Skip to main content

std\sys\thread_local\guard/
windows.rs

1//! Support for Windows TLS destructors.
2//!
3//! Windows has an API to provide a destructor for a FLS (fiber local storage) variable,
4//! which behaves similarly to a TLS variable for our purpose [1].
5//!
6//! All TLS destructors are tracked by *us*, not the Windows runtime.
7//! This means that we have a global list of destructors for
8//! each TLS key or variable that we know about.
9//!
10//! [1]: https://devblogs.microsoft.com/oldnewthing/20191011-00/?p=102989
11
12use core::ffi::c_void;
13use core::sync::atomic::{AtomicBool, AtomicU32, Ordering, fence};
14
15use crate::cell::Cell;
16use crate::ptr;
17use crate::sys::c::{self, FLS_OUT_OF_INDEXES};
18
19pub type Key = u32;
20
21unsafe fn create(dtor: c::PFLS_CALLBACK_FUNCTION) -> Key {
22    let key_result = unsafe { c::FlsAlloc(dtor) };
23
24    if key_result == c::FLS_OUT_OF_INDEXES {
25        rtabort!("out of FLS keys");
26    }
27
28    key_result
29}
30
31unsafe fn set(key: Key, ptr: *const c_void) {
32    let result = unsafe { c::FlsSetValue(key, ptr) };
33
34    if result == c::FALSE {
35        rtabort!("failed to set FLS value");
36    }
37}
38
39fn is_thread_a_fiber() -> bool {
40    let res = unsafe { c::IsThreadAFiber() };
41    res == c::TRUE
42}
43
44static KEY: AtomicU32 = AtomicU32::new(FLS_OUT_OF_INDEXES);
45
46/// Used to track whether we are currently in the critical section of `enable`.
47/// For miri, these atomic operations cause synchronization that can mask user bugs,
48/// and they are not needed as `atexit` is anyway not supported, so we can skip them.
49struct EnableGuard;
50static AT_EXIT_HOOK_CALLED: AtomicBool = AtomicBool::new(false);
51static ACTIVE_ENABLE_CALLS: AtomicU32 = AtomicU32::new(0);
52
53impl EnableGuard {
54    // Mark the start of an `enable` call, returning whether the `atexit` hook has already been called or not.
55    fn new() -> (Self, bool) {
56        if cfg!(miri) {
57            return (Self, false);
58        }
59        ACTIVE_ENABLE_CALLS.fetch_add(1, Ordering::Relaxed);
60
61        // Both `new` and `start_exit` publish state to one atomic and inspect the other.
62        // `AcqRel` is insufficient because neither read is required to observe the other's publication,
63        // so we could create the guard but `start_exit` would not see any active enable calls.
64        // `SeqCst` ensures that there's a single global order between the publish and check,
65        // so at least one side must observe the other and bail.
66        fence(Ordering::SeqCst);
67
68        let at_exit_called = AT_EXIT_HOOK_CALLED.load(Ordering::Relaxed);
69
70        (Self, at_exit_called)
71    }
72
73    /// Mark the start of process exit, returning whether we should free the FLS key or not.
74    fn start_exit() -> bool {
75        // After this hook starts, new destructor registration will be skipped,
76        // causing TLS destructors initialized after this point to leak.
77        if AT_EXIT_HOOK_CALLED.swap(true, Ordering::Relaxed) {
78            // Cleanup already started, there is nothing else to do.
79            return false;
80        }
81
82        fence(Ordering::SeqCst);
83
84        let any_active_enabled_called = ACTIVE_ENABLE_CALLS.load(Ordering::Relaxed) != 0;
85
86        if any_active_enabled_called {
87            // If another thread is currently in `enable`, it may already have loaded this key and may be about to call `FlsSetValue`.
88            // So we must *not* call free the FLS key.
89            //
90            // During real process exit this is harmless because the `cleanup` hook is always available,
91            // and the FLS callback will be triggered normally by the OS.
92            //
93            // During DLL unload, the unloader cannot safely have threads running code from the DLL except for the destructors,
94            // so there must not be any `enable` calls active anyway.
95            return false;
96        }
97
98        return true;
99    }
100}
101
102#[cfg(not(miri))]
103impl Drop for EnableGuard {
104    fn drop(&mut self) {
105        ACTIVE_ENABLE_CALLS.fetch_sub(1, Ordering::Relaxed);
106    }
107}
108
109pub fn enable() {
110    let registered = if cfg!(target_thread_local) {
111        #[thread_local]
112        static REGISTERED: Cell<bool> = Cell::new(false);
113        REGISTERED.replace(true)
114    } else {
115        // `#[thread_local]` is unavailable on windows-gnu (`target_thread_local` is off),
116        // but setting the FLS key's value is about as expensive as `TlsGet`, so we don't bother tracking registration separately.
117        false
118    };
119
120    if !registered {
121        // We are in a critical section where we are trying to register a destructor for the current thread.
122        // We need to avoid racing with the `atexit` hook that frees the FLS slot, which would cause us to call `FlsSetValue` on a freed key,
123        // or calling `atexit` during process shutdown, which would cause a deadlock.
124        let (_guard, at_exit_called) = EnableGuard::new();
125
126        if at_exit_called {
127            // We are exiting and don't want to race with the `atexit` hook, so we won't be able to run the destructors for this thread.
128            return;
129        }
130
131        let current_key = KEY.load(Ordering::Acquire);
132
133        // If we already allocated a key, we only need to set it to a non-null value so that the destructors hook is run for this thread.
134        let key = if current_key != FLS_OUT_OF_INDEXES {
135            current_key
136        } else {
137            // Otherwise, we try to allocate a key.
138            let new_key = unsafe { create(Some(cleanup)) };
139
140            // Now we need to set this key to be used by everyone else.
141            // If we won the race, our key is the right one and we can set it to non-null value.
142            // If we lost, we'll use the winning key and free our losing key.
143            match KEY.compare_exchange(current_key, new_key, Ordering::Release, Ordering::Acquire) {
144                Ok(_) => {
145                    // If the current DLL is unloaded, the registered `cleanup` hook will not be available later during thread exit,
146                    // triggering a `STATUS_ACCESS_VIOLATION`. To avoid this, we use the `atexit` hook, which is called during DLL unload
147                    // to manually free the FLS slot, triggering the destructors.
148                    //
149                    // However, calling `atexit` during process exit can cause a deadlock.
150                    // In a Rust binary, `enable` is called during the main thread startup and before any user code,
151                    // and we checked using `at_exit_called` that we aren't in process shutdown.
152                    //
153                    // In a Rust DLL, dynamic unloading can only happen safely when no other threads are
154                    // concurrently executing Rust code, so if we are here we cannot be unloading yet.
155                    //
156                    // If a main non-Rust binary is exiting, it must not be trigger the `enable` guard
157                    // for the first time during process shutdown.
158                    let res = unsafe { c::atexit(free_fls_key_at_exit) };
159                    if res != 0 {
160                        rtabort!("failed to register fls atexit hook");
161                    }
162
163                    new_key
164                }
165                Err(other_key) => {
166                    unsafe { c::FlsFree(new_key) };
167                    other_key
168                }
169            }
170        };
171
172        // Setting the key's value to non-zero will cause the dtor callback to be called when the thread exits.
173        unsafe { set(key, ptr::without_provenance(1)) };
174    }
175}
176
177extern "C" fn free_fls_key_at_exit() {
178    // The main purpose of this hook is to free the FLS slot during DLL unload.
179    // However, this hook will also be called during normal process exit, while other Rust threads are still running,
180    // so we must be careful to avoid races with `enable`.
181    let should_free_key = EnableGuard::start_exit();
182    if !should_free_key {
183        return;
184    }
185
186    let current_key = KEY.swap(c::FLS_OUT_OF_INDEXES, Ordering::AcqRel);
187    if current_key != c::FLS_OUT_OF_INDEXES {
188        // Calling `FlsFree` will cause the OS to call the `cleanup` hook, in the current thread, *for each thread* (or fiber) with a value in this FLS slot.
189        // `cleanup` is safe to run repeatedly: it only drains the current thread's TLS destructor list, and we check that we are not running in a fiber before doing so.
190        // We only call this when no `enable` call is active, so it cannot race with `FlsSetValue` using this key.
191        // Destructors of thread locals in other threads will not run and therefore leak, which is allowed since we are exiting or unloading.
192        unsafe { c::FlsFree(current_key) };
193    }
194}
195
196unsafe extern "system" fn cleanup(_ptr: *const c_void) {
197    // Avoid running the hook if we are in a fiber.
198    // This will cause destructors of thread locals to not run, leaking them.
199    // Thread-local runtime state will not be cleaned.
200    //
201    // We need to verify that we won't run the destructors *before* the thread exits,
202    // but if the fiber that registered the callback is deleted, the thread might still be running other fibers.
203    //
204    // By checking that we are not running in a fiber here, we are guaranteed that the hook is only running during the thread's exit.
205    // See also the `fiber_does_not_trigger_dtor` test.
206    if is_thread_a_fiber() {
207        return;
208    }
209
210    unsafe {
211        #[cfg(target_thread_local)]
212        super::super::destructors::run();
213        #[cfg(not(target_thread_local))]
214        super::super::key::run_dtors();
215    }
216
217    crate::rt::thread_cleanup();
218}