Debug more information when we fail to find a lock call symbol
[rust-lightning] / lightning / src / sync / debug_sync.rs
1 pub use ::alloc::sync::Arc;
2 use core::ops::{Deref, DerefMut};
3 use core::time::Duration;
4
5 use std::cell::RefCell;
6
7 use std::sync::atomic::{AtomicUsize, Ordering};
8 use std::sync::Mutex as StdMutex;
9 use std::sync::MutexGuard as StdMutexGuard;
10 use std::sync::RwLock as StdRwLock;
11 use std::sync::RwLockReadGuard as StdRwLockReadGuard;
12 use std::sync::RwLockWriteGuard as StdRwLockWriteGuard;
13 use std::sync::Condvar as StdCondvar;
14
15 pub use std::sync::WaitTimeoutResult;
16
17 use crate::prelude::*;
18
19 use super::{LockTestExt, LockHeldState};
20
21 #[cfg(feature = "backtrace")]
22 use {crate::prelude::hash_map, backtrace::Backtrace, std::sync::Once};
23
24 #[cfg(not(feature = "backtrace"))]
25 struct Backtrace{}
26 #[cfg(not(feature = "backtrace"))]
27 impl Backtrace { fn new() -> Backtrace { Backtrace {} } }
28
29 pub type LockResult<Guard> = Result<Guard, ()>;
30
31 pub struct Condvar {
32         inner: StdCondvar,
33 }
34
35 impl Condvar {
36         pub fn new() -> Condvar {
37                 Condvar { inner: StdCondvar::new() }
38         }
39
40         pub fn wait_while<'a, T, F: FnMut(&mut T) -> bool>(&'a self, guard: MutexGuard<'a, T>, condition: F)
41         -> LockResult<MutexGuard<'a, T>> {
42                 let mutex: &'a Mutex<T> = guard.mutex;
43                 self.inner.wait_while(guard.into_inner(), condition).map(|lock| MutexGuard { mutex, lock })
44                         .map_err(|_| ())
45         }
46
47         #[allow(unused)]
48         pub fn wait_timeout_while<'a, T, F: FnMut(&mut T) -> bool>(&'a self, guard: MutexGuard<'a, T>, dur: Duration, condition: F)
49         -> LockResult<(MutexGuard<'a, T>, WaitTimeoutResult)> {
50                 let mutex = guard.mutex;
51                 self.inner.wait_timeout_while(guard.into_inner(), dur, condition).map_err(|_| ())
52                         .map(|(lock, e)| (MutexGuard { mutex, lock }, e))
53         }
54
55         pub fn notify_all(&self) { self.inner.notify_all(); }
56 }
57
58 thread_local! {
59         /// We track the set of locks currently held by a reference to their `LockMetadata`
60         static LOCKS_HELD: RefCell<HashMap<u64, Arc<LockMetadata>>> = RefCell::new(new_hash_map());
61 }
62 static LOCK_IDX: AtomicUsize = AtomicUsize::new(0);
63
64 #[cfg(feature = "backtrace")]
65 static mut LOCKS: Option<StdMutex<HashMap<String, Arc<LockMetadata>>>> = None;
66 #[cfg(feature = "backtrace")]
67 static LOCKS_INIT: Once = Once::new();
68
69 /// Metadata about a single lock, by id, the set of things locked-before it, and the backtrace of
70 /// when the Mutex itself was constructed.
71 struct LockMetadata {
72         lock_idx: u64,
73         locked_before: StdMutex<HashMap<u64, LockDep>>,
74         _lock_construction_bt: Backtrace,
75 }
76
77 struct LockDep {
78         lock: Arc<LockMetadata>,
79         /// lockdep_trace is unused unless we're building with `backtrace`, so we mark it _
80         _lockdep_trace: Backtrace,
81 }
82
83 // Locates the frame preceding the earliest `debug_sync` frame in the call stack. This ensures we
84 // can properly detect a lock's construction and acquiral callsites, since the latter may contain
85 // multiple `debug_sync` frames.
86 #[cfg(feature = "backtrace")]
87 fn locate_call_symbol(backtrace: &Backtrace) -> (String, Option<u32>) {
88         // Find the earliest `debug_sync` frame (or that is in our tests) and use the frame preceding it
89         // as the callsite. Note that the first few frames may be in the `backtrace` crate, so we have
90         // to ignore those.
91         let sync_mutex_constr_regex = regex::Regex::new(r"lightning.*debug_sync").unwrap();
92         let mut found_debug_sync = false;
93         let mut symbol_after_latest_debug_sync = None;
94         for frame in backtrace.frames().iter() {
95                 for symbol in frame.symbols().iter() {
96                         if let Some(symbol_name) = symbol.name().map(|name| name.as_str()).flatten() {
97                                 if !sync_mutex_constr_regex.is_match(symbol_name) {
98                                         if found_debug_sync {
99                                                 symbol_after_latest_debug_sync = Some(symbol);
100                                                 found_debug_sync = false;
101                                         }
102                                 } else { found_debug_sync = true; }
103                         }
104                 }
105         }
106         let symbol = symbol_after_latest_debug_sync.unwrap_or_else(|| {
107                 panic!("Couldn't find lock call symbol in trace {:?}", backtrace);
108         });
109         (format!("{}:{}", symbol.filename().unwrap().display(), symbol.lineno().unwrap()), symbol.colno())
110 }
111
112 impl LockMetadata {
113         fn new() -> Arc<LockMetadata> {
114                 let backtrace = Backtrace::new();
115                 let lock_idx = LOCK_IDX.fetch_add(1, Ordering::Relaxed) as u64;
116
117                 let res = Arc::new(LockMetadata {
118                         locked_before: StdMutex::new(new_hash_map()),
119                         lock_idx,
120                         _lock_construction_bt: backtrace,
121                 });
122
123                 #[cfg(feature = "backtrace")]
124                 {
125                         let (lock_constr_location, lock_constr_colno) =
126                                 locate_call_symbol(&res._lock_construction_bt);
127                         LOCKS_INIT.call_once(|| { unsafe { LOCKS = Some(StdMutex::new(new_hash_map())); } });
128                         let mut locks = unsafe { LOCKS.as_ref() }.unwrap().lock().unwrap();
129                         match locks.entry(lock_constr_location) {
130                                 hash_map::Entry::Occupied(e) => {
131                                         assert_eq!(lock_constr_colno,
132                                                 locate_call_symbol(&e.get()._lock_construction_bt).1,
133                                                 "Because Windows doesn't support column number results in backtraces, we cannot construct two mutexes on the same line or we risk lockorder detection false positives.");
134                                         return Arc::clone(e.get())
135                                 },
136                                 hash_map::Entry::Vacant(e) => { e.insert(Arc::clone(&res)); },
137                         }
138                 }
139                 res
140         }
141
142         fn pre_lock(this: &Arc<LockMetadata>, _double_lock_self_allowed: bool) {
143                 LOCKS_HELD.with(|held| {
144                         // For each lock that is currently held, check that no lock's `locked_before` set
145                         // includes the lock we're about to hold, which would imply a lockorder inversion.
146                         for (locked_idx, _locked) in held.borrow().iter() {
147                                 if *locked_idx == this.lock_idx {
148                                         // Note that with `feature = "backtrace"` set, we may be looking at different
149                                         // instances of the same lock. Still, doing so is quite risky, a total order
150                                         // must be maintained, and doing so across a set of otherwise-identical mutexes
151                                         // is fraught with issues.
152                                         #[cfg(feature = "backtrace")]
153                                         debug_assert!(_double_lock_self_allowed,
154                                                 "Tried to acquire a lock while it was held!\nLock constructed at {}",
155                                                 locate_call_symbol(&this._lock_construction_bt).0);
156                                         #[cfg(not(feature = "backtrace"))]
157                                         panic!("Tried to acquire a lock while it was held!");
158                                 }
159                         }
160                         for (_locked_idx, locked) in held.borrow().iter() {
161                                 for (locked_dep_idx, _locked_dep) in locked.locked_before.lock().unwrap().iter() {
162                                         let is_dep_this_lock = *locked_dep_idx == this.lock_idx;
163                                         let has_same_construction = *locked_dep_idx == locked.lock_idx;
164                                         if is_dep_this_lock && !has_same_construction {
165                                                 #[allow(unused_mut, unused_assignments)]
166                                                 let mut has_same_callsite = false;
167                                                 #[cfg(feature = "backtrace")] {
168                                                         has_same_callsite = _double_lock_self_allowed &&
169                                                                 locate_call_symbol(&_locked_dep._lockdep_trace) ==
170                                                                         locate_call_symbol(&Backtrace::new());
171                                                 }
172                                                 if !has_same_callsite {
173                                                         #[cfg(feature = "backtrace")]
174                                                         panic!("Tried to violate existing lockorder.\nMutex that should be locked after the current lock was created at the following backtrace.\nNote that to get a backtrace for the lockorder violation, you should set RUST_BACKTRACE=1\nLock being taken constructed at: {} ({}):\n{:?}\nLock constructed at: {} ({})\n{:?}\n\nLock dep created at:\n{:?}\n\n",
175                                                                 locate_call_symbol(&this._lock_construction_bt).0,
176                                                                 this.lock_idx, this._lock_construction_bt,
177                                                                 locate_call_symbol(&locked._lock_construction_bt).0,
178                                                                 locked.lock_idx, locked._lock_construction_bt,
179                                                                 _locked_dep._lockdep_trace);
180                                                         #[cfg(not(feature = "backtrace"))]
181                                                         panic!("Tried to violate existing lockorder. Build with the backtrace feature for more info.");
182                                                 }
183                                         }
184                                 }
185                                 // Insert any already-held locks in our locked-before set.
186                                 let mut locked_before = this.locked_before.lock().unwrap();
187                                 if !locked_before.contains_key(&locked.lock_idx) {
188                                         let lockdep = LockDep { lock: Arc::clone(locked), _lockdep_trace: Backtrace::new() };
189                                         locked_before.insert(lockdep.lock.lock_idx, lockdep);
190                                 }
191                         }
192                         held.borrow_mut().insert(this.lock_idx, Arc::clone(this));
193                 });
194         }
195
196         fn held_by_thread(this: &Arc<LockMetadata>) -> LockHeldState {
197                 let mut res = LockHeldState::NotHeldByThread;
198                 LOCKS_HELD.with(|held| {
199                         for (locked_idx, _locked) in held.borrow().iter() {
200                                 if *locked_idx == this.lock_idx {
201                                         res = LockHeldState::HeldByThread;
202                                 }
203                         }
204                 });
205                 res
206         }
207
208         fn try_locked(this: &Arc<LockMetadata>) {
209                 LOCKS_HELD.with(|held| {
210                         // Since a try-lock will simply fail if the lock is held already, we do not
211                         // consider try-locks to ever generate lockorder inversions. However, if a try-lock
212                         // succeeds, we do consider it to have created lockorder dependencies.
213                         let mut locked_before = this.locked_before.lock().unwrap();
214                         for (locked_idx, locked) in held.borrow().iter() {
215                                 if !locked_before.contains_key(locked_idx) {
216                                         let lockdep = LockDep { lock: Arc::clone(locked), _lockdep_trace: Backtrace::new() };
217                                         locked_before.insert(*locked_idx, lockdep);
218                                 }
219                         }
220                         held.borrow_mut().insert(this.lock_idx, Arc::clone(this));
221                 });
222         }
223 }
224
225 pub struct Mutex<T: Sized> {
226         inner: StdMutex<T>,
227         deps: Arc<LockMetadata>,
228 }
229 impl<T: Sized> Mutex<T> {
230         pub(crate) fn into_inner(self) -> LockResult<T> {
231                 self.inner.into_inner().map_err(|_| ())
232         }
233 }
234
235 #[must_use = "if unused the Mutex will immediately unlock"]
236 pub struct MutexGuard<'a, T: Sized + 'a> {
237         mutex: &'a Mutex<T>,
238         lock: StdMutexGuard<'a, T>,
239 }
240
241 impl<'a, T: Sized> MutexGuard<'a, T> {
242         fn into_inner(self) -> StdMutexGuard<'a, T> {
243                 // Somewhat unclear why we cannot move out of self.lock, but doing so gets E0509.
244                 unsafe {
245                         let v: StdMutexGuard<'a, T> = std::ptr::read(&self.lock);
246                         std::mem::forget(self);
247                         v
248                 }
249         }
250 }
251
252 impl<T: Sized> Drop for MutexGuard<'_, T> {
253         fn drop(&mut self) {
254                 LOCKS_HELD.with(|held| {
255                         held.borrow_mut().remove(&self.mutex.deps.lock_idx);
256                 });
257         }
258 }
259
260 impl<T: Sized> Deref for MutexGuard<'_, T> {
261         type Target = T;
262
263         fn deref(&self) -> &T {
264                 &self.lock.deref()
265         }
266 }
267
268 impl<T: Sized> DerefMut for MutexGuard<'_, T> {
269         fn deref_mut(&mut self) -> &mut T {
270                 self.lock.deref_mut()
271         }
272 }
273
274 impl<T> Mutex<T> {
275         pub fn new(inner: T) -> Mutex<T> {
276                 Mutex { inner: StdMutex::new(inner), deps: LockMetadata::new() }
277         }
278
279         pub fn lock<'a>(&'a self) -> LockResult<MutexGuard<'a, T>> {
280                 LockMetadata::pre_lock(&self.deps, false);
281                 self.inner.lock().map(|lock| MutexGuard { mutex: self, lock }).map_err(|_| ())
282         }
283
284         pub fn try_lock<'a>(&'a self) -> LockResult<MutexGuard<'a, T>> {
285                 let res = self.inner.try_lock().map(|lock| MutexGuard { mutex: self, lock }).map_err(|_| ());
286                 if res.is_ok() {
287                         LockMetadata::try_locked(&self.deps);
288                 }
289                 res
290         }
291 }
292
293 impl<'a, T: 'a> LockTestExt<'a> for Mutex<T> {
294         #[inline]
295         fn held_by_thread(&self) -> LockHeldState {
296                 LockMetadata::held_by_thread(&self.deps)
297         }
298         type ExclLock = MutexGuard<'a, T>;
299         #[inline]
300         fn unsafe_well_ordered_double_lock_self(&'a self) -> MutexGuard<T> {
301                 LockMetadata::pre_lock(&self.deps, true);
302                 self.inner.lock().map(|lock| MutexGuard { mutex: self, lock }).unwrap()
303         }
304 }
305
306 pub struct RwLock<T: Sized> {
307         inner: StdRwLock<T>,
308         deps: Arc<LockMetadata>,
309 }
310
311 pub struct RwLockReadGuard<'a, T: Sized + 'a> {
312         lock: &'a RwLock<T>,
313         guard: StdRwLockReadGuard<'a, T>,
314 }
315
316 pub struct RwLockWriteGuard<'a, T: Sized + 'a> {
317         lock: &'a RwLock<T>,
318         guard: StdRwLockWriteGuard<'a, T>,
319 }
320
321 impl<T: Sized> Deref for RwLockReadGuard<'_, T> {
322         type Target = T;
323
324         fn deref(&self) -> &T {
325                 &self.guard.deref()
326         }
327 }
328
329 impl<T: Sized> Drop for RwLockReadGuard<'_, T> {
330         fn drop(&mut self) {
331                 LOCKS_HELD.with(|held| {
332                         held.borrow_mut().remove(&self.lock.deps.lock_idx);
333                 });
334         }
335 }
336
337 impl<T: Sized> Deref for RwLockWriteGuard<'_, T> {
338         type Target = T;
339
340         fn deref(&self) -> &T {
341                 &self.guard.deref()
342         }
343 }
344
345 impl<T: Sized> Drop for RwLockWriteGuard<'_, T> {
346         fn drop(&mut self) {
347                 LOCKS_HELD.with(|held| {
348                         held.borrow_mut().remove(&self.lock.deps.lock_idx);
349                 });
350         }
351 }
352
353 impl<T: Sized> DerefMut for RwLockWriteGuard<'_, T> {
354         fn deref_mut(&mut self) -> &mut T {
355                 self.guard.deref_mut()
356         }
357 }
358
359 impl<T> RwLock<T> {
360         pub fn new(inner: T) -> RwLock<T> {
361                 RwLock { inner: StdRwLock::new(inner), deps: LockMetadata::new() }
362         }
363
364         pub fn read<'a>(&'a self) -> LockResult<RwLockReadGuard<'a, T>> {
365                 // Note that while we could be taking a recursive read lock here, Rust's `RwLock` may
366                 // deadlock trying to take a second read lock if another thread is waiting on the write
367                 // lock. This behavior is platform dependent, but our in-tree `FairRwLock` guarantees
368                 // such a deadlock.
369                 LockMetadata::pre_lock(&self.deps, false);
370                 self.inner.read().map(|guard| RwLockReadGuard { lock: self, guard }).map_err(|_| ())
371         }
372
373         pub fn write<'a>(&'a self) -> LockResult<RwLockWriteGuard<'a, T>> {
374                 LockMetadata::pre_lock(&self.deps, false);
375                 self.inner.write().map(|guard| RwLockWriteGuard { lock: self, guard }).map_err(|_| ())
376         }
377
378         pub fn try_write<'a>(&'a self) -> LockResult<RwLockWriteGuard<'a, T>> {
379                 let res = self.inner.try_write().map(|guard| RwLockWriteGuard { lock: self, guard }).map_err(|_| ());
380                 if res.is_ok() {
381                         LockMetadata::try_locked(&self.deps);
382                 }
383                 res
384         }
385 }
386
387 impl<'a, T: 'a> LockTestExt<'a> for RwLock<T> {
388         #[inline]
389         fn held_by_thread(&self) -> LockHeldState {
390                 LockMetadata::held_by_thread(&self.deps)
391         }
392         type ExclLock = RwLockWriteGuard<'a, T>;
393         #[inline]
394         fn unsafe_well_ordered_double_lock_self(&'a self) -> RwLockWriteGuard<'a, T> {
395                 LockMetadata::pre_lock(&self.deps, true);
396                 self.inner.write().map(|guard| RwLockWriteGuard { lock: self, guard }).unwrap()
397         }
398 }
399
400 pub type FairRwLock<T> = RwLock<T>;