//! Implement thread-local storage.
use std::collections::BTreeMap;
-use std::collections::btree_map::Entry;
-use std::collections::HashSet;
+use std::collections::btree_map::Entry as BTreeEntry;
+use std::collections::hash_map::Entry as HashMapEntry;
use log::trace;
-use rustc_index::vec::Idx;
+use rustc_data_structures::fx::FxHashMap;
use rustc_middle::ty;
use rustc_target::abi::{Size, HasDataLayout};
+use rustc_target::spec::abi::Abi;
-use crate::{
- HelpersEvalContextExt, InterpResult, MPlaceTy, Scalar, StackPopCleanup, Tag, ThreadId,
- ThreadsEvalContextExt,
-};
+use crate::*;
pub type TlsKey = u128;
pub struct TlsEntry<'tcx> {
/// The data for this key. None is used to represent NULL.
/// (We normalize this early to avoid having to do a NULL-ptr-test each time we access the data.)
- /// Will eventually become a map from thread IDs to `Scalar`s, if we ever support more than one thread.
data: BTreeMap<ThreadId, Scalar<Tag>>,
dtor: Option<ty::Instance<'tcx>>,
}
+#[derive(Clone, Debug)]
+struct RunningDtorsState {
+ /// The last TlsKey used to retrieve a TLS destructor. `None` means that we
+ /// have not tried to retrieve a TLS destructor yet or that we already tried
+ /// all keys.
+ last_dtor_key: Option<TlsKey>,
+}
+
#[derive(Debug)]
pub struct TlsData<'tcx> {
/// The Key to use for the next thread-local allocation.
/// pthreads-style thread-local storage.
keys: BTreeMap<TlsKey, TlsEntry<'tcx>>,
- /// A single global per thread dtor (that's how things work on macOS) with a data argument.
- global_dtors: BTreeMap<ThreadId, (ty::Instance<'tcx>, Scalar<Tag>)>,
+ /// A single per thread destructor of the thread local storage (that's how
+ /// things work on macOS) with a data argument.
+ macos_thread_dtors: BTreeMap<ThreadId, (ty::Instance<'tcx>, Scalar<Tag>)>,
- /// Whether we are in the "destruct" phase, during which some operations are UB.
- dtors_running: HashSet<ThreadId>,
+ /// State for currently running TLS dtors. If this map contains a key for a
+ /// specific thread, it means that we are in the "destruct" phase, during
+ /// which some operations are UB.
+ dtors_running: FxHashMap<ThreadId, RunningDtorsState>,
}
impl<'tcx> Default for TlsData<'tcx> {
TlsData {
next_key: 1, // start with 1 as we must not use 0 on Windows
keys: Default::default(),
- global_dtors: Default::default(),
+ macos_thread_dtors: Default::default(),
dtors_running: Default::default(),
}
}
pub fn create_tls_key(&mut self, dtor: Option<ty::Instance<'tcx>>, max_size: Size) -> InterpResult<'tcx, TlsKey> {
let new_key = self.next_key;
self.next_key += 1;
- self.keys.insert(new_key, TlsEntry { data: Default::default(), dtor }).unwrap_none();
+ self.keys.try_insert(new_key, TlsEntry { data: Default::default(), dtor }).unwrap();
trace!("New TLS key allocated: {} with dtor {:?}", new_key, dtor);
if max_size.bits() < 128 && new_key >= (1u128 << max_size.bits() as u128) {
) -> InterpResult<'tcx, Scalar<Tag>> {
match self.keys.get(&key) {
Some(TlsEntry { data, .. }) => {
- let value = data.get(&thread_id).cloned();
+ let value = data.get(&thread_id).copied();
trace!("TLS key {} for thread {:?} loaded: {:?}", key, thread_id, value);
Ok(value.unwrap_or_else(|| Scalar::null_ptr(cx).into()))
}
pub fn store_tls(
&mut self,
- key: TlsKey, thread_id: ThreadId, new_data: Option<Scalar<Tag>>) -> InterpResult<'tcx> {
+ key: TlsKey,
+ thread_id: ThreadId,
+ new_data: Option<Scalar<Tag>>
+ ) -> InterpResult<'tcx> {
match self.keys.get_mut(&key) {
Some(TlsEntry { data, .. }) => {
match new_data {
- Some(ptr) => {
- trace!("TLS key {} for thread {:?} stored: {:?}", key, thread_id, ptr);
- data.insert(thread_id, ptr);
+ Some(scalar) => {
+ trace!("TLS key {} for thread {:?} stored: {:?}", key, thread_id, scalar);
+ data.insert(thread_id, scalar);
}
None => {
trace!("TLS key {} for thread {:?} removed", key, thread_id);
}
}
- /// Set global dtor for the given thread.
- pub fn set_global_dtor(&mut self, thread: ThreadId, dtor: ty::Instance<'tcx>, data: Scalar<Tag>) -> InterpResult<'tcx> {
- if self.dtors_running.contains(&thread) {
+ /// Set the thread wide destructor of the thread local storage for the given
+ /// thread. This function is used to implement `_tlv_atexit` shim on MacOS.
+ ///
+ /// Thread wide dtors are available only on MacOS. There is one destructor
+ /// per thread as can be guessed from the following comment in the
+ /// [`_tlv_atexit`
+ /// implementation](https://github.com/opensource-apple/dyld/blob/195030646877261f0c8c7ad8b001f52d6a26f514/src/threadLocalVariables.c#L389):
+ ///
+ /// // NOTE: this does not need locks because it only operates on current thread data
+ pub fn set_macos_thread_dtor(
+ &mut self,
+ thread: ThreadId,
+ dtor: ty::Instance<'tcx>,
+ data: Scalar<Tag>
+ ) -> InterpResult<'tcx> {
+ if self.dtors_running.contains_key(&thread) {
// UB, according to libstd docs.
- throw_ub_format!("setting global destructor while destructors are already running");
+ throw_ub_format!("setting thread's local storage destructor while destructors are already running");
}
- if self.global_dtors.insert(thread, (dtor, data)).is_some() {
- throw_unsup_format!("setting more than one global destructor for the same thread is not supported");
+ if self.macos_thread_dtors.insert(thread, (dtor, data)).is_some() {
+ throw_unsup_format!("setting more than one thread local storage destructor for the same thread is not supported");
}
Ok(())
}
key: Option<TlsKey>,
thread_id: ThreadId,
) -> Option<(ty::Instance<'tcx>, Scalar<Tag>, TlsKey)> {
- use std::collections::Bound::*;
+ use std::ops::Bound::*;
let thread_local = &mut self.keys;
let start = match key {
thread_local.range_mut((start, Unbounded))
{
match data.entry(thread_id) {
- Entry::Occupied(entry) => {
- let data_scalar = entry.remove();
+ BTreeEntry::Occupied(entry) => {
if let Some(dtor) = dtor {
+ // Set TLS data to NULL, and call dtor with old value.
+ let data_scalar = entry.remove();
let ret = Some((*dtor, data_scalar, key));
return ret;
}
}
- Entry::Vacant(_) => {}
+ BTreeEntry::Vacant(_) => {}
}
}
None
}
-}
-impl<'mir, 'tcx: 'mir> EvalContextExt<'mir, 'tcx> for crate::MiriEvalContext<'mir, 'tcx> {}
-pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriEvalContextExt<'mir, 'tcx> {
+ /// Set that dtors are running for `thread`. It is guaranteed not to change
+ /// the existing values stored in `dtors_running` for this thread. Returns
+ /// `true` if dtors for `thread` are already running.
+ fn set_dtors_running_for_thread(&mut self, thread: ThreadId) -> bool {
+ match self.dtors_running.entry(thread) {
+ HashMapEntry::Occupied(_) => true,
+ HashMapEntry::Vacant(entry) => {
+ // We cannot just do `self.dtors_running.insert` because that
+ // would overwrite `last_dtor_key` with `None`.
+ entry.insert(RunningDtorsState { last_dtor_key: None });
+ false
+ }
+ }
+ }
- /// Run TLS destructors for the main thread on Windows. The implementation
- /// assumes that we do not support concurrency on Windows yet.
- ///
- /// Note: on non-Windows OS this function is a no-op.
- fn run_windows_tls_dtors(&mut self) -> InterpResult<'tcx> {
- let this = self.eval_context_mut();
- if this.tcx.sess.target.target.target_os != "windows" {
- return Ok(());
+ /// Delete all TLS entries for the given thread. This function should be
+ /// called after all TLS destructors have already finished.
+ fn delete_all_thread_tls(&mut self, thread_id: ThreadId) {
+ for TlsEntry { data, .. } in self.keys.values_mut() {
+ data.remove(&thread_id);
}
- let active_thread = this.get_active_thread()?;
- assert_eq!(active_thread.index(), 0, "concurrency on Windows not supported");
- assert!(!this.machine.tls.dtors_running.contains(&active_thread), "running TLS dtors twice");
- this.machine.tls.dtors_running.insert(active_thread);
+ }
+}
+
+impl<'mir, 'tcx: 'mir> EvalContextPrivExt<'mir, 'tcx> for crate::MiriEvalContext<'mir, 'tcx> {}
+trait EvalContextPrivExt<'mir, 'tcx: 'mir>: crate::MiriEvalContextExt<'mir, 'tcx> {
+ /// Schedule TLS destructors for the main thread on Windows. The
+ /// implementation assumes that we do not support concurrency on Windows
+ /// yet.
+ fn schedule_windows_tls_dtors(&mut self) -> InterpResult<'tcx> {
+ let this = self.eval_context_mut();
+ let active_thread = this.get_active_thread();
+ assert_eq!(this.get_total_thread_count(), 1, "concurrency on Windows is not supported");
// Windows has a special magic linker section that is run on certain events.
// Instead of searching for that section and supporting arbitrary hooks in there
// (that would be basically https://github.com/rust-lang/miri/issues/450),
// we specifically look up the static in libstd that we know is placed
// in that section.
- let thread_callback = this.eval_path_scalar(&["std", "sys", "windows", "thread_local", "p_thread_callback"])?;
- let thread_callback = this.memory.get_fn(thread_callback.not_undef()?)?.as_instance()?;
+ let thread_callback = this.eval_path_scalar(&["std", "sys", "windows", "thread_local_key", "p_thread_callback"])?;
+ let thread_callback = this.memory.get_fn(thread_callback.check_init()?)?.as_instance()?;
// The signature of this function is `unsafe extern "system" fn(h: c::LPVOID, dwReason: c::DWORD, pv: c::LPVOID)`.
- let reason = this.eval_path_scalar(&["std", "sys", "windows", "c", "DLL_PROCESS_DETACH"])?;
+ let reason = this.eval_path_scalar(&["std", "sys", "windows", "c", "DLL_THREAD_DETACH"])?;
let ret_place = MPlaceTy::dangling(this.machine.layouts.unit, this).into();
this.call_function(
thread_callback,
+ Abi::System { unwind: false },
&[Scalar::null_ptr(this).into(), reason.into(), Scalar::null_ptr(this).into()],
- Some(ret_place),
+ Some(&ret_place),
StackPopCleanup::None { cleanup: true },
)?;
- // step until out of stackframes
- this.run()?;
-
- // Windows doesn't have other destructors.
+ this.enable_thread(active_thread);
Ok(())
}
- /// Run TLS destructors for the active thread.
+ /// Schedule the MacOS thread destructor of the thread local storage to be
+ /// executed. Returns `true` if scheduled.
///
- /// Note: on Windows OS this function is a no-op because we do not support
- /// concurrency on Windows yet.
- fn run_tls_dtors_for_active_thread(&mut self) -> InterpResult<'tcx> {
+ /// Note: It is safe to call this function also on other Unixes.
+ fn schedule_macos_tls_dtor(&mut self) -> InterpResult<'tcx, bool> {
let this = self.eval_context_mut();
- if this.tcx.sess.target.target.target_os == "windows" {
- return Ok(());
- }
- let thread_id = this.get_active_thread()?;
- assert!(!this.machine.tls.dtors_running.contains(&thread_id), "running TLS dtors twice");
- this.machine.tls.dtors_running.insert(thread_id);
-
- // The macOS global dtor runs "before any TLS slots get freed", so do that first.
- if let Some(&(instance, data)) = this.machine.tls.global_dtors.get(&thread_id) {
- trace!("Running global dtor {:?} on {:?} at {:?}", instance, data, thread_id);
+ let thread_id = this.get_active_thread();
+ if let Some((instance, data)) = this.machine.tls.macos_thread_dtors.remove(&thread_id) {
+ trace!("Running macos dtor {:?} on {:?} at {:?}", instance, data, thread_id);
let ret_place = MPlaceTy::dangling(this.machine.layouts.unit, this).into();
this.call_function(
instance,
+ Abi::C { unwind: false },
&[data.into()],
- Some(ret_place),
+ Some(&ret_place),
StackPopCleanup::None { cleanup: true },
)?;
- // step until out of stackframes
- this.run()?;
+ // Enable the thread so that it steps through the destructor which
+ // we just scheduled. Since we deleted the destructor, it is
+ // guaranteed that we will schedule it again. The `dtors_running`
+ // flag will prevent the code from adding the destructor again.
+ this.enable_thread(thread_id);
+ Ok(true)
+ } else {
+ Ok(false)
}
+ }
- assert!(this.has_terminated(thread_id)?, "running TLS dtors for non-terminated thread");
- let mut dtor = this.machine.tls.fetch_tls_dtor(None, thread_id);
- while let Some((instance, ptr, key)) = dtor {
- trace!("Running TLS dtor {:?} on {:?} at {:?}", instance, ptr, thread_id);
- assert!(!this.is_null(ptr).unwrap(), "Data can't be NULL when dtor is called!");
+ /// Schedule a pthread TLS destructor. Returns `true` if found
+ /// a destructor to schedule, and `false` otherwise.
+ fn schedule_next_pthread_tls_dtor(&mut self) -> InterpResult<'tcx, bool> {
+ let this = self.eval_context_mut();
+ let active_thread = this.get_active_thread();
+
+ assert!(this.has_terminated(active_thread), "running TLS dtors for non-terminated thread");
+ // Fetch next dtor after `key`.
+ let last_key = this.machine.tls.dtors_running[&active_thread].last_dtor_key.clone();
+ let dtor = match this.machine.tls.fetch_tls_dtor(last_key, active_thread) {
+ dtor @ Some(_) => dtor,
+ // We ran each dtor once, start over from the beginning.
+ None => {
+ this.machine.tls.fetch_tls_dtor(None, active_thread)
+ }
+ };
+ if let Some((instance, ptr, key)) = dtor {
+ this.machine.tls.dtors_running.get_mut(&active_thread).unwrap().last_dtor_key = Some(key);
+ trace!("Running TLS dtor {:?} on {:?} at {:?}", instance, ptr, active_thread);
+ assert!(!this.is_null(ptr).unwrap(), "data can't be NULL when dtor is called!");
let ret_place = MPlaceTy::dangling(this.machine.layouts.unit, this).into();
this.call_function(
instance,
+ Abi::C { unwind: false },
&[ptr.into()],
- Some(ret_place),
+ Some(&ret_place),
StackPopCleanup::None { cleanup: true },
)?;
- // step until out of stackframes
- this.run()?;
+ this.enable_thread(active_thread);
+ return Ok(true);
+ }
+ this.machine.tls.dtors_running.get_mut(&active_thread).unwrap().last_dtor_key = None;
+
+ Ok(false)
+ }
+}
+
+impl<'mir, 'tcx: 'mir> EvalContextExt<'mir, 'tcx> for crate::MiriEvalContext<'mir, 'tcx> {}
+pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriEvalContextExt<'mir, 'tcx> {
- // Fetch next dtor after `key`.
- dtor = match this.machine.tls.fetch_tls_dtor(Some(key), thread_id) {
- dtor @ Some(_) => dtor,
- // We ran each dtor once, start over from the beginning.
- None => this.machine.tls.fetch_tls_dtor(None, thread_id),
- };
+ /// Schedule an active thread's TLS destructor to run on the active thread.
+ /// Note that this function does not run the destructors itself, it just
+ /// schedules them one by one each time it is called and reenables the
+ /// thread so that it can be executed normally by the main execution loop.
+ ///
+ /// Note: we consistently run TLS destructors for all threads, including the
+ /// main thread. However, it is not clear that we should run the TLS
+ /// destructors for the main thread. See issue:
+ /// https://github.com/rust-lang/rust/issues/28129.
+ fn schedule_next_tls_dtor_for_active_thread(&mut self) -> InterpResult<'tcx> {
+ let this = self.eval_context_mut();
+ let active_thread = this.get_active_thread();
+
+ if !this.machine.tls.set_dtors_running_for_thread(active_thread) {
+ // This is the first time we got asked to schedule a destructor. The
+ // Windows schedule destructor function must be called exactly once,
+ // this is why it is in this block.
+ if this.tcx.sess.target.os == "windows" {
+ // On Windows, we signal that the thread quit by starting the
+ // relevant function, reenabling the thread, and going back to
+ // the scheduler.
+ this.schedule_windows_tls_dtors()?;
+ return Ok(())
+ }
+ }
+ // The remaining dtors make some progress each time around the scheduler loop,
+ // until they return `false` to indicate that they are done.
+
+ // The macOS thread wide destructor runs "before any TLS slots get
+ // freed", so do that first.
+ if this.schedule_macos_tls_dtor()? {
+ // We have scheduled a MacOS dtor to run on the thread. Execute it
+ // to completion and come back here. Scheduling a destructor
+ // destroys it, so we will not enter this branch again.
+ return Ok(())
}
+ if this.schedule_next_pthread_tls_dtor()? {
+ // We have scheduled a pthread destructor and removed it from the
+ // destructors list. Run it to completion and come back here.
+ return Ok(())
+ }
+
+ // All dtors done!
+ this.machine.tls.delete_all_thread_tls(active_thread);
+ this.thread_terminated()?;
Ok(())
}