Auto merge of #2699 - RalfJung:schedule-refactor, r=RalfJung

refactor scheduler

Refactors the scheduler to use something akin to a generator -- a callback that will be invoked when the stack of a thread is empty, which has the chance to push a new stack frame or do other things and then indicates whether this thread is done, or should be scheduled again. (Unfortunately I think we [cannot use actual generators](https://rust-lang.zulipchat.com/#narrow/stream/213817-t-lang/topic/Generators.20that.20borrow.20on.20each.20resume.3F) here.) The interpreter loop is now a proper infinite loop, the only way to leave it is for some kind of interrupt to be triggered (represented as `InterpError`) -- unifying how we handle 'exit when calling `process::exit`' and 'exit when main thread quits'.

The last commit implements an alternative approach to https://github.com/rust-lang/miri/pull/2660 using this new structure. Fixes https://github.com/rust-lang/miri/issues/2629.

15 files changed, 412 insertions, 327 deletions

diff --git a/src/tools/miri/cargo-miri/src/setup.rs b/src/tools/miri/cargo-miri/src/setup.rs
index 9c179e82ba1..3ec63ba0f10 100644
--- a/src/tools/miri/cargo-miri/src/setup.rs
+++ b/src/tools/miri/cargo-miri/src/setup.rs
@@ -99,12 +99,10 @@ pub fn setup(subcommand: &MiriCommand, target: &str, rustc_version: &VersionMeta
         // `config.toml`.
         command.env("RUSTC_WRAPPER", "");
 
-        if only_setup {
-            if print_sysroot {
-                // Be extra sure there is no noise on stdout.
-                command.stdout(process::Stdio::null());
-            }
+        if only_setup && !print_sysroot {
+            // Forward output.
         } else {
+            // Supress output.
             command.stdout(process::Stdio::null());
             command.stderr(process::Stdio::null());
         }
@@ -120,7 +118,9 @@ pub fn setup(subcommand: &MiriCommand, target: &str, rustc_version: &VersionMeta
     std::env::set_var("MIRI_SYSROOT", &sysroot_dir);
 
     // Do the build.
-    if only_setup {
+    if print_sysroot {
+        // Be silent.
+    } else if only_setup {
         // We want to be explicit.
         eprintln!("Preparing a sysroot for Miri (target: {target})...");
     } else {
@@ -143,7 +143,9 @@ pub fn setup(subcommand: &MiriCommand, target: &str, rustc_version: &VersionMeta
                 )
             }
         });
-    if only_setup {
+    if print_sysroot {
+        // Be silent.
+    } else if only_setup {
         eprintln!("A sysroot for Miri is now available in `{}`.", sysroot_dir.display());
     } else {
         eprintln!("done");
diff --git a/src/tools/miri/ci.sh b/src/tools/miri/ci.sh
index dd2d2abe35b..e455b482338 100755
--- a/src/tools/miri/ci.sh
+++ b/src/tools/miri/ci.sh
@@ -40,10 +40,15 @@ function run_tests {
   ./miri test
   if [ -z "${MIRI_TEST_TARGET+exists}" ]; then
     # Only for host architecture: tests with optimizations (`-O` is what cargo passes, but crank MIR
-    # optimizations up all the way).
-    # Optimizations change diagnostics (mostly backtraces), so we don't check them
-    #FIXME(#2155): we want to only run the pass and panic tests here, not the fail tests.
+    # optimizations up all the way, too).
+    # Optimizations change diagnostics (mostly backtraces), so we don't check
+    # them. Also error locations change so we don't run the failing tests.
     MIRIFLAGS="${MIRIFLAGS:-} -O -Zmir-opt-level=4" MIRI_SKIP_UI_CHECKS=1 ./miri test -- tests/{pass,panic}
+
+    # Also run some many-seeds tests. 64 seeds means this takes around a minute per test.
+    for FILE in tests/many-seeds/*.rs; do
+      MIRI_SEEDS=64 CARGO_EXTRA_FLAGS="$CARGO_EXTRA_FLAGS -q" ./miri many-seeds ./miri run "$FILE"
+    done
   fi
 
   ## test-cargo-miri
diff --git a/src/tools/miri/miri b/src/tools/miri/miri
index 38d36898768..a259576ed42 100755
--- a/src/tools/miri/miri
+++ b/src/tools/miri/miri
@@ -36,7 +36,8 @@ Mainly meant to be invoked by rust-analyzer.
 ./miri many-seeds <command>:
 Runs <command> over and over again with different seeds for Miri. The MIRIFLAGS
 variable is set to its original value appended with ` -Zmiri-seed=$SEED` for
-many different seeds.
+many different seeds. The MIRI_SEEDS variable controls how many seeds are being
+tried; MIRI_SEED_START controls the first seed to try.
 
 ./miri bench <benches>:
 Runs the benchmarks from bench-cargo-miri in hyperfine. hyperfine needs to be installed.
@@ -174,7 +175,9 @@ rustc-push)
     fi
     ;;
 many-seeds)
-    for SEED in $(seq 0 255); do
+    MIRI_SEED_START=${MIRI_SEED_START:-0} # default to 0
+    MIRI_SEEDS=${MIRI_SEEDS:-256} # default to 256
+    for SEED in $(seq $MIRI_SEED_START $(( $MIRI_SEED_START + $MIRI_SEEDS - 1 )) ); do
         echo "Trying seed: $SEED"
         MIRIFLAGS="$MIRIFLAGS -Zlayout-seed=$SEED -Zmiri-seed=$SEED" $@ || { echo "Failing seed: $SEED"; break; }
     done
@@ -249,6 +252,8 @@ export RUSTFLAGS="-C link-args=-Wl,-rpath,$LIBDIR $RUSTFLAGS"
 # Build a sysroot and set MIRI_SYSROOT to use it. Arguments are passed to `cargo miri setup`.
 build_sysroot() {
     if ! MIRI_SYSROOT="$($CARGO run $CARGO_EXTRA_FLAGS --manifest-path "$MIRIDIR"/cargo-miri/Cargo.toml -- miri setup --print-sysroot "$@")"; then
+        # Run it again so the user can see the error.
+        $CARGO run $CARGO_EXTRA_FLAGS --manifest-path "$MIRIDIR"/cargo-miri/Cargo.toml -- miri setup "$@"
         echo "'cargo miri setup' failed"
         exit 1
     fi
diff --git a/src/tools/miri/src/concurrency/thread.rs b/src/tools/miri/src/concurrency/thread.rs
index dacb3a9b88f..900d24443cc 100644
--- a/src/tools/miri/src/concurrency/thread.rs
+++ b/src/tools/miri/src/concurrency/thread.rs
@@ -3,6 +3,7 @@
 use std::cell::RefCell;
 use std::collections::hash_map::Entry;
 use std::num::TryFromIntError;
+use std::task::Poll;
 use std::time::{Duration, SystemTime};
 
 use log::trace;
@@ -16,18 +17,17 @@ use rustc_target::spec::abi::Abi;
 
 use crate::concurrency::data_race;
 use crate::concurrency::sync::SynchronizationState;
+use crate::shims::tls;
 use crate::*;
 
 #[derive(Clone, Copy, Debug, PartialEq, Eq)]
-pub enum SchedulingAction {
+enum SchedulingAction {
     /// Execute step on the active thread.
     ExecuteStep,
     /// Execute a timeout callback.
     ExecuteTimeoutCallback,
-    /// Execute destructors of the active thread.
-    ExecuteDtors,
-    /// Stop the program.
-    Stop,
+    /// Wait for a bit, until there is a timeout to be called.
+    Sleep(Duration),
 }
 
 /// Trait for callbacks that can be executed when some event happens, such as after a timeout.
@@ -41,9 +41,6 @@ type TimeoutCallback<'mir, 'tcx> = Box<dyn MachineCallback<'mir, 'tcx> + 'tcx>;
 #[derive(Clone, Copy, Debug, PartialOrd, Ord, PartialEq, Eq, Hash)]
 pub struct ThreadId(u32);
 
-/// The main thread. When it terminates, the whole application terminates.
-const MAIN_THREAD: ThreadId = ThreadId(0);
-
 impl ThreadId {
     pub fn to_u32(self) -> u32 {
         self.0
@@ -118,6 +115,12 @@ pub struct Thread<'mir, 'tcx> {
     /// The virtual call stack.
     stack: Vec<Frame<'mir, 'tcx, Provenance, FrameData<'tcx>>>,
 
+    /// The function to call when the stack ran empty, to figure out what to do next.
+    /// Conceptually, this is the interpreter implementation of the things that happen 'after' the
+    /// Rust language entry point for this thread returns (usually implemented by the C or OS runtime).
+    /// (`None` is an error, it means the callback has not been set up yet or is actively running.)
+    pub(crate) on_stack_empty: Option<StackEmptyCallback<'mir, 'tcx>>,
+
     /// The index of the topmost user-relevant frame in `stack`. This field must contain
     /// the value produced by `get_top_user_relevant_frame`.
     /// The `None` state here represents
@@ -137,19 +140,10 @@ pub struct Thread<'mir, 'tcx> {
     pub(crate) last_error: Option<MPlaceTy<'tcx, Provenance>>,
 }
 
-impl<'mir, 'tcx> Thread<'mir, 'tcx> {
-    /// Check if the thread is done executing (no more stack frames). If yes,
-    /// change the state to terminated and return `true`.
-    fn check_terminated(&mut self) -> bool {
-        if self.state == ThreadState::Enabled {
-            if self.stack.is_empty() {
-                self.state = ThreadState::Terminated;
-                return true;
-            }
-        }
-        false
-    }
+pub type StackEmptyCallback<'mir, 'tcx> =
+    Box<dyn FnMut(&mut MiriInterpCx<'mir, 'tcx>) -> InterpResult<'tcx, Poll<()>>>;
 
+impl<'mir, 'tcx> Thread<'mir, 'tcx> {
     /// Get the name of the current thread, or `<unnamed>` if it was not set.
     fn thread_name(&self) -> &[u8] {
         if let Some(ref thread_name) = self.thread_name { thread_name } else { b"<unnamed>" }
@@ -202,28 +196,21 @@ impl<'mir, 'tcx> std::fmt::Debug for Thread<'mir, 'tcx> {
     }
 }
 
-impl<'mir, 'tcx> Default for Thread<'mir, 'tcx> {
-    fn default() -> Self {
+impl<'mir, 'tcx> Thread<'mir, 'tcx> {
+    fn new(name: Option<&str>, on_stack_empty: Option<StackEmptyCallback<'mir, 'tcx>>) -> Self {
         Self {
             state: ThreadState::Enabled,
-            thread_name: None,
+            thread_name: name.map(|name| Vec::from(name.as_bytes())),
             stack: Vec::new(),
             top_user_relevant_frame: None,
             join_status: ThreadJoinStatus::Joinable,
             panic_payload: None,
             last_error: None,
+            on_stack_empty,
         }
     }
 }
 
-impl<'mir, 'tcx> Thread<'mir, 'tcx> {
-    fn new(name: &str) -> Self {
-        let mut thread = Thread::default();
-        thread.thread_name = Some(Vec::from(name.as_bytes()));
-        thread
-    }
-}
-
 impl VisitTags for Thread<'_, '_> {
     fn visit_tags(&self, visit: &mut dyn FnMut(SbTag)) {
         let Thread {
@@ -234,6 +221,7 @@ impl VisitTags for Thread<'_, '_> {
             state: _,
             thread_name: _,
             join_status: _,
+            on_stack_empty: _, // we assume the closure captures no GC-relevant state
         } = self;
 
         panic_payload.visit_tags(visit);
@@ -327,22 +315,6 @@ pub struct ThreadManager<'mir, 'tcx> {
     timeout_callbacks: FxHashMap<ThreadId, TimeoutCallbackInfo<'mir, 'tcx>>,
 }
 
-impl<'mir, 'tcx> Default for ThreadManager<'mir, 'tcx> {
-    fn default() -> Self {
-        let mut threads = IndexVec::new();
-        // Create the main thread and add it to the list of threads.
-        threads.push(Thread::new("main"));
-        Self {
-            active_thread: ThreadId::new(0),
-            threads,
-            sync: SynchronizationState::default(),
-            thread_local_alloc_ids: Default::default(),
-            yield_active_thread: false,
-            timeout_callbacks: FxHashMap::default(),
-        }
-    }
-}
-
 impl VisitTags for ThreadManager<'_, '_> {
     fn visit_tags(&self, visit: &mut dyn FnMut(SbTag)) {
         let ThreadManager {
@@ -367,8 +339,28 @@ impl VisitTags for ThreadManager<'_, '_> {
     }
 }
 
+impl<'mir, 'tcx> Default for ThreadManager<'mir, 'tcx> {
+    fn default() -> Self {
+        let mut threads = IndexVec::new();
+        // Create the main thread and add it to the list of threads.
+        threads.push(Thread::new(Some("main"), None));
+        Self {
+            active_thread: ThreadId::new(0),
+            threads,
+            sync: SynchronizationState::default(),
+            thread_local_alloc_ids: Default::default(),
+            yield_active_thread: false,
+            timeout_callbacks: FxHashMap::default(),
+        }
+    }
+}
+
 impl<'mir, 'tcx: 'mir> ThreadManager<'mir, 'tcx> {
-    pub(crate) fn init(ecx: &mut MiriInterpCx<'mir, 'tcx>) {
+    pub(crate) fn init(
+        ecx: &mut MiriInterpCx<'mir, 'tcx>,
+        on_main_stack_empty: StackEmptyCallback<'mir, 'tcx>,
+    ) {
+        ecx.machine.threads.threads[ThreadId::new(0)].on_stack_empty = Some(on_main_stack_empty);
         if ecx.tcx.sess.target.os.as_ref() != "windows" {
             // The main thread can *not* be joined on except on windows.
             ecx.machine.threads.threads[ThreadId::new(0)].join_status = ThreadJoinStatus::Detached;
@@ -411,9 +403,9 @@ impl<'mir, 'tcx: 'mir> ThreadManager<'mir, 'tcx> {
     }
 
     /// Create a new thread and returns its id.
-    fn create_thread(&mut self) -> ThreadId {
+    fn create_thread(&mut self, on_stack_empty: StackEmptyCallback<'mir, 'tcx>) -> ThreadId {
         let new_thread_id = ThreadId::new(self.threads.len());
-        self.threads.push(Default::default());
+        self.threads.push(Thread::new(None, Some(on_stack_empty)));
         new_thread_id
     }
 
@@ -458,7 +450,7 @@ impl<'mir, 'tcx: 'mir> ThreadManager<'mir, 'tcx> {
     }
 
     /// Get a mutable borrow of the currently active thread.
-    fn active_thread_mut(&mut self) -> &mut Thread<'mir, 'tcx> {
+    pub fn active_thread_mut(&mut self) -> &mut Thread<'mir, 'tcx> {
         &mut self.threads[self.active_thread]
     }
 
@@ -669,18 +661,6 @@ impl<'mir, 'tcx: 'mir> ThreadManager<'mir, 'tcx> {
     /// long as we can and switch only when we have to (the active thread was
     /// blocked, terminated, or has explicitly asked to be preempted).
     fn schedule(&mut self, clock: &Clock) -> InterpResult<'tcx, SchedulingAction> {
-        // Check whether the thread has **just** terminated (`check_terminated`
-        // checks whether the thread has popped all its stack and if yes, sets
-        // the thread state to terminated).
-        if self.threads[self.active_thread].check_terminated() {
-            return Ok(SchedulingAction::ExecuteDtors);
-        }
-        // If we get here again and the thread is *still* terminated, there are no more dtors to run.
-        if self.threads[MAIN_THREAD].state == ThreadState::Terminated {
-            // The main thread terminated; stop the program.
-            // We do *not* run TLS dtors of remaining threads, which seems to match rustc behavior.
-            return Ok(SchedulingAction::Stop);
-        }
         // This thread and the program can keep going.
         if self.threads[self.active_thread].state == ThreadState::Enabled
             && !self.yield_active_thread
@@ -688,18 +668,18 @@ impl<'mir, 'tcx: 'mir> ThreadManager<'mir, 'tcx> {
             // The currently active thread is still enabled, just continue with it.
             return Ok(SchedulingAction::ExecuteStep);
         }
-        // The active thread yielded. Let's see if there are any timeouts to take care of. We do
-        // this *before* running any other thread, to ensure that timeouts "in the past" fire before
-        // any other thread can take an action. This ensures that for `pthread_cond_timedwait`, "an
-        // error is returned if [...] the absolute time specified by abstime has already been passed
-        // at the time of the call".
+        // The active thread yielded or got terminated. Let's see if there are any timeouts to take
+        // care of. We do this *before* running any other thread, to ensure that timeouts "in the
+        // past" fire before any other thread can take an action. This ensures that for
+        // `pthread_cond_timedwait`, "an error is returned if [...] the absolute time specified by
+        // abstime has already been passed at the time of the call".
         // <https://pubs.opengroup.org/onlinepubs/9699919799/functions/pthread_cond_timedwait.html>
         let potential_sleep_time =
             self.timeout_callbacks.values().map(|info| info.call_time.get_wait_time(clock)).min();
         if potential_sleep_time == Some(Duration::new(0, 0)) {
             return Ok(SchedulingAction::ExecuteTimeoutCallback);
         }
-        // No callbacks scheduled, pick a regular thread to execute.
+        // No callbacks immediately scheduled, pick a regular thread to execute.
         // The active thread blocked or yielded. So we go search for another enabled thread.
         // Crucially, we start searching at the current active thread ID, rather than at 0, since we
         // want to avoid always scheduling threads 0 and 1 without ever making progress in thread 2.
@@ -730,15 +710,58 @@ impl<'mir, 'tcx: 'mir> ThreadManager<'mir, 'tcx> {
             // All threads are currently blocked, but we have unexecuted
             // timeout_callbacks, which may unblock some of the threads. Hence,
             // sleep until the first callback.
-
-            clock.sleep(sleep_time);
-            Ok(SchedulingAction::ExecuteTimeoutCallback)
+            Ok(SchedulingAction::Sleep(sleep_time))
         } else {
             throw_machine_stop!(TerminationInfo::Deadlock);
         }
     }
 }
 
+impl<'mir, 'tcx: 'mir> EvalContextPrivExt<'mir, 'tcx> for MiriInterpCx<'mir, 'tcx> {}
+trait EvalContextPrivExt<'mir, 'tcx: 'mir>: MiriInterpCxExt<'mir, 'tcx> {
+    /// Execute a timeout callback on the callback's thread.
+    #[inline]
+    fn run_timeout_callback(&mut self) -> InterpResult<'tcx> {
+        let this = self.eval_context_mut();
+        let (thread, callback) = if let Some((thread, callback)) =
+            this.machine.threads.get_ready_callback(&this.machine.clock)
+        {
+            (thread, callback)
+        } else {
+            // get_ready_callback can return None if the computer's clock
+            // was shifted after calling the scheduler and before the call
+            // to get_ready_callback (see issue
+            // https://github.com/rust-lang/miri/issues/1763). In this case,
+            // just do nothing, which effectively just returns to the
+            // scheduler.
+            return Ok(());
+        };
+        // This back-and-forth with `set_active_thread` is here because of two
+        // design decisions:
+        // 1. Make the caller and not the callback responsible for changing
+        //    thread.
+        // 2. Make the scheduler the only place that can change the active
+        //    thread.
+        let old_thread = this.set_active_thread(thread);
+        callback.call(this)?;
+        this.set_active_thread(old_thread);
+        Ok(())
+    }
+
+    #[inline]
+    fn run_on_stack_empty(&mut self) -> InterpResult<'tcx, Poll<()>> {
+        let this = self.eval_context_mut();
+        let mut callback = this
+            .active_thread_mut()
+            .on_stack_empty
+            .take()
+            .expect("`on_stack_empty` not set up, or already running");
+        let res = callback(this)?;
+        this.active_thread_mut().on_stack_empty = Some(callback);
+        Ok(res)
+    }
+}
+
 // Public interface to thread management.
 impl<'mir, 'tcx: 'mir> EvalContextExt<'mir, 'tcx> for crate::MiriInterpCx<'mir, 'tcx> {}
 pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriInterpCxExt<'mir, 'tcx> {
@@ -773,18 +796,9 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriInterpCxExt<'mir, 'tcx> {
         }
     }
 
+    /// Start a regular (non-main) thread.
     #[inline]
-    fn create_thread(&mut self) -> ThreadId {
-        let this = self.eval_context_mut();
-        let id = this.machine.threads.create_thread();
-        if let Some(data_race) = &mut this.machine.data_race {
-            data_race.thread_created(&this.machine.threads, id);
-        }
-        id
-    }
-
-    #[inline]
-    fn start_thread(
+    fn start_regular_thread(
         &mut self,
         thread: Option<MPlaceTy<'tcx, Provenance>>,
         start_routine: Pointer<Option<Provenance>>,
@@ -795,7 +809,13 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriInterpCxExt<'mir, 'tcx> {
         let this = self.eval_context_mut();
 
         // Create the new thread
-        let new_thread_id = this.create_thread();
+        let new_thread_id = this.machine.threads.create_thread({
+            let mut state = tls::TlsDtorsState::default();
+            Box::new(move |m| state.on_stack_empty(m))
+        });
+        if let Some(data_race) = &mut this.machine.data_race {
+            data_race.thread_created(&this.machine.threads, new_thread_id);
+        }
 
         // Write the current thread-id, switch to the next thread later
         // to treat this write operation as occuring on the current thread.
@@ -889,12 +909,6 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriInterpCxExt<'mir, 'tcx> {
     }
 
     #[inline]
-    fn has_terminated(&self, thread_id: ThreadId) -> bool {
-        let this = self.eval_context_ref();
-        this.machine.threads.has_terminated(thread_id)
-    }
-
-    #[inline]
     fn have_all_terminated(&self) -> bool {
         let this = self.eval_context_ref();
         this.machine.threads.have_all_terminated()
@@ -943,26 +957,22 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriInterpCxExt<'mir, 'tcx> {
     where
         'mir: 'c,
     {
-        let this = self.eval_context_ref();
-        this.machine.threads.get_thread_name(thread)
+        self.eval_context_ref().machine.threads.get_thread_name(thread)
     }
 
     #[inline]
     fn block_thread(&mut self, thread: ThreadId) {
-        let this = self.eval_context_mut();
-        this.machine.threads.block_thread(thread);
+        self.eval_context_mut().machine.threads.block_thread(thread);
     }
 
     #[inline]
     fn unblock_thread(&mut self, thread: ThreadId) {
-        let this = self.eval_context_mut();
-        this.machine.threads.unblock_thread(thread);
+        self.eval_context_mut().machine.threads.unblock_thread(thread);
     }
 
     #[inline]
     fn yield_active_thread(&mut self) {
-        let this = self.eval_context_mut();
-        this.machine.threads.yield_active_thread();
+        self.eval_context_mut().machine.threads.yield_active_thread();
     }
 
     #[inline]
@@ -995,49 +1005,42 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriInterpCxExt<'mir, 'tcx> {
         this.machine.threads.unregister_timeout_callback_if_exists(thread);
     }
 
-    /// Execute a timeout callback on the callback's thread.
-    #[inline]
-    fn run_timeout_callback(&mut self) -> InterpResult<'tcx> {
-        let this = self.eval_context_mut();
-        let (thread, callback) = if let Some((thread, callback)) =
-            this.machine.threads.get_ready_callback(&this.machine.clock)
-        {
-            (thread, callback)
-        } else {
-            // get_ready_callback can return None if the computer's clock
-            // was shifted after calling the scheduler and before the call
-            // to get_ready_callback (see issue
-            // https://github.com/rust-lang/miri/issues/1763). In this case,
-            // just do nothing, which effectively just returns to the
-            // scheduler.
-            return Ok(());
-        };
-        // This back-and-forth with `set_active_thread` is here because of two
-        // design decisions:
-        // 1. Make the caller and not the callback responsible for changing
-        //    thread.
-        // 2. Make the scheduler the only place that can change the active
-        //    thread.
-        let old_thread = this.set_active_thread(thread);
-        callback.call(this)?;
-        this.set_active_thread(old_thread);
-        Ok(())
-    }
-
-    /// Decide which action to take next and on which thread.
-    #[inline]
-    fn schedule(&mut self) -> InterpResult<'tcx, SchedulingAction> {
+    /// Run the core interpreter loop. Returns only when an interrupt occurs (an error or program
+    /// termination).
+    fn run_threads(&mut self) -> InterpResult<'tcx, !> {
         let this = self.eval_context_mut();
-        this.machine.threads.schedule(&this.machine.clock)
+        loop {
+            match this.machine.threads.schedule(&this.machine.clock)? {
+                SchedulingAction::ExecuteStep => {
+                    if !this.step()? {
+                        // See if this thread can do something else.
+                        match this.run_on_stack_empty()? {
+                            Poll::Pending => {} // keep going
+                            Poll::Ready(()) => this.terminate_active_thread()?,
+                        }
+                    }
+                }
+                SchedulingAction::ExecuteTimeoutCallback => {
+                    this.run_timeout_callback()?;
+                }
+                SchedulingAction::Sleep(duration) => {
+                    this.machine.clock.sleep(duration);
+                }
+            }
+        }
     }
 
     /// Handles thread termination of the active thread: wakes up threads joining on this one,
     /// and deallocated thread-local statics.
     ///
-    /// This is called from `tls.rs` after handling the TLS dtors.
+    /// This is called by the eval loop when a thread's on_stack_empty returns `Ready`.
     #[inline]
-    fn thread_terminated(&mut self) -> InterpResult<'tcx> {
+    fn terminate_active_thread(&mut self) -> InterpResult<'tcx> {
         let this = self.eval_context_mut();
+        let thread = this.active_thread_mut();
+        assert!(thread.stack.is_empty(), "only threads with an empty stack can be terminated");
+        thread.state = ThreadState::Terminated;
+
         for ptr in this.machine.threads.thread_terminated(this.machine.data_race.as_mut()) {
             this.deallocate_ptr(ptr.into(), None, MiriMemoryKind::Tls.into())?;
         }
diff --git a/src/tools/miri/src/diagnostics.rs b/src/tools/miri/src/diagnostics.rs
index 7658cea10f9..5cd0a0eeb58 100644
--- a/src/tools/miri/src/diagnostics.rs
+++ b/src/tools/miri/src/diagnostics.rs
@@ -11,7 +11,10 @@ use crate::*;
 
 /// Details of premature program termination.
 pub enum TerminationInfo {
-    Exit(i64),
+    Exit {
+        code: i64,
+        leak_check: bool,
+    },
     Abort(String),
     UnsupportedInIsolation(String),
     StackedBorrowsUb {
@@ -38,7 +41,7 @@ impl fmt::Display for TerminationInfo {
     fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
         use TerminationInfo::*;
         match self {
-            Exit(code) => write!(f, "the evaluated program completed with exit code {code}"),
+            Exit { code, .. } => write!(f, "the evaluated program completed with exit code {code}"),
             Abort(msg) => write!(f, "{msg}"),
             UnsupportedInIsolation(msg) => write!(f, "{msg}"),
             Int2PtrWithStrictProvenance =>
@@ -148,11 +151,11 @@ fn prune_stacktrace<'tcx>(
 
 /// Emit a custom diagnostic without going through the miri-engine machinery.
 ///
-/// Returns `Some` if this was regular program termination with a given exit code, `None` otherwise.
+/// Returns `Some` if this was regular program termination with a given exit code and a `bool` indicating whether a leak check should happen; `None` otherwise.
 pub fn report_error<'tcx, 'mir>(
     ecx: &InterpCx<'mir, 'tcx, MiriMachine<'mir, 'tcx>>,
     e: InterpErrorInfo<'tcx>,
-) -> Option<i64> {
+) -> Option<(i64, bool)> {
     use InterpError::*;
 
     let mut msg = vec![];
@@ -161,7 +164,7 @@ pub fn report_error<'tcx, 'mir>(
         let info = info.downcast_ref::<TerminationInfo>().expect("invalid MachineStop payload");
         use TerminationInfo::*;
         let title = match info {
-            Exit(code) => return Some(*code),
+            Exit { code, leak_check } => return Some((*code, *leak_check)),
             Abort(_) => Some("abnormal termination"),
             UnsupportedInIsolation(_) | Int2PtrWithStrictProvenance =>
                 Some("unsupported operation"),
diff --git a/src/tools/miri/src/eval.rs b/src/tools/miri/src/eval.rs
index 363b647d6c6..b8578b1277f 100644
--- a/src/tools/miri/src/eval.rs
+++ b/src/tools/miri/src/eval.rs
@@ -4,6 +4,7 @@ use std::ffi::{OsStr, OsString};
 use std::iter;
 use std::panic::{self, AssertUnwindSafe};
 use std::path::PathBuf;
+use std::task::Poll;
 use std::thread;
 
 use log::info;
@@ -20,8 +21,14 @@ use rustc_target::spec::abi::Abi;
 
 use rustc_session::config::EntryFnType;
 
+use crate::shims::tls;
 use crate::*;
 
+/// When the main thread would exit, we will yield to any other thread that is ready to execute.
+/// But we must only do that a finite number of times, or a background thread running `loop {}`
+/// will hang the program.
+const MAIN_THREAD_YIELDS_AT_SHUTDOWN: u32 = 256;
+
 #[derive(Copy, Clone, Debug, PartialEq)]
 pub enum AlignmentCheck {
     /// Do not check alignment.
@@ -172,17 +179,79 @@ impl Default for MiriConfig {
     }
 }
 
-/// Returns a freshly created `InterpCx`, along with an `MPlaceTy` representing
-/// the location where the return value of the `start` function will be
-/// written to.
+/// The state of the main thread. Implementation detail of `on_main_stack_empty`.
+#[derive(Default, Debug)]
+enum MainThreadState {
+    #[default]
+    Running,
+    TlsDtors(tls::TlsDtorsState),
+    Yield {
+        remaining: u32,
+    },
+    Done,
+}
+
+impl MainThreadState {
+    fn on_main_stack_empty<'tcx>(
+        &mut self,
+        this: &mut MiriInterpCx<'_, 'tcx>,
+    ) -> InterpResult<'tcx, Poll<()>> {
+        use MainThreadState::*;
+        match self {
+            Running => {
+                *self = TlsDtors(Default::default());
+            }
+            TlsDtors(state) =>
+                match state.on_stack_empty(this)? {
+                    Poll::Pending => {} // just keep going
+                    Poll::Ready(()) => {
+                        // Give background threads a chance to finish by yielding the main thread a
+                        // couple of times -- but only if we would also preempt threads randomly.
+                        if this.machine.preemption_rate > 0.0 {
+                            // There is a non-zero chance they will yield back to us often enough to
+                            // make Miri terminate eventually.
+                            *self = Yield { remaining: MAIN_THREAD_YIELDS_AT_SHUTDOWN };
+                        } else {
+                            // The other threads did not get preempted, so no need to yield back to
+                            // them.
+                            *self = Done;
+                        }
+                    }
+                },
+            Yield { remaining } =>
+                match remaining.checked_sub(1) {
+                    None => *self = Done,
+                    Some(new_remaining) => {
+                        *remaining = new_remaining;
+                        this.yield_active_thread();
+                    }
+                },
+            Done => {
+                // Figure out exit code.
+                let ret_place = MPlaceTy::from_aligned_ptr(
+                    this.machine.main_fn_ret_place.unwrap().ptr,
+                    this.machine.layouts.isize,
+                );
+                let exit_code = this.read_scalar(&ret_place.into())?.to_machine_isize(this)?;
+                // Need to call this ourselves since we are not going to return to the scheduler
+                // loop, and we want the main thread TLS to not show up as memory leaks.
+                this.terminate_active_thread()?;
+                // Stop interpreter loop.
+                throw_machine_stop!(TerminationInfo::Exit { code: exit_code, leak_check: true });
+            }
+        }
+        Ok(Poll::Pending)
+    }
+}
+
+/// Returns a freshly created `InterpCx`.
 /// Public because this is also used by `priroda`.
 pub fn create_ecx<'mir, 'tcx: 'mir>(
     tcx: TyCtxt<'tcx>,
     entry_id: DefId,
     entry_type: EntryFnType,
     config: &MiriConfig,
-) -> InterpResult<'tcx, (InterpCx<'mir, 'tcx, MiriMachine<'mir, 'tcx>>, MPlaceTy<'tcx, Provenance>)>
-{
+) -> InterpResult<'tcx, InterpCx<'mir, 'tcx, MiriMachine<'mir, 'tcx>>> {
     let param_env = ty::ParamEnv::reveal_all();
     let layout_cx = LayoutCx { tcx, param_env };
     let mut ecx = InterpCx::new(
@@ -193,7 +262,11 @@ pub fn create_ecx<'mir, 'tcx: 'mir>(
     );
 
     // Some parts of initialization require a full `InterpCx`.
-    MiriMachine::late_init(&mut ecx, config)?;
+    MiriMachine::late_init(&mut ecx, config, {
+        let mut state = MainThreadState::default();
+        // Cannot capture anything GC-relevant here.
+        Box::new(move |m| state.on_main_stack_empty(m))
+    })?;
 
     // Make sure we have MIR. We check MIR for some stable monomorphic function in libcore.
     let sentinel = ecx.try_resolve_path(&["core", "ascii", "escape_default"], Namespace::ValueNS);
@@ -274,6 +347,7 @@ pub fn create_ecx<'mir, 'tcx: 'mir>(
 
     // Return place (in static memory so that it does not count as leak).
     let ret_place = ecx.allocate(ecx.machine.layouts.isize, MiriMemoryKind::Machine.into())?;
+    ecx.machine.main_fn_ret_place = Some(*ret_place);
     // Call start function.
 
     match entry_type {
@@ -321,7 +395,7 @@ pub fn create_ecx<'mir, 'tcx: 'mir>(
         }
     }
 
-    Ok((ecx, ret_place))
+    Ok(ecx)
 }
 
 /// Evaluates the entry function specified by `entry_id`.
@@ -337,7 +411,7 @@ pub fn eval_entry<'tcx>(
     // Copy setting before we move `config`.
     let ignore_leaks = config.ignore_leaks;
 
-    let (mut ecx, ret_place) = match create_ecx(tcx, entry_id, entry_type, &config) {
+    let mut ecx = match create_ecx(tcx, entry_id, entry_type, &config) {
         Ok(v) => v,
         Err(err) => {
             err.print_backtrace();
@@ -346,34 +420,17 @@ pub fn eval_entry<'tcx>(
     };
 
     // Perform the main execution.
-    let res: thread::Result<InterpResult<'_, i64>> = panic::catch_unwind(AssertUnwindSafe(|| {
-        // Main loop.
-        loop {
-            match ecx.schedule()? {
-                SchedulingAction::ExecuteStep => {
-                    assert!(ecx.step()?, "a terminated thread was scheduled for execution");
-                }
-                SchedulingAction::ExecuteTimeoutCallback => {
-                    ecx.run_timeout_callback()?;
-                }
-                SchedulingAction::ExecuteDtors => {
-                    // This will either enable the thread again (so we go back
-                    // to `ExecuteStep`), or determine that this thread is done
-                    // for good.
-                    ecx.schedule_next_tls_dtor_for_active_thread()?;
-                }
-                SchedulingAction::Stop => {
-                    break;
-                }
-            }
-        }
-        let return_code = ecx.read_scalar(&ret_place.into())?.to_machine_isize(&ecx)?;
-        Ok(return_code)
-    }));
+    let res: thread::Result<InterpResult<'_, !>> =
+        panic::catch_unwind(AssertUnwindSafe(|| ecx.run_threads()));
     let res = res.unwrap_or_else(|panic_payload| {
         ecx.handle_ice();
         panic::resume_unwind(panic_payload)
     });
+    let res = match res {
+        Err(res) => res,
+        // `Ok` can never happen
+        Ok(never) => match never {},
+    };
 
     // Machine cleanup. Only do this if all threads have terminated; threads that are still running
     // might cause Stacked Borrows errors (https://github.com/rust-lang/miri/issues/2396).
@@ -386,32 +443,26 @@ pub fn eval_entry<'tcx>(
     }
 
     // Process the result.
-    match res {
-        Ok(return_code) => {
-            if !ignore_leaks {
-                // Check for thread leaks.
-                if !ecx.have_all_terminated() {
-                    tcx.sess.err(
-                        "the main thread terminated without waiting for all remaining threads",
-                    );
-                    tcx.sess.note_without_error("pass `-Zmiri-ignore-leaks` to disable this check");
-                    return None;
-                }
-                // Check for memory leaks.
-                info!("Additonal static roots: {:?}", ecx.machine.static_roots);
-                let leaks = ecx.leak_report(&ecx.machine.static_roots);
-                if leaks != 0 {
-                    tcx.sess.err("the evaluated program leaked memory");
-                    tcx.sess.note_without_error("pass `-Zmiri-ignore-leaks` to disable this check");
-                    // Ignore the provided return code - let the reported error
-                    // determine the return code.
-                    return None;
-                }
-            }
-            Some(return_code)
+    let (return_code, leak_check) = report_error(&ecx, res)?;
+    if leak_check && !ignore_leaks {
+        // Check for thread leaks.
+        if !ecx.have_all_terminated() {
+            tcx.sess.err("the main thread terminated without waiting for all remaining threads");
+            tcx.sess.note_without_error("pass `-Zmiri-ignore-leaks` to disable this check");
+            return None;
+        }
+        // Check for memory leaks.
+        info!("Additonal static roots: {:?}", ecx.machine.static_roots);
+        let leaks = ecx.leak_report(&ecx.machine.static_roots);
+        if leaks != 0 {
+            tcx.sess.err("the evaluated program leaked memory");
+            tcx.sess.note_without_error("pass `-Zmiri-ignore-leaks` to disable this check");
+            // Ignore the provided return code - let the reported error
+            // determine the return code.
+            return None;
         }
-        Err(e) => report_error(&ecx, e),
     }
+    Some(return_code)
 }
 
 /// Turns an array of arguments into a Windows command line string.
diff --git a/src/tools/miri/src/lib.rs b/src/tools/miri/src/lib.rs
index 8913f8aa10f..6f483cf2cc4 100644
--- a/src/tools/miri/src/lib.rs
+++ b/src/tools/miri/src/lib.rs
@@ -81,7 +81,7 @@ pub use crate::shims::intrinsics::EvalContextExt as _;
 pub use crate::shims::os_str::EvalContextExt as _;
 pub use crate::shims::panic::{CatchUnwindData, EvalContextExt as _};
 pub use crate::shims::time::EvalContextExt as _;
-pub use crate::shims::tls::{EvalContextExt as _, TlsData};
+pub use crate::shims::tls::TlsData;
 pub use crate::shims::EvalContextExt as _;
 
 pub use crate::clock::{Clock, Instant};
@@ -89,7 +89,7 @@ pub use crate::concurrency::{
     data_race::{AtomicFenceOrd, AtomicReadOrd, AtomicRwOrd, AtomicWriteOrd, EvalContextExt as _},
     init_once::{EvalContextExt as _, InitOnceId},
     sync::{CondvarId, EvalContextExt as _, MutexId, RwLockId, SyncId},
-    thread::{EvalContextExt as _, SchedulingAction, ThreadId, ThreadManager, ThreadState, Time},
+    thread::{EvalContextExt as _, StackEmptyCallback, ThreadId, ThreadManager, Time},
 };
 pub use crate::diagnostics::{
     report_error, EvalContextExt as _, NonHaltingDiagnostic, TerminationInfo,
@@ -107,7 +107,7 @@ pub use crate::mono_hash_map::MonoHashMap;
 pub use crate::operator::EvalContextExt as _;
 pub use crate::range_map::RangeMap;
 pub use crate::stacked_borrows::{
-    CallId, EvalContextExt as _, Item, Permission, RetagFields, SbTag, Stack, Stacks,
+    CallId, EvalContextExt as _, Item, Permission, RetagFields, SbTag,
 };
 pub use crate::tag_gc::{EvalContextExt as _, VisitTags};
 
diff --git a/src/tools/miri/src/machine.rs b/src/tools/miri/src/machine.rs
index edfef211dc6..4d3444bc39c 100644
--- a/src/tools/miri/src/machine.rs
+++ b/src/tools/miri/src/machine.rs
@@ -363,6 +363,9 @@ pub struct MiriMachine<'mir, 'tcx> {
     /// Miri does not expose env vars from the host to the emulated program.
     pub(crate) env_vars: EnvVars<'tcx>,
 
+    /// Return place of the main function.
+    pub(crate) main_fn_ret_place: Option<MemPlace<Provenance>>,
+
     /// Program arguments (`Option` because we can only initialize them after creating the ecx).
     /// These are *pointers* to argc/argv because macOS.
     /// We also need the full command line as one string because of Windows.
@@ -492,6 +495,7 @@ impl<'mir, 'tcx> MiriMachine<'mir, 'tcx> {
             intptrcast: RefCell::new(intptrcast::GlobalStateInner::new(config)),
             // `env_vars` depends on a full interpreter so we cannot properly initialize it yet.
             env_vars: EnvVars::default(),
+            main_fn_ret_place: None,
             argc: None,
             argv: None,
             cmd_line: None,
@@ -556,10 +560,11 @@ impl<'mir, 'tcx> MiriMachine<'mir, 'tcx> {
     pub(crate) fn late_init(
         this: &mut MiriInterpCx<'mir, 'tcx>,
         config: &MiriConfig,
+        on_main_stack_empty: StackEmptyCallback<'mir, 'tcx>,
     ) -> InterpResult<'tcx> {
         EnvVars::init(this, config)?;
         MiriMachine::init_extern_statics(this)?;
-        ThreadManager::init(this);
+        ThreadManager::init(this, on_main_stack_empty);
         Ok(())
     }
 
@@ -657,6 +662,7 @@ impl VisitTags for MiriMachine<'_, '_> {
             threads,
             tls,
             env_vars,
+            main_fn_ret_place,
             argc,
             argv,
             cmd_line,
@@ -702,6 +708,7 @@ impl VisitTags for MiriMachine<'_, '_> {
         data_race.visit_tags(visit);
         stacked_borrows.visit_tags(visit);
         intptrcast.visit_tags(visit);
+        main_fn_ret_place.visit_tags(visit);
         argc.visit_tags(visit);
         argv.visit_tags(visit);
         cmd_line.visit_tags(visit);
@@ -901,7 +908,13 @@ impl<'mir, 'tcx> Machine<'mir, 'tcx> for MiriMachine<'mir, 'tcx> {
 
         let alloc = alloc.into_owned();
         let stacks = ecx.machine.stacked_borrows.as_ref().map(|stacked_borrows| {
-            Stacks::new_allocation(id, alloc.size(), stacked_borrows, kind, &ecx.machine)
+            stacked_borrows::Stacks::new_allocation(
+                id,
+                alloc.size(),
+                stacked_borrows,
+                kind,
+                &ecx.machine,
+            )
         });
         let race_alloc = ecx.machine.data_race.as_ref().map(|data_race| {
             data_race::AllocExtra::new_allocation(
diff --git a/src/tools/miri/src/shims/foreign_items.rs b/src/tools/miri/src/shims/foreign_items.rs
index 058f730833b..f72521f64ad 100644
--- a/src/tools/miri/src/shims/foreign_items.rs
+++ b/src/tools/miri/src/shims/foreign_items.rs
@@ -286,7 +286,7 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriInterpCxExt<'mir, 'tcx> {
                         let [code] = this.check_shim(abi, exp_abi, link_name, args)?;
                         // it's really u32 for ExitProcess, but we have to put it into the `Exit` variant anyway
                         let code = this.read_scalar(code)?.to_i32()?;
-                        throw_machine_stop!(TerminationInfo::Exit(code.into()));
+                        throw_machine_stop!(TerminationInfo::Exit { code: code.into(), leak_check: false });
                     }
                     "abort" => {
                         let [] = this.check_shim(abi, Abi::C { unwind: false }, link_name, args)?;
diff --git a/src/tools/miri/src/shims/tls.rs b/src/tools/miri/src/shims/tls.rs
index 5fda8bd7b7d..fe278ff717f 100644
--- a/src/tools/miri/src/shims/tls.rs
+++ b/src/tools/miri/src/shims/tls.rs
@@ -1,12 +1,11 @@
 //! Implement thread-local storage.
 
 use std::collections::btree_map::Entry as BTreeEntry;
-use std::collections::hash_map::Entry as HashMapEntry;
 use std::collections::BTreeMap;
+use std::task::Poll;
 
 use log::trace;
 
-use rustc_data_structures::fx::FxHashMap;
 use rustc_middle::ty;
 use rustc_target::abi::{HasDataLayout, Size};
 use rustc_target::spec::abi::Abi;
@@ -23,12 +22,12 @@ pub struct TlsEntry<'tcx> {
     dtor: Option<ty::Instance<'tcx>>,
 }
 
-#[derive(Clone, Debug)]
-struct RunningDtorsState {
+#[derive(Default, Debug)]
+struct RunningDtorState {
     /// 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>,
+    last_key: Option<TlsKey>,
 }
 
 #[derive(Debug)]
@@ -42,11 +41,6 @@ pub struct TlsData<'tcx> {
     /// 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<Provenance>)>,
-
-    /// 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> {
@@ -55,7 +49,6 @@ impl<'tcx> Default for TlsData<'tcx> {
             next_key: 1, // start with 1 as we must not use 0 on Windows
             keys: Default::default(),
             macos_thread_dtors: Default::default(),
-            dtors_running: Default::default(),
         }
     }
 }
@@ -143,12 +136,6 @@ impl<'tcx> TlsData<'tcx> {
         dtor: ty::Instance<'tcx>,
         data: Scalar<Provenance>,
     ) -> InterpResult<'tcx> {
-        if self.dtors_running.contains_key(&thread) {
-            // UB, according to libstd docs.
-            throw_ub_format!(
-                "setting thread's local storage destructor while destructors are already running"
-            );
-        }
         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"
@@ -211,21 +198,6 @@ impl<'tcx> TlsData<'tcx> {
         None
     }
 
-    /// 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
-            }
-        }
-    }
-
     /// 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) {
@@ -237,7 +209,7 @@ impl<'tcx> TlsData<'tcx> {
 
 impl VisitTags for TlsData<'_> {
     fn visit_tags(&self, visit: &mut dyn FnMut(SbTag)) {
-        let TlsData { keys, macos_thread_dtors, next_key: _, dtors_running: _ } = self;
+        let TlsData { keys, macos_thread_dtors, next_key: _ } = self;
 
         for scalar in keys.values().flat_map(|v| v.data.values()) {
             scalar.visit_tags(visit);
@@ -248,13 +220,77 @@ impl VisitTags for TlsData<'_> {
     }
 }
 
+#[derive(Debug, Default)]
+pub struct TlsDtorsState(TlsDtorsStatePriv);
+
+#[derive(Debug, Default)]
+enum TlsDtorsStatePriv {
+    #[default]
+    Init,
+    PthreadDtors(RunningDtorState),
+    Done,
+}
+
+impl TlsDtorsState {
+    pub fn on_stack_empty<'tcx>(
+        &mut self,
+        this: &mut MiriInterpCx<'_, 'tcx>,
+    ) -> InterpResult<'tcx, Poll<()>> {
+        use TlsDtorsStatePriv::*;
+        match &mut self.0 {
+            Init => {
+                match this.tcx.sess.target.os.as_ref() {
+                    "linux" | "freebsd" | "android" => {
+                        // Run the pthread dtors.
+                        self.0 = PthreadDtors(Default::default());
+                    }
+                    "macos" => {
+                        // The macOS thread wide destructor runs "before any TLS slots get
+                        // freed", so do that first.
+                        this.schedule_macos_tls_dtor()?;
+                        // When the stack is empty again, go on with the pthread dtors.
+                        self.0 = PthreadDtors(Default::default());
+                    }
+                    "windows" => {
+                        // Run the special magic hook.
+                        this.schedule_windows_tls_dtors()?;
+                        // And move to the final state.
+                        self.0 = Done;
+                    }
+                    "wasi" | "none" => {
+                        // No OS, no TLS dtors.
+                        // FIXME: should we do something on wasi?
+                        self.0 = Done;
+                    }
+                    os => {
+                        throw_unsup_format!(
+                            "the TLS machinery does not know how to handle OS `{os}`"
+                        );
+                    }
+                }
+            }
+            PthreadDtors(state) => {
+                match this.schedule_next_pthread_tls_dtor(state)? {
+                    Poll::Pending => {} // just keep going
+                    Poll::Ready(()) => self.0 = Done,
+                }
+            }
+            Done => {
+                this.machine.tls.delete_all_thread_tls(this.get_active_thread());
+                return Ok(Poll::Ready(()));
+            }
+        }
+
+        Ok(Poll::Pending)
+    }
+}
+
 impl<'mir, 'tcx: 'mir> EvalContextPrivExt<'mir, 'tcx> for crate::MiriInterpCx<'mir, 'tcx> {}
 trait EvalContextPrivExt<'mir, 'tcx: 'mir>: crate::MiriInterpCxExt<'mir, 'tcx> {
     /// Schedule TLS destructors for Windows.
     /// On windows, TLS destructors are managed by std.
     fn schedule_windows_tls_dtors(&mut self) -> InterpResult<'tcx> {
         let this = self.eval_context_mut();
-        let active_thread = this.get_active_thread();
 
         // 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
@@ -284,16 +320,12 @@ trait EvalContextPrivExt<'mir, 'tcx: 'mir>: crate::MiriInterpCxExt<'mir, 'tcx> {
             None,
             StackPopCleanup::Root { cleanup: true },
         )?;
-
-        this.enable_thread(active_thread);
         Ok(())
     }
 
     /// Schedule the MacOS thread destructor of the thread local storage to be
-    /// executed. Returns `true` if scheduled.
-    ///
-    /// Note: It is safe to call this function also on other Unixes.
-    fn schedule_macos_tls_dtor(&mut self) -> InterpResult<'tcx, bool> {
+    /// executed.
+    fn schedule_macos_tls_dtor(&mut self) -> InterpResult<'tcx> {
         let this = self.eval_context_mut();
         let thread_id = this.get_active_thread();
         if let Some((instance, data)) = this.machine.tls.macos_thread_dtors.remove(&thread_id) {
@@ -306,35 +338,27 @@ trait EvalContextPrivExt<'mir, 'tcx: 'mir>: crate::MiriInterpCxExt<'mir, 'tcx> {
                 None,
                 StackPopCleanup::Root { cleanup: true },
             )?;
-
-            // 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)
         }
+        Ok(())
     }
 
     /// 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> {
+    fn schedule_next_pthread_tls_dtor(
+        &mut self,
+        state: &mut RunningDtorState,
+    ) -> InterpResult<'tcx, Poll<()>> {
         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;
-        let dtor = match this.machine.tls.fetch_tls_dtor(last_key, active_thread) {
+        let dtor = match this.machine.tls.fetch_tls_dtor(state.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);
+            state.last_key = Some(key);
             trace!("Running TLS dtor {:?} on {:?} at {:?}", instance, ptr, active_thread);
             assert!(
                 !ptr.to_machine_usize(this).unwrap() != 0,
@@ -349,64 +373,9 @@ trait EvalContextPrivExt<'mir, 'tcx: 'mir>: crate::MiriInterpCxExt<'mir, 'tcx> {
                 StackPopCleanup::Root { cleanup: true },
             )?;
 
-            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::MiriInterpCx<'mir, 'tcx> {}
-pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriInterpCxExt<'mir, 'tcx> {
-    /// 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();
-        trace!("schedule_next_tls_dtor_for_active_thread on thread {:?}", 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(());
-            }
+            return Ok(Poll::Pending);
         }
-        // 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(())
+        Ok(Poll::Ready(()))
     }
 }
diff --git a/src/tools/miri/src/shims/unix/thread.rs b/src/tools/miri/src/shims/unix/thread.rs
index b43682710bb..5b9dc90f0f0 100644
--- a/src/tools/miri/src/shims/unix/thread.rs
+++ b/src/tools/miri/src/shims/unix/thread.rs
@@ -19,7 +19,7 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriInterpCxExt<'mir, 'tcx> {
 
         let func_arg = this.read_immediate(arg)?;
 
-        this.start_thread(
+        this.start_regular_thread(
             Some(thread_info_place),
             start_routine,
             Abi::C { unwind: false },
diff --git a/src/tools/miri/src/shims/windows/thread.rs b/src/tools/miri/src/shims/windows/thread.rs
index 5ed0cb92f9e..25a5194caa0 100644
--- a/src/tools/miri/src/shims/windows/thread.rs
+++ b/src/tools/miri/src/shims/windows/thread.rs
@@ -46,7 +46,7 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriInterpCxExt<'mir, 'tcx> {
             throw_unsup_format!("non-null `lpThreadAttributes` in `CreateThread`")
         }
 
-        this.start_thread(
+        this.start_regular_thread(
             thread,
             start_routine,
             Abi::System { unwind: false },
diff --git a/src/tools/miri/src/stacked_borrows/diagnostics.rs b/src/tools/miri/src/stacked_borrows/diagnostics.rs
index 9970b79f8c7..023f6005419 100644
--- a/src/tools/miri/src/stacked_borrows/diagnostics.rs
+++ b/src/tools/miri/src/stacked_borrows/diagnostics.rs
@@ -5,7 +5,9 @@ use rustc_middle::mir::interpret::{alloc_range, AllocId, AllocRange};
 use rustc_span::{Span, SpanData};
 use rustc_target::abi::Size;
 
-use crate::stacked_borrows::{err_sb_ub, AccessKind, GlobalStateInner, Permission, ProtectorKind};
+use crate::stacked_borrows::{
+    err_sb_ub, AccessKind, GlobalStateInner, Permission, ProtectorKind, Stack,
+};
 use crate::*;
 
 use rustc_middle::mir::interpret::InterpError;
diff --git a/src/tools/miri/tests/many-seeds/scoped-thread-leak.rs b/src/tools/miri/tests/many-seeds/scoped-thread-leak.rs
new file mode 100644
index 00000000000..f28e43696f7
--- /dev/null
+++ b/src/tools/miri/tests/many-seeds/scoped-thread-leak.rs
@@ -0,0 +1,8 @@
+//! Regression test for https://github.com/rust-lang/miri/issues/2629
+use std::thread;
+
+fn main() {
+    thread::scope(|s| {
+        s.spawn(|| {});
+    });
+}
diff --git a/src/tools/miri/tests/pass/concurrency/scope.rs b/src/tools/miri/tests/pass/concurrency/scope.rs
new file mode 100644
index 00000000000..ce5d17f5f2d
--- /dev/null
+++ b/src/tools/miri/tests/pass/concurrency/scope.rs
@@ -0,0 +1,24 @@
+use std::thread;
+
+fn main() {
+    let mut a = vec![1, 2, 3];
+    let mut x = 0;
+
+    thread::scope(|s| {
+        s.spawn(|| {
+            // We can borrow `a` here.
+            let _s = format!("hello from the first scoped thread: {a:?}");
+        });
+        s.spawn(|| {
+            let _s = format!("hello from the second scoped thread");
+            // We can even mutably borrow `x` here,
+            // because no other threads are using it.
+            x += a[0] + a[2];
+        });
+        let _s = format!("hello from the main thread");
+    });
+
+    // After the scope, we can modify and access our variables again:
+    a.push(4);
+    assert_eq!(x, a.len());
+}