
timer_thread_wakeup_locked(rb_vm_t *vm)
{
    // should be locked before.
    ASSERT_ractor_sched_locked(vm, NULL);

    if (timer_th.created_fork_gen == current_fork_gen) {
        if (vm->ractor.sched.timeslice_wait_inf) {
            RUBY_DEBUG_LOG("wakeup with fd:%d", timer_th.comm_fds[1]);
            timer_thread_wakeup_force();
        }
        else {
            RUBY_DEBUG_LOG("will be wakeup...");
        }
    }
}

static void
timer_thread_wakeup(void)
{
    rb_vm_t *vm = GET_VM();

    ractor_sched_lock(vm, NULL);
    {
        timer_thread_wakeup_locked(vm);
    }
    ractor_sched_unlock(vm, NULL);
}

static void
rb_thread_create_timer_thread(void)
{
    rb_serial_t created_fork_gen = timer_th.created_fork_gen;

    RUBY_DEBUG_LOG("fork_gen create:%d current:%d", (int)created_fork_gen, (int)current_fork_gen);

    timer_th.created_fork_gen = current_fork_gen;

    if (created_fork_gen != current_fork_gen) {
        if (created_fork_gen != 0) {
            RUBY_DEBUG_LOG("forked child process");

            CLOSE_INVALIDATE_PAIR(timer_th.comm_fds);
#if USE_MN_THREADS
            // The parent's waiters do not exist in the child, and the armings
            // belong to the closed event backend: a stale entry would satisfy
            // fd_waiters_arm's want == armed_flags check and never arm the new
            // backend (a lost wake the old dynamic map also had).
            for (unsigned int ci = 0; ci < FDMAP_MAX_CHUNKS; ci++) {
                struct rb_fd_waiters *chunk = timer_th.fdmap_chunks[ci];
                if (chunk == NULL) continue;
                for (unsigned int i = 0; i < FDMAP_CHUNK_SIZE; i++) {
                    ccan_list_head_init(&chunk[i].waiters);
                    chunk[i].armed_flags = 0;
                    chunk[i].generation++;
                }
            }
#endif
#if HAVE_SYS_EPOLL_H && USE_MN_THREADS
            close_invalidate(&timer_th.event_fd, "close event_fd");
#elif HAVE_SYS_EVENT_H && USE_MN_THREADS
            // A kqueue is not inherited across fork: the number names a closed
            // fd in the child, and closing it could hit a reused one.
            timer_th.event_fd = -1;
#endif
            // No mutex_destroy for waiting_lock: glibc returns EBUSY (and
            // rb_native_mutex_destroy rb_bugs) for a mutex another ractor's
            // M:N thread held at the fork moment.  The initialize below
            // starts over.
        }

#if USE_MN_THREADS
        for (int lvl = 0; lvl < TIMER_WHEEL_LEVELS; lvl++) {
            timer_th.wheel[lvl].occupied = 0;
            for (int slot = 0; slot < TIMER_WHEEL_SLOTS; slot++) {
                ccan_list_head_init(&timer_th.wheel[lvl].slots[slot]);
            }
        }
        timer_th.wheel_cursor_tick = timer_wheel_tick(rb_hrtime_now());
        timer_th.next_expiry = TIMER_WHEEL_NO_EXPIRY;
        rb_native_mutex_initialize(&timer_th.waiting_lock);
        for (int i = 0; i < IO_WAIT_SHARDS; i++) {
            rb_native_mutex_initialize(&timer_th.fd_shard_locks[i]);
        }
        rb_native_mutex_initialize(&timer_th.wake_pending_lock);
        rb_native_cond_initialize(&timer_th.wake_pending_cond);
#endif

        // open communication channel
        setup_communication_pipe_internal(timer_th.comm_fds);

        // open event fd
        timer_thread_setup_mn();
    }

    int err = pthread_create(&timer_th.pthread_id, NULL, timer_thread_func, GET_VM());
    if (err != 0) {
        // The timer thread delivers signals and drives the M:N scheduler;
        // running without one only defers the failure to stranger places.
        rb_bug_errno("pthread_create (timer thread)", err);
    }
}

static int
native_stop_timer_thread(void)
{
    RUBY_ATOMIC_SET(system_working, 0);

    RUBY_DEBUG_LOG("wakeup send %d", timer_th.comm_fds[1]);
    timer_thread_wakeup_force();
    RUBY_DEBUG_LOG("wakeup sent");
    pthread_join(timer_th.pthread_id, NULL);

    if (TT_DEBUG) fprintf(stderr, "stop timer thread\n");

    return 1;
}

static void
native_reset_timer_thread(void)
{
    //
}

#ifdef HAVE_SIGALTSTACK
int
ruby_stack_overflowed_p(const rb_thread_t *th, const void *addr)
{
    void *base;
    size_t size;
    const size_t water_mark = RB_THREAD_PAGE_SIZE;
    STACK_GROW_DIR_DETECTION;

    if (th) {
        size = th->ec->machine.stack_maxsize;
        base = (char *)th->ec->machine.stack_start - STACK_DIR_UPPER(0, size);
    }
#ifdef STACKADDR_AVAILABLE
    else if (get_stack(&base, &size) == 0) {
# ifdef __APPLE__
        // th is NULL in this branch; ask about the calling thread itself.
        if (pthread_equal(pthread_self(), native_main_thread.id)) {
            struct rlimit rlim;
            if (getrlimit(RLIMIT_STACK, &rlim) == 0 && rlim.rlim_cur > size) {
                size = (size_t)rlim.rlim_cur;
            }
        }
# endif
        base = (char *)base + STACK_DIR_UPPER(+size, -size);
    }
#endif
    else {
        return 0;
    }

    if (size > water_mark) size = water_mark;
    if (IS_STACK_DIR_UPPER()) {
        if (size > ~(size_t)base+1) size = ~(size_t)base+1;
        if (addr > base && addr <= (void *)((char *)base + size)) return 1;
    }
    else {
        if (size > (size_t)base) size = (size_t)base;
        if (addr > (void *)((char *)base - size) && addr <= base) return 1;
    }
    return 0;
}
#endif

int
rb_reserved_fd_p(int fd)
{
    /* no false-positive if out-of-FD at startup */
    if (fd < 0) return 0;

    if (fd == timer_th.comm_fds[0] ||
        fd == timer_th.comm_fds[1]
#if (HAVE_SYS_EPOLL_H || HAVE_SYS_EVENT_H) && USE_MN_THREADS
        || fd == timer_th.event_fd
#endif
        ) {
        goto check_fork_gen;
    }
    return 0;

  check_fork_gen:
    if (timer_th.created_fork_gen == current_fork_gen) {
        /* async-signal-safe */
        return 1;
    }
    else {
        return 0;
    }
}

rb_nativethread_id_t
rb_nativethread_self(void)
{
    return pthread_self();
}

#if defined(USE_POLL) && !defined(HAVE_PPOLL)
/* TODO: don't ignore sigmask */
static int
ruby_ppoll(struct pollfd *fds, nfds_t nfds,
      const struct timespec *ts, const sigset_t *sigmask)
{
    int timeout_ms;

    if (ts) {
        int tmp, tmp2;

        if (ts->tv_sec > INT_MAX/1000)
            timeout_ms = INT_MAX;
        else {
            tmp = (int)(ts->tv_sec * 1000);
            /* round up 1ns to 1ms to avoid excessive wakeups for <1ms sleep */
            tmp2 = (int)((ts->tv_nsec + 999999L) / (1000L * 1000L));
            if (INT_MAX - tmp < tmp2)
                timeout_ms = INT_MAX;
            else
                timeout_ms = (int)(tmp + tmp2);
        }
    }
    else
        timeout_ms = -1;

    return poll(fds, nfds, timeout_ms);
}
#  define ppoll(fds,nfds,ts,sigmask) ruby_ppoll((fds),(nfds),(ts),(sigmask))
#endif

static void
native_sleep(rb_thread_t *th, rb_hrtime_t *rel)
{
    struct rb_thread_sched *sched = TH_SCHED(th);

    RUBY_DEBUG_LOG("rel:%d", rel ? (int)*rel : 0);

    if (rel && !th_has_dedicated_nt(th)) {
        // an M:N thread has no condvar of its own: the timer thread wakes it
        thread_sched_wait_events(sched, th, -1, thread_sched_waiting_timeout, rel);
    }
    else if (rel) {
        /* Solaris cond_timedwait() returns EINVAL if an argument is greater than
         * current_time + 100,000,000.  So cut up to 100,000,000.  This is
         * considered as a kind of spurious wakeup.  The caller to native_sleep
         * should care about spurious wakeup.
         *
         * See also [Bug #1341] [ruby-core:29702]
         * http://download.oracle.com/docs/cd/E19683-01/816-0216/6m6ngupgv/index.html
         */
        const rb_hrtime_t max = (rb_hrtime_t)100000000 * RB_HRTIME_PER_SEC;
        if (*rel > max) *rel = max;

        rb_hrtime_t end = rb_hrtime_add(rb_hrtime_now(), *rel);
        thread_sched_to_waiting_until_wakeup(sched, th, &end);
    }
    else {
        thread_sched_to_waiting_until_wakeup(sched, th, NULL);
    }

    RUBY_DEBUG_LOG("wakeup");
}

// fork read-write lock (only for pthread)
static pthread_rwlock_t rb_thread_fork_rw_lock = PTHREAD_RWLOCK_INITIALIZER;

void
rb_thread_release_fork_lock(void)
{
    int r;
    if ((r = pthread_rwlock_unlock(&rb_thread_fork_rw_lock))) {
        rb_bug_errno("pthread_rwlock_unlock", r);
    }
}

void
rb_thread_reset_fork_lock(void)
{
    int r;
    if ((r = pthread_rwlock_destroy(&rb_thread_fork_rw_lock))) {
        rb_bug_errno("pthread_rwlock_destroy", r);
    }

    if ((r = pthread_rwlock_init(&rb_thread_fork_rw_lock, NULL))) {
        rb_bug_errno("pthread_rwlock_init", r);
    }
}

void *
rb_thread_prevent_fork(void *(*func)(void *), void *data)
{
    int r;
    if ((r = pthread_rwlock_rdlock(&rb_thread_fork_rw_lock))) {
        rb_bug_errno("pthread_rwlock_rdlock", r);
    }
    void *result = func(data);
    rb_thread_release_fork_lock();
    return result;
}

void
rb_thread_acquire_fork_lock(void)
{
    int r;
    if ((r = pthread_rwlock_wrlock(&rb_thread_fork_rw_lock))) {
        rb_bug_errno("pthread_rwlock_wrlock", r);
    }
}

// thread internal event hooks (only for pthread)

struct rb_internal_thread_event_hook {
    rb_internal_thread_event_callback callback;
    rb_event_flag_t event;
    void *user_data;

    struct rb_internal_thread_event_hook *next;
};

static pthread_rwlock_t rb_internal_thread_event_hooks_rw_lock = PTHREAD_RWLOCK_INITIALIZER;

/* For the GC: whether any thread-event hook is registered right now.  The rwlock
 * makes the answer happen-after any completed registration; a hook registered
 * after this read gets no event from the asking thread (rb_thread_execute_hooks
 * skips a swept thread), so it never sees the objects the caller may collect. */
bool
rb_thread_event_hooks_registered_p(void)
{
    int r;
    if ((r = pthread_rwlock_rdlock(&rb_internal_thread_event_hooks_rw_lock))) {
        rb_bug_errno("pthread_rwlock_rdlock", r);
    }
    const bool registered = (rb_internal_thread_event_hooks != NULL);
    if ((r = pthread_rwlock_unlock(&rb_internal_thread_event_hooks_rw_lock))) {
        rb_bug_errno("pthread_rwlock_unlock", r);
    }
    return registered;
}

#if defined(HAVE_WORKING_FORK)
static void
rb_internal_thread_event_hooks_rw_lock_atfork(void)
{
  // After fork(), this rwlock may have been held by a now-dead thread.
  //
  // pthread_rwlock_destroy() on a held lock is undefined behavior, and
  // pthread_rwlock_init() on an already-initialized lock is also undefined
  // behavior
  //
  // Direct assignment of PTHREAD_RWLOCK_INITIALIZER is safe and portable.
  rb_internal_thread_event_hooks_rw_lock =
      (pthread_rwlock_t)PTHREAD_RWLOCK_INITIALIZER;
}
#endif

rb_internal_thread_event_hook_t *
rb_internal_thread_add_event_hook(rb_internal_thread_event_callback callback, rb_event_flag_t internal_event, void *user_data)
{
    rb_internal_thread_event_hook_t *hook = ALLOC_N(rb_internal_thread_event_hook_t, 1);
    hook->callback = callback;
    hook->user_data = user_data;
    hook->event = internal_event;

    int r;
    if ((r = pthread_rwlock_wrlock(&rb_internal_thread_event_hooks_rw_lock))) {
        rb_bug_errno("pthread_rwlock_wrlock", r);
    }

    hook->next = rb_internal_thread_event_hooks;
    ATOMIC_PTR_EXCHANGE(rb_internal_thread_event_hooks, hook);

    if ((r = pthread_rwlock_unlock(&rb_internal_thread_event_hooks_rw_lock))) {
        rb_bug_errno("pthread_rwlock_unlock", r);
    }
    return hook;
}

bool
rb_internal_thread_remove_event_hook(rb_internal_thread_event_hook_t * hook)
{
    int r;
    if ((r = pthread_rwlock_wrlock(&rb_internal_thread_event_hooks_rw_lock))) {
        rb_bug_errno("pthread_rwlock_wrlock", r);
    }

    bool success = FALSE;

    if (rb_internal_thread_event_hooks == hook) {
        ATOMIC_PTR_EXCHANGE(rb_internal_thread_event_hooks, hook->next);
        success = TRUE;
    }
    else {
        rb_internal_thread_event_hook_t *h = rb_internal_thread_event_hooks;

        do {
            if (h->next == hook) {
                h->next = hook->next;
                success = TRUE;
                break;
            }
        } while ((h = h->next));
    }

    if ((r = pthread_rwlock_unlock(&rb_internal_thread_event_hooks_rw_lock))) {
        rb_bug_errno("pthread_rwlock_unlock", r);
    }

    if (success) {
        SIZED_FREE(hook);
    }
    return success;
}

static void
rb_thread_execute_hooks(rb_event_flag_t event, rb_thread_t *th)
{
    int r;

    /* th->self == 0: the dying thread's final collection swept its Thread wrapper, so
     * no hook existed then; one registered since has no ordering claim to this event. */
    if (th->self == 0) return;
    if ((r = pthread_rwlock_rdlock(&rb_internal_thread_event_hooks_rw_lock))) {
        rb_bug_errno("pthread_rwlock_rdlock", r);
    }

    if (rb_internal_thread_event_hooks) {
        rb_internal_thread_event_hook_t *h = rb_internal_thread_event_hooks;
        do {
            if (h->event & event) {
                rb_internal_thread_event_data_t event_data = {
                    .thread = th->self,
                };
                (*h->callback)(event, &event_data, h->user_data);
            }
        } while((h = h->next));
    }
    if ((r = pthread_rwlock_unlock(&rb_internal_thread_event_hooks_rw_lock))) {
        rb_bug_errno("pthread_rwlock_unlock", r);
    }
}

// return true if the current thread acquires DNT.
// return false if the current thread already acquires DNT.
bool
rb_thread_lock_native_thread(void)
{
    rb_thread_t *th = GET_THREAD();
    bool is_snt = th->nt->dedicated == 0;
    native_thread_dedicated_inc(th->vm, th->ractor, th->nt);

    return is_snt;
}

void
rb_thread_malloc_stack_set(rb_thread_t *th, void *stack, size_t stack_size)
{
    th->sched.malloc_stack = true;
    th->sched.context_stack = stack;
    th->sched.context_stack_size = stack_size;
}

#endif /* THREAD_SYSTEM_DEPENDENT_IMPLEMENTATION */