std: directly use pthread in UNIX parker implementation

library/std/src/sys/unix/mod.rs

@@ -39,6 +39,7 @@ pub mod stdio;
 pub mod thread;
 pub mod thread_local_dtor;
 pub mod thread_local_key;
+pub mod thread_parker;
 pub mod time;
 
 #[cfg(target_os = "espidf")]

library/std/src/sys/unix/thread_parker.rs

@@ -0,0 +1,303 @@
+//! Thread parking without `futex` using the `pthread` synchronization primitives.
+
+#![cfg(not(any(
+    target_os = "linux",
+    target_os = "android",
+    all(target_os = "emscripten", target_feature = "atomics")
+)))]
+
+use crate::cell::UnsafeCell;
+use crate::marker::PhantomPinned;
+use crate::pin::Pin;
+use crate::ptr::addr_of_mut;
+use crate::sync::atomic::AtomicUsize;
+use crate::sync::atomic::Ordering::SeqCst;
+use crate::time::Duration;
+
+const EMPTY: usize = 0;
+const PARKED: usize = 1;
+const NOTIFIED: usize = 2;
+
+unsafe fn lock(lock: *mut libc::pthread_mutex_t) {
+    let r = libc::pthread_mutex_lock(lock);
+    debug_assert_eq!(r, 0);
+}
+
+unsafe fn unlock(lock: *mut libc::pthread_mutex_t) {
+    let r = libc::pthread_mutex_unlock(lock);
+    debug_assert_eq!(r, 0);
+}
+
+unsafe fn notify_one(cond: *mut libc::pthread_cond_t) {
+    let r = libc::pthread_cond_signal(cond);
+    debug_assert_eq!(r, 0);
+}
+
+unsafe fn wait(cond: *mut libc::pthread_cond_t, lock: *mut libc::pthread_mutex_t) {
+    let r = libc::pthread_cond_wait(cond, lock);
+    debug_assert_eq!(r, 0);
+}
+
+const TIMESPEC_MAX: libc::timespec =
+    libc::timespec { tv_sec: <libc::time_t>::MAX, tv_nsec: 1_000_000_000 - 1 };
+
+fn saturating_cast_to_time_t(value: u64) -> libc::time_t {
+    if value > <libc::time_t>::MAX as u64 { <libc::time_t>::MAX } else { value as libc::time_t }
+}
+
+// This implementation is used on systems that support pthread_condattr_setclock
+// where we configure the condition variable to use the monotonic clock (instead of
+// the default system clock). This approach avoids all problems that result
+// from changes made to the system time.
+#[cfg(not(any(target_os = "macos", target_os = "ios", target_os = "espidf")))]
+unsafe fn wait_timeout(
+    cond: *mut libc::pthread_cond_t,
+    lock: *mut libc::pthread_mutex_t,
+    dur: Duration,
+) {
+    use crate::mem;
+
+    let mut now: libc::timespec = mem::zeroed();
+    let r = libc::clock_gettime(libc::CLOCK_MONOTONIC, &mut now);
+    assert_eq!(r, 0);
+    // Nanosecond calculations can't overflow because both values are below 1e9.
+    let nsec = dur.subsec_nanos() + now.tv_nsec as u32;
+    let sec = saturating_cast_to_time_t(dur.as_secs())
+        .checked_add((nsec / 1_000_000_000) as libc::time_t)
+        .and_then(|s| s.checked_add(now.tv_sec));
+    let nsec = nsec % 1_000_000_000;
+    let timeout =
+        sec.map(|s| libc::timespec { tv_sec: s, tv_nsec: nsec as _ }).unwrap_or(TIMESPEC_MAX);
+    let r = libc::pthread_cond_timedwait(cond, lock, &timeout);
+    assert!(r == libc::ETIMEDOUT || r == 0);
+}
+
+// This implementation is modeled after libcxx's condition_variable
+// https://github.com/llvm-mirror/libcxx/blob/release_35/src/condition_variable.cpp#L46
+// https://github.com/llvm-mirror/libcxx/blob/release_35/include/__mutex_base#L367
+#[cfg(any(target_os = "macos", target_os = "ios", target_os = "espidf"))]
+unsafe fn wait_timeout(
+    cond: *mut libc::pthread_cond_t,
+    lock: *mut libc::pthread_mutex_t,
+    mut dur: Duration,
+) {
+    use crate::ptr;
+
+    // 1000 years
+    let max_dur = Duration::from_secs(1000 * 365 * 86400);
+
+    if dur > max_dur {
+        // OSX implementation of `pthread_cond_timedwait` is buggy
+        // with super long durations. When duration is greater than
+        // 0x100_0000_0000_0000 seconds, `pthread_cond_timedwait`
+        // in macOS Sierra return error 316.
+        //
+        // This program demonstrates the issue:
+        // https://gist.github.com/stepancheg/198db4623a20aad2ad7cddb8fda4a63c
+        //
+        // To work around this issue, and possible bugs of other OSes, timeout
+        // is clamped to 1000 years, which is allowable per the API of `park_timeout`
+        // because of spurious wakeups.
+        dur = max_dur;
+    }
+
+    let mut sys_now = libc::timeval { tv_sec: 0, tv_usec: 0 };
+    let r = libc::gettimeofday(&mut sys_now, ptr::null_mut());
+    debug_assert_eq!(r, 0);
+    let nsec = dur.subsec_nanos() as libc::c_long + (sys_now.tv_usec * 1000) as libc::c_long;
+    let extra = (nsec / 1_000_000_000) as libc::time_t;
+    let nsec = nsec % 1_000_000_000;
+    let seconds = saturating_cast_to_time_t(dur.as_secs());
+    let timeout = sys_now
+        .tv_sec
+        .checked_add(extra)
+        .and_then(|s| s.checked_add(seconds))
+        .map(|s| libc::timespec { tv_sec: s, tv_nsec: nsec })
+        .unwrap_or(TIMESPEC_MAX);
+    // And wait!
+    let r = libc::pthread_cond_timedwait(cond, lock, &timeout);
+    debug_assert!(r == libc::ETIMEDOUT || r == 0);
+}
+
+pub struct Parker {
+    state: AtomicUsize,
+    lock: UnsafeCell<libc::pthread_mutex_t>,
+    cvar: UnsafeCell<libc::pthread_cond_t>,
+    // The `pthread` primitives require a stable address, so make this struct `!Unpin`.
+    _pinned: PhantomPinned,
+}
+
+impl Parker {
+    /// Construct the UNIX parker in-place.
+    ///
+    /// # Safety
+    /// The constructed parker must never be moved.
+    pub unsafe fn new(parker: *mut Parker) {
+        // Use the default mutex implementation to allow for simpler initialization.
+        // This could lead to undefined behaviour when deadlocking. This is avoided
+        // by not deadlocking. Note in particular the unlocking operation before any
+        // panic, as code after the panic could try to park again.
+        addr_of_mut!((*parker).state).write(AtomicUsize::new(EMPTY));
+        addr_of_mut!((*parker).lock).write(UnsafeCell::new(libc::PTHREAD_MUTEX_INITIALIZER));
+
+        cfg_if::cfg_if! {
+            if #[cfg(any(
+                target_os = "macos",
+                target_os = "ios",
+                target_os = "l4re",
+                target_os = "android",
+                target_os = "redox"
+            ))] {
+                addr_of_mut!((*parker).cvar).write(UnsafeCell::new(libc::PTHREAD_COND_INITIALIZER));
+            } else if #[cfg(target_os = "espidf")] {
+                let r = libc::pthread_cond_init(addr_of_mut!((*parker).cvar).cast(), crate::ptr::null());
+                assert_eq!(r, 0);
+            } else {
+                use crate::mem::MaybeUninit;
+                let mut attr = MaybeUninit::<libc::pthread_condattr_t>::uninit();
+                let r = libc::pthread_condattr_init(attr.as_mut_ptr());
+                assert_eq!(r, 0);
+                let r = libc::pthread_condattr_setclock(attr.as_mut_ptr(), libc::CLOCK_MONOTONIC);
+                assert_eq!(r, 0);
+                let r = libc::pthread_cond_init(addr_of_mut!((*parker).cvar).cast(), attr.as_ptr());
+                assert_eq!(r, 0);
+                let r = libc::pthread_condattr_destroy(attr.as_mut_ptr());
+                assert_eq!(r, 0);
+            }
+        }
+    }
+
+    // This implementation doesn't require `unsafe`, but other implementations
+    // may assume this is only called by the thread that owns the Parker.
+    pub unsafe fn park(self: Pin<&Self>) {
+        // If we were previously notified then we consume this notification and
+        // return quickly.
+        if self.state.compare_exchange(NOTIFIED, EMPTY, SeqCst, SeqCst).is_ok() {
+            return;
+        }
+
+        // Otherwise we need to coordinate going to sleep
+        lock(self.lock.get());
+        match self.state.compare_exchange(EMPTY, PARKED, SeqCst, SeqCst) {
+            Ok(_) => {}
+            Err(NOTIFIED) => {
+                // We must read here, even though we know it will be `NOTIFIED`.
+                // This is because `unpark` may have been called again since we read
+                // `NOTIFIED` in the `compare_exchange` above. We must perform an
+                // acquire operation that synchronizes with that `unpark` to observe
+                // any writes it made before the call to unpark. To do that we must
+                // read from the write it made to `state`.
+                let old = self.state.swap(EMPTY, SeqCst);
+
+                unlock(self.lock.get());
+
+                assert_eq!(old, NOTIFIED, "park state changed unexpectedly");
+                return;
+            } // should consume this notification, so prohibit spurious wakeups in next park.
+            Err(_) => {
+                unlock(self.lock.get());
+
+                panic!("inconsistent park state")
+            }
+        }
+
+        loop {
+            wait(self.cvar.get(), self.lock.get());
+
+            match self.state.compare_exchange(NOTIFIED, EMPTY, SeqCst, SeqCst) {
+                Ok(_) => break, // got a notification
+                Err(_) => {}    // spurious wakeup, go back to sleep
+            }
+        }
+
+        unlock(self.lock.get());
+    }
+
+    // This implementation doesn't require `unsafe`, but other implementations
+    // may assume this is only called by the thread that owns the Parker. Use
+    // `Pin` to guarantee a stable address for the mutex and condition variable.
+    pub unsafe fn park_timeout(self: Pin<&Self>, dur: Duration) {
+        // Like `park` above we have a fast path for an already-notified thread, and
+        // afterwards we start coordinating for a sleep.
+        // return quickly.
+        if self.state.compare_exchange(NOTIFIED, EMPTY, SeqCst, SeqCst).is_ok() {
+            return;
+        }
+
+        lock(self.lock.get());
+        match self.state.compare_exchange(EMPTY, PARKED, SeqCst, SeqCst) {
+            Ok(_) => {}
+            Err(NOTIFIED) => {
+                // We must read again here, see `park`.
+                let old = self.state.swap(EMPTY, SeqCst);
+                unlock(self.lock.get());
+
+                assert_eq!(old, NOTIFIED, "park state changed unexpectedly");
+                return;
+            } // should consume this notification, so prohibit spurious wakeups in next park.
+            Err(_) => {
+                unlock(self.lock.get());
+                panic!("inconsistent park_timeout state")
+            }
+        }
+
+        // Wait with a timeout, and if we spuriously wake up or otherwise wake up
+        // from a notification we just want to unconditionally set the state back to
+        // empty, either consuming a notification or un-flagging ourselves as
+        // parked.
+        wait_timeout(self.cvar.get(), self.lock.get(), dur);
+
+        match self.state.swap(EMPTY, SeqCst) {
+            NOTIFIED => unlock(self.lock.get()), // got a notification, hurray!
+            PARKED => unlock(self.lock.get()),   // no notification, alas
+            n => {
+                unlock(self.lock.get());
+                panic!("inconsistent park_timeout state: {n}")
+            }
+        }
+    }
+
+    pub fn unpark(self: Pin<&Self>) {
+        // To ensure the unparked thread will observe any writes we made
+        // before this call, we must perform a release operation that `park`
+        // can synchronize with. To do that we must write `NOTIFIED` even if
+        // `state` is already `NOTIFIED`. That is why this must be a swap
+        // rather than a compare-and-swap that returns if it reads `NOTIFIED`
+        // on failure.
+        match self.state.swap(NOTIFIED, SeqCst) {
+            EMPTY => return,    // no one was waiting
+            NOTIFIED => return, // already unparked
+            PARKED => {}        // gotta go wake someone up
+            _ => panic!("inconsistent state in unpark"),
+        }
+
+        // There is a period between when the parked thread sets `state` to
+        // `PARKED` (or last checked `state` in the case of a spurious wake
+        // up) and when it actually waits on `cvar`. If we were to notify
+        // during this period it would be ignored and then when the parked
+        // thread went to sleep it would never wake up. Fortunately, it has
+        // `lock` locked at this stage so we can acquire `lock` to wait until
+        // it is ready to receive the notification.
+        //
+        // Releasing `lock` before the call to `notify_one` means that when the
+        // parked thread wakes it doesn't get woken only to have to wait for us
+        // to release `lock`.
+        unsafe {
+            lock(self.lock.get());
+            unlock(self.lock.get());
+            notify_one(self.cvar.get());
+        }
+    }
+}
+
+impl Drop for Parker {
+    fn drop(&mut self) {
+        unsafe {
+            libc::pthread_cond_destroy(self.cvar.get_mut());
+            libc::pthread_mutex_destroy(self.lock.get_mut());
+        }
+    }
+}
+
+unsafe impl Sync for Parker {}
+unsafe impl Send for Parker {}

library/std/src/sys/windows/thread_parker.rs

@@ -58,6 +58,7 @@
 // [4]: Windows Internals, Part 1, ISBN 9780735671300
 
 use crate::convert::TryFrom;
+use crate::pin::Pin;
 use crate::ptr;
 use crate::sync::atomic::{
     AtomicI8, AtomicPtr,
@@ -95,13 +96,16 @@ const NOTIFIED: i8 = 1;
 // Ordering::Release when writing NOTIFIED (the 'token') in unpark(), and using
 // Ordering::Acquire when reading this state in park() after waking up.
 impl Parker {
-    pub fn new() -> Self {
-        Self { state: AtomicI8::new(EMPTY) }
+    /// Construct the Windows parker. The UNIX parker implementation
+    /// requires this to happen in-place.
+    pub unsafe fn new(parker: *mut Parker) {
+        parker.write(Self { state: AtomicI8::new(EMPTY) });
     }
 
     // Assumes this is only called by the thread that owns the Parker,
-    // which means that `self.state != PARKED`.
-    pub unsafe fn park(&self) {
+    // which means that `self.state != PARKED`. This implementation doesn't require `Pin`,
+    // but other implementations do.
+    pub unsafe fn park(self: Pin<&Self>) {
         // Change NOTIFIED=>EMPTY or EMPTY=>PARKED, and directly return in the
         // first case.
         if self.state.fetch_sub(1, Acquire) == NOTIFIED {
@@ -132,8 +136,9 @@ impl Parker {
     }
 
     // Assumes this is only called by the thread that owns the Parker,
-    // which means that `self.state != PARKED`.
-    pub unsafe fn park_timeout(&self, timeout: Duration) {
+    // which means that `self.state != PARKED`. This implementation doesn't require `Pin`,
+    // but other implementations do.
+    pub unsafe fn park_timeout(self: Pin<&Self>, timeout: Duration) {
         // Change NOTIFIED=>EMPTY or EMPTY=>PARKED, and directly return in the
         // first case.
         if self.state.fetch_sub(1, Acquire) == NOTIFIED {
@@ -184,7 +189,8 @@ impl Parker {
         }
     }
 
-    pub fn unpark(&self) {
+    // This implementation doesn't require `Pin`, but other implementations do.
+    pub fn unpark(self: Pin<&Self>) {
         // Change PARKED=>NOTIFIED, EMPTY=>NOTIFIED, or NOTIFIED=>NOTIFIED, and
         // wake the thread in the first case.
         //