farce 0.3.0-x86-linux-gnu → 0.4.1-x86-linux-gnu

data/ext/farce/map.c

@@ -1,16 +1,14 @@
 #include "containers.h"
+#include "native_wait.h"
 #include "transaction.h"
 #include "ruby/fiber/scheduler.h"
-#include "ruby/io.h"
 
 #include <errno.h>
-#include <fcntl.h>
 #include <math.h>
 #include <stdint.h>
 #include <stdatomic.h>
 #include <stdlib.h>
 #include <time.h>
-#include <unistd.h>
 
 #define MAP_INITIAL_CAPACITY 16
 
@@ -42,9 +40,7 @@ typedef struct {
 } map_slot_t;
 
 typedef struct map_waiter {
-    int read_fd;
-    int write_fd;
-    bool notified;
+    containers_waiter_t notification;
     struct map_waiter *next;
 } map_waiter_t;
 
@@ -141,46 +137,12 @@ map_parse_timeout(VALUE timeout)
     return parsed;
 }
 
-static void
-map_set_fd_flags(int fd)
-{
-#ifdef _WIN32
-    (void)fd;
-#else
-    int flags = fcntl(fd, F_GETFL, 0);
-    if (flags >= 0) (void)fcntl(fd, F_SETFL, flags | O_NONBLOCK);
-    flags = fcntl(fd, F_GETFD, 0);
-    if (flags >= 0) (void)fcntl(fd, F_SETFD, flags | FD_CLOEXEC);
-#endif
-}
-
-static void
-map_waiter_initialize(map_waiter_t *waiter)
-{
-    int descriptors[2];
-    if (pipe(descriptors) != 0) rb_sys_fail("pipe");
-    waiter->read_fd = descriptors[0];
-    waiter->write_fd = descriptors[1];
-    waiter->notified = false;
-    waiter->next = NULL;
-    map_set_fd_flags(waiter->read_fd);
-    map_set_fd_flags(waiter->write_fd);
-}
-
 static void
 map_notify_waiters_locked(map_t *map)
 {
-    unsigned char byte = 1;
     map->generation++;
     for (map_waiter_t *waiter = map->waiters; waiter; waiter = waiter->next) {
-        if (waiter->notified) continue;
-        ssize_t result;
-        do {
-            result = write(waiter->write_fd, &byte, 1);
-        } while (result < 0 && errno == EINTR);
-        if (result == 1 || (result < 0 && (errno == EAGAIN || errno == EWOULDBLOCK))) {
-            waiter->notified = true;
-        }
+        (void)containers_waiter_notify(&waiter->notification);
     }
 }
 
@@ -260,7 +222,7 @@ map_memsize(const void *pointer)
 }
 
 static const rb_data_type_t map_type = {
-    .wrap_struct_name = "Ractor::Containers::Map",
+    .wrap_struct_name = "Farce::Internal::Map",
     .function = {
         .dmark = map_mark,
         .dfree = map_free,
@@ -430,13 +392,10 @@ static VALUE
 map_wait_body(VALUE opaque)
 {
     map_wait_context_t *context = (map_wait_context_t *)opaque;
-    VALUE wait_timeout = Qnil;
-    if (context->timeout->finite) {
-        double remaining = context->timeout->deadline - map_monotonic_now();
-        if (remaining <= 0) return Qfalse;
-        wait_timeout = DBL2NUM(remaining);
-    }
-    return containers_wait_for_readable(context->waiter->read_fd, wait_timeout) ? Qtrue : Qfalse;
+    return containers_waiter_wait(
+        &context->waiter->notification, &context->map->lock,
+        context->timeout->finite, context->timeout->deadline, NULL
+    ) ? Qtrue : Qfalse;
 }
 
 static VALUE
@@ -451,8 +410,7 @@ map_wait_cleanup(VALUE opaque)
         pthread_mutex_unlock(&context->map->lock);
         context->registered = false;
     }
-    close(context->waiter->read_fd);
-    close(context->waiter->write_fd);
+    containers_waiter_close(&context->waiter->notification);
     return Qnil;
 }
 
@@ -460,13 +418,11 @@ static bool
 map_wait_once(map_t *map, uint64_t generation, map_timeout_t *timeout)
 {
     map_waiter_t waiter;
-    map_waiter_initialize(&waiter);
+    containers_waiter_init(&waiter.notification);
 
     pthread_mutex_lock(&map->lock);
     if (map->generation != generation) {
         pthread_mutex_unlock(&map->lock);
-        close(waiter.read_fd);
-        close(waiter.write_fd);
         return true;
     }
     waiter.next = map->waiters;

data/ext/farce/native_wait.c

@@ -0,0 +1,139 @@
+#include "native_wait.h"
+#include "ruby/fiber/scheduler.h"
+
+#include <errno.h>
+#include <fcntl.h>
+#include <time.h>
+#include <unistd.h>
+
+static double
+waiter_monotonic_now(void)
+{
+    struct timespec now;
+#ifdef CLOCK_MONOTONIC
+    if (clock_gettime(CLOCK_MONOTONIC, &now) == 0) {
+        return (double)now.tv_sec + (double)now.tv_nsec / 1000000000.0;
+    }
+#endif
+    struct timeval fallback;
+    gettimeofday(&fallback, NULL);
+    return (double)fallback.tv_sec + (double)fallback.tv_usec / 1000000.0;
+}
+
+static void
+waiter_set_fd_flags(int fd)
+{
+#ifndef _WIN32
+    int flags = fcntl(fd, F_GETFL, 0);
+    if (flags >= 0) (void)fcntl(fd, F_SETFL, flags | O_NONBLOCK);
+    flags = fcntl(fd, F_GETFD, 0);
+    if (flags >= 0) (void)fcntl(fd, F_SETFD, flags | FD_CLOEXEC);
+#endif
+}
+
+void
+containers_waiter_init(containers_waiter_t *waiter)
+{
+    containers_futex_init(&waiter->futex);
+    waiter->read_fd = waiter->write_fd = -1;
+    waiter->mode = CONTAINERS_WAITER_PENDING;
+    waiter->notified = false;
+}
+
+bool
+containers_waiter_notify(containers_waiter_t *waiter)
+{
+    if (waiter->notified) return true;
+    if (waiter->mode == CONTAINERS_WAITER_FUTEX) {
+        waiter->notified = true;
+        containers_futex_notify(&waiter->futex);
+        return true;
+    }
+
+    /* A pending waiter has not selected its backend yet. Latch its wake so
+     * setup can return immediately or a future kernel wait sees a mismatch. */
+    atomic_store_explicit(&waiter->futex.value, 1, memory_order_release);
+    if (waiter->mode == CONTAINERS_WAITER_PENDING) {
+        waiter->notified = true;
+        return true;
+    }
+
+    unsigned char byte = 1;
+    ssize_t result;
+    do {
+        result = write(waiter->write_fd, &byte, 1);
+    } while (result < 0 && errno == EINTR);
+    if (result == 1 || (result < 0 && (errno == EAGAIN || errno == EWOULDBLOCK))) {
+        waiter->notified = true;
+    }
+    return waiter->notified;
+}
+
+bool
+containers_waiter_wait(
+    containers_waiter_t *waiter,
+    pthread_mutex_t *guard,
+    bool finite,
+    double deadline,
+    int *pipe_error
+)
+{
+    if (pipe_error) *pipe_error = 0;
+    if (atomic_load_explicit(&waiter->futex.value, memory_order_acquire) == 1) return true;
+    if (finite && deadline <= waiter_monotonic_now()) return false;
+
+    /* Selection can enter Ruby or probe the kernel. The registration stays
+     * pending, with no native container mutex held, throughout this work. */
+    VALUE scheduler = rb_fiber_scheduler_current();
+    /* Blocking futex waits failed the cross-core thread handoff performance
+     * gate. Keep Ruby's managed IO path for threads and use futexes only when
+     * the owning event loop can complete them asynchronously. */
+    bool use_futex = !NIL_P(scheduler) && containers_futex_scheduler_supported(scheduler);
+
+    int descriptors[2] = {-1, -1};
+    if (!use_futex) {
+        if (pipe(descriptors) != 0) {
+            int error = errno;
+            if (atomic_load_explicit(&waiter->futex.value, memory_order_acquire) == 1) return true;
+            if (pipe_error) {
+                *pipe_error = error;
+                return false;
+            }
+            rb_syserr_fail(error, "pipe");
+        }
+        waiter_set_fd_flags(descriptors[0]);
+        waiter_set_fd_flags(descriptors[1]);
+    }
+
+    /* No Ruby calls occur between creating descriptors and recording ownership.
+     * A notifier either latches a pending wake or observes the complete backend. */
+    if (guard) pthread_mutex_lock(guard);
+    waiter->read_fd = descriptors[0];
+    waiter->write_fd = descriptors[1];
+    waiter->mode = use_futex ? CONTAINERS_WAITER_FUTEX : CONTAINERS_WAITER_IO;
+    bool notified = waiter->notified;
+    if (guard) pthread_mutex_unlock(guard);
+    if (notified) return true;
+
+    VALUE timeout = Qnil;
+    if (finite) {
+        double remaining = deadline - waiter_monotonic_now();
+        if (remaining <= 0) {
+            return atomic_load_explicit(&waiter->futex.value, memory_order_acquire) == 1;
+        }
+        timeout = DBL2NUM(remaining);
+    }
+
+    if (use_futex) {
+        return containers_futex_scheduler_wait(scheduler, &waiter->futex, timeout);
+    }
+    return containers_wait_for_readable(waiter->read_fd, timeout);
+}
+
+void
+containers_waiter_close(containers_waiter_t *waiter)
+{
+    if (waiter->read_fd >= 0) close(waiter->read_fd);
+    if (waiter->write_fd >= 0) close(waiter->write_fd);
+    waiter->read_fd = waiter->write_fd = -1;
+}

data/ext/farce/native_wait.h

@@ -0,0 +1,37 @@
+#ifndef FARCE_NATIVE_WAIT_H
+#define FARCE_NATIVE_WAIT_H
+
+#include "containers.h"
+#include "futex.h"
+
+typedef enum {
+    CONTAINERS_WAITER_PENDING,
+    CONTAINERS_WAITER_FUTEX,
+    CONTAINERS_WAITER_IO
+} containers_waiter_mode_t;
+
+typedef struct {
+    containers_futex_t futex;
+    int read_fd, write_fd;
+    containers_waiter_mode_t mode;
+    bool notified;
+} containers_waiter_t;
+
+/* Initialize before linking under the container's guard. Notification is
+ * native-only and uses that same guard. A NULL guard denotes owner-local state. */
+void containers_waiter_init(containers_waiter_t *waiter);
+bool containers_waiter_notify(containers_waiter_t *waiter);
+/* Run outside the guard and inside the caller's cleanup ensure. The absolute
+ * monotonic deadline includes backend setup. Optional pipe_error lets callers
+ * retain their own descriptor-pressure recovery; other errors raise normally. */
+bool containers_waiter_wait(
+    containers_waiter_t *waiter,
+    pthread_mutex_t *guard,
+    bool finite,
+    double deadline,
+    int *pipe_error
+);
+/* Unlink under the guard first. Waiting has already drained native requests. */
+void containers_waiter_close(containers_waiter_t *waiter);
+
+#endif

data/ext/farce/priority_queue.c

@@ -1,4 +1,5 @@
 #include "containers.h"
+#include "native_wait.h"
 #include "unshared_wait.h"
 #include <ruby/atomic.h>
 #include <ruby/fiber/scheduler.h>
@@ -348,9 +349,7 @@ priority_queue_bucket_unlink(
 }
 
 typedef struct priority_queue_lock_waiter {
-    int read_fd;
-    int write_fd;
-    bool notified;
+    containers_waiter_t notification;
     struct priority_queue_lock_waiter *next;
 } priority_queue_lock_waiter_t;
 
@@ -753,19 +752,6 @@ priority_queue_get(VALUE self)
     return queue;
 }
 
-static void
-priority_queue_lock_set_fd_flags(int fd)
-{
-#ifdef _WIN32
-    (void)fd;
-#else
-    int flags = fcntl(fd, F_GETFL, 0);
-    if (flags >= 0) (void)fcntl(fd, F_SETFL, flags | O_NONBLOCK);
-    flags = fcntl(fd, F_GETFD, 0);
-    if (flags >= 0) (void)fcntl(fd, F_SETFD, flags | FD_CLOEXEC);
-#endif
-}
-
 /* Called with queue->lock held. A logical unlock only needs to make one
  * waiter runnable. The selected waiter remains linked until ensure cleanup. */
 static void
@@ -777,20 +763,9 @@ priority_queue_lock_notify_one_locked(priority_queue_t *queue)
          * next waiter. Reentrant operations can register for a later unlock. */
         farce_unshared_wait_notify_all(&queue->unshared_lock_waiters);
     }
-    unsigned char byte = 1;
-    priority_queue_lock_waiter_t *waiter;
-
-    for (waiter = queue->waiters; waiter; waiter = waiter->next) {
-        ssize_t result;
-        if (waiter->notified) continue;
-        do {
-            result = write(waiter->write_fd, &byte, 1);
-        } while (result < 0 && errno == EINTR);
-        if (result == 1 ||
-            (result < 0 && (errno == EAGAIN || errno == EWOULDBLOCK))) {
-            waiter->notified = true;
-            return;
-        }
+    for (priority_queue_lock_waiter_t *waiter = queue->waiters; waiter; waiter = waiter->next) {
+        if (waiter->notification.notified) continue;
+        if (containers_waiter_notify(&waiter->notification)) return;
     }
 }
 
@@ -799,18 +774,16 @@ typedef struct {
     priority_queue_lock_waiter_t waiter;
 } priority_queue_lock_wait_context_t;
 
-static bool
-priority_queue_lock_wait_for_descriptor(int fd)
-{
-    return containers_wait_for_readable(fd, Qnil);
-}
-
 static VALUE
 priority_queue_lock_wait_body(VALUE opaque)
 {
     priority_queue_lock_wait_context_t *context =
         (priority_queue_lock_wait_context_t *)opaque;
-    bool awakened = priority_queue_lock_wait_for_descriptor(context->waiter.read_fd);
+    bool awakened = containers_waiter_wait(
+        &context->waiter.notification,
+        context->queue->shared ? &context->queue->lock : NULL,
+        false, 0, NULL
+    );
     return awakened ? Qtrue : Qfalse;
 }
 
@@ -835,10 +808,10 @@ priority_queue_lock_wait_cleanup(VALUE opaque)
     }
     /* A notification only makes this waiter runnable; it does not reserve
      * ownership. Pass the baton whenever a notified waiter leaves while the
-     * queue is still ownerless. This covers both interruption inside rb_io_wait
+     * queue is still ownerless. This covers both interruption while waiting
      * and cancellation after a normal wake but before the acquisition loop can
      * claim ownership, so later waiters cannot be stranded. */
-    if (context->waiter.notified &&
+    if (context->waiter.notification.notified &&
         NIL_P(context->queue->owner_fiber) &&
         !priority_queue_publication_busy(
             RUBY_ATOMIC_LOAD(context->queue->publication_state)
@@ -846,8 +819,7 @@ priority_queue_lock_wait_cleanup(VALUE opaque)
         priority_queue_lock_notify_one_locked(context->queue);
     }
     priority_queue_mutex_unlock(context->queue);
-    close(context->waiter.read_fd);
-    close(context->waiter.write_fd);
+    containers_waiter_close(&context->waiter.notification);
     return Qnil;
 }
 
@@ -859,25 +831,11 @@ priority_queue_lock_wait_once(priority_queue_t *queue)
         (void)farce_unshared_wait(&queue->unshared_lock_waiters, queue->unshared_self, false, 0);
         return;
     }
-    int descriptors[2];
-    priority_queue_lock_wait_context_t context;
-
-    if (pipe(descriptors) != 0) {
-        priority_queue_mutex_unlock(queue);
-        rb_sys_fail("pipe");
-    }
-    priority_queue_lock_set_fd_flags(descriptors[0]);
-    priority_queue_lock_set_fd_flags(descriptors[1]);
-
-    context = (priority_queue_lock_wait_context_t){
+    priority_queue_lock_wait_context_t context = {
         .queue = queue,
-        .waiter = {
-            .read_fd = descriptors[0],
-            .write_fd = descriptors[1],
-            .notified = false,
-            .next = NULL,
-        },
+        .waiter = {.next = NULL},
     };
+    containers_waiter_init(&context.waiter.notification);
     if (queue->last_waiter) queue->last_waiter->next = &context.waiter;
     else queue->waiters = &context.waiter;
     queue->last_waiter = &context.waiter;

data/ext/farce/queue.c

@@ -224,7 +224,7 @@ queue_memsize(const void *pointer)
 }
 
 static const rb_data_type_t queue_type = {
-    .wrap_struct_name = "Ractor::Containers::Queue",
+    .wrap_struct_name = "Farce::Internal::Queue",
     .function = {
         .dmark = queue_mark,
         .dfree = queue_free,

data/ext/farce/reactor.c

@@ -345,6 +345,9 @@ void fs_watch_io(fs_state *s, fs_wait *w, int fd, int events, int group, long in
     }
 }
 
+// Cancel the outstanding futex operation before its caller releases the word.
+static void fs_cancel_futex(VALUE self, fs_wait *w);
+
 // Finish a wait and remove its watches and ready-queue entry.
 // A pending receive must stop using the buffer before cleanup can continue.
 // Clear Ruby references and keep the fiber's slot available for its next wait.
@@ -352,6 +355,12 @@ VALUE fs_retire(VALUE self, VALUE token) {
     fs_state *s = fs_get(self); fs_wait *w = fs_lookup(s, token);
     if (!w) return Qnil;
     if (w->receive_pending) fs_cancel_receive(self, w);
+    if (w->futex_pending) {
+        // Logical cancellation must not discard the lookup required by the CQE.
+        // Suppress resumption first, then consume the original completion.
+        w->status = 2;
+        fs_cancel_futex(self, w);
+    }
     ready_remove(s, w);
     fs_watch *watch = w->watches;
     while (watch) {
@@ -381,6 +390,56 @@ VALUE fs_retire(VALUE self, VALUE token) {
     return Qnil;
 }
 
+// Select asynchronous futex parking only for an installed, live native ring.
+// External schedulers require a later integration instead of a duck-typed hook.
+bool containers_futex_scheduler_supported(VALUE scheduler) {
+#ifdef FARCE_HAVE_URING_FUTEX
+    if (NIL_P(scheduler) || !rb_typeddata_is_kind_of(scheduler, &fs_type)) return false;
+    fs_state *s = fs_get(scheduler);
+    return s->driver == FS_URING_DRIVER && s->ring_live && s->futex_supported &&
+        !s->closed && !s->stopping && s->owner == rb_thread_current() &&
+        s->active == rb_fiber_current() && rb_fiber_scheduler_get() == scheduler &&
+        containers_futex_available();
+#else
+    return false;
+#endif
+}
+
+typedef struct {
+    VALUE self, token, timeout;
+    containers_futex_t *word;
+} fs_futex_wait;
+
+// Submit and park using one suspension. The surrounding ensure retains the word
+// until the kernel request has completed, even if timer setup or parking raises.
+static VALUE futex_wait_body(VALUE data) {
+    fs_futex_wait *wait = (fs_futex_wait *)data;
+    fs_state *s = fs_get(wait->self);
+    fs_wait *w = fs_lookup(s, wait->token);
+    int error = fs_driver_futex(s, w, wait->word);
+    if (error) rb_syserr_fail(error, "scheduler futex submission");
+    return rb_funcall(wait->self, rb_intern("wait_with_timeout"), 3,
+        wait->token, wait->timeout, Qfalse);
+}
+
+static VALUE futex_wait_cleanup(VALUE data) {
+    fs_futex_wait *wait = (fs_futex_wait *)data;
+    return fs_retire(wait->self, wait->token);
+}
+
+bool containers_futex_scheduler_wait(VALUE scheduler, containers_futex_t *word, VALUE timeout) {
+    if (!containers_futex_scheduler_supported(scheduler)) return true;
+    fs_futex_wait wait = {scheduler, fs_arm(scheduler), timeout, word};
+    return RTEST(rb_ensure(futex_wait_body, (VALUE)&wait, futex_wait_cleanup, (VALUE)&wait));
+}
+
+// Expose the backend capability privately so native-path validation cannot pass
+// by silently exercising a descriptor fallback.
+static VALUE futex_supported(VALUE self) {
+    fs_state *s = fs_get(self);
+    return !s->closed && !s->stopping && s->futex_supported && containers_futex_available() ? Qtrue : Qfalse;
+}
+
 // Switch from the root to a ready fiber.
 // If a newly started child finishes, return control to its waiting parent
 // before returning to the root dispatcher.
@@ -600,6 +659,27 @@ static int snapshot_fallback(st_data_t key, st_data_t value, st_data_t arg) {
 // Wake-pipe events ask the dispatcher to return to Ruby policy code.
 static void accept_event(VALUE self, fs_event *e) {
     fs_state *s = fs_get(self);
+    if (e->generation & FS_FUTEX) {
+        fs_wait *w = fs_lookup(s, ULL2NUM(e->generation & ~FS_OPERATION));
+        if (!w) return;
+        w->futex_pending = 0;
+        int error = e->error < 0 ? -e->error : 0;
+        if (error == ENOSYS || error == EOPNOTSUPP) s->futex_supported = 0;
+        if (!w->status) {
+            if (error == ENOSYS || error == EOPNOTSUPP) {
+                // Retry with the now-disabled capability, rather than reporting
+                // a timeout to consumers that have not reached their deadline.
+                fs_resume(self, w->token, Qtrue, Qnil);
+            } else {
+                // Early notification is reported as EAGAIN. Other interruptions
+                // also return to the container's condition/deadline check.
+                if (error != EAGAIN && error != EINTR && error != ECANCELED)
+                    w->native_error = error;
+                fs_resume(self, w->token, Qtrue, Qnil);
+            }
+        }
+        return;
+    }
     if (e->generation & FS_RECEIVE) {
         fs_wait *w = fs_lookup(s, ULL2NUM(e->generation & ~FS_RECEIVE));
         if (!w) return;
@@ -660,36 +740,44 @@ static void poll_ready(VALUE self) {
     if (s->poll_error || s->selected_pending) s->policy_pending = 1;
 }
 
-// Cancel an io_uring receive and wait until the original operation has finished.
-// Keep the caller's buffer locked throughout, because a cancellation reply alone
-// does not prove the kernel has stopped writing to it.
-void fs_cancel_receive(VALUE self, fs_wait *w) {
+// Cancel an operation and consume its original CQE before releasing storage.
+// Cancellation acknowledgements use zero userdata, so a delayed acknowledgement
+// cannot refer to an operation record or a subsequently reused suspension.
+static void cancel_operation(VALUE self, fs_wait *w, uint64_t kind, int *pending) {
 #ifdef FARCE_URING
     fs_state *s = fs_get(self);
     struct io_uring_sqe *entry = io_uring_get_sqe(&s->ring);
     if (!entry) {
         int result;
         do { result = io_uring_submit(&s->ring); } while (result == -EINTR);
-        if (result < 0) rb_bug("cannot submit pending scheduler receive");
+        if (result < 0) rb_bug("cannot submit pending scheduler operation");
         entry = io_uring_get_sqe(&s->ring);
         if (!entry) rb_bug("cannot reserve scheduler cancellation");
     }
-    io_uring_prep_cancel64(entry, FS_RECEIVE | NUM2ULL(w->token), 0);
+    io_uring_prep_cancel64(entry, kind | NUM2ULL(w->token), 0);
     entry->user_data = 0;
-    // Wait for the original receive CQE, not just the cancellation CQE. No Ruby
-    // fiber runs here, and the caller's buffer remains locked throughout.
-    while (w->receive_pending) {
+    // Do not run fibers here. The caller's buffer or futex word stays alive
+    // until the original completion, rather than just the cancellation reply.
+    while (*pending) {
         int result = io_uring_submit_and_wait(&s->ring, 1);
-        if (result < 0 && result != -EINTR) rb_bug("cannot acknowledge scheduler receive cancellation");
+        if (result < 0 && result != -EINTR) rb_bug("cannot acknowledge scheduler cancellation");
         s->timeout = 0;
         fs_driver_poll(s);
-        if (s->poll_error && s->poll_error != EINTR) rb_bug("cannot drain scheduler receive cancellation");
+        if (s->poll_error && s->poll_error != EINTR) rb_bug("cannot drain scheduler cancellation");
         for (int i = 0; i < s->event_count; i++) accept_event(self, &s->events[i]);
         rearm_events(s);
     }
 #endif
 }
 
+void fs_cancel_receive(VALUE self, fs_wait *w) {
+    cancel_operation(self, w, FS_RECEIVE, &w->receive_pending);
+}
+
+static void fs_cancel_futex(VALUE self, fs_wait *w) {
+    cancel_operation(self, w, FS_FUTEX, &w->futex_pending);
+}
+
 // Call a helper for every active wait, skipping empty slots and idle fibers.
 static void each_wait(fs_state *s, int (*callback)(st_data_t, st_data_t, st_data_t), st_data_t arg) {
     for (uint32_t i = 0; i < s->slot_count; i++) {
@@ -836,6 +924,7 @@ void farce_init_fiber_scheduler(VALUE internal) {
     rb_define_private_method(klass, "retire_wait", fs_retire, 1);
     rb_define_private_method(klass, "register_select", register_select, 3);
     rb_define_private_method(klass, "ready?", ready, 0);
+    rb_define_private_method(klass, "futex_supported?", futex_supported, 0);
     rb_define_private_method(klass, "pending?", pending, 0);
     rb_define_private_method(klass, "pending_count", pending_count, 0);
     rb_define_private_method(klass, "wakeup", wakeup, 0);

data/ext/farce/reactor.h

@@ -1,6 +1,7 @@
 #ifndef FARCE_FIBER_REACTOR_H
 #define FARCE_FIBER_REACTOR_H
 #include "containers.h"
+#include "futex.h"
 #include <ruby/io.h>
 #include <ruby/io/buffer.h>
 #include <ruby/thread.h>
@@ -21,6 +22,8 @@
 #define FS_WRITE 4
 #define FS_EVENTS 1024
 #define FS_RECEIVE (UINT64_C(1) << 63)
+#define FS_FUTEX (UINT64_C(1) << 62)
+#define FS_OPERATION (FS_RECEIVE | FS_FUTEX)
 #define FS_SLOT_BITS 20
 #define FS_SLOT_MASK ((1U << FS_SLOT_BITS) - 1)
 
@@ -45,7 +48,7 @@ struct fs_reg {
 struct fs_wait {
     VALUE token, fiber, value, error, groups, ios;
     uint32_t slot;
-    int status, events, selected, queued, native_error, receive_pending;
+    int status, events, selected, queued, native_error, receive_pending, futex_pending;
     fs_wait *ready_next, *ready_prev;
     fs_watch *watches;
     fs_watch first_watch;
@@ -62,7 +65,7 @@ typedef struct {
     size_t pending;
     uint64_t serial, generation, suspensions;
     unsigned immediate;
-    int selected_pending, receive_disabled;
+    int selected_pending, receive_disabled, futex_supported;
     int driver, descriptor, wake[2], closed;
     pthread_mutex_t wake_mutex;
 #ifdef FARCE_URING
@@ -85,6 +88,7 @@ static inline fs_state *fs_get(VALUE self) {
 int fs_driver_init(fs_state *s, int driver);
 int fs_driver_update(fs_state *s, fs_reg *reg, int events);
 int fs_driver_receive(fs_state *s, fs_wait *w, int fd, void *buffer, size_t size);
+int fs_driver_futex(fs_state *s, fs_wait *w, containers_futex_t *word);
 void fs_cancel_receive(VALUE self, fs_wait *w);
 void fs_driver_close(fs_state *s);
 void *fs_driver_poll(void *data);