Windows: Implement mutex using futex

Well, the Windows equivalent: `WaitOnAddress`, `WakeByAddressSingle` and `WakeByAddressAll`.

Author: ChrisDenton

library/std/src/sys/locks/condvar/mod.rs

@@ -1,5 +1,6 @@
 cfg_if::cfg_if! {
     if #[cfg(any(
+        all(target_os = "windows", not(target_vendor="win7")),
         target_os = "linux",
         target_os = "android",
         target_os = "freebsd",
@@ -14,9 +15,9 @@ cfg_if::cfg_if! {
     } else if #[cfg(target_family = "unix")] {
         mod pthread;
         pub use pthread::Condvar;
-    } else if #[cfg(target_os = "windows")] {
-        mod windows;
-        pub use windows::Condvar;
+    } else if #[cfg(all(target_os = "windows", target_vendor = "win7"))] {
+        mod windows7;
+        pub use windows7::Condvar;
     } else if #[cfg(all(target_vendor = "fortanix", target_env = "sgx"))] {
         mod sgx;
         pub use sgx::Condvar;

library/std/src/sys/locks/condvar/windows.rs renamed to library/std/src/sys/locks/condvar/windows7.rs

@@ -1,36 +1,89 @@
 use crate::sync::atomic::{
-    AtomicU32,
+    self,
     Ordering::{Acquire, Relaxed, Release},
 };
-use crate::sys::futex::{futex_wait, futex_wake};
+use crate::sys::futex::futex_wake;
+
+cfg_if::cfg_if! {
+if #[cfg(windows)] {
+    // On Windows we can have a smol futex
+    type Atomic = atomic::AtomicU8;
+    type State = u8;
+} else {
+    type Atomic = atomic::AtomicU32;
+    type State = u32;
+}
+}
 
 pub struct Mutex {
-    /// 0: unlocked
-    /// 1: locked, no other threads waiting
-    /// 2: locked, and other threads waiting (contended)
-    futex: AtomicU32,
+    futex: Atomic,
 }
 
+const UNLOCKED: State = 0;
+const LOCKED: State = 1; // locked, no other threads waiting
+const CONTENDED: State = 2; // locked, and other threads waiting (contended)
+
 impl Mutex {
     #[inline]
-    pub const fn new() -> Self {
-        Self { futex: AtomicU32::new(0) }
+    pub const fn new() -> Mutex {
+        Mutex { futex: Atomic::new(UNLOCKED) }
+    }
+
+    #[inline]
+    pub fn lock(&self) {
+        if let Err(state) = self.futex.compare_exchange(UNLOCKED, LOCKED, Acquire, Relaxed) {
+            self.lock_contended(state)
+        }
     }
 
     #[inline]
     pub fn try_lock(&self) -> bool {
-        self.futex.compare_exchange(0, 1, Acquire, Relaxed).is_ok()
+        self.futex.compare_exchange(UNLOCKED, LOCKED, Acquire, Relaxed).is_ok()
     }
 
     #[inline]
-    pub fn lock(&self) {
-        if self.futex.compare_exchange(0, 1, Acquire, Relaxed).is_err() {
-            self.lock_contended();
+    pub unsafe fn unlock(&self) {
+        if self.futex.swap(UNLOCKED, Release) == CONTENDED {
+            // We only wake up one thread. When that thread locks the mutex, it
+            // will mark the mutex as CONTENDED (see lock_contended above),
+            // which makes sure that any other waiting threads will also be
+            // woken up eventually.
+            self.wake();
         }
     }
 
     #[cold]
-    fn lock_contended(&self) {
+    fn wake(&self) {
+        futex_wake(&self.futex);
+    }
+}
+
+#[cfg(windows)]
+impl Mutex {
+    #[cold]
+    fn lock_contended(&self, mut state: State) {
+        use crate::sys::futex::wait_on_address;
+        // Note: WaitOnAddress is already quite spin-happy so we don't do any further spinning on top.
+        loop {
+            // Put the lock in contended state.
+            // We avoid an unnecessary write if it as already set to CONTENDED,
+            // to be friendlier for the caches.
+            if state != CONTENDED && self.futex.swap(CONTENDED, Acquire) == UNLOCKED {
+                // We changed it from UNLOCKED to CONTENDED, so we just successfully locked it.
+                return;
+            }
+            // Wait for the futex to change state, assuming it is still CONTENDED.
+            wait_on_address(&self.futex, CONTENDED, None);
+            state = self.futex.load(Relaxed);
+        }
+    }
+}
+
+#[cfg(not(windows))]
+impl Mutex {
+    #[cold]
+    fn lock_contended(&self, _state: State) {
+        use crate::sys::futex::futex_wait;
         // Spin first to speed things up if the lock is released quickly.
         let mut state = self.spin();
 
@@ -77,20 +130,4 @@ impl Mutex {
             spin -= 1;
         }
     }
-
-    #[inline]
-    pub unsafe fn unlock(&self) {
-        if self.futex.swap(0, Release) == 2 {
-            // We only wake up one thread. When that thread locks the mutex, it
-            // will mark the mutex as contended (2) (see lock_contended above),
-            // which makes sure that any other waiting threads will also be
-            // woken up eventually.
-            self.wake();
-        }
-    }
-
-    #[cold]
-    fn wake(&self) {
-        futex_wake(&self.futex);
-    }
 }

library/std/src/sys/locks/mutex/futex.rs

@@ -1,5 +1,6 @@
 cfg_if::cfg_if! {
     if #[cfg(any(
+        all(target_os = "windows", not(target_vendor = "win7")),
         target_os = "linux",
         target_os = "android",
         target_os = "freebsd",
@@ -19,9 +20,9 @@ cfg_if::cfg_if! {
     ))] {
         mod pthread;
         pub use pthread::{Mutex, raw};
-    } else if #[cfg(target_os = "windows")] {
-        mod windows;
-        pub use windows::{Mutex, raw};
+    } else if #[cfg(all(target_os = "windows", target_vendor = "win7"))] {
+        mod windows7;
+        pub use windows7::{Mutex, raw};
     } else if #[cfg(all(target_vendor = "fortanix", target_env = "sgx"))] {
         mod sgx;
         pub use sgx::Mutex;

library/std/src/sys/locks/mutex/mod.rs

@@ -291,7 +291,10 @@ impl RwLock {
     }
 
     /// Spin for a while, but stop directly at the given condition.
+    ///
+    /// We avoid spinning on Windows because the futex implementation spins enough.
     #[inline]
+    #[cfg(not(windows))]
     fn spin_until(&self, f: impl Fn(u32) -> bool) -> u32 {
         let mut spin = 100; // Chosen by fair dice roll.
         loop {
@@ -304,6 +307,12 @@ impl RwLock {
         }
     }
 
+    #[inline]
+    #[cfg(windows)]
+    fn spin_until(&self, _f: impl Fn(u32) -> bool) -> u32 {
+        self.state.load(Relaxed)
+    }
+
     #[inline]
     fn spin_write(&self) -> u32 {
         // Stop spinning when it's unlocked or when there's waiting writers, to keep things somewhat fair.

library/std/src/sys/locks/mutex/windows.rs renamed to library/std/src/sys/locks/mutex/windows7.rs

@@ -1,5 +1,6 @@
 cfg_if::cfg_if! {
     if #[cfg(any(
+        all(target_os = "windows", not(target_vendor = "win7")),
         target_os = "linux",
         target_os = "android",
         target_os = "freebsd",
@@ -14,9 +15,9 @@ cfg_if::cfg_if! {
     } else if #[cfg(target_family = "unix")] {
         mod queue;
         pub use queue::RwLock;
-    } else if #[cfg(target_os = "windows")] {
-        mod windows;
-        pub use windows::RwLock;
+    } else if #[cfg(all(target_os = "windows", target_vendor = "win7"))] {
+        mod windows7;
+        pub use windows7::RwLock;
     } else if #[cfg(all(target_vendor = "fortanix", target_env = "sgx"))] {
         mod sgx;
         pub use sgx::RwLock;

library/std/src/sys/locks/rwlock/futex.rs

@@ -36,6 +36,7 @@ pub type LPVOID = *mut c_void;
 pub type LPWCH = *mut WCHAR;
 pub type LPWSTR = *mut WCHAR;
 
+#[cfg(target_vendor = "win7")]
 pub type PSRWLOCK = *mut SRWLOCK;
 
 pub type socklen_t = c_int;
@@ -50,7 +51,9 @@ pub const INVALID_HANDLE_VALUE: HANDLE = ::core::ptr::without_provenance_mut(-1i
 pub const EXIT_SUCCESS: u32 = 0;
 pub const EXIT_FAILURE: u32 = 1;
 
+#[cfg(target_vendor = "win7")]
 pub const CONDITION_VARIABLE_INIT: CONDITION_VARIABLE = CONDITION_VARIABLE { Ptr: ptr::null_mut() };
+#[cfg(target_vendor = "win7")]
 pub const SRWLOCK_INIT: SRWLOCK = SRWLOCK { Ptr: ptr::null_mut() };
 pub const INIT_ONCE_STATIC_INIT: INIT_ONCE = INIT_ONCE { Ptr: ptr::null_mut() };
 
@@ -373,6 +376,7 @@ extern "system" {
         dwmilliseconds: u32,
     ) -> BOOL;
     pub fn WakeByAddressSingle(address: *const c_void);
+    pub fn WakeByAddressAll(address: *const c_void);
 }
 
 #[cfg(target_vendor = "win7")]

library/std/src/sys/locks/rwlock/mod.rs

@@ -0,0 +1,54 @@
+use crate::sys::api;
+use crate::sys::c;
+use crate::sys::dur2timeout;
+use core::ffi::c_void;
+use core::mem;
+use core::ptr;
+use core::time::Duration;
+
+#[inline(always)]
+pub fn wait_on_address<T, U>(address: &T, compare: U, timeout: Option<Duration>) -> bool {
+    assert_eq!(mem::size_of::<T>(), mem::size_of::<U>());
+    unsafe {
+        let addr = ptr::addr_of!(*address).cast::<c_void>();
+        let size = mem::size_of::<T>();
+        let compare_addr = ptr::addr_of!(compare).cast::<c_void>();
+        let timeout = timeout.map(dur2timeout).unwrap_or(c::INFINITE);
+        c::WaitOnAddress(addr, compare_addr, size, timeout) == c::TRUE
+    }
+}
+
+#[inline(always)]
+pub fn wake_by_address_single<T>(address: &T) -> bool {
+    unsafe {
+        let addr = ptr::addr_of!(*address).cast::<c_void>();
+        c::WakeByAddressSingle(addr);
+        false
+    }
+}
+
+#[inline(always)]
+pub fn wake_by_address_all<T>(address: &T) {
+    unsafe {
+        let addr = ptr::addr_of!(*address).cast::<c_void>();
+        c::WakeByAddressAll(addr);
+    }
+}
+
+#[inline(always)]
+pub fn futex_wait<T, U>(futex: &T, expected: U, timeout: Option<Duration>) -> bool {
+    // return false only on timeout
+    if wait_on_address(futex, expected, timeout) {
+        true
+    } else {
+        api::get_last_error().code != c::ERROR_TIMEOUT
+    }
+}
+#[inline(always)]
+pub fn futex_wake<T>(futex: &T) -> bool {
+    wake_by_address_single(futex);
+    false
+}
+pub fn futex_wake_all<T>(futex: &T) {
+    wake_by_address_all(futex)
+}

library/std/src/sys/locks/rwlock/windows.rs renamed to library/std/src/sys/locks/rwlock/windows7.rs

@@ -17,6 +17,8 @@ pub mod args;
 pub mod c;
 pub mod env;
 pub mod fs;
+#[cfg(not(target_vendor = "win7"))]
+pub mod futex;
 pub mod handle;
 pub mod io;
 pub mod net;
@@ -39,7 +41,7 @@ cfg_if::cfg_if! {
     }
 }
 
-mod api;
+pub(in crate::sys) mod api;
 
 /// Map a Result<T, WinError> to io::Result<T>.
 trait IoResult<T> {

library/std/src/sys/pal/windows/c.rs

@@ -366,6 +366,12 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriInterpCxExt<'mir, 'tcx> {
 
                 this.WakeByAddressSingle(ptr_op)?;
             }
+            "WakeByAddressAll" => {
+                let [ptr_op] =
+                    this.check_shim(abi, Abi::System { unwind: false }, link_name, args)?;
+
+                this.WakeByAddressAll(ptr_op)?;
+            }
 
             // Dynamic symbol loading
             "GetProcAddress" => {

library/std/src/sys/pal/windows/futex.rs

@@ -384,6 +384,21 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriInterpCxExt<'mir, 'tcx> {
 
         Ok(())
     }
+    fn WakeByAddressAll(&mut self, ptr_op: &OpTy<'tcx, Provenance>) -> InterpResult<'tcx> {
+        let this = self.eval_context_mut();
+
+        let ptr = this.read_pointer(ptr_op)?;
+
+        // See the Linux futex implementation for why this fence exists.
+        this.atomic_fence(AtomicFenceOrd::SeqCst)?;
+
+        while let Some(thread) = this.futex_wake(ptr.addr().bytes(), u32::MAX) {
+            this.unblock_thread(thread);
+            this.unregister_timeout_callback_if_exists(thread);
+        }
+
+        Ok(())
+    }
 
     fn SleepConditionVariableSRW(
         &mut self,